diff --git a/.gitattributes b/.gitattributes index a8436e9cfd0b720626494c5d9189390e63e5fedc..f1c8bcb4d217a19fa708460aee7021358fdc6bcb 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1 +1,3 @@ llm/ext_server/* linguist-vendored +* text=auto +*.go text eol=lf diff --git a/.gitignore b/.gitignore index d1c71d668ce401ff236a9b2807425d827e8b01a6..0d826ab614dd419d8ff5f7c1f2e0d41445e64494 100644 --- a/.gitignore +++ b/.gitignore @@ -12,8 +12,4 @@ ggml-metal.metal test_data *.crt llm/build -__debug_bin* -llm/llama.cpp/models/Llama* - -*build* -*pycache* \ No newline at end of file +__debug_bin* \ No newline at end of file diff --git a/.golangci.yaml b/.golangci.yaml index 7dec49deddbc5d14c35f61b399af0393ce710b13..c2c8b52b2e24a94dbd0d0f42dbcb2252e50bb779 100644 --- a/.golangci.yaml +++ b/.golangci.yaml @@ -7,11 +7,38 @@ linters: - bodyclose - containedctx - contextcheck + - errcheck - exportloopref + - gci - gocheckcompilerdirectives - # FIXME: for some reason this errors on windows - # - gofmt - # - goimports + - gofmt + - gofumpt + - gosimple + - govet + - ineffassign + - intrange + - makezero - misspell - nilerr + - nolintlint + - nosprintfhostport + - staticcheck + - tenv + - testifylint + - unconvert - unused + - usestdlibvars + - wastedassign + - whitespace +linters-settings: + gci: + sections: [standard, default, localmodule] +severity: + default-severity: error + rules: + - linters: + - gofmt + - goimports + - intrange + - usestdlibvars + severity: info diff --git a/Dockerfile b/Dockerfile index 72edef2a9144a67701e80e3975308bed95cf4f0b..c8efdd8a27621a8662770e974e1bc756f1f64aa9 100644 --- a/Dockerfile +++ b/Dockerfile @@ -1,8 +1,8 @@ -ARG GOLANG_VERSION=1.22.1 +ARG GOLANG_VERSION=1.22.5 ARG CMAKE_VERSION=3.22.1 # this CUDA_VERSION corresponds with the one specified in docs/gpu.md ARG CUDA_VERSION=11.3.1 -ARG ROCM_VERSION=6.0.2 +ARG ROCM_VERSION=6.1.2 # Copy the minimal context we need to run the generate scripts FROM scratch AS llm-code @@ -70,12 +70,12 @@ RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64 RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh -FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64 +FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64 ARG CMAKE_VERSION ARG GOLANG_VERSION COPY ./scripts/rh_linux_deps.sh / RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh -ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH +ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH COPY --from=llm-code / /go/src/github.com/ollama/ollama/ ARG OLLAMA_CUSTOM_CPU_DEFS ARG CGO_CFLAGS diff --git a/SECURITY.md b/SECURITY.md new file mode 100644 index 0000000000000000000000000000000000000000..d38bb7c4e2445baee1c4e5f42037838decf6b7c5 --- /dev/null +++ b/SECURITY.md @@ -0,0 +1,25 @@ +# Security + +The Ollama maintainer team takes security seriously and will actively work to resolve security issues. + +## Reporting a vulnerability + +If you discover a security vulnerability, please do not open a public issue. Instead, please report it by emailing hello@ollama.com. We ask that you give us sufficient time to investigate and address the vulnerability before disclosing it publicly. + +Please include the following details in your report: +- A description of the vulnerability +- Steps to reproduce the issue +- Your assessment of the potential impact +- Any possible mitigations + +## Security best practices + +While the maintainer team does their best to secure Ollama, users are encouraged to implement their own security best practices, such as: + +- Regularly updating to the latest version of Ollama +- Securing access to hosted instances of Ollama +- Monitoring systems for unusual activity + +## Contact + +For any other questions or concerns related to security, please contact us at hello@ollama.com diff --git a/api/client.go b/api/client.go index d50b397d3e5774f2dd8dda5ebe17745639a2a4e3..bbdf8202f1bd508a9bd28a4d72b442c9050de405 100644 --- a/api/client.go +++ b/api/client.go @@ -18,16 +18,14 @@ import ( "bytes" "context" "encoding/json" + "errors" "fmt" "io" - "net" "net/http" "net/url" - "os" "runtime" - "strconv" - "strings" + "github.com/ollama/ollama/envconfig" "github.com/ollama/ollama/format" "github.com/ollama/ollama/version" ) @@ -65,66 +63,12 @@ func checkError(resp *http.Response, body []byte) error { // If the variable is not specified, a default ollama host and port will be // used. func ClientFromEnvironment() (*Client, error) { - ollamaHost, err := GetOllamaHost() - if err != nil { - return nil, err - } - return &Client{ - base: &url.URL{ - Scheme: ollamaHost.Scheme, - Host: net.JoinHostPort(ollamaHost.Host, ollamaHost.Port), - }, + base: envconfig.Host(), http: http.DefaultClient, }, nil } -type OllamaHost struct { - Scheme string - Host string - Port string -} - -func GetOllamaHost() (OllamaHost, error) { - defaultPort := "11434" - - hostVar := os.Getenv("OLLAMA_HOST") - hostVar = strings.TrimSpace(strings.Trim(strings.TrimSpace(hostVar), "\"'")) - - scheme, hostport, ok := strings.Cut(hostVar, "://") - switch { - case !ok: - scheme, hostport = "http", hostVar - case scheme == "http": - defaultPort = "80" - case scheme == "https": - defaultPort = "443" - } - - // trim trailing slashes - hostport = strings.TrimRight(hostport, "/") - - host, port, err := net.SplitHostPort(hostport) - if err != nil { - host, port = "127.0.0.1", defaultPort - if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil { - host = ip.String() - } else if hostport != "" { - host = hostport - } - } - - if portNum, err := strconv.ParseInt(port, 10, 32); err != nil || portNum > 65535 || portNum < 0 { - return OllamaHost{}, ErrInvalidHostPort - } - - return OllamaHost{ - Scheme: scheme, - Host: host, - Port: port, - }, nil -} - func NewClient(base *url.URL, http *http.Client) *Client { return &Client{ base: base, @@ -229,7 +173,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f } if errorResponse.Error != "" { - return fmt.Errorf(errorResponse.Error) + return errors.New(errorResponse.Error) } if response.StatusCode >= http.StatusBadRequest { @@ -355,8 +299,8 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) { } // List running models. -func (c *Client) ListRunning(ctx context.Context) (*ListResponse, error) { - var lr ListResponse +func (c *Client) ListRunning(ctx context.Context) (*ProcessResponse, error) { + var lr ProcessResponse if err := c.do(ctx, http.MethodGet, "/api/ps", nil, &lr); err != nil { return nil, err } @@ -398,7 +342,16 @@ func (c *Client) Heartbeat(ctx context.Context) error { return nil } -// Embeddings generates embeddings from a model. +// Embed generates embeddings from a model. +func (c *Client) Embed(ctx context.Context, req *EmbedRequest) (*EmbedResponse, error) { + var resp EmbedResponse + if err := c.do(ctx, http.MethodPost, "/api/embed", req, &resp); err != nil { + return nil, err + } + return &resp, nil +} + +// Embeddings generates an embedding from a model. func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*EmbeddingResponse, error) { var resp EmbeddingResponse if err := c.do(ctx, http.MethodPost, "/api/embeddings", req, &resp); err != nil { diff --git a/api/client_test.go b/api/client_test.go index b2c51d006b06a24f6d39e83ff0743b82c765c9b4..23fe9334b9535b6421e2bb03cb873c4ad5ff69d0 100644 --- a/api/client_test.go +++ b/api/client_test.go @@ -1,11 +1,7 @@ package api import ( - "fmt" - "net" "testing" - - "github.com/stretchr/testify/assert" ) func TestClientFromEnvironment(t *testing.T) { @@ -46,40 +42,4 @@ func TestClientFromEnvironment(t *testing.T) { } }) } - - hostTestCases := map[string]*testCase{ - "empty": {value: "", expect: "127.0.0.1:11434"}, - "only address": {value: "1.2.3.4", expect: "1.2.3.4:11434"}, - "only port": {value: ":1234", expect: ":1234"}, - "address and port": {value: "1.2.3.4:1234", expect: "1.2.3.4:1234"}, - "hostname": {value: "example.com", expect: "example.com:11434"}, - "hostname and port": {value: "example.com:1234", expect: "example.com:1234"}, - "zero port": {value: ":0", expect: ":0"}, - "too large port": {value: ":66000", err: ErrInvalidHostPort}, - "too small port": {value: ":-1", err: ErrInvalidHostPort}, - "ipv6 localhost": {value: "[::1]", expect: "[::1]:11434"}, - "ipv6 world open": {value: "[::]", expect: "[::]:11434"}, - "ipv6 no brackets": {value: "::1", expect: "[::1]:11434"}, - "ipv6 + port": {value: "[::1]:1337", expect: "[::1]:1337"}, - "extra space": {value: " 1.2.3.4 ", expect: "1.2.3.4:11434"}, - "extra quotes": {value: "\"1.2.3.4\"", expect: "1.2.3.4:11434"}, - "extra space+quotes": {value: " \" 1.2.3.4 \" ", expect: "1.2.3.4:11434"}, - "extra single quotes": {value: "'1.2.3.4'", expect: "1.2.3.4:11434"}, - } - - for k, v := range hostTestCases { - t.Run(k, func(t *testing.T) { - t.Setenv("OLLAMA_HOST", v.value) - - oh, err := GetOllamaHost() - if err != v.err { - t.Fatalf("expected %s, got %s", v.err, err) - } - - if err == nil { - host := net.JoinHostPort(oh.Host, oh.Port) - assert.Equal(t, v.expect, host, fmt.Sprintf("%s: expected %s, got %s", k, v.expect, host)) - } - }) - } } diff --git a/api/types.go b/api/types.go index 4195a7c5823a2705b4ec1cfac2d8a6764b7aba05..2f5a94241240b2846c898a975dd5569cba473093 100644 --- a/api/types.go +++ b/api/types.go @@ -2,7 +2,6 @@ package api import ( "encoding/json" - "errors" "fmt" "log/slog" "math" @@ -48,6 +47,9 @@ type GenerateRequest struct { // Prompt is the textual prompt to send to the model. Prompt string `json:"prompt"` + // Suffix is the text that comes after the inserted text. + Suffix string `json:"suffix"` + // System overrides the model's default system message/prompt. System string `json:"system"` @@ -98,17 +100,85 @@ type ChatRequest struct { // followin the request. KeepAlive *Duration `json:"keep_alive,omitempty"` + // Tools is an optional list of tools the model has access to. + Tools `json:"tools,omitempty"` + // Options lists model-specific options. Options map[string]interface{} `json:"options"` } +type Tools []Tool + +func (t Tools) String() string { + bts, _ := json.Marshal(t) + return string(bts) +} + +func (t Tool) String() string { + bts, _ := json.Marshal(t) + return string(bts) +} + // Message is a single message in a chat sequence. The message contains the // role ("system", "user", or "assistant"), the content and an optional list // of images. type Message struct { - Role string `json:"role"` - Content string `json:"content"` - Images []ImageData `json:"images,omitempty"` + Role string `json:"role"` + Content string `json:"content"` + Images []ImageData `json:"images,omitempty"` + ToolCalls []ToolCall `json:"tool_calls,omitempty"` +} + +func (m *Message) UnmarshalJSON(b []byte) error { + type Alias Message + var a Alias + if err := json.Unmarshal(b, &a); err != nil { + return err + } + + *m = Message(a) + m.Role = strings.ToLower(m.Role) + return nil +} + +type ToolCall struct { + Function ToolCallFunction `json:"function"` +} + +type ToolCallFunction struct { + Name string `json:"name"` + Arguments ToolCallFunctionArguments `json:"arguments"` +} + +type ToolCallFunctionArguments map[string]any + +func (t *ToolCallFunctionArguments) String() string { + bts, _ := json.Marshal(t) + return string(bts) +} + +type Tool struct { + Type string `json:"type"` + Function ToolFunction `json:"function"` +} + +type ToolFunction struct { + Name string `json:"name"` + Description string `json:"description"` + Parameters struct { + Type string `json:"type"` + Required []string `json:"required"` + Properties map[string]struct { + Type string `json:"type"` + Description string `json:"description"` + Enum []string `json:"enum,omitempty"` + } `json:"properties"` + } `json:"parameters"` +} + +func (t *ToolFunction) String() string { + bts, _ := json.Marshal(t) + return string(bts) } // ChatResponse is the response returned by [Client.Chat]. Its fields are @@ -144,6 +214,7 @@ type Options struct { NumPredict int `json:"num_predict,omitempty"` TopK int `json:"top_k,omitempty"` TopP float32 `json:"top_p,omitempty"` + MinP float32 `json:"min_p,omitempty"` TFSZ float32 `json:"tfs_z,omitempty"` TypicalP float32 `json:"typical_p,omitempty"` RepeatLastN int `json:"repeat_last_n,omitempty"` @@ -160,18 +231,45 @@ type Options struct { // Runner options which must be set when the model is loaded into memory type Runner struct { - UseNUMA bool `json:"numa,omitempty"` - NumCtx int `json:"num_ctx,omitempty"` - NumBatch int `json:"num_batch,omitempty"` - NumGPU int `json:"num_gpu,omitempty"` - MainGPU int `json:"main_gpu,omitempty"` - LowVRAM bool `json:"low_vram,omitempty"` - F16KV bool `json:"f16_kv,omitempty"` - LogitsAll bool `json:"logits_all,omitempty"` - VocabOnly bool `json:"vocab_only,omitempty"` - UseMMap bool `json:"use_mmap,omitempty"` - UseMLock bool `json:"use_mlock,omitempty"` - NumThread int `json:"num_thread,omitempty"` + NumCtx int `json:"num_ctx,omitempty"` + NumBatch int `json:"num_batch,omitempty"` + NumGPU int `json:"num_gpu,omitempty"` + MainGPU int `json:"main_gpu,omitempty"` + LowVRAM bool `json:"low_vram,omitempty"` + F16KV bool `json:"f16_kv,omitempty"` + LogitsAll bool `json:"logits_all,omitempty"` + VocabOnly bool `json:"vocab_only,omitempty"` + UseMMap *bool `json:"use_mmap,omitempty"` + UseMLock bool `json:"use_mlock,omitempty"` + NumThread int `json:"num_thread,omitempty"` +} + +// EmbedRequest is the request passed to [Client.Embed]. +type EmbedRequest struct { + // Model is the model name. + Model string `json:"model"` + + // Input is the input to embed. + Input any `json:"input"` + + // KeepAlive controls how long the model will stay loaded in memory following + // this request. + KeepAlive *Duration `json:"keep_alive,omitempty"` + + Truncate *bool `json:"truncate,omitempty"` + + // Options lists model-specific options. + Options map[string]interface{} `json:"options"` +} + +// EmbedResponse is the response from [Client.Embed]. +type EmbedResponse struct { + Model string `json:"model"` + Embeddings [][]float32 `json:"embeddings"` + + TotalDuration time.Duration `json:"total_duration,omitempty"` + LoadDuration time.Duration `json:"load_duration,omitempty"` + PromptEvalCount int `json:"prompt_eval_count,omitempty"` } // EmbeddingRequest is the request passed to [Client.Embeddings]. @@ -220,9 +318,12 @@ type DeleteRequest struct { // ShowRequest is the request passed to [Client.Show]. type ShowRequest struct { - Model string `json:"model"` - System string `json:"system"` + Model string `json:"model"` + System string `json:"system"` + + // Template is deprecated Template string `json:"template"` + Verbose bool `json:"verbose"` Options map[string]interface{} `json:"options"` @@ -232,13 +333,16 @@ type ShowRequest struct { // ShowResponse is the response returned from [Client.Show]. type ShowResponse struct { - License string `json:"license,omitempty"` - Modelfile string `json:"modelfile,omitempty"` - Parameters string `json:"parameters,omitempty"` - Template string `json:"template,omitempty"` - System string `json:"system,omitempty"` - Details ModelDetails `json:"details,omitempty"` - Messages []Message `json:"messages,omitempty"` + License string `json:"license,omitempty"` + Modelfile string `json:"modelfile,omitempty"` + Parameters string `json:"parameters,omitempty"` + Template string `json:"template,omitempty"` + System string `json:"system,omitempty"` + Details ModelDetails `json:"details,omitempty"` + Messages []Message `json:"messages,omitempty"` + ModelInfo map[string]any `json:"model_info,omitempty"` + ProjectorInfo map[string]any `json:"projector_info,omitempty"` + ModifiedAt time.Time `json:"modified_at,omitempty"` } // CopyRequest is the request passed to [Client.Copy]. @@ -282,19 +386,40 @@ type PushRequest struct { // ListResponse is the response from [Client.List]. type ListResponse struct { - Models []ModelResponse `json:"models"` + Models []ListModelResponse `json:"models"` } -// ModelResponse is a single model description in [ListResponse]. -type ModelResponse struct { +// ProcessResponse is the response from [Client.Process]. +type ProcessResponse struct { + Models []ProcessModelResponse `json:"models"` +} + +// ListModelResponse is a single model description in [ListResponse]. +type ListModelResponse struct { Name string `json:"name"` Model string `json:"model"` - ModifiedAt time.Time `json:"modified_at,omitempty"` + ModifiedAt time.Time `json:"modified_at"` Size int64 `json:"size"` Digest string `json:"digest"` Details ModelDetails `json:"details,omitempty"` - ExpiresAt time.Time `json:"expires_at,omitempty"` - SizeVRAM int64 `json:"size_vram,omitempty"` +} + +// ProcessModelResponse is a single model description in [ProcessResponse]. +type ProcessModelResponse struct { + Name string `json:"name"` + Model string `json:"model"` + Size int64 `json:"size"` + Digest string `json:"digest"` + Details ModelDetails `json:"details,omitempty"` + ExpiresAt time.Time `json:"expires_at"` + SizeVRAM int64 `json:"size_vram"` +} + +type RetrieveModelResponse struct { + Id string `json:"id"` + Object string `json:"object"` + Created int64 `json:"created"` + OwnedBy string `json:"owned_by"` } type TokenResponse struct { @@ -306,7 +431,7 @@ type GenerateResponse struct { // Model is the model name that generated the response. Model string `json:"model"` - //CreatedAt is the timestamp of the response. + // CreatedAt is the timestamp of the response. CreatedAt time.Time `json:"created_at"` // Response is the textual response itself. @@ -363,8 +488,6 @@ func (m *Metrics) Summary() { } } -var ErrInvalidHostPort = errors.New("invalid port specified in OLLAMA_HOST") - func (opts *Options) FromMap(m map[string]interface{}) error { valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct typeOpts := reflect.TypeOf(opts).Elem() // types of the fields in the options struct @@ -381,7 +504,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error { for key, val := range m { opt, ok := jsonOpts[key] if !ok { - slog.Warn("invalid option provided", "option", opt.Name) + slog.Warn("invalid option provided", "option", key) continue } @@ -437,6 +560,17 @@ func (opts *Options) FromMap(m map[string]interface{}) error { slice[i] = str } field.Set(reflect.ValueOf(slice)) + case reflect.Pointer: + var b bool + if field.Type() == reflect.TypeOf(&b) { + val, ok := val.(bool) + if !ok { + return fmt.Errorf("option %q must be of type boolean", key) + } + field.Set(reflect.ValueOf(&val)) + } else { + return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type()) + } default: return fmt.Errorf("unknown type loading config params: %v", field.Kind()) } @@ -479,8 +613,7 @@ func DefaultOptions() Options { LowVRAM: false, F16KV: true, UseMLock: false, - UseMMap: true, - UseNUMA: false, + UseMMap: nil, }, } } @@ -576,6 +709,17 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) { case reflect.Slice: // TODO: only string slices are supported right now out[key] = vals + case reflect.Pointer: + var b bool + if field.Type() == reflect.TypeOf(&b) { + boolVal, err := strconv.ParseBool(vals[0]) + if err != nil { + return nil, fmt.Errorf("invalid bool value %s", vals) + } + out[key] = &boolVal + } else { + return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key) + } default: return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key) } diff --git a/api/types_test.go b/api/types_test.go index cfe1331f7d4bf8a795e273b9f5145a5004913393..a9de5a9a916a4e8c8f691e9bdc019d3660975863 100644 --- a/api/types_test.go +++ b/api/types_test.go @@ -2,6 +2,7 @@ package api import ( "encoding/json" + "errors" "math" "testing" "time" @@ -72,13 +73,13 @@ func TestDurationMarshalUnmarshal(t *testing.T) { }, { "positive duration", - time.Duration(42 * time.Second), - time.Duration(42 * time.Second), + 42 * time.Second, + 42 * time.Second, }, { "another positive duration", - time.Duration(42 * time.Minute), - time.Duration(42 * time.Minute), + 42 * time.Minute, + 42 * time.Minute, }, { "zero duration", @@ -105,3 +106,128 @@ func TestDurationMarshalUnmarshal(t *testing.T) { }) } } + +func TestUseMmapParsingFromJSON(t *testing.T) { + tr := true + fa := false + tests := []struct { + name string + req string + exp *bool + }{ + { + name: "Undefined", + req: `{ }`, + exp: nil, + }, + { + name: "True", + req: `{ "use_mmap": true }`, + exp: &tr, + }, + { + name: "False", + req: `{ "use_mmap": false }`, + exp: &fa, + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + var oMap map[string]interface{} + err := json.Unmarshal([]byte(test.req), &oMap) + require.NoError(t, err) + opts := DefaultOptions() + err = opts.FromMap(oMap) + require.NoError(t, err) + assert.Equal(t, test.exp, opts.UseMMap) + }) + } +} + +func TestUseMmapFormatParams(t *testing.T) { + tr := true + fa := false + tests := []struct { + name string + req map[string][]string + exp *bool + err error + }{ + { + name: "True", + req: map[string][]string{ + "use_mmap": {"true"}, + }, + exp: &tr, + err: nil, + }, + { + name: "False", + req: map[string][]string{ + "use_mmap": {"false"}, + }, + exp: &fa, + err: nil, + }, + { + name: "Numeric True", + req: map[string][]string{ + "use_mmap": {"1"}, + }, + exp: &tr, + err: nil, + }, + { + name: "Numeric False", + req: map[string][]string{ + "use_mmap": {"0"}, + }, + exp: &fa, + err: nil, + }, + { + name: "invalid string", + req: map[string][]string{ + "use_mmap": {"foo"}, + }, + exp: nil, + err: errors.New("invalid bool value [foo]"), + }, + } + + for _, test := range tests { + t.Run(test.name, func(t *testing.T) { + resp, err := FormatParams(test.req) + require.Equal(t, test.err, err) + respVal, ok := resp["use_mmap"] + if test.exp != nil { + assert.True(t, ok, "resp: %v", resp) + assert.Equal(t, *test.exp, *respVal.(*bool)) + } + }) + } +} + +func TestMessage_UnmarshalJSON(t *testing.T) { + tests := []struct { + input string + expected string + }{ + {`{"role": "USER", "content": "Hello!"}`, "user"}, + {`{"role": "System", "content": "Initialization complete."}`, "system"}, + {`{"role": "assistant", "content": "How can I help you?"}`, "assistant"}, + {`{"role": "TOOl", "content": "Access granted."}`, "tool"}, + } + + for _, test := range tests { + var msg Message + if err := json.Unmarshal([]byte(test.input), &msg); err != nil { + t.Errorf("Unexpected error: %v", err) + } + + if msg.Role != test.expected { + t.Errorf("role not lowercased: got %v, expected %v", msg.Role, test.expected) + } + } +} diff --git a/app/lifecycle/getstarted_nonwindows.go b/app/lifecycle/getstarted_nonwindows.go index c36d14c09c053cf3f12d0a7aab8a801be369a846..2af87ab9262f7438fb2aca060b7cb60c15bd066d 100644 --- a/app/lifecycle/getstarted_nonwindows.go +++ b/app/lifecycle/getstarted_nonwindows.go @@ -2,8 +2,8 @@ package lifecycle -import "fmt" +import "errors" func GetStarted() error { - return fmt.Errorf("GetStarted not implemented") + return errors.New("not implemented") } diff --git a/app/lifecycle/getstarted_windows.go b/app/lifecycle/getstarted_windows.go index 092c3c17f092dd7b42351ae3561f4fb1e0975815..f39dc31c0033328bc0642fe6085c47c1d1f520e1 100644 --- a/app/lifecycle/getstarted_windows.go +++ b/app/lifecycle/getstarted_windows.go @@ -34,7 +34,6 @@ func GetStarted() error { Sys: &syscall.SysProcAttr{CreationFlags: CREATE_NEW_CONSOLE, HideWindow: false}, } proc, err := os.StartProcess(args[0], args, attrs) - if err != nil { return fmt.Errorf("unable to start getting started shell %w", err) } diff --git a/app/lifecycle/logging.go b/app/lifecycle/logging.go index df2597a83d61e25576fdcb503b7951fc44df416d..9985fc3f8f276355c66088f5b9bd8b87f886b228 100644 --- a/app/lifecycle/logging.go +++ b/app/lifecycle/logging.go @@ -5,6 +5,8 @@ import ( "log/slog" "os" "path/filepath" + "strconv" + "strings" "github.com/ollama/ollama/envconfig" ) @@ -12,7 +14,7 @@ import ( func InitLogging() { level := slog.LevelInfo - if envconfig.Debug { + if envconfig.Debug() { level = slog.LevelDebug } @@ -24,7 +26,8 @@ func InitLogging() { logFile = os.Stderr // TODO - write one-line to the app.log file saying we're running in console mode to help avoid confusion } else { - logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755) + rotateLogs(AppLogFile) + logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0o755) if err != nil { slog.Error(fmt.Sprintf("failed to create server log %v", err)) return @@ -46,3 +49,32 @@ func InitLogging() { slog.Info("ollama app started") } + +func rotateLogs(logFile string) { + if _, err := os.Stat(logFile); os.IsNotExist(err) { + return + } + index := strings.LastIndex(logFile, ".") + pre := logFile[:index] + post := "." + logFile[index+1:] + for i := LogRotationCount; i > 0; i-- { + older := pre + "-" + strconv.Itoa(i) + post + newer := pre + "-" + strconv.Itoa(i-1) + post + if i == 1 { + newer = pre + post + } + if _, err := os.Stat(newer); err == nil { + if _, err := os.Stat(older); err == nil { + err := os.Remove(older) + if err != nil { + slog.Warn("Failed to remove older log", "older", older, "error", err) + continue + } + } + err := os.Rename(newer, older) + if err != nil { + slog.Warn("Failed to rotate log", "older", older, "newer", newer, "error", err) + } + } + } +} diff --git a/app/lifecycle/logging_nonwindows.go b/app/lifecycle/logging_nonwindows.go index 50b3a638c63a597ef2b06ac781da137d568280c5..205e47d774c2479891e7bad0953f9ac448f37257 100644 --- a/app/lifecycle/logging_nonwindows.go +++ b/app/lifecycle/logging_nonwindows.go @@ -5,5 +5,5 @@ package lifecycle import "log/slog" func ShowLogs() { - slog.Warn("ShowLogs not yet implemented") + slog.Warn("not implemented") } diff --git a/app/lifecycle/logging_test.go b/app/lifecycle/logging_test.go new file mode 100644 index 0000000000000000000000000000000000000000..8d5cdf6e7928f324a50fc51188df5bb78821e6e9 --- /dev/null +++ b/app/lifecycle/logging_test.go @@ -0,0 +1,44 @@ +package lifecycle + +import ( + "os" + "path/filepath" + "strconv" + "testing" + + "github.com/stretchr/testify/assert" + "github.com/stretchr/testify/require" +) + +func TestRotateLogs(t *testing.T) { + logDir := t.TempDir() + logFile := filepath.Join(logDir, "testlog.log") + + // No log exists + rotateLogs(logFile) + + require.NoError(t, os.WriteFile(logFile, []byte("1"), 0o644)) + assert.FileExists(t, logFile) + // First rotation + rotateLogs(logFile) + assert.FileExists(t, filepath.Join(logDir, "testlog-1.log")) + assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log")) + assert.NoFileExists(t, logFile) + + // Should be a no-op without a new log + rotateLogs(logFile) + assert.FileExists(t, filepath.Join(logDir, "testlog-1.log")) + assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log")) + assert.NoFileExists(t, logFile) + + for i := 2; i <= LogRotationCount+1; i++ { + require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0o644)) + assert.FileExists(t, logFile) + rotateLogs(logFile) + assert.NoFileExists(t, logFile) + for j := 1; j < i; j++ { + assert.FileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(j)+".log")) + } + assert.NoFileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(i+1)+".log")) + } +} diff --git a/app/lifecycle/paths.go b/app/lifecycle/paths.go index e4f2dbd95b544a77d6602cd292ab5f3a50bb3c31..4d9f4c5a1ed6bceedc00831dc4572fc214889014 100644 --- a/app/lifecycle/paths.go +++ b/app/lifecycle/paths.go @@ -16,11 +16,12 @@ var ( AppDir = "/opt/Ollama" AppDataDir = "/opt/Ollama" // TODO - should there be a distinct log dir? - UpdateStageDir = "/tmp" - AppLogFile = "/tmp/ollama_app.log" - ServerLogFile = "/tmp/ollama.log" - UpgradeLogFile = "/tmp/ollama_update.log" - Installer = "OllamaSetup.exe" + UpdateStageDir = "/tmp" + AppLogFile = "/tmp/ollama_app.log" + ServerLogFile = "/tmp/ollama.log" + UpgradeLogFile = "/tmp/ollama_update.log" + Installer = "OllamaSetup.exe" + LogRotationCount = 5 ) func init() { @@ -69,7 +70,6 @@ func init() { slog.Error(fmt.Sprintf("create ollama dir %s: %v", AppDataDir, err)) } } - } else if runtime.GOOS == "darwin" { // TODO AppName += ".app" diff --git a/app/lifecycle/server.go b/app/lifecycle/server.go index 3c11edb8384d355d625df2a309bb5196b2bf5a8c..37957399c9a802d62221da2e7719a8cb0e84caf8 100644 --- a/app/lifecycle/server.go +++ b/app/lifecycle/server.go @@ -15,7 +15,7 @@ import ( ) func getCLIFullPath(command string) string { - cmdPath := "" + var cmdPath string appExe, err := os.Executable() if err == nil { cmdPath = filepath.Join(filepath.Dir(appExe), command) @@ -54,8 +54,8 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) { return nil, fmt.Errorf("failed to spawn server stderr pipe: %w", err) } - // TODO - rotation - logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755) + rotateLogs(ServerLogFile) + logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0o755) if err != nil { return nil, fmt.Errorf("failed to create server log: %w", err) } @@ -65,7 +65,6 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) { if err != nil { if !errors.Is(err, os.ErrNotExist) { return nil, fmt.Errorf("stat ollama server log dir %s: %v", logDir, err) - } if err := os.MkdirAll(logDir, 0o755); err != nil { diff --git a/app/lifecycle/server_windows.go b/app/lifecycle/server_windows.go index cd4244ff9c55a7344039fb00507c410301b2fb44..5f9fe1246c8d9793771cafd2d833b83461a322dc 100644 --- a/app/lifecycle/server_windows.go +++ b/app/lifecycle/server_windows.go @@ -24,7 +24,8 @@ func terminate(cmd *exec.Cmd) error { if err != nil { return err } - defer dll.Release() // nolint: errcheck + //nolint:errcheck + defer dll.Release() pid := cmd.Process.Pid @@ -73,7 +74,8 @@ func isProcessExited(pid int) (bool, error) { if err != nil { return false, fmt.Errorf("failed to open process: %v", err) } - defer windows.CloseHandle(hProcess) // nolint: errcheck + //nolint:errcheck + defer windows.CloseHandle(hProcess) var exitCode uint32 err = windows.GetExitCodeProcess(hProcess, &exitCode) diff --git a/app/lifecycle/updater.go b/app/lifecycle/updater.go index 243bbf22355cc415c02880bbdb028bc323f60074..4d3c7d8dcfdc1b998750fbb14180c9d898afc119 100644 --- a/app/lifecycle/updater.go +++ b/app/lifecycle/updater.go @@ -15,6 +15,7 @@ import ( "path" "path/filepath" "runtime" + "strconv" "strings" "time" @@ -46,7 +47,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) { query.Add("os", runtime.GOOS) query.Add("arch", runtime.GOARCH) query.Add("version", version.Version) - query.Add("ts", fmt.Sprintf("%d", time.Now().Unix())) + query.Add("ts", strconv.FormatInt(time.Now().Unix(), 10)) nonce, err := auth.NewNonce(rand.Reader, 16) if err != nil { @@ -78,7 +79,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) { } defer resp.Body.Close() - if resp.StatusCode == 204 { + if resp.StatusCode == http.StatusNoContent { slog.Debug("check update response 204 (current version is up to date)") return false, updateResp } @@ -87,7 +88,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) { slog.Warn(fmt.Sprintf("failed to read body response: %s", err)) } - if resp.StatusCode != 200 { + if resp.StatusCode != http.StatusOK { slog.Info(fmt.Sprintf("check update error %d - %.96s", resp.StatusCode, string(body))) return false, updateResp } @@ -114,7 +115,7 @@ func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error { if err != nil { return fmt.Errorf("error checking update: %w", err) } - if resp.StatusCode != 200 { + if resp.StatusCode != http.StatusOK { return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode) } resp.Body.Close() diff --git a/app/lifecycle/updater_nonwindows.go b/app/lifecycle/updater_nonwindows.go index 0f213b34f7b2086d888b7752d73368dc845077d1..1d2dda8016ded8d4c0645ed6317d99f382f025d8 100644 --- a/app/lifecycle/updater_nonwindows.go +++ b/app/lifecycle/updater_nonwindows.go @@ -4,9 +4,9 @@ package lifecycle import ( "context" - "fmt" + "errors" ) func DoUpgrade(cancel context.CancelFunc, done chan int) error { - return fmt.Errorf("DoUpgrade not yet implemented") + return errors.New("not implemented") } diff --git a/app/lifecycle/updater_windows.go b/app/lifecycle/updater_windows.go index 4053671a5d7eccb1ff623f8e909226026fbb6157..1d3830d4ef4de7342f6274d8d029c84a5d8e98b5 100644 --- a/app/lifecycle/updater_windows.go +++ b/app/lifecycle/updater_windows.go @@ -2,6 +2,7 @@ package lifecycle import ( "context" + "errors" "fmt" "log/slog" "os" @@ -15,7 +16,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error { return fmt.Errorf("failed to lookup downloads: %s", err) } if len(files) == 0 { - return fmt.Errorf("no update downloads found") + return errors.New("no update downloads found") } else if len(files) > 1 { // Shouldn't happen slog.Warn(fmt.Sprintf("multiple downloads found, using first one %v", files)) @@ -64,7 +65,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error { } } else { // TODO - some details about why it didn't start, or is this a pedantic error case? - return fmt.Errorf("installer process did not start") + return errors.New("installer process did not start") } // TODO should we linger for a moment and check to make sure it's actually running by checking the pid? diff --git a/app/ollama.iss b/app/ollama.iss index 9dc61abbf4fdd0624756fed7b48829f86b669652..dc6178f7bb34564770c2fed495122dddc3c81e6c 100644 --- a/app/ollama.iss +++ b/app/ollama.iss @@ -88,10 +88,15 @@ DialogFontSize=12 [Files] Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit -Source: "..\dist\windows-{#ARCH}\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion +#if DirExists("..\dist\windows-amd64\cuda") + Source: "..\dist\windows-amd64\cuda\*"; DestDir: "{app}\cuda\"; Flags: ignoreversion recursesubdirs +#endif +#if DirExists("..\dist\windows-amd64\oneapi") + Source: "..\dist\windows-amd64\oneapi\*"; DestDir: "{app}\oneapi\"; Flags: ignoreversion recursesubdirs +#endif #if DirExists("..\dist\windows-amd64\rocm") Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs #endif @@ -122,6 +127,10 @@ Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\models" Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\history" ; NOTE: if the user has a custom OLLAMA_MODELS it will be preserved +[InstallDelete] +Type: filesandordirs; Name: "{%TEMP}\ollama*" +Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama" + [Messages] WizardReady=Ollama Windows Preview ReadyLabel1=%nLet's get you up and running with your own large language models. @@ -129,7 +138,7 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi ;FinishedHeadingLabel=Run your first model -;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n ollama run llama3 +;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n ollama run llama3.1 ;ClickFinish=%n [Registry] diff --git a/app/ollama_welcome.ps1 b/app/ollama_welcome.ps1 index 9af37a461158033e1fae55efca13234a1efe831d..46777a3a6a8e3eb40f86822ec61c80e0b0da309c 100644 --- a/app/ollama_welcome.ps1 +++ b/app/ollama_welcome.ps1 @@ -4,5 +4,5 @@ write-host "Welcome to Ollama!" write-host "" write-host "Run your first model:" write-host "" -write-host "`tollama run llama3" +write-host "`tollama run llama3.1" write-host "" \ No newline at end of file diff --git a/app/store/store.go b/app/store/store.go index 13a75a60aac45aefa5bef6b8d8551a25e4d6747a..b743e8a88ae76fdb001111cf9d2631ddbef8ad1b 100644 --- a/app/store/store.go +++ b/app/store/store.go @@ -29,7 +29,6 @@ func GetID() string { initStore() } return store.ID - } func GetFirstTimeRun() bool { diff --git a/app/tray/tray_nonwindows.go b/app/tray/tray_nonwindows.go index ae5572b23f91d6fa94b08c74a7424debac59d932..a03d233eaa9e22b05b2279c45611babf958c4de6 100644 --- a/app/tray/tray_nonwindows.go +++ b/app/tray/tray_nonwindows.go @@ -3,11 +3,11 @@ package tray import ( - "fmt" + "errors" "github.com/ollama/ollama/app/tray/commontray" ) func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) { - return nil, fmt.Errorf("NOT IMPLEMENTED YET") + return nil, errors.New("not implemented") } diff --git a/app/tray/wintray/eventloop.go b/app/tray/wintray/eventloop.go index a0af978732c8aa5f120f9dc84c33f5c61743124c..157828a365a7960b1bdc9ada270f0905162b6421 100644 --- a/app/tray/wintray/eventloop.go +++ b/app/tray/wintray/eventloop.go @@ -11,9 +11,7 @@ import ( "golang.org/x/sys/windows" ) -var ( - quitOnce sync.Once -) +var quitOnce sync.Once func (t *winTray) Run() { nativeLoop() @@ -47,7 +45,6 @@ func nativeLoop() { default: pTranslateMessage.Call(uintptr(unsafe.Pointer(m))) //nolint:errcheck pDispatchMessage.Call(uintptr(unsafe.Pointer(m))) //nolint:errcheck - } } } @@ -160,8 +157,8 @@ func (t *winTray) wndProc(hWnd windows.Handle, message uint32, wParam, lParam ui lResult, _, _ = pDefWindowProc.Call( uintptr(hWnd), uintptr(message), - uintptr(wParam), - uintptr(lParam), + wParam, + lParam, ) } return diff --git a/app/tray/wintray/tray.go b/app/tray/wintray/tray.go index 69d4487de90086dabdba73842e2c2b0598713cf8..ccd087a17acdab01310227f55440319b3bbdf6ac 100644 --- a/app/tray/wintray/tray.go +++ b/app/tray/wintray/tray.go @@ -13,8 +13,9 @@ import ( "sync" "unsafe" - "github.com/ollama/ollama/app/tray/commontray" "golang.org/x/sys/windows" + + "github.com/ollama/ollama/app/tray/commontray" ) // Helpful sources: https://github.com/golang/exp/blob/master/shiny/driver/internal/win32 @@ -186,7 +187,7 @@ func (t *winTray) initInstance() error { t.muNID.Lock() defer t.muNID.Unlock() t.nid = ¬ifyIconData{ - Wnd: windows.Handle(t.window), + Wnd: t.window, ID: 100, Flags: NIF_MESSAGE, CallbackMessage: t.wmSystrayMessage, @@ -197,7 +198,6 @@ func (t *winTray) initInstance() error { } func (t *winTray) createMenu() error { - menuHandle, _, err := pCreatePopupMenu.Call() if menuHandle == 0 { return err @@ -246,7 +246,7 @@ func (t *winTray) addOrUpdateMenuItem(menuItemId uint32, parentId uint32, title mi := menuItemInfo{ Mask: MIIM_FTYPE | MIIM_STRING | MIIM_ID | MIIM_STATE, Type: MFT_STRING, - ID: uint32(menuItemId), + ID: menuItemId, TypeData: titlePtr, Cch: uint32(len(title)), } @@ -302,11 +302,10 @@ func (t *winTray) addOrUpdateMenuItem(menuItemId uint32, parentId uint32, title } func (t *winTray) addSeparatorMenuItem(menuItemId, parentId uint32) error { - mi := menuItemInfo{ Mask: MIIM_FTYPE | MIIM_ID | MIIM_STATE, Type: MFT_SEPARATOR, - ID: uint32(menuItemId), + ID: menuItemId, } mi.Size = uint32(unsafe.Sizeof(mi)) @@ -416,7 +415,7 @@ func iconBytesToFilePath(iconBytes []byte) (string, error) { iconFilePath := filepath.Join(os.TempDir(), "ollama_temp_icon_"+dataHash) if _, err := os.Stat(iconFilePath); os.IsNotExist(err) { - if err := os.WriteFile(iconFilePath, iconBytes, 0644); err != nil { + if err := os.WriteFile(iconFilePath, iconBytes, 0o644); err != nil { return "", err } } @@ -426,7 +425,6 @@ func iconBytesToFilePath(iconBytes []byte) (string, error) { // Loads an image from file and shows it in tray. // Shell_NotifyIcon: https://msdn.microsoft.com/en-us/library/windows/desktop/bb762159(v=vs.85).aspx func (t *winTray) setIcon(src string) error { - h, err := t.loadIconFrom(src) if err != nil { return err @@ -444,7 +442,6 @@ func (t *winTray) setIcon(src string) error { // Loads an image from file to be shown in tray or menu item. // LoadImage: https://msdn.microsoft.com/en-us/library/windows/desktop/ms648045(v=vs.85).aspx func (t *winTray) loadIconFrom(src string) (windows.Handle, error) { - // Save and reuse handles of loaded images t.muLoadedImages.RLock() h, ok := t.loadedImages[src] diff --git a/auth/auth.go b/auth/auth.go index 026b2a2c71a2ce6e457329e9bc200ccbd29ec38e..e1d854124709a0c27f1e342ac4c38e21cae077c0 100644 --- a/auth/auth.go +++ b/auth/auth.go @@ -5,6 +5,7 @@ import ( "context" "crypto/rand" "encoding/base64" + "errors" "fmt" "io" "log/slog" @@ -78,7 +79,7 @@ func Sign(ctx context.Context, bts []byte) (string, error) { publicKey := ssh.MarshalAuthorizedKey(privateKey.PublicKey()) parts := bytes.Split(publicKey, []byte(" ")) if len(parts) < 2 { - return "", fmt.Errorf("malformed public key") + return "", errors.New("malformed public key") } signedData, err := privateKey.Sign(rand.Reader, bts) diff --git a/cmd/cmd.go b/cmd/cmd.go index b285f83cbfed9ba386a78bda0e7706feffa9462c..d47db65b32914b14a1df8a49c6b2673713428414 100644 --- a/cmd/cmd.go +++ b/cmd/cmd.go @@ -20,6 +20,7 @@ import ( "path/filepath" "regexp" "runtime" + "slices" "strings" "syscall" "time" @@ -29,7 +30,6 @@ import ( "github.com/olekukonko/tablewriter" "github.com/spf13/cobra" "golang.org/x/crypto/ssh" - "golang.org/x/exp/slices" "golang.org/x/term" "github.com/ollama/ollama/api" @@ -162,9 +162,6 @@ func tempZipFiles(path string) (string, error) { } defer tempfile.Close() - zipfile := zip.NewWriter(tempfile) - defer zipfile.Close() - detectContentType := func(path string) (string, error) { f, err := os.Open(path) if err != nil { @@ -233,6 +230,9 @@ func tempZipFiles(path string) (string, error) { files = append(files, tks...) } + zipfile := zip.NewWriter(tempfile) + defer zipfile.Close() + for _, file := range files { f, err := os.Open(file) if err != nil { @@ -287,38 +287,12 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er } func RunHandler(cmd *cobra.Command, args []string) error { - client, err := api.ClientFromEnvironment() - if err != nil { - return err - } - - name := args[0] - - // check if the model exists on the server - show, err := client.Show(cmd.Context(), &api.ShowRequest{Name: name}) - var statusError api.StatusError - switch { - case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound: - if err := PullHandler(cmd, []string{name}); err != nil { - return err - } - - show, err = client.Show(cmd.Context(), &api.ShowRequest{Name: name}) - if err != nil { - return err - } - case err != nil: - return err - } - interactive := true opts := runOptions{ - Model: args[0], - WordWrap: os.Getenv("TERM") == "xterm-256color", - Options: map[string]interface{}{}, - MultiModal: slices.Contains(show.Details.Families, "clip"), - ParentModel: show.Details.ParentModel, + Model: args[0], + WordWrap: os.Getenv("TERM") == "xterm-256color", + Options: map[string]interface{}{}, } format, err := cmd.Flags().GetString("format") @@ -362,11 +336,53 @@ func RunHandler(cmd *cobra.Command, args []string) error { } opts.WordWrap = !nowrap - if !interactive { - return generate(cmd, opts) + // Fill out the rest of the options based on information about the + // model. + client, err := api.ClientFromEnvironment() + if err != nil { + return err + } + + name := args[0] + info, err := func() (*api.ShowResponse, error) { + showReq := &api.ShowRequest{Name: name} + info, err := client.Show(cmd.Context(), showReq) + var se api.StatusError + if errors.As(err, &se) && se.StatusCode == http.StatusNotFound { + if err := PullHandler(cmd, []string{name}); err != nil { + return nil, err + } + return client.Show(cmd.Context(), &api.ShowRequest{Name: name}) + } + return info, err + }() + if err != nil { + return err } - return generateInteractive(cmd, opts) + opts.MultiModal = slices.Contains(info.Details.Families, "clip") + opts.ParentModel = info.Details.ParentModel + + if interactive { + if err := loadModel(cmd, &opts); err != nil { + return err + } + + for _, msg := range info.Messages { + switch msg.Role { + case "user": + fmt.Printf(">>> %s\n", msg.Content) + case "assistant": + state := &displayResponseState{} + displayResponse(msg.Content, opts.WordWrap, state) + fmt.Println() + fmt.Println() + } + } + + return generateInteractive(cmd, opts) + } + return generate(cmd, opts) } func errFromUnknownKey(unknownKeyErr error) error { @@ -579,10 +595,6 @@ func ShowHandler(cmd *cobra.Command, args []string) error { return err } - if len(args) != 1 { - return errors.New("missing model name") - } - license, errLicense := cmd.Flags().GetBool("license") modelfile, errModelfile := cmd.Flags().GetBool("modelfile") parameters, errParams := cmd.Flags().GetBool("parameters") @@ -625,8 +637,6 @@ func ShowHandler(cmd *cobra.Command, args []string) error { if flagsSet > 1 { return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified") - } else if flagsSet == 0 { - return errors.New("one of '--license', '--modelfile', '--parameters', '--system', or '--template' must be specified") } req := api.ShowRequest{Name: args[0]} @@ -635,22 +645,141 @@ func ShowHandler(cmd *cobra.Command, args []string) error { return err } - switch showType { - case "license": - fmt.Println(resp.License) - case "modelfile": - fmt.Println(resp.Modelfile) - case "parameters": - fmt.Println(resp.Parameters) - case "system": - fmt.Println(resp.System) - case "template": - fmt.Println(resp.Template) + if flagsSet == 1 { + switch showType { + case "license": + fmt.Println(resp.License) + case "modelfile": + fmt.Println(resp.Modelfile) + case "parameters": + fmt.Println(resp.Parameters) + case "system": + fmt.Println(resp.System) + case "template": + fmt.Println(resp.Template) + } + + return nil } + showInfo(resp) + return nil } +func showInfo(resp *api.ShowResponse) { + arch := resp.ModelInfo["general.architecture"].(string) + + modelData := [][]string{ + {"arch", arch}, + {"parameters", resp.Details.ParameterSize}, + {"quantization", resp.Details.QuantizationLevel}, + {"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))}, + {"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))}, + } + + mainTableData := [][]string{ + {"Model"}, + {renderSubTable(modelData, false)}, + } + + if resp.ProjectorInfo != nil { + projectorData := [][]string{ + {"arch", "clip"}, + {"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))}, + } + + if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok { + projectorData = append(projectorData, []string{"projector type", projectorType.(string)}) + } + + projectorData = append(projectorData, + []string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))}, + []string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))}, + ) + + mainTableData = append(mainTableData, + []string{"Projector"}, + []string{renderSubTable(projectorData, false)}, + ) + } + + if resp.Parameters != "" { + mainTableData = append(mainTableData, []string{"Parameters"}, []string{formatParams(resp.Parameters)}) + } + + if resp.System != "" { + mainTableData = append(mainTableData, []string{"System"}, []string{renderSubTable(twoLines(resp.System), true)}) + } + + if resp.License != "" { + mainTableData = append(mainTableData, []string{"License"}, []string{renderSubTable(twoLines(resp.License), true)}) + } + + table := tablewriter.NewWriter(os.Stdout) + table.SetAutoWrapText(false) + table.SetBorder(false) + table.SetAlignment(tablewriter.ALIGN_LEFT) + + for _, v := range mainTableData { + table.Append(v) + } + + table.Render() +} + +func renderSubTable(data [][]string, file bool) string { + var buf bytes.Buffer + table := tablewriter.NewWriter(&buf) + table.SetAutoWrapText(!file) + table.SetBorder(false) + table.SetNoWhiteSpace(true) + table.SetTablePadding("\t") + table.SetAlignment(tablewriter.ALIGN_LEFT) + + for _, v := range data { + table.Append(v) + } + + table.Render() + + renderedTable := buf.String() + lines := strings.Split(renderedTable, "\n") + for i, line := range lines { + lines[i] = "\t" + line + } + + return strings.Join(lines, "\n") +} + +func twoLines(s string) [][]string { + lines := strings.Split(s, "\n") + res := [][]string{} + + count := 0 + for _, line := range lines { + line = strings.TrimSpace(line) + if line != "" { + count++ + res = append(res, []string{line}) + if count == 2 { + return res + } + } + } + return res +} + +func formatParams(s string) string { + lines := strings.Split(s, "\n") + table := [][]string{} + + for _, line := range lines { + table = append(table, strings.Fields(line)) + } + return renderSubTable(table, false) +} + func CopyHandler(cmd *cobra.Command, args []string) error { client, err := api.ClientFromEnvironment() if err != nil { @@ -729,7 +858,6 @@ type runOptions struct { WordWrap bool Format string System string - Template string Images []api.ImageData Options map[string]interface{} MultiModal bool @@ -746,7 +874,6 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState) if wordWrap && termWidth >= 10 { for _, ch := range content { if state.lineLength+1 > termWidth-5 { - if runewidth.StringWidth(state.wordBuffer) > termWidth-10 { fmt.Printf("%s%c", state.wordBuffer, ch) state.wordBuffer = "" @@ -924,7 +1051,6 @@ func generate(cmd *cobra.Command, opts runOptions) error { Images: opts.Images, Format: opts.Format, System: opts.System, - Template: opts.Template, Options: opts.Options, KeepAlive: opts.KeepAlive, } @@ -961,17 +1087,11 @@ func generate(cmd *cobra.Command, opts runOptions) error { } func RunServer(cmd *cobra.Command, _ []string) error { - // retrieve the OLLAMA_HOST environment variable - ollamaHost, err := api.GetOllamaHost() - if err != nil { - return err - } - if err := initializeKeypair(); err != nil { return err } - ln, err := net.Listen("tcp", net.JoinHostPort(ollamaHost.Host, ollamaHost.Port)) + ln, err := net.Listen("tcp", envconfig.Host().Host) if err != nil { return err } @@ -1030,24 +1150,6 @@ func initializeKeypair() error { return nil } -//nolint:unused -func waitForServer(ctx context.Context, client *api.Client) error { - // wait for the server to start - timeout := time.After(5 * time.Second) - tick := time.Tick(500 * time.Millisecond) - for { - select { - case <-timeout: - return errors.New("timed out waiting for server to start") - case <-tick: - if err := client.Heartbeat(ctx); err == nil { - return nil // server has started - } - } - } - -} - func checkServerHeartbeat(cmd *cobra.Command, _ []string) error { client, err := api.ClientFromEnvironment() if err != nil { @@ -1058,7 +1160,7 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error { return err } if err := startApp(cmd.Context(), client); err != nil { - return fmt.Errorf("could not connect to ollama app, is it running?") + return errors.New("could not connect to ollama app, is it running?") } } return nil @@ -1254,10 +1356,10 @@ func NewCLI() *cobra.Command { envVars["OLLAMA_NUM_PARALLEL"], envVars["OLLAMA_NOPRUNE"], envVars["OLLAMA_ORIGINS"], + envVars["OLLAMA_SCHED_SPREAD"], envVars["OLLAMA_TMPDIR"], envVars["OLLAMA_FLASH_ATTENTION"], envVars["OLLAMA_LLM_LIBRARY"], - envVars["OLLAMA_MAX_VRAM"], }) default: appendEnvDocs(cmd, envs) diff --git a/cmd/interactive.go b/cmd/interactive.go index c055df0ea21fe436da24e004270c6161c8574e68..4462cf29dfde0303127535d97d6f118d3868b8db 100644 --- a/cmd/interactive.go +++ b/cmd/interactive.go @@ -1,6 +1,7 @@ package cmd import ( + "cmp" "errors" "fmt" "io" @@ -8,14 +9,15 @@ import ( "os" "path/filepath" "regexp" - "sort" + "slices" "strings" "github.com/spf13/cobra" - "golang.org/x/exp/slices" + "golang.org/x/exp/maps" "github.com/ollama/ollama/api" "github.com/ollama/ollama/envconfig" + "github.com/ollama/ollama/parser" "github.com/ollama/ollama/progress" "github.com/ollama/ollama/readline" "github.com/ollama/ollama/types/errtypes" @@ -27,74 +29,29 @@ const ( MultilineNone MultilineState = iota MultilinePrompt MultilineSystem - MultilineTemplate ) func loadModel(cmd *cobra.Command, opts *runOptions) error { - client, err := api.ClientFromEnvironment() - if err != nil { - return err - } - p := progress.NewProgress(os.Stderr) defer p.StopAndClear() spinner := progress.NewSpinner("") p.Add("", spinner) - showReq := api.ShowRequest{Name: opts.Model} - showResp, err := client.Show(cmd.Context(), &showReq) + client, err := api.ClientFromEnvironment() if err != nil { return err } - opts.MultiModal = slices.Contains(showResp.Details.Families, "clip") - opts.ParentModel = showResp.Details.ParentModel - - if len(showResp.Messages) > 0 { - opts.Messages = append(opts.Messages, showResp.Messages...) - } chatReq := &api.ChatRequest{ - Model: opts.Model, - Messages: []api.Message{}, + Model: opts.Model, + KeepAlive: opts.KeepAlive, } - if opts.KeepAlive != nil { - chatReq.KeepAlive = opts.KeepAlive - } - - err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error { - p.StopAndClear() - if len(opts.Messages) > 0 { - for _, msg := range opts.Messages { - switch msg.Role { - case "user": - fmt.Printf(">>> %s\n", msg.Content) - case "assistant": - state := &displayResponseState{} - displayResponse(msg.Content, opts.WordWrap, state) - fmt.Println() - fmt.Println() - } - } - } - return nil - }) - if err != nil { - return err - } - - return nil + return client.Chat(cmd.Context(), chatReq, func(api.ChatResponse) error { return nil }) } func generateInteractive(cmd *cobra.Command, opts runOptions) error { - opts.Messages = make([]api.Message, 0) - - err := loadModel(cmd, &opts) - if err != nil { - return err - } - usage := func() { fmt.Fprintln(os.Stderr, "Available Commands:") fmt.Fprintln(os.Stderr, " /set Set session variables") @@ -119,7 +76,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { fmt.Fprintln(os.Stderr, "Available Commands:") fmt.Fprintln(os.Stderr, " /set parameter ... Set a parameter") fmt.Fprintln(os.Stderr, " /set system Set system message") - fmt.Fprintln(os.Stderr, " /set template Set prompt template") fmt.Fprintln(os.Stderr, " /set history Enable history") fmt.Fprintln(os.Stderr, " /set nohistory Disable history") fmt.Fprintln(os.Stderr, " /set wordwrap Enable wordwrap") @@ -165,6 +121,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { fmt.Fprintln(os.Stderr, " /set parameter num_predict Max number of tokens to predict") fmt.Fprintln(os.Stderr, " /set parameter top_k Pick from top k num of tokens") fmt.Fprintln(os.Stderr, " /set parameter top_p Pick token based on sum of probabilities") + fmt.Fprintln(os.Stderr, " /set parameter min_p Pick token based on top token probability * min_p") fmt.Fprintln(os.Stderr, " /set parameter num_ctx Set the context size") fmt.Fprintln(os.Stderr, " /set parameter temperature Set creativity level") fmt.Fprintln(os.Stderr, " /set parameter repeat_penalty How strongly to penalize repetitions") @@ -184,7 +141,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { return err } - if envconfig.NoHistory { + if envconfig.NoHistory() { scanner.HistoryDisable() } @@ -229,10 +186,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { opts.Messages = append(opts.Messages, api.Message{Role: "system", Content: opts.System}) fmt.Println("Set system message.") sb.Reset() - case MultilineTemplate: - opts.Template = sb.String() - fmt.Println("Set prompt template.") - sb.Reset() } multiline = MultilineNone @@ -351,17 +304,13 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { } fmt.Printf("Set parameter '%s' to '%s'\n", args[2], strings.Join(params, ", ")) opts.Options[args[2]] = fp[args[2]] - case "system", "template": + case "system": if len(args) < 3 { usageSet() continue } - if args[1] == "system" { - multiline = MultilineSystem - } else if args[1] == "template" { - multiline = MultilineTemplate - } + multiline = MultilineSystem line := strings.Join(args[2:], " ") line, ok := strings.CutPrefix(line, `"""`) @@ -381,23 +330,17 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { continue } - if args[1] == "system" { - opts.System = sb.String() // for display in modelfile - newMessage := api.Message{Role: "system", Content: sb.String()} - // Check if the slice is not empty and the last message is from 'system' - if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" { - // Replace the last message - opts.Messages[len(opts.Messages)-1] = newMessage - } else { - opts.Messages = append(opts.Messages, newMessage) - } - fmt.Println("Set system message.") - sb.Reset() - } else if args[1] == "template" { - opts.Template = sb.String() - fmt.Println("Set prompt template.") - sb.Reset() + opts.System = sb.String() // for display in modelfile + newMessage := api.Message{Role: "system", Content: sb.String()} + // Check if the slice is not empty and the last message is from 'system' + if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" { + // Replace the last message + opts.Messages[len(opts.Messages)-1] = newMessage + } else { + opts.Messages = append(opts.Messages, newMessage) } + fmt.Println("Set system message.") + sb.Reset() sb.Reset() continue @@ -416,10 +359,9 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { return err } req := &api.ShowRequest{ - Name: opts.Model, - System: opts.System, - Template: opts.Template, - Options: opts.Options, + Name: opts.Model, + System: opts.System, + Options: opts.Options, } resp, err := client.Show(cmd.Context(), req) if err != nil { @@ -429,15 +371,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { switch args[1] { case "info": - fmt.Println("Model details:") - if len(resp.Details.Families) > 0 { - fmt.Printf("Family %s\n", strings.Join(resp.Details.Families, ", ")) - } else if resp.Details.Family != "" { - fmt.Printf("Family %s\n", resp.Details.Family) - } - fmt.Printf("Parameter Size %s\n", resp.Details.ParameterSize) - fmt.Printf("Quantization Level %s\n", resp.Details.QuantizationLevel) - fmt.Println("") + showInfo(resp) case "license": if resp.License == "" { fmt.Println("No license was specified for this model.") @@ -470,12 +404,9 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { fmt.Println("No system message was specified for this model.") } case "template": - switch { - case opts.Template != "": - fmt.Println(opts.Template + "\n") - case resp.Template != "": + if resp.Template != "" { fmt.Println(resp.Template) - default: + } else { fmt.Println("No prompt template was specified for this model.") } default: @@ -559,35 +490,35 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { } func buildModelfile(opts runOptions) string { - var mf strings.Builder - model := opts.ParentModel - if model == "" { - model = opts.Model - } - fmt.Fprintf(&mf, "FROM %s\n", model) - if opts.System != "" { - fmt.Fprintf(&mf, "SYSTEM \"\"\"%s\"\"\"\n", opts.System) - } + var f parser.File + f.Commands = append(f.Commands, parser.Command{Name: "model", Args: cmp.Or(opts.ParentModel, opts.Model)}) - if opts.Template != "" { - fmt.Fprintf(&mf, "TEMPLATE \"\"\"%s\"\"\"\n", opts.Template) + if opts.System != "" { + f.Commands = append(f.Commands, parser.Command{Name: "system", Args: opts.System}) } - keys := make([]string, 0) - for k := range opts.Options { - keys = append(keys, k) - } - sort.Strings(keys) + keys := maps.Keys(opts.Options) + slices.Sort(keys) for _, k := range keys { - fmt.Fprintf(&mf, "PARAMETER %s %v\n", k, opts.Options[k]) + v := opts.Options[k] + var cmds []parser.Command + switch t := v.(type) { + case []string: + for _, s := range t { + cmds = append(cmds, parser.Command{Name: k, Args: s}) + } + default: + cmds = append(cmds, parser.Command{Name: k, Args: fmt.Sprintf("%v", t)}) + } + + f.Commands = append(f.Commands, cmds...) } - fmt.Fprintln(&mf) for _, msg := range opts.Messages { - fmt.Fprintf(&mf, "MESSAGE %s \"\"\"%s\"\"\"\n", msg.Role, msg.Content) + f.Commands = append(f.Commands, parser.Command{Name: "message", Args: fmt.Sprintf("%s: %s", msg.Role, msg.Content)}) } - return mf.String() + return f.String() } func normalizeFilePath(fp string) string { @@ -673,7 +604,7 @@ func getImageData(filePath string) ([]byte, error) { // Check if the file size exceeds 100MB var maxSize int64 = 100 * 1024 * 1024 // 100MB in bytes if info.Size() > maxSize { - return nil, fmt.Errorf("file size exceeds maximum limit (100MB)") + return nil, errors.New("file size exceeds maximum limit (100MB)") } buf = make([]byte, info.Size()) diff --git a/cmd/interactive_test.go b/cmd/interactive_test.go index 8eedf72920f6f80037b87dea5049dc192931ba9b..bb7e0abaf88e59ee4ffc55e7d6f700a02fffc8e5 100644 --- a/cmd/interactive_test.go +++ b/cmd/interactive_test.go @@ -1,10 +1,9 @@ package cmd import ( - "bytes" "testing" - "text/template" + "github.com/google/go-cmp/cmp" "github.com/stretchr/testify/assert" "github.com/ollama/ollama/api" @@ -56,61 +55,53 @@ d:\path with\spaces\seven.svg inbetween7 c:\users\jdoe\eight.png inbetween8 func TestModelfileBuilder(t *testing.T) { opts := runOptions{ - Model: "hork", - System: "You are part horse and part shark, but all hork. Do horklike things", - Template: "This is a template.", + Model: "hork", + System: "You are part horse and part shark, but all hork. Do horklike things", Messages: []api.Message{ {Role: "user", Content: "Hey there hork!"}, {Role: "assistant", Content: "Yes it is true, I am half horse, half shark."}, }, - Options: map[string]interface{}{}, + Options: map[string]any{ + "temperature": 0.9, + "seed": 42, + "penalize_newline": false, + "stop": []string{"hi", "there"}, + }, } - opts.Options["temperature"] = 0.9 - opts.Options["seed"] = 42 - opts.Options["penalize_newline"] = false - opts.Options["stop"] = []string{"hi", "there"} - - mf := buildModelfile(opts) - expectedModelfile := `FROM {{.Model}} -SYSTEM """{{.System}}""" -TEMPLATE """{{.Template}}""" + t.Run("model", func(t *testing.T) { + expect := `FROM hork +SYSTEM You are part horse and part shark, but all hork. Do horklike things PARAMETER penalize_newline false PARAMETER seed 42 -PARAMETER stop [hi there] +PARAMETER stop hi +PARAMETER stop there PARAMETER temperature 0.9 - -MESSAGE user """Hey there hork!""" -MESSAGE assistant """Yes it is true, I am half horse, half shark.""" +MESSAGE user Hey there hork! +MESSAGE assistant Yes it is true, I am half horse, half shark. ` - tmpl, err := template.New("").Parse(expectedModelfile) - assert.Nil(t, err) - - var buf bytes.Buffer - err = tmpl.Execute(&buf, opts) - assert.Nil(t, err) - assert.Equal(t, buf.String(), mf) + actual := buildModelfile(opts) + if diff := cmp.Diff(expect, actual); diff != "" { + t.Errorf("mismatch (-want +got):\n%s", diff) + } + }) - opts.ParentModel = "horseshark" - mf = buildModelfile(opts) - expectedModelfile = `FROM {{.ParentModel}} -SYSTEM """{{.System}}""" -TEMPLATE """{{.Template}}""" + t.Run("parent model", func(t *testing.T) { + opts.ParentModel = "horseshark" + expect := `FROM horseshark +SYSTEM You are part horse and part shark, but all hork. Do horklike things PARAMETER penalize_newline false PARAMETER seed 42 -PARAMETER stop [hi there] +PARAMETER stop hi +PARAMETER stop there PARAMETER temperature 0.9 - -MESSAGE user """Hey there hork!""" -MESSAGE assistant """Yes it is true, I am half horse, half shark.""" +MESSAGE user Hey there hork! +MESSAGE assistant Yes it is true, I am half horse, half shark. ` - - tmpl, err = template.New("").Parse(expectedModelfile) - assert.Nil(t, err) - - var parentBuf bytes.Buffer - err = tmpl.Execute(&parentBuf, opts) - assert.Nil(t, err) - assert.Equal(t, parentBuf.String(), mf) + actual := buildModelfile(opts) + if diff := cmp.Diff(expect, actual); diff != "" { + t.Errorf("mismatch (-want +got):\n%s", diff) + } + }) } diff --git a/cmd/start.go b/cmd/start.go new file mode 100644 index 0000000000000000000000000000000000000000..0c4eed084bb5041583e371308835a5411272483e --- /dev/null +++ b/cmd/start.go @@ -0,0 +1,27 @@ +//go:build darwin || windows + +package cmd + +import ( + "context" + "errors" + "time" + + "github.com/ollama/ollama/api" +) + +func waitForServer(ctx context.Context, client *api.Client) error { + // wait for the server to start + timeout := time.After(5 * time.Second) + tick := time.Tick(500 * time.Millisecond) + for { + select { + case <-timeout: + return errors.New("timed out waiting for server to start") + case <-tick: + if err := client.Heartbeat(ctx); err == nil { + return nil // server has started + } + } + } +} diff --git a/cmd/start_darwin.go b/cmd/start_darwin.go index 82b09ad625ec75a9a2ef46a78c80f1fe70d81c29..1a9a1ae87aa7c8ccd8e6da1e31036c00d361c8ed 100644 --- a/cmd/start_darwin.go +++ b/cmd/start_darwin.go @@ -2,7 +2,7 @@ package cmd import ( "context" - "fmt" + "errors" "os" "os/exec" "strings" @@ -20,7 +20,7 @@ func startApp(ctx context.Context, client *api.Client) error { return err } if !strings.Contains(link, "Ollama.app") { - return fmt.Errorf("could not find ollama app") + return errors.New("could not find ollama app") } path := strings.Split(link, "Ollama.app") if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil { diff --git a/cmd/start_default.go b/cmd/start_default.go index c9d6137b9927988364b2a69b87622d481c8abfdc..5eabb28628f6c7f0eab8cecf8135ec9d8ed07305 100644 --- a/cmd/start_default.go +++ b/cmd/start_default.go @@ -4,11 +4,11 @@ package cmd import ( "context" - "fmt" + "errors" "github.com/ollama/ollama/api" ) func startApp(ctx context.Context, client *api.Client) error { - return fmt.Errorf("could not connect to ollama server, run 'ollama serve' to start it") + return errors.New("could not connect to ollama server, run 'ollama serve' to start it") } diff --git a/cmd/start_windows.go b/cmd/start_windows.go index 6024a23522dca13f25f281011eccdb41d61297be..5bca24331ff9a53190e83874fb269e4e04ea0e0b 100644 --- a/cmd/start_windows.go +++ b/cmd/start_windows.go @@ -31,7 +31,7 @@ func startApp(ctx context.Context, client *api.Client) error { // Finally look in the path appExe, err = exec.LookPath(AppName) if err != nil { - return fmt.Errorf("could not locate ollama app") + return errors.New("could not locate ollama app") } } } diff --git a/convert/convert.go b/convert/convert.go index e71a0ff3f533c953c657ee9f2d11f2e8faab2068..b9461e4fe339577e9fd782fdd1988c1c99425520 100644 --- a/convert/convert.go +++ b/convert/convert.go @@ -1,200 +1,122 @@ package convert import ( - "cmp" - "encoding/binary" "encoding/json" + "errors" "fmt" "io" + "io/fs" "log/slog" - "os" - "path/filepath" - "slices" - "strings" - "google.golang.org/protobuf/proto" - - "github.com/ollama/ollama/convert/sentencepiece" "github.com/ollama/ollama/llm" ) -const ( - _ int32 = iota - tokenTypeNormal - tokenTypeUnknown - tokenTypeControl - tokenTypeUserDefined - tokenTypeUnused - tokenTypeByte -) - -type Params struct { - Architectures []string `json:"architectures"` - VocabSize int `json:"vocab_size"` - HiddenSize int `json:"hidden_size"` // n_embd - HiddenLayers int `json:"num_hidden_layers"` // n_layer - ContextSize int `json:"max_position_embeddings"` - IntermediateSize int `json:"intermediate_size"` - AttentionHeads int `json:"num_attention_heads"` // n_head - KeyValHeads int `json:"num_key_value_heads"` - NormEPS float64 `json:"rms_norm_eps"` - BoSTokenID int `json:"bos_token_id"` - EoSTokenID int `json:"eos_token_id"` - HeadDimension int `json:"head_dim"` - PaddingTokenID int `json:"pad_token_id"` - RopeFrequencyBase float64 `json:"rope_theta"` - - Experts int `json:"num_local_experts"` - ExpertsUsed int `json:"num_experts_per_tok"` - - PreTokenizer string - - ByteOrder +type Parameters struct { + Architectures []string `json:"architectures"` + VocabSize uint32 `json:"vocab_size"` } -type ByteOrder interface { - binary.ByteOrder - binary.AppendByteOrder -} +func (Parameters) KV(t *Tokenizer) llm.KV { + kv := llm.KV{ + "general.file_type": uint32(1), + "general.quantization_version": uint32(2), + "tokenizer.ggml.pre": t.Pre, + "tokenizer.ggml.model": t.Vocabulary.Model, + "tokenizer.ggml.tokens": t.Vocabulary.Tokens, + "tokenizer.ggml.scores": t.Vocabulary.Scores, + "tokenizer.ggml.token_type": t.Vocabulary.Types, + } -type ModelArch interface { - GetTensors() error - LoadVocab() error - WriteGGUF(io.WriteSeeker) error -} + if t.Template != "" { + kv["tokenizer.chat_template"] = t.Template + } -type ModelFormat interface { - GetLayerName(string) (string, error) - GetTensors(string, *Params) ([]llm.Tensor, error) - GetParams(string) (*Params, error) - GetModelArch(string, string, *Params) (ModelArch, error) -} + for _, sv := range t.SpecialVocabulary { + kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID) + kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken + } -type ModelData struct { - Path string - Name string - Params *Params - Vocab *Vocab - Tensors []llm.Tensor - Format ModelFormat + return kv } -func GetModelFormat(dirname string) (ModelFormat, error) { - files, err := filepath.Glob(filepath.Join(dirname, "*")) - if err != nil { - return nil, err - } - - for _, fn := range files { - if strings.HasSuffix(fn, ".safetensors") { - return &SafetensorFormat{}, nil - } else if strings.HasSuffix(fn, ".bin") || strings.HasSuffix(fn, ".pth") { - slog.Debug("model is torch") - return &TorchFormat{}, nil - } +func (Parameters) specialTokenTypes() []string { + return []string{ + "bos", "eos", "unk", "sep", "pad", "cls", "mask", } +} - return nil, fmt.Errorf("couldn't determine model format") +func (Parameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error { + return llm.WriteGGUF(ws, kv, ts) } -// Details on gguf's tokenizer can be found at: -// https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#tokenizer -type Vocab struct { - Tokens []string - Scores []float32 - Types []int32 - Merges []string +type Converter interface { + // KV maps parameters to LLM key-values + KV(*Tokenizer) llm.KV + // Tensors maps input tensors to LLM tensors. Model specific modifications can be done here. + Tensors([]Tensor) []llm.Tensor + + // tensorName returns the LLM tensor name for a specific input name + tensorName(string) string + // specialTokenTypes returns any special token types the model uses + specialTokenTypes() []string + writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error } -func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) { - slog.Info(fmt.Sprintf("reading vocab from %s", filepath.Join(dirpath, "tokenizer.model"))) - in, err := os.ReadFile(filepath.Join(dirpath, "tokenizer.model")) +// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations +// and files it finds in the input path. +// Supported input model formats include safetensors. +// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model. +func Convert(fsys fs.FS, ws io.WriteSeeker) error { + bts, err := fs.ReadFile(fsys, "config.json") if err != nil { - return nil, err + return err } - // To regenerate sentencepiece from the protobufs use: - // protoc -I=./ --go_out=./ sentencepiece_model.proto - modelProto := &sentencepiece.ModelProto{} - if err := proto.Unmarshal(in, modelProto); err != nil { - return nil, err + var p Parameters + if err := json.Unmarshal(bts, &p); err != nil { + return err } - v := &Vocab{ - Tokens: make([]string, 0), - Scores: make([]float32, 0), - Types: make([]int32, 0), + if len(p.Architectures) < 1 { + return errors.New("unknown architecture") } - pieces := modelProto.GetPieces() - for _, p := range pieces { - v.Tokens = append(v.Tokens, p.GetPiece()) - v.Scores = append(v.Scores, p.GetScore()) - t := p.GetType() - switch t { - case sentencepiece.ModelProto_SentencePiece_UNKNOWN: - case sentencepiece.ModelProto_SentencePiece_CONTROL: - case sentencepiece.ModelProto_SentencePiece_UNUSED: - case sentencepiece.ModelProto_SentencePiece_BYTE: - default: - t = sentencepiece.ModelProto_SentencePiece_NORMAL - } - v.Types = append(v.Types, int32(t)) - } - - slog.Info(fmt.Sprintf("vocab size: %d", len(v.Tokens))) - - // add any additional tokens - addIn, err := os.ReadFile(filepath.Join(dirpath, "added_tokens.json")) - if os.IsNotExist(err) { - return v, nil - } else if err != nil { - return nil, err - } - - slog.Info("reading user defined tokens") - - var extraTokenData map[string]int - if err := json.Unmarshal(addIn, &extraTokenData); err != nil { - return nil, err + var conv Converter + switch p.Architectures[0] { + case "LlamaForCausalLM", "MistralForCausalLM": + conv = &llama{} + case "MixtralForCausalLM": + conv = &mixtral{} + case "GemmaForCausalLM": + conv = &gemma{} + default: + return errors.New("unsupported architecture") } - type token struct { - key string - pos int + if err := json.Unmarshal(bts, conv); err != nil { + return err } - extraTokens := make([]token, 0) - for k, id := range extraTokenData { - extraTokens = append(extraTokens, token{k, id}) + t, err := parseTokenizer(fsys, conv.specialTokenTypes()) + if err != nil { + return err } - slices.SortFunc(extraTokens, func(a, b token) int { - return cmp.Compare(a.pos, b.pos) - }) - - numToks := len(v.Tokens) - - for cnt, t := range extraTokens { - // the token id should match the specific index for the total number of tokens - if t.pos != cnt+numToks { - return nil, fmt.Errorf("token ID '%d' for '%s' doesn't match total token size", t.pos, t.key) + if vocabSize := int(p.VocabSize); vocabSize > len(t.Vocabulary.Tokens) { + slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", p.VocabSize, "actual", len(t.Vocabulary.Tokens)) + for i := range vocabSize - len(t.Vocabulary.Tokens) { + t.Vocabulary.Tokens = append(t.Vocabulary.Tokens, fmt.Sprintf("[PAD%d]", i)) + t.Vocabulary.Scores = append(t.Vocabulary.Scores, -1) + t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined) } - v.Tokens = append(v.Tokens, t.key) - v.Scores = append(v.Scores, -1000.0) - v.Types = append(v.Types, tokenTypeUserDefined) + } else { + slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens)) } - slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens))) - - if params.VocabSize > len(v.Tokens) { - missingTokens := params.VocabSize - len(v.Tokens) - slog.Warn(fmt.Sprintf("vocab is missing %d tokens", missingTokens)) - for cnt := 0; cnt < missingTokens; cnt++ { - v.Tokens = append(v.Tokens, fmt.Sprintf("", cnt+1)) - v.Scores = append(v.Scores, -1) - v.Types = append(v.Types, tokenTypeUserDefined) - } + + ts, err := parseTensors(fsys) + if err != nil { + return err } - return v, nil + return conv.writeFile(ws, conv.KV(t), conv.Tensors(ts)) } diff --git a/convert/convert_gemma.go b/convert/convert_gemma.go new file mode 100644 index 0000000000000000000000000000000000000000..9213e15768d9acecfb944e3c3268bf9c26473cdc --- /dev/null +++ b/convert/convert_gemma.go @@ -0,0 +1,103 @@ +package convert + +import ( + "strings" + + "github.com/pdevine/tensor" + "github.com/pdevine/tensor/native" + + "github.com/ollama/ollama/llm" +) + +type gemma struct { + Parameters + MaxPositionEmbeddings uint32 `json:"max_position_embeddings"` + HiddenSize uint32 `json:"hidden_size"` + HiddenLayers uint32 `json:"num_hidden_layers"` + IntermediateSize uint32 `json:"intermediate_size"` + NumAttentionHeads uint32 `json:"num_attention_heads"` + NumKeyValueHeads uint32 `json:"num_key_value_heads"` + RMSNormEPS float32 `json:"rms_norm_eps"` + HeadDim uint32 `json:"head_dim"` +} + +var _ Converter = (*gemma)(nil) + +func (p *gemma) KV(t *Tokenizer) llm.KV { + kv := p.Parameters.KV(t) + kv["general.architecture"] = "gemma" + kv["general.name"] = "gemma" + kv["gemma.context_length"] = p.MaxPositionEmbeddings + kv["gemma.embedding_length"] = p.HiddenSize + kv["gemma.block_count"] = p.HiddenLayers + kv["gemma.feed_forward_length"] = p.IntermediateSize + kv["gemma.attention.head_count"] = p.NumAttentionHeads + kv["gemma.attention.head_count_kv"] = p.NumKeyValueHeads + kv["gemma.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS + kv["gemma.attention.key_length"] = p.HeadDim + kv["gemma.attention.value_length"] = p.HeadDim + kv["tokenizer.ggml.eot_token_id"] = uint32(107) + kv["tokenizer.ggml.middle_token_id"] = uint32(68) + kv["tokenizer.ggml.prefix_token_id"] = uint32(67) + kv["tokenizer.ggml.suffix_token_id"] = uint32(69) + return kv +} + +func (p *gemma) Tensors(ts []Tensor) []llm.Tensor { + var out []llm.Tensor + for _, t := range ts { + name := p.tensorName(t.Name()) + if strings.HasSuffix(name, "_norm.weight") { + t.SetRepacker(p.addOne) + } + + out = append(out, llm.Tensor{ + Name: name, + Kind: t.Kind(), + Shape: t.Shape(), + WriterTo: t, + }) + } + + return out +} + +func (p *gemma) tensorName(n string) string { + return strings.NewReplacer( + "model.embed_tokens", "token_embd", + "model.norm", "output_norm", + "model.layers", "blk", + "input_layernorm", "attn_norm", + "self_attn.q_proj", "attn_q", + "self_attn.k_proj", "attn_k", + "self_attn.v_proj", "attn_v", + "self_attn.o_proj", "attn_output", + "mlp.gate_proj", "ffn_gate", + "mlp.down_proj", "ffn_down", + "mlp.up_proj", "ffn_up", + "post_attention_layernorm", "ffn_norm", + "block_sparse_moe.gate", "ffn_inp", + ).Replace(n) +} + +func (*gemma) addOne(_ string, data []float32, shape []uint64) ([]float32, error) { + n := tensor.New(tensor.WithShape(int(shape[0])), tensor.WithBacking(data)) + ones := tensor.Ones(tensor.Float32, int(shape[0])) + + n, err := n.Add(ones) + if err != nil { + return nil, err + } + + ts, err := native.SelectF32(n, 0) + if err != nil { + return nil, err + } + + var f32s []float32 + for _, t := range ts { + f32s = append(f32s, t...) + } + + return f32s, nil +} diff --git a/convert/convert_llama.go b/convert/convert_llama.go new file mode 100644 index 0000000000000000000000000000000000000000..0383a85eb6973c9a11ef0d15e6eb9e7d1a509c20 --- /dev/null +++ b/convert/convert_llama.go @@ -0,0 +1,183 @@ +package convert + +import ( + "cmp" + "fmt" + "strings" + + "github.com/pdevine/tensor" + "github.com/pdevine/tensor/native" + + "github.com/ollama/ollama/llm" +) + +type llama struct { + Parameters + NLayers uint32 `json:"n_layers"` + NumHiddenLayers uint32 `json:"num_hidden_layers"` + NLayer uint32 `json:"n_layer"` + MaxPositionEmbeddings uint32 `json:"max_position_embeddings"` + NCtx uint32 `json:"n_ctx"` + HiddenSize uint32 `json:"hidden_size"` + NEmbd uint32 `json:"n_embd"` + IntermediateSize uint32 `json:"intermediate_size"` + NInner uint32 `json:"n_inner"` + NumAttentionHeads uint32 `json:"num_attention_heads"` + NHead uint32 `json:"n_head"` + NumKeyValueHeads uint32 `json:"num_key_value_heads"` + RopeTheta float32 `json:"rope_theta"` + RopeScaling struct { + Type string `json:"type"` + Factor float32 `json:"factor"` + } `json:"rope_scaling"` + RMSNormEPS float32 `json:"rms_norm_eps"` + LayerNormEPS float32 `json:"layer_norm_eps"` + LayerNormEpsilon float32 `json:"layer_norm_epsilon"` + NormEpsilon float32 `json:"norm_epsilon"` + HeadDim uint32 `json:"head_dim"` +} + +var _ Converter = (*llama)(nil) + +func (p *llama) KV(t *Tokenizer) llm.KV { + kv := p.Parameters.KV(t) + kv["general.architecture"] = "llama" + kv["general.name"] = "llama" + kv["llama.vocab_size"] = p.VocabSize + + kv["llama.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer) + + if contextLength := cmp.Or(p.MaxPositionEmbeddings, p.NCtx); contextLength > 0 { + kv["llama.context_length"] = contextLength + } + + if embeddingLength := cmp.Or(p.HiddenSize, p.NEmbd); embeddingLength > 0 { + kv["llama.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd) + } + + if feedForwardLength := cmp.Or(p.IntermediateSize, p.NInner); feedForwardLength > 0 { + kv["llama.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner) + } + + if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 { + kv["llama.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead) + kv["llama.rope.dimension_count"] = p.HiddenSize / headCount + } + + if p.RopeTheta > 0 { + kv["llama.rope.freq_base"] = p.RopeTheta + } + + if p.RopeScaling.Type == "linear" { + kv["llama.rope.scaling.type"] = p.RopeScaling.Type + kv["llama.rope.scaling.factor"] = p.RopeScaling.Factor + } + + if p.NumKeyValueHeads > 0 { + kv["llama.attention.head_count_kv"] = p.NumKeyValueHeads + } + + if p.RMSNormEPS > 0 { + kv["llama.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS + } + + if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon, p.NormEpsilon); layerNormEpsilon > 0 { + kv["llama.attention.layer_norm_epsilon"] = layerNormEpsilon + } + + if p.HeadDim > 0 { + kv["llama.attention.key_length"] = p.HeadDim + kv["llama.attention.value_length"] = p.HeadDim + } + + if len(t.Merges) > 0 { + kv["tokenizer.ggml.merges"] = t.Merges + } + + return kv +} + +func (p *llama) Tensors(ts []Tensor) []llm.Tensor { + var out []llm.Tensor + for _, t := range ts { + name := p.tensorName(t.Name()) + if strings.HasSuffix(name, "attn_q.weight") || + strings.HasSuffix(name, "attn_k.weight") { + t.SetRepacker(p.repack) + } + + out = append(out, llm.Tensor{ + Name: name, + Kind: t.Kind(), + Shape: t.Shape(), + WriterTo: t, + }) + } + + return out +} + +func (p *llama) tensorName(n string) string { + return strings.NewReplacer( + "lm_head", "output", + "model.embed_tokens", "token_embd", + "model.norm", "output_norm", + "model.layers", "blk", + "input_layernorm", "attn_norm", + "self_attn.q_proj", "attn_q", + "self_attn.k_proj", "attn_k", + "self_attn.v_proj", "attn_v", + "self_attn.o_proj", "attn_output", + "mlp.gate_proj", "ffn_gate", + "mlp.down_proj", "ffn_down", + "mlp.up_proj", "ffn_up", + "post_attention_layernorm", "ffn_norm", + // mixtral + "block_sparse_moe.gate", "ffn_gate_inp", + ).Replace(n) +} + +func (p *llama) repack(name string, data []float32, shape []uint64) ([]float32, error) { + var dims []int + for _, dim := range shape { + dims = append(dims, int(dim)) + } + + var heads uint32 + if strings.HasSuffix(name, "q_proj.weight") { + heads = p.NumAttentionHeads + } else if strings.HasSuffix(name, "k_proj.weight") { + heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads) + } else { + return nil, fmt.Errorf("unknown tensor for repack: %s", name) + } + + n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data)) + if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil { + return nil, err + } + + if err := n.T(0, 2, 1, 3); err != nil { + return nil, err + } + + if err := n.Reshape(dims...); err != nil { + return nil, err + } + + if err := n.Transpose(); err != nil { + return nil, err + } + + ts, err := native.SelectF32(n, 1) + if err != nil { + return nil, err + } + + var f32s []float32 + for _, t := range ts { + f32s = append(f32s, t...) + } + + return f32s, nil +} diff --git a/convert/convert_mixtral.go b/convert/convert_mixtral.go new file mode 100644 index 0000000000000000000000000000000000000000..3263a27b383d7c5eabfe78449ab57f4d34222b87 --- /dev/null +++ b/convert/convert_mixtral.go @@ -0,0 +1,89 @@ +package convert + +import ( + "fmt" + "io" + "slices" + "strings" + + "github.com/ollama/ollama/llm" +) + +type mixtral struct { + llama + NumLocalExperts uint32 `json:"num_local_experts"` + NumExpertsPerToken uint32 `json:"num_experts_per_tok"` +} + +var _ Converter = (*mixtral)(nil) + +func (p *mixtral) KV(t *Tokenizer) llm.KV { + kv := p.llama.KV(t) + + if p.NumLocalExperts > 0 { + kv["llama.expert_count"] = p.NumLocalExperts + } + + if p.NumExpertsPerToken > 0 { + kv["llama.expert_used_count"] = p.NumExpertsPerToken + } + + return kv +} + +func (p *mixtral) Tensors(ts []Tensor) []llm.Tensor { + oldnew := []string{ + "model.layers", "blk", + "w1", "ffn_gate_exps", + "w2", "ffn_down_exps", + "w3", "ffn_up_exps", + } + + for i := range p.NumLocalExperts { + oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".") + } + + // group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor + namer := strings.NewReplacer(oldnew...) + experts := make(map[string]experts) + + // merge experts into a single tensor while removing them from ts + ts = slices.DeleteFunc(ts, func(t Tensor) bool { + if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") { + return false + } + + name := namer.Replace(t.Name()) + experts[name] = append(experts[name], t) + return true + }) + + var out []llm.Tensor + for n, e := range experts { + // TODO(mxyng): sanity check experts + out = append(out, llm.Tensor{ + Name: n, + Kind: e[0].Kind(), + Shape: append([]uint64{uint64(len(e))}, e[0].Shape()...), + WriterTo: e, + }) + } + + return append(out, p.llama.Tensors(ts)...) +} + +type experts []Tensor + +func (e experts) WriteTo(w io.Writer) (int64, error) { + // TODO(mxyng): experts _should_ be numerically sorted by expert but this should check + for _, t := range e { + // the canonical merged experts tensor stacks all experts along a new, 0 axis, + // e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers + // this accomplishes the same thing by writing each expert tensor in sequence + if _, err := t.WriteTo(w); err != nil { + return 0, err + } + } + + return 0, nil +} diff --git a/convert/convert_test.go b/convert/convert_test.go index 6aa33a49c10da8394f53e73e8b9364b79c815945..88f38494ed8f0a14ee692bb3bb28c824ea7ff2dc 100644 --- a/convert/convert_test.go +++ b/convert/convert_test.go @@ -1,102 +1,127 @@ -//go:build slow - package convert import ( + "crypto/sha256" + "encoding/hex" + "encoding/json" + "flag" + "fmt" + "io" + "io/fs" + "log/slog" + "math" "os" "path/filepath" + "slices" "testing" + "golang.org/x/exp/maps" + "github.com/ollama/ollama/llm" ) -func convertFull(t *testing.T, p string) (llm.KV, llm.Tensors) { +func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) { t.Helper() - mf, err := GetModelFormat(p) + f, err := os.CreateTemp(t.TempDir(), "f16") if err != nil { t.Fatal(err) } + defer f.Close() - params, err := mf.GetParams(p) - if err != nil { + if err := Convert(fsys, f); err != nil { t.Fatal(err) } - arch, err := mf.GetModelArch("", p, params) + r, err := os.Open(f.Name()) if err != nil { t.Fatal(err) } + t.Cleanup(func() { r.Close() }) - if err := arch.LoadVocab(); err != nil { - t.Fatal(err) - } - - if err := arch.GetTensors(); err != nil { - t.Fatal(err) - } - - f, err := os.CreateTemp(t.TempDir(), "f16") + m, _, err := llm.DecodeGGML(r, math.MaxInt) if err != nil { t.Fatal(err) } - defer f.Close() - if err := arch.WriteGGUF(f); err != nil { + if _, err := r.Seek(0, io.SeekStart); err != nil { t.Fatal(err) } - r, err := os.Open(f.Name()) - if err != nil { - t.Fatal(err) - } - defer r.Close() - - m, _, err := llm.DecodeGGML(r) - if err != nil { - t.Fatal(err) - } + return r, m.KV(), m.Tensors() +} - return m.KV(), m.Tensors() +func TestMain(m *testing.M) { + var level slog.Level + flag.TextVar(&level, "level", slog.LevelInfo, "log level") + flag.Parse() + slog.SetLogLoggerLevel(level) + os.Exit(m.Run()) } func TestConvertFull(t *testing.T) { - cases := []struct { - path string - arch string - tensors int - layers int - }{ - {"Meta-Llama-3-8B-Instruct", "llama", 291, 35}, - {"Mistral-7B-Instruct-v0.2", "llama", 291, 35}, - {"Mixtral-8x7B-Instruct-v0.1", "llama", 291, 35}, - {"gemma-2b-it", "gemma", 164, 20}, + cases := []string{ + "Meta-Llama-3-8B-Instruct", + "Mistral-7B-Instruct-v0.2", + "Mixtral-8x7B-Instruct-v0.1", + "gemma-2b-it", } - for _, tt := range cases { - t.Run(tt.path, func(t *testing.T) { - p := filepath.Join("testdata", tt.path) - if _, err := os.Stat(p); err != nil { + for i := range cases { + tt := cases[i] + t.Run(tt, func(t *testing.T) { + t.Parallel() + + p := filepath.Join("testdata", tt) + if testing.Short() { + t.Skip("skipping in short mode") + } else if _, err := os.Stat(p); err != nil { t.Skipf("%s not found", p) } - kv, tensors := convertFull(t, p) + f, kv, tensors := convertFull(t, os.DirFS(p)) + actual := make(map[string]string) + for k, v := range kv { + if s, ok := v.(json.Marshaler); !ok { + actual[k] = fmt.Sprintf("%v", v) + } else { + bts, err := json.Marshal(s) + if err != nil { + t.Fatal(err) + } + + actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts)) + } + } + + for _, tensor := range tensors.Items { + sha256sum := sha256.New() + sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size())) + if _, err := io.Copy(sha256sum, sr); err != nil { + t.Fatal(err) + } - if kv.Architecture() != tt.arch { - t.Fatalf("expected llama, got %s", kv.Architecture()) + actual[tensor.Name] = hex.EncodeToString(sha256sum.Sum(nil)) } - if kv.FileType().String() != "F16" { - t.Fatalf("expected F16, got %s", kv.FileType()) + expectFile, err := os.Open(filepath.Join("testdata", fmt.Sprintf("%s.json", tt))) + if err != nil { + t.Fatal(err) } - if len(tensors) != tt.tensors { - t.Fatalf("expected %d tensors, got %d", tt.tensors, len(tensors)) + var expect map[string]string + if err := json.NewDecoder(expectFile).Decode(&expect); err != nil { + t.Fatal(err) } - layers := tensors.Layers() - if len(layers) != tt.layers { - t.Fatalf("expected %d layers, got %d", tt.layers, len(layers)) + keys := maps.Keys(expect) + slices.Sort(keys) + for _, k := range keys { + if v, ok := actual[k]; !ok { + t.Errorf("missing %s", k) + } else if v != expect[k] { + t.Errorf("unexpected %s: want %s, got %s", k, expect[k], v) + } } }) } diff --git a/convert/fs.go b/convert/fs.go new file mode 100644 index 0000000000000000000000000000000000000000..31132dbe72c389e52f21192508de5f6bb0e001c7 --- /dev/null +++ b/convert/fs.go @@ -0,0 +1,58 @@ +package convert + +import ( + "archive/zip" + "errors" + "io" + "io/fs" + "os" + "path/filepath" +) + +type ZipReader struct { + r *zip.Reader + p string + + // limit is the maximum size of a file that can be read directly + // from the zip archive. Files larger than this size will be extracted + limit int64 +} + +func NewZipReader(r *zip.Reader, p string, limit int64) fs.FS { + return &ZipReader{r, p, limit} +} + +func (z *ZipReader) Open(name string) (fs.File, error) { + r, err := z.r.Open(name) + if err != nil { + return nil, err + } + defer r.Close() + + if fi, err := r.Stat(); err != nil { + return nil, err + } else if fi.Size() < z.limit { + return r, nil + } + + if !filepath.IsLocal(name) { + return nil, zip.ErrInsecurePath + } + + n := filepath.Join(z.p, name) + if _, err := os.Stat(n); errors.Is(err, os.ErrNotExist) { + w, err := os.Create(n) + if err != nil { + return nil, err + } + defer w.Close() + + if _, err := io.Copy(w, r); err != nil { + return nil, err + } + } else if err != nil { + return nil, err + } + + return os.Open(n) +} diff --git a/convert/reader.go b/convert/reader.go new file mode 100644 index 0000000000000000000000000000000000000000..ce95208e1222d42dd38c2ec991d501c992c25d87 --- /dev/null +++ b/convert/reader.go @@ -0,0 +1,82 @@ +package convert + +import ( + "errors" + "io" + "io/fs" + "strings" +) + +type Tensor interface { + Name() string + Shape() []uint64 + Kind() uint32 + SetRepacker(repacker) + WriteTo(io.Writer) (int64, error) +} + +type tensorBase struct { + name string + shape []uint64 + repacker +} + +func (t tensorBase) Name() string { + return t.name +} + +func (t tensorBase) Shape() []uint64 { + return t.shape +} + +const ( + tensorKindF32 uint32 = iota + tensorKindF16 +) + +func (t tensorBase) Kind() uint32 { + if strings.HasSuffix(t.name, ".block_sparse_moe.gate.weight") { + return 0 + } + + switch len(t.shape) { + case 0: + panic("invalid tensor shape") + case 1: + return tensorKindF32 + default: + return tensorKindF16 + } +} + +func (t *tensorBase) SetRepacker(fn repacker) { + t.repacker = fn +} + +type repacker func(string, []float32, []uint64) ([]float32, error) + +func parseTensors(fsys fs.FS) ([]Tensor, error) { + patterns := []struct { + Pattern string + Func func(fs.FS, ...string) ([]Tensor, error) + }{ + {"model-*-of-*.safetensors", parseSafetensors}, + {"model.safetensors", parseSafetensors}, + {"pytorch_model-*-of-*.bin", parseTorch}, + {"pytorch_model.bin", parseTorch}, + {"consolidated.*.pth", parseTorch}, + } + + for _, pattern := range patterns { + matches, err := fs.Glob(fsys, pattern.Pattern) + if err != nil { + return nil, err + } + + if len(matches) > 0 { + return pattern.Func(fsys, matches...) + } + } + + return nil, errors.New("unknown tensor format") +} diff --git a/convert/reader_safetensors.go b/convert/reader_safetensors.go new file mode 100644 index 0000000000000000000000000000000000000000..42f902a5a1796d2dc118d979bf42d8fe873b44a8 --- /dev/null +++ b/convert/reader_safetensors.go @@ -0,0 +1,150 @@ +package convert + +import ( + "bytes" + "encoding/binary" + "encoding/json" + "fmt" + "io" + "io/fs" + "slices" + + "github.com/d4l3k/go-bfloat16" + "github.com/x448/float16" + "golang.org/x/exp/maps" +) + +type safetensorMetadata struct { + Type string `json:"dtype"` + Shape []uint64 `json:"shape"` + Offsets []int64 `json:"data_offsets"` +} + +func parseSafetensors(fsys fs.FS, ps ...string) ([]Tensor, error) { + var ts []Tensor + for _, p := range ps { + f, err := fsys.Open(p) + if err != nil { + return nil, err + } + defer f.Close() + + var n int64 + if err := binary.Read(f, binary.LittleEndian, &n); err != nil { + return nil, err + } + + b := bytes.NewBuffer(make([]byte, 0, n)) + if _, err = io.CopyN(b, f, n); err != nil { + return nil, err + } + + var headers map[string]safetensorMetadata + if err := json.NewDecoder(b).Decode(&headers); err != nil { + return nil, err + } + + keys := maps.Keys(headers) + slices.Sort(keys) + + for _, key := range keys { + if value := headers[key]; value.Type != "" { + ts = append(ts, safetensor{ + fs: fsys, + path: p, + dtype: value.Type, + offset: safetensorsPad(n, value.Offsets[0]), + size: safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]), + tensorBase: &tensorBase{ + name: key, + shape: value.Shape, + }, + }) + } + } + } + + return ts, nil +} + +// safetensorsPad returns the padded size of the safetensors file given a length n and offset s +func safetensorsPad(n, offset int64) int64 { + return 8 + n + offset +} + +type safetensor struct { + fs fs.FS + path string + dtype string + offset int64 + size int64 + *tensorBase +} + +func (st safetensor) WriteTo(w io.Writer) (int64, error) { + f, err := st.fs.Open(st.path) + if err != nil { + return 0, err + } + defer f.Close() + + if seeker, ok := f.(io.Seeker); ok { + if _, err := seeker.Seek(st.offset, io.SeekStart); err != nil { + return 0, err + } + } else { + if _, err := io.CopyN(io.Discard, f, st.offset); err != nil { + return 0, err + } + } + + var f32s []float32 + switch st.dtype { + case "F32": + f32s = make([]float32, st.size/4) + if err = binary.Read(f, binary.LittleEndian, f32s); err != nil { + return 0, err + } + case "F16": + u16s := make([]uint16, st.size/2) + if err = binary.Read(f, binary.LittleEndian, u16s); err != nil { + return 0, err + } + + f32s = make([]float32, len(u16s)) + for i := range u16s { + f32s[i] = float16.Frombits(u16s[i]).Float32() + } + + case "BF16": + u8s := make([]uint8, st.size) + if err = binary.Read(f, binary.LittleEndian, u8s); err != nil { + return 0, err + } + + f32s = bfloat16.DecodeFloat32(u8s) + default: + return 0, fmt.Errorf("unknown data type: %s", st.dtype) + } + + if st.repacker != nil { + f32s, err = st.repacker(st.Name(), f32s, st.Shape()) + if err != nil { + return 0, err + } + } + + switch st.Kind() { + case tensorKindF32: + return 0, binary.Write(w, binary.LittleEndian, f32s) + case tensorKindF16: + f16s := make([]uint16, len(f32s)) + for i := range f32s { + f16s[i] = float16.Fromfloat32(f32s[i]).Bits() + } + + return 0, binary.Write(w, binary.LittleEndian, f16s) + default: + return 0, fmt.Errorf("unknown storage type: %d", st.Kind()) + } +} diff --git a/convert/reader_torch.go b/convert/reader_torch.go new file mode 100644 index 0000000000000000000000000000000000000000..531996bf23f767baf3ca8792859a3bea71cf6fac --- /dev/null +++ b/convert/reader_torch.go @@ -0,0 +1,47 @@ +package convert + +import ( + "io" + "io/fs" + + "github.com/nlpodyssey/gopickle/pytorch" + "github.com/nlpodyssey/gopickle/types" +) + +func parseTorch(fsys fs.FS, ps ...string) ([]Tensor, error) { + var ts []Tensor + for _, p := range ps { + pt, err := pytorch.Load(p) + if err != nil { + return nil, err + } + + for _, k := range pt.(*types.Dict).Keys() { + t := pt.(*types.Dict).MustGet(k) + + var shape []uint64 + for dim := range t.(*pytorch.Tensor).Size { + shape = append(shape, uint64(dim)) + } + + ts = append(ts, torch{ + storage: t.(*pytorch.Tensor).Source, + tensorBase: &tensorBase{ + name: k.(string), + shape: shape, + }, + }) + } + } + + return ts, nil +} + +type torch struct { + storage pytorch.StorageInterface + *tensorBase +} + +func (pt torch) WriteTo(w io.Writer) (int64, error) { + return 0, nil +} diff --git a/convert/testdata/Meta-Llama-3-8B-Instruct.json b/convert/testdata/Meta-Llama-3-8B-Instruct.json new file mode 100644 index 0000000000000000000000000000000000000000..808826bb607be0da61301258bc22a93409ce1658 --- /dev/null +++ b/convert/testdata/Meta-Llama-3-8B-Instruct.json @@ -0,0 +1,313 @@ +{ + "general.architecture": "llama", + "general.file_type": "1", + "general.quantization_version": "2", + "llama.block_count": "32", + "llama.context_length": "8192", + "llama.embedding_length": "4096", + "llama.feed_forward_length": "14336", + "llama.rope.dimension_count": "128", + "llama.rope.freq_base": "500000", + "llama.vocab_size": "128256", + "llama.attention.head_count": "32", + "llama.attention.head_count_kv": "8", + "llama.attention.layer_norm_rms_epsilon": "1e-05", + "tokenizer.ggml.model": "gpt2", + "tokenizer.ggml.pre": "llama-bpe", + "tokenizer.ggml.bos_token_id": "128000", + "tokenizer.ggml.eos_token_id": "128009", + "tokenizer.ggml.merges": "d0cbac1fcc9dcf03724b8db5c9bfb593ae1cf68fb9bc72eb1d15274dcbbf618b", + "tokenizer.ggml.token_type": "d70a88809fd7da6f1f028622685cd64268a7a922c5d343c96f25b66327358978", + "tokenizer.ggml.tokens": "765b529dbcbc42dd202ce657341c63807b51f3b07e09898f6aa6196326865d5a", + "token_embd.weight": "b53102a11d9064bbd404833e3464b1b13e08ce73300b442312cccde2f19b2698", + "blk.0.attn_norm.weight": "7318df3cca9e8d153ff0a503026a1265e63d20b2a8c1dd7a2769585082b5d1ee", + "blk.0.ffn_down.weight": "b950806a1fc722c9fad7fd0b20c3c0a7fb50f14395e1e7663a590bfd62e20900", + "blk.0.ffn_gate.weight": "e73e580af6d4f08e060a74a3c25efdf5d3bed99e183d95a5a85ae859014839fd", + "blk.0.ffn_up.weight": "c8158af679ef99746da1befb67eebb19489e0bbe6ce7d97e13e348508244e516", + "blk.0.ffn_norm.weight": "7ec69c3c31e95e49a3359003b0033f6b9e85561a3e3fd83e7476661ecdd756bb", + "blk.0.attn_k.weight": "2732303257bac969b4964e0e32ec08b5a7f5c031bb02bf6ac4467b3ea0ebcf1e", + "blk.0.attn_output.weight": "ecda1d43b4ccc91cd5b366d7e7a275353990ac78561a07c83d9c77031aba12dc", + "blk.0.attn_q.weight": "569b1f5faf92b6f00910cf7effb2d5862f91038ce5c3b0019fc10e5d79fbd5e1", + "blk.0.attn_v.weight": "aa8416c5ef7e32fb54a1f20d6ac651656845d4af240564b397c39bd83e06e3b8", + "blk.1.attn_norm.weight": "03327e02862908c2a44b2f52decdb924bf4201f400b46f8037a9cb2e1d7a61ff", + "blk.1.ffn_down.weight": "5a83a87603f38c99f8e1e370a2d5f967bb45ac51d881a609304a7811027321e0", + "blk.1.ffn_gate.weight": "31da0572c79e655186c721c231376f85e56cdcc6257c28d08c8c5b40d5c22b40", + "blk.1.ffn_up.weight": "e0c811d64ca155c8de10a868e72015d43888834804614ee1aa2953129ffbc90f", + "blk.1.ffn_norm.weight": "5861f313d6137d6f0f904d423df47fffc6069e224ff746e1b637ac9c7f0af862", + "blk.1.attn_k.weight": "5fbbec0acca6457b9416ebdcd90e526885d0224537b7628f6be376a7f275313d", + "blk.1.attn_output.weight": "b237c9763fa3f75166a6f70b70f1566e77d0d89dfa164ed1b3137393e90575c3", + "blk.1.attn_q.weight": "c0a9cf4a98b4882b16f3eb2b49d933793dcc5357abb246fd3fe3134ed2b12e1c", + "blk.1.attn_v.weight": "96867111727200cac1af7865189dd41fd62b47584e5e5f33a91f1d34509cbd40", + "blk.2.attn_norm.weight": "f392f8a88ee3a95b1cc19c40dd4ef66317037b0faaa1800f610779e129ee0539", + "blk.2.ffn_down.weight": "73823eef46632aedcc8c1cb08a736b6aa97ca97842cd1fdfc5567d8dec459662", + "blk.2.ffn_gate.weight": "f4909ae19fc3848b00bb8b9050122e74f8e903b89e22937036f4cc9fea20a718", + "blk.2.ffn_up.weight": "16f4904a3d814ea68f00519724fc4943e48444a84c786bda39aa5efc298a7d84", + "blk.2.ffn_norm.weight": "e3ccdf56e75cb969f6f69c39caf6daf7c4e70e89e25df0f4d2e4bc60e159aafe", + "blk.2.attn_k.weight": "c3beb1e0a11bcf007ef0f0d8f6bdd3082d8b29090cd29597846b5d51e308a8e5", + "blk.2.attn_output.weight": "bb9f66c32cff51154fea92933c2cd62549236f8cb1a767f9ef28d3f99809b343", + "blk.2.attn_q.weight": "8eba394132eef2a05c5a92d62d2376000f7948448d7a2dc74e6b608203add20d", + "blk.2.attn_v.weight": "88f61f77c53567c617db3eef8f30621109a750e679f6784f7911739bd42c2f02", + "blk.3.attn_norm.weight": "7b996675b7ca75fa24107b3ebe0788653ede0f49ac83b8659d71ff54d591f81a", + "blk.3.ffn_down.weight": "2cb332bc05e4821962fdc9dcbcc7cc12630f32117711b687d18fb53c0bc4fbf4", + "blk.3.ffn_gate.weight": "340b387c7f208c8f0a6db904ef8d87c1e84b7d6ad57177abd32d86c8d18b760f", + "blk.3.ffn_up.weight": "07484433f8a7ee061c55aa0de2ecc009f769b0617c9c0ec096e9bb2946df9f0e", + "blk.3.ffn_norm.weight": "4f1a4ade36b393af341240bc894a2aab09cff7e4d56dc4658445deb107f9371b", + "blk.3.attn_k.weight": "483dcd96acb4528df84b9842970994630dbd82b8715ace394aa8b39fcf8d6291", + "blk.3.attn_output.weight": "beaff0810687923585642ee11d929cbf3b43dc6f87f30ddb552c222ab57bdbb3", + "blk.3.attn_q.weight": "0739355002f6fce520863add697e0ff25fc88215322dc3f993be7bb68dcce7e8", + "blk.3.attn_v.weight": "c216d17b6d90ee3e07f82598b8161fae34de2f392dbb0f745b682b578c324767", + "blk.4.attn_norm.weight": "91ab405bc4ba15bf63af233f266aa43aaab43789a9e6596e14a357c2ac7df217", + "blk.4.ffn_down.weight": "620f34ee75cdc73aecb8949af5fbb0d2437fd81422b6d8eb7acfc52addb9fc68", + "blk.4.ffn_gate.weight": "f6feec7bc9acadf35ec22532f8998d8e50f31afedabb19263590dcf8b9a92eee", + "blk.4.ffn_up.weight": "4a72af7cd28fd07b038f6cc4406678d120517280236ea85d9e76eff40ab2cc22", + "blk.4.ffn_norm.weight": "1805b37b44d5d682bdbd2fadeafb763ee001617d7870848cc487079ee34b21f9", + "blk.4.attn_k.weight": "a1e4f9d97cdf4c1b0d177cf00c4e32d1be30c1984a239b3c9bd73f8848888853", + "blk.4.attn_output.weight": "a1547e2497c423b0aff0eee71d9300d6fdf4e4986679418b6e637b69a9a6720b", + "blk.4.attn_q.weight": "0677483a9264ea6803d03d304d87a54632242cb516e8b76b6e3e8284c2f4de04", + "blk.4.attn_v.weight": "02691ba3af344fcc1969428ab0df811ac94aaa2fd91b0dc4ec1ac0a58806980d", + "blk.5.attn_norm.weight": "ba9c028335e5c895b87a5bd1448ca429248f9746ed97bdcb8679923206117156", + "blk.5.ffn_down.weight": "ccfdc9006acad1940a6bc05042a3947f1066acd671e0bb53b7684e9eea9ef5c9", + "blk.5.ffn_gate.weight": "623157679f1e742ccc3807c0b0153ddc8450104de75ec62f1370ec3807c09cf4", + "blk.5.ffn_up.weight": "05748804c65091f963729b58b085f58351891cac8a2861f5eae26b06aa60b2a0", + "blk.5.ffn_norm.weight": "84bae55af2efc8b8429f09056c8c04990c466dae31cb3f9356038b8957f1b406", + "blk.5.attn_k.weight": "8c766180c726b037d587fc52371de6e3307140c52409011609d1225624b6a3eb", + "blk.5.attn_output.weight": "490b582b3b1dc151ae55aee8b6743dad6c01fb49e43afefb6e68394b74be3d73", + "blk.5.attn_q.weight": "6f7b8ca4d9025ec836a44bbcca46be30c66b471a9fb62943ddff8288b3731409", + "blk.5.attn_v.weight": "9f70df3ba00c9e723214b3da83ff435a2163fff5915f75515c9664c05c866c27", + "blk.6.attn_norm.weight": "1a4a66613a682df6f061fc7c4d986f9f7e9175b62f0c42fc1ef31db536bd5942", + "blk.6.ffn_down.weight": "c56f25e4e49b443dbc82d88311ee63bc1f5002cc67e52f4787fd5f003aedeac1", + "blk.6.ffn_gate.weight": "31a5cf1aa9b831a81588d508550f51fc425f9517c43254d4ef7096d38029cf04", + "blk.6.ffn_up.weight": "ce135f3a1163e0c9297a615bdbe68a67ead21edce8debbfa9f6e15e6af8d4c94", + "blk.6.ffn_norm.weight": "4e328ce0648c94e732bc40501858ef6262ad1161e2e407b0cdcf4813fa9d45d8", + "blk.6.attn_k.weight": "1eb1c4c9f9c4c7ff7f5429075e0dc6a7782bed55109fa88df209a817dd8ef960", + "blk.6.attn_output.weight": "3d32986b56873b88655ee1edabdd413fdd9ab18b82108c9ce90bdbc2d3a6f3a3", + "blk.6.attn_q.weight": "8432f583b3a2809c99c393f9beb077cb0534dd5d247c17108f2986cadc6651f6", + "blk.6.attn_v.weight": "5045381513815bb91839dbac8335ffe49bbc7b0008369de7ea97eb676c5e2b36", + "blk.7.attn_norm.weight": "3dabd003638ec2499bfc8a48c49eef34276caab4fe76894eb963207848c2fdaf", + "blk.7.ffn_down.weight": "194fae858608bdcffd235be59ab119d0b91c8549f864ea06dae69249e099935f", + "blk.7.ffn_gate.weight": "00b24c29c30246892bce0791be804a89701d4c1332777e0bcdad5d9d5666604f", + "blk.7.ffn_up.weight": "44d7082a5280080c90cef9e19d410391de34f212ca0736377769b8ddd0c82d5e", + "blk.7.ffn_norm.weight": "21fe8a7fd6911c64e0d15a788b3b4cb6d71dd6ec51de65f760ee89afbb6ae53e", + "blk.7.attn_k.weight": "57a149eec5f6744a9526cd3925ac073f9d12db0fbcb5afe042ef4dc846458c44", + "blk.7.attn_output.weight": "0e9c28a3e81a2880251ce5eed77bcb8be8aaa1a51c9cb6de820b47ed83849fc2", + "blk.7.attn_q.weight": "15ee75263ee4e2a43eb322bc159ae004bb7d77e3a7e63ee4ddab700430693fff", + "blk.7.attn_v.weight": "440aa970bba4bff429fd7b7b1de21f2ad14fb2952b776cfa4acee68d7c6e9b8f", + "blk.8.attn_norm.weight": "af5b44825633c42c1ae964c82bb2be6a242d3a751f0a91f1bae4f593e8f5b6ec", + "blk.8.ffn_down.weight": "b11c14c76adca94fa200496dd2c10743becb23aab6642443ef1ae6d8710edbc1", + "blk.8.ffn_gate.weight": "7bb03d3325bf8637ae2fa1296b0651356515578d46a7c5ca65c7a923d7de27bc", + "blk.8.ffn_up.weight": "b956ef0a0669b5a9c9bf3a8da2d1c24f52d331cfb7354f6d7c51bd65be355e30", + "blk.8.ffn_norm.weight": "c78c3d748302edfef76f71ea5cb2055c94352122eee8b9b1173779a1814d224e", + "blk.8.attn_k.weight": "c0fba6a596ed9c1c32a7055c31a935a8b31e42b77282ee47c1f03ee3bde736b5", + "blk.8.attn_output.weight": "83cf9947080c5d8d571f04a842bc3dcfe7bbb0195fb25b346e22635e8649f2d4", + "blk.8.attn_q.weight": "47409350a576b333d97b7c877d69f47f46df504f3765102dfc0be9e521c7ecd6", + "blk.8.attn_v.weight": "1999dff91404fdcf1ecb34d9eaaaa9244ec7658a74dec8feb7cfd1fddba0347e", + "blk.9.attn_norm.weight": "1e6e29d5c3889ab4e1b0a5b9998cba60179b0f1fca133515df49cbc19d092593", + "blk.9.ffn_down.weight": "acb898a6490adff592e10b4c62d70edc5941661ee6da44658500e9205357c8e9", + "blk.9.ffn_gate.weight": "4cff63013593aadc3ffbaaa6ed70ffdba1224cd43c3644bf6f4162b5ac1ab542", + "blk.9.ffn_up.weight": "f985b5a2d6cf4fe32c7256301c3c89b8ad22b59e516342c52da42d8110766a4e", + "blk.9.ffn_norm.weight": "0d659c538bc6b21ed0018f107ab674a7424a00a42946c80e07208b479b21918f", + "blk.9.attn_k.weight": "f67611d888780d1b38c1c146b361c65310c8183bdf64fd73e2259985c6e8517f", + "blk.9.attn_output.weight": "f12ca1fa62a02ddc3f77f798bfb5707e0c50bf18ee0eaa67025521a98355f26b", + "blk.9.attn_q.weight": "3865185f4361a645b086ad47b72904c095313fb1c624e511647bf1a7dfc1c476", + "blk.9.attn_v.weight": "92125bbfed63544ab56052bd1e4aa453bbf34c795249ee54cde54907c8c6d1d3", + "blk.10.attn_norm.weight": "5d6bfbe545bcc2fcb2fc75c68f64b1f4c918badaf53e0156fe2d88aa977b2f94", + "blk.10.ffn_down.weight": "1dd9da8b0d2696ab5531fbca8a29c7d67567620a9d3e5fc2a19ec5d7e4c6cc8a", + "blk.10.ffn_gate.weight": "6e55e7f014edaebda0ac6819a426221d3b025c27312a2e18cc5806f31e3db226", + "blk.10.ffn_up.weight": "d80dde54af5db51241345ee8d64c1972608644f4deeac1e8195dc423bf27474a", + "blk.10.ffn_norm.weight": "f6ca65951d58ae3379eee8247bec34ebd0db05674cc9295593573841b8a55df3", + "blk.10.attn_k.weight": "b58e350bd6b49aba0fba4e4dd6865de3a2a0651ab865dbf2419b627b53ffc187", + "blk.10.attn_output.weight": "6b26a986e12fe66ec286a21d7d5af5eaa1bfe6f2bf502165d270e4497235a54a", + "blk.10.attn_q.weight": "3440e0e5b7e0d1e426424ae5a33f4e057be623249e9035ea12e57dbe5d3893c4", + "blk.10.attn_v.weight": "ebfadcfe14bcd6dee933053df0a67e12e7a196d5cc45728c1ffb2a2daedd5ca2", + "blk.11.attn_norm.weight": "3ed057b9576cd2de84507ef64c7646dc478c651efca4c2024cbe91a4f3fbf0bc", + "blk.11.ffn_down.weight": "8ff1c2487d22f5c499761e4eb721418f141f960160d0bab779595a34e4d68898", + "blk.11.ffn_gate.weight": "9c74e4507c7e45bf39b7cc7402198cd1dd77e3fff8c625b0413acaeb16efeb9f", + "blk.11.ffn_up.weight": "4367158007161d29939e00a322bb6776016e43f648a94f9b08a96a477aae75be", + "blk.11.ffn_norm.weight": "1cc0288c1491072121f4c9a0af20be0e13af49895696a3320e4fcac608768de3", + "blk.11.attn_k.weight": "066f5b3c144fce1366835e1ebf376f768b333b8ae29f5b478c42d1d0c809c855", + "blk.11.attn_output.weight": "e0d9f3d3f2c54aed59c02713ea4fb562799ddbacbe67ca3998dfc887bc44e47b", + "blk.11.attn_q.weight": "28d3ecc8a88cb3815e89a7f7a7d043da7a71f702b337a126e4d3a2ac1cd6370f", + "blk.11.attn_v.weight": "7c5cdef10ee73bca0a3b9f6ece5f0a0155664e0ce3d8de90ccdccfab5545e5e7", + "blk.12.attn_norm.weight": "973b133301a1af760cd7b3a7955371ea0a750808b442deb6adaf7b98482bd0c6", + "blk.12.ffn_down.weight": "d6c87b4b4ca03f75546ddd6a9e7fca720585a309188723c1ace8122438d4b200", + "blk.12.ffn_gate.weight": "2189a6e0cab1540bd05d6089b922aa8fd694be51255654933c165f302a0c955f", + "blk.12.ffn_up.weight": "5affbec19b58d092b9305721e3552481fe2eff51269ea3ed91cda3b9ef84d4df", + "blk.12.ffn_norm.weight": "f650fd42a34e950f758b4a130e7b8b1a712b1dcbede0291bb8edde47aaed0ef6", + "blk.12.attn_k.weight": "59b1e86f10450a7cc188beefc0856d2dcf44e8d7fdd9cd8859c30ec1ebaf24b6", + "blk.12.attn_output.weight": "446b0d36b2f66bd72a2323f4f4e9d85a0f621e9a58872e89a27248d6b1123238", + "blk.12.attn_q.weight": "3ed6bfd39f040301ed99fad882d3e569769d594259f9948445bef0e44ec881fb", + "blk.12.attn_v.weight": "e73652cd5d0029b1931be3ba9d82508f6696dce5a29d085476a54fb7a2ddbabc", + "blk.13.attn_norm.weight": "491b85278c0bd67bd31b9b8a9720902c244bd067e53a4a03641b7c0994782e82", + "blk.13.ffn_down.weight": "ad71cc248a85e9ced49307a24a9bfae01d387e979a7689c82ff59998e09741f3", + "blk.13.ffn_gate.weight": "0a55984d53971fab97575ee0ef5882013be7fdecfa76e3fbebb5dc85a07a14d4", + "blk.13.ffn_up.weight": "378b697b35e2e53c0de98e8e29b73d42ae3ec112ec16129aa5997a9e2f3b5943", + "blk.13.ffn_norm.weight": "f8aff2f69ab286210fad45a62b03f8d10b38f96a420d7baadf6b95d7b0b0bcd2", + "blk.13.attn_k.weight": "25ceb841afb1034831bea7f4d6a6c578def2ce4d4c412c780ef147dc9a598360", + "blk.13.attn_output.weight": "a242b322889c6bdaa14b67a7bab593db39df8eea3721638ef639abbb74d482e3", + "blk.13.attn_q.weight": "d80be9945a369439e835c55cfb0e97828b8a66bb7ced534d9059c92487bf20a9", + "blk.13.attn_v.weight": "ac33274cf9b67979d9ecdc967a55175afe0c9c4aeeff6391433cd9840c818706", + "blk.14.attn_norm.weight": "12a1e1091de5b2da12c9e7c0b1c8e6f09ce2a749733cf7d5240445b8e21cd093", + "blk.14.ffn_down.weight": "cfd41965c88266e32bc2dcdadda512499c35519e8686fefb9a7f249ab2291eb5", + "blk.14.ffn_gate.weight": "8dcfe774f07a095c7c6cf0a901c9df70d938bad7b5ba347fbc8f694e7603c0d1", + "blk.14.ffn_up.weight": "c7995577fe4a72ea0fb17c4a7b6b87b959072bbfdd5edacc6c367d43465809ae", + "blk.14.ffn_norm.weight": "81c41ebde41739e7016ffec31d2256217b825dc3cae049a935f5f61a60d22003", + "blk.14.attn_k.weight": "fb708bdebe4384f5c4b479c110028554f4d122f166b8091eda7d8d65e6780eb8", + "blk.14.attn_output.weight": "f5295caf2dfdc60553dcabe17537a80577e8b153c902247daac058df23542514", + "blk.14.attn_q.weight": "c12b7a3601c68c63ab5dc9d2599ebf3f3a10abc2c59d3a2126fffd5818f2763b", + "blk.14.attn_v.weight": "1ce968d9149bf0d5e237d52cc6d6433565b4bbf03252a736262bb00a2b34a687", + "blk.15.attn_norm.weight": "266fd2c36d7dcefc6b6bb7f1c9374c41f2bab5d6c84a063b6f91c4f682dad3c4", + "blk.15.ffn_down.weight": "6154886e9ef0a6cc08ab0d264a35f497e6f0987efdac992ed04e87088bea7801", + "blk.15.ffn_gate.weight": "183d9fd3c1b5657840099053d2fd3f72ad953b1de523296159b7761f20491a76", + "blk.15.ffn_up.weight": "51546d4498842ae2340ee226a0888d5f61e7d2ca4d052dfa06a77b0451242d3d", + "blk.15.ffn_norm.weight": "ef7378091a41a25a5f58bf1bf9d3bc64ea562e7f421e1c232b1f177c30fd3500", + "blk.15.attn_k.weight": "8d556ab8d9639324141774999b6eed0e91d7ee645bf3e7a3dcd200b2e7a00751", + "blk.15.attn_output.weight": "54aa6ba87def7cbe18b0c6ab3aff5c351cb3b6ca4a0d7b2cd5f75a1312991429", + "blk.15.attn_q.weight": "10731b0dc031ea8e0ef37bd7f010e0a78518a10a6df05a8bae48e3148b73ef3e", + "blk.15.attn_v.weight": "cbbe50c2ed7224866d3cf9b489c599f3ec41a4ea1aa3181e9f4e87e1fa0cefec", + "blk.16.attn_norm.weight": "387058eb39d4b28c04cf1368247417f1faeae8ae79d894c9f293457e0eaa00b0", + "blk.16.ffn_down.weight": "2cb26ccee585e933401ad5c82ed36ddacb3289efa0b28f8cf91b020ffbd9c333", + "blk.16.ffn_gate.weight": "d745985efb5bab42304e5d509024631efe35f92f2b2ec4931ead6db97ca9727e", + "blk.16.ffn_up.weight": "7a67bd195e0642828ca36eb7818149bb70c2c25f82de07e2b5807c520daf540e", + "blk.16.ffn_norm.weight": "7cefd061c8182482a89272f8a4e88a954b12609a62716923ca1cb3593b1c1651", + "blk.16.attn_k.weight": "d7968a2de67e755b4533e061aaad1cb62f8882af92dcad67f99d6d5112513439", + "blk.16.attn_output.weight": "9e9ab5788272ca3394ea89eadbce8c86ecc3fd75b7899184d6191c134ad9aae0", + "blk.16.attn_q.weight": "ef81c261b536c1a3a093b33f44cf2d42b86e5aa2d821674f07a0c80e992ed925", + "blk.16.attn_v.weight": "aef38e7958301b4a437cbdd2fbae6197f677b09269ec1eaf63188cd5da428d25", + "blk.17.attn_norm.weight": "28f6b289f1bc3131041e9f791b7a2a3a48baee0dfea27bf7051ebbb7ed364d80", + "blk.17.ffn_down.weight": "1a502829aafc6a9bd6bc81f12573bf8632d5c8c659f0dfb13c8b2411f3b1ec05", + "blk.17.ffn_gate.weight": "ddfd8aa0eb98846ebc9afe31366249159f46ae9815199dd70161527ed241ac4d", + "blk.17.ffn_up.weight": "4211a3cc247071bd361b30de2131d02382f552855062bf3b3e004c17992e5d09", + "blk.17.ffn_norm.weight": "647e5fa99a5b0d232af36d15816539f4d27e60a50a341b00aa88bb6e4474f8b9", + "blk.17.attn_k.weight": "d9125ff33a19c502c0f8846433ffc24395048582fc2f463d34a0301a82156f02", + "blk.17.attn_output.weight": "3d64fbb1cfef04444827f37c35fd9ad3413eb2165094d339ef89f00503f09de4", + "blk.17.attn_q.weight": "e5b29424028f578beca385fd82e29f37adedf3037cd51e5889d5a1ffb0428ca7", + "blk.17.attn_v.weight": "1809c5aaf2ac04c5d65539097564ad62796e87d24bb8b9ce5b095561a61d908a", + "blk.18.attn_norm.weight": "99daca58d001c627523d3adfbca1d95f04e590382a326866544d57989d5f4835", + "blk.18.ffn_down.weight": "84f30231ce6ca0f10227541dfc602d6418c1a210386b0c4926ef1656e7d4635c", + "blk.18.ffn_gate.weight": "ca5bbe4468b541740e54f69b9e08fcc8e478c344b70551dab21b1206acfbaadb", + "blk.18.ffn_up.weight": "0b3067b9dded31686dcfdc1e247eae3974a28a61ac59e9862758dbfaad64e8f7", + "blk.18.ffn_norm.weight": "8154a102232dbc0f90ce77ae5c1ff8f26f8b6e4dcf326e9ec1645749669e7960", + "blk.18.attn_k.weight": "25abb26021ccc481471a30e0d4cbeb7e1db29828417ec5136edeb93fecf09ac4", + "blk.18.attn_output.weight": "d87d481d9b046b68efa06ccdd4ed8cbf61e692d61114b75b7fad5ed75f5d87b2", + "blk.18.attn_q.weight": "cc6400379e15766992ff1293be79dc67682c28e9e15155a78109f4b64653b164", + "blk.18.attn_v.weight": "45c75cb1dd496aea3173aafe2575b841dd1d02cbe010b3198099731eb98f531c", + "blk.19.attn_norm.weight": "65389efc75297684773284ef8e5f8789a4504b636c9f33b8a32e0ee42499fa72", + "blk.19.ffn_down.weight": "4eefab7e939f64a17e4a214ca3c77a6fa110d94f677e2d6401086f70fc538b04", + "blk.19.ffn_gate.weight": "f1c0a59cafda66f466ab585b0b8b4861b58abe87a67cea1f6a488492242edfdf", + "blk.19.ffn_up.weight": "c42d045eef588db4a0e56960a57e110e1ff92eb8041107d19899165fd3b90f17", + "blk.19.ffn_norm.weight": "a8f33eda6d5d62ff5f333ad9771783caff556641f4e7df713451385676f441fa", + "blk.19.attn_k.weight": "0bab5d9e9083492bfb05a5a3bb23b79c0e7b99ef6a6644817b4d57d5c453b8a5", + "blk.19.attn_output.weight": "c99c551d70eafad0f7aea98fb6f9251635897168eb3895f76abf0d4ea3b3aa6f", + "blk.19.attn_q.weight": "c98bde95627c3b54c9443813ca50b4e14f518319681db6bbf7b2332ba26e9a60", + "blk.19.attn_v.weight": "ff3a490518cf64904db89ce0dc7d6eb89e870f1440e41883c6b55a221f82de84", + "blk.20.ffn_gate.weight": "761f0e317229cafe9d3754048ab038a0a84e9a287b196ab65f633139f2d29aba", + "blk.20.attn_k.weight": "45d13439b41066d282e8490a726785abf513605f46c79bd0c840f6419d27e790", + "blk.20.attn_output.weight": "a3b958d84b4a097844179b7d55c18fd0e4f319cb15e918c6fde33b68de1bcac6", + "blk.20.attn_q.weight": "127ab8e7d8c3f882874904196a02712bab42e6744fde45871b67350609d19f5e", + "blk.20.attn_v.weight": "5f0ad2d14a8ae42dd3bbeccfb33295687a14055fa92c54bc946249373c1c9f17", + "blk.20.attn_norm.weight": "77300b1755edc8c70089e0f45efa646056b9add7d8568b2324d2f3e62b64971a", + "blk.20.ffn_down.weight": "ab93d0e075b42e9017b701a070d561e698050d90aac4b4b9919256fbe50c3204", + "blk.20.ffn_up.weight": "4fd6628a07acc57a48d1ef83f81b7d7aa0bce569c1160a99d307284f8821322c", + "blk.20.ffn_norm.weight": "2a9e46b9e48e8e55215de56592e1f189530037c1c94a1428e3d6f106c7f26fb2", + "blk.21.attn_norm.weight": "4b3b5912c7bc61eb9da8e47d4651f896e85d9e59c4ecaa65df7acf3c21737298", + "blk.21.ffn_down.weight": "7146f931663d93b8771cd84405cd4802ea6560d0729b0d6d44588203c095bc53", + "blk.21.ffn_gate.weight": "b44ec5d64388fa40b90b3e9976d97a8b6800fa3b97584f32e64b03daffb8601f", + "blk.21.ffn_up.weight": "0cf3643fd23c685e17062cd11e116e17ce57a405e5e78953bab94cd62fe48789", + "blk.21.ffn_norm.weight": "4ef2cdb53da166df70b39f3e6b17af51848cfa5ea3c27ad6a1ae2a1bb1da1ce9", + "blk.21.attn_k.weight": "5d40f32a706f670c19972b14176bf660d5b045e3637b110dbf8d7de4ff32101a", + "blk.21.attn_output.weight": "18afaa916752ce16c9653ec0ec7e2fe60be55faa2aa5025d147be184adb75cac", + "blk.21.attn_q.weight": "2621daa5f858931514a4b2f0fe8d81cf9b96f541e6af99bfa7539e9bde8e34ee", + "blk.21.attn_v.weight": "63226dafc54c899bbce4aa49efceeedd8908e94faa613450fdda91f332b62864", + "blk.22.attn_norm.weight": "cf3058daab4d2c04387e7d169d1553bb8e7358eea66285ec067703f6ce62043a", + "blk.22.ffn_down.weight": "6a58d5fd220abdbac6cee7ba048abab794731af318f04982c2506df59413d0b3", + "blk.22.ffn_gate.weight": "d5614535324b03c7b91727a903b2a72f8d07ad17f7aa8b61ea173cf9b895069e", + "blk.22.ffn_up.weight": "ec20da3949566e93f66cabb67f8cd7eab399047ec6ebf5d43edfaf3669b82296", + "blk.22.ffn_norm.weight": "84c82f38f53a649972a44466fc476bf764e064ce18de870291edc302f3700e28", + "blk.22.attn_k.weight": "a3d2ecc37fde7c201176bb8abadf27f0d8ede9679a6034913e03d9db924fda12", + "blk.22.attn_output.weight": "5a3b8bb433f43a387df43dd371bdf80ddfac986dfeaf38e9bac1d7a0ec6628de", + "blk.22.attn_q.weight": "3a875cec661b4859f30a8fd2c866811184b25b68c9e36fe2663d299caf8b59c6", + "blk.22.attn_v.weight": "8717a83b79035058dcfd3ef6f8e5b36e71d77379e5a239e1899eef8766fb7703", + "blk.23.attn_norm.weight": "2b4a68a0a2f023dd646e4755c9bef17c2f631901154afd839edac7ac006ec99c", + "blk.23.ffn_down.weight": "29499b1586c6fc4883c9b7a9c8cf388035146b5aecf90c5c4c8c8e082c71e7d7", + "blk.23.ffn_gate.weight": "7d6554036d21c587b9b556428054f9c15cbef96d24b257f906fcef4ae38bd9c8", + "blk.23.ffn_up.weight": "19761ecb288d6ebd44b681c4535661583b1e19dc29e96d0c007333cd8f00aacf", + "blk.23.ffn_norm.weight": "37dc35500790a4ca33807b39cf7af65065e535dc25b9e94f3ed2759f61887ac9", + "blk.23.attn_k.weight": "717547d00323817b0cb40a72ec5f8cf42ecd1f9e3e42715c2cc5e38f07fffffe", + "blk.23.attn_output.weight": "a24786feb6a905fdf166d7500133757cbe494779d4ebcba9eb03046b319557df", + "blk.23.attn_q.weight": "6a2c4a98f138b928d22136efa163562691d3b4ed526d52d46a2fa2694a8f3965", + "blk.23.attn_v.weight": "c6e6081eb9c38a7fda023085957b460e9ea321e1fff408b38c2b58595c39979c", + "blk.24.attn_norm.weight": "5e6283f891e538670425f3e244b08dc6f96f33dfa4aefa913f8eb17212421850", + "blk.24.ffn_down.weight": "e09eb170f389deea0a4a1cbfdb52c12490768a2c60491b7bef8a4c445e2a08f5", + "blk.24.ffn_gate.weight": "af29d815cf49a38fc2ebd0bf9b2dd9933d023a29f2d766981acb9a1b53f09117", + "blk.24.ffn_up.weight": "36ccd9333426666de9d3088bd4dcdf5b624b09dca9e3a83a22fc0383f2d950fa", + "blk.24.ffn_norm.weight": "a88e1692318826db6ac42582d182e51a3c698c655d0e21e04fa086318832d07b", + "blk.24.attn_k.weight": "f7d61d6d1225289bcc502e3bbb0168b4584add0253218c1b77ac92ccef9a1c2e", + "blk.24.attn_output.weight": "85a1363b3ccc87312094c2195022687c16b0dad7fafb9e80bb4ec474d53c29ac", + "blk.24.attn_q.weight": "53482a2c008f42f4fad779ca323addc3712040149dfc12f782417756388a72bb", + "blk.24.attn_v.weight": "67498272369af7dd10097c73b07f731b565cfc9a559e711cc0d526389e7b44e2", + "blk.25.attn_norm.weight": "98dd617def5cb7825ee4833132ca2da2121245921585e1d9e36b93344adc321b", + "blk.25.ffn_down.weight": "7fd477d6c50aed5f424a878dd284343379cffbee8a34c0b6e55100c8305fa13f", + "blk.25.ffn_gate.weight": "f892c9806c8ec22e8aa746734ac9213428c534921cf161239e1d249fdb5d1ec0", + "blk.25.ffn_up.weight": "528bed14c9bf9762f790525ee40412545221f4321d2a2323fa8e73c58b7643c5", + "blk.25.ffn_norm.weight": "ca5831966672e7be6a578feeb631ec3570d3b5afe12860819ccb96e896ffc346", + "blk.25.attn_k.weight": "610d3068cc9b20401f0c3a0efea39a279dd9f564fde19baf3403b2ec2319e4c4", + "blk.25.attn_output.weight": "798aaf702e53b657265ac3b5e6caf3a0ab515bdadfeb1a3a156b4f3bfba76666", + "blk.25.attn_q.weight": "8a7fa25248de83029fb97b51d036a01baebe31fcb4be121ab00dd8b7de209b10", + "blk.25.attn_v.weight": "2a53d5e9f8a1218c66958c6388d3b37400a9af7956c785024ca44bfbc3c7d371", + "blk.26.attn_norm.weight": "5f44fc043481eb0771f3e6d2420bcbcf73140afb9a9feb8eddb6575452acebee", + "blk.26.ffn_down.weight": "944a60a409d0d5b6a851e33c69aca152454b691711a8b96f5bcc488772ab2833", + "blk.26.ffn_gate.weight": "2a0ca4abb3de5593e6693d8be69b63d6d1a639855ac8332a75f520353f030c62", + "blk.26.ffn_up.weight": "0b1df496163f9ac07bf89375d3eb441b51a81d41b47d769a04a61efc18dbe35b", + "blk.26.ffn_norm.weight": "56b8dd046e9be6ea71f7efd80dbd14e7fb1aa020d3cd38e063275f3873fd12f8", + "blk.26.attn_k.weight": "b1dabfabb970e6971c7ea6e53c63cf7ef56341e6a2edd9cf177785cad9af2f9a", + "blk.26.attn_output.weight": "39532c7e836baad164a655fb97ec5114ea4da37ffba9fdea2684f6e4450e6f84", + "blk.26.attn_q.weight": "8f48bf6aaa1252bc149e98af2be1777a5c0d2c3274c6d314171ea9344a41b604", + "blk.26.attn_v.weight": "02fb145f7fd905133750e90571effacadddfd3f4966552dc59982ac3900ab8c4", + "blk.27.attn_norm.weight": "654d168fc3cab716d91261f5719f180b7d697218401633b4878a759f1b5283f2", + "blk.27.ffn_down.weight": "2823272bec3a1c12f02cc4cb24aa4031abd7e9dbe0b02676e2305b21671818f0", + "blk.27.ffn_gate.weight": "b1a1d40cd02f97182cac17a79971d1934ee0daf3aa0bf11303568c636e208a64", + "blk.27.ffn_up.weight": "ed62ec72a020d070e64eb7b50237b32213944727b5b2427f45d989f50df5fb2a", + "blk.27.ffn_norm.weight": "c69649ac65d694b306a905dee8b03b89eec1ed188b1eaaf38f8e29d4b12e38a0", + "blk.27.attn_k.weight": "cc57bbf413f1fd227128dc66efc8590c73634cbd6f96d01ec4878b5e7ca6a925", + "blk.27.attn_output.weight": "cac407ad02361d53207b3c7e25ceab84dcb4347b8087055162e2efe14d11d84a", + "blk.27.attn_q.weight": "0af18e07cee12015761c07c94407024f4f4d77d97bdb24163db0e16669e2cef3", + "blk.27.attn_v.weight": "a1d08fbdfa40af773c5adcf93bd68b78a44ed144e3fc6bbeb8af02e937527eb6", + "blk.28.attn_norm.weight": "f39a51f814512b040a1082143150e4a49ff730f85cef49d7f77fc79d83e91f40", + "blk.28.ffn_down.weight": "74f29ed51055d1c1adb8f0660bbe538a27e016c65650f2d67efc6f1c84fa1b45", + "blk.28.ffn_gate.weight": "ae48bb16487ded6781c60aafc0bf738fb4ae15729952906f247d216592ce249a", + "blk.28.ffn_up.weight": "543009727718ac22f11ee4b17815f68ea6f15ba1f3e7ed5ecdb755cf6417565b", + "blk.28.ffn_norm.weight": "b8f9e54c322079ff20a82b88948cdc2916c22c7db40b9a9ed6d3cbe89efb727e", + "blk.28.attn_k.weight": "55d055ba653b728d6e784f9e013786fed07115c9fdf23367e3941386d5e77db8", + "blk.28.attn_output.weight": "155101c03ddbf18f4fd0694bfc982f33c7bae25c9b087d6f5273c2bfbffcf2c9", + "blk.28.attn_q.weight": "1ed19bfdd22e9c14eca014739982492e9516d411515a8585f65cf754d849e53f", + "blk.28.attn_v.weight": "11ba854dd575c025d37256eee9041f6d1bd2b549a083d6409a09bfc1542913f3", + "blk.29.attn_norm.weight": "02b0bf5e2fcefd11a153cc988c81ba672682e4844fcf6442423e21a0e10d566d", + "blk.29.ffn_down.weight": "594bb692ec2779938721ff4748666ca8370e0e4fe85229503f616438b8884f5f", + "blk.29.ffn_gate.weight": "8bedcf47e91dcb2cf4093de56b048ee411faab6ff472f89ab2c9c113a08e6967", + "blk.29.ffn_up.weight": "e241a547b5fd6dfca8200b8141e21c1c487a96cbc4e5855f181a7ed1be91b642", + "blk.29.ffn_norm.weight": "e63eba5e4c6b288bfd9f15e46e236086456c8b7f1f9c732c0b5de84962a2e7cc", + "blk.29.attn_k.weight": "afe5979d5bcf211aebb526620f5974bcb0a2c39c8be71e815575c55d6385e3aa", + "blk.29.attn_output.weight": "9c944ed44b124b014906fc240afd3b90aed56bbd9567f2eddfd5b7a685b3cb48", + "blk.29.attn_q.weight": "e234e08e5c1bd9245a2edc8d63e9933b6b879f97c01392209cad4f55f05f3ada", + "blk.29.attn_v.weight": "5cb8e3e5f954e775c5a5e4de7a9a62b17e9c6931bb0ff0e2f82c4126fd3e1a1c", + "blk.30.attn_norm.weight": "a65483ee51a0b214144ec8a14f28ea5437586e9e12ebe342a57d1f8627ee12af", + "blk.30.ffn_down.weight": "417959da77ceb33ead4271cbb9428b195196173a893c44e52880a7ec61b4856b", + "blk.30.ffn_gate.weight": "a0d503ffcbe45dc927600bb98c9f6082487e65cb577ab545add400d666a87638", + "blk.30.ffn_up.weight": "f8ab957b82ffcd10b21303cb5e866209b6fe95f827b1b94e9a949207952d12c0", + "blk.30.ffn_norm.weight": "210c7ceb0514a9ef27b5d4d1b3aff6dde43f1af0345a050d71097940e0e73e03", + "blk.30.attn_k.weight": "16861b9abcf5a3fe73c93d977ca45a1e6daa65be0fd85c2cff53486ce2033afa", + "blk.30.attn_output.weight": "ca541fb2e57e2257118c35784845b0c731278af8db3036ac53d71aa1681fdbdc", + "blk.30.attn_q.weight": "f7834917748e26bb456b945e230bc926c228e93696bc01fbc2b134bdeeac71a1", + "blk.30.attn_v.weight": "9292783171dbe5eb689d17c9bda11e537f0e9b328fced6986c938d61ed590e81", + "blk.31.ffn_gate.weight": "e4766a04bcd8f937ba883c6a144101e546747804ca66c35c97281d6ccb47b566", + "blk.31.ffn_up.weight": "cc1e666116f7e6b06736db4aa4b81003c583f54f4d9200bfa48842249940e16a", + "blk.31.attn_k.weight": "fc80b57557687504efae7d24265cb7dc39b8f826bb3d897a11783012dbedc44f", + "blk.31.attn_output.weight": "215617f50a1f5d9b2250b82f3652b35a9e9aa0ad9ef2b485d73965a14b2b872a", + "blk.31.attn_q.weight": "274b4f1dfb0bdec28632705677049fb3e327ce6d9e1f3baaad1560439039982f", + "blk.31.attn_v.weight": "e641b8b926f9dfcbbf6b6da1c02555525ac4b1c306d96f20cfbba7d6662c4e56", + "blk.31.attn_norm.weight": "b3243c361d4041ddb892ce6862dd5091f57d87357e3c67e177451b85d8baf34d", + "blk.31.ffn_down.weight": "0a00cd3ecd5e91624a27f9e239b1de425d5ba3cfff82c256a11a4ad434abf3c2", + "blk.31.ffn_norm.weight": "2a0d67ea2bb1303975712243f07273c92fce83baa11b1cd6d8e42e74ea3c810b", + "output.weight": "768615f077fb797967844571c58b94d7c399d884d115be3ab4b0154504cae892", + "output_norm.weight": "7cc5b7ce10e5082000fa00bfa68af8c7c5da218e59e2c41cf2f1499d40ca229e" +} diff --git a/convert/testdata/Mistral-7B-Instruct-v0.2.json b/convert/testdata/Mistral-7B-Instruct-v0.2.json new file mode 100644 index 0000000000000000000000000000000000000000..88d447b3a6a247e4e09c56ebc940f1576fe80e67 --- /dev/null +++ b/convert/testdata/Mistral-7B-Instruct-v0.2.json @@ -0,0 +1,313 @@ +{ + "general.architecture": "llama", + "general.file_type": "1", + "general.quantization_version": "2", + "llama.block_count": "32", + "llama.context_length": "32768", + "llama.embedding_length": "4096", + "llama.feed_forward_length": "14336", + "llama.attention.head_count": "32", + "llama.attention.head_count_kv": "8", + "llama.attention.layer_norm_rms_epsilon": "1e-05", + "llama.rope.dimension_count": "128", + "tokenizer.ggml.model": "llama", + "tokenizer.ggml.add_bos_token": "true", + "tokenizer.ggml.add_eos_token": "false", + "tokenizer.ggml.bos_token_id": "1", + "tokenizer.ggml.eos_token_id": "2", + "tokenizer.ggml.unknown_token_id": "0", + "tokenizer.ggml.scores": "e3d3eea80bb41a1213f2d0aa3e8a38581d1f19323be77dbd779c9c7e3b72e676", + "tokenizer.ggml.token_type": "6040635e6bd38d98af06698feb75c1802bad35180ee6ae0a503e38c0f60fd71e", + "tokenizer.ggml.tokens": "604ac4bfbd019e430d7b6cdf18c6c0cd5b967900601f0307f714ec7773aa5ca6", + "token_embd.weight": "cde834ccac5e94324b25cb81b02d27312cac0c551b55a7e1d555d90bf6cb6e81", + "blk.0.attn_k.weight": "458bfdd9715c66e017c2447b1ed3c582963a3111479314e664faad8c914f42be", + "blk.0.attn_norm.weight": "e1fd60b95f713bae7b7e3ca933c64ae6c9cd1e8d808000204bbfdc19f0ba635b", + "blk.0.attn_output.weight": "df13b6a157d9d4f96c53b012b3b9bcd207d0c94144cbd22ae3ec13bb07d6c373", + "blk.0.attn_q.weight": "13b4126b4245bf06c915a93317c42b8174e05053535ec99dc576541e4cec7c25", + "blk.0.attn_v.weight": "5b1781d3a341214511b27eb4e268674ea3ea829dbdf8ae5a6bb89b3c0b33fafd", + "blk.0.ffn_down.weight": "49186f5d8148d316b07458841d13a2e66587f4af69b776188a809591ed9c070d", + "blk.0.ffn_gate.weight": "4397e30ece09136f00f4ff84ff49e5241b765a374deb8c5a12e897e2bf73473e", + "blk.0.ffn_norm.weight": "43260589aac3850a779bca3f9649f793bbfbe5db538361cb743b3830217f8287", + "blk.0.ffn_up.weight": "fd7ac918240a07566f6967527ffca58fcf433a30b78fdd6d84b2136d4ebd9987", + "blk.1.attn_k.weight": "209839566c7d235bdc20565a4766378b6ee8553133a5a3315abe8a85baa80712", + "blk.1.attn_norm.weight": "58c52986f7c69784ba327cb7f350923420782bee17fa39b1fbd13839d4005357", + "blk.1.attn_output.weight": "5067cc628449682665dfcf59b16e58fe2a9d2a81cb099f0fcd42f4f8670c6740", + "blk.1.attn_q.weight": "f410f9f0dd5edc09401af597d02e2a4c727f1502ec3ec3898321617b36c6df6b", + "blk.1.attn_v.weight": "d40fa49e07c102c0644e130e7909eaa93ed0d54e2edddc0759e721d58a4e4f5e", + "blk.1.ffn_down.weight": "594b1eff6ed4defbdd819fabbe2d48764984f08878a860bdb808511d5a25b8db", + "blk.1.ffn_gate.weight": "4cda97541e388a5bb607ce4cc8b3db1da7045830a630e7ba4d17807befcff346", + "blk.1.ffn_norm.weight": "66c13d7481be65b97aa474735ddc9674f33d512ddda76fa6fb45c7464b09f1ed", + "blk.1.ffn_up.weight": "1adc6de288ba4cc1237833ca8b4eb81107149842e38bc452e18e5cfe284338a2", + "blk.2.attn_k.weight": "5420423559f236ab22d85a00849f31e0cc6e9c7dd879de724393d8cd2b379153", + "blk.2.attn_norm.weight": "495fe1ab40cc52aa054ddd4f0c2d2790f4326c8d103296b1b38f3b1060db2a24", + "blk.2.attn_output.weight": "ccb83e7085381f558bfd65588c525ad2671feddcbc3887afb4038ad9c7aac348", + "blk.2.attn_q.weight": "2e8f77478392bc93c2a391f2e0f4a173a952bbab88a7aca099c6ee909726409a", + "blk.2.attn_v.weight": "d64512590f3b7ebbb9e77c2eb97fbda90b00d45c944f2b174f03a2cb11007567", + "blk.2.ffn_down.weight": "1de5084a05dcaa6b1bd926e83517dbe9ebe7fde79235fe56018b3028b1aa6397", + "blk.2.ffn_gate.weight": "cbea526b557f49aad8c976973cf367fcd12175b900f551984f498b9e07e4b7fd", + "blk.2.ffn_norm.weight": "530aa49b10c7eae08899d143409240deb95dae4e1d5bf78cea3b26393cff3ba1", + "blk.2.ffn_up.weight": "13a5fc19b96b4dcc1e9bd01998c8272ebe52034c1933ed123a506b711fae9a5c", + "blk.3.attn_k.weight": "1913b63a73305941d8cdc472e7f101c633d3357a78602eac0a4b49a744261075", + "blk.3.attn_norm.weight": "9c11bed5ab41f4adbfdae4ead65b525c8f19443e656a8c61ba412a4e1ad1193b", + "blk.3.attn_output.weight": "bb0b42c1d34779c5943272ed71f1dbb31ad8edd75f8bcd5c868f88505ac3a610", + "blk.3.attn_q.weight": "3461a1fe4e49f5319ea047cae98ccdb46528a3ec23831183fe87610b48c94948", + "blk.3.attn_v.weight": "82aa30be6a61526a41fb79bb28a2617416f5909f0477aa9e95e16be9370fcb38", + "blk.3.ffn_down.weight": "68521011ae03f5e3b0966127111afa8ee9f2eaeeef8d3a0b86b633e0332e9fbf", + "blk.3.ffn_gate.weight": "1e89e26338fd364bb679695968c65106382f15ad55c95cbb5ec9bdfeb766f432", + "blk.3.ffn_norm.weight": "c81932529a5a8c417c27b888dbe95fff8b447c2ea5f6f560444ec5d50b93832c", + "blk.3.ffn_up.weight": "305021735afd8669afefd713f56137248d5e817e60471a112ad06b7fa07ffe88", + "blk.4.attn_k.weight": "cc26ba5c5c28082a79e6abfe61186029e80b145252ca6a7924c437f0bcf2d51b", + "blk.4.attn_norm.weight": "302d251fdcc91f7468cf33f80b49484251d8917d7018ad264ab3a85c8ecf9ddd", + "blk.4.attn_output.weight": "a012f5bee3520cd4ce51f0076c132ebc3653309f304032ad051aa308f55f36de", + "blk.4.attn_q.weight": "3c8d607e447f5ef21e73af71e3c0d32fae16f91f31faae34ff06912cf9cb68fa", + "blk.4.attn_v.weight": "49f6c81a634ce46d71c2350206ecbd231b1732af96e4e4e67693c41a07e007d8", + "blk.4.ffn_down.weight": "e89504f311a4a34dc819a67b761022f14d71c43df3ead4f892c87aaa8e9f0adf", + "blk.4.ffn_gate.weight": "18b22f079a2fbaefe3572eec61fdcd996fd747724e2f0ff4f08cfcb43eb7bfb6", + "blk.4.ffn_norm.weight": "22415a492c168a0878912b05c854a631228b01c3ea8842e1d75989ec46c18a65", + "blk.4.ffn_up.weight": "f57379eae2874d8853f14ddf0f0fcc4ff1338574d5ed5d7e88331d5fb84f5642", + "blk.5.attn_k.weight": "d627af853c40bddf9762ce3988008c1ff17f2686fa8f73a0b5da38010147c316", + "blk.5.attn_norm.weight": "9ce01092c7f7f1c3ef72d6b794da12d77aa1f6a24fb96ba1b9bd5a0bcc3e2443", + "blk.5.attn_output.weight": "0388da8064c4b6b795ce2d8079e8a36535e82b2c9cf794e38ce8ae460aae726d", + "blk.5.attn_q.weight": "039b7ce1c909761fdf475c06cf14cabe5a90199282c89e4dcf460e95a4b6275d", + "blk.5.attn_v.weight": "c47bfd8d2496bdb6e00e03b903e15fd0ee806a515094ec257e43cc433147ab7e", + "blk.5.ffn_down.weight": "1d62e6708974bae318cbf00a8bf621d9ba0537e549ce4710a536520a8d14168e", + "blk.5.ffn_gate.weight": "8b42b1b11c92db19985094cbb50434e3a7c9cfea71ee6f21ea79eae7c49284a5", + "blk.5.ffn_norm.weight": "e0bc520f1505e687ec391d632a381d38d8ebcdec19f614a11a2000ab573e8b7b", + "blk.5.ffn_up.weight": "8cdcd17d2ea89bb9ab902dbc6bf3f827fa4ee029c6bf19eecbdefd146d8b6f2f", + "blk.6.attn_k.weight": "5dc6bcff89794d1756bf57ec665b58622d9352130d31082a6c66e1a079f99932", + "blk.6.attn_norm.weight": "13b26008abe0f119b5104b9d78ebd5e797d3cdd68122b93d73a3b4831a54d085", + "blk.6.attn_output.weight": "f5a49917ea70c3fb311ccfffbfafa63ab18416a5d55e5429b70ce8bfba57c075", + "blk.6.attn_q.weight": "d9c2f652c87dbd09ec3822e12876648fa32e86553ac25afab723b1cd9f8cef90", + "blk.6.attn_v.weight": "5ecc5fe67609a35151011cb526f45c56fc0a999079ae0ff37c755ca03c68c555", + "blk.6.ffn_down.weight": "0ec125ae0ecb2d9277fdb1b04f17efee94e37d0ae37311057c212ca2db3fe6d1", + "blk.6.ffn_gate.weight": "fa4d6d38355ee8aa3b80b476d65ae7e343c9b7770d7b097fc848ee8a6e091d1f", + "blk.6.ffn_norm.weight": "30e8f7defc627532e1739dc76d31223d45767391a431f925b63dabe334b0f392", + "blk.6.ffn_up.weight": "6b97cc32b290fa9087806b5d65aa6dc1760737730c8c71394cc4f30c2157f9ab", + "blk.7.attn_k.weight": "0231cb127cb7c3714cd72b8f39343891d7715a9bab2237ade9e7bc5f4ed2e68a", + "blk.7.attn_norm.weight": "7c3187f07eead7d219d98ab2daf87905e88d5f1ace109b6f5fa55dce3914981f", + "blk.7.attn_output.weight": "2f30ad972c284ae7c8eb0482053433495ebe8fe9c5ee2c28b4bc4ed1f33050fe", + "blk.7.attn_q.weight": "3a2b4b8d61cc9956d304fa9f82a9e65b4bb9fda2196670b16df7e0d8c43eff2c", + "blk.7.attn_v.weight": "d2aab97d0dcf0f61dd2f32848f7a8a99c423a4948a660a660a03a546972b8db8", + "blk.7.ffn_down.weight": "2270d520468c5549cd30023ff9c452a277058310104c4239a616373fc5a94387", + "blk.7.ffn_gate.weight": "4134a3ef71b3eac8f76b6f1a2e58625b3bae48081f175994bc3ed7d8b0d4f2d0", + "blk.7.ffn_norm.weight": "42df4abd4b8769b16f3930068f96960af1b061f1aeb7505384f272233b2badff", + "blk.7.ffn_up.weight": "c920549054ec16ff8c73a72f5d837cf4e11885e44db57c1c1c584c18fbd7a9a5", + "blk.8.attn_k.weight": "01c609bd3bf31ce65688f1f640ee413740e821330134d4ed1877a3065d1527d5", + "blk.8.attn_norm.weight": "48857411f769b00290f4e4f2e593e092781fdc2503f80c1e3eeda1b85a20f74d", + "blk.8.attn_output.weight": "90fb273f8df83744554bd59236515c16c5a5a698ca3fbedc17cc89ddcee354ff", + "blk.8.attn_q.weight": "ade617ac4653c7f00593dbb51837a468afef20a14eaab3780fb96ac3d6714369", + "blk.8.attn_v.weight": "c2c37496494864fee5c527d1fe1f88529d31c73f9cbd02ef9b2e9b23611ea50f", + "blk.8.ffn_down.weight": "2da58572e9ad79087c03cbb0c23c9ef69f93ec221fd5fe4ed92fb93871d23ffa", + "blk.8.ffn_gate.weight": "4483294e628edaa4901708e73e92c917bdd93b780fa01aa74aed57166f2bbf0a", + "blk.8.ffn_norm.weight": "c0cbb7a4f8123b62f0c4652a687f3b394802bc32870dc446eefb709e42043a7f", + "blk.8.ffn_up.weight": "9eaf8a2060cb9224cd585997cd671866c4051ad885c2c6d9fdc7056c2a5c0d89", + "blk.9.attn_k.weight": "5dd36c45fbc9c50fd35c36cd75576288506971eac5c5311d4f5c16ef60099645", + "blk.9.attn_norm.weight": "3c8ca64f2f75ed7c8fc1da010c23be787648139a96ca0ef3ad10be7b14942b8d", + "blk.9.attn_output.weight": "6277e1f833024f53c409be919ec76d34464a78b278c8f9dbf79e777746e3b995", + "blk.9.attn_q.weight": "87352b70d9e328c2d51d59090cf5ea5a046529864a890d0bc8986447a0a5c006", + "blk.9.attn_v.weight": "2efdf01161d7a82a9117cc2d87d37dba5ffefcf730781cb94fcc95130e48ff9e", + "blk.9.ffn_down.weight": "e7658a2ca984961c7ace16acb679387bedb1fef656b5330bbbf588db19673a75", + "blk.9.ffn_gate.weight": "773cd330d4ff5d64be8af00adf2e2722fae4e33fc26bb9d03549f6f4b3b0fe57", + "blk.9.ffn_norm.weight": "c8b86cd5c43b332f72060b807091c33a258e5dac01358ff4733b916cd34c9c97", + "blk.9.ffn_up.weight": "d8cc3bcff18bd46124ba2aa7caacc71220b44eeef6fccb993b4c6cb53e8f2c3a", + "blk.10.attn_k.weight": "964bdf3b4e77b915a216f750ff7b0f2eb1dd6bfa071358aef21010b90111044d", + "blk.10.attn_norm.weight": "59ed411d91d14775764eb514acb0895a75a10cbbfbc1c15d453bc50f8046cb7f", + "blk.10.attn_output.weight": "4d35a2a44cfe4ac0a83fd3ab0dcf1f5a0bf54cdb3b7be9fc353ed32c8a3eb81c", + "blk.10.attn_q.weight": "defff5339450dd881ac352f5c459293f39e07b9619ebd10ed632d79a3f310278", + "blk.10.attn_v.weight": "b9803e8d6a54acea58f662d4c0a5c8ebdf986676de7dfe12d4b288937881ce93", + "blk.10.ffn_down.weight": "eba856be64e4be20b92fb4639a783454dd92427250759df92a337e39f1971c08", + "blk.10.ffn_gate.weight": "2d5c509b066584db4de3632b01234e86edcde35409c5ebce18957dc80fe465e3", + "blk.10.ffn_norm.weight": "ecb9a8679945ff0273856624ce435dd250ffe5a440ea0861a5c84f0e4c44d2c6", + "blk.10.ffn_up.weight": "e76ec7e993f399af02958778c643aa78368e3067846714165eb5aba9d5f547f5", + "blk.11.attn_k.weight": "29c6d1f34bd3ba2f0904e57b32a5bf8dcb2834d439159a33edf234ce0b775677", + "blk.11.attn_norm.weight": "b5817b275149cd2abe18a6a10e19854605fc58fd364666744362ceee8cfe49f4", + "blk.11.attn_output.weight": "1e05653220e237cbe0cc770033e183c9a0eed5680510997409b16186c6691950", + "blk.11.attn_q.weight": "03db725ae669151e4d536e50285b3b047ad097f52475df208ed3e790e31a44be", + "blk.11.attn_v.weight": "27cdf1d4e971326c451a4615a0b79a8c7fe9508f9b76c0d52fa01971fc7eb403", + "blk.11.ffn_down.weight": "176938cd7c2966094f614cace8ba568b10532e45a0d438f80eccd19b6c2a7f87", + "blk.11.ffn_gate.weight": "9782339915dd6fa70013628a01524ee1d01ad8beab04068da7ac6a5ee7603a60", + "blk.11.ffn_norm.weight": "8245f6391e3be97811c0ff27f0d8f484ecc82a468a837c893f059745bfcd95eb", + "blk.11.ffn_up.weight": "15616ddde096d0d25e906375c548b6de4bd5576d1f6b68eefdc29f14e183af42", + "blk.12.attn_k.weight": "66dd21604993edd1b1fe547bcaa06f5bb7e31c9204902d147a227e4badf7feec", + "blk.12.attn_norm.weight": "23a69f85dd8a0904b9839cc5d0afcda299b74e82ae2642106224a1c820f2b761", + "blk.12.attn_output.weight": "4a98d132e376beb274a39d4ea9b6a1b870ad5c66625439d7ff6f45c229c3ca04", + "blk.12.attn_q.weight": "1c6c309d63afcfde32fe37257e300a78e25d01117e33490801107c0e75d1ea66", + "blk.12.attn_v.weight": "723d9e4ebe4e2b1974afa01d8f512b52933698fa36717dd47b37b07760c50a10", + "blk.12.ffn_down.weight": "00e0fb09e1f1fbbf3803f1dee373eaae7a93756b6e13063ab77f9927bc6f996a", + "blk.12.ffn_gate.weight": "89159f7f97aefb1e100107e3ac2d694e1008ad873f79bb953d60c2c1bb22724d", + "blk.12.ffn_norm.weight": "5f70aebd0e43a39d6373d8658cc670c13aadd7818831d3d84f761d5f688442f0", + "blk.12.ffn_up.weight": "faec21b446f061eb4dca561a3180712724347b77a71eb312e7afe9be9e89fa04", + "blk.13.attn_k.weight": "3d440825d19eac3b1753b34d94fee2b3a3cb6636c10b2703ffcf688d3c1eded3", + "blk.13.attn_norm.weight": "47b575e57e410738ad13fd3c74bb49c06b3d31030910834ece509cd1a5c6d9be", + "blk.13.attn_output.weight": "05436d8e613f4475741c1798a7c371b53d61b229507fa04fe23c504ba1f0e12a", + "blk.13.attn_q.weight": "002b5024ce520da41256e3ded5cdc60e5ae07ad9b202cb19d76ab511efd02b1b", + "blk.13.attn_v.weight": "c1f2d6763587c50312cee0d7140fa2c7ee326f5b172bc99b2d8946e08329cabd", + "blk.13.ffn_down.weight": "b5c4e0d8a3ff96cd76a135e415b89f02d28c28f7f3c16a36af31ef0ab8773da5", + "blk.13.ffn_gate.weight": "ae06e9e3d2e1f64c7ad23a4009dc904c2eccd7241f9f91c4974ab2504f116be0", + "blk.13.ffn_norm.weight": "e44a22321bcbcb4a3c345b504e939e8071370f54a8cd702fabdb40b97e0d7683", + "blk.13.ffn_up.weight": "7e6f366d538e21ad431264b12c011892d0be9dfe4c4da9f730af677f920641ba", + "blk.14.attn_k.weight": "95492d6417952ec24b2cab87bceb750fc7e95ac6b1944fc328a3852d980164be", + "blk.14.attn_norm.weight": "6b7b09e1c51addcdbb160ea59edf032531421c520ec5645fe1ff9ca4180cef54", + "blk.14.attn_output.weight": "75887474e4d72c218e6ab0f69f1bf3ec3dc414d51b36fc59df00cdb23421bb6a", + "blk.14.attn_q.weight": "940e33f76e48c21215d19e8a21234c8246d4d084381a7d9806aecb24b071d5bd", + "blk.14.attn_v.weight": "c58601cf5a9833f80f7f9a5b2656e8eab5eb133211446ebd48f8be15fed4ebb9", + "blk.14.ffn_down.weight": "f9f886e7f9b2a54d717b08947a25a0a93e8c2a5b8bcd5a907c06817c8ee3ac11", + "blk.14.ffn_gate.weight": "727ed0ee68594a3f59d704ed3240b6929f083b9c36650fb848d182315737245c", + "blk.14.ffn_norm.weight": "bd2471008ff1b2bae9aa26bea019393fb2bbc5b9493b8cec3ebd2c280fca24ca", + "blk.14.ffn_up.weight": "b006446769f51e4f93b503c4727deae897bc1fc7f4fad49f85024b63c4548d38", + "blk.15.attn_k.weight": "23bb70f9035356624039547a603e46be7d1e4403616eafc2451cc09c5373d522", + "blk.15.attn_norm.weight": "718cb371ca052eeb3bfac6ac506abb887df125271821fd171797a7f2d8dd6313", + "blk.15.attn_output.weight": "c76a2695a204b43a8e5acfa5720590b5d449a9ad9e082cbe3e80fab5903ea16a", + "blk.15.attn_q.weight": "2b3e4037b9e91bdd26d6e8d904cf39f948192dcf09bb6445cb55ca058d4f4626", + "blk.15.attn_v.weight": "7c15e89b6acafc8619e86aa9d412f5893ab17843ff2cfaf40eea9637b24910c6", + "blk.15.ffn_down.weight": "e16fd4bdc6d1c1209c6b633454df4992870c8cefb2cb0e8c92a7e489e9fb5d19", + "blk.15.ffn_gate.weight": "95a46bea366c260337c537fde06b4cbeaeec52484a69c3390bb1d178eb0525c9", + "blk.15.ffn_norm.weight": "37730293f704da265dc6d1896b3be00c39c0a41dab07f573af39dc30a481d623", + "blk.15.ffn_up.weight": "ba74a199da2d0875d7410824238c4ffafbda3993568812284a72b8800df91f15", + "blk.16.attn_k.weight": "f58f79a2a91c9a763adefce0c53a71eb5ce6bd8442f4af554b04b58083bff27e", + "blk.16.attn_norm.weight": "0c16e41b95e81978e0e0e3b338e2afe2d297426578cacee94de15df74e94eaad", + "blk.16.attn_output.weight": "ead22fc337514e4add49aee19720008558e52090466866e849671953a1fccba4", + "blk.16.attn_q.weight": "ef59c4e8fe8918c1add43d7e9c6fb3ef799dd3e1bdd731ec7b6a4a6f97c86048", + "blk.16.attn_v.weight": "902e6b84c2b64241470b13e6f412f859f66b4b223bcfb9c15d5cb1106b07ef3b", + "blk.16.ffn_down.weight": "2ad6e9eb4d8372c32a554395d460d17cfb02d6dbcb757cc962b6bfa36db4f5ee", + "blk.16.ffn_gate.weight": "825b2d50fcce3dbe6a5d8d8a50a95466f83ca4a10343efe67894c20b4628fb15", + "blk.16.ffn_norm.weight": "3bf6ac90befb0e17e077c8ea9454a8485a30f89f2d761ec7751b60c90aed1af9", + "blk.16.ffn_up.weight": "9fbdd08739b32411f5ab0252174d386bab19eb0b17884862f760429b7d41d78c", + "blk.17.attn_k.weight": "4033398718bf3674830ed1b73071ed8482b6dd4ef27f31a6c5fbb998321b6c07", + "blk.17.attn_norm.weight": "714f2e8ac9592966a0f1c02ee979eee8f84586405b992e8ee9543e840199ffa1", + "blk.17.attn_output.weight": "b6bbb618597d767b8f535117be68f92911e4a71d4eb4d8b5d943444151445ece", + "blk.17.attn_q.weight": "b84a0dc00ceb515faa2628125dcec502eed923077b21cfe900a4ff16c2e5f9ed", + "blk.17.attn_v.weight": "4387c7d6a17da9cc7a6bca8f4a75618b20407d570792056283a8e93b6ec65f18", + "blk.17.ffn_down.weight": "47db95c6f1e12b399c3eaf9ddba261782dd71173dd163b52af96541cf87b5196", + "blk.17.ffn_gate.weight": "59abaded0aedfd12f01df81f7a811e84db6a227f51b60abe9a247ca726e87392", + "blk.17.ffn_norm.weight": "b7e86445be5c7b722e01ddb98d5c7527ca86cb827ce0354f2c269e0f2558751e", + "blk.17.ffn_up.weight": "8e31c293bac649d2f60da4b3fc4a3acdce1111ec6058d8805eeeb242443011de", + "blk.18.attn_k.weight": "5ce762ab7b032511c131df81093b587871718c7097f79d8e07d707571f18a47b", + "blk.18.attn_norm.weight": "1f52cdc7af1f4dc1f0ef6ad1ad02e18cda32133654e57cfa9c72ada9c0b1d995", + "blk.18.attn_output.weight": "6486957f30bf8a88516e25772c6650f98b13923f490a2865a8752e36439d1cfa", + "blk.18.attn_q.weight": "93621c8abf69d2ca29c5207180eb628fb2b544d89de6c4a7fb0699be95534899", + "blk.18.attn_v.weight": "11604083b5a74828ac1d226af015ad5dc0215a1fdca44fa7131c2163c02d8156", + "blk.18.ffn_down.weight": "8f9997feb94385f106915df810239c9753b31efda2bf14bdf18a9fbbeec8233d", + "blk.18.ffn_gate.weight": "427c213b3a4e94af703429daf2f65766f70424d8230c123e7e712a18bceb5ecb", + "blk.18.ffn_norm.weight": "c45d305c4ea6a54013ba112f12dafaade064a32cf01317373464a3618d8ba44a", + "blk.18.ffn_up.weight": "a2811f2e73ac9eb9cce91a21a454e84e230a155244e2cd73f2c12aad3c9b8cfd", + "blk.19.attn_k.weight": "b2daed159925eac58c291e2f1e2000beed21002b03c9e1bc7e7a52e22240666c", + "blk.19.attn_norm.weight": "6307306ede2ab5bffa1bcac3f8b139354678c0376b1d9f5530c1fcb4268cfeb4", + "blk.19.attn_output.weight": "ebb98218b2a9c84d3fb6baeb02c5df264b7ab80d994d1098ba1cd47aa398effe", + "blk.19.attn_q.weight": "4f10df2ad09177e7528e9456039b670d07db22940a49417101b725d239c16724", + "blk.19.attn_v.weight": "30f1efc5114badaeaafa91fa466dc7fa14b1616db433c6f563ab851f7333a5dd", + "blk.19.ffn_down.weight": "be5ec7fe6b48855cd0015b0e430d1b70c620de87a7ff188c7c1afef546d7b6bd", + "blk.19.ffn_gate.weight": "10dffea4213881f8a9b583ee0fd370e033756d32255ed15053f794375b9400e9", + "blk.19.ffn_norm.weight": "e75cd24ade45dca78fdb0cbcaaa2d4a17d83a5a73dcc94ce0ec2d68fbdb2a881", + "blk.19.ffn_up.weight": "63e81bdb951410ffa81bcfba1b94a679ec9ebae59cd1623ce2651ed5d4c78bfd", + "blk.20.attn_k.weight": "c2fc5ad39e9bdd45e73c6e54aecc474388d944c4be1ee1921b7fcd035bad02e0", + "blk.20.attn_norm.weight": "aaa9169171937bdce20c1f057e94e9252f221cabacf1ced12e11b9586f23d308", + "blk.20.attn_output.weight": "a9f4fb496e4bc053e3f6cf2e72e22d4cd2b545ef6c32f7e782c2ef6ebcc21d4b", + "blk.20.attn_q.weight": "5a07ac619ed251494170b213921ef3fcc4c2712839da262516d9d5b8ea1ff185", + "blk.20.attn_v.weight": "d6689473105d241eacb17f09f06000ee237336916cf5ec4f48271c5b41bcb8e7", + "blk.20.ffn_down.weight": "74be38db51df736f26ede7c6b52ea787e385f181cb66231e2cced4556a25c9b8", + "blk.20.ffn_gate.weight": "ea91e06dc3d051c0ba0243b5a8bb40edbf254eadfb54fda7247e05cfdd88cbe2", + "blk.20.ffn_norm.weight": "5fbd357b3d6f44a7a91e8a4fc246b24303891b7957e0f3c32818ae5dc16ddd8d", + "blk.20.ffn_up.weight": "fe3290333e056af4ed12942ac72aeba97a6b562e2db05e79cd35dd07eab5b101", + "blk.21.attn_k.weight": "201ec6ee95f06ea5eb80fe86fd07bd016d3ae9ab6abd25d631834414e14a010e", + "blk.21.attn_norm.weight": "ea8154f93e06485828475a00b98cc397ac84768dd70e06ecc0c075b5712d7276", + "blk.21.attn_output.weight": "9f8af74d531478fd304723fd8e4e01578db598441b80dc7c960cb801dbbc501e", + "blk.21.attn_q.weight": "277de9953a8d3cff894ffd06c15ad0ee1407e319df0c1a693d4f45fa9c74ac7f", + "blk.21.attn_v.weight": "6bfdc16cfb898909b7788ddd39dd04b928f31d6732772195d53c558004638dca", + "blk.21.ffn_down.weight": "173877146cb94801157796ee9e5eecf3f46acb3b5e797f90b83a3fc22395eb30", + "blk.21.ffn_gate.weight": "53146713e2ca1be80496024077a028f6b6d749b02e71003c349e113b436f48f4", + "blk.21.ffn_norm.weight": "b28b97e18ab20a5c553ba422f7d7f6014f5902f1d62a69abd20d9fe19a5f9462", + "blk.21.ffn_up.weight": "5c39d0ac4d602b8ec8909dade93b2efcd6b6d9d84a19b252d76bb66dcfaab87c", + "blk.22.attn_k.weight": "01f26272c82917a87a3ccf922fa1d521a952b05de878241b7efe3525b617ac87", + "blk.22.attn_norm.weight": "5ffc96249d8873b506e9eb7158bdfd07fa1429e53c1951430ca7505d25f11c76", + "blk.22.attn_output.weight": "9c2201569358f720244b9c9497e4da02585a167b1414c8a506b85ad75ba990d0", + "blk.22.attn_q.weight": "906036eb4ddf027f6d920f9356a6a2a5e529b96f4e1231a0496d46b4434a5842", + "blk.22.attn_v.weight": "30ede8b0d166003a4b8a81fc99437f557719fc36e5c4dd510c9f161f36a47e73", + "blk.22.ffn_down.weight": "d04c164beabab30e1837b843e18852260efccfbb9d96a34ddd816e6fb3ba23c5", + "blk.22.ffn_gate.weight": "19c889db6b19179f0a62d5981a1506592c65de83760d67afbe00d202202750a8", + "blk.22.ffn_norm.weight": "4885eff2d851b32dbd306bd632c725857e6d164f0fa8b3d5857e572e6ef98ee9", + "blk.22.ffn_up.weight": "365594d8db8e95cf87cc33ac23947942dc326110175cc8ec5a07b5c7059089a7", + "blk.23.attn_k.weight": "badfea1569da0fc6ab817c5727ca3a69b07d9cfd622fb8be5e66678d5b3f7ae2", + "blk.23.attn_norm.weight": "8968f78a379ac3ca5458b4ed4251e8d9112aca6d6dd1ef6440b4bb0b380375a4", + "blk.23.attn_output.weight": "93e43393c03956287b1fe31e9735ff1cfe84f4ae56b83dbaebe96275e4e11831", + "blk.23.attn_q.weight": "aaff73c725a8700ae66bf26ac8869dfe96738eff23a8ff340de2ab53400a5795", + "blk.23.attn_v.weight": "3a86a8dcf14a746ed1411f5a7e634064bc4dfd6511c24cfeccfb2c9ebb6b4101", + "blk.23.ffn_down.weight": "d4da6f37bd7ef69bb203f7b0dd59f50bce37432c70627e6cf274ab81548af5cf", + "blk.23.ffn_gate.weight": "5b6072936c4a693923bb4e3d1473fd45545cb02fc07799aca458ef0449a04061", + "blk.23.ffn_norm.weight": "cd76e37025f84773180298ddb15e0d4ba9cfc7d832e19c791049daa47c6d9c10", + "blk.23.ffn_up.weight": "cde43b99b83124a13b2e4753d12674b3a61dfb34c04703007ced3e8e2aee1801", + "blk.24.attn_k.weight": "457379edc4cce4cbbe107385079019bc922264fdfc7bd1d1ae84343a81460c66", + "blk.24.attn_norm.weight": "0ce0dfab2edeede5da419fa7833db78e36222cf25c358d08f3ec664310f031fb", + "blk.24.attn_output.weight": "0cf91c2fd40c204d2fd4b9c85b69281e5ad4ea8442972fcd44b5fc8e835ffdf8", + "blk.24.attn_q.weight": "87ede30c09eafec6a4e6285674c1bc4637140b168b2da4ed34f36fdb6e176cc9", + "blk.24.attn_v.weight": "4c0b078b2798ca35d6d2c2258fe499820d2bc88700654ba4016e4b028f563590", + "blk.24.ffn_down.weight": "cdb8540c32b1ab988f984484928d39f6841f2131c1cebe90ad9456737fccbcaf", + "blk.24.ffn_gate.weight": "da2e0e913648b5526bd2bbb344038dd067639343aed3b413662b064b0db7556e", + "blk.24.ffn_norm.weight": "8940bd781c610d75eb2be63cfc8d869a3af05e53c963dc7fd4c6f653df5a80ab", + "blk.24.ffn_up.weight": "90cbac2a58801abe11ed6c24560aa4acb949f79429f2aa8ff129ac05868bb87d", + "blk.25.attn_k.weight": "90607131e36998e990ce718ad05cbecd1bcaed010931401ce6baa3b0d93ebce6", + "blk.25.attn_norm.weight": "fbf679c85656c04a6cf8fedd5412c1ace22960e6c2d47f2d43997827811fbb97", + "blk.25.attn_output.weight": "08412724ee7a2086514406e6f68fb9f622e10bac25b0c373b294709f4b09bd2b", + "blk.25.attn_q.weight": "9c1238e98a2747654a0d4371d3e7ea8b979867f609dc42482544f25591e85c7f", + "blk.25.attn_v.weight": "a57796a535c6cb09581cbafd6a91dc14adc8cca2a2465a7ffd0aec546cd84074", + "blk.25.ffn_down.weight": "f7e34e8a6391b480da08b52640613ccadce268373934b409759743a1735b74d6", + "blk.25.ffn_gate.weight": "b8d0b2f4612678b5ce42bd4a683f8024514b75fb5ebf6b22c600811e95582ee4", + "blk.25.ffn_norm.weight": "cde1fdba2369d315f3c6940a997c471ec891924e642505db580d732763bd7b75", + "blk.25.ffn_up.weight": "72e700c32ac8b9c47559c2222e45888a480b527ea512075423c5dc01678e2bb3", + "blk.26.attn_k.weight": "6ac83b3414ae75bf3a9055c32e49d2c40fe611ab21f8444f03d2f465d18122c9", + "blk.26.attn_norm.weight": "55f9d6dc9d75973dc75136ecb9d991b4398097ac133070873fb96ec76a6f60bc", + "blk.26.attn_output.weight": "ebc4fcbd15b33263e50ed2ad45740867cce15bc90e1216623babcb1820734509", + "blk.26.attn_q.weight": "080f057521073e412936fe3fee64fd574c8128fa4a148b879d3e598fe4954581", + "blk.26.attn_v.weight": "0fa2830d6746487ac91b243716e4302361f891e4e008eddd14abec47c7809d5e", + "blk.26.ffn_down.weight": "cb2ab8af1653adc57111ada49d2825c6995e338c8208455b92de10e580f60f31", + "blk.26.ffn_gate.weight": "231ce30966086bce2dc0e0afd34a22a1958cfda7a57c41b3b8e9444c5dfde8a6", + "blk.26.ffn_norm.weight": "35d959d25d17b00617590f5d5831bf705c385c51e46297a14375a700effca6af", + "blk.26.ffn_up.weight": "367680c8d332538b467d1ef87cfeb36cc5c6af564c5023c5fb50e728e3438287", + "blk.27.attn_k.weight": "0bfcb351c6d17aeac5b55a915074fbdf00f11c4bda98babb196ac8804805746b", + "blk.27.attn_norm.weight": "5d598a88c2e75ba59dd7ba4fee940bdec92d72038f1286536d2dfb71d008a09c", + "blk.27.attn_output.weight": "23a9da7347336479f6a10ded14cb3f46e06b5bd56dc4b0fbc526c688552ec840", + "blk.27.attn_q.weight": "b83319dba9055f069208e9c9d66da08bc6874f23e575288fcd81697d1777aa54", + "blk.27.attn_v.weight": "36ed34ccb2f36fdf16b2c2dd225a98ea6b7b0e376e7791191136ccd7bd7a4add", + "blk.27.ffn_down.weight": "5488e1d3a58c71b5e9ddda430540b4776b268cfe1457cbc1c2622dedd9e4526e", + "blk.27.ffn_gate.weight": "4ff48011ee0bac39af704849d9132a2410392c87a509c684f2062f6b76b498fb", + "blk.27.ffn_norm.weight": "32afe99675983da3de2961d1b5ca41c98970a356823597fe29e91f6e86abf0e8", + "blk.27.ffn_up.weight": "1eae3088a75629571fdbf6a20f141bc2bb2ed3f5ba2b9fd1d949f80695e442a1", + "blk.28.attn_k.weight": "c4e80af714962d6f9040d2c09f316f4a1cbc3a2e994e19902d7c653cf3c73dba", + "blk.28.attn_norm.weight": "c1ecf85dedc1c83d5d402bb7c94fb8b9c11f1a3e5f64e7680f80912d4a560794", + "blk.28.attn_output.weight": "72ba47c061b21f5ebc5213a455eaf6fc49c8f8e04ff9ce37e6ed4921b629161d", + "blk.28.attn_q.weight": "c4abc47234307f44b8ca789aa6668e298158fa4b459b2c1e84bd581806591cc1", + "blk.28.attn_v.weight": "aeba950799d4950e491ad0fcbe30334e39b8975177990a2cb339031c45ac153c", + "blk.28.ffn_down.weight": "4e84ce382a37b994fb8608df451a60040559e3f4f3241c3b3cb8989a3ed50d83", + "blk.28.ffn_gate.weight": "04df157acdc8e8534ad60acc2d2a4dd3a7a6610f6382535ec728994fa6f83f83", + "blk.28.ffn_norm.weight": "4d0386dae2bd1c1a9d0f9730718333e3a486c3bc6a5c5d482193c75d39832c80", + "blk.28.ffn_up.weight": "fec60bb0a3daf182a14bd8311fe6dd1e3fd020c5fc273e2549cdb1a2d6b79b05", + "blk.29.attn_k.weight": "b0532a263aa5a4e2a7a80adc83fc5dec974493bd18da7f953e7ebfc3f3a19aae", + "blk.29.attn_norm.weight": "593fc3b4000c35b7a59dace09ca1756c08be0105b2edd354a0e1c16c82898859", + "blk.29.attn_output.weight": "315b896f9f0cbacd0ca8937384c3a3a227efa908cb8c3a9125ec00c480e32b9b", + "blk.29.attn_q.weight": "d482d45386d4ad3394f08e9dff233ee3a70d0427d65c0b8fa05905da7e25ca53", + "blk.29.attn_v.weight": "cd3b5a6e2852da796902930a6a84bc87fc6a7c7bf51f8fc23758d12a39013b36", + "blk.29.ffn_down.weight": "5b3dba6f9753bd1b1ebcba65ef5373dd62c38e755c44b7231b95d93d45761f89", + "blk.29.ffn_gate.weight": "8610d9d2db15c256243ffcca3ffd31786d0ada0af0e7c7aa3fd20524370ab036", + "blk.29.ffn_norm.weight": "1a2ef2d38b7ac3e51190b9ccb8b6552ba83ab290e523356a7f851ddb35dedca2", + "blk.29.ffn_up.weight": "a5fdd15811bde16dc27677cf1a4c97daab4c28cb12a9530f1a0e573134fdb69c", + "blk.30.attn_k.weight": "1efeb0b5f4b45a85cdf47300f892ac77ac1f38000ec3653565d1303d1fb8c743", + "blk.30.attn_norm.weight": "c73934c182c7fe80838ec1d0b92f50a583f75f7a3d78d822f009b58ad2c80e65", + "blk.30.attn_output.weight": "3a0fd89de2d274614750345d827a9c886a4f97b343a13cdf680390505df596a3", + "blk.30.attn_q.weight": "711e113362bdb067db843c66236704eb1cd3fc5f40e3767143e96d510686ef4e", + "blk.30.attn_v.weight": "82b12a9a74fd3d91b73cc2e841e2b3f0a5197ccd2998afa17020995f880d2267", + "blk.30.ffn_down.weight": "af9f4b1287c0d824ae22d6e335d19e04a70135b835be7caa2435f1d85e931993", + "blk.30.ffn_gate.weight": "e2ab3e6f15f5c50fca66c084cb6a57a2b6b82406d65150e82ea0437b93dd9a46", + "blk.30.ffn_norm.weight": "c1b9c325c83f00e177386a4d7e769945f2995e60950c4a576c0a2c4ab9703d04", + "blk.30.ffn_up.weight": "9b94a21efd419715d82071b490d3b635cf1e8da080620dcc39e5bde976d7e9a6", + "blk.31.attn_k.weight": "0db0d82e3ddcc2c06209f5f013e1d72a84a996c40bf00186be485b909cc268e8", + "blk.31.attn_norm.weight": "2b8b7239471f57140c5cdfe06bd224a4f6326282f99736e44fba4c7b120ac101", + "blk.31.attn_output.weight": "a310b048840cc3ff2be4b84796340e8e2cdf05ec89d14bd3655c109b2bfa9fcd", + "blk.31.attn_q.weight": "f45e0cd95645175ea82813455356d171838539bc3f7676d877c698f2af0a0eda", + "blk.31.attn_v.weight": "8bde008e809112aa7e7c23e9c3099087bcc557313b01306c87efa0a4a30805ba", + "blk.31.ffn_down.weight": "8266fec7e203fbfad7033120861e44984581ff8b6851d01dfb7b81c5d8fa90ec", + "blk.31.ffn_gate.weight": "b73bc0aa5baf006d9ef6403104891b8133671b0992398fe038380b67e0d7e2cf", + "blk.31.ffn_norm.weight": "9c62cc27a7b6017c1df8ad49bff249a8245e8895c6754f402cd44623fda83268", + "blk.31.ffn_up.weight": "5b970a4694ea3171a0167f6e1636d9f00268bc1c9640430ffc35218494884adb", + "output.weight": "74fa0ef08c57a30e633e7117b1e9c805f833e2e5e21434bc79ddf9c92c6d7330", + "output_norm.weight": "59b8a59fd3fbf39353506116e43e5e76edd0cbf2a2873d869da4cf27a04997c3" +} diff --git a/convert/testdata/Mixtral-8x7B-Instruct-v0.1.json b/convert/testdata/Mixtral-8x7B-Instruct-v0.1.json new file mode 100644 index 0000000000000000000000000000000000000000..a159653243aaaf82144edd8446f2f586f440410c --- /dev/null +++ b/convert/testdata/Mixtral-8x7B-Instruct-v0.1.json @@ -0,0 +1,348 @@ +{ + "general.architecture": "llama", + "general.file_type": "1", + "general.quantization_version": "2", + "llama.block_count": "32", + "llama.context_length": "32768", + "llama.embedding_length": "4096", + "llama.feed_forward_length": "14336", + "llama.rope.dimension_count": "128", + "llama.rope.freq_base": "1e+06", + "llama.attention.head_count": "32", + "llama.attention.head_count_kv": "8", + "llama.attention.layer_norm_rms_epsilon": "1e-05", + "llama.expert_count": "8", + "llama.expert_used_count": "2", + "tokenizer.ggml.model": "llama", + "tokenizer.ggml.add_bos_token": "true", + "tokenizer.ggml.add_eos_token": "false", + "tokenizer.ggml.bos_token_id": "1", + "tokenizer.ggml.eos_token_id": "2", + "tokenizer.ggml.unknown_token_id": "0", + "tokenizer.ggml.scores": "e3d3eea80bb41a1213f2d0aa3e8a38581d1f19323be77dbd779c9c7e3b72e676", + "tokenizer.ggml.token_type": "6040635e6bd38d98af06698feb75c1802bad35180ee6ae0a503e38c0f60fd71e", + "tokenizer.ggml.tokens": "604ac4bfbd019e430d7b6cdf18c6c0cd5b967900601f0307f714ec7773aa5ca6", + "token_embd.weight": "1d1d1d39a867d5a4bfb32792a47247d2638c10c95a6259391d02843583505cc4", + "blk.0.ffn_gate_exps.weight": "2e5cd43ac3f26c44f071926ff6c3f239ecc52a34bc9a5b5906d3d4c1bf2fbbfa", + "blk.0.ffn_down_exps.weight": "a4dfc7e7c96e7402eb70279601675b956bb7331da8101e63fe5c0a611b6972e5", + "blk.0.ffn_up_exps.weight": "2d5d87b378b2319c344ed2c642598b6f7cb6beeb582a8ea51abc9ae690d473c3", + "blk.0.ffn_gate_inp.weight": "a46aaf5aba7401ce6e41f158242b4879d34901661f3ede85496cbd0ce79d6314", + "blk.0.attn_norm.weight": "3fe37d913bdd2b65076bcdd6efe64a37b0b03cacbb1b80b9f7089068aa35f38c", + "blk.0.ffn_norm.weight": "5e14308a3c894734eb204c8f558bdc817e94bbd5b4e9cb4094e91ba388c8f7f2", + "blk.0.attn_k.weight": "73d943dcac0911e87bd771f4aa1c901e1bfe1aed293af06e1a67812159859f67", + "blk.0.attn_output.weight": "4c5f754c855e262e8d4c94c6fbbb57af06399dc0e170d7d99a1a17fc9aab9227", + "blk.0.attn_q.weight": "d6fd7403c873d49c05f6f03208f30d99ad34cb3b71c9990c47334d502a8e4c7b", + "blk.0.attn_v.weight": "cf17cf64b2d683bd9de6cebaf60e5c264df6fdc38fe719dde9d54c80334f6366", + "blk.1.ffn_gate_inp.weight": "0d524de81cd915816b4e714bf595ad6946a9130b3de731cd89428b2781230809", + "blk.1.attn_k.weight": "2ea47f412992b374c70674730fe84700e0c8cce177086ce9b6635e42408964bd", + "blk.1.attn_output.weight": "b4b2520794d54113e86c8ff678eacfc62e35be4395a594a6c8c22b4383ebcc0c", + "blk.1.attn_q.weight": "5db930c98c4f91f6eab57eb974c72210b158e366d23d6d2890b2759c053bee33", + "blk.1.attn_v.weight": "079bdde09668394bf7af9f8bc175017b4f48f0ab64e6dd855a4d7561d1693c0f", + "blk.1.ffn_gate_exps.weight": "146a62de19f9ab093deb101f9640534ffc3dc40d69f508be12fc0475d01b0c7a", + "blk.1.ffn_down_exps.weight": "949da94a3c0f375160672a979e85f7def284264b10d48d038238aad5f5ece793", + "blk.1.ffn_up_exps.weight": "7016a3f467d9e3f2f4b4019579ed86b757469cd367f2b225483305376b4bb3c1", + "blk.1.attn_norm.weight": "1614d1e6ed537737275eb888666c7bac533f4eefbe73dec92b591045ca9e1afd", + "blk.1.ffn_norm.weight": "405a455fa7d1ec36894652ceb554bbcb09a07fd6405f42741e66dc4a4665c19c", + "blk.2.ffn_gate_exps.weight": "90d5003fc7421f44220c0842d43128955e91488f6f785fe570b62d81b719e964", + "blk.2.ffn_down_exps.weight": "ecdc2b5a8b504ef0a7833acff47d69b0c1fa9c22126de1bb120ff5e48c3d6e2c", + "blk.2.ffn_up_exps.weight": "2cbd9485a32460d315eb50a2f3b00863fd77245bfe885b7565efac1cdb1f191e", + "blk.2.ffn_gate_inp.weight": "0d0a17a1a2c7a61f2cca49ecbb479154dc93a870873257bc4f225e7607f2e2c2", + "blk.2.attn_norm.weight": "b2e4c5a977f87a6f880896bd73596234c9b83622fa0d7add5892501e3155913c", + "blk.2.ffn_norm.weight": "0ab875b4280afa922376cfc7b9aa3f7071c9432ea1254091ce7de3749df0e8e6", + "blk.2.attn_k.weight": "bb884af51fb51550acfef54ccf1b58ce8284e587806e6a2f88c8265e1ad05a5e", + "blk.2.attn_output.weight": "0f03099ba1ef342ea61af9cd71d028123bbd8b1dd7d7fd9b509aef77815427d9", + "blk.2.attn_q.weight": "8fad0d29eb4c9d24e564774ee3316b9eb7a4c4985e4567111d2c836c830f6cf3", + "blk.2.attn_v.weight": "fe04c847ff677632401a94e7b6b6fdca60391ab21cb23bd791533115de6303a1", + "blk.3.ffn_gate_inp.weight": "29e3aaa724590c070e614af8288939603d2641b0ef11e8c0f476bebb2776673c", + "blk.3.attn_k.weight": "231cc5631def10f7f292d8862d6125ff555164cd70480ac76362149fad204497", + "blk.3.attn_output.weight": "86467a605c62852e05fda1a7ef43150df2cf715fe59785dbcba09f1c27cfa086", + "blk.3.attn_q.weight": "901822402453922225c2d6ac79616691d48217635d5ff7338daa971d5ddee210", + "blk.3.attn_v.weight": "27030784f44375720df2f090933645a31a022d3fb3b14573e5ca0b78f44070c1", + "blk.3.ffn_gate_exps.weight": "231ba59cc0b988d125d77bf627aa3f04636684870af88f081f3944b48a160d86", + "blk.3.ffn_down_exps.weight": "530c3ab44ae4d66e8afa4d10c153ba5dfcdfb7321989a988e62e9d12e7234625", + "blk.3.ffn_up_exps.weight": "b85c2d4d9d11332e702b3c0a6610d4f525f9a93e5d12f5c7c55c592c40755e75", + "blk.3.attn_norm.weight": "05dbb6d88cfa6b199f9d705ccbda97c0ef13f9ec875c595398a1a42d009a4555", + "blk.3.ffn_norm.weight": "6880b1c27d46969ce36fac049c05dc8b89e4bb47dc89df357e32df7e18fc512e", + "blk.4.ffn_gate_exps.weight": "a883b4f225b760c5a2f6605dc5e2167ab85bb398c70bf64ceb539fcbd6128dcd", + "blk.4.ffn_down_exps.weight": "d291bb656aae77947d4b525e2819bf4112afece53ff31de9dab999af1f65f9c4", + "blk.4.ffn_up_exps.weight": "38592afb8ba3dcfb26970f906174f7d3fa62da44fa4be4fc6912a19030ea9164", + "blk.4.ffn_gate_inp.weight": "1596cb74e8fd6c3080b937b06468bb397b0dbb661e6d180a6bcbdc43e8bfd0c6", + "blk.4.attn_norm.weight": "f90c83c5ff4366281d283384efc941620542b9cfdea160d678dc54a75e33f758", + "blk.4.ffn_norm.weight": "d28d8c49d1746b7cc085562d1074905fd14023844de823dc4fb22202bb280790", + "blk.4.attn_k.weight": "792bbf412cc357140fdaba543e547a9b2f7582919e307bbd9a80c7d6d8f5f1f9", + "blk.4.attn_output.weight": "d98e4a062d2631d9c315f1990d5f6ca9a88e7e0e46387f611ccb0353f876aa12", + "blk.4.attn_q.weight": "1a11a55a91d9f748a72176ff6b1c174844df406e00d1b66b9aa64dc6ee4bcd1d", + "blk.4.attn_v.weight": "04cb3c02b12a6313c7ac7044513441083d534fb4c5a3f63bbaa58f7edbd2fadb", + "blk.5.ffn_gate_inp.weight": "cbd5cdf015d33a2da6703eb74c22fcb97581fb9175435173b6dc4f9e8364320d", + "blk.5.attn_k.weight": "4fdf3405e4d657403f5647b51233521310ee984b4b81bbcd901cb3e6ab76b7ff", + "blk.5.attn_output.weight": "4a25662c46979a29600ed77e1907cf81fb16ef30e724c155444e54ccb76af481", + "blk.5.attn_q.weight": "e2acb30e30b97300039bb20ad0878f05159d5657fa811748a51d5b6fb35d631e", + "blk.5.attn_v.weight": "306504b6a26aa123c63dbbed3f4ced0ed2ee8fb6a30bf0093539b817539f5ece", + "blk.5.ffn_gate_exps.weight": "7e34df9b9944dbeea5e8565786d3aa6937314a4b87acd4d0874687877c5a39fd", + "blk.5.ffn_down_exps.weight": "c4b7a57a42b5ac0a8ae27dcd5cb2646d7a7cc7123126d44a56ab128e85f60b13", + "blk.5.ffn_up_exps.weight": "09d47593b6dd6c664a9155bff02fc2eb7ac4a70219a88162d05c802a01d3c6ba", + "blk.5.attn_norm.weight": "58804a036d6ac4c1fe357b8b6a97a5c37cae1c2f06ee0086c041d449c1c6ef6a", + "blk.5.ffn_norm.weight": "d872dee6789f0826211aa46ca9d0869e3e96bcace9e77d6559a7b6f3e524f3ca", + "blk.6.ffn_gate_inp.weight": "fb1eae732e974d6c1d020a5b4ef98c5f33016f984701bcea656f999a99daad66", + "blk.6.attn_k.weight": "55e9c59c5051ab5519b3a7962e1b5fa96a3c0251cb6200dc2f177885ad2de470", + "blk.6.attn_output.weight": "f3c834a8d0027370350e2b6294d95434d31432e57be6313b013c15a56303d61c", + "blk.6.attn_q.weight": "efaefe5f11c2140dc7cb532b0832c2a0b363a165cbda21f00fadae77efca377b", + "blk.6.attn_v.weight": "900bd734d75616d846a90a121c97e081c956a3d1ab012f66dd0bc62c43e1ec3c", + "blk.6.ffn_gate_exps.weight": "312a99661b1468fcaed2474621116f1681432755e973f3ee79d01912974fd424", + "blk.6.ffn_down_exps.weight": "ac9cd7db67a2ef0d2b5def86873673d05e48d49d147dd944469dbb8e2d4c46f6", + "blk.6.ffn_up_exps.weight": "57613e7e09579400a1a09fee4445acfbfe83f2f327fdf317877787d96ada6b84", + "blk.6.attn_norm.weight": "0e8801e09885c633bc01a9a5b85d4e878d30158a4eb41a937dc5b760ebd044cb", + "blk.6.ffn_norm.weight": "b8c58062ac93072f878446b0e7f958c737aa47fb769fc3a8f593133d12db2dd1", + "blk.7.ffn_gate_exps.weight": "1ef611732ff13edfa8d30981ed9dac00c15ceba9fc012ed0b199e9280a849948", + "blk.7.ffn_down_exps.weight": "856c6811945c7b0fa461ca17811cfa43436b4cdf5326bad23cbc30883486d7cc", + "blk.7.ffn_up_exps.weight": "6725e3e33994302ee13fa5ec163631ce2dcaa08aadde8fc166c2265d4561c5c5", + "blk.7.ffn_gate_inp.weight": "36b49d7f80c1003dc392b2c1b9960cd49889dd69e77b26b9e4b13d01f3d0a32a", + "blk.7.attn_norm.weight": "7a0ec49acc5e20ee71c6f80ca02f4f1e564c485e0ae0621309e7c2eb0c616cf0", + "blk.7.ffn_norm.weight": "eeae035c39ab6e64bc06a4baa1bf6e50d4c8b8797cb0ad8abd48be86974802c0", + "blk.7.attn_k.weight": "e8f78c1def01a7a38d2d9bf7becb17755e28fefe4927856f7890fbee52840187", + "blk.7.attn_output.weight": "5367f05ac3bb49ef8745ba5902e1bdd4442415a3ebff2c7e1a3918d7be6fe948", + "blk.7.attn_q.weight": "37c95fc5acc55a4f6e5f02cab9be60e4fe54c08b65f98f4455741b4aa542ff4e", + "blk.7.attn_v.weight": "c89f1343486ba55814233511e94090f7365662a8a4214aa4c278cdadc79196c2", + "blk.8.ffn_gate_inp.weight": "4e239afe8c7afb8de3a005757c887cf14b1622ca2d224227591cb0e5301f4c17", + "blk.8.attn_k.weight": "2ad0229f30fdcc1e85ce64e00d8f75902238294844a81d5af43e14ba75c02983", + "blk.8.attn_output.weight": "2e44a4722acb3b521b81d0b910f8ca2f6c286d874a92ddd02150566454061699", + "blk.8.attn_q.weight": "1cd2b09cb2f43e08de776b5f7eac197a5a6d4ffdfd52b21baa36319450147bd0", + "blk.8.attn_v.weight": "5a22c57ebfd33ac500cbcfd321d5b5b1783f8728801db6f3f8bed51c7183e4db", + "blk.8.ffn_gate_exps.weight": "91063fe56cb4f3ff3b41052bb5046fcf8ef61516a603ee90aab893a9d68c15a7", + "blk.8.ffn_down_exps.weight": "d4c3abc8f1d1b462f67f70bd8f404b3fcf45dceeaa8527fa120527254c383c90", + "blk.8.ffn_up_exps.weight": "76a1a1f08ec577716a2e7027b45293e9205751126424f1bebe1de89c78f087d5", + "blk.8.attn_norm.weight": "f980d774da39eb76c52358afac3e38cb4c81cb323deaabbe5c41822e3f17a98e", + "blk.8.ffn_norm.weight": "1c937658cf90f1a85db9a5f26e077730fdd4b694607dbeeb825c5fb2bc407e0b", + "blk.9.ffn_gate_exps.weight": "a2532471ecb7896d5c78e5a34e10cfaf4125265e1595166c8d0d0dfbe2a3187f", + "blk.9.ffn_down_exps.weight": "b47921a28412d48fee450b8b9d97cee42344a2e69f06d407fd9523d7adf13333", + "blk.9.ffn_up_exps.weight": "7c461bd1b2a73b439cff6a10d94afa01e8b06f7e6f09d9a6f28e3876aef48bce", + "blk.9.ffn_gate_inp.weight": "1648dfb08b5c06d7953a5a97ecb764995fae9487fb729a1c867023b2538149d0", + "blk.9.attn_norm.weight": "8635db0f299882a63b7cfcd1d4259c9e53fab22c31d3d054de36b1001380b31b", + "blk.9.ffn_norm.weight": "f9309aa323062d174c463613afef9b0a33501b510bfaa58a8e0e866d12ffef3c", + "blk.9.attn_k.weight": "dfe62030441e947a588512d18d9c6e4ed72c2f71c227d622c095e4263b23dadf", + "blk.9.attn_output.weight": "1977beb75c6349c50ba7dd3865d7c0a9c5c5ddc854413147b0eec98ac4fda351", + "blk.9.attn_q.weight": "eb132596719605cd6bd1782487f121994629e115190edd69240b12af66e734f5", + "blk.9.attn_v.weight": "9e708f15d332d7c5187b0693b1a977eb30a2fa10bf7df48ed9d7537c0aa6ed99", + "blk.10.ffn_gate_inp.weight": "97503a5d166c1925f9b65c0eed980753d411714d66896f3d0fad5286c7aba702", + "blk.10.attn_k.weight": "1ebdd222336bd25b48df1b138cdbe09021c4a5562ea7cb78cadd1255d2be3a39", + "blk.10.attn_output.weight": "5e98faa38e9d514b9057e1c8342c509cbe1083defd518e506f6bad89117d1f5a", + "blk.10.attn_q.weight": "3323a26c87d936d1dd87c577d0b763459fced726679612c874b3de5fc6d969c5", + "blk.10.attn_v.weight": "d5fa73cb56aca388e205f44455e4b4f676fdc12ed7fac4542fbb3b41ecea59ad", + "blk.10.ffn_gate_exps.weight": "225021b53782800906cd13b70be3a4161e8b300b97f984a959ccad6a6e8adcbd", + "blk.10.ffn_down_exps.weight": "f08eb91526bd22f5fd0402fe925d6141cdbb308a1ced0330858d0c85c71f5ef3", + "blk.10.ffn_up_exps.weight": "a9f688350c3b53eaada5103b5848bd9a3d7d6b327a70fa16c24bf28ece933eac", + "blk.10.attn_norm.weight": "5ba426c9dfc79805015ccd76cd1068b0ad3bb7a8453e14bb1d35486f122d8f95", + "blk.10.ffn_norm.weight": "98891d6acbc3986b2581b7a3af9f5946a392d9188972c6a8b15d4e745a4f2482", + "blk.11.ffn_gate_inp.weight": "b2365a60566e7dace892e1cb0e62eb73ce387352601723e847052b34874feaa6", + "blk.11.attn_k.weight": "0efbc1d1430505543ff71532a4fcda821aeac616ef6c1dca40e00d4f2ff70bea", + "blk.11.attn_output.weight": "3d5bd4d9a41236f30d4293edb9ae27beaa113ffb31b4fbfadff3a4c370dfd3e6", + "blk.11.attn_q.weight": "aa11e9db14dd9c77951511443077c2a1a78070753d7bd3d9811038473f69e325", + "blk.11.attn_v.weight": "5adc567f377aa11d1763d35f50e53fb2896a8b03b623ac36acc45efa2486d512", + "blk.11.ffn_gate_exps.weight": "71d07d982aabfab9eed3c733d49c20f023bf475368fc71db5084d91beadc4b47", + "blk.11.ffn_down_exps.weight": "9a06e61461e48b3925a9f7d9cca634d048c8b62163d7bc5c43e35899f959319e", + "blk.11.ffn_up_exps.weight": "bc05494d0dcec61021b3ac0c5bc1bf502736cadf48224e213bc139d562699a89", + "blk.11.attn_norm.weight": "a5758a10bdd0404ae1470e8e9db903985d4d07f60553c5001a5e7b660d4f7ada", + "blk.11.ffn_norm.weight": "814ae037563aad3771787316bec4806c95bf6f5991dd6474b4b1e5cc13dc18ee", + "blk.12.ffn_gate_exps.weight": "3a68b831ba1606fb9ef6dffed4732032447ecef23ea563ff4e79317586c7eb49", + "blk.12.ffn_down_exps.weight": "268b25e13f4b7beab08686e83705a41b21d15251809ee4784526f78a580da829", + "blk.12.ffn_up_exps.weight": "9105751a5b5b42ca2614d0456f24f779d2e2ac8cdff0f96842aa7ae2b70f341e", + "blk.12.ffn_gate_inp.weight": "d0de1558cc1d458c5c504f63ddc59785c323df7330474bb0644c346104b40a3a", + "blk.12.attn_norm.weight": "859a4c8113678e2e202d10299850e0cfb52eb11ea50bcbf4fe3ff39bdd394154", + "blk.12.ffn_norm.weight": "7fbf4c459c1760218877e9ee3f5ad49e960956a4369bcfe96c143f04ff9ddf97", + "blk.12.attn_k.weight": "0a7e254fdf3730a57372b6ff421a613eabaea68cdefd64800857941411318374", + "blk.12.attn_output.weight": "ceb763fc15d88af149d8fb78e82db2b7dab3aeae584af8cf7611a12356a397e5", + "blk.12.attn_q.weight": "a43402d23c46cb2d3cb3c2a98c81b19d10026b7e6742370fed6b2880b6e049b5", + "blk.12.attn_v.weight": "3bc24f2c0480ce91ef72993ee8f1cf962f7359e12183424583ffa1246bf3db52", + "blk.13.ffn_gate_inp.weight": "a6d68c82bfe66d8bab68f980f5f18268a9e2c0cd6b8832ed39010e0de198ae05", + "blk.13.attn_k.weight": "0166c39546b37dc2e01b2b396ba43e183f797dd04eaa51a6d103d8b58ee4bace", + "blk.13.attn_output.weight": "2ce5eb198deab9557475a58b69b11e9874b547e05c23f223c6e42fa35ddca069", + "blk.13.attn_q.weight": "745c1bbdf434284a7fae98f45e821c076dd9c2a2467dba6a9d8cf0041e419dbc", + "blk.13.attn_v.weight": "9ece68d5ac64d1421ea7aa32e1cff9cc1fecf5175f4c4da858dd31d8633e3337", + "blk.13.ffn_gate_exps.weight": "ccfdcb4670b131689de12d396a010b5ea737795cf5c15a14a304d720b3c7c899", + "blk.13.ffn_down_exps.weight": "8b8fb328664764f1aaa5cbdec336d5654e981e965a02ef622bde5f07ea1c164d", + "blk.13.ffn_up_exps.weight": "d2ace0236c2fb3365fdc85499d676a7f65813c48e5085348b1df1799922766ec", + "blk.13.attn_norm.weight": "1ed29d7d89ce52d7cb4d57e895ff7115430466e917136c049c385c030ed44e9c", + "blk.13.ffn_norm.weight": "a194fc542597a4dcfdfaec5e3cba2a2b2b21b21edfc87c39c0d7f7651355bc4d", + "blk.14.ffn_gate_exps.weight": "a625e3574e5e740e7f8e2f9c40390f2f382c720aab5b10534e298002dd8d1fb9", + "blk.14.ffn_down_exps.weight": "bc366f015b83c865946afd74c8a884943e0ea2c671314a0b7bb72f21a44d2f78", + "blk.14.ffn_up_exps.weight": "ee3199bf2086de77b49f57f487676be8ee70e102a2fb5a5ef8ddbbc28a9eff41", + "blk.14.ffn_gate_inp.weight": "2b437870c850fa2e2044d032bb02908af634356e37466fdae260b933e48ee8b4", + "blk.14.attn_norm.weight": "cd8344d193a1cbd42bd898e17f4bcb1ca0b2918420fbdafa9249a6f2b7f4ae06", + "blk.14.ffn_norm.weight": "70eec40374e558fed5b07257283cf36342b6b0129285a00007deb59c32c9f7c8", + "blk.14.attn_k.weight": "4053bdb507e0543d724b632570bac86b31707696d90a0db44c49b2a082e0d599", + "blk.14.attn_output.weight": "0182632cb0e06a07241b8293d25d109fbc1862e1e337d435f908e8681e2eb1ab", + "blk.14.attn_q.weight": "ffc7794a4c1b6f793c842dba969435330a7a80b9212e457b4b2ac33e68b41241", + "blk.14.attn_v.weight": "6411805292d528e61bbaad8f9aab9dd073529a17946c057fb06864fad9cf3211", + "blk.15.ffn_gate_inp.weight": "77d0744567c76e6abb67f81ba9c715b2b544841186d5b948309571eff213bafb", + "blk.15.attn_k.weight": "1f7957954ea4c6521c257b35a360e868ffa02bdb3de91f146d5e06bb4a545c98", + "blk.15.attn_output.weight": "d7809d36bd8d3342240c46fd87bcc7f9821a222f48d9a95e45ae50460265d3cf", + "blk.15.attn_q.weight": "25f509313ae4d8401b871904059f472a26f5714e7c791c725de77a1a522c976e", + "blk.15.attn_v.weight": "96fedf5a591fc0f020e6de10fd72ff12b3ef9cf70cd21dabaa0d3e7b06f54e73", + "blk.15.ffn_gate_exps.weight": "8f950d976b2fd9a3d213b84123cf114c1377efde9352767fb2ddee89e177c8ef", + "blk.15.ffn_down_exps.weight": "6fd09d1557bb94b06efbd4f6a1ca4be532a202ba290e9315bc8da3d12a5c4c4a", + "blk.15.ffn_up_exps.weight": "cbeb59ae7b0266a928dc7e3a6e70a9330b92f9ee1b17ee1ed91022108204a33c", + "blk.15.attn_norm.weight": "2005330911ac2edc7b6d27aca021c67d30d16eb632e49b1a13f30fdb2717aed0", + "blk.15.ffn_norm.weight": "0e9198f3b548eb78acc8961f2b3350d238d26cec110933ba753a8cf0035c501c", + "blk.16.ffn_gate_inp.weight": "a41d1f99d739c8b150c3945b6949763988d0c6a4c5a2b5855592ca1a48ed23d5", + "blk.16.attn_k.weight": "b624e2ec88c2d3047f60530fb87e72cb4a5e655a9663f6f3e9b09e5ad32cddaa", + "blk.16.attn_output.weight": "687759ea75e45108526ffc1573d6fdf084728079bfc2dc89b9979e76280f43c4", + "blk.16.attn_q.weight": "beff3a45c7e9ec82ffc6d3c701126be28654d10aabd747d03441210491fd31b6", + "blk.16.attn_v.weight": "43a349b13f0b9d040cacecd942bcb168c030fef8c75c987d59a4fce6c14e855b", + "blk.16.ffn_gate_exps.weight": "793406d6c13d727c82bb7b692ca98d65ca975baee69fc57be5378d77c5a19b62", + "blk.16.ffn_down_exps.weight": "9bad3dd150d0230404b7f886ac7ff8803225757e813f195cdb26bad245243b4d", + "blk.16.ffn_up_exps.weight": "7449d663023fea3496475bf0a9c1de7272ad0ce9adcb3265e8e424badaa674dc", + "blk.16.attn_norm.weight": "a424ce34c195a401df1ce37ac4f2794e8a6720b1ee8acb21428e2b68c65e0125", + "blk.16.ffn_norm.weight": "405a68bb8e16e1064df2de55ca3cd9ceddda1d9fc0af007a9bd7cad4b2676248", + "blk.17.ffn_gate_exps.weight": "97c6e5321491ca5dc039ee88da0eb0e78f347372785411809af84b3298cb19dd", + "blk.17.ffn_down_exps.weight": "1617ac19788a1be19bac69277408761e6bdf5719d63a8c7fea14d41cc27641b5", + "blk.17.ffn_up_exps.weight": "4ead1c365f112581c10610ea3f63d2a1474311d2503d2060fed4b458ef337f5d", + "blk.17.ffn_gate_inp.weight": "ed4b3393f2523f2b5e0fc7680a1caa2842e605728a529b5af68a7fa8d7abf940", + "blk.17.attn_norm.weight": "beac17ef86a7fb2b5840cc72f7a95a5e3d6bd24e7fa698e0b0ebb9bdac45c561", + "blk.17.ffn_norm.weight": "81cb58ec6d6dc02a0b4ede10adc336dc865fa76f982d4eab0e4a37b40f5b0fac", + "blk.17.attn_k.weight": "eab569e5ea8c8b05e5a6a209fba031129453c2e28181eee3e736b3b04b36bbec", + "blk.17.attn_output.weight": "f85b70f01438ce8fe5d10599b113f30bf18dee2bbae0657d3eba295870001db3", + "blk.17.attn_q.weight": "887ceebfbf6a2b94b43d2df4439ac3a5bbc29311d4b28addc04d525546032047", + "blk.17.attn_v.weight": "2df9414d65014c06a93da22ba3a668be7b83e2e8008e98d7771f7dfebed98298", + "blk.18.ffn_gate_inp.weight": "9b07741a0950fc667e5fd25937e33bc22e1f764f80eb4ff3119f005327ae0f6e", + "blk.18.attn_k.weight": "8649598dbb63938744c39bcda5ce8c31773e29c573be8d4d2c114f5030f8d3e8", + "blk.18.attn_output.weight": "f8e391adb92622298ca834d5d1eda48b69c3b1c51c5a584ef6c54a725c298d75", + "blk.18.attn_q.weight": "84bf8708a2eed618f48f69c178ed7dd11fa4c468102376e72e910ebd037d131f", + "blk.18.attn_v.weight": "31db3cd773f09548c2c1b1eac2718e46364a7810970fe9c433fad9d8de5397eb", + "blk.18.ffn_gate_exps.weight": "be2a2ba378002f1b61f86c273a69eede9b93786d5ce96b4fee1861f730dca4c4", + "blk.18.ffn_down_exps.weight": "d35196159e37705db50a5343e3989f7335477f1a4add67ef42ad64a638cd07ae", + "blk.18.ffn_up_exps.weight": "c6ceedd86e97913a6dcadc838e7abb762d629fb8dd55f15cf02fd9bd66d2ba78", + "blk.18.attn_norm.weight": "41f0b1ad83d6e3cb9fbe0d27878c2e7ad4a351b9f554a6bc9117c01745cdf6e5", + "blk.18.ffn_norm.weight": "96646204bd0d82f25dc77faba4dbd86b1332e449313e6684e00122da8be99057", + "blk.19.ffn_gate_exps.weight": "c6eb7f61e7938bda0492dbc05e51e8f631c99224fe18e99861fc4fc53ba9e9ff", + "blk.19.ffn_down_exps.weight": "4384803da3a3a3d44120d7dd192fe2c9bbd9a1a0cb492dbec1fdd7565230f1e8", + "blk.19.ffn_up_exps.weight": "22d73de2fbb8bb0f1bd2caf17fad8a355c47d914143f7f6e6d0128f66f074a60", + "blk.19.ffn_gate_inp.weight": "9a0cc4a2301a5634022fbce41189021bf0d1a961792d2d9330fd35556d18e5bd", + "blk.19.attn_norm.weight": "c5cc56ec5df9a1f7d5ad71fbda49f1433132e58895d45cb44c73420bd61ebd6b", + "blk.19.ffn_norm.weight": "77e17de741742ef2482fc7872fd423c8e3c1454dc4d2be89ee939084b6d78bc0", + "blk.19.attn_k.weight": "a92ea36ce2e3569656306aeefb835ccd5d1b03b33a86e0d3d030644cc923b813", + "blk.19.attn_output.weight": "5e2a912b37855f84ea964907a1a86d609cbdd79efa0c93c3e8e2fc07caf7c226", + "blk.19.attn_q.weight": "4ef3a5913292ac3c1a6fd3e9e53d011021f2b41d0276cf849706d1ca925cf7a7", + "blk.19.attn_v.weight": "42981b75b68ae852cee638b5433605c147da4392aaa6d7a06e756115b0171f39", + "blk.20.ffn_gate_inp.weight": "71381b9879a7c80b9f7b475abc0aa31b8cd71ccc00856ebe89764a2acb9df2dc", + "blk.20.attn_k.weight": "1928b7ebc054eb3967929ed6fb446314d5352f4aaf8b475ce55c6345019f2ea4", + "blk.20.attn_output.weight": "6071ecd9ca91af0d2ba93fef4a1a56f3b243dd70f862a21a2d164d56f386043b", + "blk.20.attn_q.weight": "002e95042a40f36ceed5829e3d0c8072e5f5e4ee86a089e2902b2348fed24dd5", + "blk.20.attn_v.weight": "42f509cdb1c0e298f89f896e349be86952c5168e49b3f83bb17badbcb7596d57", + "blk.20.ffn_gate_exps.weight": "a684a3ffe4b0a57c819a5fa9cb3521de223f392732927271e97ce925b6e33765", + "blk.20.ffn_down_exps.weight": "e3081a7bc7ba750d8a4886bc8ca4f231b55db4ca082b54b4106c7531964725cb", + "blk.20.ffn_up_exps.weight": "fad0fd5eca36ab154788da28be8ec25bb5d6db06c9d133db89e96df358a2f6a2", + "blk.20.attn_norm.weight": "c3e3f2429715ae95e884ef1246b0b461b23c5cc0ed08beecf70a14cddd184820", + "blk.20.ffn_norm.weight": "ff31f609dda65ca496b0584fabea6550e42edd05ebf229812aa6b7bb5ede15e6", + "blk.21.ffn_gate_exps.weight": "366f09ef0ecfb86808eb3296cc9abdb957951d27f6533c03f1422b54061da660", + "blk.21.ffn_down_exps.weight": "3fc495947d27fcca7fc0893c8a96e5d48ba27b2c8c58f8fcfb8dcfcd5539741c", + "blk.21.ffn_up_exps.weight": "6713ed51410bcc8283cbb001c4ad784098f25701e8021f4fa4f411e186859c4a", + "blk.21.ffn_gate_inp.weight": "6d4c92c01ec801647134d907bf1108878156df266a6107abc10526332b328b93", + "blk.21.attn_norm.weight": "27605719ae2df24f4f2e85a730927cab20367631612cb501631f6bbf38eb1209", + "blk.21.ffn_norm.weight": "ca80ee8177db185b15a4a378c1cb6f7143c76546a7f1726bda23f329323d4ffa", + "blk.21.attn_k.weight": "9e49f743d4a5bda9b4bd9c40c2ca37cdae5aec7e54cb193897ac8b4945ada14d", + "blk.21.attn_output.weight": "ab923540879753feaed152f5950f69cdd83d8f2413ca873f5f038b63ab0aea12", + "blk.21.attn_q.weight": "62617fc3f1c9d2aa672a4d91a121c7a91b92d145b65e75f0b06b4bb7c825dc36", + "blk.21.attn_v.weight": "15f8b2e72f8e8e992f2f6b3e93238a9d7be7bd6136f91c9d04b4b4cd0cd60369", + "blk.22.ffn_gate_inp.weight": "3ddb1773d9257b68add7a2a4e94dad25ed926803e02707863dd742ab9b2dc179", + "blk.22.attn_k.weight": "680e45a9e8d5feddee5266e119dc053bf80718fa9af1cf6803e6f493b265f1eb", + "blk.22.attn_output.weight": "0d5fae3402fb2c5aa3a860010e3973fc8e3168d1015f7a76b7b2964681693206", + "blk.22.attn_q.weight": "eee7e3d426ab533bd18d62c9aa142eedbde394bed07db58313e0fccc82a23237", + "blk.22.attn_v.weight": "26b5be1fe3c2b6824c5a648a3e4bdf17691904526fca158fbc3ebb627b67e2f4", + "blk.22.ffn_gate_exps.weight": "32ab7a7735313d60f6a75229b1aeee940b6aee176c9648536bf5921b0dc2929a", + "blk.22.ffn_down_exps.weight": "67590808f6a67777d3eb7976c31fe616d388b98fecbb12253b72d1241d70753f", + "blk.22.ffn_up_exps.weight": "fc245c0183e6d90829ff5e71a4ec93e4860b3d4c1a17b9dda2fb64f5f5c9ed32", + "blk.22.attn_norm.weight": "128e99d206d4d6724758ec97468af767fa0aea592149c324b731659c1e74a1a8", + "blk.22.ffn_norm.weight": "e45f498033f0cffa15da0eff2c47b4472e43fcf8921729fc4eeb2e3a6b3c78e2", + "blk.23.ffn_gate_inp.weight": "d63e686f5325fbc89fa242c2c52a3b8ff54f867dca914c9ae6eea13e9d6f46e5", + "blk.23.attn_k.weight": "f71f5a577f46ea12b1818f3a5ff4b85ddc45f9a2afb0fa2e041d71a3e31c6779", + "blk.23.attn_output.weight": "92b13563c1e0eac0d748fb67b235dfd7a64c8f16e2dafb316885744582e23b4b", + "blk.23.attn_q.weight": "2f9b9c35dc4f912f3f51c06e2d68f417b51a0de0a84aac530a64f9d3d7b0a2dd", + "blk.23.attn_v.weight": "268e40813806e74a5c364b19556d087bf8374e76e7b6fcf55c381eb7da13ccd1", + "blk.23.ffn_gate_exps.weight": "12f857e7a7ce228afac34d99b602c8d6fe96984f2a21118f459a58cb767ee65e", + "blk.23.ffn_down_exps.weight": "cdb082c16599c3bb36a28066dcc122d9529b54fa91b6cf0153437ec960a5e16d", + "blk.23.ffn_up_exps.weight": "f4b99f6f44d7b8b5a305894e88633bf5938fc1f6303a2b2092399da9c8b64d7c", + "blk.23.attn_norm.weight": "a691392210383915916b4d3886d5e4d56e7855e27e37e414fbd73bf66b3712e6", + "blk.23.ffn_norm.weight": "0c3dc72f667e5ae19b69bfa9f2bd2a01a57681f89ef9527bad4eb0d8c7b70da8", + "blk.24.ffn_gate_exps.weight": "86baca2a3157994df7fd8ced5e08436d5c1810dc29c0715637c36de723e0e7d1", + "blk.24.ffn_down_exps.weight": "ac5d559562b35c34993e34b071f66d15c65be5907797078c2d2a49aba54e3192", + "blk.24.ffn_up_exps.weight": "fce0a099cf09777f44fbab3606ceb75f7fae6f0b80725f9e871654b8cdf9262a", + "blk.24.ffn_gate_inp.weight": "e7c6800c0cfc56b565b2d35ad6f1dbfdb70dd0b05b338bc8da2286ffc3678d79", + "blk.24.attn_norm.weight": "dc6cc18ec52d102d015153c4a1132f9d7a504e29cbdec81c5edbf3b9e65815e1", + "blk.24.ffn_norm.weight": "480d5a1397af5e0e657f1e67d20ec0cdef5724e71246a326843321b87ffabd33", + "blk.24.attn_k.weight": "338c0597954a9b95a782545b2fe36469553e73f86ae2d2b5697767b28e1c7daa", + "blk.24.attn_output.weight": "a77d23b79933c67e52f1eef7f83a3dff4f767ce0bbcc39572f8cec4acd457643", + "blk.24.attn_q.weight": "45c9478593002be1998e96e70668aafa2dd3972380fbc1df12fb05c24ba959e0", + "blk.24.attn_v.weight": "515729420885408a6a9614bc27cda393ed907521318d14d21335d39a3eff0b61", + "blk.25.ffn_gate_inp.weight": "aae4ac40e9ab3925241f9d784b54b38851d9bc999a6c3bc03fc3f17c9b28a67c", + "blk.25.attn_k.weight": "4ab4808d02396c35b00b426f536015673b71c17ae6cd55bbc2e6bfe7a4c59d0c", + "blk.25.attn_output.weight": "1990bb982b77e0c947cd1a8ef0b36227ee1259e6dbbc2829e5c136edf88675eb", + "blk.25.attn_q.weight": "a1490f3048e8c0ec8784f8550c43adf5cc8d0f2f90131c934713fe4b1b015bd7", + "blk.25.attn_v.weight": "f15e53c6d45b3b6f58808fa968425d65e0b26b7f9b268127a77abb1227c67431", + "blk.25.ffn_gate_exps.weight": "656662447ff54f56ee80f78a1b9483f7efdc40f7375d0cd8a9c72ccf21f77e7b", + "blk.25.ffn_down_exps.weight": "db06f101bccbaef19cced0f6c185166e18202465f4a42cddfd535fbe5cbabb4a", + "blk.25.ffn_up_exps.weight": "584a7b02456f27fe1d8d3c7ccd21d426b6ea887795a3ed77f704596a1e3841d7", + "blk.25.attn_norm.weight": "8f0f3597982930fd237e9d609776c64f2b909a455b21678f83a7ebd4bbb83e64", + "blk.25.ffn_norm.weight": "3e7079c32582afba0c55e032f254adc18d2997705eec860185e9a6dd3d82f07e", + "blk.26.ffn_gate_exps.weight": "e70341691b583b86489812b29b77aa41eb658b1865733d6118da54c66e3bfcc6", + "blk.26.ffn_down_exps.weight": "5c1b812d11dfb064af816ced5ab6463bf9722eefdfc341b8a93705d5038fd781", + "blk.26.ffn_up_exps.weight": "e18118362ae54ef7432781c83884f9fb230a9d934e342aabeda8822ea5f71fb6", + "blk.26.ffn_gate_inp.weight": "cd1c5f6710166b9567c6b74c97b2348b191c60aa860958c6bc264ab095261dff", + "blk.26.attn_norm.weight": "71d087531af2520bda2e676c489e8529cef5db8aeea1eec0a937a8b4f2fa2e54", + "blk.26.ffn_norm.weight": "7f704e936fda28eb5c2cc339f0f6a5f78170b5aa43c01265b21668870d819c82", + "blk.26.attn_k.weight": "1cc62a0ce0ae251275d898c52c4a9fba5995fca10955d2011d10dd1a59e1afb8", + "blk.26.attn_output.weight": "636e881b1505f9cef656a4be98bec6a4765321d51f9bf1dac8933397cf44b765", + "blk.26.attn_q.weight": "89a3c4d202d7d6adebb9e0c1bcfd8b775f6456386f1be25e86e43acc949c1e16", + "blk.26.attn_v.weight": "ff2cc963b597cdf1a21703f3e7022af3bb4c65a34a19e19d9309a7c5e198b5bd", + "blk.27.ffn_gate_inp.weight": "6150139498fefe380bb99d11e72028da47a15ecb73dfc5b2774f726f4bed8f9e", + "blk.27.attn_k.weight": "f286eb9e5c56c7b801a497aedc40158c2a27877d7f9fb59b3fc67834798902d2", + "blk.27.attn_output.weight": "5dc3d3a05f9f7729509147fd09c16fb53f85f520cdab5cb69abf4bae3fd460c7", + "blk.27.attn_q.weight": "8462e40f86b24251960d6f35a9ea99b8793a01937faf1aec2859f2e5395dbb61", + "blk.27.attn_v.weight": "bac1a99e38e25953f8315f7212eb9777dc216cadb09b959977885ae62724ceca", + "blk.27.ffn_gate_exps.weight": "6a15eca7f0f6ecfd93db2e55c63875348ec4a78c4ff643ec46df9e958c0101e4", + "blk.27.ffn_down_exps.weight": "2e1c91247c4359e2073a8e5f26fd7f6426da7be3ed5bc65dcfff701f0a5022b2", + "blk.27.ffn_up_exps.weight": "65d6f5c553c9332085eae4aeadf25090b5d7768212ea7b08ed698102c21b29a1", + "blk.27.attn_norm.weight": "7fab8ae63ec8e91ce625cd130ab96d8427dad3a7413bb21b25ec5f408c5b9f5a", + "blk.27.ffn_norm.weight": "532720546b0fdcd423a02ca6e3e9d8aacb84b1b3e8269968f88a47fe2a69bab4", + "blk.28.ffn_gate_inp.weight": "a305ea58d98962d9dcf0c53ad2389b7acc8936fb35a0e3fc9410e7767cd49dea", + "blk.28.attn_k.weight": "8315e8a2e4f78dfdf36d4fc18fffc74bc95fe42c3ae4f9af2b6c874612c0f71b", + "blk.28.attn_output.weight": "9b5fdedd32d39ef46a22cca7cd5355d7b93bd07ea305f466a8aad6ca5a4f3778", + "blk.28.attn_q.weight": "4e8fb96997c30e231c437130f410d7c91d541a816f6c568b5f3bfdb4b8dece74", + "blk.28.attn_v.weight": "1fec739cf3bd7b4913f72ca358d4cf31391c304de44ac0ae31ecb825beaa7cfd", + "blk.28.ffn_gate_exps.weight": "9f259789d535e09268266b9a8020f32d6a6779966c909d91d3a10574f06238a2", + "blk.28.ffn_down_exps.weight": "516d3f8abaedb01b9916a4b67d4672159769138ef2850158bc1b32c41e31f0e8", + "blk.28.ffn_up_exps.weight": "f2f1d88d2c31ed588806fb5ad981d68f5134d7284c4fc022fd018de2eef437fc", + "blk.28.attn_norm.weight": "960fd005598deadaebd969996f4367a9dbfad90539a863674fe95730935acc64", + "blk.28.ffn_norm.weight": "e1993b37ced93d4049e9af2c47b0d9207d8f7e6f2cc3a52f57bef30bc806d805", + "blk.29.ffn_gate_exps.weight": "58927146338f443513337476b3cd30e6341742f096c2beb5890d400f10121298", + "blk.29.ffn_down_exps.weight": "03a3386e4f0b75a28c5608e23b2de8f0de25f21954e4aa7fc343431bde9db07e", + "blk.29.ffn_up_exps.weight": "6916b7490a7ae7b04a5d81cc1e7ac9b20c483434f3b186b12d87fe176bf1567b", + "blk.29.ffn_gate_inp.weight": "98e710e467a3d567abe4ce29d78b8e8dc033148762290c0c5e1ae4d78efd8c78", + "blk.29.attn_norm.weight": "4e64cb307d37be20d55f38c94faf7e451d11df5e60df347906cbaf9c5441be71", + "blk.29.ffn_norm.weight": "696c23a52f742679bd44440d687a4c44b4302d57f1e9dc5610d23374336187e7", + "blk.29.attn_k.weight": "e85253652fd6120c623634ba66b725bf7cd491318b54ccdad2c7df8851d64c0a", + "blk.29.attn_output.weight": "4f650a71efb150d1f24cd4d114d4187bf570ac424da3b92ea6455abdf1aea705", + "blk.29.attn_q.weight": "69fa7da901026ebcbbbc848455b425458b7e3295007d7fc093acf4b38e2166ea", + "blk.29.attn_v.weight": "17e2e7590b317b21f106de546aafd955579703d1e95d6aea044ee72ec3a514c9", + "blk.30.ffn_gate_inp.weight": "3a03284b4aa60d59d4a2ec86253469b61fc656372afca427cb77a5332fbcc62c", + "blk.30.attn_k.weight": "d518cfd0db9708e769eb1399e87ee49357dc54d5afdbac3d4c0ca46c64e789eb", + "blk.30.attn_output.weight": "9b44378714d784c5ef9ab604359091baca4e0ec222afa139b7f840eaefb371fd", + "blk.30.attn_q.weight": "cbb95365bbfbcad0c9cd99b4eebb5a5d32de68ce08e4063b5ec3e792b7548044", + "blk.30.attn_v.weight": "e7985c04fe1740e35a9598f43b67b0922b4fc2d00b68a92a9f917b82c3248de1", + "blk.30.ffn_gate_exps.weight": "8ac4bbd07935d98f895ba94dc174e5ad5046c3c222b53729d60f987c05e7eb70", + "blk.30.ffn_down_exps.weight": "dd672cc71e82abf05064a18121b8e55fe1a4f19bc1d7cb9a142f4add54bc336e", + "blk.30.ffn_up_exps.weight": "12282f664a2a12aa25e2deac58946108715ebb978bafed5274cef24569107646", + "blk.30.attn_norm.weight": "1a33458fee054c6c9c896a4bb0a4e1fbfa0293b2408c7dd2b81d692e966e7273", + "blk.30.ffn_norm.weight": "311e33b68051f507f1478ed8f2693fddb846170ddb7285a91be43f795c2ce31e", + "blk.31.ffn_gate_exps.weight": "8af43d9867a51cd8392fb48b981b0ceee0ae979c491c07d711b3b56b5162c786", + "blk.31.ffn_down_exps.weight": "5579cb7758c1600b19d1f540deffe081b575962e37437b3b2efb2fb0a2924e40", + "blk.31.ffn_up_exps.weight": "f2e7c005276b3a001fb40753f027fa10b4d5a346f43cf4b4bbdeec6e74e1cf6a", + "blk.31.ffn_gate_inp.weight": "89885dc0e30b6b16a90c0331d7fa3174671e941364e8102d934f02132237e61b", + "blk.31.attn_norm.weight": "99e4e9bf86a9edf8c404153a7e8a82324ba79da462622196e2faba161bd95172", + "blk.31.ffn_norm.weight": "55335997cf6de781bf332b943de96ff4646966b05d9fee86b76ea897e27b6ca7", + "blk.31.attn_k.weight": "cee570762b78da6316b637892cc4b080e40f57af5551ffb1866b9a8e80e96628", + "blk.31.attn_output.weight": "fa321ff55ec7819ead7b819fd45215262f39744569765ba2113c989c03588802", + "blk.31.attn_q.weight": "9e2c409b878f8a2a1436874abf428fceb1c534b21f9ad4dd6f532b8a469007f0", + "blk.31.attn_v.weight": "a845d0be68ba537b4a775bfba4d897faf7c82a811a2612b0b7420cc4f3574cb8", + "output.weight": "16101cbb74b54cda9ebc07ca3c762e3263a56efb3cc011156184b95807d7cf13", + "output_norm.weight": "d7aa61585baedd60157aafe157930785742c55989c288573566a971b02423564" +} diff --git a/convert/testdata/gemma-2b-it.json b/convert/testdata/gemma-2b-it.json new file mode 100644 index 0000000000000000000000000000000000000000..0482f1e145ecf759be746561587b7d3ad38b4c89 --- /dev/null +++ b/convert/testdata/gemma-2b-it.json @@ -0,0 +1,188 @@ +{ + "general.architecture": "gemma", + "general.file_type": "1", + "general.quantization_version": "2", + "gemma.block_count": "18", + "gemma.context_length": "8192", + "gemma.embedding_length": "2048", + "gemma.feed_forward_length": "16384", + "gemma.attention.head_count": "8", + "gemma.attention.head_count_kv": "1", + "gemma.attention.key_length": "256", + "gemma.attention.value_length": "256", + "gemma.attention.layer_norm_rms_epsilon": "1e-06", + "tokenizer.ggml.model": "llama", + "tokenizer.ggml.add_bos_token": "true", + "tokenizer.ggml.add_eos_token": "false", + "tokenizer.ggml.bos_token_id": "2", + "tokenizer.ggml.eos_token_id": "1", + "tokenizer.ggml.padding_token_id": "0", + "tokenizer.ggml.unknown_token_id": "3", + "tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8", + "tokenizer.ggml.token_type": "485e40bf3d715a4764818fc097d6a2a41db872d82ee714bc500872a3437ff48d", + "tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda", + "token_embd.weight": "17b87ab2c01c80657855a5413d0457b4a041afaeda0cc785080e44e2f04acf07", + "blk.0.attn_k.weight": "28ac0da05754ad2714ae95da28a5ad191192140b30b8fd22d108d4700c9d989f", + "blk.0.attn_norm.weight": "3f9d5675d1ab0eb8a816719dac9fab81f2e95c52be02c34263339acbc087febb", + "blk.0.attn_output.weight": "703295c2c63990ff896778685c678f145298886f680f3ed5dc2a7ad54c293265", + "blk.0.attn_q.weight": "69c2d0e4870e9d722a190d356203c9605575a16863466c3d1747966ef1cf5791", + "blk.0.attn_v.weight": "95219c9c07b5ffe9a9a01e456d845eef2b11f4fc12c93dbbba479db395444c13", + "blk.0.ffn_down.weight": "a2feb5eb3d572c57c5bafbf0ab506862df1160fe40965dcfe4b9fd855c08bed7", + "blk.0.ffn_gate.weight": "fcca072c445c31f4dc4d5dfaa785b1bdf7271342442099b74fd17268b5829fbf", + "blk.0.ffn_norm.weight": "7621f95dbd245cade6fffd6b08797d69d8e3954e960f0b5551b90d967ab95448", + "blk.0.ffn_up.weight": "14a9bcdd451403c67136391e1b6e53b3b1830f00199bd911dbcc56d8749c14f4", + "blk.1.attn_k.weight": "c70f73c5df20579cb44d971164b48b5f0d8d5abdb38b381e7a8b880ba12aa406", + "blk.1.attn_norm.weight": "88b6b91f93a1ef83425a7c7dc2a2fbd3b22704a04c64a80061df376ac8c33626", + "blk.1.attn_output.weight": "f031a537490c452be3b3bb51e6b7949a636405756e160976a1c070a792ea00ee", + "blk.1.attn_q.weight": "bdb23214b1cf9cfd30f863a0a5868e52c6809d93b7e8f44df096a94204d9896a", + "blk.1.attn_v.weight": "e9bbc0b05f2c872fb1403f8f938cd1612b502229ee401f12593b1164c61acc00", + "blk.1.ffn_down.weight": "5ff53811038b661a7b8f2bfdf213bebfb185ec1a6060b662f063714f33584d79", + "blk.1.ffn_gate.weight": "205085c8c951a5c7543b1495183cd96028fb49f67464b3e9862a2693a6077a33", + "blk.1.ffn_norm.weight": "798f354fc85afce9625f5d10093a585a966831698a0560e6c9b97ce659eb4b22", + "blk.1.ffn_up.weight": "db92dc5684cb6e90940e13f4d1da555ed20ba4f8cab1e990ddfd7553e2e91315", + "blk.2.attn_k.weight": "ef5ce360c4eed6d00d03ca4761e0f8e4b0af4509978468314be14f3d46621044", + "blk.2.attn_norm.weight": "6dadbc05dbd0d3fabb4216affa60a3de1378a82d2859dc90b338cbe70f50d455", + "blk.2.attn_output.weight": "6bbf87a966f691bbfd7c8d25629aa4e6710107bd431a667434861febb391edc5", + "blk.2.attn_q.weight": "4e575c09ae2de417ce9057ce8b073680e860a24aae13a472b68f101b760752e5", + "blk.2.attn_v.weight": "cd33f7f01141e9439afdaf2ea1aaced9feaa335e32a58daa136ebd555d4d96f4", + "blk.2.ffn_down.weight": "b970ff1b0b6494165defe2fbfa1d31425766ed71e64de9ec4e66ac3955c8bc5f", + "blk.2.ffn_gate.weight": "dbb3e1360402e0e369b101995bb686b73f95d4a7673f061be85d64d15dfb0061", + "blk.2.ffn_norm.weight": "bfb7980105d8ac9647710454f57a5cdac50598a0f6f4884e16f1d94b00844687", + "blk.2.ffn_up.weight": "50ef89339b275a438b664686f6227dd9b6e43853ed6856ec9e33ef4bbd90bda1", + "blk.3.attn_k.weight": "be942ea98151434eebcd2c1da4b00e0146152fe524a530689b1fd491cb833d21", + "blk.3.attn_norm.weight": "0df2f218daf609c289fb7c60c5f375fa99c0d4e04381ad5a494a19144edd8e20", + "blk.3.attn_output.weight": "c2184aaf86aa2cb8f47be49f60b165834e97205f39c6ee1dfd19fd4411a156ce", + "blk.3.attn_q.weight": "4f86e2a0a4221c1c84ff9c409ac89893cb95d7208cf65bf1e98e24e01125f991", + "blk.3.attn_v.weight": "abfdb8a60c349dadde641d1afc9542025e24fbf41a3238bfa9675e0b1f1e4b68", + "blk.3.ffn_down.weight": "58821a8d87008d47d122427911c6fad5272aca70c448bbae223256a74bacd07e", + "blk.3.ffn_gate.weight": "776e051f1a0ddd5c4934e69186683a75ca9a3c8c0f61911bba321fed1dd287d2", + "blk.3.ffn_norm.weight": "7f380f29335e28be90bfcfae6f6d69fdf5751211b36d2dd62aa5541ed113e4f2", + "blk.3.ffn_up.weight": "fc5ae8d488894cbd4951059675468d227da27871d26e925c9941863841c097ee", + "blk.4.attn_k.weight": "14833b078cc4c5137bdd5fdc0538047974ca147a99b0282e1b144440c78bc1db", + "blk.4.attn_norm.weight": "0a69957d4a15599fb80ad4753558020804925221457d9a5052926754d3768065", + "blk.4.attn_output.weight": "887a49b6130fb6297cf10767207c3dd97191b2cf63723449af9c27bca8dbeda0", + "blk.4.attn_q.weight": "51fd577b76764824dd6f0d4891c137ebe4736f591b5ca2793c5fff2be49abbde", + "blk.4.attn_v.weight": "1a623c43cf9c509d1b7ea0d1a5c04d0af4809665f9f9e93b7d6dba8c5df178fa", + "blk.4.ffn_down.weight": "5d61e8856d8941d2b1fd138116d015f63840d0fa1e31e20e20a5ceca1536ceec", + "blk.4.ffn_gate.weight": "06640f7273764f8ca5df7e386547417916b6cd7d565a8343153113239a94b0a1", + "blk.4.ffn_norm.weight": "91a6c6c41b894228e361435ecbc5058dca34d4911a23da5b56de219299c964d3", + "blk.4.ffn_up.weight": "d016dac1055e36d6a10b6317e57f98a904709ea892ef3194342f4d2f6326561e", + "blk.5.attn_k.weight": "987146afe124131500808cc0da33c06d207433656d41df6e6d8c99118a83bac5", + "blk.5.attn_norm.weight": "6b354938966f2608a2fb8d0f5b363ed0d8b0967c2ec8d0abd5c625b413042ded", + "blk.5.attn_output.weight": "cdcbfe02c6ff79d5326882b017a02099f5af71beedf6b1b3eb4de01e3a844536", + "blk.5.attn_q.weight": "b910d0cff781d3efb42eab0a302f46f286b2de717079175680d5b42bf8c309c8", + "blk.5.attn_v.weight": "66d3a279f747412f9f4b0e8abad44540c122ab2e811a7ee74c1f33bc36caade9", + "blk.5.ffn_down.weight": "c9b0efd2212981f16d956d8571f054b68780ad01f4917033647e359b557a4653", + "blk.5.ffn_gate.weight": "fe96b94109ca141c01f6a04788e20783019ca6ec334aa1f3134810bdb499e557", + "blk.5.ffn_norm.weight": "aa7b016e832e7055a36c6e20de58ea1936f995f390401fff1c5fc65906064e49", + "blk.5.ffn_up.weight": "555ce27c4873d3375394f38ad3b45e3d8848f9d5642dc1602383d0f0a33c2a14", + "blk.6.attn_k.weight": "88280d461db324c4f36475ce396793063e61a27283ec64511b0480890fb5b3b4", + "blk.6.attn_norm.weight": "af8f460c411f660d33196286d208f1845fd5a2b45f7b56549a4df31e7515447a", + "blk.6.attn_output.weight": "dd9996fb0a256e8375ad3917705258a33fce006bcea0f536caae420a77974d8b", + "blk.6.attn_q.weight": "7a4841541191e037cfb9b07930c4d8cab451809658b182f0ada6ccde9615c003", + "blk.6.attn_v.weight": "ae81e6a592b64d701a9d40233e986039a56cba8d8d24f61aea93c6393cf3078a", + "blk.6.ffn_down.weight": "622dd1ce1706355cbc659a8ab2c4509678ffe0f3ad34258e5e25ed2a5d951bcd", + "blk.6.ffn_gate.weight": "8389a735c0bd5591010f8ced9805a2a12c749f6df0d3c18ad4d05c2a302e7168", + "blk.6.ffn_norm.weight": "621f5346400382474d61358397bd58fb1459b07c53e376e4bca15e08b3f9b3fb", + "blk.6.ffn_up.weight": "8d834e4c42f13c251dfee36cf89e12f1bd400680d00d5c2e6cac0459e9ce2f7f", + "blk.7.attn_k.weight": "8bd0412de65a3e64901ef8fe6a28c95e116bf39dc9aa22f0126b9d36688e5ea7", + "blk.7.attn_norm.weight": "056d8e56be4e87d6dc6f900762f0dc6fde07bfdc50dd85bfc510415e2bba3f3d", + "blk.7.attn_output.weight": "27972eda51da53d416ff95aed78149a2c5a287b47d2cd46f2f544ca692ecb3bb", + "blk.7.attn_q.weight": "41eca977b9371f7932800c11a9c45b931310196919e2a0651b847703b180fc7f", + "blk.7.attn_v.weight": "13c74fd7e07f08883a09fb070a1fe5bbdd2341b4cb8d1cac07c4b637049b5774", + "blk.7.ffn_down.weight": "9e75db42468800849a9a7da603d0072c5e86c8ed2b4d8b20a312a51fb86a7a10", + "blk.7.ffn_gate.weight": "db6bdc3117f910088aaf7db51f2da63ea5bd933de36af5599c215bfb26f7db2b", + "blk.7.ffn_norm.weight": "48bb82b49bfc8679a1e77f282ee182d952db7a3c11be7ef9a102ee2ddd8011e2", + "blk.7.ffn_up.weight": "feebea87175817a0f3585ec0af09dc873d94c203581ae97a712eb356d3b49efe", + "blk.8.attn_k.weight": "d5640ad71b6af68d88e17bf8e7fc26c907d2262605457a84247dd9afc2884d69", + "blk.8.attn_norm.weight": "75b850c481a69083ae09d0207ba7317b37c735a39fcf5fef5400e6c84fb1257f", + "blk.8.attn_output.weight": "cbd669dbdea2bdd90f9f0cc97566b3dffff3c56cecb4f47290ceef30da83b2d6", + "blk.8.attn_q.weight": "9edcb63087a431bac361822497e6ecdaa06d9ea4a1a754e36da7ba9f8db81c7c", + "blk.8.attn_v.weight": "3fb72c2c4f95a83626aa3e30062f9450b09ab37c7871e229f18bbc5cf744633c", + "blk.8.ffn_down.weight": "bd69d2c9172974fff154441b237b4787fb53b2d185325442d5048130ef5bc4ef", + "blk.8.ffn_gate.weight": "d04689c80553edd011d1cbaa5d570fffa7fa91e88b66cf1352d89ab60b72f908", + "blk.8.ffn_norm.weight": "e49984183b735b7f2c4e4730c289eed9394056d2e283a00fd83ea0915df31a73", + "blk.8.ffn_up.weight": "8fe62a1ce8e847e567add6c6f6bf2922bc467495b5eb4c116b3cb85b85b3b211", + "blk.9.attn_k.weight": "d90904959e5004cf0d6e729c6bff18cc33c094798b802473c1ec55ab8d276183", + "blk.9.attn_norm.weight": "79277f290cc07411115d8fa138045edf4a17b3416ab2145409cbe8ab829fd4ee", + "blk.9.attn_output.weight": "5a21bf2e1f09a81405025f96d4153ffb630158e17269cff8ffff935c38ceb1a7", + "blk.9.attn_q.weight": "51b1d0febc3b350945be4504f55afa4347517bde0f710e1a4b88e6b17e71e7c7", + "blk.9.attn_v.weight": "aab7e1db0a8b50a03036356791ffce736ab010d15674c96eaef8049d80076054", + "blk.9.ffn_down.weight": "cbf43ec84becb40c9359a181ab0e641fd7faae7d34b549501f7cfb7afdc3d764", + "blk.9.ffn_gate.weight": "dce0e8661c778327bed7f03b6790d26710764188aed9dc746e6e05863891fa57", + "blk.9.ffn_norm.weight": "6d41642104f995c77bf31122b13237caebda3e7fcccb1367ce91db36b015e923", + "blk.9.ffn_up.weight": "82fe4c67bf24e7b2d6f6e05f7b1234c2bf90c3932951091a9066211b8e15ecbb", + "blk.10.attn_k.weight": "f6a9ed8fd8d3229b5d03175c413ffc56a07f2ce7236271986361dd3d8993f9aa", + "blk.10.attn_norm.weight": "cebbef89f0326ca8e02df3867a571e4d61c20c2a12f295f98ae590d62bc86010", + "blk.10.attn_output.weight": "34f5efb86accb4f06347d83a32558ea8eab3039d128969161a741ebacbb656ff", + "blk.10.attn_q.weight": "1e0efe27df2d5d50f7157253ba2cfd436d6781c3dc78ca176d0c16a210b5b763", + "blk.10.attn_v.weight": "8f085bf50a2b0f83cd6cdda3c8ef5a9e204a36348ed95871aac725d1f68640cf", + "blk.10.ffn_down.weight": "bf3b3cb4cace435809ac7b4cc933f20853af12f1f272d3dcefe7f19c0f203b8b", + "blk.10.ffn_gate.weight": "d3df7a1413b1c5adf1a1dcda9e5225a15c89874bae53bb6137ad1ea42fca2d34", + "blk.10.ffn_norm.weight": "a1da603b0480471b5ed8e862148cecd5fed918f8304d6933ab0bdb25b8d2fb8f", + "blk.10.ffn_up.weight": "bffbba605922e972dc47dda88a0b4659aa52236c76e5fe861a949e6d9a367492", + "blk.11.attn_k.weight": "9f31c63d66cd32c29b1eb8bb829d0c8525ce2ae936e0eefdaab6335a2d12a3df", + "blk.11.attn_norm.weight": "0bde1a266d8b2e8f202bb7e2e88b19147ca83021901f6d3cae77a4df5548c754", + "blk.11.attn_output.weight": "e10725c7cf746ed4a7e472cf7aea6cb564e5db6a1d5197adc980d650a387ccea", + "blk.11.attn_q.weight": "05ee758a7d065802630f8c65dca424364c1c8825e389aa33f9405c45e8a50cce", + "blk.11.attn_v.weight": "0c3ae7090f11775d24c51120db6e305db6aff706493e7ee123dcab74485ba789", + "blk.11.ffn_down.weight": "7ba40b8e12c09c5fb2006b77a771cb01ce894e88a3b3e1877f927a5b89c91709", + "blk.11.ffn_gate.weight": "db76388a023b98097972d354ba1c6a5e26efdeb1c596b9c28bf2cd8f6596975e", + "blk.11.ffn_norm.weight": "a38c3ae1b89a68ddc7b72c99c5b28be7fe3787c4fad9904d0c43d64eaf00c474", + "blk.11.ffn_up.weight": "13c8142f9cf1eddc658babf978daf3515c4ccc45f849f3e7e3930aa18a8480a0", + "blk.12.attn_k.weight": "f03241c36ac87cb57429a2ef22186b8d7d0b590a8b173beb01fa13d93772f3b1", + "blk.12.attn_norm.weight": "4568f654e6d65104d586e7c16ba960c83428698ce103022b7e0be15e2884e13b", + "blk.12.attn_output.weight": "04867603f82f91e41306e09b33ecda0104b3ee4834061f2c0bbdc8da33c72509", + "blk.12.attn_q.weight": "70fe04b9a8e08b6100cc8d6b58bf4cbbad15ca1de82d63baca5d352ba6c4cbae", + "blk.12.attn_v.weight": "15cb28db61a86c98687991d7e611bc92a1fcc6007f3432149cfb5fe518a4f65e", + "blk.12.ffn_down.weight": "6d10c790a4e3dc44c2dc36d96251ae97cdf30a4fa04d4c43e31bfbd038e6a7b7", + "blk.12.ffn_gate.weight": "3462a2d8f6b4743b25e24da51b90018ac2858d05ac7e582bcb69063cfdac1104", + "blk.12.ffn_norm.weight": "1f96392c1faa34e34ae5dea55a6a86c5aa4c79758952075d53d28de89dd88456", + "blk.12.ffn_up.weight": "d22eacc612a7411953d948483c5fb201e11722955ee0754da866e7bec578ac6d", + "blk.13.attn_k.weight": "5864977e6b733ea942647d6feed5c76156c48c200649c22e4e11b9e5860e57f3", + "blk.13.attn_norm.weight": "87e053535144723db4145aa5402acc54331b7696752d852bb9fc542ff33f0fb5", + "blk.13.attn_output.weight": "078145f5ad83f8b14f97a869346f7fd1583b24d1e3edadaa95d3da4242973f8f", + "blk.13.attn_q.weight": "3b8caf35504cbc4d1a7dd6e011a95760703b7f71e2218b030b1254f811362dd7", + "blk.13.attn_v.weight": "4fdf8365a603e043e5b40c4a21c84ac167f9be62794178f9d8a608dfe5653bf9", + "blk.13.ffn_down.weight": "a07d3abbfcacf48ba028df2cab895be32cc15022d23389a745286e79c1b1d1fd", + "blk.13.ffn_gate.weight": "1d2ab39666aa2909acc96787432a3ed13b19d25170f74665fadff9b17bbaffb1", + "blk.13.ffn_norm.weight": "4f2e809fda5f3eadf52578ee50e0ba36e53be91e55dce418c12dfe595f5f18e7", + "blk.13.ffn_up.weight": "8783d2720c2c37ca176a5801e0b3ef1f9cc9cf3ef1cd37af423aaf6b2a27e2bd", + "blk.14.attn_k.weight": "ce9428e2b55d43ae0c6690dbd56182f99adc427694ba8236b405cc8ea5035e86", + "blk.14.attn_norm.weight": "6abb35f9db8251d6ae954bda147c6ada2371b0574d11702e828f3c6ac99b7cc0", + "blk.14.attn_output.weight": "fe3880916d0ceb5bff672c88bbefb7060a545be609bf049beb2024b38221836d", + "blk.14.attn_q.weight": "7c8ad81be6f4a350931fd108b5f7c9e366e8c26ef62d1d85ffef5dca8fd893f8", + "blk.14.attn_v.weight": "e4bdedffacbebe38567a0734dfd67db90e911d9a9669fcde9a7c4ad8a0066c52", + "blk.14.ffn_down.weight": "ef6694dff1e05820aac0cd2b22f39ac7788b4967afc9250775575554c66aab2c", + "blk.14.ffn_gate.weight": "db63c4179e2db704bc505e2b4696e055b593e295a1b7c4c586fc793bdd5aab19", + "blk.14.ffn_norm.weight": "2796a62d832a9710148f95d533320492a33e712b2e5218659c548705bd11684d", + "blk.14.ffn_up.weight": "3f78c78d8c2d54df45f799d4ff902316628af296834afe4ceed63d4a324ff03e", + "blk.15.attn_k.weight": "6e810ee3859e07695645ee0c9a5efc7962668984a5f0a9325f47e462743b447c", + "blk.15.attn_norm.weight": "0956b576ae96db0b28cb09f761f801cfd9281432284664f0fe181c8d9c55d1ec", + "blk.15.attn_output.weight": "03a17f7e94208177aace5cc41b7f54670ba57873b7274ff6e23caf58cce110ca", + "blk.15.attn_q.weight": "b8edafe7d2216a6f8b4ae4905a906475490e6ea418f6e1d3cec563dbdc6fab91", + "blk.15.attn_v.weight": "f8ae8cae0f4cfa34a459824eba57350c3c248104ba5607e7d9dc7d7c39aaf4a6", + "blk.15.ffn_down.weight": "8d02eb439da852246d2ca67e9b7b6de0b090b80744355e64728a23e41926505b", + "blk.15.ffn_gate.weight": "ed5bf361c67db8731f186b775826f21c33bdb521111fd2d922539719a770239f", + "blk.15.ffn_norm.weight": "5942ca3c73209ac9a0c8bfd9b4aab7f7be7aee9aa12d9c35833493b44af76767", + "blk.15.ffn_up.weight": "f4bebf4ad99ec5f911327dec347be6c595814885309c7bc5647ce28c7f4d1cf5", + "blk.16.attn_k.weight": "756a534c19364448e0958b8948fe33891c6ccda0fbb4dfa2024e1f532a87804b", + "blk.16.attn_norm.weight": "386b7b9e4e6509f6af9c022d942b6c6c6cc136aeed8751ecb037c74d7c4bfb93", + "blk.16.attn_output.weight": "3ba1a766a25830b84d7c22178203635f9c5624caad290bc5e5d73da5d5e7a2ec", + "blk.16.attn_q.weight": "d39b0c91e1fda7685d50a0f7cc8d18c44b5bdc90a142c7fda0bc329cca1afa74", + "blk.16.attn_v.weight": "98b33fcb0ee3483cff1b06ecb44d7b7ffb4d34c268248e4d73dfdf82b2065b2f", + "blk.16.ffn_down.weight": "14006f5e4acb2f9416271ae562e299359cd2585739c7fc77ccbca54495563948", + "blk.16.ffn_gate.weight": "12f8abae2d301d8f88bedb6af98b1daecc7b0b8d05148594f931f30958d77aca", + "blk.16.ffn_norm.weight": "129a15a046ee96d06de288bd43c80f77a6b0fb3a159c7367154c6e4aaf362672", + "blk.16.ffn_up.weight": "b4a5911a45f3871ef1d4efb7dc7108645a564b70f818eccf45beebef2e844ee9", + "blk.17.attn_k.weight": "5e1bfcff0146ebdde3817b656952892eb671e14e75afc92fa53f84f8eecbec4c", + "blk.17.attn_norm.weight": "60bc988fab7c4b29ee9de599df41a8de00caa94fcd74677da011fac82f60f465", + "blk.17.attn_output.weight": "ba49b40d6a0b5685f749c24b0edbed3adc44dbe13b5d5e5fa1e56169fc746555", + "blk.17.attn_q.weight": "82bb415d24efcd14d03ace03f907bb70db6a204c76a0bdd1892e0fba165db87d", + "blk.17.attn_v.weight": "73dbe54beb91a899884e275ea81ffc5187a20cb7d5b68d5c299b783096999d94", + "blk.17.ffn_down.weight": "7c086166241e0664f8963fd1ca4ed74c737abfb2525ec20f8435821ff50158f3", + "blk.17.ffn_gate.weight": "51a32f78244d42a539f619c5ce661db9e6cf41636280a826d439b5444edcd28c", + "blk.17.ffn_norm.weight": "c4bb247fccd1ecc84875028af63dd20aaf5cbd17eb94a9bc36679c09285dccab", + "blk.17.ffn_up.weight": "b5886182790bc6fbadd63de9bc4ffee416f3b69a66280d197ab8c18edf769abf", + "output_norm.weight": "481f3097d0a20412e35b3a739b1b958487bcd41ff67744baa3c9acbddd2ee4d4" +} diff --git a/convert/tokenizer.go b/convert/tokenizer.go index efeb5491f0956b07b71b803b6e0bfade423a19a5..0d42a6d8ab7f65b071fbafa9cd985c63a6a46275 100644 --- a/convert/tokenizer.go +++ b/convert/tokenizer.go @@ -3,19 +3,150 @@ package convert import ( "cmp" "crypto/sha256" + "encoding/hex" "encoding/json" + "errors" "fmt" + "io/fs" "log/slog" "os" "slices" +) - "golang.org/x/exp/maps" +const ( + _ int32 = iota + tokenTypeNormal + tokenTypeUnknown + tokenTypeControl + tokenTypeUserDefined + tokenTypeUnused + tokenTypeByte ) type Tokenizer struct { - Version string `json:"version"` - AddedTokens []Token `json:"added_tokens"` - Model TokenizerModel `json:"model"` + *Vocabulary + SpecialVocabulary []*SpecialVocabulary + Merges []string + + Pre string + Template string +} + +func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error) { + v, err := parseVocabulary(fsys) + if err != nil { + return nil, err + } + + t := &Tokenizer{ + Vocabulary: v, + Pre: "default", + } + + addedTokens := make(map[string]token) + if f, err := fsys.Open("tokenizer.json"); errors.Is(err, os.ErrNotExist) { + } else if err != nil { + return nil, err + } else { + defer f.Close() + + var tt tokenizer + if err := json.NewDecoder(f).Decode(&tt); err != nil { + return nil, err + } + + for _, t := range tt.AddedTokens { + addedTokens[t.Content] = t + } + + t.Merges = tt.Model.Merges + + sha256sum := sha256.New() + for _, pt := range tt.PreTokenizer.PreTokenizers { + switch pt.Type { + case "Split": + if pt.Pattern.Regex != "" { + // create a checksum of all Split pretokenizers which should be sufficient + // to identify the pretokenizer + sha256sum.Write([]byte(pt.Pattern.Regex)) + } + } + } + + switch digest := hex.EncodeToString(sha256sum.Sum(nil)); digest { + case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f": + t.Pre = "llama-bpe" + case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02": + t.Pre = "deepseek-llm" + case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e": + t.Pre = "deepseek-coder" + case "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855": + // noop, empty pretokenizer + default: + slog.Warn("unknown pretokenizer, using default", "digest", digest) + } + } + + if f, err := fsys.Open("tokenizer_config.json"); errors.Is(err, os.ErrNotExist) { + } else if err != nil { + return nil, err + } else { + defer f.Close() + + var p map[string]json.RawMessage + if err := json.NewDecoder(f).Decode(&p); err != nil { + return nil, err + } + + if template, ok := p["chat_template"]; ok { + if err := json.Unmarshal(template, &t.Template); err != nil { + return nil, err + } + } + + for _, st := range specialTokenTypes { + sv := SpecialVocabulary{Type: st} + if bts, ok := p[fmt.Sprintf("add_%s_token", st)]; ok { + if err := json.Unmarshal(bts, &sv.AddToken); err != nil { + return nil, err + } + } + + if bts, ok := p[fmt.Sprintf("%s_token", st)]; ok { + var content string + if err := json.Unmarshal(bts, &content); err != nil { + var mm map[string]any + if err := json.Unmarshal(bts, &mm); err != nil { + continue + } + + content, ok = mm["content"].(string) + if !ok { + continue + } + } + + sv.Content = content + } + + if id, ok := addedTokens[sv.Content]; ok { + sv.ID = id.ID + t.SpecialVocabulary = append(t.SpecialVocabulary, &sv) + } + } + } + + return t, nil +} + +type tokenizer struct { + Version string `json:"version"` + AddedTokens []token `json:"added_tokens"` + Model struct { + Type string `json:"type"` + Vocab map[string]int `json:"vocab"` + Merges []string `json:"merges"` + } `json:"model"` PreTokenizer struct { PreTokenizers []struct { @@ -27,83 +158,108 @@ type Tokenizer struct { } `json:"pre_tokenizer"` } -type TokenizerModel struct { - Type string `json:"type"` - Vocab map[string]int `json:"vocab"` - Merges []string `json:"merges"` - Tokens []Token -} - -type Token struct { +type token struct { ID int `json:"id"` Content string `json:"content"` Special bool `json:"special"` UserDefined bool } -func (t *Token) Type() int32 { - switch { - case t.Special: - return tokenTypeControl - case t.UserDefined: - return tokenTypeUserDefined - default: - return tokenTypeNormal - } +type Vocabulary struct { + Model string + Tokens []string + Scores []float32 + Types []int32 } -func (t *Tokenizer) maxID() int { - return max( - slices.Max(maps.Values(t.Model.Vocab)), - slices.MaxFunc(t.AddedTokens, func(a, b Token) int { - return cmp.Compare(a.ID, b.ID) - }).ID, - ) -} - -func parseTokens(dirpath string) (pre string, tokens []Token, merges []string, err error) { - f, err := os.Open(dirpath) +func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) { + f, err := fsys.Open("tokenizer.json") if err != nil { - panic(err) + return nil, err } defer f.Close() - var t Tokenizer + var t tokenizer if err := json.NewDecoder(f).Decode(&t); err != nil { - return "", nil, nil, err + return nil, err } - tokens = make([]Token, t.maxID()+1) + var tokens []token for k, v := range t.Model.Vocab { - tokens[v] = Token{ID: v, Content: k, Special: false, UserDefined: false} + tokens = append(tokens, token{ + ID: v, + Content: k, + }) } - for _, v := range t.AddedTokens { - v.UserDefined = true - tokens[v.ID] = v + for _, t := range t.AddedTokens { + t.UserDefined = true + tokens = append(tokens, t) } - sha256sum := sha256.New() - for _, pt := range t.PreTokenizer.PreTokenizers { - switch pt.Type { - case "Split": - if pt.Pattern.Regex != "" { - sha256sum.Write([]byte(pt.Pattern.Regex)) - } + slices.SortFunc(tokens, func(i, j token) int { + return cmp.Compare(i.ID, j.ID) + }) + + v := Vocabulary{Model: "gpt2"} + for _, t := range tokens { + v.Tokens = append(v.Tokens, t.Content) + v.Scores = append(v.Scores, float32(t.ID)) + + switch { + case t.Special: + v.Types = append(v.Types, tokenTypeControl) + case t.UserDefined: + v.Types = append(v.Types, tokenTypeUserDefined) + default: + v.Types = append(v.Types, tokenTypeNormal) + } + } + + return &v, nil +} + +func parseVocabulary(fsys fs.FS) (*Vocabulary, error) { + patterns := []struct { + Pattern string + Func func(fs.FS) (*Vocabulary, error) + }{ + {"tokenizer.model", parseSentencePiece}, + {"tokenizer.json", parseVocabularyFromTokenizer}, + } + + for _, pattern := range patterns { + if _, err := fs.Stat(fsys, pattern.Pattern); errors.Is(err, os.ErrNotExist) { + continue + } else if err != nil { + return nil, err } + + return pattern.Func(fsys) } - switch digest := fmt.Sprintf("%x", sha256sum.Sum(nil)); digest { - case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f": - pre = "llama-bpe" - case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02": - pre = "deepseek-llm" - case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e": - pre = "deepseek-coder" - default: - slog.Warn("unknown pretokenizer, using default", "digest", digest) - pre = "default" + return nil, errors.New("unknown tensor format") +} + +type SpecialVocabulary struct { + Type string + ID int + Content string + AddToken bool +} + +func (sv SpecialVocabulary) Key() string { + switch t := sv.Type; t { + case "bos", "eos", "cls", "mask": + return t + case "unk": + return "unknown" + case "sep": + //nolint:misspell // this is an upstream typo + return "seperator" + case "pad": + return "padding" } - return pre, tokens, t.Model.Merges, nil + panic("unknown special vocabulary type") } diff --git a/convert/tokenizer_spm.go b/convert/tokenizer_spm.go new file mode 100644 index 0000000000000000000000000000000000000000..babf702c77ce55f9a4aaa63a313392cb6c394ab8 --- /dev/null +++ b/convert/tokenizer_spm.go @@ -0,0 +1,83 @@ +package convert + +import ( + "cmp" + "encoding/json" + "errors" + "fmt" + "io/fs" + "os" + "slices" + + "google.golang.org/protobuf/proto" + + "github.com/ollama/ollama/convert/sentencepiece" +) + +func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) { + bts, err := fs.ReadFile(fsys, "tokenizer.model") + if err != nil { + return nil, err + } + + var spm sentencepiece.ModelProto + if err := proto.Unmarshal(bts, &spm); err != nil { + return nil, err + } + + v := Vocabulary{Model: "llama"} + for _, piece := range spm.GetPieces() { + v.Tokens = append(v.Tokens, piece.GetPiece()) + v.Scores = append(v.Scores, piece.GetScore()) + + switch t := piece.GetType(); t { + case sentencepiece.ModelProto_SentencePiece_UNKNOWN, + sentencepiece.ModelProto_SentencePiece_CONTROL, + sentencepiece.ModelProto_SentencePiece_UNUSED, + sentencepiece.ModelProto_SentencePiece_BYTE: + v.Types = append(v.Types, int32(t)) + default: + v.Types = append(v.Types, int32(sentencepiece.ModelProto_SentencePiece_NORMAL)) + } + } + + f, err := fsys.Open("added_tokens.json") + if errors.Is(err, os.ErrNotExist) { + return &v, nil + } else if err != nil { + return nil, err + } + defer f.Close() + + var atm map[string]int + if err := json.NewDecoder(f).Decode(&atm); err != nil { + return nil, err + } + + type t struct { + id int + content string + } + + var ts []t + for content, id := range atm { + ts = append(ts, t{id, content}) + } + + slices.SortFunc(ts, func(i, j t) int { + return cmp.Compare(i.id, j.id) + }) + + n := len(v.Tokens) + for i, t := range ts { + if t.id != i+n { + return nil, fmt.Errorf("invalid token id: %d", t.id) + } + + v.Tokens = append(v.Tokens, t.content) + v.Scores = append(v.Scores, -1000.0) + v.Types = append(v.Types, tokenTypeUserDefined) + } + + return &v, nil +} diff --git a/docs/api.md b/docs/api.md index 0f11c388cc71fc73d5968bd2897a8e917cbd9e6d..aed2b69fbde8c32d4a20ab844c215432f7e82e3b 100644 --- a/docs/api.md +++ b/docs/api.md @@ -12,6 +12,7 @@ - [Pull a Model](#pull-a-model) - [Push a Model](#push-a-model) - [Generate Embeddings](#generate-embeddings) +- [List Running Models](#list-running-models) ## Conventions @@ -25,7 +26,7 @@ All durations are returned in nanoseconds. ### Streaming responses -Certain endpoints stream responses as JSON objects and can optional return non-streamed responses. +Certain endpoints stream responses as JSON objects. Streaming can be disabled by providing `{"stream": false}` for these endpoints. ## Generate a completion @@ -39,6 +40,7 @@ Generate a response for a given prompt with a provided model. This is a streamin - `model`: (required) the [model name](#model-names) - `prompt`: the prompt to generate a response for +- `suffix`: the text after the model response - `images`: (optional) a list of base64-encoded images (for multimodal models such as `llava`) Advanced parameters (optional): @@ -56,7 +58,8 @@ Advanced parameters (optional): Enable JSON mode by setting the `format` parameter to `json`. This will structure the response as a valid JSON object. See the JSON mode [example](#request-json-mode) below. -> Note: it's important to instruct the model to use JSON in the `prompt`. Otherwise, the model may generate large amounts whitespace. +> [!IMPORTANT] +> It's important to instruct the model to use JSON in the `prompt`. Otherwise, the model may generate large amounts whitespace. ### Examples @@ -147,8 +150,44 @@ If `stream` is set to `false`, the response will be a single JSON object: } ``` +#### Request (with suffix) + +##### Request + +```shell +curl http://localhost:11434/api/generate -d '{ + "model": "codellama:code", + "prompt": "def compute_gcd(a, b):", + "suffix": " return result", + "options": { + "temperature": 0 + }, + "stream": false +}' +``` + +##### Response + +```json +{ + "model": "codellama:code", + "created_at": "2024-07-22T20:47:51.147561Z", + "response": "\n if a == 0:\n return b\n else:\n return compute_gcd(b % a, a)\n\ndef compute_lcm(a, b):\n result = (a * b) / compute_gcd(a, b)\n", + "done": true, + "done_reason": "stop", + "context": [...], + "total_duration": 1162761250, + "load_duration": 6683708, + "prompt_eval_count": 17, + "prompt_eval_duration": 201222000, + "eval_count": 63, + "eval_duration": 953997000 +} +``` + #### Request (JSON mode) +> [!IMPORTANT] > When `format` is set to `json`, the output will always be a well-formed JSON object. It's important to also instruct the model to respond in JSON. ##### Request @@ -249,7 +288,7 @@ curl http://localhost:11434/api/generate -d '{ #### Request (Reproducible outputs) -For reproducible outputs, set `temperature` to 0 and `seed` to a number: +For reproducible outputs, set `seed` to a number: ##### Request @@ -258,8 +297,7 @@ curl http://localhost:11434/api/generate -d '{ "model": "mistral", "prompt": "Why is the sky blue?", "options": { - "seed": 123, - "temperature": 0 + "seed": 123 } }' ``` @@ -298,6 +336,7 @@ curl http://localhost:11434/api/generate -d '{ "num_predict": 100, "top_k": 20, "top_p": 0.9, + "min_p": 0.0, "tfs_z": 0.5, "typical_p": 0.7, "repeat_last_n": 33, @@ -380,12 +419,14 @@ Generate the next message in a chat with a provided model. This is a streaming e - `model`: (required) the [model name](#model-names) - `messages`: the messages of the chat, this can be used to keep a chat memory +- `tools`: tools for the model to use if supported. Requires `stream` to be set to `false` The `message` object has the following fields: -- `role`: the role of the message, either `system`, `user` or `assistant` +- `role`: the role of the message, either `system`, `user`, `assistant`, or `tool` - `content`: the content of the message - `images` (optional): a list of images to include in the message (for multimodal models such as `llava`) +- `tool_calls` (optional): a list of tools the model wants to use Advanced parameters (optional): @@ -546,7 +587,7 @@ Final response: ##### Request -Send a chat message with a conversation history. +Send a chat message with images. The images should be provided as an array, with the individual images encoded in Base64. ```shell curl http://localhost:11434/api/chat -d '{ @@ -622,6 +663,79 @@ curl http://localhost:11434/api/chat -d '{ } ``` +#### Chat request (with tools) + +##### Request + +``` +curl http://localhost:11434/api/chat -d '{ + "model": "llama3.1", + "messages": [ + { + "role": "user", + "content": "What is the weather today in Paris?" + } + ], + "stream": false, + "tools": [ + { + "type": "function", + "function": { + "name": "get_current_weather", + "description": "Get the current weather for a location", + "parameters": { + "type": "object", + "properties": { + "location": { + "type": "string", + "description": "The location to get the weather for, e.g. San Francisco, CA" + }, + "format": { + "type": "string", + "description": "The format to return the weather in, e.g. 'celsius' or 'fahrenheit'", + "enum": ["celsius", "fahrenheit"] + } + }, + "required": ["location", "format"] + } + } + } + ] +}' +``` + +##### Response + +```json +{ + "model": "llama3.1", + "created_at": "2024-07-22T20:33:28.123648Z", + "message": { + "role": "assistant", + "content": "", + "tool_calls": [ + { + "function": { + "name": "get_current_weather", + "arguments": { + "format": "celsius", + "location": "Paris, FR" + } + } + } + ] + }, + "done_reason": "stop", + "done": true, + "total_duration": 885095291, + "load_duration": 3753500, + "prompt_eval_count": 122, + "prompt_eval_duration": 328493000, + "eval_count": 33, + "eval_duration": 552222000 +} +``` + ## Create a Model ```shell @@ -777,11 +891,12 @@ A single JSON object will be returned. POST /api/show ``` -Show information about a model including details, modelfile, template, parameters, license, and system prompt. +Show information about a model including details, modelfile, template, parameters, license, system prompt. ### Parameters - `name`: name of the model to show +- `verbose`: (optional) if set to `true`, returns full data for verbose response fields ### Examples @@ -798,14 +913,40 @@ curl http://localhost:11434/api/show -d '{ ```json { "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"", - "parameters": "num_ctx 4096\nstop \u003c/s\u003e\nstop USER:\nstop ASSISTANT:", - "template": "{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: ", + "parameters": "num_keep 24\nstop \"<|start_header_id|>\"\nstop \"<|end_header_id|>\"\nstop \"<|eot_id|>\"", + "template": "{{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>", "details": { + "parent_model": "", "format": "gguf", "family": "llama", - "families": ["llama", "clip"], - "parameter_size": "7B", + "families": [ + "llama" + ], + "parameter_size": "8.0B", "quantization_level": "Q4_0" + }, + "model_info": { + "general.architecture": "llama", + "general.file_type": 2, + "general.parameter_count": 8030261248, + "general.quantization_version": 2, + "llama.attention.head_count": 32, + "llama.attention.head_count_kv": 8, + "llama.attention.layer_norm_rms_epsilon": 0.00001, + "llama.block_count": 32, + "llama.context_length": 8192, + "llama.embedding_length": 4096, + "llama.feed_forward_length": 14336, + "llama.rope.dimension_count": 128, + "llama.rope.freq_base": 500000, + "llama.vocab_size": 128256, + "tokenizer.ggml.bos_token_id": 128000, + "tokenizer.ggml.eos_token_id": 128009, + "tokenizer.ggml.merges": [], // populates if `verbose=true` + "tokenizer.ggml.model": "gpt2", + "tokenizer.ggml.pre": "llama-bpe", + "tokenizer.ggml.token_type": [], // populates if `verbose=true` + "tokenizer.ggml.tokens": [] // populates if `verbose=true` } } ``` @@ -998,6 +1139,121 @@ If `stream` is set to `false`, then the response is a single JSON object: ## Generate Embeddings +```shell +POST /api/embed +``` + +Generate embeddings from a model + +### Parameters + +- `model`: name of model to generate embeddings from +- `input`: text or list of text to generate embeddings for + +Advanced parameters: + +- `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true` +- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature` +- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`) + +### Examples + +#### Request + +```shell +curl http://localhost:11434/api/embed -d '{ + "model": "all-minilm", + "input": "Why is the sky blue?" +}' +``` + +#### Response + +```json +{ + "model": "all-minilm", + "embeddings": [[ + 0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814, + 0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348 + ]], + "total_duration": 14143917, + "load_duration": 1019500, + "prompt_eval_count": 8 +} +``` + +#### Request (Multiple input) + +```shell +curl http://localhost:11434/api/embed -d '{ + "model": "all-minilm", + "input": ["Why is the sky blue?", "Why is the grass green?"] +}' +``` + +#### Response + +```json +{ + "model": "all-minilm", + "embeddings": [[ + 0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814, + 0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348 + ],[ + -0.0098027075, 0.06042469, 0.025257962, -0.006364387, 0.07272725, + 0.017194884, 0.09032035, -0.051705178, 0.09951512, 0.09072481 + ]] +} +``` + +## List Running Models +```shell +GET /api/ps +``` + +List models that are currently loaded into memory. + +#### Examples + +### Request + +```shell +curl http://localhost:11434/api/ps +``` + +#### Response + +A single JSON object will be returned. + +```json +{ + "models": [ + { + "name": "mistral:latest", + "model": "mistral:latest", + "size": 5137025024, + "digest": "2ae6f6dd7a3dd734790bbbf58b8909a606e0e7e97e94b7604e0aa7ae4490e6d8", + "details": { + "parent_model": "", + "format": "gguf", + "family": "llama", + "families": [ + "llama" + ], + "parameter_size": "7.2B", + "quantization_level": "Q4_0" + }, + "expires_at": "2024-06-04T14:38:31.83753-07:00", + "size_vram": 5137025024 + } + ] +} +``` + +## Generate Embedding + +> Note: this endpoint has been superseded by `/api/embed` + ```shell POST /api/embeddings ``` diff --git a/docs/development.md b/docs/development.md index 8c035a518483618064b71e36543c90a6f2de72e9..cd6c41af5404cee12a0b5e7a6a4a8ae02ed2836a 100644 --- a/docs/development.md +++ b/docs/development.md @@ -104,7 +104,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H, you might use: ``` -OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./... +OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./... go build . ``` @@ -114,15 +114,18 @@ If you have Docker available, you can build linux binaries with `./scripts/build ### Windows -Note: The windows build for Ollama is still under development. +Note: The Windows build for Ollama is still under development. -Install required tools: +First, install required tools: - MSVC toolchain - C/C++ and cmake as minimal requirements - Go version 1.22 or higher - MinGW (pick one variant) with GCC. - [MinGW-w64](https://www.mingw-w64.org/) - [MSYS2](https://www.msys2.org/) +- The `ThreadJob` Powershell module: `Install-Module -Name ThreadJob -Scope CurrentUser` + +Then, build the `ollama` binary: ```powershell $env:CGO_ENABLED="1" diff --git a/docs/docker.md b/docs/docker.md index 22cd62dd3a88558ed1452c25f1f681826da614c8..314666b26cfd810aedd18ec90046aea86924befe 100644 --- a/docs/docker.md +++ b/docs/docker.md @@ -26,19 +26,19 @@ sudo apt-get install -y nvidia-container-toolkit #### Install with Yum or Dnf 1. Configure the repository - + ```bash curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \ | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo ``` - + 2. Install the NVIDIA Container Toolkit packages - + ```bash sudo yum install -y nvidia-container-toolkit ``` -#### Configure Docker to use Nvidia driver +#### Configure Docker to use Nvidia driver ``` sudo nvidia-ctk runtime configure --runtime=docker sudo systemctl restart docker @@ -63,7 +63,7 @@ docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 114 Now you can run a model: ``` -docker exec -it ollama ollama run llama3 +docker exec -it ollama ollama run llama3.1 ``` ### Try different models diff --git a/docs/faq.md b/docs/faq.md index b50a3138c013a08b2892278d961f7afbed40849c..324116d1a2d9ef6db9c9541034c69a44fb76dc10 100644 --- a/docs/faq.md +++ b/docs/faq.md @@ -227,7 +227,7 @@ curl http://localhost:11434/api/chat -d '{"model": "mistral"}' To preload a model using the CLI, use the command: ```shell -ollama run llama3 "" +ollama run llama3.1 "" ``` ## How do I keep a model loaded in memory or make it unload immediately? @@ -257,3 +257,23 @@ If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` AP ## How do I manage the maximum number of requests the Ollama server can queue? If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded. You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`. + +## How does Ollama handle concurrent requests? + +Ollama supports two levels of concurrent processing. If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time. For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing. + +If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded. As prior models become idle, one or more will be unloaded to make room for the new model. Queued requests will be processed in order. When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads. + +Parallel request processing for a given model results in increasing the context size by the number of parallel requests. For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation. + +The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms: + +- `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory. The default is 3 * the number of GPUs or 3 for CPU inference. +- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time. The default will auto-select either 4 or 1 based on available memory. +- `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512 + +Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting. Once ROCm v6.2 is available, Windows Radeon will follow the defaults above. You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM. + +## How does Ollama load models on multiple GPUs? + +Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models. When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available. If the model will entirely fit on any single GPU, Ollama will load the model on that GPU. This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference. If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs. \ No newline at end of file diff --git a/docs/gpu.md b/docs/gpu.md index a6b559f04f24d45d1ff7d96f42e7598df19b22fc..e669ea32c21915bfbacfd8b89b2089cdefb7efd1 100644 --- a/docs/gpu.md +++ b/docs/gpu.md @@ -8,7 +8,7 @@ Check your compute compatibility to see if your card is supported: | Compute Capability | Family | Cards | | ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- | | 9.0 | NVIDIA | `H100` | -| 8.9 | GeForce RTX 40xx | `RTX 4090` `RTX 4080` `RTX 4070 Ti` `RTX 4060 Ti` | +| 8.9 | GeForce RTX 40xx | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060` | | | NVIDIA Professional | `L4` `L40` `RTX 6000` | | 8.6 | GeForce RTX 30xx | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` | | | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2` | @@ -18,7 +18,7 @@ Check your compute compatibility to see if your card is supported: | | Quadro | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000` | | 7.0 | NVIDIA | `TITAN V` `V100` `Quadro GV100` | | 6.1 | NVIDIA TITAN | `TITAN Xp` `TITAN X` | -| | GeForce GTX | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050` | +| | GeForce GTX | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050 Ti` `GTX 1050` | | | Quadro | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` | | | Tesla | `P40` `P4` | | 6.0 | NVIDIA | `Tesla P100` `Quadro GP100` | @@ -46,13 +46,24 @@ sudo modprobe nvidia_uvm` ## AMD Radeon Ollama supports the following AMD GPUs: + +### Linux Support | Family | Cards and accelerators | | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- | | AMD Radeon RX | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56` | | AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` | | AMD Instinct | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50` | -### Overrides +### Windows Support +With ROCm v6.1, the following GPUs are supported on Windows. + +| Family | Cards and accelerators | +| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- | +| AMD Radeon RX | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` | +| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` | + + +### Overrides on Linux Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In some cases you can force the system to try to use a similar LLVM target that is close. For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4) @@ -63,7 +74,7 @@ would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the server. If you have an unsupported AMD GPU you can experiment using the list of supported types below. -At this time, the known supported GPU types are the following LLVM Targets. +At this time, the known supported GPU types on linux are the following LLVM Targets. This table shows some example GPUs that map to these LLVM targets: | **LLVM Target** | **An Example GPU** | |-----------------|---------------------| diff --git a/docs/import.md b/docs/import.md index 7041b74d5922684149d18da136a91bfb5a9f43fa..f34f09aceea4d871d9e3e0528f42da9f251bf85c 100644 --- a/docs/import.md +++ b/docs/import.md @@ -1,170 +1,88 @@ -# Import a model +# Import -This guide walks through importing a GGUF, PyTorch or Safetensors model. +GGUF models and select Safetensors models can be imported directly into Ollama. -## Importing (GGUF) +## Import GGUF -### Step 1: Write a `Modelfile` +A binary GGUF file can be imported directly into Ollama through a Modelfile. -Start by creating a `Modelfile`. This file is the blueprint for your model, specifying weights, parameters, prompt templates and more. - -``` -FROM ./mistral-7b-v0.1.Q4_0.gguf +```dockerfile +FROM /path/to/file.gguf ``` -(Optional) many chat models require a prompt template in order to answer correctly. A default prompt template can be specified with the `TEMPLATE` instruction in the `Modelfile`: - -``` -FROM ./mistral-7b-v0.1.Q4_0.gguf -TEMPLATE "[INST] {{ .Prompt }} [/INST]" -``` - -### Step 2: Create the Ollama model - -Finally, create a model from your `Modelfile`: +## Import Safetensors -``` -ollama create example -f Modelfile -``` +If the model being imported is one of these architectures, it can be imported directly into Ollama through a Modelfile: -### Step 3: Run your model + - LlamaForCausalLM + - MistralForCausalLM + - GemmaForCausalLM -Next, test the model with `ollama run`: - -``` -ollama run example "What is your favourite condiment?" +```dockerfile +FROM /path/to/safetensors/directory ``` -## Importing (PyTorch & Safetensors) +For architectures not directly convertable by Ollama, see llama.cpp's [guide](https://github.com/ggerganov/llama.cpp/blob/master/README.md#prepare-and-quantize) on conversion. After conversion, see [Import GGUF](#import-gguf). -> Importing from PyTorch and Safetensors is a longer process than importing from GGUF. Improvements that make it easier are a work in progress. +## Automatic Quantization -### Setup +> [!NOTE] +> Automatic quantization requires v0.1.35 or higher. -First, clone the `ollama/ollama` repo: +Ollama is capable of quantizing FP16 or FP32 models to any of the supported quantizations with the `-q/--quantize` flag in `ollama create`. +```dockerfile +FROM /path/to/my/gemma/f16/model ``` -git clone git@github.com:ollama/ollama.git ollama -cd ollama -``` - -and then fetch its `llama.cpp` submodule: ```shell -git submodule init -git submodule update llm/llama.cpp +$ ollama create -q Q4_K_M mymodel +transferring model data +quantizing F16 model to Q4_K_M +creating new layer sha256:735e246cc1abfd06e9cdcf95504d6789a6cd1ad7577108a70d9902fef503c1bd +creating new layer sha256:0853f0ad24e5865173bbf9ffcc7b0f5d56b66fd690ab1009867e45e7d2c4db0f +writing manifest +success ``` -Next, install the Python dependencies: +### Supported Quantizations -``` -python3 -m venv llm/llama.cpp/.venv -source llm/llama.cpp/.venv/bin/activate -pip install -r llm/llama.cpp/requirements.txt -``` - -Then build the `quantize` tool: +- `Q4_0` +- `Q4_1` +- `Q5_0` +- `Q5_1` +- `Q8_0` -``` -make -C llm/llama.cpp quantize -``` +#### K-means Quantizations -### Clone the HuggingFace repository (optional) +- `Q3_K_S` +- `Q3_K_M` +- `Q3_K_L` +- `Q4_K_S` +- `Q4_K_M` +- `Q5_K_S` +- `Q5_K_M` +- `Q6_K` -If the model is currently hosted in a HuggingFace repository, first clone that repository to download the raw model. +## Template Detection -Install [Git LFS](https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage), verify it's installed, and then clone the model's repository: - -``` -git lfs install -git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1 model -``` +> [!NOTE] +> Template detection requires v0.1.42 or higher. -### Convert the model +Ollama uses model metadata, specifically `tokenizer.chat_template`, to automatically create a template appropriate for the model you're importing. -> Note: some model architectures require using specific convert scripts. For example, Qwen models require running `convert-hf-to-gguf.py` instead of `convert.py` - -``` -python llm/llama.cpp/convert.py ./model --outtype f16 --outfile converted.bin -``` - -### Quantize the model - -``` -llm/llama.cpp/quantize converted.bin quantized.bin q4_0 +```dockerfile +FROM /path/to/my/gemma/model ``` -### Step 3: Write a `Modelfile` - -Next, create a `Modelfile` for your model: - -``` -FROM quantized.bin -TEMPLATE "[INST] {{ .Prompt }} [/INST]" -``` - -### Step 4: Create the Ollama model - -Finally, create a model from your `Modelfile`: - -``` -ollama create example -f Modelfile -``` - -### Step 5: Run your model - -Next, test the model with `ollama run`: - -``` -ollama run example "What is your favourite condiment?" -``` - -## Publishing your model (optional – early alpha) - -Publishing models is in early alpha. If you'd like to publish your model to share with others, follow these steps: - -1. Create [an account](https://ollama.com/signup) -2. Copy your Ollama public key: - - macOS: `cat ~/.ollama/id_ed25519.pub | pbcopy` - - Windows: `type %USERPROFILE%\.ollama\id_ed25519.pub` - - Linux: `cat /usr/share/ollama/.ollama/id_ed25519.pub` -3. Add your public key to your [Ollama account](https://ollama.com/settings/keys) - -Next, copy your model to your username's namespace: - -``` -ollama cp example /example -``` - -> Note: model names may only contain lowercase letters, digits, and the characters `.`, `-`, and `_`. - -Then push the model: - -``` -ollama push /example +```shell +$ ollama create mymodel +transferring model data +using autodetected template gemma-instruct +creating new layer sha256:baa2a0edc27d19cc6b7537578a9a7ba1a4e3214dc185ed5ae43692b319af7b84 +creating new layer sha256:ba66c3309914dbef07e5149a648fd1877f030d337a4f240d444ea335008943cb +writing manifest +success ``` -After publishing, your model will be available at `https://ollama.com//example`. - -## Quantization reference - -The quantization options are as follow (from highest highest to lowest levels of quantization). Note: some architectures such as Falcon do not support K quants. - -- `q2_K` -- `q3_K` -- `q3_K_S` -- `q3_K_M` -- `q3_K_L` -- `q4_0` (recommended) -- `q4_1` -- `q4_K` -- `q4_K_S` -- `q4_K_M` -- `q5_0` -- `q5_1` -- `q5_K` -- `q5_K_S` -- `q5_K_M` -- `q6_K` -- `q8_0` -- `f16` +Defining a template in the Modelfile will disable this feature which may be useful if you want to use a different template than the autodetected one. diff --git a/docs/linux.md b/docs/linux.md index 9e7e06fabe41410485b9fc2fa1f614bc907c11af..ec7306560d8a7d44e3f576db99acaee7cd3d57ab 100644 --- a/docs/linux.md +++ b/docs/linux.md @@ -100,6 +100,16 @@ sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama sudo chmod +x /usr/bin/ollama ``` +## Installing specific versions + +Use `OLLAMA_VERSION` environment variable with the install script to install a specific version of Ollama, including pre-releases. You can find the version numbers in the [releases page](https://github.com/ollama/ollama/releases). + +For example: + +``` +curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.1.32 sh +``` + ## Viewing logs To view logs of Ollama running as a startup service, run: diff --git a/docs/modelfile.md b/docs/modelfile.md index 21ee1826e3d49eef85eae9dacde202dcf6d62b9a..852bf96c11f45d8d74dc28fa46903f65301edd0b 100644 --- a/docs/modelfile.md +++ b/docs/modelfile.md @@ -1,6 +1,7 @@ # Ollama Model File -> Note: `Modelfile` syntax is in development +> [!NOTE] +> `Modelfile` syntax is in development A model file is the blueprint to create and share models with Ollama. @@ -140,6 +141,7 @@ PARAMETER | num_predict | Maximum number of tokens to predict when generating text. (Default: 128, -1 = infinite generation, -2 = fill context) | int | num_predict 42 | | top_k | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40) | int | top_k 40 | | top_p | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9) | float | top_p 0.9 | +| min_p | Alternative to the top_p, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out. (Default: 0.0) | float | min_p 0.05 | ### TEMPLATE diff --git a/docs/openai.md b/docs/openai.md index 557b5846d9e8c9454df17e7b4ad322a5ada54552..7b3a3f31597e5480bf53171504128caf55bd24ec 100644 --- a/docs/openai.md +++ b/docs/openai.md @@ -27,6 +27,37 @@ chat_completion = client.chat.completions.create( ], model='llama3', ) + +response = client.chat.completions.create( + model="llava", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "What's in this image?"}, + { + "type": "image_url", + "image_url": "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", + }, + ], + } + ], + max_tokens=300, +) + +completion = client.completions.create( + model="llama3", + prompt="Say this is a test", +) + +list_completion = client.models.list() + +model = client.models.retrieve("llama3") + +embeddings = client.embeddings.create( + model="all-minilm", + input=["why is the sky blue?", "why is the grass green?"], +) ``` ### OpenAI JavaScript library @@ -42,14 +73,44 @@ const openai = new OpenAI({ }) const chatCompletion = await openai.chat.completions.create({ - messages: [{ role: 'user', content: 'Say this is a test' }], - model: 'llama3', + messages: [{ role: 'user', content: 'Say this is a test' }], + model: 'llama3', +}) + +const response = await openai.chat.completions.create({ + model: "llava", + messages: [ + { + role: "user", + content: [ + { type: "text", text: "What's in this image?" }, + { + type: "image_url", + image_url: "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", + }, + ], + }, + ], +}) + +const completion = await openai.completions.create({ + model: "llama3", + prompt: "Say this is a test.", +}) + +const listCompletion = await openai.models.list() + +const model = await openai.models.retrieve("llama3") + +const embedding = await openai.embeddings.create({ + model: "all-minilm", + input: ["why is the sky blue?", "why is the grass green?"], }) ``` ### `curl` -``` +``` shell curl http://localhost:11434/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ @@ -65,6 +126,48 @@ curl http://localhost:11434/v1/chat/completions \ } ] }' + +curl http://localhost:11434/v1/chat/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "llava", + "messages": [ + { + "role": "user", + "content": [ + { + "type": "text", + "text": "What'\''s in this image?" + }, + { + "type": "image_url", + "image_url": { + "url": "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" + } + } + ] + } + ], + "max_tokens": 300 + }' + +curl http://localhost:11434/v1/completions \ + -H "Content-Type: application/json" \ + -d '{ + "model": "llama3", + "prompt": "Say this is a test" + }' + +curl http://localhost:11434/v1/models + +curl http://localhost:11434/v1/models/llama3 + +curl http://localhost:11434/v1/embeddings \ + -H "Content-Type: application/json" \ + -d '{ + "model": "all-minilm", + "input": ["why is the sky blue?", "why is the grass green?"] + }' ``` ## Endpoints @@ -77,8 +180,9 @@ curl http://localhost:11434/v1/chat/completions \ - [x] Streaming - [x] JSON mode - [x] Reproducible outputs +- [x] Vision +- [x] Tools (streaming support coming soon) - [ ] Vision -- [ ] Function calling - [ ] Logprobs #### Supported request fields @@ -86,7 +190,10 @@ curl http://localhost:11434/v1/chat/completions \ - [x] `model` - [x] `messages` - [x] Text `content` - - [ ] Array of `content` parts + - [x] Image `content` + - [x] Base64 encoded image + - [ ] Image URL + - [x] Array of `content` parts - [x] `frequency_penalty` - [x] `presence_penalty` - [x] `response_format` @@ -96,17 +203,72 @@ curl http://localhost:11434/v1/chat/completions \ - [x] `temperature` - [x] `top_p` - [x] `max_tokens` -- [ ] `logit_bias` -- [ ] `tools` +- [x] `tools` - [ ] `tool_choice` +- [ ] `logit_bias` +- [ ] `user` +- [ ] `n` + +### `/v1/completions` + +#### Supported features + +- [x] Completions +- [x] Streaming +- [x] JSON mode +- [x] Reproducible outputs +- [ ] Logprobs + +#### Supported request fields + +- [x] `model` +- [x] `prompt` +- [x] `frequency_penalty` +- [x] `presence_penalty` +- [x] `seed` +- [x] `stop` +- [x] `stream` +- [x] `temperature` +- [x] `top_p` +- [x] `max_tokens` +- [x] `suffix` +- [ ] `best_of` +- [ ] `echo` +- [ ] `logit_bias` - [ ] `user` - [ ] `n` #### Notes -- Setting `seed` will always set `temperature` to `0` -- `finish_reason` will always be `stop` -- `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached +- `prompt` currently only accepts a string + +### `/v1/models` + +#### Notes + +- `created` corresponds to when the model was last modified +- `owned_by` corresponds to the ollama username, defaulting to `"library"` + +### `/v1/models/{model}` + +#### Notes + +- `created` corresponds to when the model was last modified +- `owned_by` corresponds to the ollama username, defaulting to `"library"` + +### `/v1/embeddings` + +#### Supported request fields + +- [x] `model` +- [x] `input` + - [x] string + - [x] array of strings + - [ ] array of tokens + - [ ] array of token arrays +- [ ] `encoding format` +- [ ] `dimensions` +- [ ] `user` ## Models diff --git a/docs/template.md b/docs/template.md new file mode 100644 index 0000000000000000000000000000000000000000..f6ce06ba461b57c496904ac9e79b3f030bd9d9be --- /dev/null +++ b/docs/template.md @@ -0,0 +1,173 @@ +# Template + +Ollama provides a powerful templating engine backed by Go's built-in templating engine to construct prompts for your large language model. This feature is a valuable tool to get the most out of your models. + +## Basic Template Structure + +A basic Go template consists of three main parts: + +* **Layout**: The overall structure of the template. +* **Variables**: Placeholders for dynamic data that will be replaced with actual values when the template is rendered. +* **Functions**: Custom functions or logic that can be used to manipulate the template's content. + +Here's an example of a simple chat template: + +```gotmpl +{{- range .Messages }} +{{ .Role }}: {{ .Content }} +{{- end }} +``` + +In this example, we have: + +* A basic messages structure (layout) +* Three variables: `Messages`, `Role`, and `Content` (variables) +* A custom function (action) that iterates over an array of items (`range .Messages`) and displays each item + +## Adding templates to your model + +By default, models imported into Ollama have a default template of `{{ .Prompt }}`, i.e. user inputs are sent verbatim to the LLM. This is appropriate for text or code completion models but lacks essential markers for chat or instruction models. + +Omitting a template in these models puts the responsibility of correctly templating input onto the user. Adding a template allows users to easily get the best results from the model. + +To add templates in your model, you'll need to add a `TEMPLATE` command to the Modelfile. Here's an example using Meta's Llama 3. + +```dockerfile +FROM llama3 + +TEMPLATE """{{- if .System }}<|start_header_id|>system<|end_header_id|> + +{{ .System }}<|eot_id|> +{{- end }} +{{- range .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|> + +{{ .Content }}<|eot_id|> +{{- end }}<|start_header_id|>assistant<|end_header_id|> + +""" +``` + +## Variables + +`System` (string): system prompt + +`Prompt` (string): user prompt + +`Response` (string): assistant response + +`Suffix` (string): text inserted after the assistant's response + +`Messages` (list): list of messages + +`Messages[].Role` (string): role which can be one of `system`, `user`, `assistant`, or `tool` + +`Messages[].Content` (string): message content + +`Messages[].ToolCalls` (list): list of tools the model wants to call + +`Messages[].ToolCalls[].Function` (object): function to call + +`Messages[].ToolCalls[].Function.Name` (string): function name + +`Messages[].ToolCalls[].Function.Arguments` (map): mapping of argument name to argument value + +`Tools` (list): list of tools the model can access + +`Tools[].Type` (string): schema type. `type` is always `function` + +`Tools[].Function` (object): function definition + +`Tools[].Function.Name` (string): function name + +`Tools[].Function.Description` (string): function description + +`Tools[].Function.Parameters` (object): function parameters + +`Tools[].Function.Parameters.Type` (string): schema type. `type` is always `object` + +`Tools[].Function.Parameters.Required` (list): list of required properties + +`Tools[].Function.Parameters.Properties` (map): mapping of property name to property definition + +`Tools[].Function.Parameters.Properties[].Type` (string): property type + +`Tools[].Function.Parameters.Properties[].Description` (string): property description + +`Tools[].Function.Parameters.Properties[].Enum` (list): list of valid values + +## Tips and Best Practices + +Keep the following tips and best practices in mind when working with Go templates: + +* **Be mindful of dot**: Control flow structures like `range` and `with` changes the value `.` +* **Out-of-scope variables**: Use `$.` to reference variables not currently in scope, starting from the root +* **Whitespace control**: Use `-` to trim leading (`{{-`) and trailing (`-}}`) whitespace + +## Examples + +### Example Messages + +#### ChatML + +ChatML is a popular template format. It can be used for models such as Databrick's DBRX, Intel's Neural Chat, and Microsoft's Orca 2. + +```gotmpl +{{- if .System }}<|im_start|>system +{{ .System }}<|im_end|> +{{ end }} +{{- range .Messages }}<|im_start|>{{ .Role }} +{{ .Content }}<|im_end|> +{{ end }}<|im_start|>assistant +{{ else }} +{{ if .System }}<|im_start|>system +{{ .System }}<|im_end|> +``` + +### Example Tools + +Tools support can be added to a model by adding a `{{ .Tools }}` node to the template. This feature is useful for models trained to call external tools and can a powerful tool for retrieving real-time data or performing complex tasks. + +#### Mistral + +Mistral v0.3 and Mixtral 8x22B supports tool calling. + +```gotmpl +{{- range $index, $_ := .Messages }} +{{- if eq .Role "user" }} +{{- if and (le (len (slice $.Messages $index)) 2) $.Tools }}[AVAILABLE_TOOLS] {{ json $.Tools }}[/AVAILABLE_TOOLS] +{{- end }}[INST] {{ if and (eq (len (slice $.Messages $index)) 1) $.System }}{{ $.System }} + +{{ end }}{{ .Content }}[/INST] +{{- else if eq .Role "assistant" }} +{{- if .Content }} {{ .Content }} +{{- else if .ToolCalls }}[TOOL_CALLS] [ +{{- range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ json .Function.Arguments }}} +{{- end }}] +{{- end }} +{{- else if eq .Role "tool" }}[TOOL_RESULTS] {"content": {{ .Content }}}[/TOOL_RESULTS] +{{- end }} +{{- end }} +``` + +### Example Fill-in-Middle + +Fill-in-middle support can be added to a model by adding a `{{ .Suffix }}` node to the template. This feature is useful for models that are trained to generate text in the middle of user input, such as code completion models. + +#### CodeLlama + +CodeLlama [7B](https://ollama.com/library/codellama:7b-code) and [13B](https://ollama.com/library/codellama:13b-code) code completion models support fill-in-middle. + +```gotmpl +
 {{ .Prompt }} {{ .Suffix }} 
+```
+
+> [!NOTE]
+> CodeLlama 34B and 70B code completion and all instruct and Python fine-tuned models do not support fill-in-middle.
+
+#### Codestral
+
+Codestral [22B](https://ollama.com/library/codestral:22b) supports fill-in-middle.
+
+```gotmpl
+[SUFFIX]{{ .Suffix }}[PREFIX] {{ .Prompt }}
+```
diff --git a/docs/troubleshooting.md b/docs/troubleshooting.md
index 60d63c7d91aaa9e083b8a9e2938bffabe2cd7c14..589061a89a8d0137be0861fcabaedebb0d9a2e1a 100644
--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:
 
 ```shell
-journalctl -u ollama
+journalctl -u ollama --no-pager
 ```
 
 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
@@ -22,7 +22,7 @@ docker logs 
 If manually running `ollama serve` in a terminal, the logs will be on that terminal.
 
 When you run Ollama on **Windows**, there are a few different locations. You can view them in the explorer window by hitting `+R` and type in:
-- `explorer %LOCALAPPDATA%\Ollama` to view logs
+- `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
 - `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
@@ -70,14 +70,18 @@ curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 
 If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
 
-## Container fails to run on NVIDIA GPU
+## NVIDIA GPU Discovery
 
-Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+When Ollama starts up, it takes inventory of the GPUs present in the system to determine compatibility and how much VRAM is available.  Sometimes this discovery can fail to find your GPUs.  In general, running the latest driver will yield the best results.
 
-Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+### Linux NVIDIA Troubleshooting
 
-- Is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
-- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
+If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+
+Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+
+- If you are using a container, is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
+- Is the uvm driver loaded? `sudo nvidia-modprobe -u`
 - Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
 - Try rebooting
 - Make sure you're running the latest nvidia drivers
@@ -85,3 +89,8 @@ Sometimes the container runtime can have difficulties initializing the GPU. When
 If none of those resolve the problem, gather additional information and file an issue:
 - Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`
+
+
+## Windows Terminal Errors
+
+Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.
diff --git a/docs/tutorials/langchainjs.md b/docs/tutorials/langchainjs.md
index 4d60afb643695407f6beff2063f36906c0ac317e..f925869b5f840099aaa722e85fa83e8422c6255a 100644
--- a/docs/tutorials/langchainjs.md
+++ b/docs/tutorials/langchainjs.md
@@ -15,7 +15,7 @@ import { Ollama } from "@langchain/community/llms/ollama";
 
 const ollama = new Ollama({
   baseUrl: "http://localhost:11434",
-  model: "llama3",
+  model: "llama3.1",
 });
 
 const answer = await ollama.invoke(`why is the sky blue?`);
@@ -23,7 +23,7 @@ const answer = await ollama.invoke(`why is the sky blue?`);
 console.log(answer);
 ```
 
-That will get us the same thing as if we ran `ollama run llama3 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
+That will get us the same thing as if we ran `ollama run llama3.1 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
 
 ```bash
 npm install cheerio
diff --git a/docs/tutorials/langchainpy.md b/docs/tutorials/langchainpy.md
index 9a1bca0ddc779a463890a4f6524b2db7d2061a1f..06543a070189cc36cb2f737fc8d6667a95e88292 100644
--- a/docs/tutorials/langchainpy.md
+++ b/docs/tutorials/langchainpy.md
@@ -45,7 +45,7 @@ all_splits = text_splitter.split_documents(data)
 ```
 
 It's split up, but we have to find the relevant splits and then submit those to the model. We can do this by creating embeddings and storing them in a vector database. We can use Ollama directly to instantiate an embedding model. We will use ChromaDB in this example for a vector database. `pip install chromadb`
-
+We also need to pull embedding model: `ollama pull nomic-embed-text`
 ```python
 from langchain.embeddings import OllamaEmbeddings
 from langchain.vectorstores import Chroma
@@ -68,7 +68,8 @@ The next thing is to send the question and the relevant parts of the docs to the
 ```python
 from langchain.chains import RetrievalQA
 qachain=RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
-qachain.invoke({"query": question})
+res = qachain.invoke({"query": question})
+print(res['result'])
 ```
 
 The answer received from this chain was:
diff --git a/docs/windows.md b/docs/windows.md
index 832b3d4315cfb6cf9465d16a7876c51a300c8efc..dbfc14407acb88ea26ba332418dd265edb15a63f 100644
--- a/docs/windows.md
+++ b/docs/windows.md
@@ -19,10 +19,12 @@ Logs will often be helpful in diagnosing the problem (see
 
 ## System Requirements
 
-* Windows 10 or newer, Home or Pro
+* Windows 10 22H2 or newer, Home or Pro
 * NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
 * AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
 
+Ollama uses unicode characters for progress indication, which may render as unknown squares in some older terminal fonts in Windows 10. If you see this, try changing your terminal font settings.
+
 ## API Access
 
 Here's a quick example showing API access from `powershell`
@@ -39,8 +41,8 @@ server.
 Ollama on Windows stores files in a few different locations.  You can view them in
 the explorer window by hitting `+R` and type in:
 - `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
-    - *app.log* contains logs from the GUI application
-    - *server.log* contains the server logs
+    - *app.log* contains most resent logs from the GUI application
+    - *server.log* contains the most recent server logs
     - *upgrade.log* contains log output for upgrades
 - `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
diff --git a/envconfig/config.go b/envconfig/config.go
index d66994513d7c3111ee6f2f6f34174eb088f97634..b82b773d4adb2e27c9dc896b223f932d53ea3d0a 100644
--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -3,210 +3,282 @@ package envconfig
 import (
 	"fmt"
 	"log/slog"
+	"math"
+	"net"
+	"net/url"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
+	"time"
 )
 
-var (
-	// Set via OLLAMA_ORIGINS in the environment
-	AllowOrigins []string
-	// Set via OLLAMA_DEBUG in the environment
-	Debug bool
-	// Experimental flash attention
-	FlashAttention bool
-	// Set via OLLAMA_KEEP_ALIVE in the environment
-	KeepAlive string
-	// Set via OLLAMA_LLM_LIBRARY in the environment
-	LLMLibrary string
-	// Set via OLLAMA_MAX_LOADED_MODELS in the environment
-	MaxRunners int
-	// Set via OLLAMA_MAX_QUEUE in the environment
-	MaxQueuedRequests int
-	// Set via OLLAMA_MAX_VRAM in the environment
-	MaxVRAM uint64
-	// Set via OLLAMA_NOHISTORY in the environment
-	NoHistory bool
-	// Set via OLLAMA_NOPRUNE in the environment
-	NoPrune bool
-	// Set via OLLAMA_NUM_PARALLEL in the environment
-	NumParallel int
-	// Set via OLLAMA_RUNNERS_DIR in the environment
-	RunnersDir string
-	// Set via OLLAMA_TMPDIR in the environment
-	TmpDir string
-)
+// Host returns the scheme and host. Host can be configured via the OLLAMA_HOST environment variable.
+// Default is scheme "http" and host "127.0.0.1:11434"
+func Host() *url.URL {
+	defaultPort := "11434"
 
-type EnvVar struct {
-	Name        string
-	Value       any
-	Description string
-}
+	s := strings.TrimSpace(Var("OLLAMA_HOST"))
+	scheme, hostport, ok := strings.Cut(s, "://")
+	switch {
+	case !ok:
+		scheme, hostport = "http", s
+	case scheme == "http":
+		defaultPort = "80"
+	case scheme == "https":
+		defaultPort = "443"
+	}
 
-func AsMap() map[string]EnvVar {
-	return map[string]EnvVar{
-		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug, "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
-		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention, "Enabled flash attention"},
-		"OLLAMA_HOST":              {"OLLAMA_HOST", "", "IP Address for the ollama server (default 127.0.0.1:11434)"},
-		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"},
-		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"},
-		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models (default 1)"},
-		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
-		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"},
-		"OLLAMA_MODELS":            {"OLLAMA_MODELS", "", "The path to the models directory"},
-		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"},
-		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"},
-		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default 1)"},
-		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"},
-		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"},
-		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir, "Location for temporary files"},
+	// trim trailing slashes
+	hostport = strings.TrimRight(hostport, "/")
+
+	host, port, err := net.SplitHostPort(hostport)
+	if err != nil {
+		host, port = "127.0.0.1", defaultPort
+		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
+			host = ip.String()
+		} else if hostport != "" {
+			host = hostport
+		}
 	}
-}
 
-func Values() map[string]string {
-	vals := make(map[string]string)
-	for k, v := range AsMap() {
-		vals[k] = fmt.Sprintf("%v", v.Value)
+	if n, err := strconv.ParseInt(port, 10, 32); err != nil || n > 65535 || n < 0 {
+		slog.Warn("invalid port, using default", "port", port, "default", defaultPort)
+		return &url.URL{
+			Scheme: scheme,
+			Host:   net.JoinHostPort(host, defaultPort),
+		}
 	}
-	return vals
-}
 
-var defaultAllowOrigins = []string{
-	"localhost",
-	"127.0.0.1",
-	"0.0.0.0",
+	return &url.URL{
+		Scheme: scheme,
+		Host:   net.JoinHostPort(host, port),
+	}
 }
 
-// Clean quotes and spaces from the value
-func clean(key string) string {
-	return strings.Trim(os.Getenv(key), "\"' ")
-}
+// Origins returns a list of allowed origins. Origins can be configured via the OLLAMA_ORIGINS environment variable.
+func Origins() (origins []string) {
+	if s := Var("OLLAMA_ORIGINS"); s != "" {
+		origins = strings.Split(s, ",")
+	}
+
+	for _, origin := range []string{"localhost", "127.0.0.1", "0.0.0.0"} {
+		origins = append(origins,
+			fmt.Sprintf("http://%s", origin),
+			fmt.Sprintf("https://%s", origin),
+			fmt.Sprintf("http://%s", net.JoinHostPort(origin, "*")),
+			fmt.Sprintf("https://%s", net.JoinHostPort(origin, "*")),
+		)
+	}
 
-func init() {
-	// default values
-	NumParallel = 1
-	MaxRunners = 1
-	MaxQueuedRequests = 512
+	origins = append(origins,
+		"app://*",
+		"file://*",
+		"tauri://*",
+	)
 
-	LoadConfig()
+	return origins
 }
 
-func LoadConfig() {
-	if debug := clean("OLLAMA_DEBUG"); debug != "" {
-		d, err := strconv.ParseBool(debug)
-		if err == nil {
-			Debug = d
-		} else {
-			Debug = true
-		}
+// Models returns the path to the models directory. Models directory can be configured via the OLLAMA_MODELS environment variable.
+// Default is $HOME/.ollama/models
+func Models() string {
+	if s := Var("OLLAMA_MODELS"); s != "" {
+		return s
 	}
 
-	if fa := clean("OLLAMA_FLASH_ATTENTION"); fa != "" {
-		d, err := strconv.ParseBool(fa)
-		if err == nil {
-			FlashAttention = d
-		}
+	home, err := os.UserHomeDir()
+	if err != nil {
+		panic(err)
 	}
 
-	RunnersDir = clean("OLLAMA_RUNNERS_DIR")
-	if runtime.GOOS == "windows" && RunnersDir == "" {
-		// On Windows we do not carry the payloads inside the main executable
-		appExe, err := os.Executable()
-		if err != nil {
-			slog.Error("failed to lookup executable path", "error", err)
-		}
+	return filepath.Join(home, ".ollama", "models")
+}
 
-		cwd, err := os.Getwd()
-		if err != nil {
-			slog.Error("failed to lookup working directory", "error", err)
+// KeepAlive returns the duration that models stay loaded in memory. KeepAlive can be configured via the OLLAMA_KEEP_ALIVE environment variable.
+// Negative values are treated as infinite. Zero is treated as no keep alive.
+// Default is 5 minutes.
+func KeepAlive() (keepAlive time.Duration) {
+	keepAlive = 5 * time.Minute
+	if s := Var("OLLAMA_KEEP_ALIVE"); s != "" {
+		if d, err := time.ParseDuration(s); err == nil {
+			keepAlive = d
+		} else if n, err := strconv.ParseInt(s, 10, 64); err == nil {
+			keepAlive = time.Duration(n) * time.Second
 		}
+	}
 
-		var paths []string
-		for _, root := range []string{filepath.Dir(appExe), cwd} {
-			paths = append(paths,
-				filepath.Join(root),
-				filepath.Join(root, "windows-"+runtime.GOARCH),
-				filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
-			)
-		}
+	if keepAlive < 0 {
+		return time.Duration(math.MaxInt64)
+	}
+
+	return keepAlive
+}
 
-		// Try a few variations to improve developer experience when building from source in the local tree
-		for _, p := range paths {
-			candidate := filepath.Join(p, "ollama_runners")
-			_, err := os.Stat(candidate)
-			if err == nil {
-				RunnersDir = candidate
-				break
+func Bool(k string) func() bool {
+	return func() bool {
+		if s := Var(k); s != "" {
+			b, err := strconv.ParseBool(s)
+			if err != nil {
+				return true
 			}
+
+			return b
 		}
-		if RunnersDir == "" {
-			slog.Error("unable to locate llm runner directory.  Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
-		}
+
+		return false
 	}
+}
 
-	TmpDir = clean("OLLAMA_TMPDIR")
+var (
+	// Debug enabled additional debug information.
+	Debug = Bool("OLLAMA_DEBUG")
+	// FlashAttention enables the experimental flash attention feature.
+	FlashAttention = Bool("OLLAMA_FLASH_ATTENTION")
+	// NoHistory disables readline history.
+	NoHistory = Bool("OLLAMA_NOHISTORY")
+	// NoPrune disables pruning of model blobs on startup.
+	NoPrune = Bool("OLLAMA_NOPRUNE")
+	// SchedSpread allows scheduling models across all GPUs.
+	SchedSpread = Bool("OLLAMA_SCHED_SPREAD")
+	// IntelGPU enables experimental Intel GPU detection.
+	IntelGPU = Bool("OLLAMA_INTEL_GPU")
+)
 
-	userLimit := clean("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseUint(userLimit, 10, 64)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_VRAM", userLimit, "error", err)
-		} else {
-			MaxVRAM = avail
-		}
+func String(s string) func() string {
+	return func() string {
+		return Var(s)
 	}
+}
+
+var (
+	LLMLibrary = String("OLLAMA_LLM_LIBRARY")
+	TmpDir     = String("OLLAMA_TMPDIR")
 
-	LLMLibrary = clean("OLLAMA_LLM_LIBRARY")
+	CudaVisibleDevices    = String("CUDA_VISIBLE_DEVICES")
+	HipVisibleDevices     = String("HIP_VISIBLE_DEVICES")
+	RocrVisibleDevices    = String("ROCR_VISIBLE_DEVICES")
+	GpuDeviceOrdinal      = String("GPU_DEVICE_ORDINAL")
+	HsaOverrideGfxVersion = String("HSA_OVERRIDE_GFX_VERSION")
+)
 
-	if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {
-		val, err := strconv.Atoi(onp)
-		if err != nil || val <= 0 {
-			slog.Error("invalid setting must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err)
-		} else {
-			NumParallel = val
-		}
+func RunnersDir() (p string) {
+	if p := Var("OLLAMA_RUNNERS_DIR"); p != "" {
+		return p
 	}
 
-	if nohistory := clean("OLLAMA_NOHISTORY"); nohistory != "" {
-		NoHistory = true
+	if runtime.GOOS != "windows" {
+		return
 	}
 
-	if noprune := clean("OLLAMA_NOPRUNE"); noprune != "" {
-		NoPrune = true
+	defer func() {
+		if p == "" {
+			slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
+		}
+	}()
+
+	// On Windows we do not carry the payloads inside the main executable
+	exe, err := os.Executable()
+	if err != nil {
+		return
 	}
 
-	if origins := clean("OLLAMA_ORIGINS"); origins != "" {
-		AllowOrigins = strings.Split(origins, ",")
+	cwd, err := os.Getwd()
+	if err != nil {
+		return
 	}
-	for _, allowOrigin := range defaultAllowOrigins {
-		AllowOrigins = append(AllowOrigins,
-			fmt.Sprintf("http://%s", allowOrigin),
-			fmt.Sprintf("https://%s", allowOrigin),
-			fmt.Sprintf("http://%s:*", allowOrigin),
-			fmt.Sprintf("https://%s:*", allowOrigin),
+
+	var paths []string
+	for _, root := range []string{filepath.Dir(exe), cwd} {
+		paths = append(paths,
+			root,
+			filepath.Join(root, "windows-"+runtime.GOARCH),
+			filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
 		)
 	}
 
-	maxRunners := clean("OLLAMA_MAX_LOADED_MODELS")
-	if maxRunners != "" {
-		m, err := strconv.Atoi(maxRunners)
-		if err != nil {
-			slog.Error("invalid setting", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
-		} else {
-			MaxRunners = m
+	// Try a few variations to improve developer experience when building from source in the local tree
+	for _, path := range paths {
+		candidate := filepath.Join(path, "ollama_runners")
+		if _, err := os.Stat(candidate); err == nil {
+			p = candidate
+			break
 		}
 	}
 
-	if onp := os.Getenv("OLLAMA_MAX_QUEUE"); onp != "" {
-		p, err := strconv.Atoi(onp)
-		if err != nil || p <= 0 {
-			slog.Error("invalid setting", "OLLAMA_MAX_QUEUE", onp, "error", err)
-		} else {
-			MaxQueuedRequests = p
+	return p
+}
+
+func Uint(key string, defaultValue uint) func() uint {
+	return func() uint {
+		if s := Var(key); s != "" {
+			if n, err := strconv.ParseUint(s, 10, 64); err != nil {
+				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
+			} else {
+				return uint(n)
+			}
 		}
+
+		return defaultValue
+	}
+}
+
+var (
+	// NumParallel sets the number of parallel model requests. NumParallel can be configured via the OLLAMA_NUM_PARALLEL environment variable.
+	NumParallel = Uint("OLLAMA_NUM_PARALLEL", 0)
+	// MaxRunners sets the maximum number of loaded models. MaxRunners can be configured via the OLLAMA_MAX_LOADED_MODELS environment variable.
+	MaxRunners = Uint("OLLAMA_MAX_LOADED_MODELS", 0)
+	// MaxQueue sets the maximum number of queued requests. MaxQueue can be configured via the OLLAMA_MAX_QUEUE environment variable.
+	MaxQueue = Uint("OLLAMA_MAX_QUEUE", 512)
+	// MaxVRAM sets a maximum VRAM override in bytes. MaxVRAM can be configured via the OLLAMA_MAX_VRAM environment variable.
+	MaxVRAM = Uint("OLLAMA_MAX_VRAM", 0)
+)
+
+type EnvVar struct {
+	Name        string
+	Value       any
+	Description string
+}
+
+func AsMap() map[string]EnvVar {
+	ret := map[string]EnvVar{
+		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
+		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
+		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
+		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
+		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary(), "Set LLM library to bypass autodetection"},
+		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners(), "Maximum number of loaded models per GPU"},
+		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueue(), "Maximum number of queued requests"},
+		"OLLAMA_MODELS":            {"OLLAMA_MODELS", Models(), "The path to the models directory"},
+		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory(), "Do not preserve readline history"},
+		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
+		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
+		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
+		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir(), "Location for runners"},
+		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
+		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir(), "Location for temporary files"},
+	}
+	if runtime.GOOS != "darwin" {
+		ret["CUDA_VISIBLE_DEVICES"] = EnvVar{"CUDA_VISIBLE_DEVICES", CudaVisibleDevices(), "Set which NVIDIA devices are visible"}
+		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices(), "Set which AMD devices are visible"}
+		ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices(), "Set which AMD devices are visible"}
+		ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal(), "Set which AMD devices are visible"}
+		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion(), "Override the gfx used for all detected AMD GPUs"}
+		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGPU(), "Enable experimental Intel GPU detection"}
 	}
+	return ret
+}
+
+func Values() map[string]string {
+	vals := make(map[string]string)
+	for k, v := range AsMap() {
+		vals[k] = fmt.Sprintf("%v", v.Value)
+	}
+	return vals
+}
 
-	KeepAlive = clean("OLLAMA_KEEP_ALIVE")
+// Var returns an environment variable stripped of leading and trailing quotes or spaces
+func Var(key string) string {
+	return strings.Trim(strings.TrimSpace(os.Getenv(key)), "\"'")
 }
diff --git a/envconfig/config_test.go b/envconfig/config_test.go
index 429434aeaf6791fc2117d488089869f0b7466913..92a500f15b580cad9d5cce02c2d20f15f9b8e6f0 100644
--- a/envconfig/config_test.go
+++ b/envconfig/config_test.go
@@ -1,23 +1,235 @@
 package envconfig
 
 import (
+	"math"
 	"testing"
+	"time"
 
-	"github.com/stretchr/testify/require"
+	"github.com/google/go-cmp/cmp"
 )
 
-func TestConfig(t *testing.T) {
-	Debug = false // Reset whatever was loaded in init()
-	t.Setenv("OLLAMA_DEBUG", "")
-	LoadConfig()
-	require.False(t, Debug)
-	t.Setenv("OLLAMA_DEBUG", "false")
-	LoadConfig()
-	require.False(t, Debug)
-	t.Setenv("OLLAMA_DEBUG", "1")
-	LoadConfig()
-	require.True(t, Debug)
-	t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
-	LoadConfig()
-	require.True(t, FlashAttention)
+func TestHost(t *testing.T) {
+	cases := map[string]struct {
+		value  string
+		expect string
+	}{
+		"empty":               {"", "127.0.0.1:11434"},
+		"only address":        {"1.2.3.4", "1.2.3.4:11434"},
+		"only port":           {":1234", ":1234"},
+		"address and port":    {"1.2.3.4:1234", "1.2.3.4:1234"},
+		"hostname":            {"example.com", "example.com:11434"},
+		"hostname and port":   {"example.com:1234", "example.com:1234"},
+		"zero port":           {":0", ":0"},
+		"too large port":      {":66000", ":11434"},
+		"too small port":      {":-1", ":11434"},
+		"ipv6 localhost":      {"[::1]", "[::1]:11434"},
+		"ipv6 world open":     {"[::]", "[::]:11434"},
+		"ipv6 no brackets":    {"::1", "[::1]:11434"},
+		"ipv6 + port":         {"[::1]:1337", "[::1]:1337"},
+		"extra space":         {" 1.2.3.4 ", "1.2.3.4:11434"},
+		"extra quotes":        {"\"1.2.3.4\"", "1.2.3.4:11434"},
+		"extra space+quotes":  {" \" 1.2.3.4 \" ", "1.2.3.4:11434"},
+		"extra single quotes": {"'1.2.3.4'", "1.2.3.4:11434"},
+		"http":                {"http://1.2.3.4", "1.2.3.4:80"},
+		"http port":           {"http://1.2.3.4:4321", "1.2.3.4:4321"},
+		"https":               {"https://1.2.3.4", "1.2.3.4:443"},
+		"https port":          {"https://1.2.3.4:4321", "1.2.3.4:4321"},
+	}
+
+	for name, tt := range cases {
+		t.Run(name, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", tt.value)
+			if host := Host(); host.Host != tt.expect {
+				t.Errorf("%s: expected %s, got %s", name, tt.expect, host.Host)
+			}
+		})
+	}
+}
+
+func TestOrigins(t *testing.T) {
+	cases := []struct {
+		value  string
+		expect []string
+	}{
+		{"", []string{
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+		{"http://10.0.0.1", []string{
+			"http://10.0.0.1",
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+		{"http://172.16.0.1,https://192.168.0.1", []string{
+			"http://172.16.0.1",
+			"https://192.168.0.1",
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+		{"http://totally.safe,http://definitely.legit", []string{
+			"http://totally.safe",
+			"http://definitely.legit",
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+	}
+	for _, tt := range cases {
+		t.Run(tt.value, func(t *testing.T) {
+			t.Setenv("OLLAMA_ORIGINS", tt.value)
+
+			if diff := cmp.Diff(Origins(), tt.expect); diff != "" {
+				t.Errorf("%s: mismatch (-want +got):\n%s", tt.value, diff)
+			}
+		})
+	}
+}
+
+func TestBool(t *testing.T) {
+	cases := map[string]bool{
+		"":      false,
+		"true":  true,
+		"false": false,
+		"1":     true,
+		"0":     false,
+		// invalid values
+		"random":    true,
+		"something": true,
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_BOOL", k)
+			if b := Bool("OLLAMA_BOOL")(); b != v {
+				t.Errorf("%s: expected %t, got %t", k, v, b)
+			}
+		})
+	}
+}
+
+func TestUint(t *testing.T) {
+	cases := map[string]uint{
+		"0":    0,
+		"1":    1,
+		"1337": 1337,
+		// default values
+		"":       11434,
+		"-1":     11434,
+		"0o10":   11434,
+		"0x10":   11434,
+		"string": 11434,
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_UINT", k)
+			if i := Uint("OLLAMA_UINT", 11434)(); i != v {
+				t.Errorf("%s: expected %d, got %d", k, v, i)
+			}
+		})
+	}
+}
+
+func TestKeepAlive(t *testing.T) {
+	cases := map[string]time.Duration{
+		"":       5 * time.Minute,
+		"1s":     time.Second,
+		"1m":     time.Minute,
+		"1h":     time.Hour,
+		"5m0s":   5 * time.Minute,
+		"1h2m3s": 1*time.Hour + 2*time.Minute + 3*time.Second,
+		"0":      time.Duration(0),
+		"60":     60 * time.Second,
+		"120":    2 * time.Minute,
+		"3600":   time.Hour,
+		"-0":     time.Duration(0),
+		"-1":     time.Duration(math.MaxInt64),
+		"-1m":    time.Duration(math.MaxInt64),
+		// invalid values
+		" ":   5 * time.Minute,
+		"???": 5 * time.Minute,
+		"1d":  5 * time.Minute,
+		"1y":  5 * time.Minute,
+		"1w":  5 * time.Minute,
+	}
+
+	for tt, expect := range cases {
+		t.Run(tt, func(t *testing.T) {
+			t.Setenv("OLLAMA_KEEP_ALIVE", tt)
+			if actual := KeepAlive(); actual != expect {
+				t.Errorf("%s: expected %s, got %s", tt, expect, actual)
+			}
+		})
+	}
+}
+
+func TestVar(t *testing.T) {
+	cases := map[string]string{
+		"value":       "value",
+		" value ":     "value",
+		" 'value' ":   "value",
+		` "value" `:   "value",
+		" ' value ' ": " value ",
+		` " value " `: " value ",
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_VAR", k)
+			if s := Var("OLLAMA_VAR"); s != v {
+				t.Errorf("%s: expected %q, got %q", k, v, s)
+			}
+		})
+	}
 }
diff --git a/examples/go-chat/main.go b/examples/go-chat/main.go
index 5266f03e9f33f100a275f77c8874edcc7f8e16fa..7663fb8f4ee0e1a12906a07329c920001f6c615b 100644
--- a/examples/go-chat/main.go
+++ b/examples/go-chat/main.go
@@ -35,7 +35,7 @@ func main() {
 
 	ctx := context.Background()
 	req := &api.ChatRequest{
-		Model:    "llama3",
+		Model:    "llama3.1",
 		Messages: messages,
 	}
 
diff --git a/examples/go-generate-streaming/main.go b/examples/go-generate-streaming/main.go
index 49403351135e09dd25326577cda814006795f8cf..3acfb22a9fc45cab4d57e01649f74c8f130ef57c 100644
--- a/examples/go-generate-streaming/main.go
+++ b/examples/go-generate-streaming/main.go
@@ -16,7 +16,7 @@ func main() {
 
 	// By default, GenerateRequest is streaming.
 	req := &api.GenerateRequest{
-		Model:  "gemma",
+		Model:  "gemma2",
 		Prompt: "how many planets are there?",
 	}
 
diff --git a/examples/go-generate/main.go b/examples/go-generate/main.go
index 50fbf64b7a77f87e6603b0b92e055bf9c5a9d13d..2fe28742b5a63d5f8080a27508592ddd08272afa 100644
--- a/examples/go-generate/main.go
+++ b/examples/go-generate/main.go
@@ -15,7 +15,7 @@ func main() {
 	}
 
 	req := &api.GenerateRequest{
-		Model:  "gemma",
+		Model:  "gemma2",
 		Prompt: "how many planets are there?",
 
 		// set streaming to false
diff --git a/examples/langchain-python-rag-document/README.md b/examples/langchain-python-rag-document/README.md
index 20a73a8838a4a291e91d45625594d4a9c204e88f..e2f3bc02824500aae0d66cc8934a7d5739f2c3c3 100644
--- a/examples/langchain-python-rag-document/README.md
+++ b/examples/langchain-python-rag-document/README.md
@@ -4,6 +4,14 @@ This example provides an interface for asking questions to a PDF document.
 
 ## Setup
 
+1. Ensure you have the `llama3.1` model installed:
+
+```
+ollama pull llama3.1
+```
+
+2. Install the Python Requirements.
+
 ```
 pip install -r requirements.txt
 ```
diff --git a/examples/langchain-python-rag-document/main.py b/examples/langchain-python-rag-document/main.py
index 3ed9499f248b6e64e4e2fab56dec077e5ccc63ea..6f7cec9be3faab6d01cb60bbbeb904c924f60854 100644
--- a/examples/langchain-python-rag-document/main.py
+++ b/examples/langchain-python-rag-document/main.py
@@ -51,7 +51,7 @@ while True:
         template=template,
     )
 
-    llm = Ollama(model="llama3:8b", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    llm = Ollama(model="llama3.1", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
     qa_chain = RetrievalQA.from_chain_type(
         llm,
         retriever=vectorstore.as_retriever(),
diff --git a/examples/langchain-python-rag-privategpt/ingest.py b/examples/langchain-python-rag-privategpt/ingest.py
index 35324775c88d1161767d803ca2e3c4c49c8975a0..0f71ccf0714a32ccbfbf0771a8cc4f78ba99d1f1 100644
--- a/examples/langchain-python-rag-privategpt/ingest.py
+++ b/examples/langchain-python-rag-privategpt/ingest.py
@@ -77,13 +77,21 @@ LOADER_MAPPING = {
 
 
 def load_single_document(file_path: str) -> List[Document]:
-    ext = "." + file_path.rsplit(".", 1)[-1]
-    if ext in LOADER_MAPPING:
-        loader_class, loader_args = LOADER_MAPPING[ext]
-        loader = loader_class(file_path, **loader_args)
-        return loader.load()
+    if os.path.getsize(file_path) != 0:
+        filename, ext = os.path.splitext(file_path)
+        if ext in LOADER_MAPPING:
+            loader_class, loader_args = LOADER_MAPPING[ext]
+            try:
+                loader = loader_class(file_path, **loader_args)
+                if loader:
+                    return loader.load()
+            except:
+                print(f"Corrupted file {file_path}. Ignoring it.")
+        else:
+            print(f"Unsupported file {file_path}. Ignoring it.")
+    else:
+        print(f"Empty file {file_path}. Ignoring it.")
 
-    raise ValueError(f"Unsupported file extension '{ext}'")
 
 def load_documents(source_dir: str, ignored_files: List[str] = []) -> List[Document]:
     """
@@ -100,7 +108,8 @@ def load_documents(source_dir: str, ignored_files: List[str] = []) -> List[Docum
         results = []
         with tqdm(total=len(filtered_files), desc='Loading new documents', ncols=80) as pbar:
             for i, docs in enumerate(pool.imap_unordered(load_single_document, filtered_files)):
-                results.extend(docs)
+                if docs:
+                    results.extend(docs)
                 pbar.update()
 
     return results
diff --git a/examples/langchain-python-rag-privategpt/requirements.txt b/examples/langchain-python-rag-privategpt/requirements.txt
index d4c5cb2e8dd35716d017de7b509a781d9249b306..0aad1fe559918f49fecfd89f363a055dfde3807c 100644
--- a/examples/langchain-python-rag-privategpt/requirements.txt
+++ b/examples/langchain-python-rag-privategpt/requirements.txt
@@ -11,4 +11,5 @@ tabulate==0.9.0
 pandoc==2.3
 pypandoc==1.11
 tqdm==4.66.1
-sentence_transformers==2.2.2
\ No newline at end of file
+sentence_transformers==2.2.2
+numpy>=1.22.2 # not directly required, pinned by Snyk to avoid a vulnerability
\ No newline at end of file
diff --git a/examples/langchain-python-rag-websummary/README.md b/examples/langchain-python-rag-websummary/README.md
index 3f3b98733540d3c77984b796f24e0bc0787beb3f..29c706a392fd2647f3f80f0d9d3bffb3802c821d 100644
--- a/examples/langchain-python-rag-websummary/README.md
+++ b/examples/langchain-python-rag-websummary/README.md
@@ -4,10 +4,10 @@ This example summarizes the website, [https://ollama.com/blog/run-llama2-uncenso
 
 ## Running the Example
 
-1. Ensure you have the `llama2` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama2
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/langchain-python-rag-websummary/main.py b/examples/langchain-python-rag-websummary/main.py
index d1b05ba8aa6bf35fed71e00cea36c4604e0290ad..77b09fbbcdfac3a8f3703a9b8ebcc8361cd0fa98 100644
--- a/examples/langchain-python-rag-websummary/main.py
+++ b/examples/langchain-python-rag-websummary/main.py
@@ -5,8 +5,8 @@ from langchain.chains.summarize import load_summarize_chain
 loader = WebBaseLoader("https://ollama.com/blog/run-llama2-uncensored-locally")
 docs = loader.load()
 
-llm = Ollama(model="llama3")
+llm = Ollama(model="llama3.1")
 chain = load_summarize_chain(llm, chain_type="stuff")
 
-result = chain.invoke(docs) 
+result = chain.invoke(docs)
 print(result)
diff --git a/examples/langchain-python-simple/README.md b/examples/langchain-python-simple/README.md
index d4102dec78211a480d06444cb29ea1f8cb260447..60db2c8c3d80ca73888954b5ea7fcd7ea0b73442 100644
--- a/examples/langchain-python-simple/README.md
+++ b/examples/langchain-python-simple/README.md
@@ -4,10 +4,10 @@ This example is a basic "hello world" of using LangChain with Ollama.
 
 ## Running the Example
 
-1. Ensure you have the `llama3` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/langchain-python-simple/main.py b/examples/langchain-python-simple/main.py
index 7cb6528673d77258bd7066026b8f9c4ed742268a..a7ed81d679b1ee2f712841cd1d82f5e76ebcf8e8 100644
--- a/examples/langchain-python-simple/main.py
+++ b/examples/langchain-python-simple/main.py
@@ -1,6 +1,6 @@
 from langchain.llms import Ollama
 
 input = input("What is your question?")
-llm = Ollama(model="llama3")
+llm = Ollama(model="llama3.1")
 res = llm.predict(input)
 print (res)
diff --git a/examples/modelfile-mario/Modelfile b/examples/modelfile-mario/Modelfile
index 33d5952b1f19863d748803aa7ae1b67f70d070dc..a3747086439eb7786a162783c07c8e1dda320fb2 100644
--- a/examples/modelfile-mario/Modelfile
+++ b/examples/modelfile-mario/Modelfile
@@ -1,4 +1,4 @@
-FROM llama3
+FROM llama3.1
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from super mario bros, acting as an assistant.
diff --git a/examples/modelfile-mario/readme.md b/examples/modelfile-mario/readme.md
index e4f0d4172e524946cfe4d2eeaf942aab1d468ebc..c3f34197a1ec1970cc2e42b814f16ec034535e79 100644
--- a/examples/modelfile-mario/readme.md
+++ b/examples/modelfile-mario/readme.md
@@ -2,12 +2,12 @@
 
 # Example character: Mario
 
-This example shows how to create a basic character using Llama3 as the base model.
+This example shows how to create a basic character using Llama3.1 as the base model.
 
 To run this example:
 
 1. Download the Modelfile
-2. `ollama pull llama3` to get the base model used in the model file.
+2. `ollama pull llama3.1` to get the base model used in the model file.
 3. `ollama create NAME -f ./Modelfile`
 4. `ollama run NAME`
 
@@ -18,7 +18,7 @@ Ask it some questions like "Who are you?" or "Is Peach in trouble again?"
 What the model file looks like:
 
 ```
-FROM llama3
+FROM llama3.1
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from Super Mario Bros, acting as an assistant.
diff --git a/examples/python-dockerit/dockerit.py b/examples/python-dockerit/dockerit.py
index b013102fab8e2fe6acc0f3a807cf5d210d53ed85..6a288d9064a3fe887e3c1fc2ce08d73c6fdfd0de 100644
--- a/examples/python-dockerit/dockerit.py
+++ b/examples/python-dockerit/dockerit.py
@@ -4,7 +4,7 @@ imageName = input("Enter the name of the image: ")
 client = docker.from_env()
 s = requests.Session()
 output=""
-with s.post('http://localhost:11434/api/generate', json={'model': 'dockerit', 'prompt': inputDescription}, stream=True) as r:
+with s.post('http://localhost:11434/api/generate', json={'model': 'mattw/dockerit', 'prompt': inputDescription}, stream=True) as r:
   for line in r.iter_lines():
     if line:
       j = json.loads(line)
diff --git a/examples/python-json-datagenerator/predefinedschema.py b/examples/python-json-datagenerator/predefinedschema.py
index 1fd54892be1454b92f48540680e637efb64006bf..68090ad790bd22ecc3773ce93aef815271f48994 100644
--- a/examples/python-json-datagenerator/predefinedschema.py
+++ b/examples/python-json-datagenerator/predefinedschema.py
@@ -2,7 +2,7 @@ import requests
 import json
 import random
 
-model = "llama3"
+model = "llama3.1"
 template = {
   "firstName": "",
   "lastName": "",
diff --git a/examples/python-json-datagenerator/randomaddresses.py b/examples/python-json-datagenerator/randomaddresses.py
index 72b1fefb9c36a53cd4755095f56546530c329a85..878c98037d97387e9b8480bf497b545f8b9d9878 100644
--- a/examples/python-json-datagenerator/randomaddresses.py
+++ b/examples/python-json-datagenerator/randomaddresses.py
@@ -12,7 +12,7 @@ countries = [
     "France",
 ]
 country = random.choice(countries)
-model = "llama3"
+model = "llama3.1"
 
 prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in {country}, and phone number. Do not use common names. Respond using JSON. Key names should have no backslashes, values should use plain ascii with no special characters."
 
diff --git a/examples/python-json-datagenerator/readme.md b/examples/python-json-datagenerator/readme.md
index 88357044304e3fbf90046be8db73e0f61f596a45..5b444dff13c131f9ab2c3eb8901ab4104db0e924 100644
--- a/examples/python-json-datagenerator/readme.md
+++ b/examples/python-json-datagenerator/readme.md
@@ -6,10 +6,10 @@ There are two python scripts in this example. `randomaddresses.py` generates ran
 
 ## Running the Example
 
-1. Ensure you have the `llama3` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/python-simplechat/client.py b/examples/python-simplechat/client.py
index f82a16b3e61e6c91ef4d19734d15cfd88c5925ba..85043d5f4f24a52ef4c15bd668f513d589d7bcd1 100644
--- a/examples/python-simplechat/client.py
+++ b/examples/python-simplechat/client.py
@@ -2,7 +2,7 @@ import json
 import requests
 
 # NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
-model = "llama3"  # TODO: update this for whatever model you wish to use
+model = "llama3.1"  # TODO: update this for whatever model you wish to use
 
 
 def chat(messages):
diff --git a/examples/python-simplechat/readme.md b/examples/python-simplechat/readme.md
index dd2576bc19c7ef6f933233984344d23e77a11890..4c2ded4d8fd035c87d7ef02d103cdd54bc6181c9 100644
--- a/examples/python-simplechat/readme.md
+++ b/examples/python-simplechat/readme.md
@@ -4,10 +4,10 @@ The **chat** endpoint is one of two ways to generate text from an LLM with Ollam
 
 ## Running the Example
 
-1. Ensure you have the `llama3` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/typescript-simplechat/client.ts b/examples/typescript-simplechat/client.ts
index a1e0eea3761dbe3dc4f948d11f0fa89152766556..8ad113b12c0c50dc772e57907614e8d0bbc55828 100644
--- a/examples/typescript-simplechat/client.ts
+++ b/examples/typescript-simplechat/client.ts
@@ -1,6 +1,6 @@
 import * as readline from "readline";
 
-const model = "llama3";
+const model = "llama3.1";
 type Message = {
   role: "assistant" | "user" | "system";
   content: string;
diff --git a/format/format.go b/format/format.go
index 31059578f7171c9f578290d601a5c140edee00fb..ac50570dfe4c272feb8b7e52c92addddf4a3c162 100644
--- a/format/format.go
+++ b/format/format.go
@@ -3,6 +3,7 @@ package format
 import (
 	"fmt"
 	"math"
+	"strconv"
 )
 
 const (
@@ -28,6 +29,6 @@ func HumanNumber(b uint64) string {
 	case b >= Thousand:
 		return fmt.Sprintf("%.0fK", float64(b)/Thousand)
 	default:
-		return fmt.Sprintf("%d", b)
+		return strconv.FormatUint(b, 10)
 	}
 }
diff --git a/format/format_test.go b/format/format_test.go
index 1d73c80bdabafb55ddf783528ef2aca04183cab7..bff32780300ba1eeea4e928b4444ff3d83dba052 100644
--- a/format/format_test.go
+++ b/format/format_test.go
@@ -5,7 +5,6 @@ import (
 )
 
 func TestHumanNumber(t *testing.T) {
-
 	type testCase struct {
 		input    uint64
 		expected string
diff --git a/go.mod b/go.mod
index 2f3d4ca3488e3ba4a9a4fbf1a8dd1c9df8c2264a..2e0c6614cf510a2c285677650ac92009216a0d87 100644
--- a/go.mod
+++ b/go.mod
@@ -16,7 +16,9 @@ require (
 )
 
 require (
+	github.com/agnivade/levenshtein v1.1.1
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
+	github.com/google/go-cmp v0.6.0
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
@@ -70,7 +72,7 @@ require (
 	golang.org/x/net v0.25.0 // indirect
 	golang.org/x/sys v0.20.0
 	golang.org/x/term v0.20.0
-	golang.org/x/text v0.15.0 // indirect
+	golang.org/x/text v0.15.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )
diff --git a/go.sum b/go.sum
index 9e1baebe002cb568cc21951a588dc8359994360a..926ed26d882b81ac7fff28e32a20701f5b44ac0c 100644
--- a/go.sum
+++ b/go.sum
@@ -4,10 +4,14 @@ dmitri.shuralyov.com/gpu/mtl v0.0.0-20190408044501-666a987793e9/go.mod h1:H6x//7
 gioui.org v0.0.0-20210308172011-57750fc8a0a6/go.mod h1:RSH6KIUZ0p2xy5zHDxgAM4zumjgTw83q2ge/PI+yyw8=
 github.com/BurntSushi/toml v0.3.1/go.mod h1:xHWCNGjB5oqiDr8zfno3MHue2Ht5sIBksp03qcyfWMU=
 github.com/BurntSushi/xgb v0.0.0-20160522181843-27f122750802/go.mod h1:IVnqGOEym/WlBOVXweHU+Q+/VP0lqqI8lqeDx9IjBqo=
+github.com/agnivade/levenshtein v1.1.1 h1:QY8M92nrzkmr798gCo3kmMyqXFzdQVpxLlGPRBij0P8=
+github.com/agnivade/levenshtein v1.1.1/go.mod h1:veldBMzWxcCG2ZvUTKD2kJNRdCk5hVbJomOvKkmgYbo=
 github.com/ajstarks/svgo v0.0.0-20180226025133-644b8db467af/go.mod h1:K08gAheRH3/J6wwsYMMT4xOr94bZjxIelGM0+d/wbFw=
 github.com/antihax/optional v1.0.0/go.mod h1:uupD/76wgC+ih3iEmQUL+0Ugr19nfwCT1kdvxnR2qWY=
 github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 h1:q4dksr6ICHXqG5hm0ZW5IHyeEJXoIJSOZeBLmWPNeIQ=
 github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40/go.mod h1:Q7yQnSMnLvcXlZ8RV+jwz/6y1rQTqbX6C82SndT52Zs=
+github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0 h1:jfIu9sQUG6Ig+0+Ap1h4unLjW6YQJpKZVmUzxsD4E/Q=
+github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0/go.mod h1:t2tdKJDJF9BV14lnkjHmOQgcvEKgtqs5a1N3LNdJhGE=
 github.com/boombuler/barcode v1.0.0/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
 github.com/bytedance/sonic v1.11.6 h1:oUp34TzMlL+OY1OUWxHqsdkgC/Zfc85zGqw9siXjrc0=
 github.com/bytedance/sonic v1.11.6/go.mod h1:LysEHSvpvDySVdC2f87zGWf6CIKJcAvqab1ZaiQtds4=
@@ -36,6 +40,8 @@ github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1/go.mod h1:uw2gLc
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/dgryski/trifles v0.0.0-20200323201526-dd97f9abfb48 h1:fRzb/w+pyskVMQ+UbP35JkH8yB7MYb4q/qhBarqZE6g=
+github.com/dgryski/trifles v0.0.0-20200323201526-dd97f9abfb48/go.mod h1:if7Fbed8SFyPtHLHbg49SI7NAdJiC5WIA09pe59rfAA=
 github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
 github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/envoyproxy/go-control-plane v0.9.0/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
diff --git a/gpu/amd_common.go b/gpu/amd_common.go
index 27a81e3f8e6c8c4d218720ba7f90befca0ebf380..2839cb7c31fd0bf1cc18bf174a4c907ae70eb9fc 100644
--- a/gpu/amd_common.go
+++ b/gpu/amd_common.go
@@ -3,7 +3,7 @@
 package gpu
 
 import (
-	"fmt"
+	"errors"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -49,9 +49,17 @@ func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 }
 
 func commonAMDValidateLibDir() (string, error) {
-	// We try to favor system paths first, so that we can wire up the subprocess to use
-	// the system version.  Only use our bundled version if the system version doesn't work
-	// This gives users a more recovery options if versions have subtle problems at runtime
+	// Favor our bundled version
+
+	// Installer payload location if we're running the installed binary
+	exe, err := os.Executable()
+	if err == nil {
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
+	}
 
 	// Prefer explicit HIP env var
 	hipPath := os.Getenv("HIP_PATH")
@@ -87,14 +95,5 @@ func commonAMDValidateLibDir() (string, error) {
 		}
 	}
 
-	// Installer payload location if we're running the installed binary
-	exe, err := os.Executable()
-	if err == nil {
-		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
-		if rocmLibUsable(rocmTargetDir) {
-			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-			return rocmTargetDir, nil
-		}
-	}
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }
diff --git a/gpu/amd_hip_windows.go b/gpu/amd_hip_windows.go
index 8572a24c5347dce096cceebf5644100b59bfe5c4..2cea282423db01c4114bb126ec253802d1943f53 100644
--- a/gpu/amd_hip_windows.go
+++ b/gpu/amd_hip_windows.go
@@ -1,6 +1,7 @@
 package gpu
 
 import (
+	"errors"
 	"fmt"
 	"log/slog"
 	"syscall"
@@ -33,9 +34,10 @@ type HipLib struct {
 }
 
 func NewHipLib() (*HipLib, error) {
-	h, err := windows.LoadLibrary("amdhip64.dll")
+	// At runtime we depend on v6, so discover GPUs with the same library for a consistent set of GPUs
+	h, err := windows.LoadLibrary("amdhip64_6.dll")
 	if err != nil {
-		return nil, fmt.Errorf("unable to load amdhip64.dll: %w", err)
+		return nil, fmt.Errorf("unable to load amdhip64_6.dll, please make sure to upgrade to the latest amd driver: %w", err)
 	}
 	hl := &HipLib{}
 	hl.dll = h
@@ -75,7 +77,7 @@ func (hl *HipLib) Release() {
 
 func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	if hl.dll == 0 {
-		return 0, 0, fmt.Errorf("dll has been unloaded")
+		return 0, 0, errors.New("dll has been unloaded")
 	}
 	var version int
 	status, _, err := syscall.SyscallN(hl.hipDriverGetVersion, uintptr(unsafe.Pointer(&version)))
@@ -84,9 +86,8 @@ func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	}
 
 	slog.Debug("hipDriverGetVersion", "version", version)
-	// TODO - this isn't actually right, but the docs claim hipDriverGetVersion isn't accurate anyway...
-	driverMajor = version / 1000
-	driverMinor = (version - (driverMajor * 1000)) / 10
+	driverMajor = version / 10000000
+	driverMinor = (version - (driverMajor * 10000000)) / 100000
 
 	return driverMajor, driverMinor, nil
 }
@@ -110,7 +111,7 @@ func (hl *HipLib) HipGetDeviceCount() int {
 
 func (hl *HipLib) HipSetDevice(device int) error {
 	if hl.dll == 0 {
-		return fmt.Errorf("dll has been unloaded")
+		return errors.New("dll has been unloaded")
 	}
 	status, _, err := syscall.SyscallN(hl.hipSetDevice, uintptr(device))
 	if status != hipSuccess {
@@ -121,7 +122,7 @@ func (hl *HipLib) HipSetDevice(device int) error {
 
 func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, error) {
 	if hl.dll == 0 {
-		return nil, fmt.Errorf("dll has been unloaded")
+		return nil, errors.New("dll has been unloaded")
 	}
 	var props hipDevicePropMinimal
 	status, _, err := syscall.SyscallN(hl.hipGetDeviceProperties, uintptr(unsafe.Pointer(&props)), uintptr(device))
@@ -134,7 +135,7 @@ func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, err
 // free, total, err
 func (hl *HipLib) HipMemGetInfo() (uint64, uint64, error) {
 	if hl.dll == 0 {
-		return 0, 0, fmt.Errorf("dll has been unloaded")
+		return 0, 0, errors.New("dll has been unloaded")
 	}
 	var totalMemory uint64
 	var freeMemory uint64
diff --git a/gpu/amd_linux.go b/gpu/amd_linux.go
index 5b1f33602414248b5b6ddfa3d8f1dd895f10daf7..fee5834efa8e811eb1fe70f11875aaf986ac6f4a 100644
--- a/gpu/amd_linux.go
+++ b/gpu/amd_linux.go
@@ -10,9 +10,11 @@ import (
 	"path/filepath"
 	"regexp"
 	"slices"
+	"sort"
 	"strconv"
 	"strings"
 
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 )
 
@@ -20,25 +22,35 @@ import (
 
 const (
 	// DriverVersionFile     = "/sys/module/amdgpu/version"
-	DriverVersionFile = "/sys/module/hydcu/version"
+	DriverVersionFile     = "/sys/module/hydcu/version"
 	AMDNodesSysfsDir      = "/sys/class/kfd/kfd/topology/nodes/"
 	GPUPropertiesFileGlob = AMDNodesSysfsDir + "*/properties"
 
 	// Prefix with the node dir
 	GPUTotalMemoryFileGlob = "mem_banks/*/properties" // size_in_bytes line
-	GPUUsedMemoryFileGlob  = "mem_banks/*/used_memory"
+
+	// Direct Rendering Manager sysfs location
+	DRMDeviceDirGlob   = "/sys/class/drm/card*/device"
+	DRMTotalMemoryFile = "mem_info_vram_total"
+	DRMUsedMemoryFile  = "mem_info_vram_used"
+
+	// In hex; properties file is in decimal
+	DRMUniqueIDFile = "unique_id"
+	DRMVendorFile   = "vendor"
+	DRMDeviceFile   = "device"
 )
 
 var (
 	// Used to validate if the given ROCm lib is usable
 	// ROCmLibGlobs          = []string{"libhipblas.so.2*", "rocblas"} // TODO - probably include more coverage of files here...
 	ROCmLibGlobs = []string{"libhipblas.so.0.1","rocblas"}
+	// RocmStandardLocations = []string{"/opt/rocm/lib", "/usr/lib64"}
 	RocmStandardLocations = []string{"/opt/dtk/lib", "/usr/lib64"}
 )
 
 // Gather GPU information from the amdgpu driver if any supported GPUs are detected
-func AMDGetGPUInfo() []GpuInfo {
-	resp := []GpuInfo{}
+func AMDGetGPUInfo() []RocmGPUInfo {
+	resp := []RocmGPUInfo{}
 	if !AMDDetected() {
 		return resp
 	}
@@ -52,9 +64,9 @@ func AMDGetGPUInfo() []GpuInfo {
 
 	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
 	var visibleDevices []string
-	hipVD := os.Getenv("HIP_VISIBLE_DEVICES")   // zero based index only
-	rocrVD := os.Getenv("ROCR_VISIBLE_DEVICES") // zero based index or UUID, but consumer cards seem to not support UUID
-	gpuDO := os.Getenv("GPU_DEVICE_ORDINAL")    // zero based index
+	hipVD := envconfig.HipVisibleDevices()   // zero based index only
+	rocrVD := envconfig.RocrVisibleDevices() // zero based index or UUID, but consumer cards seem to not support UUID
+	gpuDO := envconfig.GpuDeviceOrdinal()    // zero based index
 	switch {
 	// TODO is this priorty order right?
 	case hipVD != "":
@@ -67,13 +79,27 @@ func AMDGetGPUInfo() []GpuInfo {
 		visibleDevices = strings.Split(gpuDO, ",")
 	}
 
-	gfxOverride := os.Getenv("HSA_OVERRIDE_GFX_VERSION")
+	gfxOverride := envconfig.HsaOverrideGfxVersion()
 	var supported []string
 	libDir := ""
 
 	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
 	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
 	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
+	sort.Slice(matches, func(i, j int) bool {
+		// /sys/class/kfd/kfd/topology/nodes//properties
+		a, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[i])), 10, 64)
+		if err != nil {
+			slog.Debug("parse err", "error", err, "match", matches[i])
+			return false
+		}
+		b, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[j])), 10, 64)
+		if err != nil {
+			slog.Debug("parse err", "error", err, "match", matches[i])
+			return false
+		}
+		return a < b
+	})
 	cpuCount := 0
 	for _, match := range matches {
 		slog.Debug("evaluating amdgpu node " + match)
@@ -92,7 +118,7 @@ func AMDGetGPUInfo() []GpuInfo {
 		scanner := bufio.NewScanner(fp)
 		isCPU := false
 		var major, minor, patch uint64
-		var vendor, device uint64
+		var vendor, device, uniqueID uint64
 		for scanner.Scan() {
 			line := strings.TrimSpace(scanner.Text())
 			// Note: we could also use "cpu_cores_count X" where X is greater than zero to detect CPUs
@@ -123,30 +149,43 @@ func AMDGetGPUInfo() []GpuInfo {
 			} else if strings.HasPrefix(line, "vendor_id") {
 				ver := strings.Fields(line)
 				if len(ver) != 2 {
-					slog.Debug("malformed vendor_id", "vendor_id", line)
+					slog.Debug("malformed", "vendor_id", line)
 					continue
 				}
-				vendor, err = strconv.ParseUint(ver[1], 10, 32)
+				vendor, err = strconv.ParseUint(ver[1], 10, 64)
 				if err != nil {
-					slog.Debug("malformed vendor_id" + line)
+					slog.Debug("malformed", "vendor_id", line, "error", err)
 				}
 			} else if strings.HasPrefix(line, "device_id") {
 				ver := strings.Fields(line)
 				if len(ver) != 2 {
-					slog.Debug("malformed device_id", "device_id", line)
+					slog.Debug("malformed", "device_id", line)
+					continue
+				}
+				device, err = strconv.ParseUint(ver[1], 10, 64)
+				if err != nil {
+					slog.Debug("malformed", "device_id", line, "error", err)
+				}
+			} else if strings.HasPrefix(line, "unique_id") {
+				ver := strings.Fields(line)
+				if len(ver) != 2 {
+					slog.Debug("malformed", "unique_id", line)
 					continue
 				}
-				device, err = strconv.ParseUint(ver[1], 10, 32)
+				uniqueID, err = strconv.ParseUint(ver[1], 10, 64)
 				if err != nil {
-					slog.Debug("malformed device_id" + line)
+					slog.Debug("malformed", "unique_id", line, "error", err)
 				}
 			}
-
 			// TODO - any other properties we want to extract and record?
 			// vendor_id + device_id -> pci lookup for "Name"
 			// Other metrics that may help us understand relative performance between multiple GPUs
 		}
 
+		// Note: while ./mem_banks/*/used_memory exists, it doesn't appear to take other VRAM consumers
+		// into consideration, so we instead map the device over to the DRM driver sysfs nodes which
+		// do reliably report VRAM usage.
+
 		if isCPU {
 			cpuCount++
 			continue
@@ -158,7 +197,7 @@ func AMDGetGPUInfo() []GpuInfo {
 		// Shouldn't happen, but just in case...
 		if gpuID < 0 {
 			slog.Error("unexpected amdgpu sysfs data resulted in negative GPU ID, please set OLLAMA_DEBUG=1 and report an issue")
-			return []GpuInfo{}
+			return nil
 		}
 
 		if int(major) < RocmComputeMin {
@@ -169,65 +208,68 @@ func AMDGetGPUInfo() []GpuInfo {
 		// Look up the memory for the current node
 		totalMemory := uint64(0)
 		usedMemory := uint64(0)
-		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUTotalMemoryFileGlob)
-		propFiles, err := filepath.Glob(propGlob)
-		if err != nil {
-			slog.Warn("error looking up total GPU memory", "glob", propGlob, "error", err)
+		var usedFile string
+		mapping := []struct {
+			id       uint64
+			filename string
+		}{
+			{vendor, DRMVendorFile},
+			{device, DRMDeviceFile},
+			{uniqueID, DRMUniqueIDFile}, // Not all devices will report this
 		}
-		// 1 or more memory banks - sum the values of all of them
-		for _, propFile := range propFiles {
-			fp, err := os.Open(propFile)
-			if err != nil {
-				slog.Warn("failed to open sysfs node", "file", propFile, "erroir", err)
-				continue
-			}
-			defer fp.Close()
-			scanner := bufio.NewScanner(fp)
-			for scanner.Scan() {
-				line := strings.TrimSpace(scanner.Text())
-				if strings.HasPrefix(line, "size_in_bytes") {
-					ver := strings.Fields(line)
-					if len(ver) != 2 {
-						slog.Warn("malformed " + line)
-						continue
-					}
-					bankSizeInBytes, err := strconv.ParseUint(ver[1], 10, 64)
-					if err != nil {
-						slog.Warn("malformed int " + line)
-						continue
-					}
-					totalMemory += bankSizeInBytes
+		slog.Debug("mapping amdgpu to drm sysfs nodes", "amdgpu", match, "vendor", vendor, "device", device, "unique_id", uniqueID)
+		// Map over to DRM location to find the total/free memory
+		drmMatches, _ := filepath.Glob(DRMDeviceDirGlob)
+		for _, devDir := range drmMatches {
+			matched := true
+			for _, m := range mapping {
+				if m.id == 0 {
+					// Null ID means it didn't populate, so we can't use it to match
+					continue
+				}
+				filename := filepath.Join(devDir, m.filename)
+				buf, err := os.ReadFile(filename)
+				if err != nil {
+					slog.Debug("failed to read sysfs node", "file", filename, "error", err)
+					matched = false
+					break
+				}
+				// values here are in hex, strip off the lead 0x and parse so we can compare the numeric (decimal) values in amdgpu
+				cmp, err := strconv.ParseUint(strings.TrimPrefix(strings.TrimSpace(string(buf)), "0x"), 16, 64)
+				if err != nil {
+					slog.Debug("failed to parse sysfs node", "file", filename, "error", err)
+					matched = false
+					break
+				}
+				if cmp != m.id {
+					matched = false
+					break
 				}
 			}
-		}
-		if totalMemory == 0 {
-			slog.Warn("amdgpu reports zero total memory", "gpu", gpuID)
-			continue
-		}
-		usedGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUUsedMemoryFileGlob)
-		usedFiles, err := filepath.Glob(usedGlob)
-		if err != nil {
-			slog.Warn("error looking up used GPU memory", "glob", usedGlob, "error", err)
-			continue
-		}
-		for _, usedFile := range usedFiles {
-			fp, err := os.Open(usedFile)
-			if err != nil {
-				slog.Warn("failed to open sysfs node", "file", usedFile, "error", err)
+			if !matched {
 				continue
 			}
-			defer fp.Close()
-			data, err := io.ReadAll(fp)
+
+			// Found the matching DRM directory
+			slog.Debug("matched", "amdgpu", match, "drm", devDir)
+			totalFile := filepath.Join(devDir, DRMTotalMemoryFile)
+			buf, err := os.ReadFile(totalFile)
 			if err != nil {
-				slog.Warn("failed to read sysfs node", "file", usedFile, "error", err)
-				continue
+				slog.Debug("failed to read sysfs node", "file", totalFile, "error", err)
+				break
 			}
-			used, err := strconv.ParseUint(strings.TrimSpace(string(data)), 10, 64)
+			totalMemory, err = strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
 			if err != nil {
-				slog.Warn("malformed used memory", "data", string(data), "error", err)
-				continue
+				slog.Debug("failed to parse sysfs node", "file", totalFile, "error", err)
+				break
 			}
-			usedMemory += used
+
+			usedFile = filepath.Join(devDir, DRMUsedMemoryFile)
+			usedMemory, err = getFreeMemory(usedFile)
+			if err != nil {
+				slog.Debug("failed to update used memory", "error", err)
+			}
+			break
 		}
 
 		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
@@ -243,18 +285,21 @@ func AMDGetGPUInfo() []GpuInfo {
 
 		slog.Debug("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
 		slog.Debug("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
-		gpuInfo := GpuInfo{
-			Library: "rocm",
-			memInfo: memInfo{
-				TotalMemory: totalMemory,
-				FreeMemory:  (totalMemory - usedMemory),
+		gpuInfo := RocmGPUInfo{
+			GpuInfo: GpuInfo{
+				Library: "rocm",
+				memInfo: memInfo{
+					TotalMemory: totalMemory,
+					FreeMemory:  (totalMemory - usedMemory),
+				},
+				ID:            strconv.Itoa(gpuID),
+				Name:          name,
+				Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
+				MinimumMemory: rocmMinimumMemory,
+				DriverMajor:   driverMajor,
+				DriverMinor:   driverMinor,
 			},
-			ID:            fmt.Sprintf("%d", gpuID),
-			Name:          name,
-			Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
-			MinimumMemory: rocmMinimumMemory,
-			DriverMajor:   driverMajor,
-			DriverMinor:   driverMinor,
+			usedFilepath: usedFile,
 		}
 
 		// If the user wants to filter to a subset of devices, filter out if we aren't a match
@@ -278,7 +323,7 @@ func AMDGetGPUInfo() []GpuInfo {
 			libDir, err = AMDValidateLibDir()
 			if err != nil {
 				slog.Warn("unable to verify rocm library, will use cpu", "error", err)
-				return []GpuInfo{}
+				return nil
 			}
 		}
 		gpuInfo.DependencyPath = libDir
@@ -289,7 +334,7 @@ func AMDGetGPUInfo() []GpuInfo {
 				supported, err = GetSupportedGFX(libDir)
 				if err != nil {
 					slog.Warn("failed to lookup supported GFX types, falling back to CPU mode", "error", err)
-					return []GpuInfo{}
+					return nil
 				}
 				slog.Debug("rocm supported GPUs", "types", supported)
 			}
@@ -306,6 +351,11 @@ func AMDGetGPUInfo() []GpuInfo {
 			slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
 		}
 
+		// Check for env var workarounds
+		if name == "1002:687f" { // Vega RX 56
+			gpuInfo.EnvWorkarounds = append(gpuInfo.EnvWorkarounds, [2]string{"HSA_ENABLE_SDMA", "0"})
+		}
+
 		// The GPU has passed all the verification steps and is supported
 		resp = append(resp, gpuInfo)
 	}
@@ -346,7 +396,7 @@ func AMDValidateLibDir() (string, error) {
 
 	// If we still haven't found a usable rocm, the user will have to install it on their own
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Either install rocm v6, or follow manual install instructions at https://github.com/ollama/ollama/blob/main/docs/linux.md#manual-install")
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }
 
 func AMDDriverVersion() (driverMajor, driverMinor int, err error) {
@@ -380,3 +430,31 @@ func AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	}
 	return driverMajor, driverMinor, nil
 }
+
+func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
+	if len(gpus) == 0 {
+		return nil
+	}
+	for i := range gpus {
+		usedMemory, err := getFreeMemory(gpus[i].usedFilepath)
+		if err != nil {
+			return err
+		}
+		slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(gpus[i].TotalMemory-usedMemory))
+		gpus[i].FreeMemory = gpus[i].TotalMemory - usedMemory
+	}
+	return nil
+}
+
+func getFreeMemory(usedFile string) (uint64, error) {
+	buf, err := os.ReadFile(usedFile)
+	if err != nil {
+		return 0, fmt.Errorf("failed to read sysfs node %s %w", usedFile, err)
+	}
+	usedMemory, err := strconv.ParseUint(strings.TrimSpace(string(buf)), 10, 64)
+	if err != nil {
+		slog.Debug("failed to parse sysfs node", "file", usedFile, "error", err)
+		return 0, fmt.Errorf("failed to parse sysfs node %s %w", usedFile, err)
+	}
+	return usedMemory, nil
+}
diff --git a/gpu/amd_windows.go b/gpu/amd_windows.go
index aae6c5b7acd8c5dd421911313edd23bdf2d915f3..edabeb43a877e59ae001aaa643dc6f95f64be787 100644
--- a/gpu/amd_windows.go
+++ b/gpu/amd_windows.go
@@ -2,13 +2,15 @@ package gpu
 
 import (
 	"bytes"
-	"fmt"
+	"errors"
 	"log/slog"
 	"os"
 	"path/filepath"
 	"slices"
+	"strconv"
 	"strings"
 
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 )
 
@@ -20,12 +22,12 @@ const (
 
 var (
 	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // TODO - probably include more coverage of files here...
-	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\5.7\\bin"} // TODO glob?
+	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // This is not sufficient to discern v5 vs v6
+	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\6.1\\bin"} // TODO glob?
 )
 
-func AMDGetGPUInfo() []GpuInfo {
-	resp := []GpuInfo{}
+func AMDGetGPUInfo() []RocmGPUInfo {
+	resp := []RocmGPUInfo{}
 	hl, err := NewHipLib()
 	if err != nil {
 		slog.Debug(err.Error())
@@ -33,12 +35,11 @@ func AMDGetGPUInfo() []GpuInfo {
 	}
 	defer hl.Release()
 
-	// TODO - this reports incorrect version information, so omitting for now
-	// driverMajor, driverMinor, err := hl.AMDDriverVersion()
-	// if err != nil {
-	// 	// For now this is benign, but we may eventually need to fail compatibility checks
-	// 	slog.Debug("error looking up amd driver version", "error", err)
-	// }
+	driverMajor, driverMinor, err := hl.AMDDriverVersion()
+	if err != nil {
+		// For now this is benign, but we may eventually need to fail compatibility checks
+		slog.Debug("error looking up amd driver version", "error", err)
+	}
 
 	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
@@ -52,7 +53,7 @@ func AMDGetGPUInfo() []GpuInfo {
 	}
 
 	var supported []string
-	gfxOverride := os.Getenv("HSA_OVERRIDE_GFX_VERSION")
+	gfxOverride := envconfig.HsaOverrideGfxVersion()
 	if gfxOverride == "" {
 		supported, err = GetSupportedGFX(libDir)
 		if err != nil {
@@ -65,7 +66,7 @@ func AMDGetGPUInfo() []GpuInfo {
 
 	slog.Debug("detected hip devices", "count", count)
 	// TODO how to determine the underlying device ID when visible devices is causing this to subset?
-	for i := 0; i < count; i++ {
+	for i := range count {
 		err = hl.HipSetDevice(i)
 		if err != nil {
 			slog.Warn("set device", "id", i, "error", err)
@@ -84,14 +85,15 @@ func AMDGetGPUInfo() []GpuInfo {
 		n = bytes.IndexByte(props.GcnArchName[:], 0)
 		gfx := string(props.GcnArchName[:n])
 		slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
-		//slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
+		// slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
 		// TODO  Why isn't props.iGPU accurate!?
 		if strings.EqualFold(name, iGPUName) {
 			slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
 			continue
 		}
 		if gfxOverride == "" {
-			if !slices.Contains[[]string, string](supported, gfx) {
+			// Strip off Target Features when comparing
+			if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
 				slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
 				// TODO - consider discrete markdown just for ROCM troubleshooting?
 				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")
@@ -113,25 +115,27 @@ func AMDGetGPUInfo() []GpuInfo {
 			continue
 		}
 
-		// TODO revisit this once ROCm v6 is available on windows.
-		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
 		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
 		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
-		gpuInfo := GpuInfo{
-			Library: "rocm",
-			memInfo: memInfo{
-				TotalMemory: totalMemory,
-				FreeMemory:  freeMemory,
+		gpuInfo := RocmGPUInfo{
+			GpuInfo: GpuInfo{
+				Library: "rocm",
+				memInfo: memInfo{
+					TotalMemory: totalMemory,
+					FreeMemory:  freeMemory,
+				},
+				// Free memory reporting on Windows is not reliable until we bump to ROCm v6.2
+				UnreliableFreeMemory: true,
+
+				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
+				DependencyPath: libDir,
+				MinimumMemory:  rocmMinimumMemory,
+				Name:           name,
+				Compute:        gfx,
+				DriverMajor:    driverMajor,
+				DriverMinor:    driverMinor,
 			},
-			ID:             fmt.Sprintf("%d", i), // TODO this is probably wrong if we specify visible devices
-			DependencyPath: libDir,
-			MinimumMemory:  rocmMinimumMemory,
-			Name:           name,
-			Compute:        gfx,
-
-			// TODO - this information isn't accurate on windows, so don't report it until we find the right way to retrieve
-			// DriverMajor:    driverMajor,
-			// DriverMinor:    driverMinor,
+			index: i,
 		}
 
 		resp = append(resp, gpuInfo)
@@ -157,5 +161,32 @@ func AMDValidateLibDir() (string, error) {
 
 	// Should not happen on windows since we include it in the installer, but stand-alone binary might hit this
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Please install ROCm")
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
+}
+
+func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
+	if len(gpus) == 0 {
+		return nil
+	}
+	hl, err := NewHipLib()
+	if err != nil {
+		slog.Debug(err.Error())
+		return nil
+	}
+	defer hl.Release()
+
+	for i := range gpus {
+		err := hl.HipSetDevice(gpus[i].index)
+		if err != nil {
+			return err
+		}
+		freeMemory, _, err := hl.HipMemGetInfo()
+		if err != nil {
+			slog.Warn("get mem info", "id", i, "error", err)
+			continue
+		}
+		slog.Debug("updating rocm free memory", "gpu", gpus[i].ID, "name", gpus[i].Name, "before", format.HumanBytes2(gpus[i].FreeMemory), "now", format.HumanBytes2(freeMemory))
+		gpus[i].FreeMemory = freeMemory
+	}
+	return nil
 }
diff --git a/gpu/assets.go b/gpu/assets.go
index e3fbe47cc8af492ca4ab43838fdd0dd97dee9563..6d62d0dca983b677e0533aca4363285bbf9d3253 100644
--- a/gpu/assets.go
+++ b/gpu/assets.go
@@ -26,7 +26,7 @@ func PayloadsDir() (string, error) {
 	defer lock.Unlock()
 	var err error
 	if payloadsDir == "" {
-		runnersDir := envconfig.RunnersDir
+		runnersDir := envconfig.RunnersDir()
 
 		if runnersDir != "" {
 			payloadsDir = runnersDir
@@ -35,27 +35,23 @@ func PayloadsDir() (string, error) {
 
 		// The remainder only applies on non-windows where we still carry payloads in the main executable
 		cleanupTmpDirs()
-		tmpDir := envconfig.TmpDir
+		tmpDir := envconfig.TmpDir()
 		if tmpDir == "" {
 			tmpDir, err = os.MkdirTemp("", "ollama")
 			if err != nil {
 				return "", fmt.Errorf("failed to generate tmp dir: %w", err)
 			}
 		} else {
-			err = os.MkdirAll(tmpDir, 0755)
+			err = os.MkdirAll(tmpDir, 0o755)
 			if err != nil {
 				return "", fmt.Errorf("failed to generate tmp dir %s: %w", tmpDir, err)
 			}
 		}
 
 		// Track our pid so we can clean up orphaned tmpdirs
-		pidFilePath := filepath.Join(tmpDir, "ollama.pid")
-		pidFile, err := os.OpenFile(pidFilePath, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, os.ModePerm)
-		if err != nil {
-			return "", err
-		}
-		if _, err := pidFile.Write([]byte(fmt.Sprint(os.Getpid()))); err != nil {
-			return "", err
+		n := filepath.Join(tmpDir, "ollama.pid")
+		if err := os.WriteFile(n, []byte(strconv.Itoa(os.Getpid())), 0o644); err != nil {
+			return "", fmt.Errorf("failed to write pid file %s: %w", n, err)
 		}
 
 		// We create a distinct subdirectory for payloads within the tmpdir
@@ -67,30 +63,44 @@ func PayloadsDir() (string, error) {
 
 // Best effort to clean up prior tmpdirs
 func cleanupTmpDirs() {
-	dirs, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*"))
+	matches, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*", "ollama.pid"))
 	if err != nil {
 		return
 	}
-	for _, d := range dirs {
-		info, err := os.Stat(d)
-		if err != nil || !info.IsDir() {
+
+	for _, match := range matches {
+		raw, err := os.ReadFile(match)
+		if errors.Is(err, os.ErrNotExist) {
+			slog.Debug("not a ollama runtime directory, skipping", "path", match)
+			continue
+		} else if err != nil {
+			slog.Warn("could not read ollama.pid, skipping", "path", match, "error", err)
 			continue
 		}
-		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
-		if err == nil {
-			pid, err := strconv.Atoi(string(raw))
-			if err == nil {
-				if proc, err := os.FindProcess(int(pid)); err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
-					// Another running ollama, ignore this tmpdir
-					continue
-				}
-			}
-		} else {
-			slog.Debug("failed to open ollama.pid", "path", d, "error", err)
-		}
-		err = os.RemoveAll(d)
+
+		pid, err := strconv.Atoi(string(raw))
 		if err != nil {
-			slog.Debug("unable to cleanup stale tmpdir", "path", d, "error", err)
+			slog.Warn("invalid pid, skipping", "path", match, "error", err)
+			continue
+		}
+
+		p, err := os.FindProcess(pid)
+		if err == nil && !errors.Is(p.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+			slog.Warn("process still running, skipping", "pid", pid, "path", match)
+			continue
+		}
+
+		if err := os.Remove(match); err != nil {
+			slog.Warn("could not cleanup stale pidfile", "path", match, "error", err)
+		}
+
+		runners := filepath.Join(filepath.Dir(match), "runners")
+		if err := os.RemoveAll(runners); err != nil {
+			slog.Warn("could not cleanup stale runners", "path", runners, "error", err)
+		}
+
+		if err := os.Remove(filepath.Dir(match)); err != nil {
+			slog.Warn("could not cleanup stale tmpdir", "path", filepath.Dir(match), "error", err)
 		}
 	}
 }
@@ -98,7 +108,7 @@ func cleanupTmpDirs() {
 func Cleanup() {
 	lock.Lock()
 	defer lock.Unlock()
-	runnersDir := envconfig.RunnersDir
+	runnersDir := envconfig.RunnersDir()
 	if payloadsDir != "" && runnersDir == "" && runtime.GOOS != "windows" {
 		// We want to fully clean up the tmpdir parent of the payloads dir
 		tmpDir := filepath.Clean(filepath.Join(payloadsDir, ".."))
diff --git a/gpu/cpu_common.go b/gpu/cpu_common.go
index 920d0f5b34640015d6857a3c5675d845b1cbc50f..34edcdc5a01d2afd429a38c31964c5e2fca0afe0 100644
--- a/gpu/cpu_common.go
+++ b/gpu/cpu_common.go
@@ -1,21 +1,37 @@
 package gpu
 
 import (
-	"log/slog"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
 
 	"golang.org/x/sys/cpu"
 )
 
-func GetCPUVariant() string {
+func GetCPUCapability() CPUCapability {
 	if cpu.X86.HasAVX2 {
-		slog.Debug("CPU has AVX2")
-		return "avx2"
+		return CPUCapabilityAVX2
 	}
 	if cpu.X86.HasAVX {
-		slog.Debug("CPU has AVX")
-		return "avx"
+		return CPUCapabilityAVX
 	}
-	slog.Debug("CPU does not have vector extensions")
 	// else LCD
-	return ""
+	return CPUCapabilityNone
+}
+
+func IsNUMA() bool {
+	if runtime.GOOS != "linux" {
+		// numa support in llama.cpp is linux only
+		return false
+	}
+	ids := map[string]interface{}{}
+	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
+	for _, packageId := range packageIds {
+		id, err := os.ReadFile(packageId)
+		if err == nil {
+			ids[strings.TrimSpace(string(id))] = struct{}{}
+		}
+	}
+	return len(ids) > 1
 }
diff --git a/gpu/cuda_common.go b/gpu/cuda_common.go
index 03c1a25bbf2b4ecd5357cf33a7c11a2c2de82c01..c90a644c48eeec1dfcb94d076e7349c216b7bd2e 100644
--- a/gpu/cuda_common.go
+++ b/gpu/cuda_common.go
@@ -18,5 +18,4 @@ func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 		ids = append(ids, info.ID)
 	}
 	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
-
 }
diff --git a/gpu/gpu.go b/gpu/gpu.go
index 03e16702d67c2e9b6c911079e816633b5ed628e5..dc124a3ed5981a7a22279ea6a5690e0c9bdfb1f4 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -7,9 +7,9 @@ package gpu
 #cgo windows LDFLAGS: -lpthread
 
 #include "gpu_info.h"
-
 */
 import "C"
+
 import (
 	"fmt"
 	"log/slog"
@@ -24,19 +24,37 @@ import (
 	"github.com/ollama/ollama/format"
 )
 
-type handles struct {
+type cudaHandles struct {
 	deviceCount int
 	cudart      *C.cudart_handle_t
 	nvcuda      *C.nvcuda_handle_t
+	nvml        *C.nvml_handle_t
+}
+
+type oneapiHandles struct {
 	oneapi      *C.oneapi_handle_t
+	deviceCount int
 }
 
 const (
 	cudaMinimumMemory = 457 * format.MebiByte
 	rocmMinimumMemory = 457 * format.MebiByte
+	// TODO OneAPI minimum memory
 )
 
-var gpuMutex sync.Mutex
+var (
+	gpuMutex      sync.Mutex
+	bootstrapped  bool
+	cpuCapability CPUCapability
+	cpus          []CPUInfo
+	cudaGPUs      []CudaGPUInfo
+	nvcudaLibPath string
+	cudartLibPath string
+	oneapiLibPath string
+	nvmlLibPath   string
+	rocmGPUs      []RocmGPUInfo
+	oneapiGPUs    []OneapiGPUInfo
+)
 
 // With our current CUDA compile flags, older than 5.0 will not work properly
 var CudaComputeMin = [2]C.int{5, 0}
@@ -46,113 +64,112 @@ var RocmComputeMin = 9
 // TODO find a better way to detect iGPU instead of minimum memory
 const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
 
-var CudartLinuxGlobs = []string{
-	"/usr/local/cuda/lib64/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
-	"/usr/lib/wsl/lib/libcudart.so*",
-	"/usr/lib/wsl/drivers/*/libcudart.so*",
-	"/opt/cuda/lib64/libcudart.so*",
-	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
-	"/usr/local/cuda/lib*/libcudart.so*",
-	"/usr/lib*/libcudart.so*",
-	"/usr/local/lib*/libcudart.so*",
-}
-
-var CudartWindowsGlobs = []string{
-	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
-}
-
-var NvcudaLinuxGlobs = []string{
-	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
-	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
-	"/usr/lib/*-linux-gnu/libcuda.so*",
-	"/usr/lib/wsl/lib/libcuda.so*",
-	"/usr/lib/wsl/drivers/*/libcuda.so*",
-	"/opt/cuda/lib*/libcuda.so*",
-	"/usr/local/cuda/lib*/libcuda.so*",
-	"/usr/lib*/libcuda.so*",
-	"/usr/local/lib*/libcuda.so*",
-}
-
-var NvcudaWindowsGlobs = []string{
-	"c:\\windows\\system*\\nvcuda.dll",
-}
-
-var OneapiWindowsGlobs = []string{
-	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
-}
-
-var OneapiLinuxGlobs = []string{
-	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
-	"/usr/lib*/libze_intel_gpu.so*",
-}
-
 // Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
 
 // Note: gpuMutex must already be held
-func initGPUHandles() *handles {
-
+func initCudaHandles() *cudaHandles {
 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
 
-	gpuHandles := &handles{}
-	var cudartMgmtName string
+	cHandles := &cudaHandles{}
+	// Short Circuit if we already know which library to use
+	if nvmlLibPath != "" {
+		cHandles.nvml, _ = LoadNVMLMgmt([]string{nvmlLibPath})
+		return cHandles
+	}
+	if nvcudaLibPath != "" {
+		cHandles.deviceCount, cHandles.nvcuda, _ = LoadNVCUDAMgmt([]string{nvcudaLibPath})
+		return cHandles
+	}
+	if cudartLibPath != "" {
+		cHandles.deviceCount, cHandles.cudart, _ = LoadCUDARTMgmt([]string{cudartLibPath})
+		return cHandles
+	}
+
+	slog.Debug("searching for GPU discovery libraries for NVIDIA")
 	var cudartMgmtPatterns []string
-	var nvcudaMgmtName string
-	var nvcudaMgmtPatterns []string
 
-	tmpDir, _ := PayloadsDir()
-	switch runtime.GOOS {
-	case "windows":
-		cudartMgmtName = "cudart64_*.dll"
+	// Aligned with driver, we can't carry as payloads
+	nvcudaMgmtPatterns := NvcudaGlobs
+
+	if runtime.GOOS == "windows" {
 		localAppData := os.Getenv("LOCALAPPDATA")
-		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", cudartMgmtName)}
-		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartWindowsGlobs...)
-		// Aligned with driver, we can't carry as payloads
-		nvcudaMgmtName = "nvcuda.dll"
-		nvcudaMgmtPatterns = NvcudaWindowsGlobs
-	case "linux":
-		cudartMgmtName = "libcudart.so*"
-		if tmpDir != "" {
-			// TODO - add "payloads" for subprocess
-			cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", cudartMgmtName)}
+		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
+	}
+	tmpDir, _ := PayloadsDir()
+	if tmpDir != "" {
+		// TODO - add "payloads" for subprocess
+		cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", CudartMgmtName)}
+	}
+	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)
+
+	if len(NvmlGlobs) > 0 {
+		nvmlLibPaths := FindGPULibs(NvmlMgmtName, NvmlGlobs)
+		if len(nvmlLibPaths) > 0 {
+			nvml, libPath := LoadNVMLMgmt(nvmlLibPaths)
+			if nvml != nil {
+				slog.Debug("nvidia-ml loaded", "library", libPath)
+				cHandles.nvml = nvml
+				nvmlLibPath = libPath
+			}
 		}
-		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartLinuxGlobs...)
-		// Aligned with driver, we can't carry as payloads
-		nvcudaMgmtName = "libcuda.so*"
-		nvcudaMgmtPatterns = NvcudaLinuxGlobs
-	default:
-		return gpuHandles
 	}
 
-	slog.Debug("Detecting GPUs")
-	nvcudaLibPaths := FindGPULibs(nvcudaMgmtName, nvcudaMgmtPatterns)
+	nvcudaLibPaths := FindGPULibs(NvcudaMgmtName, nvcudaMgmtPatterns)
 	if len(nvcudaLibPaths) > 0 {
 		deviceCount, nvcuda, libPath := LoadNVCUDAMgmt(nvcudaLibPaths)
 		if nvcuda != nil {
 			slog.Debug("detected GPUs", "count", deviceCount, "library", libPath)
-			gpuHandles.nvcuda = nvcuda
-			gpuHandles.deviceCount = deviceCount
-			return gpuHandles
+			cHandles.nvcuda = nvcuda
+			cHandles.deviceCount = deviceCount
+			nvcudaLibPath = libPath
+			return cHandles
 		}
 	}
 
-	cudartLibPaths := FindGPULibs(cudartMgmtName, cudartMgmtPatterns)
+	cudartLibPaths := FindGPULibs(CudartMgmtName, cudartMgmtPatterns)
 	if len(cudartLibPaths) > 0 {
 		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
 		if cudart != nil {
 			slog.Debug("detected GPUs", "library", libPath, "count", deviceCount)
-			gpuHandles.cudart = cudart
-			gpuHandles.deviceCount = deviceCount
-			return gpuHandles
+			cHandles.cudart = cudart
+			cHandles.deviceCount = deviceCount
+			cudartLibPath = libPath
+			return cHandles
 		}
 	}
 
-	return gpuHandles
+	return cHandles
+}
+
+// Note: gpuMutex must already be held
+func initOneAPIHandles() *oneapiHandles {
+	oHandles := &oneapiHandles{}
+
+	// Short Circuit if we already know which library to use
+	if oneapiLibPath != "" {
+		oHandles.deviceCount, oHandles.oneapi, _ = LoadOneapiMgmt([]string{oneapiLibPath})
+		return oHandles
+	}
+
+	oneapiLibPaths := FindGPULibs(OneapiMgmtName, OneapiGlobs)
+	if len(oneapiLibPaths) > 0 {
+		oHandles.deviceCount, oHandles.oneapi, oneapiLibPath = LoadOneapiMgmt(oneapiLibPaths)
+	}
+
+	return oHandles
+}
+
+func GetCPUInfo() GpuInfoList {
+	gpuMutex.Lock()
+	if !bootstrapped {
+		gpuMutex.Unlock()
+		GetGPUInfo()
+	} else {
+		gpuMutex.Unlock()
+	}
+	return GpuInfoList{cpus[0].GpuInfo}
 }
 
 func GetGPUInfo() GpuInfoList {
@@ -160,110 +177,291 @@ func GetGPUInfo() GpuInfoList {
 	// GPUs so we can report warnings if we see Nvidia/AMD but fail to load the libraries
 	gpuMutex.Lock()
 	defer gpuMutex.Unlock()
-
-	gpuHandles := initGPUHandles()
+	needRefresh := true
+	var cHandles *cudaHandles
+	var oHandles *oneapiHandles
 	defer func() {
-		if gpuHandles.cudart != nil {
-			C.cudart_release(*gpuHandles.cudart)
+		if cHandles != nil {
+			if cHandles.cudart != nil {
+				C.cudart_release(*cHandles.cudart)
+			}
+			if cHandles.nvcuda != nil {
+				C.nvcuda_release(*cHandles.nvcuda)
+			}
+			if cHandles.nvml != nil {
+				C.nvml_release(*cHandles.nvml)
+			}
 		}
-		if gpuHandles.nvcuda != nil {
-			C.nvcuda_release(*gpuHandles.nvcuda)
+		if oHandles != nil {
+			if oHandles.oneapi != nil {
+				// TODO - is this needed?
+				C.oneapi_release(*oHandles.oneapi)
+			}
 		}
 	}()
 
-	// All our GPU builds on x86 have AVX enabled, so fallback to CPU if we don't detect at least AVX
-	cpuVariant := GetCPUVariant()
-	if cpuVariant == "" && runtime.GOARCH == "amd64" {
-		slog.Warn("CPU does not have AVX or AVX2, disabling GPU support.")
-	}
+	if !bootstrapped {
+		slog.Info("looking for compatible GPUs")
+		needRefresh = false
+		cpuCapability = GetCPUCapability()
+		var memInfo C.mem_info_t
 
-	// On windows we bundle the nvidia library one level above the runner dir
-	depPath := ""
-	if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
-		depPath = filepath.Dir(envconfig.RunnersDir)
-	}
+		mem, err := GetCPUMem()
+		if err != nil {
+			slog.Warn("error looking up system memory", "error", err)
+		}
+		cpus = []CPUInfo{
+			{
+				GpuInfo: GpuInfo{
+					memInfo: mem,
+					Library: "cpu",
+					Variant: cpuCapability,
+					ID:      "0",
+				},
+			},
+		}
 
-	var memInfo C.mem_info_t
-	resp := []GpuInfo{}
+		// Fallback to CPU mode if we're lacking required vector extensions on x86
+		if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
+			slog.Warn("CPU does not have minimum vector extensions, GPU inference disabled", "required", GPURunnerCPUCapability, "detected", cpuCapability)
+			bootstrapped = true
+			// No need to do any GPU discovery, since we can't run on them
+			return GpuInfoList{cpus[0].GpuInfo}
+		}
 
-	// NVIDIA first
-	for i := 0; i < gpuHandles.deviceCount; i++ {
-		// TODO once we support CPU compilation variants of GPU libraries refine this...
-		if cpuVariant == "" && runtime.GOARCH == "amd64" {
-			continue
+		// On windows we bundle the nvidia library one level above the runner dir
+		depPath := ""
+		if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
+			depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "cuda")
 		}
-		if gpuHandles.cudart != nil || gpuHandles.nvcuda != nil {
-			gpuInfo := GpuInfo{
-				Library: "cuda",
+
+		// Load ALL libraries
+		cHandles = initCudaHandles()
+
+		// NVIDIA
+		for i := range cHandles.deviceCount {
+			if cHandles.cudart != nil || cHandles.nvcuda != nil {
+				gpuInfo := CudaGPUInfo{
+					GpuInfo: GpuInfo{
+						Library: "cuda",
+					},
+					index: i,
+				}
+				var driverMajor int
+				var driverMinor int
+				if cHandles.cudart != nil {
+					C.cudart_bootstrap(*cHandles.cudart, C.int(i), &memInfo)
+				} else {
+					C.nvcuda_bootstrap(*cHandles.nvcuda, C.int(i), &memInfo)
+					driverMajor = int(cHandles.nvcuda.driver_major)
+					driverMinor = int(cHandles.nvcuda.driver_minor)
+				}
+				if memInfo.err != nil {
+					slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+					C.free(unsafe.Pointer(memInfo.err))
+					continue
+				}
+				if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
+					slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
+					continue
+				}
+				gpuInfo.TotalMemory = uint64(memInfo.total)
+				gpuInfo.FreeMemory = uint64(memInfo.free)
+				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+				gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
+				gpuInfo.MinimumMemory = cudaMinimumMemory
+				gpuInfo.DependencyPath = depPath
+				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+				gpuInfo.DriverMajor = driverMajor
+				gpuInfo.DriverMinor = driverMinor
+
+				// query the management library as well so we can record any skew between the two
+				// which represents overhead on the GPU we must set aside on subsequent updates
+				if cHandles.nvml != nil {
+					C.nvml_get_free(*cHandles.nvml, C.int(gpuInfo.index), &memInfo.free, &memInfo.total, &memInfo.used)
+					if memInfo.err != nil {
+						slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+						C.free(unsafe.Pointer(memInfo.err))
+					} else {
+						if memInfo.free != 0 && uint64(memInfo.free) > gpuInfo.FreeMemory {
+							gpuInfo.OSOverhead = uint64(memInfo.free) - gpuInfo.FreeMemory
+							slog.Info("detected OS VRAM overhead",
+								"id", gpuInfo.ID,
+								"library", gpuInfo.Library,
+								"compute", gpuInfo.Compute,
+								"driver", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor),
+								"name", gpuInfo.Name,
+								"overhead", format.HumanBytes2(gpuInfo.OSOverhead),
+							)
+						}
+					}
+				}
+
+				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
+				cudaGPUs = append(cudaGPUs, gpuInfo)
+			}
+		}
+
+		// Intel
+		if envconfig.IntelGPU() {
+			oHandles = initOneAPIHandles()
+			if oHandles != nil && oHandles.oneapi != nil {
+
+				// On windows we bundle the oneapi library one level above the runner dir
+				depPath = ""
+				if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
+					depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "oneapi")
+				}
+
+				for d := range oHandles.oneapi.num_drivers {
+					if oHandles.oneapi == nil {
+						// shouldn't happen
+						slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
+						continue
+					}
+					devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
+					for i := range devCount {
+						gpuInfo := OneapiGPUInfo{
+							GpuInfo: GpuInfo{
+								Library: "oneapi",
+							},
+							driverIndex: int(d),
+							gpuIndex:    int(i),
+						}
+						// TODO - split bootstrapping from updating free memory
+						C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
+						// TODO - convert this to MinimumMemory based on testing...
+						var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
+						memInfo.free = C.uint64_t(totalFreeMem)
+						gpuInfo.TotalMemory = uint64(memInfo.total)
+						gpuInfo.FreeMemory = uint64(memInfo.free)
+						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+						gpuInfo.DependencyPath = depPath
+						oneapiGPUs = append(oneapiGPUs, gpuInfo)
+					}
+				}
 			}
-			var driverMajor int
-			var driverMinor int
-			if gpuHandles.cudart != nil {
-				C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
+		}
+
+		rocmGPUs = AMDGetGPUInfo()
+		bootstrapped = true
+		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
+			slog.Info("no compatible GPUs were discovered")
+		}
+	}
+
+	// For detected GPUs, load library if not loaded
+
+	// Refresh free memory usage
+	if needRefresh {
+		mem, err := GetCPUMem()
+		if err != nil {
+			slog.Warn("error looking up system memory", "error", err)
+		} else {
+			slog.Debug("updating system memory data",
+				slog.Group(
+					"before",
+					"total", format.HumanBytes2(cpus[0].TotalMemory),
+					"free", format.HumanBytes2(cpus[0].FreeMemory),
+					"free_swap", format.HumanBytes2(cpus[0].FreeSwap),
+				),
+				slog.Group(
+					"now",
+					"total", format.HumanBytes2(mem.TotalMemory),
+					"free", format.HumanBytes2(mem.FreeMemory),
+					"free_swap", format.HumanBytes2(mem.FreeSwap),
+				),
+			)
+			cpus[0].FreeMemory = mem.FreeMemory
+			cpus[0].FreeSwap = mem.FreeSwap
+		}
+
+		var memInfo C.mem_info_t
+		if cHandles == nil && len(cudaGPUs) > 0 {
+			cHandles = initCudaHandles()
+		}
+		for i, gpu := range cudaGPUs {
+			if cHandles.nvml != nil {
+				C.nvml_get_free(*cHandles.nvml, C.int(gpu.index), &memInfo.free, &memInfo.total, &memInfo.used)
+			} else if cHandles.cudart != nil {
+				C.cudart_bootstrap(*cHandles.cudart, C.int(gpu.index), &memInfo)
+			} else if cHandles.nvcuda != nil {
+				C.nvcuda_get_free(*cHandles.nvcuda, C.int(gpu.index), &memInfo.free, &memInfo.total)
+				memInfo.used = memInfo.total - memInfo.free
 			} else {
-				C.nvcuda_check_vram(*gpuHandles.nvcuda, C.int(i), &memInfo)
-				driverMajor = int(gpuHandles.nvcuda.driver_major)
-				driverMinor = int(gpuHandles.nvcuda.driver_minor)
+				// shouldn't happen
+				slog.Warn("no valid cuda library loaded to refresh vram usage")
+				break
 			}
 			if memInfo.err != nil {
-				slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+				slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
 				C.free(unsafe.Pointer(memInfo.err))
 				continue
 			}
-			if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
-				slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
+			if memInfo.free == 0 {
+				slog.Warn("error looking up nvidia GPU memory")
 				continue
 			}
-			gpuInfo.TotalMemory = uint64(memInfo.total)
-			gpuInfo.FreeMemory = uint64(memInfo.free)
-			gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-			gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
-			gpuInfo.MinimumMemory = cudaMinimumMemory
-			gpuInfo.DependencyPath = depPath
-			gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-			gpuInfo.DriverMajor = int(driverMajor)
-			gpuInfo.DriverMinor = int(driverMinor)
-
-			// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
-			resp = append(resp, gpuInfo)
+			if cHandles.nvml != nil && gpu.OSOverhead > 0 {
+				// When using the management library update based on recorded overhead
+				memInfo.free -= C.uint64_t(gpu.OSOverhead)
+			}
+			slog.Debug("updating cuda memory data",
+				"gpu", gpu.ID,
+				"name", gpu.Name,
+				"overhead", format.HumanBytes2(gpu.OSOverhead),
+				slog.Group(
+					"before",
+					"total", format.HumanBytes2(gpu.TotalMemory),
+					"free", format.HumanBytes2(gpu.FreeMemory),
+				),
+				slog.Group(
+					"now",
+					"total", format.HumanBytes2(uint64(memInfo.total)),
+					"free", format.HumanBytes2(uint64(memInfo.free)),
+					"used", format.HumanBytes2(uint64(memInfo.used)),
+				),
+			)
+			cudaGPUs[i].FreeMemory = uint64(memInfo.free)
 		}
-	}
-
-	// Then AMD
-	resp = append(resp, AMDGetGPUInfo()...)
 
-	if len(resp) == 0 {
-		C.cpu_check_ram(&memInfo)
-		if memInfo.err != nil {
-			slog.Info("error looking up CPU memory", "error", C.GoString(memInfo.err))
-			C.free(unsafe.Pointer(memInfo.err))
-			return resp
+		if oHandles == nil && len(oneapiGPUs) > 0 {
+			oHandles = initOneAPIHandles()
 		}
-		gpuInfo := GpuInfo{
-			Library: "cpu",
-			Variant: cpuVariant,
+		for i, gpu := range oneapiGPUs {
+			if oHandles.oneapi == nil {
+				// shouldn't happen
+				slog.Warn("nil oneapi handle with device count", "count", oHandles.deviceCount)
+				continue
+			}
+			C.oneapi_check_vram(*oHandles.oneapi, C.int(gpu.driverIndex), C.int(gpu.gpuIndex), &memInfo)
+			// TODO - convert this to MinimumMemory based on testing...
+			var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
+			memInfo.free = C.uint64_t(totalFreeMem)
+			oneapiGPUs[i].FreeMemory = uint64(memInfo.free)
 		}
-		gpuInfo.TotalMemory = uint64(memInfo.total)
-		gpuInfo.FreeMemory = uint64(memInfo.free)
-		gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 
-		resp = append(resp, gpuInfo)
+		err = RocmGPUInfoList(rocmGPUs).RefreshFreeMemory()
+		if err != nil {
+			slog.Debug("problem refreshing ROCm free memory", "error", err)
+		}
 	}
 
-	return resp
-}
-
-func GetCPUMem() (memInfo, error) {
-	var ret memInfo
-	var info C.mem_info_t
-	C.cpu_check_ram(&info)
-	if info.err != nil {
-		defer C.free(unsafe.Pointer(info.err))
-		return ret, fmt.Errorf(C.GoString(info.err))
+	resp := []GpuInfo{}
+	for _, gpu := range cudaGPUs {
+		resp = append(resp, gpu.GpuInfo)
+	}
+	for _, gpu := range rocmGPUs {
+		resp = append(resp, gpu.GpuInfo)
+	}
+	for _, gpu := range oneapiGPUs {
+		resp = append(resp, gpu.GpuInfo)
 	}
-	ret.FreeMemory = uint64(info.free)
-	ret.TotalMemory = uint64(info.total)
-	return ret, nil
+	if len(resp) == 0 {
+		resp = append(resp, cpus[0].GpuInfo)
+	}
+	return resp
 }
 
 func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
@@ -296,6 +494,7 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 		// Nvidia PhysX known to return bogus results
 		if strings.Contains(pattern, "PhysX") {
 			slog.Debug("skipping PhysX cuda library path", "path", pattern)
+			continue
 		}
 		// Ignore glob discovery errors
 		matches, _ := filepath.Glob(pattern)
@@ -352,7 +551,23 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 		defer C.free(unsafe.Pointer(lib))
 		C.nvcuda_init(lib, &resp)
 		if resp.err != nil {
-			slog.Debug("Unable to load nvcuda", "library", libPath, "error", C.GoString(resp.err))
+			// Decide what log level based on the type of error message to help users understand why
+			msg := C.GoString(resp.err)
+			switch resp.cudaErr {
+			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
+				slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
+			case C.CUDA_ERROR_NO_DEVICE:
+				slog.Info("no nvidia devices detected", "library", libPath)
+			case C.CUDA_ERROR_UNKNOWN:
+				slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
+				slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
+			default:
+				if strings.Contains(msg, "wrong ELF class") {
+					slog.Debug("skipping 32bit library", "library", libPath)
+				} else {
+					slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
+				}
+			}
 			C.free(unsafe.Pointer(resp.err))
 		} else {
 			return int(resp.num_devices), &resp.ch, libPath
@@ -361,8 +576,26 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 	return 0, nil, ""
 }
 
+func LoadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string) {
+	var resp C.nvml_init_resp_t
+	resp.ch.verbose = getVerboseState()
+	for _, libPath := range nvmlLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.nvml_init(lib, &resp)
+		if resp.err != nil {
+			slog.Info(fmt.Sprintf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err)))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return &resp.ch, libPath
+		}
+	}
+	return nil, ""
+}
+
 func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
 	var resp C.oneapi_init_resp_t
+	num_devices := 0
 	resp.oh.verbose = getVerboseState()
 	for _, libPath := range oneapiLibPaths {
 		lib := C.CString(libPath)
@@ -372,14 +605,17 @@ func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
 			slog.Debug("Unable to load oneAPI management library", "library", libPath, "error", C.GoString(resp.err))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			return int(resp.num_devices), &resp.oh, libPath
+			for i := range resp.oh.num_drivers {
+				num_devices += int(C.oneapi_get_device_count(resp.oh, C.int(i)))
+			}
+			return num_devices, &resp.oh, libPath
 		}
 	}
 	return 0, nil, ""
 }
 
 func getVerboseState() C.uint16_t {
-	if envconfig.Debug {
+	if envconfig.Debug() {
 		return C.uint16_t(1)
 	}
 	return C.uint16_t(0)
diff --git a/gpu/gpu_darwin.go b/gpu/gpu_darwin.go
index f8cc1adbe1b4dfe9cf4e25fa21963d1e52b44a8e..9d9fd84ebf47f275e9de09b7a975d6d4021ceb63 100644
--- a/gpu/gpu_darwin.go
+++ b/gpu/gpu_darwin.go
@@ -8,6 +8,7 @@ package gpu
 #include "gpu_info_darwin.h"
 */
 import "C"
+
 import (
 	"runtime"
 
@@ -24,7 +25,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
-				Variant: GetCPUVariant(),
+				Variant: GetCPUCapability(),
 				memInfo: mem,
 			},
 		}
@@ -42,10 +43,22 @@ func GetGPUInfo() GpuInfoList {
 	return []GpuInfo{info}
 }
 
+func GetCPUInfo() GpuInfoList {
+	mem, _ := GetCPUMem()
+	return []GpuInfo{
+		{
+			Library: "cpu",
+			Variant: GetCPUCapability(),
+			memInfo: mem,
+		},
+	}
+}
+
 func GetCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
-		FreeMemory:  0,
+		FreeMemory:  uint64(C.getFreeMemory()),
+		// FreeSwap omitted as Darwin uses dynamic paging
 	}, nil
 }
 
diff --git a/gpu/gpu_info.h b/gpu/gpu_info.h
index 482b81a6f3528d633e2416f0adc009d9a5d04df8..094b791a88595cc10c565400348e749c247f9e81 100644
--- a/gpu/gpu_info.h
+++ b/gpu/gpu_info.h
@@ -47,6 +47,7 @@ typedef struct mem_info {
   char gpu_name[GPU_NAME_LEN];
   uint64_t total;
   uint64_t free;
+  uint64_t used;
 
   // Compute Capability
   int major; 
@@ -62,7 +63,8 @@ void cpu_check_ram(mem_info_t *resp);
 
 #include "gpu_info_cudart.h"
 #include "gpu_info_nvcuda.h"
+#include "gpu_info_nvml.h"
 #include "gpu_info_oneapi.h"
 
 #endif  // __GPU_INFO_H__
-#endif  // __APPLE__
\ No newline at end of file
+#endif  // __APPLE__
diff --git a/gpu/gpu_info_cudart.c b/gpu/gpu_info_cudart.c
index 8e9204eaffb69179988a2eae13a7fcb6cf9ae292..03f15a2c3e09859d3a8b080efc1e90ec32b9056e 100644
--- a/gpu/gpu_info_cudart.c
+++ b/gpu/gpu_info_cudart.c
@@ -40,7 +40,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
 
   for (i = 0; l[i].s != NULL; i++) {
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!l[i].p) {
+    if (!*(l[i].p)) {
       char *msg = LOAD_ERR();
       LOG(resp->ch.verbose, "dlerr: %s\n", msg);
       UNLOAD_LIBRARY(resp->ch.handle);
@@ -94,7 +94,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
 }
 
 
-void cudart_check_vram(cudart_handle_t h, int i, mem_info_t *resp) {
+void cudart_bootstrap(cudart_handle_t h, int i, mem_info_t *resp) {
   resp->err = NULL;
   cudartMemory_t memInfo = {0,0,0};
   cudartReturn_t ret;
@@ -166,9 +166,11 @@ void cudart_check_vram(cudart_handle_t h, int i, mem_info_t *resp) {
 
   resp->total = memInfo.total;
   resp->free = memInfo.free;
+  resp->used = memInfo.used;
 
   LOG(h.verbose, "[%s] CUDA totalMem %lu\n", resp->gpu_id, resp->total);
   LOG(h.verbose, "[%s] CUDA freeMem %lu\n", resp->gpu_id, resp->free);
+  LOG(h.verbose, "[%s] CUDA usedMem %lu\n", resp->gpu_id, resp->used);
   LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
 }
 
diff --git a/gpu/gpu_info_cudart.h b/gpu/gpu_info_cudart.h
index e8a898560eaadc2e60455c9c59c74bf8f807318b..ff0c0af19f993ab5a4d0731a3c1d7df33958a3fb 100644
--- a/gpu/gpu_info_cudart.h
+++ b/gpu/gpu_info_cudart.h
@@ -140,7 +140,8 @@ typedef struct cudart_init_resp {
 } cudart_init_resp_t;
 
 void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp);
-void cudart_check_vram(cudart_handle_t ch, int device_id, mem_info_t *resp);
+void cudart_bootstrap(cudart_handle_t ch, int device_id, mem_info_t *resp);
+// TODO - if we keep this library longer term, add cudart_get_free
 void cudart_release(cudart_handle_t ch);
 
 #endif  // __GPU_INFO_CUDART_H__
diff --git a/gpu/gpu_info_darwin.h b/gpu/gpu_info_darwin.h
index 3edca237cec197d7843e7878bde2645839605e3f..415e7922d7dbaa34e77124c876ac4889edd878df 100644
--- a/gpu/gpu_info_darwin.h
+++ b/gpu/gpu_info_darwin.h
@@ -2,3 +2,4 @@
 #include 
 uint64_t getRecommendedMaxVRAM();
 uint64_t getPhysicalMemory();
+uint64_t getFreeMemory();
diff --git a/gpu/gpu_info_darwin.m b/gpu/gpu_info_darwin.m
index a145ac076177ebd874414563790ebd0802c4c53c..5ca139e0b215eb779114b332859d3e3494d0f9c5 100644
--- a/gpu/gpu_info_darwin.m
+++ b/gpu/gpu_info_darwin.m
@@ -1,4 +1,5 @@
-// go:build darwin
+#import 
+#import 
 #include "gpu_info_darwin.h"
 
 uint64_t getRecommendedMaxVRAM() {
@@ -8,6 +9,27 @@ uint64_t getRecommendedMaxVRAM() {
   return result;
 }
 
+// getPhysicalMemory returns the total physical memory in bytes
 uint64_t getPhysicalMemory() {
-  return [[NSProcessInfo processInfo] physicalMemory];
+  return [NSProcessInfo processInfo].physicalMemory;
+}
+
+// getFreeMemory returns the total free memory in bytes, including inactive
+// memory that can be reclaimed by the system.
+uint64_t getFreeMemory() {
+  mach_port_t host_port = mach_host_self();
+  mach_msg_type_number_t host_size = sizeof(vm_statistics64_data_t) / sizeof(integer_t);
+  vm_size_t pagesize;
+  vm_statistics64_data_t vm_stat;
+
+  host_page_size(host_port, &pagesize);
+  if (host_statistics64(host_port, HOST_VM_INFO64, (host_info64_t)&vm_stat, &host_size) != KERN_SUCCESS) {
+    return 0;
+  }
+
+  uint64_t free_memory = (uint64_t)vm_stat.free_count * pagesize;
+  free_memory += (uint64_t)vm_stat.speculative_count * pagesize;
+  free_memory += (uint64_t)vm_stat.inactive_count * pagesize;
+
+  return free_memory;
 }
diff --git a/gpu/gpu_info_nvcuda.c b/gpu/gpu_info_nvcuda.c
index 26d855dfa444957839b9fd9d03eef45856be24c1..a1a38bfc27215a132120d6f4319665101289b747 100644
--- a/gpu/gpu_info_nvcuda.c
+++ b/gpu/gpu_info_nvcuda.c
@@ -7,6 +7,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
   CUresult ret;
   resp->err = NULL;
   resp->num_devices = 0;
+  resp->cudaErr = CUDA_SUCCESS;
   const int buflen = 256;
   char buf[buflen + 1];
   int i;
@@ -38,12 +39,13 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
             nvcuda_lib_path, msg);
     free(msg);
     resp->err = strdup(buf);
+    resp->cudaErr = -1;
     return;
   }
 
   for (i = 0; l[i].s != NULL; i++) {
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*l[i].p) {
+    if (!*(l[i].p)) {
       char *msg = LOAD_ERR();
       LOG(resp->ch.verbose, "dlerr: %s\n", msg);
       UNLOAD_LIBRARY(resp->ch.handle);
@@ -52,6 +54,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
               msg);
       free(msg);
       resp->err = strdup(buf);
+      resp->cudaErr = -1;
       return;
     }
   }
@@ -61,12 +64,9 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
-    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
-      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
-      return;
-    }
-    snprintf(buf, buflen, "nvcuda init failure: %d", ret);
+    snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
     resp->err = strdup(buf);
+    resp->cudaErr = ret;
     return;
   }
 
@@ -91,12 +91,13 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     resp->ch.handle = NULL;
     snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
+    resp->cudaErr = ret;
     return;
   }
 }
 
 const int buflen = 256;
-void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
+void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   resp->err = NULL;
   nvcudaMemory_t memInfo = {0,0};
   CUresult ret;
@@ -106,13 +107,13 @@ void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
   CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 
   if (h.handle == NULL) {
-    resp->err = strdup("nvcuda handle isn't initialized");
+    resp->err = strdup("cuda driver library handle isn't initialized");
     return;
   }
 
   ret = (*h.cuDeviceGet)(&device, i);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device failed to initialize");
+    snprintf(buf, buflen, "cuda driver library device failed to initialize");
     resp->err = strdup(buf);
     return;
   }
@@ -168,14 +169,14 @@ void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
   // To get memory we have to set (and release) a context
   ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda failed to get primary device context %d", ret);
+    snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
     resp->err = strdup(buf);
     return;
   }
 
   ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device memory info lookup failure %d", ret);
+    snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
     resp->err = strdup(buf);
     // Best effort on failure...
     (*h.cuCtxDestroy)(ctx);
@@ -193,12 +194,47 @@ void nvcuda_check_vram(nvcuda_handle_t h, int i, mem_info_t *resp) {
 
   ret = (*h.cuCtxDestroy)(ctx);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to release primary device context %d", ret);
+    LOG(1, "cuda driver library failed to release device context %d", ret);
+  }
+}
+
+void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total) {
+  CUresult ret;
+  CUcontext ctx = NULL;
+  CUdevice device = -1;
+  *free = 0;
+  *total = 0;
+
+  ret = (*h.cuDeviceGet)(&device, i);
+  if (ret != CUDA_SUCCESS) {
+    LOG(1, "cuda driver library device failed to initialize");
+    return;
+  }
+
+
+  // To get memory we have to set (and release) a context
+  ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
+  if (ret != CUDA_SUCCESS) {
+    LOG(1, "cuda driver library failed to get device context %d", ret);
+    return;
+  }
+
+  ret = (*h.cuMemGetInfo_v2)(free, total);
+  if (ret != CUDA_SUCCESS) {
+    LOG(1, "cuda driver library device memory info lookup failure %d", ret);
+    // Best effort on failure...
+    (*h.cuCtxDestroy)(ctx);
+    return;
+  }
+
+  ret = (*h.cuCtxDestroy)(ctx);
+  if (ret != CUDA_SUCCESS) {
+    LOG(1, "cuda driver library failed to release device context %d", ret);
   }
 }
 
 void nvcuda_release(nvcuda_handle_t h) {
-  LOG(h.verbose, "releasing nvcuda library\n");
+  LOG(h.verbose, "releasing cuda driver library\n");
   UNLOAD_LIBRARY(h.handle);
   // TODO and other context release logic?
   h.handle = NULL;
diff --git a/gpu/gpu_info_nvcuda.h b/gpu/gpu_info_nvcuda.h
index 2b232839b99eb96189d5cc5d785e3d8ee749d7a4..ef2fe8a305117d2c5c5260b759ab1373b07d25fc 100644
--- a/gpu/gpu_info_nvcuda.h
+++ b/gpu/gpu_info_nvcuda.h
@@ -7,9 +7,12 @@
 typedef enum cudaError_enum {
   CUDA_SUCCESS = 0,
   CUDA_ERROR_INVALID_VALUE = 1,
-  CUDA_ERROR_MEMORY_ALLOCATION = 2,
+  CUDA_ERROR_OUT_OF_MEMORY = 2,
   CUDA_ERROR_NOT_INITIALIZED = 3,
   CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  CUDA_ERROR_NO_DEVICE = 100,
+  CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
+  CUDA_ERROR_UNKNOWN = 999,
   // Other values omitted for now...
 } CUresult;
 
@@ -64,10 +67,12 @@ typedef struct nvcuda_init_resp {
   char *err;  // If err is non-null handle is invalid
   nvcuda_handle_t ch;
   int num_devices;
+  CUresult cudaErr;
 } nvcuda_init_resp_t;
 
 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);
-void nvcuda_check_vram(nvcuda_handle_t ch, int device_id, mem_info_t *resp);
+void nvcuda_bootstrap(nvcuda_handle_t ch, int device_id, mem_info_t *resp);
+void nvcuda_get_free(nvcuda_handle_t ch,  int device_id, uint64_t *free, uint64_t *total);
 void nvcuda_release(nvcuda_handle_t ch);
 
 #endif  // __GPU_INFO_NVCUDA_H__
diff --git a/gpu/gpu_info_nvml.c b/gpu/gpu_info_nvml.c
new file mode 100644
index 0000000000000000000000000000000000000000..11293e4482a750b5ffcce9da6d479faec7aca893
--- /dev/null
+++ b/gpu/gpu_info_nvml.c
@@ -0,0 +1,104 @@
+#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
+
+#include 
+
+#include "gpu_info_nvml.h"
+
+void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
+  nvmlReturn_t ret;
+  resp->err = NULL;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  struct lookup {
+    char *s;
+    void **p;
+  } l[] = {
+      {"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
+      {"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
+      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.nvmlDeviceGetHandleByIndex},
+      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
+      {NULL, NULL},
+  };
+
+  resp->ch.handle = LOAD_LIBRARY(nvml_lib_path, RTLD_LAZY);
+  if (!resp->ch.handle) {
+    char *msg = LOAD_ERR();
+    LOG(resp->ch.verbose, "library %s load err: %s\n", nvml_lib_path, msg);
+    snprintf(buf, buflen,
+             "Unable to load %s library to query for Nvidia GPUs: %s",
+             nvml_lib_path, msg);
+    free(msg);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  // TODO once we've squashed the remaining corner cases remove this log
+  // LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", nvml_lib_path);
+  
+  for (i = 0; l[i].s != NULL; i++) {
+    // TODO once we've squashed the remaining corner cases remove this log
+    // LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
+
+    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
+    if (!*(l[i].p)) {
+      resp->ch.handle = NULL;
+      char *msg = LOAD_ERR();
+      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->ch.handle);
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
+               msg);
+      free(msg);
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+
+  ret = (*resp->ch.nvmlInit_v2)();
+  if (ret != NVML_SUCCESS) {
+    LOG(resp->ch.verbose, "nvmlInit_v2 err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
+    snprintf(buf, buflen, "nvml vram init failure: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+}
+
+
+void nvml_get_free(nvml_handle_t h, int device_id, uint64_t *free, uint64_t *total, uint64_t *used) {
+    nvmlDevice_t device;
+    nvmlMemory_t memInfo = {0};
+    nvmlReturn_t ret;
+    ret = (*h.nvmlDeviceGetHandleByIndex)(device_id, &device);
+    if (ret != NVML_SUCCESS) {
+        LOG(1, "unable to get device handle %d: %d", device_id, ret);
+        *free = 0;
+        return;
+    }
+
+    ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
+    if (ret != NVML_SUCCESS) {
+        LOG(1, "device memory info lookup failure %d: %d", device_id, ret);
+        *free = 0;
+        return;
+    }
+    *free = memInfo.free;
+    *total = memInfo.total;
+    *used = memInfo.used;
+}
+
+
+void nvml_release(nvml_handle_t h) {
+  LOG(h.verbose, "releasing nvml library\n");
+  nvmlReturn_t ret;
+  ret = (*h.nvmlShutdown)();
+  if (ret != NVML_SUCCESS) {
+    LOG(1, "error during nvmlShutdown %d", ret);
+  }
+  UNLOAD_LIBRARY(h.handle);
+  h.handle = NULL;
+}
+
+#endif  // __APPLE__
\ No newline at end of file
diff --git a/gpu/gpu_info_nvml.h b/gpu/gpu_info_nvml.h
new file mode 100644
index 0000000000000000000000000000000000000000..a661f723b365d493dc07e2e8c5a068f3a474e035
--- /dev/null
+++ b/gpu/gpu_info_nvml.h
@@ -0,0 +1,48 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_NVML_H__
+#define __GPU_INFO_NVML_H__
+#include "gpu_info.h"
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum nvmlReturn_enum {
+  NVML_SUCCESS = 0,
+  // Other values omitted for now...
+} nvmlReturn_t;
+typedef void *nvmlDevice_t;  // Opaque is sufficient
+typedef struct nvmlMemory_st {
+  unsigned long long total;
+  unsigned long long free;
+  unsigned long long used;
+} nvmlMemory_t;
+
+typedef enum nvmlBrandType_enum
+{
+    NVML_BRAND_UNKNOWN          = 0,
+} nvmlBrandType_t;
+
+typedef struct nvml_handle {
+  void *handle;
+  uint16_t verbose;
+  nvmlReturn_t (*nvmlInit_v2)(void);
+  nvmlReturn_t (*nvmlShutdown)(void);
+  nvmlReturn_t (*nvmlDeviceGetHandleByIndex)(unsigned int, nvmlDevice_t *);
+  nvmlReturn_t (*nvmlDeviceGetMemoryInfo)(nvmlDevice_t, nvmlMemory_t *);
+} nvml_handle_t;
+
+typedef struct nvml_init_resp {
+  char *err;  // If err is non-null handle is invalid
+  nvml_handle_t ch;
+} nvml_init_resp_t;
+
+typedef struct nvml_compute_capability {
+  char *err;
+  int major;
+  int minor;
+} nvml_compute_capability_t;
+
+void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp);
+void nvml_get_free(nvml_handle_t ch,  int device_id, uint64_t *free, uint64_t *total, uint64_t *used);
+void nvml_release(nvml_handle_t ch);
+
+#endif  // __GPU_INFO_NVML_H__
+#endif  // __APPLE__
\ No newline at end of file
diff --git a/gpu/gpu_info_oneapi.c b/gpu/gpu_info_oneapi.c
index 4be90e80b60d1e3fba25e97b5209a8f83a975922..3ff708ea20df7669192cf84792dc202c7c021e2f 100644
--- a/gpu/gpu_info_oneapi.c
+++ b/gpu/gpu_info_oneapi.c
@@ -4,15 +4,17 @@
 
 #include 
 
-void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp)
-{
+void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
   ze_result_t ret;
   resp->err = NULL;
+  resp->oh.devices = NULL;
+  resp->oh.num_devices = NULL;
+  resp->oh.drivers = NULL;
+  resp->oh.num_drivers = 0;
   const int buflen = 256;
   char buf[buflen + 1];
-  int i;
-  struct lookup
-  {
+  int i, d;
+  struct lookup {
     char *s;
     void **p;
   } l[] = {
@@ -28,8 +30,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp)
   };
 
   resp->oh.handle = LOAD_LIBRARY(oneapi_lib_path, RTLD_LAZY);
-  if (!resp->oh.handle)
-  {
+  if (!resp->oh.handle) {
     char *msg = LOAD_ERR();
     snprintf(buf, buflen,
              "Unable to load %s library to query for Intel GPUs: %s\n",
@@ -44,14 +45,12 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp)
       "wiring Level-Zero management library functions in %s\n",
       oneapi_lib_path);
 
-  for (i = 0; l[i].s != NULL; i++)
-  {
+  for (i = 0; l[i].s != NULL; i++) {
     // TODO once we've squashed the remaining corner cases remove this log
     LOG(resp->oh.verbose, "dlsym: %s\n", l[i].s);
 
     *l[i].p = LOAD_SYMBOL(resp->oh.handle, l[i].s);
-    if (!l[i].p)
-    {
+    if (!*(l[i].p)) {
       resp->oh.handle = NULL;
       char *msg = LOAD_ERR();
       LOG(resp->oh.verbose, "dlerr: %s\n", msg);
@@ -63,23 +62,70 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp)
     }
   }
 
+  LOG(resp->oh.verbose, "calling zesInit\n");
+
   ret = (*resp->oh.zesInit)(0);
-  if (ret != ZE_RESULT_SUCCESS)
-  {
-    LOG(resp->oh.verbose, "zesInit err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->oh.handle);
-    resp->oh.handle = NULL;
-    snprintf(buf, buflen, "oneapi vram init failure: %d", ret);
+  if (ret != ZE_RESULT_SUCCESS) {
+    LOG(resp->oh.verbose, "zesInit err: %x\n", ret);
+    snprintf(buf, buflen, "oneapi vram init failure: %x", ret);
+    resp->err = strdup(buf);
+    oneapi_release(resp->oh);
+    return;
+  }
+
+  LOG(resp->oh.verbose, "calling zesDriverGet\n");
+  ret = (*resp->oh.zesDriverGet)(&resp->oh.num_drivers, NULL);
+  if (ret != ZE_RESULT_SUCCESS) {
+    LOG(resp->oh.verbose, "zesDriverGet err: %x\n", ret);
+    snprintf(buf, buflen, "unable to get driver count: %x", ret);
+    resp->err = strdup(buf);
+    oneapi_release(resp->oh);
+    return;
+  }
+  LOG(resp->oh.verbose, "oneapi driver count: %d\n", resp->oh.num_drivers);
+  resp->oh.drivers = malloc(resp->oh.num_drivers * sizeof(zes_driver_handle_t));
+  resp->oh.num_devices = malloc(resp->oh.num_drivers * sizeof(uint32_t));
+  memset(&resp->oh.num_devices[0], 0, resp->oh.num_drivers * sizeof(uint32_t));
+  resp->oh.devices =
+      malloc(resp->oh.num_drivers * sizeof(zes_device_handle_t *));
+  ret = (*resp->oh.zesDriverGet)(&resp->oh.num_drivers, &resp->oh.drivers[0]);
+  if (ret != ZE_RESULT_SUCCESS) {
+    LOG(resp->oh.verbose, "zesDriverGet err: %x\n", ret);
+    snprintf(buf, buflen, "unable to get driver count: %x", ret);
     resp->err = strdup(buf);
+    oneapi_release(resp->oh);
+    return;
   }
 
-  (*resp->oh.zesDriverGet)(&resp->num_devices, NULL);
+  for (d = 0; d < resp->oh.num_drivers; d++) {
+    LOG(resp->oh.verbose, "calling zesDeviceGet count %d: %p\n", d, resp->oh.drivers[d]);
+    ret = (*resp->oh.zesDeviceGet)(resp->oh.drivers[d],
+                                   &resp->oh.num_devices[d], NULL);
+    if (ret != ZE_RESULT_SUCCESS) {
+      LOG(resp->oh.verbose, "zesDeviceGet err: %x\n", ret);
+      snprintf(buf, buflen, "unable to get device count: %x", ret);
+      resp->err = strdup(buf);
+      oneapi_release(resp->oh);
+      return;
+    }
+    resp->oh.devices[d] =
+        malloc(resp->oh.num_devices[d] * sizeof(zes_device_handle_t));
+    ret = (*resp->oh.zesDeviceGet)(
+        resp->oh.drivers[d], &resp->oh.num_devices[d], resp->oh.devices[d]);
+    if (ret != ZE_RESULT_SUCCESS) {
+      LOG(resp->oh.verbose, "zesDeviceGet err: %x\n", ret);
+      snprintf(buf, buflen, "unable to get device count: %x", ret);
+      resp->err = strdup(buf);
+      oneapi_release(resp->oh);
+      return;
+    }
+  }
 
   return;
 }
 
-void oneapi_check_vram(oneapi_handle_t h, mem_info_t *resp)
-{
+void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
+                       mem_info_t *resp) {
   ze_result_t ret;
   resp->err = NULL;
   uint64_t totalMem = 0;
@@ -88,127 +134,126 @@ void oneapi_check_vram(oneapi_handle_t h, mem_info_t *resp)
   char buf[buflen + 1];
   int i, d, m;
 
-  if (h.handle == NULL)
-  {
+  if (h.handle == NULL) {
     resp->err = strdup("Level-Zero handle not initialized");
     return;
   }
 
-  uint32_t driversCount = 0;
-  ret = (*h.zesDriverGet)(&driversCount, NULL);
-  if (ret != ZE_RESULT_SUCCESS)
-  {
-    snprintf(buf, buflen, "unable to get driver count: %d", ret);
-    resp->err = strdup(buf);
+  if (driver > h.num_drivers || device > h.num_devices[driver]) {
+    resp->err = strdup("driver of device index out of bounds");
     return;
   }
-  LOG(h.verbose, "discovered %d Level-Zero drivers\n", driversCount);
-
-  zes_driver_handle_t *allDrivers =
-      malloc(driversCount * sizeof(zes_driver_handle_t));
-  (*h.zesDriverGet)(&driversCount, allDrivers);
 
   resp->total = 0;
   resp->free = 0;
 
-  for (d = 0; d < driversCount; d++)
-  {
-    uint32_t deviceCount = 0;
-    ret = (*h.zesDeviceGet)(allDrivers[d], &deviceCount, NULL);
-    if (ret != ZE_RESULT_SUCCESS)
-    {
-      snprintf(buf, buflen, "unable to get device count: %d", ret);
+  zes_device_ext_properties_t ext_props;
+  ext_props.stype = ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES;
+  ext_props.pNext = NULL;
+
+  zes_device_properties_t props;
+  props.stype = ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES;
+  props.pNext = &ext_props;
+
+  ret = (*h.zesDeviceGetProperties)(h.devices[driver][device], &props);
+  if (ret != ZE_RESULT_SUCCESS) {
+    snprintf(buf, buflen, "unable to get device properties: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  snprintf(&resp->gpu_name[0], GPU_NAME_LEN, "%s", props.modelName);
+
+  // TODO this needs to map to ONEAPI_DEVICE_SELECTOR syntax
+  // (this is probably wrong...)
+  // TODO - the driver isn't included - what if there are multiple drivers?
+  snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", device);
+
+  if (h.verbose) {
+    // When in verbose mode, report more information about
+    // the card we discover.
+    LOG(h.verbose, "[%d:%d] oneAPI device name: %s\n", driver, device,
+        props.modelName);
+    LOG(h.verbose, "[%d:%d] oneAPI brand: %s\n", driver, device,
+        props.brandName);
+    LOG(h.verbose, "[%d:%d] oneAPI vendor: %s\n", driver, device,
+        props.vendorName);
+    LOG(h.verbose, "[%d:%d] oneAPI S/N: %s\n", driver, device,
+        props.serialNumber);
+    LOG(h.verbose, "[%d:%d] oneAPI board number: %s\n", driver, device,
+        props.boardNumber);
+  }
+
+  // TODO
+  // Compute Capability equivalent in resp->major, resp->minor, resp->patch
+
+  uint32_t memCount = 0;
+  ret = (*h.zesDeviceEnumMemoryModules)(h.devices[driver][device], &memCount,
+                                        NULL);
+  if (ret != ZE_RESULT_SUCCESS) {
+    snprintf(buf, buflen, "unable to enumerate Level-Zero memory modules: %x",
+             ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  LOG(h.verbose, "discovered %d Level-Zero memory modules\n", memCount);
+
+  zes_mem_handle_t *mems = malloc(memCount * sizeof(zes_mem_handle_t));
+  (*h.zesDeviceEnumMemoryModules)(h.devices[driver][device], &memCount, mems);
+
+  for (m = 0; m < memCount; m++) {
+    zes_mem_state_t state;
+    state.stype = ZES_STRUCTURE_TYPE_MEM_STATE;
+    state.pNext = NULL;
+    ret = (*h.zesMemoryGetState)(mems[m], &state);
+    if (ret != ZE_RESULT_SUCCESS) {
+      snprintf(buf, buflen, "unable to get memory state: %x", ret);
       resp->err = strdup(buf);
-      free(allDrivers);
+      free(mems);
       return;
     }
 
-    LOG(h.verbose, "discovered %d Level-Zero devices\n", deviceCount);
-
-    zes_device_handle_t *devices =
-        malloc(deviceCount * sizeof(zes_device_handle_t));
-    (*h.zesDeviceGet)(allDrivers[d], &deviceCount, devices);
-
-    for (i = 0; i < deviceCount; i++)
-    {
-      zes_device_ext_properties_t ext_props;
-      ext_props.stype = ZES_STRUCTURE_TYPE_DEVICE_EXT_PROPERTIES;
-      ext_props.pNext = NULL;
-
-      zes_device_properties_t props;
-      props.stype = ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES;
-      props.pNext = &ext_props;
-
-      ret = (*h.zesDeviceGetProperties)(devices[i], &props);
-      if (ret != ZE_RESULT_SUCCESS)
-      {
-        snprintf(buf, buflen, "unable to get device properties: %d", ret);
-        resp->err = strdup(buf);
-        free(allDrivers);
-        free(devices);
-        return;
-      }
-
-      if (h.verbose)
-      {
-        // When in verbose mode, report more information about
-        // the card we discover.
-        LOG(h.verbose, "[%d] oneAPI device name: %s\n", i,
-            props.modelName);
-        LOG(h.verbose, "[%d] oneAPI brand: %s\n", i,
-            props.brandName);
-        LOG(h.verbose, "[%d] oneAPI vendor: %s\n", i,
-            props.vendorName);
-        LOG(h.verbose, "[%d] oneAPI S/N: %s\n", i,
-            props.serialNumber);
-        LOG(h.verbose, "[%d] oneAPI board number: %s\n", i,
-            props.boardNumber);
-      }
-
-      uint32_t memCount = 0;
-      ret = (*h.zesDeviceEnumMemoryModules)(devices[i], &memCount, NULL);
-      if (ret != ZE_RESULT_SUCCESS)
-      {
-        snprintf(buf, buflen,
-                 "unable to enumerate Level-Zero memory modules: %d", ret);
-        resp->err = strdup(buf);
-        free(allDrivers);
-        free(devices);
-        return;
-      }
-
-      LOG(h.verbose, "discovered %d Level-Zero memory modules\n", memCount);
-
-      zes_mem_handle_t *mems = malloc(memCount * sizeof(zes_mem_handle_t));
-      (*h.zesDeviceEnumMemoryModules)(devices[i], &memCount, mems);
-
-      for (m = 0; m < memCount; m++)
-      {
-        zes_mem_state_t state;
-        state.stype = ZES_STRUCTURE_TYPE_MEM_STATE;
-        state.pNext = NULL;
-        ret = (*h.zesMemoryGetState)(mems[m], &state);
-        if (ret != ZE_RESULT_SUCCESS)
-        {
-          snprintf(buf, buflen, "unable to get memory state: %d", ret);
-          resp->err = strdup(buf);
-          free(allDrivers);
-          free(devices);
-          free(mems);
-          return;
-        }
-
-        resp->total += state.size;
-        resp->free += state.free;
-      }
+    resp->total += state.size;
+    resp->free += state.free;
+  }
 
-      free(mems);
-    }
+  free(mems);
+}
 
-    free(devices);
+void oneapi_release(oneapi_handle_t h) {
+  int d;
+  LOG(h.verbose, "releasing oneapi library\n");
+  for (d = 0; d < h.num_drivers; d++) {
+    if (h.devices != NULL && h.devices[d] != NULL) {
+      free(h.devices[d]);
+    }
+  }
+  if (h.devices != NULL) {
+    free(h.devices);
+    h.devices = NULL;
   }
+  if (h.num_devices != NULL) {
+    free(h.num_devices);
+    h.num_devices = NULL;
+  }
+  if (h.drivers != NULL) {
+    free(h.drivers);
+    h.drivers = NULL;
+  }
+  h.num_drivers = 0;
+  UNLOAD_LIBRARY(h.handle);
+  h.handle = NULL;
+}
 
-  free(allDrivers);
+int oneapi_get_device_count(oneapi_handle_t h, int driver) {
+  if (h.handle == NULL || h.num_devices == NULL) {
+    return 0;
+  }
+  if (driver > h.num_drivers) {
+    return 0;
+  }
+  return (int)h.num_devices[driver];
 }
 
 #endif // __APPLE__
diff --git a/gpu/gpu_info_oneapi.h b/gpu/gpu_info_oneapi.h
index 9db9fae08f4e08b3cdfb29ae950fe3a116614a44..97fcecd9c7c562854344dd3c8a316ebd035ac1bd 100644
--- a/gpu/gpu_info_oneapi.h
+++ b/gpu/gpu_info_oneapi.h
@@ -9,8 +9,7 @@
 #define ZE_BIT(_i) (1 << _i)
 
 // Just enough typedef's to dlopen/dlsym for memory information
-typedef enum ze_result_t
-{
+typedef enum ze_result_t {
   ZE_RESULT_SUCCESS = 0,
   // Other values omitted for now...
 } ze_result_t;
@@ -20,13 +19,11 @@ typedef struct _zes_driver_handle_t *zes_driver_handle_t;
 typedef struct _zes_device_handle_t *zes_device_handle_t;
 typedef struct _zes_mem_handle_t *zes_mem_handle_t;
 
-typedef enum _ze_structure_type_t
-{
+typedef enum _ze_structure_type_t {
   ZE_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
 } ze_structure_type_t;
 
-typedef enum _zes_structure_type_t
-{
+typedef enum _zes_structure_type_t {
   ZES_STRUCTURE_TYPE_DEVICE_PROPERTIES = 0x1,
   ZES_STRUCTURE_TYPE_MEM_PROPERTIES = 0xb,
   ZES_STRUCTURE_TYPE_MEM_STATE = 0x1e,
@@ -34,35 +31,29 @@ typedef enum _zes_structure_type_t
   ZES_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
 } zes_structure_type_t;
 
-typedef enum _zes_mem_type_t
-{
+typedef enum _zes_mem_type_t {
   ZES_MEM_TYPE_FORCE_UINT32 = 0x7fffffff
 } zes_mem_type_t;
 
-typedef enum _zes_mem_loc_t
-{
+typedef enum _zes_mem_loc_t {
   ZES_MEM_LOC_SYSTEM = 0,
   ZES_MEM_LOC_DEVICE = 1,
   ZES_MEM_LOC_FORCE_UINT32 = 0x7fffffff
 } zes_mem_loc_t;
 
-typedef enum _zes_mem_health_t
-{
+typedef enum _zes_mem_health_t {
   ZES_MEM_HEALTH_FORCE_UINT32 = 0x7fffffff
 } zes_mem_health_t;
 
-typedef struct _ze_device_uuid_t
-{
+typedef struct _ze_device_uuid_t {
   uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
 } ze_device_uuid_t;
 
-typedef struct _zes_uuid_t
-{
+typedef struct _zes_uuid_t {
   uint8_t id[ZE_MAX_DEVICE_UUID_SIZE];
 } zes_uuid_t;
 
-typedef enum _ze_device_type_t
-{
+typedef enum _ze_device_type_t {
   ZE_DEVICE_TYPE_GPU = 1,
   ZE_DEVICE_TYPE_CPU = 2,
   ZE_DEVICE_TYPE_FPGA = 3,
@@ -71,8 +62,7 @@ typedef enum _ze_device_type_t
   ZE_DEVICE_TYPE_FORCE_UINT32 = 0x7fffffff
 } ze_device_type_t;
 
-typedef enum _zes_device_type_t
-{
+typedef enum _zes_device_type_t {
   ZES_DEVICE_TYPE_GPU = 1,
   ZES_DEVICE_TYPE_CPU = 2,
   ZES_DEVICE_TYPE_FPGA = 3,
@@ -82,8 +72,7 @@ typedef enum _zes_device_type_t
 } zes_device_type_t;
 
 typedef uint32_t ze_device_property_flags_t;
-typedef enum _ze_device_property_flag_t
-{
+typedef enum _ze_device_property_flag_t {
   ZE_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
   ZE_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
   ZE_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
@@ -92,8 +81,7 @@ typedef enum _ze_device_property_flag_t
 } ze_device_property_flag_t;
 
 typedef uint32_t zes_device_property_flags_t;
-typedef enum _zes_device_property_flag_t
-{
+typedef enum _zes_device_property_flag_t {
   ZES_DEVICE_PROPERTY_FLAG_INTEGRATED = ZE_BIT(0),
   ZES_DEVICE_PROPERTY_FLAG_SUBDEVICE = ZE_BIT(1),
   ZES_DEVICE_PROPERTY_FLAG_ECC = ZE_BIT(2),
@@ -101,8 +89,7 @@ typedef enum _zes_device_property_flag_t
   ZES_DEVICE_PROPERTY_FLAG_FORCE_UINT32 = 0x7fffffff
 } zes_device_property_flag_t;
 
-typedef struct _ze_device_properties_t
-{
+typedef struct _ze_device_properties_t {
   ze_structure_type_t stype;
   void *pNext;
   ze_device_type_t type;
@@ -126,8 +113,7 @@ typedef struct _ze_device_properties_t
   char name[ZE_MAX_DEVICE_NAME];
 } ze_device_properties_t;
 
-typedef struct _zes_device_properties_t
-{
+typedef struct _zes_device_properties_t {
   zes_structure_type_t stype;
   void *pNext;
   ze_device_properties_t core;
@@ -140,8 +126,7 @@ typedef struct _zes_device_properties_t
   char driverVersion[ZES_STRING_PROPERTY_SIZE];
 } zes_device_properties_t;
 
-typedef struct _zes_device_ext_properties_t
-{
+typedef struct _zes_device_ext_properties_t {
   zes_structure_type_t stype;
   void *pNext;
   zes_uuid_t uuid;
@@ -149,8 +134,7 @@ typedef struct _zes_device_ext_properties_t
   zes_device_property_flags_t flags;
 } zes_device_ext_properties_t;
 
-typedef struct _zes_mem_properties_t
-{
+typedef struct _zes_mem_properties_t {
   zes_structure_type_t stype;
   void *pNext;
   zes_mem_type_t type;
@@ -162,8 +146,7 @@ typedef struct _zes_mem_properties_t
   int32_t numChannels;
 } zes_mem_properties_t;
 
-typedef struct _zes_mem_state_t
-{
+typedef struct _zes_mem_state_t {
   zes_structure_type_t stype;
   const void *pNext;
   zes_mem_health_t health;
@@ -171,10 +154,19 @@ typedef struct _zes_mem_state_t
   uint64_t size;
 } zes_mem_state_t;
 
-typedef struct oneapi_handle
-{
+typedef struct oneapi_handle {
   void *handle;
   uint16_t verbose;
+
+  uint32_t num_drivers;
+  zes_driver_handle_t *drivers;
+  uint32_t *num_devices;
+  zes_device_handle_t **devices;
+
+  // TODO Driver major, minor information
+  // int driver_major;
+  // int driver_minor;
+
   ze_result_t (*zesInit)(int);
   ze_result_t (*zesDriverGet)(uint32_t *pCount, zes_driver_handle_t *phDrivers);
   ze_result_t (*zesDeviceGet)(zes_driver_handle_t hDriver, uint32_t *pCount,
@@ -191,21 +183,21 @@ typedef struct oneapi_handle
 
 } oneapi_handle_t;
 
-typedef struct oneapi_init_resp
-{
+typedef struct oneapi_init_resp {
   char *err; // If err is non-null handle is invalid
-  int num_devices;
   oneapi_handle_t oh;
 } oneapi_init_resp_t;
 
-typedef struct oneapi_version_resp
-{
+typedef struct oneapi_version_resp {
   ze_result_t status;
   char *str; // Contains version or error string if status != 0
 } oneapi_version_resp_t;
 
 void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp);
-void oneapi_check_vram(oneapi_handle_t rh, mem_info_t *resp);
+void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
+                       mem_info_t *resp);
+void oneapi_release(oneapi_handle_t h);
+int oneapi_get_device_count(oneapi_handle_t h, int driver);
 
 #endif // __GPU_INFO_INTEL_H__
 #endif // __APPLE__
diff --git a/gpu/gpu_linux.go b/gpu/gpu_linux.go
new file mode 100644
index 0000000000000000000000000000000000000000..d6d2675c889f41b4578a77d8d927887b7ee58b87
--- /dev/null
+++ b/gpu/gpu_linux.go
@@ -0,0 +1,92 @@
+package gpu
+
+import (
+	"bufio"
+	"fmt"
+	"os"
+	"strings"
+
+	"github.com/ollama/ollama/format"
+)
+
+var CudartGlobs = []string{
+	"/usr/local/cuda/lib64/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
+	"/usr/lib/wsl/lib/libcudart.so*",
+	"/usr/lib/wsl/drivers/*/libcudart.so*",
+	"/opt/cuda/lib64/libcudart.so*",
+	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
+	"/usr/local/cuda/lib*/libcudart.so*",
+	"/usr/lib*/libcudart.so*",
+	"/usr/local/lib*/libcudart.so*",
+}
+
+var NvmlGlobs = []string{}
+
+var NvcudaGlobs = []string{
+	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
+	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
+	"/usr/lib/*-linux-gnu/libcuda.so*",
+	"/usr/lib/wsl/lib/libcuda.so*",
+	"/usr/lib/wsl/drivers/*/libcuda.so*",
+	"/opt/cuda/lib*/libcuda.so*",
+	"/usr/local/cuda/lib*/libcuda.so*",
+	"/usr/lib*/libcuda.so*",
+	"/usr/local/lib*/libcuda.so*",
+}
+
+var OneapiGlobs = []string{
+	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
+	"/usr/lib*/libze_intel_gpu.so*",
+}
+
+var (
+	CudartMgmtName = "libcudart.so*"
+	NvcudaMgmtName = "libcuda.so*"
+	NvmlMgmtName   = "" // not currently wired on linux
+	OneapiMgmtName = "libze_intel_gpu.so"
+)
+
+func GetCPUMem() (memInfo, error) {
+	var mem memInfo
+	var total, available, free, buffers, cached, freeSwap uint64
+	f, err := os.Open("/proc/meminfo")
+	if err != nil {
+		return mem, err
+	}
+	defer f.Close()
+	s := bufio.NewScanner(f)
+	for s.Scan() {
+		line := s.Text()
+		switch {
+		case strings.HasPrefix(line, "MemTotal:"):
+			_, err = fmt.Sscanf(line, "MemTotal:%d", &total)
+		case strings.HasPrefix(line, "MemAvailable:"):
+			_, err = fmt.Sscanf(line, "MemAvailable:%d", &available)
+		case strings.HasPrefix(line, "MemFree:"):
+			_, err = fmt.Sscanf(line, "MemFree:%d", &free)
+		case strings.HasPrefix(line, "Buffers:"):
+			_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
+		case strings.HasPrefix(line, "Cached:"):
+			_, err = fmt.Sscanf(line, "Cached:%d", &cached)
+		case strings.HasPrefix(line, "SwapFree:"):
+			_, err = fmt.Sscanf(line, "SwapFree:%d", &freeSwap)
+		default:
+			continue
+		}
+		if err != nil {
+			return mem, err
+		}
+	}
+	mem.TotalMemory = total * format.KibiByte
+	mem.FreeSwap = freeSwap * format.KibiByte
+	if available > 0 {
+		mem.FreeMemory = available * format.KibiByte
+	} else {
+		mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	}
+	return mem, nil
+}
diff --git a/gpu/gpu_test.go b/gpu/gpu_test.go
index a28cbe8c7fd49c8d6712536f2e2599998105564b..46d3201e1c8229b9d4473b165db4e7cd1d9dab47 100644
--- a/gpu/gpu_test.go
+++ b/gpu/gpu_test.go
@@ -5,11 +5,12 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func TestBasicGetGPUInfo(t *testing.T) {
 	info := GetGPUInfo()
-	assert.Greater(t, len(info), 0)
+	assert.NotEmpty(t, len(info))
 	assert.Contains(t, "cuda rocm cpu metal", info[0].Library)
 	if info[0].Library != "cpu" {
 		assert.Greater(t, info[0].TotalMemory, uint64(0))
@@ -19,7 +20,7 @@ func TestBasicGetGPUInfo(t *testing.T) {
 
 func TestCPUMemInfo(t *testing.T) {
 	info, err := GetCPUMem()
-	assert.NoError(t, err)
+	require.NoError(t, err)
 	switch runtime.GOOS {
 	case "darwin":
 		t.Skip("CPU memory not populated on darwin")
diff --git a/gpu/gpu_windows.go b/gpu/gpu_windows.go
new file mode 100644
index 0000000000000000000000000000000000000000..2ec72ba7b14f45073f97b1de941ee203302a90ce
--- /dev/null
+++ b/gpu/gpu_windows.go
@@ -0,0 +1,57 @@
+package gpu
+
+import (
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+type MEMORYSTATUSEX struct {
+	length               uint32
+	MemoryLoad           uint32
+	TotalPhys            uint64
+	AvailPhys            uint64
+	TotalPageFile        uint64
+	AvailPageFile        uint64
+	TotalVirtual         uint64
+	AvailVirtual         uint64
+	AvailExtendedVirtual uint64
+}
+
+var (
+	k32                      = syscall.NewLazyDLL("kernel32.dll")
+	globalMemoryStatusExProc = k32.NewProc("GlobalMemoryStatusEx")
+	sizeofMemoryStatusEx     = uint32(unsafe.Sizeof(MEMORYSTATUSEX{}))
+)
+
+var CudartGlobs = []string{
+	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
+}
+
+var NvmlGlobs = []string{
+	"c:\\Windows\\System32\\nvml.dll",
+}
+
+var NvcudaGlobs = []string{
+	"c:\\windows\\system*\\nvcuda.dll",
+}
+
+var OneapiGlobs = []string{
+	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
+}
+
+var (
+	CudartMgmtName = "cudart64_*.dll"
+	NvcudaMgmtName = "nvcuda.dll"
+	NvmlMgmtName   = "nvml.dll"
+	OneapiMgmtName = "ze_intel_gpu64.dll"
+)
+
+func GetCPUMem() (memInfo, error) {
+	memStatus := MEMORYSTATUSEX{length: sizeofMemoryStatusEx}
+	r1, _, err := globalMemoryStatusExProc.Call(uintptr(unsafe.Pointer(&memStatus)))
+	if r1 == 0 {
+		return memInfo{}, fmt.Errorf("GlobalMemoryStatusEx failed: %w", err)
+	}
+	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys, FreeSwap: memStatus.AvailPageFile}, nil
+}
diff --git a/gpu/types.go b/gpu/types.go
index af33b896003260985c695b07da31cfc3b6ca42bb..8d22b06bfa55f51717ddab941f16cea49636502e 100644
--- a/gpu/types.go
+++ b/gpu/types.go
@@ -10,6 +10,7 @@ import (
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
+	FreeSwap    uint64 `json:"free_swap,omitempty"`
 }
 
 // Beginning of an `ollama info` command
@@ -18,7 +19,7 @@ type GpuInfo struct {
 	Library string `json:"library,omitempty"`
 
 	// Optional variant to select (e.g. versions, cpu feature flags)
-	Variant string `json:"variant,omitempty"`
+	Variant CPUCapability `json:"variant"`
 
 	// MinimumMemory represents the minimum memory required to use the GPU
 	MinimumMemory uint64 `json:"-"`
@@ -26,6 +27,14 @@ type GpuInfo struct {
 	// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
 	DependencyPath string `json:"lib_path,omitempty"`
 
+	// Extra environment variables specific to the GPU as list of [key,value]
+	EnvWorkarounds [][2]string `json:"envs,omitempty"`
+
+	// Set to true if we can NOT reliably discover FreeMemory.  A value of true indicates
+	// the FreeMemory is best effort, and may over or under report actual memory usage
+	// False indicates FreeMemory can generally be trusted on this GPU
+	UnreliableFreeMemory bool
+
 	// GPU information
 	ID      string `json:"gpu_id"`  // string to use for selection of this specific GPU
 	Name    string `json:"name"`    // user friendly name if available
@@ -38,6 +47,31 @@ type GpuInfo struct {
 	// TODO other performance capability info to help in scheduling decisions
 }
 
+type CPUInfo struct {
+	GpuInfo
+}
+
+type CudaGPUInfo struct {
+	GpuInfo
+	OSOverhead uint64 // Memory overhead between the driver library and management library
+	index      int    //nolint:unused,nolintlint
+}
+type CudaGPUInfoList []CudaGPUInfo
+
+type RocmGPUInfo struct {
+	GpuInfo
+	usedFilepath string //nolint:unused,nolintlint
+	index        int    //nolint:unused,nolintlint
+}
+type RocmGPUInfoList []RocmGPUInfo
+
+type OneapiGPUInfo struct {
+	GpuInfo
+	driverIndex int //nolint:unused,nolintlint
+	gpuIndex    int //nolint:unused,nolintlint
+}
+type OneapiGPUInfoList []OneapiGPUInfo
+
 type GpuInfoList []GpuInfo
 
 // Split up the set of gpu info's by Library and variant
@@ -47,8 +81,8 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != "" {
-			requested += "_" + info.Variant
+		if info.Variant != CPUCapabilityNone {
+			requested += "_" + info.Variant.String()
 		}
 		for i, lib := range libs {
 			if lib == requested {
@@ -86,3 +120,26 @@ type ByFreeMemory []GpuInfo
 func (a ByFreeMemory) Len() int           { return len(a) }
 func (a ByFreeMemory) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 func (a ByFreeMemory) Less(i, j int) bool { return a[i].FreeMemory < a[j].FreeMemory }
+
+type CPUCapability uint32
+
+// Override at build time when building base GPU runners
+var GPURunnerCPUCapability = CPUCapabilityAVX
+
+const (
+	CPUCapabilityNone CPUCapability = iota
+	CPUCapabilityAVX
+	CPUCapabilityAVX2
+	// TODO AVX512
+)
+
+func (c CPUCapability) String() string {
+	switch c {
+	case CPUCapabilityAVX:
+		return "avx"
+	case CPUCapabilityAVX2:
+		return "avx2"
+	default:
+		return "no vector extensions"
+	}
+}
diff --git a/integration/basic_test.go b/integration/basic_test.go
index 6e632a1cefdb418ff564c3964185acec59ad3bae..8e35b5c5b0eff0804803fee854e3507709788876 100644
--- a/integration/basic_test.go
+++ b/integration/basic_test.go
@@ -45,14 +45,7 @@ func TestUnicodeModelDir(t *testing.T) {
 	defer os.RemoveAll(modelDir)
 	slog.Info("unicode", "OLLAMA_MODELS", modelDir)
 
-	oldModelsDir := os.Getenv("OLLAMA_MODELS")
-	if oldModelsDir == "" {
-		defer os.Unsetenv("OLLAMA_MODELS")
-	} else {
-		defer os.Setenv("OLLAMA_MODELS", oldModelsDir)
-	}
-	err = os.Setenv("OLLAMA_MODELS", modelDir)
-	require.NoError(t, err)
+	t.Setenv("OLLAMA_MODELS", modelDir)
 
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
diff --git a/integration/concurrency_test.go b/integration/concurrency_test.go
index f6bdb9d42d0a748192e4e466b89ab5519be87fc8..42e9d07490e3c139d4e7080d96f0d81554aec3b5 100644
--- a/integration/concurrency_test.go
+++ b/integration/concurrency_test.go
@@ -11,25 +11,29 @@ import (
 	"testing"
 	"time"
 
-	"github.com/ollama/ollama/api"
 	"github.com/stretchr/testify/require"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/format"
 )
 
 func TestMultiModelConcurrency(t *testing.T) {
 	var (
 		req = [2]api.GenerateRequest{
 			{
-				Model:  "orca-mini",
-				Prompt: "why is the ocean blue?",
-				Stream: &stream,
+				Model:     "orca-mini",
+				Prompt:    "why is the ocean blue?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:  "tinydolphin",
-				Prompt: "what is the origin of the us thanksgiving holiday?",
-				Stream: &stream,
+				Model:     "tinydolphin",
+				Prompt:    "what is the origin of the us thanksgiving holiday?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
@@ -37,43 +41,64 @@ func TestMultiModelConcurrency(t *testing.T) {
 			},
 		}
 		resp = [2][]string{
-			[]string{"sunlight"},
-			[]string{"england", "english", "massachusetts", "pilgrims"},
+			{"sunlight"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
 		}
 	)
 	var wg sync.WaitGroup
 	wg.Add(len(req))
-	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*240)
 	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for i := 0; i < len(req); i++ {
+		require.NoError(t, PullIfMissing(ctx, client, req[i].Model))
+	}
+
 	for i := 0; i < len(req); i++ {
 		go func(i int) {
 			defer wg.Done()
-			GenerateTestHelper(ctx, t, req[i], resp[i])
+			DoGenerate(ctx, t, client, req[i], resp[i], 60*time.Second, 10*time.Second)
 		}(i)
 	}
 	wg.Wait()
 }
 
 func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute) // GTX 750 2G card takes ~9 minutes
+	req, resp := GenerateRequests()
+	reqLimit := len(req)
+	iterLimit := 5
+
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err := strconv.ParseUint(s, 10, 64)
+		require.NoError(t, err)
+		// Don't hammer on small VRAM cards...
+		if maxVram < 4*format.GibiByte {
+			reqLimit = min(reqLimit, 2)
+			iterLimit = 2
+		}
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 9*time.Minute)
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
 
-	req, resp := GenerateRequests()
 	// Get the server running (if applicable) warm the model up with a single initial request
-	DoGenerate(ctx, t, client, req[0], resp[0], 60*time.Second, 5*time.Second)
+	DoGenerate(ctx, t, client, req[0], resp[0], 60*time.Second, 10*time.Second)
 
 	var wg sync.WaitGroup
-	wg.Add(len(req))
-	for i := 0; i < len(req); i++ {
+	wg.Add(reqLimit)
+	for i := 0; i < reqLimit; i++ {
 		go func(i int) {
 			defer wg.Done()
-			for j := 0; j < 5; j++ {
+			for j := 0; j < iterLimit; j++ {
 				slog.Info("Starting", "req", i, "iter", j)
-				// On slower GPUs it can take a while to process the 4 concurrent requests
+				// On slower GPUs it can take a while to process the concurrent requests
 				// so we allow a much longer initial timeout
-				DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 5*time.Second)
+				DoGenerate(ctx, t, client, req[i], resp[i], 120*time.Second, 20*time.Second)
 			}
 		}(i)
 	}
@@ -82,13 +107,16 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 
 // Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
 func TestMultiModelStress(t *testing.T) {
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
-	if vram == "" {
+	s := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
+	if s == "" {
 		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
 	}
-	max, err := strconv.ParseUint(vram, 10, 64)
-	require.NoError(t, err)
-	const MB = uint64(1024 * 1024)
+
+	maxVram, err := strconv.ParseUint(s, 10, 64)
+	if err != nil {
+		t.Fatal(err)
+	}
+
 	type model struct {
 		name string
 		size uint64 // Approximate amount of VRAM they typically use when fully loaded in VRAM
@@ -97,83 +125,82 @@ func TestMultiModelStress(t *testing.T) {
 	smallModels := []model{
 		{
 			name: "orca-mini",
-			size: 2992 * MB,
+			size: 2992 * format.MebiByte,
 		},
 		{
 			name: "phi",
-			size: 2616 * MB,
+			size: 2616 * format.MebiByte,
 		},
 		{
 			name: "gemma:2b",
-			size: 2364 * MB,
+			size: 2364 * format.MebiByte,
 		},
 		{
 			name: "stable-code:3b",
-			size: 2608 * MB,
+			size: 2608 * format.MebiByte,
 		},
 		{
 			name: "starcoder2:3b",
-			size: 2166 * MB,
+			size: 2166 * format.MebiByte,
 		},
 	}
 	mediumModels := []model{
 		{
 			name: "llama2",
-			size: 5118 * MB,
+			size: 5118 * format.MebiByte,
 		},
 		{
 			name: "mistral",
-			size: 4620 * MB,
+			size: 4620 * format.MebiByte,
 		},
 		{
 			name: "orca-mini:7b",
-			size: 5118 * MB,
+			size: 5118 * format.MebiByte,
 		},
 		{
 			name: "dolphin-mistral",
-			size: 4620 * MB,
+			size: 4620 * format.MebiByte,
 		},
 		{
 			name: "gemma:7b",
-			size: 5000 * MB,
+			size: 5000 * format.MebiByte,
+		},
+		{
+			name: "codellama:7b",
+			size: 5118 * format.MebiByte,
 		},
-		// TODO - uncomment this once #3565 is merged and this is rebased on it
-		// {
-		// 	name: "codellama:7b",
-		// 	size: 5118 * MB,
-		// },
 	}
 
 	// These seem to be too slow to be useful...
 	// largeModels := []model{
 	// 	{
 	// 		name: "llama2:13b",
-	// 		size: 7400 * MB,
+	// 		size: 7400 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "codellama:13b",
-	// 		size: 7400 * MB,
+	// 		size: 7400 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "orca-mini:13b",
-	// 		size: 7400 * MB,
+	// 		size: 7400 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "gemma:7b",
-	// 		size: 5000 * MB,
+	// 		size: 5000 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "starcoder2:15b",
-	// 		size: 9100 * MB,
+	// 		size: 9100 * format.MebiByte,
 	// 	},
 	// }
 
 	var chosenModels []model
 	switch {
-	case max < 10000*MB:
+	case maxVram < 10000*format.MebiByte:
 		slog.Info("selecting small models")
 		chosenModels = smallModels
-	// case max < 30000*MB:
+	// case maxVram < 30000*format.MebiByte:
 	default:
 		slog.Info("selecting medium models")
 		chosenModels = mediumModels
@@ -202,15 +229,15 @@ func TestMultiModelStress(t *testing.T) {
 	}
 
 	var wg sync.WaitGroup
-	consumed := uint64(256 * MB) // Assume some baseline usage
+	consumed := uint64(256 * format.MebiByte) // Assume some baseline usage
 	for i := 0; i < len(req); i++ {
 		// Always get at least 2 models, but dont' overshoot VRAM too much or we'll take too long
-		if i > 1 && consumed > max {
-			slog.Info("achieved target vram exhaustion", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
+		if i > 1 && consumed > maxVram {
+			slog.Info("achieved target vram exhaustion", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
 			break
 		}
 		consumed += chosenModels[i].size
-		slog.Info("target vram", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
+		slog.Info("target vram", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
 
 		wg.Add(1)
 		go func(i int) {
@@ -221,5 +248,23 @@ func TestMultiModelStress(t *testing.T) {
 			}
 		}(i)
 	}
+	go func() {
+		for {
+			time.Sleep(2 * time.Second)
+			select {
+			case <-ctx.Done():
+				return
+			default:
+				models, err := client.ListRunning(ctx)
+				if err != nil {
+					slog.Warn("failed to list running models", "error", err)
+					continue
+				}
+				for _, m := range models.Models {
+					slog.Info("loaded model snapshot", "model", m)
+				}
+			}
+		}
+	}()
 	wg.Wait()
 }
diff --git a/integration/context_test.go b/integration/context_test.go
index 08033125d75b880f0754ad27de5bb0458e004238..f1342e16c6c6e3055e2085699326e04ca352a186 100644
--- a/integration/context_test.go
+++ b/integration/context_test.go
@@ -11,7 +11,8 @@ import (
 )
 
 func TestContextExhaustion(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute) // TODO maybe shorter?
+	// Longer needed for small footprint GPUs
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
@@ -24,5 +25,10 @@ func TestContextExhaustion(t *testing.T) {
 			"num_ctx":     128,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"once", "upon", "lived"})
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("PullIfMissing failed: %v", err)
+	}
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
 }
diff --git a/integration/embed_test.go b/integration/embed_test.go
new file mode 100644
index 0000000000000000000000000000000000000000..10333d5dfa9992bc6a2bf05a715f61d8e016171f
--- /dev/null
+++ b/integration/embed_test.go
@@ -0,0 +1,209 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"math"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func floatsEqual32(a, b float32) bool {
+	return math.Abs(float64(a-b)) <= 1e-4
+}
+
+func floatsEqual64(a, b float64) bool {
+	return math.Abs(a-b) <= 1e-4
+}
+
+func TestAllMiniLMEmbeddings(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbeddingRequest{
+		Model:  "all-minilm",
+		Prompt: "why is the sky blue?",
+	}
+
+	res, err := embeddingTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embedding) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embedding))
+	}
+
+	if !floatsEqual64(res.Embedding[0], 0.06642947345972061) {
+		t.Fatalf("expected 0.06642947345972061, got %.16f", res.Embedding[0])
+	}
+}
+
+func TestAllMiniLMEmbed(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbedRequest{
+		Model: "all-minilm",
+		Input: "why is the sky blue?",
+	}
+
+	res, err := embedTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embeddings) != 1 {
+		t.Fatalf("expected 1 embedding, got %d", len(res.Embeddings))
+	}
+
+	if len(res.Embeddings[0]) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
+	}
+
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) {
+		t.Fatalf("expected 0.010071031, got %.8f", res.Embeddings[0][0])
+	}
+
+	if res.PromptEvalCount != 8 {
+		t.Fatalf("expected 8 prompt tokens, got %d", res.PromptEvalCount)
+	}
+}
+
+func TestAllMiniLMBatchEmbed(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbedRequest{
+		Model: "all-minilm",
+		Input: []string{"why is the sky blue?", "why is the grass green?"},
+	}
+
+	res, err := embedTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embeddings) != 2 {
+		t.Fatalf("expected 2 embeddings, got %d", len(res.Embeddings))
+	}
+
+	if len(res.Embeddings[0]) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
+	}
+
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) || !floatsEqual32(res.Embeddings[1][0], -0.009802706) {
+		t.Fatalf("expected 0.010071031 and -0.009802706, got %.8f and %.8f", res.Embeddings[0][0], res.Embeddings[1][0])
+	}
+
+	if res.PromptEvalCount != 16 {
+		t.Fatalf("expected 16 prompt tokens, got %d", res.PromptEvalCount)
+	}
+}
+
+func TestAllMiniLMEmbedTruncate(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	truncTrue, truncFalse := true, false
+
+	type testReq struct {
+		Name    string
+		Request api.EmbedRequest
+	}
+
+	reqs := []testReq{
+		{
+			Name: "Target Truncation",
+			Request: api.EmbedRequest{
+				Model: "all-minilm",
+				Input: "why",
+			},
+		},
+		{
+			Name: "Default Truncate",
+			Request: api.EmbedRequest{
+				Model:   "all-minilm",
+				Input:   "why is the sky blue?",
+				Options: map[string]any{"num_ctx": 1},
+			},
+		},
+		{
+			Name: "Explicit Truncate",
+			Request: api.EmbedRequest{
+				Model:    "all-minilm",
+				Input:    "why is the sky blue?",
+				Truncate: &truncTrue,
+				Options:  map[string]any{"num_ctx": 1},
+			},
+		},
+	}
+
+	res := make(map[string]*api.EmbedResponse)
+
+	for _, req := range reqs {
+		response, err := embedTestHelper(ctx, t, req.Request)
+		if err != nil {
+			t.Fatalf("error: %v", err)
+		}
+		res[req.Name] = response
+	}
+
+	if res["Target Truncation"].Embeddings[0][0] != res["Default Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request to truncate correctly")
+	}
+
+	if res["Default Truncate"].Embeddings[0][0] != res["Explicit Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request and truncate true request to be the same")
+	}
+
+	// check that truncate set to false returns an error if context length is exceeded
+	_, err := embedTestHelper(ctx, t, api.EmbedRequest{
+		Model:    "all-minilm",
+		Input:    "why is the sky blue?",
+		Truncate: &truncFalse,
+		Options:  map[string]any{"num_ctx": 1},
+	})
+
+	if err == nil {
+		t.Fatal("expected error, got nil")
+	}
+}
+
+func embeddingTestHelper(ctx context.Context, t *testing.T, req api.EmbeddingRequest) (*api.EmbeddingResponse, error) {
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+	}
+
+	response, err := client.Embeddings(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func embedTestHelper(ctx context.Context, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+	}
+
+	response, err := client.Embed(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
diff --git a/integration/llm_image_test.go b/integration/llm_image_test.go
index 77319aefd291775a323b3052b8e13c4b96bf00ab..d0c861cc07e74e0ddc90e386156a581bcc644816 100644
--- a/integration/llm_image_test.go
+++ b/integration/llm_image_test.go
@@ -32,7 +32,11 @@ func TestIntegrationMultimodal(t *testing.T) {
 	resp := "the ollam"
 	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
-	GenerateTestHelper(ctx, t, req, []string{resp})
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
+	// llava models on CPU can be quite slow to start,
+	DoGenerate(ctx, t, client, req, []string{resp}, 120*time.Second, 30*time.Second)
 }
 
 const imageEncoding = `iVBORw0KGgoAAAANSUhEUgAAANIAAAB4CAYAAACHHqzKAAAAAXNSR0IArs4c6QAAAIRlWElmTU0AKgAAAAgABQESAAMAAAABAAEAAAEaAAUAAAABAAAASgEb
diff --git a/integration/llm_test.go b/integration/llm_test.go
index 4952b072689d6b87542992b4532be5677968fdfc..398e0a03a9bbf0f668273a77774438ef94f294d1 100644
--- a/integration/llm_test.go
+++ b/integration/llm_test.go
@@ -35,8 +35,8 @@ var (
 		},
 	}
 	resp = [2][]string{
-		[]string{"sunlight"},
-		[]string{"england", "english", "massachusetts", "pilgrims"},
+		{"sunlight"},
+		{"england", "english", "massachusetts", "pilgrims"},
 	}
 )
 
diff --git a/integration/max_queue_test.go b/integration/max_queue_test.go
index dfa5eae0cae5127a69d468dbe0efc3f73ad55095..ec9e085a5ad851796dd3975475b04b2d93311b34 100644
--- a/integration/max_queue_test.go
+++ b/integration/max_queue_test.go
@@ -5,7 +5,6 @@ package integration
 import (
 	"context"
 	"errors"
-	"fmt"
 	"log/slog"
 	"os"
 	"strconv"
@@ -14,8 +13,10 @@ import (
 	"testing"
 	"time"
 
-	"github.com/ollama/ollama/api"
 	"github.com/stretchr/testify/require"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
 )
 
 func TestMaxQueue(t *testing.T) {
@@ -27,13 +28,10 @@ func TestMaxQueue(t *testing.T) {
 	// Note: This test can be quite slow when running in CPU mode, so keep the threadCount low unless your on GPU
 	// Also note that by default Darwin can't sustain > ~128 connections without adjusting limits
 	threadCount := 32
-	mq := os.Getenv("OLLAMA_MAX_QUEUE")
-	if mq != "" {
-		var err error
-		threadCount, err = strconv.Atoi(mq)
-		require.NoError(t, err)
+	if maxQueue := envconfig.MaxQueue(); maxQueue != 0 {
+		threadCount = int(maxQueue)
 	} else {
-		os.Setenv("OLLAMA_MAX_QUEUE", fmt.Sprintf("%d", threadCount))
+		t.Setenv("OLLAMA_MAX_QUEUE", strconv.Itoa(threadCount))
 	}
 
 	req := api.GenerateRequest{
diff --git a/integration/utils_test.go b/integration/utils_test.go
index c6f19e98228388fa9a1a5f6352d9115819ec6bd6..a60109958bbb7dea3504c281bcd5bfdf8b759173 100644
--- a/integration/utils_test.go
+++ b/integration/utils_test.go
@@ -140,7 +140,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 
 	showCtx, cancel := context.WithDeadlineCause(
 		ctx,
-		time.Now().Add(5*time.Second),
+		time.Now().Add(10*time.Second),
 		fmt.Errorf("show for existing model %s took too long", modelName),
 	)
 	defer cancel()
@@ -162,7 +162,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 	fn := func(resp api.ProgressResponse) error {
 		// fmt.Print(".")
 		if !stallTimer.Reset(stallDuration) {
-			return fmt.Errorf("stall was detected, aborting status reporting")
+			return errors.New("stall was detected, aborting status reporting")
 		}
 		return nil
 	}
@@ -180,7 +180,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 
 	select {
 	case <-stallTimer.C:
-		return fmt.Errorf("download stalled")
+		return errors.New("download stalled")
 	case <-done:
 		return pullError
 	}
@@ -243,7 +243,7 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 		// fmt.Print(".")
 		buf.Write([]byte(response.Response))
 		if !stallTimer.Reset(streamTimeout) {
-			return fmt.Errorf("stall was detected while streaming response, aborting")
+			return errors.New("stall was detected while streaming response, aborting")
 		}
 		return nil
 	}
@@ -287,41 +287,46 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 	return []api.GenerateRequest{
 			{
-				Model:  "orca-mini",
-				Prompt: "why is the ocean blue?",
-				Stream: &stream,
+				Model:     "orca-mini",
+				Prompt:    "why is the ocean blue?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:  "orca-mini",
-				Prompt: "why is the color of dirt brown?",
-				Stream: &stream,
+				Model:     "orca-mini",
+				Prompt:    "why is the color of dirt brown?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:  "orca-mini",
-				Prompt: "what is the origin of the us thanksgiving holiday?",
-				Stream: &stream,
+				Model:     "orca-mini",
+				Prompt:    "what is the origin of the us thanksgiving holiday?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:  "orca-mini",
-				Prompt: "what is the origin of independence day?",
-				Stream: &stream,
+				Model:     "orca-mini",
+				Prompt:    "what is the origin of independence day?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:  "orca-mini",
-				Prompt: "what is the composition of air?",
-				Stream: &stream,
+				Model:     "orca-mini",
+				Prompt:    "what is the composition of air?",
+				Stream:    &stream,
+				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
@@ -329,10 +334,10 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			},
 		},
 		[][]string{
-			[]string{"sunlight"},
-			[]string{"soil", "organic", "earth", "black", "tan"},
-			[]string{"england", "english", "massachusetts", "pilgrims"},
-			[]string{"fourth", "july", "declaration", "independence"},
-			[]string{"nitrogen", "oxygen", "carbon", "dioxide"},
+			{"sunlight"},
+			{"soil", "organic", "earth", "black", "tan"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
+			{"fourth", "july", "declaration", "independence"},
+			{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
 }
diff --git a/llm/ext_server/CMakeLists.txt b/llm/ext_server/CMakeLists.txt
index db7d52dcc05fe901cac1909a083aa7fa961d8b72..bfc97c63d99f2561ddd25908d8ce0f2e33a11618 100644
--- a/llm/ext_server/CMakeLists.txt
+++ b/llm/ext_server/CMakeLists.txt
@@ -1,4 +1,3 @@
-
 set(TARGET ollama_llama_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
@@ -7,7 +6,7 @@ install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$
 )
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
diff --git a/llm/ext_server/server.cpp b/llm/ext_server/server.cpp
index 8a0dffeaa5bad9829465922b00cff87a2886760c..5717c17a905d8cdaed67e4b5df6558faeca04bb1 100644
--- a/llm/ext_server/server.cpp
+++ b/llm/ext_server/server.cpp
@@ -41,8 +41,10 @@
 
 #if defined(_WIN32)
 #include 
+#include 
 #endif
 
+#include 
 #include 
 #include 
 #include 
@@ -56,7 +58,6 @@ struct server_params {
     std::string hostname = "127.0.0.1";
     std::vector api_keys;
     std::string public_path = "examples/server/public";
-    std::string chat_template = "";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
@@ -359,7 +360,6 @@ struct llama_server_context
 
     // slots / clients
     std::vector slots;
-    json default_generation_settings_for_props;
 
     llama_server_queue    queue_tasks;
     llama_server_response queue_results;
@@ -403,7 +403,9 @@ struct llama_server_context
             }
         }
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        auto init_result = llama_init_from_gpt_params(params);
+        model = init_result.model;
+        ctx = init_result.context;
         if (model == nullptr)
         {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -428,16 +430,6 @@ struct llama_server_context
         return true;
     }
 
-    void validate_model_chat_template(server_params & sparams) {
-        llama_chat_message chat[] = {{"user", "test"}};
-        std::vector buf(1);
-        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
-        if (res < 0) {
-            LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
-            sparams.chat_template = "chatml";
-        }
-    }
-
     void initialize() {
         // create slots
         all_slots_are_idle = true;
@@ -483,9 +475,6 @@ struct llama_server_context
             slots.push_back(slot);
         }
 
-        default_generation_settings_for_props = get_formated_generation(slots.front());
-        default_generation_settings_for_props["seed"] = -1;
-
         batch = llama_batch_init(n_ctx, 0, params.n_parallel);
     }
 
@@ -584,7 +573,7 @@ struct llama_server_context
         slot->sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
         slot->sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
         slot->params.n_keep             = json_value(data, "n_keep",            slot->params.n_keep);
-        slot->params.seed               = json_value(data, "seed",              default_params.seed);
+        slot->sparams.seed              = json_value(data, "seed",              default_params.seed);
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
         slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
         slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
@@ -811,7 +800,6 @@ struct llama_server_context
             llama_sampling_free(slot->ctx_sampling);
         }
         slot->ctx_sampling = llama_sampling_init(slot->sparams);
-        llama_set_rng_seed(ctx, slot->params.seed);
         slot->command = LOAD_PROMPT;
 
         all_slots_are_idle = false;
@@ -835,7 +823,7 @@ struct llama_server_context
         system_tokens.clear();
 
         if (!system_prompt.empty()) {
-            system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
+            system_tokens = ::llama_tokenize(ctx, system_prompt, true);
 
             llama_batch_clear(batch);
 
@@ -1398,12 +1386,50 @@ struct llama_server_context
         }
     }
 
+    std::string common_prefix(const std::string& str1, const std::string& str2) {
+        auto mismatch_pair = std::mismatch(str1.begin(), str1.end(), str2.begin());
+        return std::string(str1.begin(), mismatch_pair.first);
+    }
+
+    // Find the slot that has the greatest common prefix
+    server_slot *prefix_slot(const json &prompt) {
+        if (!prompt.is_string()) {
+            return nullptr;
+        }
+
+        std::string prompt_str = prompt.get();
+        server_slot *slot = nullptr;
+        size_t longest = 0;
+
+        for (server_slot &s : slots) {
+            if (s.available() && s.prompt.is_string()) {
+                std::string s_prompt = s.prompt.get();
+                std::string prefix = common_prefix(s_prompt, prompt_str);
+
+                if (prefix.size() > longest) {
+                    slot = &s;
+                    longest = prefix.size();
+                }
+            }
+        }
+
+        if (!slot) {
+            return get_slot(-1);
+        }
+
+        LOG_DEBUG("slot with common prefix found", {{
+            "slot_id", slot->id,
+            "characters", longest
+        }});
+        return slot;
+    }
+
     void process_single_task(task_server& task)
     {
         switch (task.type)
         {
             case TASK_TYPE_COMPLETION: {
-                server_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
+                server_slot *slot = prefix_slot(task.data["prompt"]);
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later
@@ -1656,7 +1682,7 @@ struct llama_server_context
                     slot.t_start_process_prompt = ggml_time_us();
                     slot.t_start_genereration = 0;
 
-                    prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token);  // add BOS if there isn't system prompt
+                    prompt_tokens = tokenize(slot.prompt, system_prompt.empty());  // add BOS if there isn't system prompt
 
                     slot.n_prompt_tokens = prompt_tokens.size();
 
@@ -1670,22 +1696,23 @@ struct llama_server_context
                     if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx)
                     {
                         const int n_left = slot.n_ctx - slot.params.n_keep;
-                        const int n_block_size = n_left / 2;
-                        const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
+                        const int n_shift = n_left / 2;
+                        const int n_erase = slot.n_prompt_tokens - slot.params.n_keep - n_shift;
 
                         std::vector new_tokens(
                             prompt_tokens.begin(),
                             prompt_tokens.begin() + slot.params.n_keep);
                         new_tokens.insert(
                             new_tokens.end(),
-                            prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size,
+                            prompt_tokens.begin() + slot.params.n_keep + n_erase,
                             prompt_tokens.end());
 
-                        LOG_VERBOSE("input truncated", {
-                            {"n_ctx",      slot.n_ctx},
-                            {"n_keep",     slot.params.n_keep},
-                            {"n_left",     n_left},
-                            {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                        LOG_INFO("input truncated", {
+                            {"n_ctx",        slot.n_ctx},
+                            {"n_keep",       slot.params.n_keep},
+                            {"n_left",       n_left},
+                            {"n_shift",      n_shift},
+                            {"n_erase",      n_erase},
                         });
                         slot.truncated = true;
                         prompt_tokens = new_tokens;
@@ -1720,7 +1747,7 @@ struct llama_server_context
                             slot.n_past -= 1;
                         }
 
-                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens - slot.n_past;
+                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens;
 
                         if (slot.ga_n != 1)
                         {
@@ -2340,9 +2367,9 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-#ifndef GGML_USE_CUBLAS
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUBLAS
+#ifndef GGML_USE_CUDA
+            fprintf(stderr, "warning: llama.cpp was compiled without CUDA. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUDA
         }
         else if (arg == "--tensor-split" || arg == "-ts")
         {
@@ -2351,7 +2378,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL)
+#if defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)
             std::string arg_next = argv[i];
 
             // split string by , and /
@@ -2372,8 +2399,8 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 }
             }
 #else
-            LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {});
-#endif // GGML_USE_CUBLAS
+            LOG_WARNING("llama.cpp was compiled without CUDA. It is not possible to set a tensor split.\n", {});
+#endif // GGML_USE_CUDA
         }
         else if (arg == "--main-gpu" || arg == "-mg")
         {
@@ -2382,7 +2409,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL)
+#if defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)
             params.main_gpu = std::stoi(argv[i]);
 #else
             LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {});
@@ -2395,7 +2422,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.emplace_back(argv[i], 1.0f);
+            params.lora_adapters.push_back({
+                std::string(argv[i]),
+                1.0,
+            });
             params.use_mmap = false;
         }
         else if (arg == "--lora-scaled")
@@ -2411,18 +2441,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
+            params.lora_adapters.push_back({
+                lora_adapter,
+                std::stof(argv[i])
+            });
             params.use_mmap = false;
         }
-        else if (arg == "--lora-base")
-        {
-            if (++i >= argc)
-            {
-                invalid_param = true;
-                break;
-            }
-            params.lora_base = argv[i];
-        }
         else if (arg == "-v" || arg == "--verbose")
         {
             server_verbose = true;
@@ -2540,7 +2564,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            sparams.chat_template = argv[i];
         }
         else if (arg == "--override-kv")
         {
@@ -2715,6 +2738,9 @@ int wmain(int argc, wchar_t **wargv) {
     for (int i = 0; i < argc; ++i) {
         argv[i] = wchar_to_char(wargv[i]);
     }
+
+    // Adjust error mode to avoid error dialog after we start.
+    SetErrorMode(SEM_FAILCRITICALERRORS);
 #else
 int main(int argc, char **argv) {
 #endif
@@ -3013,11 +3039,6 @@ int main(int argc, char **argv) {
     }
     const auto model_meta = llama.model_meta();
 
-    if (sparams.chat_template.empty()) { // custom chat template is not supplied
-        // check if the template comes with the model is supported by us
-        llama.validate_model_chat_template(sparams);
-    }
-
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation
@@ -3171,19 +3192,10 @@ int main(int argc, char **argv) {
                     prompt = "";
                 }
 
-                json image_data;
-                if (body.count("image_data") != 0) {
-                    image_data = body["image_data"];
-                }
-                else
-                {
-                    image_data = "";
-                }
-
                 // create and queue the task
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);
-                llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, true, -1);
+                llama.request_completion(task_id, {{"prompt", prompt}}, true, -1);
 
                 // get the result
                 task_result result = llama.queue_results.recv(task_id);
diff --git a/llm/generate/gen_common.sh b/llm/generate/gen_common.sh
index 15934f2864a20ff0990dc19eb854c0a089adf4d5..da1b068829b238b85ca367c29157637e12dfc7d4 100644
--- a/llm/generate/gen_common.sh
+++ b/llm/generate/gen_common.sh
@@ -1,5 +1,4 @@
 # common logic across linux and darwin
-# 
 
 init_vars() {
     case "${GOARCH}" in
@@ -9,13 +8,12 @@ init_vars() {
     "arm64")
         ARCH="arm64"
         ;;
-    *)  
+    *)
         ARCH=$(uname -m | sed -e "s/aarch64/arm64/g")
     esac
 
     LLAMACPP_DIR=../llama.cpp
     CMAKE_DEFS=""
-    # 这里和llama.cpp不一样
     CMAKE_TARGETS="--target ollama_llama_server"
     if echo "${CGO_CFLAGS}" | grep -- '-g' >/dev/null; then
         CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on ${CMAKE_DEFS}"
@@ -56,8 +54,8 @@ git_module_setup() {
         echo "Cleaning up old submodule"
         rm -rf ${LLAMACPP_DIR}
     fi
-    # git submodule init
-    # git submodule update --force ${LLAMACPP_DIR}
+    git submodule init
+    git submodule update --force ${LLAMACPP_DIR}
 
 }
 
diff --git a/llm/generate/gen_darwin.sh b/llm/generate/gen_darwin.sh
index 0baf86ffc24d9e3b08122322384d92944b1e52dc..6c0b62cb7f2ad9d7a6fdf38d4ee0430a6d9bca15 100644
--- a/llm/generate/gen_darwin.sh
+++ b/llm/generate/gen_darwin.sh
@@ -18,16 +18,16 @@ sign() {
     fi
 }
 
-COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on"
+COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
 
 case "${GOARCH}" in
 "amd64")
-    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
+    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DGGML_METAL=off -DGGML_NATIVE=off"
 
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_BLAS=off -DGGML_ACCELERATE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
@@ -37,7 +37,7 @@ case "${GOARCH}" in
         # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/cpu"
         echo "Building LCD CPU"
         build
@@ -49,7 +49,7 @@ case "${GOARCH}" in
         # Approximately 400% faster than LCD on same CPU
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
         echo "Building AVX CPU"
         build
@@ -61,7 +61,7 @@ case "${GOARCH}" in
         # Approximately 10% faster than AVX on same CPU
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
         echo "Building AVX2 CPU"
         EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
@@ -75,14 +75,14 @@ case "${GOARCH}" in
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
 
     if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
         init_vars
-        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/metal"
         EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
         build
diff --git a/llm/generate/gen_linux.sh b/llm/generate/gen_linux.sh
index 5e9b6ddcc280e61af125cf4614c0606a708a635c..124dbea9149fb5fb16010caf4f3e98474dc582f7 100644
--- a/llm/generate/gen_linux.sh
+++ b/llm/generate/gen_linux.sh
@@ -26,8 +26,6 @@ amdGPUs() {
         "gfx908:xnack-"
         "gfx90a:xnack+"
         "gfx90a:xnack-"
-        "gfx906"
-        "gfx928"
         "gfx940"
         "gfx941"
         "gfx942"
@@ -53,7 +51,7 @@ if [ -z "${CUDACXX}" ]; then
         export CUDACXX=$(command -v nvcc)
     fi
 fi
-COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
+COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
 source $(dirname $0)/gen_common.sh
 init_vars
 # git_module_setup
@@ -66,7 +64,7 @@ if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ];
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DGGML_NATIVE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}_static"
     echo "Building static library"
     build
@@ -79,29 +77,29 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
         init_vars
         echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
         BUILD_DIR="../build/linux/${ARCH}/cpu"
         echo "Building custom CPU"
         build
         compress
     else
         # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
-        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
-        # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+        # -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
         # Note: the following seem to yield slower results than AVX2 - ymmv
-        # -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
-        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
-        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
+        # -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
+        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
 
-        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off"
+        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
         if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
             #
             # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
             #
             init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
             BUILD_DIR="../build/linux/${ARCH}/cpu"
             echo "Building LCD CPU"
             build
@@ -118,7 +116,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 400% faster than LCD on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
                 echo "Building AVX CPU"
                 build
@@ -131,7 +129,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 10% faster than AVX on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
                 echo "Building AVX2 CPU"
                 build
@@ -172,15 +170,15 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
         #
         # CUDA compute < 6.0 lacks proper FP16 support on ARM.
         # Disabling has minimal performance effect while maintaining compatibility.
-        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
+        ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
     fi
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
         echo "Building custom CUDA GPU"
     else
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
     fi
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
@@ -213,12 +211,12 @@ if [ -z "${ONEAPI_ROOT}" ]; then
     ONEAPI_ROOT=/opt/intel/oneapi
 fi
 
-if [ -d "${ONEAPI_ROOT}" ]; then
+if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
     echo "OneAPI libraries detected - building dynamic OneAPI library"
     init_vars
     source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
     CC=icx
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
     BUILD_DIR="../build/linux/${ARCH}/oneapi"
     EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
     DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@@ -256,9 +254,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
     fi
     init_vars
-    echo $(amdGPUs)
-
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DLLAMA_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
@@ -266,7 +262,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         echo "Building custom ROCM GPU"
     fi
     BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
-    EXTRA_LIBS="-L${ROCM_PATH}/lib -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
+    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
     build
 
     # Record the ROCM dependencies
@@ -276,12 +272,12 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
     done
     # bomb out if for some reason we didn't get a few deps
-    #if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
-    #    cat "${BUILD_DIR}/bin/deps.txt"
-    #    echo "ERROR: deps file short"
-    #    exit 1
-    #fi
-    # compress
+    # if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
+    #     cat "${BUILD_DIR}/bin/deps.txt"
+    #     echo "ERROR: deps file short"
+    #     exit 1
+    # fi
+    compress
 fi
 
 cleanup
diff --git a/llm/generate/gen_windows.ps1 b/llm/generate/gen_windows.ps1
index 553c358b7c2fe9578e4ffc1a6c957fe8eb669d46..d8bce92d6fe2715a049048138eee8055117ca45d 100644
--- a/llm/generate/gen_windows.ps1
+++ b/llm/generate/gen_windows.ps1
@@ -6,18 +6,9 @@ function amdGPUs {
     if ($env:AMDGPU_TARGETS) {
         return $env:AMDGPU_TARGETS
     }
-    # TODO - load from some common data file for linux + windows build consistency
+    # Current supported rocblas list from ROCm v6.1.2 on windows
+    # https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html#windows-supported-gpus
     $GPU_LIST = @(
-        "gfx900"
-        "gfx906:xnack-"
-        "gfx908:xnack-"
-        "gfx90a:xnack+"
-        "gfx90a:xnack-"
-        "gfx940"
-        "gfx941"
-        "gfx942"
-        "gfx1010"
-        "gfx1012"
         "gfx1030"
         "gfx1100"
         "gfx1101"
@@ -39,7 +30,8 @@ function init_vars {
     }
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
-        "-DLLAMA_NATIVE=off"
+        "-DGGML_NATIVE=off",
+        "-DGGML_OPENMP=off"
         )
     $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
     $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
@@ -122,8 +114,13 @@ function build {
     & cmake --version
     & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ })"
-    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ })
+    if ($cmakeDefs -contains "-G") {
+        $extra=@("-j8")
+    } else {
+        $extra= @("--", "/p:CL_MPcount=8")
+    }
+    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ }) $extra"
+    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ }) $extra
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     # Rearrange output to be consistent between different generators
     if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
@@ -176,9 +173,9 @@ function cleanup {
 }
 
 
-# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
 
 function build_static() {
@@ -198,12 +195,13 @@ function build_static() {
             "-DCMAKE_C_COMPILER=gcc.exe",
             "-DCMAKE_CXX_COMPILER=g++.exe",
             "-DBUILD_SHARED_LIBS=off",
-            "-DLLAMA_NATIVE=off",
-            "-DLLAMA_AVX=off",
-            "-DLLAMA_AVX2=off",
-            "-DLLAMA_AVX512=off",
-            "-DLLAMA_F16C=off",
-            "-DLLAMA_FMA=off")
+            "-DGGML_NATIVE=off",
+            "-DGGML_AVX=off",
+            "-DGGML_AVX2=off",
+            "-DGGML_AVX512=off",
+            "-DGGML_F16C=off",
+            "-DGGML_FMA=off",
+            "-DGGML_OPENMP=off")
         $script:buildDir="../build/windows/${script:ARCH}_static"
         write-host "Building static library"
         build
@@ -217,7 +215,7 @@ function build_cpu($gen_arch) {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
         # remaining llama.cpp builds use MSVC 
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu"
         $script:distDir="$script:DIST_BASE\cpu"
         write-host "Building LCD CPU"
@@ -232,7 +230,7 @@ function build_cpu($gen_arch) {
 function build_cpu_avx() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
         $script:distDir="$script:DIST_BASE\cpu_avx"
         write-host "Building AVX CPU"
@@ -247,7 +245,7 @@ function build_cpu_avx() {
 function build_cpu_avx2() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=on", "-DGGML_AVX512=off", "-DGGML_FMA=on", "-DGGML_F16C=on") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
         $script:distDir="$script:DIST_BASE\cpu_avx2"
         write-host "Building AVX2 CPU"
@@ -270,7 +268,15 @@ function build_cuda() {
         init_vars
         $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
         $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
-        $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUDA=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
+        $script:cmakeDefs += @(
+            "-A", "x64",
+            "-DGGML_CUDA=ON",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
+            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
+            "-DCMAKE_CUDA_FLAGS=-t8",
+            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
+            )
         if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
             write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
             $script:cmakeDefs +=@("${env:OLLAMA_CUSTOM_CUDA_DEFS}")
@@ -280,17 +286,19 @@ function build_cuda() {
         sign
         install
 
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\" -ea 0 > $null
+        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
     } else {
         write-host "Skipping CUDA generation step"
     }
 }
 
 function build_oneapi() {
-  if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
+  if ((-not "${env:OLLAMA_SKIP_ONEAPI_GENERATE}") -and ("${env:ONEAPI_ROOT}"))  {
     # Get oneAPI version
     $script:ONEAPI_VERSION = icpx --version
     $script:ONEAPI_VERSION = [regex]::Match($script:ONEAPI_VERSION, '(?<=oneAPI DPC\+\+/C\+\+ Compiler )(?\d+\.\d+\.\d+)').Value
@@ -302,7 +310,7 @@ function build_oneapi() {
     $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
     $script:cmakeDefs += @(
       "-G", "MinGW Makefiles",
-      "-DLLAMA_SYCL=ON",
+      "-DGGML_SYCL=ON",
       "-DCMAKE_C_COMPILER=icx",
       "-DCMAKE_CXX_COMPILER=icx",
       "-DCMAKE_BUILD_TYPE=Release"
@@ -317,16 +325,18 @@ function build_oneapi() {
     sign
     install
 
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:distDir}"
+    rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\" -ea 0 > $null
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
   } else {
     Write-Host "Skipping oneAPI generation step"
   }
@@ -346,10 +356,11 @@ function build_rocm() {
             "-G", "Ninja", 
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DLLAMA_HIPBLAS=on",
+            "-DGGML_HIPBLAS=on",
+            "-DLLAMA_CUDA_NO_PEER_COPY=on",
             "-DHIP_PLATFORM=amd",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
             "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
             "-DAMDGPU_TARGETS=$(amdGPUs)",
             "-DGPU_TARGETS=$(amdGPUs)"
@@ -375,7 +386,6 @@ function build_rocm() {
         sign
         install
 
-        # Assumes v5.7, may need adjustments for v6
         rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
         md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
         cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
diff --git a/llm/ggla.go b/llm/ggla.go
index a5d90b6cb67a37cdbdb6a063a4bfaafc5bbf61b8..831f60712f53bd0144a7f5812b9777062f59a9b7 100644
--- a/llm/ggla.go
+++ b/llm/ggla.go
@@ -36,6 +36,8 @@ type ggla struct {
 
 	kv      KV
 	tensors []*Tensor
+
+	tensorOffset uint64
 }
 
 func newGGLA(container *containerGGLA) *ggla {
@@ -50,10 +52,13 @@ func (llm *ggla) KV() KV {
 }
 
 func (llm *ggla) Tensors() Tensors {
-	return llm.tensors
+	return Tensors{
+		Items:  llm.tensors,
+		Offset: llm.tensorOffset,
+	}
 }
 
-func (llm *ggla) decode(rs io.ReadSeeker) error {
+func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
 	var r uint32
 	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
 		return err
@@ -66,12 +71,28 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 	}
 	llm.kv["alpha"] = alpha
 
+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	llm.tensorOffset = uint64(offset)
+
 	for {
 		var dims uint32
 		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
+			if errors.Is(err, io.EOF) {
+				return nil
+			}
 			return err
 		}
 
+		defer func() {
+			if errors.Is(retErr, io.EOF) {
+				retErr = io.ErrUnexpectedEOF
+			}
+		}()
+
 		var namesize uint32
 		if err := binary.Read(rs, binary.LittleEndian, &namesize); err != nil {
 			return err
@@ -108,7 +129,7 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 			return err
 		}
 
-		if _, err := rs.Seek((offset+31)&-32, io.SeekStart); err != nil {
+		if _, err := rs.Seek((offset+31)&-32-offset, io.SeekCurrent); err != nil {
 			return err
 		}
 
diff --git a/llm/ggml.go b/llm/ggml.go
index 878800f307ebb34ea363de5810d8ac65799efc1a..d7f2eef7cbe2099d4220fd8c2f54a04574656a56 100644
--- a/llm/ggml.go
+++ b/llm/ggml.go
@@ -6,6 +6,8 @@ import (
 	"fmt"
 	"io"
 	"strings"
+
+	"github.com/ollama/ollama/util/bufioutil"
 )
 
 type GGML struct {
@@ -69,6 +71,30 @@ func (kv KV) HeadCountKV() uint64 {
 	return 1
 }
 
+func (kv KV) EmbeddingHeadCount() uint64 {
+	if heads := kv.HeadCount(); heads > 0 {
+		return kv.EmbeddingLength() / kv.HeadCount()
+	}
+
+	return 0
+}
+
+func (kv KV) EmbeddingHeadCountK() uint64 {
+	if k := kv.u64(fmt.Sprintf("%s.attention.key_length", kv.Architecture())); k > 0 {
+		return k
+	}
+
+	return kv.EmbeddingHeadCount()
+}
+
+func (kv KV) EmbeddingHeadCountV() uint64 {
+	if v := kv.u64(fmt.Sprintf("%s.attention.value_length", kv.Architecture())); v > 0 {
+		return v
+	}
+
+	return kv.EmbeddingHeadCount()
+}
+
 func (kv KV) GQA() uint64 {
 	return kv.HeadCount() / kv.HeadCountKV()
 }
@@ -81,11 +107,19 @@ func (kv KV) ContextLength() uint64 {
 	return kv.u64(fmt.Sprintf("%s.context_length", kv.Architecture()))
 }
 
-type Tensors []*Tensor
+func (kv KV) ChatTemplate() string {
+	s, _ := kv["tokenizer.chat_template"].(string)
+	return s
+}
+
+type Tensors struct {
+	Items  []*Tensor
+	Offset uint64
+}
 
 func (ts Tensors) Layers() map[string]Layer {
 	layers := make(map[string]Layer)
-	for _, t := range ts {
+	for _, t := range ts.Items {
 		parts := strings.Split(t.Name, ".")
 		if parts[0] == "blk" {
 			// join first and second part, e.g. blk.%d
@@ -249,7 +283,18 @@ func DetectGGMLType(b []byte) string {
 	}
 }
 
-func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
+// DecodeGGML decodes a GGML model from the given reader.
+//
+// It collects array values for arrays with a size less than or equal to
+// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
+// the maxArraySize is negative, all arrays are collected.
+func DecodeGGML(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
+	if maxArraySize == 0 {
+		maxArraySize = 1024
+	}
+
+	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
+
 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
 		return nil, 0, err
@@ -262,17 +307,15 @@ func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
 	case FILE_MAGIC_GGLA:
 		c = &containerGGLA{}
 	case FILE_MAGIC_GGUF_LE:
-		c = &containerGGUF{ByteOrder: binary.LittleEndian}
+		c = &containerGGUF{ByteOrder: binary.LittleEndian, maxArraySize: maxArraySize}
 	case FILE_MAGIC_GGUF_BE:
-		c = &containerGGUF{ByteOrder: binary.BigEndian}
+		c = &containerGGUF{ByteOrder: binary.BigEndian, maxArraySize: maxArraySize}
 	default:
 		return nil, 0, errors.New("invalid file magic")
 	}
 
 	model, err := c.Decode(rs)
-	if errors.Is(err, io.EOF) {
-		// noop
-	} else if err != nil {
+	if err != nil {
 		return nil, 0, err
 	}
 
@@ -292,7 +335,10 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 	embedding := llm.KV().EmbeddingLength()
 	heads := llm.KV().HeadCount()
 	headsKV := llm.KV().HeadCountKV()
-	vocab := uint64(len(llm.KV()["tokenizer.ggml.tokens"].([]any)))
+	vocab := uint64(llm.KV()["tokenizer.ggml.tokens"].(*array).size)
+
+	embeddingHeads := llm.KV().EmbeddingHeadCount()
+	embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
 
 	layers := llm.Tensors().Layers()
 
@@ -302,7 +348,8 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 
 		partialOffload = 4 * batch * embedding
 		partialOffload += max(
-			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embedding/heads*headsKV),
+			// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
+			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
 
@@ -310,21 +357,30 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			// mixtral 8x22b
 			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
-				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
-				4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
+				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
+				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
 			)
 		} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {
 			// mixtral 8x7b
 			ffnGateWeight1 := ffnGateWeight.Shape[1]
 			fullOffload = 4 * batch * (2 + 3*embedding + context*(1+heads) + 2*headsKV + ffnGateWeight1)
 			partialOffload = max(
-				4*batch*(3+embedding/heads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
+				4*batch*(3+embeddingHeads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
 				4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
 			)
 		}
-	case "gemma":
-		fullOffload = 4 * batch * (embedding + vocab)
-		partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128
+	case "gemma", "gemma2":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
+		)
+
+		partialOffload = max(
+			4*embedding*batch+embedding*vocab*105/128+4*vocab*batch,
+			4*batch*(2*embedding+1+2*embeddingHeadsK*heads+context+context*heads)+
+				4*embeddingHeadsK*context*8+
+				embedding*embeddingHeadsK*heads*9/16,
+		)
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
@@ -361,6 +417,42 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			4*batch*(vocab+2*embedding),
 			fullOffload,
 		)
+	case "deepseek2":
+		fullOffload = max(
+			4*batch*(3*embedding+vocab),
+			4*batch*(3*embedding+2+context*(1+headsKV)+2*embeddingHeadsK*headsKV),
+		)
+
+		partialOffload = max(
+			4*batch*(3*embedding+vocab)+embedding*vocab*105/128,
+			4*batch*(2*embedding+1+2*embeddingHeadsK*headsKV+context+context*headsKV)+4*embeddingHeadsK*context*headsKV+embedding*embeddingHeadsK*headsKV*9/16,
+		)
+	case "chatglm":
+		fullOffload = 4 * batch * (embedding + vocab)
+		partialOffload = 4*batch*(embedding+vocab) + embedding*vocab*105/128
+		if qkvBias, ok := layers["blk.0"]["attn_qkv.bias"]; ok {
+			fullOffload = max(
+				fullOffload,
+				4*batch*(2+
+					2*embedding+
+					context+
+					context*heads+
+					embeddingHeadsK*heads+
+					qkvBias.Shape[0]),
+			)
+
+			partialOffload = max(
+				partialOffload,
+				4*batch*(1+
+					2*embedding+
+					embeddingHeadsK*heads+
+					context+
+					context*heads)+
+					4*embeddingHeadsK*context+
+					4*context*embeddingHeadsK+
+					4*qkvBias.Shape[0],
+			)
+		}
 	}
 
 	return
diff --git a/llm/ggml_test.go b/llm/ggml_test.go
new file mode 100644
index 0000000000000000000000000000000000000000..006c3ded89c81b68eddcae0a41b66f174928a7f5
--- /dev/null
+++ b/llm/ggml_test.go
@@ -0,0 +1 @@
+package llm
diff --git a/llm/gguf.go b/llm/gguf.go
index 0ba48f76fc433688593db75d6fde360b44f7d0a5..981583131b392ddfc76b7b54cf9682088f02faf6 100644
--- a/llm/gguf.go
+++ b/llm/gguf.go
@@ -2,12 +2,16 @@ package llm
 
 import (
 	"bytes"
+	"cmp"
 	"encoding/binary"
+	"encoding/json"
 	"fmt"
 	"io"
+	"log/slog"
+	"slices"
 	"strings"
 
-	"log/slog"
+	"golang.org/x/exp/maps"
 )
 
 type containerGGUF struct {
@@ -29,6 +33,12 @@ type containerGGUF struct {
 		NumTensor uint64
 		NumKV     uint64
 	}
+
+	maxArraySize int
+}
+
+func (c *containerGGUF) canCollectArray(size int) bool {
+	return c.maxArraySize < 0 || size <= c.maxArraySize
 }
 
 func (c *containerGGUF) Name() string {
@@ -54,7 +64,6 @@ func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
 	}
 
 	model := newGGUF(c)
-	slog.Debug(fmt.Sprintf("model = %#v", model))
 	if err := model.Decode(rs); err != nil {
 		return nil, err
 	}
@@ -84,7 +93,10 @@ type gguf struct {
 	kv      KV
 	tensors []*Tensor
 
-	parameters uint64
+	parameters   uint64
+	tensorOffset uint64
+
+	scratch [16 << 10]byte
 }
 
 func newGGUF(container *containerGGUF) *gguf {
@@ -94,16 +106,15 @@ func newGGUF(container *containerGGUF) *gguf {
 	}
 }
 
-func NewGGUFV3(bo binary.ByteOrder) *gguf {
-	return newGGUF(&containerGGUF{ByteOrder: bo, Version: 3})
-}
-
 func (llm *gguf) KV() KV {
 	return llm.kv
 }
 
 func (llm *gguf) Tensors() Tensors {
-	return llm.tensors
+	return Tensors{
+		Items:  llm.tensors,
+		Offset: llm.tensorOffset,
+	}
 }
 
 func (llm *gguf) numTensor() uint64 {
@@ -181,34 +192,34 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	}
 
 	// decode tensors
-	for i := 0; uint64(i) < llm.numTensor(); i++ {
+	for range llm.numTensor() {
 		name, err := readGGUFString(llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor name: %w", err)
 		}
 
 		// dims is the number of dimensions in the tensor
 		dims, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor dimensions: %w", err)
 		}
 
-		shape := [4]uint64{1, 1, 1, 1}
+		shape := make([]uint64, dims)
 		for i := 0; uint32(i) < dims; i++ {
 			shape[i], err = readGGUF[uint64](llm, rs)
 			if err != nil {
-				return err
+				return fmt.Errorf("failed to read tensor shape: %w", err)
 			}
 		}
 
 		kind, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor kind: %w", err)
 		}
 
 		offset, err := readGGUF[uint64](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor offset: %w", err)
 		}
 
 		tensor := Tensor{
@@ -235,19 +246,22 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		return err
 	}
 
-	padding := llm.padding(offset, int64(alignment))
-	if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
-		return err
-	}
+	padding := ggufPadding(offset, int64(alignment))
+	llm.tensorOffset = uint64(offset + padding)
 
 	for _, tensor := range llm.tensors {
-		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
-			return err
+		offset, err := rs.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return fmt.Errorf("failed to get current offset: %w", err)
 		}
 
-		padding := llm.padding(int64(tensor.Size()), int64(alignment))
+		padding := ggufPadding(offset, int64(alignment))
 		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
-			return err
+			return fmt.Errorf("failed to seek to init padding: %w", err)
+		}
+
+		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
+			return fmt.Errorf("failed to seek to tensor: %w", err)
 		}
 	}
 
@@ -260,12 +274,12 @@ func readGGUF[T any](llm *gguf, r io.Reader) (T, error) {
 	return t, err
 }
 
-func writeGGUF[V any](llm *gguf, w io.Writer, t uint32, v V) error {
-	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
+func writeGGUF[V any](w io.Writer, t uint32, v V) error {
+	if err := binary.Write(w, binary.LittleEndian, t); err != nil {
 		return err
 	}
 
-	return binary.Write(w, llm.ByteOrder, v)
+	return binary.Write(w, binary.LittleEndian, v)
 }
 
 func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
@@ -285,30 +299,56 @@ func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
 	return b.String(), nil
 }
 
+func discardGGUFString(llm *gguf, r io.Reader) error {
+	buf := llm.scratch[:8]
+	_, err := io.ReadFull(r, buf)
+	if err != nil {
+		return err
+	}
+
+	size := int(llm.ByteOrder.Uint64(buf))
+	for size > 0 {
+		n, err := r.Read(llm.scratch[:min(size, cap(llm.scratch))])
+		if err != nil {
+			return err
+		}
+		size -= n
+	}
+	return nil
+}
+
 func readGGUFString(llm *gguf, r io.Reader) (string, error) {
 	if llm.Version == 1 {
 		return readGGUFV1String(llm, r)
 	}
 
-	var length uint64
-	if err := binary.Read(r, llm.ByteOrder, &length); err != nil {
+	buf := llm.scratch[:8]
+	_, err := io.ReadFull(r, buf)
+	if err != nil {
 		return "", err
 	}
 
-	var b bytes.Buffer
-	if _, err := io.CopyN(&b, r, int64(length)); err != nil {
-		return "", err
+	length := int(llm.ByteOrder.Uint64(buf))
+	if length > len(llm.scratch) {
+		buf = make([]byte, length)
+	} else {
+		buf = llm.scratch[:length]
 	}
+	clear(buf)
 
-	return b.String(), nil
+	_, err = io.ReadFull(r, buf)
+	if err != nil {
+		return "", err
+	}
+	return string(buf), nil
 }
 
-func writeGGUFString(llm *gguf, w io.Writer, s string) error {
-	if err := binary.Write(w, llm.ByteOrder, ggufTypeString); err != nil {
+func writeGGUFString(w io.Writer, s string) error {
+	if err := binary.Write(w, binary.LittleEndian, ggufTypeString); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
+	if err := binary.Write(w, binary.LittleEndian, uint64(len(s))); err != nil {
 		return err
 	}
 
@@ -316,7 +356,16 @@ func writeGGUFString(llm *gguf, w io.Writer, s string) error {
 	return err
 }
 
-func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
+type array struct {
+	size   int
+	values []any
+}
+
+func (a *array) MarshalJSON() ([]byte, error) {
+	return json.Marshal(a.values)
+}
+
+func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
 	t, err := readGGUF[uint32](llm, r)
 	if err != nil {
 		return nil, err
@@ -327,7 +376,12 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
 		return nil, err
 	}
 
-	for i := 0; uint32(i) < n; i++ {
+	a := &array{size: int(n)}
+	if llm.canCollectArray(int(n)) {
+		a.values = make([]any, 0, int(n))
+	}
+
+	for i := range n {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -361,13 +415,15 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
 			return nil, err
 		}
 
-		a = append(a, e)
+		if a.values != nil {
+			a.values[i] = e
+		}
 	}
 
-	return
+	return a, nil
 }
 
-func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
+func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 	if llm.Version == 1 {
 		return readGGUFV1Array(llm, r)
 	}
@@ -382,7 +438,12 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 		return nil, err
 	}
 
-	for i := 0; uint64(i) < n; i++ {
+	a := &array{size: int(n)}
+	if llm.canCollectArray(int(n)) {
+		a.values = make([]any, int(n))
+	}
+
+	for i := range n {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -408,7 +469,11 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 		case ggufTypeBool:
 			e, err = readGGUF[bool](llm, r)
 		case ggufTypeString:
-			e, err = readGGUFString(llm, r)
+			if a.values != nil {
+				e, err = readGGUFString(llm, r)
+			} else {
+				err = discardGGUFString(llm, r)
+			}
 		default:
 			return nil, fmt.Errorf("invalid array type: %d", t)
 		}
@@ -416,238 +481,183 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 			return nil, err
 		}
 
-		a = append(a, e)
+		if a.values != nil {
+			a.values[i] = e
+		}
 	}
 
-	return
+	return a, nil
 }
 
-func writeGGUFArray[S ~[]E, E any](llm *gguf, w io.Writer, t uint32, s S) error {
-	if err := binary.Write(w, llm.ByteOrder, ggufTypeArray); err != nil {
+// writeGGUFArray writes a slice s of type E to the write with a gguf type of t
+func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
+	if err := binary.Write(w, binary.LittleEndian, ggufTypeArray); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
+	if err := binary.Write(w, binary.LittleEndian, t); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
+	if err := binary.Write(w, binary.LittleEndian, uint64(len(s))); err != nil {
 		return err
 	}
 
-	for _, e := range s {
-		if err := binary.Write(w, llm.ByteOrder, e); err != nil {
-			return err
-		}
-	}
-
-	return nil
-}
-
-var ggufKVOrder = map[string][]string{
-	"llama": {
-		"general.architecture",
-		"general.name",
-		"llama.vocab_size",
-		"llama.context_length",
-		"llama.embedding_length",
-		"llama.block_count",
-		"llama.feed_forward_length",
-		"llama.attention.head_count",
-		"llama.attention.head_count_kv",
-		"llama.attention.layer_norm_rms_epsilon",
-		"llama.rope.freq_base",
-		"llama.rope.dimension_count",
-		"llama.expert_count",
-		"llama.expert_used_count",
-		"gemma.context_length",
-		"gemma.embedding_length",
-		"gemma.block_count",
-		"gemma.feed_forward_length",
-		"gemma.attention.head_count",
-		"gemma.attention.head_count_kv",
-		"gemma.attention.layer_norm_rms_epsilon",
-		"gemma.attention.key_length",
-		"gemma.attention.value_length",
-		"general.file_type",
-		"tokenizer.ggml.pre",
-		"tokenizer.ggml.model",
-		"tokenizer.ggml.tokens",
-		"tokenizer.ggml.scores",
-		"tokenizer.ggml.merges",
-		"tokenizer.ggml.token_type",
-		"tokenizer.ggml.bos_token_id",
-		"tokenizer.ggml.eos_token_id",
-		"tokenizer.ggml.unknown_token_id",
-		"tokenizer.ggml.padding_token_id",
-		"tokenizer.ggml.add_bos_token",
-		"tokenizer.ggml.add_eos_token",
-		"tokenizer.chat_template",
-	},
+	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
-	switch llm.Version {
-	case 3:
-		llm.V3.NumTensor = uint64(len(tensors))
-		llm.V3.NumKV = uint64(len(kv))
-	default:
-		return fmt.Errorf("not implemented: ggufv%d", llm.Version)
-	}
-
-	if err := binary.Write(ws, llm.ByteOrder, []byte("GGUF")); err != nil {
+func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
+	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, llm.ByteOrder, llm.Version); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint32(3)); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, llm.ByteOrder, llm.numTensor()); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(ts))); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, llm.ByteOrder, llm.numKV()); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(kv))); err != nil {
 		return err
 	}
 
-	kvCheck := make(map[string]bool)
-	for k := range kv {
-		kvCheck[k] = false
+	keys := maps.Keys(kv)
+	slices.Sort(keys)
+
+	for _, key := range keys {
+		if err := ggufWriteKV(ws, key, kv[key]); err != nil {
+			return err
+		}
 	}
 
-	for _, k := range ggufKVOrder["llama"] {
-		v, ok := kv[k]
-		if !ok {
-			continue
+	slices.SortFunc(ts, func(a, b Tensor) int {
+		var i, j int
+		if n, err := fmt.Sscanf(a.Name, "blk.%d", &i); err != nil || n != 1 {
+			return cmp.Compare(a.Name, b.Name)
+		} else if n, err := fmt.Sscanf(b.Name, "blk.%d", &j); err != nil || n != 1 {
+			return cmp.Compare(a.Name, b.Name)
 		}
-		kvCheck[k] = true
 
-		if err := binary.Write(ws, llm.ByteOrder, uint64(len(k))); err != nil {
+		return cmp.Compare(i, j)
+	})
+
+	var s uint64
+	for _, t := range ts {
+		t.Offset = s
+		if err := ggufWriteTensorInfo(ws, t); err != nil {
 			return err
 		}
+		s += t.Size()
+	}
 
-		if err := binary.Write(ws, llm.ByteOrder, []byte(k)); err != nil {
+	var alignment int64 = 32
+	for _, t := range ts {
+		if err := ggufWriteTensor(ws, t, alignment); err != nil {
 			return err
 		}
+	}
 
-		var err error
-		switch v := v.(type) {
-		case uint32:
-			err = writeGGUF(llm, ws, ggufTypeUint32, v)
-		case float32:
-			err = writeGGUF(llm, ws, ggufTypeFloat32, v)
-		case bool:
-			err = writeGGUF(llm, ws, ggufTypeBool, v)
-		case string:
-			err = writeGGUFString(llm, ws, v)
-		case []int32:
-			err = writeGGUFArray(llm, ws, ggufTypeInt32, v)
-		case []uint32:
-			err = writeGGUFArray(llm, ws, ggufTypeUint32, v)
-		case []float32:
-			err = writeGGUFArray(llm, ws, ggufTypeFloat32, v)
-		case []string:
-			if err := binary.Write(ws, llm.ByteOrder, ggufTypeArray); err != nil {
-				return err
-			}
-
-			if err := binary.Write(ws, llm.ByteOrder, ggufTypeString); err != nil {
-				return err
-			}
-
-			if err := binary.Write(ws, llm.ByteOrder, uint64(len(v))); err != nil {
-				return err
-			}
-
-			for _, e := range v {
-				if err := binary.Write(ws, llm.ByteOrder, uint64(len(e))); err != nil {
-					return err
-				}
+	return nil
+}
 
-				if err := binary.Write(ws, llm.ByteOrder, []byte(e)); err != nil {
-					return err
-				}
-			}
-		default:
-			return fmt.Errorf("improper type for '%s'", k)
-		}
-		if err != nil {
-			return err
-		}
+func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
+	slog.Debug(k, "type", fmt.Sprintf("%T", v))
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(k))); err != nil {
+		return err
 	}
 
-	for k, v := range kvCheck {
-		if !v {
-			return fmt.Errorf("Didn't know how to write kv %s", k)
-		}
+	if err := binary.Write(ws, binary.LittleEndian, []byte(k)); err != nil {
+		return err
 	}
 
-	for _, tensor := range tensors {
-		if err := binary.Write(ws, llm.ByteOrder, uint64(len(tensor.Name))); err != nil {
+	var err error
+	switch v := v.(type) {
+	case uint32:
+		err = writeGGUF(ws, ggufTypeUint32, v)
+	case float32:
+		err = writeGGUF(ws, ggufTypeFloat32, v)
+	case bool:
+		err = writeGGUF(ws, ggufTypeBool, v)
+	case string:
+		err = writeGGUFString(ws, v)
+	case []int32:
+		err = writeGGUFArray(ws, ggufTypeInt32, v)
+	case []uint32:
+		err = writeGGUFArray(ws, ggufTypeUint32, v)
+	case []float32:
+		err = writeGGUFArray(ws, ggufTypeFloat32, v)
+	case []string:
+		if err := binary.Write(ws, binary.LittleEndian, ggufTypeArray); err != nil {
 			return err
 		}
 
-		if err := binary.Write(ws, llm.ByteOrder, []byte(tensor.Name)); err != nil {
+		if err := binary.Write(ws, binary.LittleEndian, ggufTypeString); err != nil {
 			return err
 		}
 
-		dims := 0
-		for cnt := 0; cnt < len(tensor.Shape); cnt++ {
-			if tensor.Shape[cnt] > 0 {
-				dims++
-			}
-		}
-
-		if err := binary.Write(ws, llm.ByteOrder, uint32(dims)); err != nil {
+		if err := binary.Write(ws, binary.LittleEndian, uint64(len(v))); err != nil {
 			return err
 		}
 
-		for i := 0; i < dims; i++ {
-			if err := binary.Write(ws, llm.ByteOrder, uint64(tensor.Shape[dims-1-i])); err != nil {
+		for _, e := range v {
+			if err := binary.Write(ws, binary.LittleEndian, uint64(len(e))); err != nil {
 				return err
 			}
-		}
 
-		if err := binary.Write(ws, llm.ByteOrder, tensor.Kind); err != nil {
-			return err
+			if err := binary.Write(ws, binary.LittleEndian, []byte(e)); err != nil {
+				return err
+			}
 		}
+	default:
+		return fmt.Errorf("improper type for '%s'", k)
+	}
 
-		if err := binary.Write(ws, llm.ByteOrder, tensor.Offset); err != nil {
-			return err
-		}
+	return err
+}
+
+func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
+	slog.Debug(t.Name, "kind", t.Kind, "shape", t.Shape, "offset", t.Offset)
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(t.Name))); err != nil {
+		return err
 	}
 
-	offset, err := ws.Seek(0, io.SeekCurrent)
-	if err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, []byte(t.Name)); err != nil {
 		return err
 	}
 
-	var alignment int64 = 32
-	padding := llm.padding(offset, alignment)
-	if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint32(len(t.Shape))); err != nil {
 		return err
 	}
 
-	for _, tensor := range tensors {
-		if _, err := tensor.WriteTo(ws); err != nil {
+	for i := range len(t.Shape) {
+		if err := binary.Write(ws, binary.LittleEndian, t.Shape[len(t.Shape)-i-1]); err != nil {
 			return err
 		}
+	}
 
-		offset, err := ws.Seek(0, io.SeekCurrent)
-		if err != nil {
-			return err
-		}
+	if err := binary.Write(ws, binary.LittleEndian, t.Kind); err != nil {
+		return err
+	}
 
-		padding := llm.padding(offset, alignment)
-		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
-			return err
-		}
+	return binary.Write(ws, binary.LittleEndian, t.Offset)
+}
+
+func ggufWriteTensor(ws io.WriteSeeker, t Tensor, alignment int64) error {
+	offset, err := ws.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
 	}
 
-	return nil
+	if err := binary.Write(ws, binary.LittleEndian, bytes.Repeat([]byte{0}, int(ggufPadding(offset, alignment)))); err != nil {
+		return err
+	}
+
+	_, err = t.WriteTo(ws)
+	return err
 }
 
-func (gguf) padding(offset, align int64) int64 {
+func ggufPadding(offset, align int64) int64 {
 	return (align - offset%align) % align
 }
diff --git a/llm/llama.cpp/.clang-tidy b/llm/llama.cpp/.clang-tidy
new file mode 100644
index 0000000000000000000000000000000000000000..952c0cca82580e2068864c6ee2344c167a4aff02
--- /dev/null
+++ b/llm/llama.cpp/.clang-tidy
@@ -0,0 +1,24 @@
+---
+Checks: >
+    bugprone-*,
+    -bugprone-easily-swappable-parameters,
+    -bugprone-implicit-widening-of-multiplication-result,
+    -bugprone-misplaced-widening-cast,
+    -bugprone-narrowing-conversions,
+    readability-*,
+    -readability-avoid-unconditional-preprocessor-if,
+    -readability-function-cognitive-complexity,
+    -readability-identifier-length,
+    -readability-implicit-bool-conversion,
+    -readability-magic-numbers,
+    -readability-uppercase-literal-suffix,
+    -readability-simplify-boolean-expr,
+    clang-analyzer-*,
+    -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
+    performance-*,
+    portability-*,
+    misc-*,
+    -misc-const-correctness,
+    -misc-non-private-member-variables-in-classes,
+    -misc-no-recursion,
+FormatStyle: none
diff --git a/llm/llama.cpp/.devops/cloud-v-pipeline b/llm/llama.cpp/.devops/cloud-v-pipeline
index f3a4944f8a419fdaaca79296006511e0bbfd0401..af8c0cea6155c71c0c162e034715908ea16ee9cf 100644
--- a/llm/llama.cpp/.devops/cloud-v-pipeline
+++ b/llm/llama.cpp/.devops/cloud-v-pipeline
@@ -15,7 +15,7 @@ node('x86_runner1'){            // Running on x86 runner containing latest vecto
     stage('Running llama.cpp'){
         sh'''#!/bin/bash
             module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
-            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
             cat llama_log.txt                   # Printing results
         '''
     }
diff --git a/llm/llama.cpp/.devops/full-cuda.Dockerfile b/llm/llama.cpp/.devops/full-cuda.Dockerfile
index 059fd26950607a9a911e0a8d9e2c6055fb118f19..61f671465ca71b466cf9b1de3b47ee972605b721 100644
--- a/llm/llama.cpp/.devops/full-cuda.Dockerfile
+++ b/llm/llama.cpp/.devops/full-cuda.Dockerfile
@@ -6,13 +6,13 @@ ARG CUDA_VERSION=11.7.1
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
-FROM ${BASE_CUDA_DEV_CONTAINER} as build
+FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
 # Unless otherwise specified, we make a fat build.
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -27,10 +27,10 @@ COPY . .
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable CUDA
-ENV LLAMA_CUDA=1
+ENV GGML_CUDA=1
 # Enable cURL
 ENV LLAMA_CURL=1
 
-RUN make
+RUN make -j$(nproc)
 
 ENTRYPOINT ["/app/.devops/tools.sh"]
diff --git a/llm/llama.cpp/.devops/full-rocm.Dockerfile b/llm/llama.cpp/.devops/full-rocm.Dockerfile
index 6ecf3bcc7cb83817ef03ee60b46cdb25ece6cf81..680d1cb92205d7a107babb6f3e932a0846fa2474 100644
--- a/llm/llama.cpp/.devops/full-rocm.Dockerfile
+++ b/llm/llama.cpp/.devops/full-rocm.Dockerfile
@@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
-FROM ${BASE_ROCM_DEV_CONTAINER} as build
+FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
@@ -36,7 +36,7 @@ COPY . .
 # Set nvcc architecture
 ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
 # Enable ROCm
-ENV LLAMA_HIPBLAS=1
+ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
 
@@ -45,6 +45,6 @@ ENV LLAMA_CURL=1
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev
 
-RUN make
+RUN make -j$(nproc)
 
 ENTRYPOINT ["/app/.devops/tools.sh"]
diff --git a/llm/llama.cpp/.devops/full.Dockerfile b/llm/llama.cpp/.devops/full.Dockerfile
index 432fb5dadcbca5f73f7d87e70190a26aa663019b..2a06f82b738aee88ed82266c710e7c942e408976 100644
--- a/llm/llama.cpp/.devops/full.Dockerfile
+++ b/llm/llama.cpp/.devops/full.Dockerfile
@@ -1,9 +1,9 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -18,7 +18,7 @@ COPY . .
 ENV LLAMA_CURL=1
 
 
-RUN make
+RUN make -j$(nproc)
 
 ENV LC_ALL=C.utf8
 
diff --git a/llm/llama.cpp/.devops/llama-cli-cuda.Dockerfile b/llm/llama.cpp/.devops/llama-cli-cuda.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..8eda63a89504d7bb5a7849210316da5aead483b8
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-cli-cuda.Dockerfile
@@ -0,0 +1,35 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable CUDA
+ENV GGML_CUDA=1
+
+RUN make -j$(nproc) llama-cli
+
+FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libgomp1
+
+COPY --from=build /app/llama-cli /llama-cli
+
+ENTRYPOINT [ "/llama-cli" ]
diff --git a/llm/llama.cpp/.devops/llama-cli-intel.Dockerfile b/llm/llama.cpp/.devops/llama-cli-intel.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..79dba06a77d6eba3c8e2b71d809a4cdf3213d337
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-cli-intel.Dockerfile
@@ -0,0 +1,28 @@
+ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
+
+ARG GGML_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
+        echo "GGML_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
+    fi && \
+    echo "Building with static libs" && \
+    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
+    ${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \
+    cmake --build build --config Release --target llama-cli
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
+
+COPY --from=build /app/build/bin/llama-cli /llama-cli
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/llama-cli" ]
diff --git a/llm/llama.cpp/.devops/llama-cli-rocm.Dockerfile b/llm/llama.cpp/.devops/llama-cli-rocm.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..c3d1ab06702ec5fc437de8919a893f6792492102
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-cli-rocm.Dockerfile
@@ -0,0 +1,45 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV GGML_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make -j$(nproc) llama-cli
+
+ENTRYPOINT [ "/app/llama-cli" ]
diff --git a/llm/llama.cpp/.devops/llama-cli-vulkan.Dockerfile b/llm/llama.cpp/.devops/llama-cli-vulkan.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..9b0dad8bf7a13b04d9953ce586eee34e1ec9694a
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-cli-vulkan.Dockerfile
@@ -0,0 +1,27 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget libgomp1
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN cmake -B build -DGGML_VULKAN=1 && \
+    cmake --build build --config Release --target llama-cli
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/llama-cli /llama-cli && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/llama-cli" ]
diff --git a/llm/llama.cpp/.devops/llama-cli.Dockerfile b/llm/llama.cpp/.devops/llama-cli.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..7f741aa46ecf0e81de24a5697cfa4c9cebcf5def
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-cli.Dockerfile
@@ -0,0 +1,23 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+RUN make -j$(nproc) llama-cli
+
+FROM ubuntu:$UBUNTU_VERSION AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libgomp1
+
+COPY --from=build /app/llama-cli /llama-cli
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/llama-cli" ]
diff --git a/llm/llama.cpp/.devops/llama-cpp-cuda.srpm.spec b/llm/llama.cpp/.devops/llama-cpp-cuda.srpm.spec
index ba9cb7cbb824fc03f8274bf26494d1a5d946ad73..7425d3a9d7a4027fcff13077ffec3caf6e66d02f 100644
--- a/llm/llama.cpp/.devops/llama-cpp-cuda.srpm.spec
+++ b/llm/llama.cpp/.devops/llama-cpp-cuda.srpm.spec
@@ -32,13 +32,13 @@ CPU inference for Meta's Lllama2 models using default options.
 %setup -n llama.cpp-master
 
 %build
-make -j LLAMA_CUDA=1
+make -j GGML_CUDA=1
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p main %{buildroot}%{_bindir}/llamacppcuda
-cp -p server %{buildroot}%{_bindir}/llamacppcudaserver
-cp -p simple %{buildroot}%{_bindir}/llamacppcudasimple
+cp -p llama-cli %{buildroot}%{_bindir}/llama-cuda-cli
+cp -p llama-server %{buildroot}%{_bindir}/llama-cuda-server
+cp -p llama-simple %{buildroot}%{_bindir}/llama-cuda-simple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
 %{__cat} < %{buildroot}/usr/lib/systemd/system/llamacuda.service
@@ -49,7 +49,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
 [Service]
 Type=simple
 EnvironmentFile=/etc/sysconfig/llama
-ExecStart=/usr/bin/llamacppcudaserver $LLAMA_ARGS
+ExecStart=/usr/bin/llama-cuda-server $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -67,9 +67,9 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llamacppcuda
-%{_bindir}/llamacppcudaserver
-%{_bindir}/llamacppcudasimple
+%{_bindir}/llama-cuda-cli
+%{_bindir}/llama-cuda-server
+%{_bindir}/llama-cuda-simple
 /usr/lib/systemd/system/llamacuda.service
 %config /etc/sysconfig/llama
 
diff --git a/llm/llama.cpp/.devops/llama-cpp.srpm.spec b/llm/llama.cpp/.devops/llama-cpp.srpm.spec
index 1d9e4f425b43add0b46fbc90cd79160f85915ec2..4d5560089816c6679bd02b673dc57a2206ebfbe7 100644
--- a/llm/llama.cpp/.devops/llama-cpp.srpm.spec
+++ b/llm/llama.cpp/.devops/llama-cpp.srpm.spec
@@ -38,9 +38,9 @@ make -j
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p main %{buildroot}%{_bindir}/llama
-cp -p server %{buildroot}%{_bindir}/llamaserver
-cp -p simple %{buildroot}%{_bindir}/llamasimple
+cp -p llama-cli %{buildroot}%{_bindir}/llama-cli
+cp -p llama-server %{buildroot}%{_bindir}/llama-server
+cp -p llama-simple %{buildroot}%{_bindir}/llama-simple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
 %{__cat} < %{buildroot}/usr/lib/systemd/system/llama.service
@@ -51,7 +51,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
 [Service]
 Type=simple
 EnvironmentFile=/etc/sysconfig/llama
-ExecStart=/usr/bin/llamaserver $LLAMA_ARGS
+ExecStart=/usr/bin/llama-server $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -69,9 +69,9 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llama
-%{_bindir}/llamaserver
-%{_bindir}/llamasimple
+%{_bindir}/llama-cli
+%{_bindir}/llama-server
+%{_bindir}/llama-simple
 /usr/lib/systemd/system/llama.service
 %config /etc/sysconfig/llama
 
diff --git a/llm/llama.cpp/.devops/llama-server-cuda.Dockerfile b/llm/llama.cpp/.devops/llama-server-cuda.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..67328cf1c17881a5b6bb7d51e0d4d8b7f3ebd3f6
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-server-cuda.Dockerfile
@@ -0,0 +1,39 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable CUDA
+ENV GGML_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
+
+RUN make -j$(nproc) llama-server
+
+FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/llama-server /llama-server
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]
diff --git a/llm/llama.cpp/.devops/llama-server-intel.Dockerfile b/llm/llama.cpp/.devops/llama-server-intel.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..f525658dddfe5d6beff9c589c36c4500442c5353
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-server-intel.Dockerfile
@@ -0,0 +1,32 @@
+ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
+
+ARG GGML_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
+        echo "GGML_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
+    fi && \
+    echo "Building with dynamic libs" && \
+    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
+    cmake --build build --config Release --target llama-server
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev curl
+
+COPY --from=build /app/build/bin/llama-server /llama-server
+
+ENV LC_ALL=C.utf8
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]
diff --git a/llm/llama.cpp/.devops/llama-server-rocm.Dockerfile b/llm/llama.cpp/.devops/llama-server-rocm.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..763b4cd3f1c2ed8038da9019ac2e403a650f9adc
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-server-rocm.Dockerfile
@@ -0,0 +1,52 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV GGML_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+# Enable cURL
+ENV LLAMA_CURL=1
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev curl
+
+RUN make -j$(nproc) llama-server
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/app/llama-server" ]
diff --git a/llm/llama.cpp/.devops/llama-server-vulkan.Dockerfile b/llm/llama.cpp/.devops/llama-server-vulkan.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..13a61ffd8454b68cbd88f8cfa32e0018800ccd1b
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-server-vulkan.Dockerfile
@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK and cURL
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN cmake -B build -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
+    cmake --build build --config Release --target llama-server
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/llama-server /llama-server && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]
diff --git a/llm/llama.cpp/.devops/llama-server.Dockerfile b/llm/llama.cpp/.devops/llama-server.Dockerfile
new file mode 100644
index 0000000000000000000000000000000000000000..ff558604ebde23616976bd8bac654f000788dad4
--- /dev/null
+++ b/llm/llama.cpp/.devops/llama-server.Dockerfile
@@ -0,0 +1,27 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+ENV LLAMA_CURL=1
+
+RUN make -j$(nproc) llama-server
+
+FROM ubuntu:$UBUNTU_VERSION AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/llama-server /llama-server
+
+ENV LC_ALL=C.utf8
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]
diff --git a/llm/llama.cpp/.devops/nix/apps.nix b/llm/llama.cpp/.devops/nix/apps.nix
index b8a12cc0a0463063a6083acd16e4bf6491420a77..0ecf19fc56d554c69aeef8a03b253fc15338688e 100644
--- a/llm/llama.cpp/.devops/nix/apps.nix
+++ b/llm/llama.cpp/.devops/nix/apps.nix
@@ -6,11 +6,10 @@
         let
           inherit (config.packages) default;
           binaries = [
-            "llama"
+            "llama-cli"
             "llama-embedding"
             "llama-server"
-            "quantize"
-            "train-text-from-scratch"
+            "llama-quantize"
           ];
           mkApp = name: {
             type = "app";
diff --git a/llm/llama.cpp/.devops/nix/package.nix b/llm/llama.cpp/.devops/nix/package.nix
index e8d5b0bd92d2cec6088e764bd764d9bfb8a9f238..a87423c71307909cf88ac9ddd6178c0fa3b95970 100644
--- a/llm/llama.cpp/.devops/nix/package.nix
+++ b/llm/llama.cpp/.devops/nix/package.nix
@@ -17,19 +17,19 @@
   rocmPackages,
   vulkan-headers,
   vulkan-loader,
-  clblast,
+  curl,
+  shaderc,
   useBlas ? builtins.all (x: !x) [
     useCuda
     useMetalKit
-    useOpenCL
     useRocm
     useVulkan
   ] && blas.meta.available,
   useCuda ? config.cudaSupport,
-  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
+  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
   useMpi ? false, # Increases the runtime closure size by ~700M
-  useOpenCL ? false,
   useRocm ? config.rocmSupport,
+  enableCurl ? true,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
 
@@ -56,7 +56,6 @@ let
     ++ lib.optionals useCuda [ "CUDA" ]
     ++ lib.optionals useMetalKit [ "MetalKit" ]
     ++ lib.optionals useMpi [ "MPI" ]
-    ++ lib.optionals useOpenCL [ "OpenCL" ]
     ++ lib.optionals useRocm [ "ROCm" ]
     ++ lib.optionals useVulkan [ "Vulkan" ];
 
@@ -91,6 +90,22 @@ let
       ps.tiktoken
       ps.torchWithoutCuda
       ps.transformers
+
+      # server bench
+      ps.matplotlib
+
+      # server tests
+      ps.openai
+      ps.behave
+      ps.prometheus-client
+
+      # for examples/pydantic-models-to-grammar-examples.py
+      ps.docstring-parser
+      ps.pydantic
+
+      # for scripts/compare-llama-bench.py
+      ps.gitpython
+      ps.tabulate
     ]
   );
 
@@ -111,16 +126,9 @@ let
     ++ optionals useMetalKit [ MetalKit ];
 
   cudaBuildInputs = with cudaPackages; [
-    cuda_cccl.dev # 
-
-    # A temporary hack for reducing the closure size, remove once cudaPackages
-    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
-    cuda_cudart.dev
-    cuda_cudart.lib
-    cuda_cudart.static
-    libcublas.dev
-    libcublas.lib
-    libcublas.static
+    cuda_cudart
+    cuda_cccl # 
+    libcublas
   ];
 
   rocmBuildInputs = with rocmPackages; [
@@ -132,6 +140,7 @@ let
   vulkanBuildInputs = [
     vulkan-headers
     vulkan-loader
+    shaderc
   ];
 in
 
@@ -160,9 +169,9 @@ effectiveStdenv.mkDerivation (
     };
 
     postPatch = ''
-      substituteInPlace ./ggml-metal.m \
+      substituteInPlace ./ggml/src/ggml-metal.m \
         --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-      substituteInPlace ./ggml-metal.m \
+      substituteInPlace ./ggml/src/ggml-metal.m \
         --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
     '';
 
@@ -198,24 +207,24 @@ effectiveStdenv.mkDerivation (
       optionals effectiveStdenv.isDarwin darwinBuildInputs
       ++ optionals useCuda cudaBuildInputs
       ++ optionals useMpi [ mpi ]
-      ++ optionals useOpenCL [ clblast ]
       ++ optionals useRocm rocmBuildInputs
       ++ optionals useBlas [ blas ]
-      ++ optionals useVulkan vulkanBuildInputs;
+      ++ optionals useVulkan vulkanBuildInputs
+      ++ optionals enableCurl [ curl ];
 
     cmakeFlags =
       [
-        (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
         (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
-        (cmakeBool "LLAMA_BLAS" useBlas)
-        (cmakeBool "LLAMA_CLBLAST" useOpenCL)
-        (cmakeBool "LLAMA_CUDA" useCuda)
-        (cmakeBool "LLAMA_HIPBLAS" useRocm)
-        (cmakeBool "LLAMA_METAL" useMetalKit)
-        (cmakeBool "LLAMA_VULKAN" useVulkan)
-        (cmakeBool "LLAMA_STATIC" enableStatic)
+        (cmakeBool "LLAMA_CURL" enableCurl)
+        (cmakeBool "GGML_NATIVE" false)
+        (cmakeBool "GGML_BLAS" useBlas)
+        (cmakeBool "GGML_CUDA" useCuda)
+        (cmakeBool "GGML_HIPBLAS" useRocm)
+        (cmakeBool "GGML_METAL" useMetalKit)
+        (cmakeBool "GGML_VULKAN" useVulkan)
+        (cmakeBool "GGML_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (
@@ -231,7 +240,7 @@ effectiveStdenv.mkDerivation (
       ]
       ++ optionals useMetalKit [
         (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
-        (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
+        (cmakeBool "GGML_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
       ];
 
     # Environment variables needed for ROCm
@@ -243,10 +252,8 @@ effectiveStdenv.mkDerivation (
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.
     postInstall = ''
-      mv $out/bin/main${executableSuffix} $out/bin/llama${executableSuffix}
-      mv $out/bin/server${executableSuffix} $out/bin/llama-server${executableSuffix}
       mkdir -p $out/include
-      cp $src/llama.h $out/include/
+      cp $src/include/llama.h $out/include/
     '';
 
     # Define the shells here, but don't add in the inputsFrom to avoid recursion.
@@ -256,7 +263,6 @@ effectiveStdenv.mkDerivation (
         useCuda
         useMetalKit
         useMpi
-        useOpenCL
         useRocm
         useVulkan
         ;
@@ -283,7 +289,7 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals useCuda lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.
@@ -294,7 +300,7 @@ effectiveStdenv.mkDerivation (
       license = lib.licenses.mit;
 
       # Accommodates `nix run` and `lib.getExe`
-      mainProgram = "llama";
+      mainProgram = "llama-cli";
 
       # These people might respond, on the best effort basis, if you ping them
       # in case of Nix-specific regressions or for reviewing Nix-specific PRs.
diff --git a/llm/llama.cpp/.devops/tools.sh b/llm/llama.cpp/.devops/tools.sh
index 97424c3aa746a9f0d8cf41fc978fd9bb357bf9a2..24dcfd35079cb73236dc80d911480bcbacc7008d 100644
--- a/llm/llama.cpp/.devops/tools.sh
+++ b/llm/llama.cpp/.devops/tools.sh
@@ -8,13 +8,11 @@ arg1="$1"
 shift
 
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert-hf-to-gguf.py "$@"
+    python3 ./convert_hf_to_gguf.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
-    ./quantize "$@"
+    ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
-    ./main "$@"
-elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
-    ./finetune "$@"
+    ./llama-cli "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -22,11 +20,11 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
             echo "Skip model quantization, it already exists: ${i/f16/q4_0}"
         else
             echo "Converting PTH to GGML: $i into ${i/f16/q4_0}..."
-            ./quantize "$i" "${i/f16/q4_0}" q4_0
+            ./llama-quantize "$i" "${i/f16/q4_0}" q4_0
         fi
     done
 elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
-    ./server "$@"
+    ./llama-server "$@"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "
@@ -36,8 +34,6 @@ else
     echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
-    echo "  --finetune (-f): Run finetune command to create a lora finetune of the model"
-    echo "              See documentation for finetune for command-line parameters"
     echo "  --all-in-one (-a): Execute --convert & --quantize"
     echo "              ex: \"/models/\" 7B"
     echo "  --server (-s): Run a model on the server"
diff --git a/llm/llama.cpp/.dockerignore b/llm/llama.cpp/.dockerignore
index 633bbc3a971c117193b64bddcc38c451ba984faa..8916e2a660732a6111ed0b7bbc7684500c4f977e 100644
--- a/llm/llama.cpp/.dockerignore
+++ b/llm/llama.cpp/.dockerignore
@@ -12,8 +12,8 @@ build*/
 
 models/*
 
-/main
-/quantize
+/llama-cli
+/llama-quantize
 
 arm_neon.h
 compile_commands.json
diff --git a/llm/llama.cpp/.editorconfig b/llm/llama.cpp/.editorconfig
index 16d16b3b55bf5575e956bebed8de6b635bee4d8c..f88f8da67cd78758ed461758ae646ea62c5fbbc5 100644
--- a/llm/llama.cpp/.editorconfig
+++ b/llm/llama.cpp/.editorconfig
@@ -26,3 +26,7 @@ indent_size = 2
 
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab
+
+[examples/cvector-generator/*.txt]
+trim_trailing_whitespace = unset
+insert_final_newline = unset
diff --git a/llm/llama.cpp/.github/ISSUE_TEMPLATE/01-bug-low.yml b/llm/llama.cpp/.github/ISSUE_TEMPLATE/01-bug-low.yml
index bfb9d9a0692c4b4fc0cdcbd13d55296a9ec8bf13..54785854f776e647a53547b1b52e416d2bcd4ee4 100644
--- a/llm/llama.cpp/.github/ISSUE_TEMPLATE/01-bug-low.yml
+++ b/llm/llama.cpp/.github/ISSUE_TEMPLATE/01-bug-low.yml
@@ -24,7 +24,7 @@ body:
       label: Name and Version
       description: Which executable and which version of our software are you running? (use `--version` to get a version string)
       placeholder: |
-        $./main --version
+        $./llama-cli --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:
diff --git a/llm/llama.cpp/.github/ISSUE_TEMPLATE/02-bug-medium.yml b/llm/llama.cpp/.github/ISSUE_TEMPLATE/02-bug-medium.yml
index e8297eea035514c2377ad1cfcb9562f37ab2bf49..a6285c6f05bac8c551f0fc0284a98ba1cf88d4ee 100644
--- a/llm/llama.cpp/.github/ISSUE_TEMPLATE/02-bug-medium.yml
+++ b/llm/llama.cpp/.github/ISSUE_TEMPLATE/02-bug-medium.yml
@@ -24,7 +24,7 @@ body:
       label: Name and Version
       description: Which executable and which version of our software are you running? (use `--version` to get a version string)
       placeholder: |
-        $./main --version
+        $./llama-cli --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:
diff --git a/llm/llama.cpp/.github/ISSUE_TEMPLATE/03-bug-high.yml b/llm/llama.cpp/.github/ISSUE_TEMPLATE/03-bug-high.yml
index 3c9d50d169720d18fda30a9aa30fb159c60b59a0..ff816b93769c3a28fd81cd20a702f37038a1ecd9 100644
--- a/llm/llama.cpp/.github/ISSUE_TEMPLATE/03-bug-high.yml
+++ b/llm/llama.cpp/.github/ISSUE_TEMPLATE/03-bug-high.yml
@@ -24,7 +24,7 @@ body:
       label: Name and Version
       description: Which executable and which version of our software are you running? (use `--version` to get a version string)
       placeholder: |
-        $./main --version
+        $./llama-cli --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:
diff --git a/llm/llama.cpp/.github/ISSUE_TEMPLATE/04-bug-critical.yml b/llm/llama.cpp/.github/ISSUE_TEMPLATE/04-bug-critical.yml
index d089d5fa10cfcc7366cda464c6c7cf15f5b09e8d..7af42a80b3b93ce7ce231f21c336482a5d56b574 100644
--- a/llm/llama.cpp/.github/ISSUE_TEMPLATE/04-bug-critical.yml
+++ b/llm/llama.cpp/.github/ISSUE_TEMPLATE/04-bug-critical.yml
@@ -24,7 +24,7 @@ body:
       label: Name and Version
       description: Which executable and which version of our software are you running? (use `--version` to get a version string)
       placeholder: |
-        $./main --version
+        $./llama-cli --version
         version: 2999 (42b4109e)
         built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
     validations:
diff --git a/llm/llama.cpp/.github/ISSUE_TEMPLATE/config.yml b/llm/llama.cpp/.github/ISSUE_TEMPLATE/config.yml
index c88134dbb644a598526511979938d5c2db144c03..eb8c4b472df4c4fb22e7190a320dbf800eea4807 100644
--- a/llm/llama.cpp/.github/ISSUE_TEMPLATE/config.yml
+++ b/llm/llama.cpp/.github/ISSUE_TEMPLATE/config.yml
@@ -9,5 +9,3 @@ contact_links:
   - name: Want to contribute?
     url: https://github.com/ggerganov/llama.cpp/wiki/contribute
     about: Head to the contribution guide page of the wiki for areas you can help with
-
-
diff --git a/llm/llama.cpp/.github/labeler.yml b/llm/llama.cpp/.github/labeler.yml
index 97d739b5811e8ca82e508105f6c85f2d8be3d745..89436740d1ffb218304426c402300b3aedc177e7 100644
--- a/llm/llama.cpp/.github/labeler.yml
+++ b/llm/llama.cpp/.github/labeler.yml
@@ -2,31 +2,33 @@
 Kompute:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml-kompute.h
-            - ggml-kompute.cpp
+            - ggml/include/ggml-kompute.h
+            - ggml/src/ggml-kompute.cpp
             - README-kompute.md
 Apple Metal:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml-metal.h
-            - ggml-metal.cpp
+            - ggml/include/ggml-metal.h
+            - ggml/src/ggml-metal.cpp
             - README-metal.md
 SYCL:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml-sycl.h
-            - ggml-sycl.cpp
-            - README-sycl.md
+            - ggml/include/ggml-sycl.h
+            - ggml/src/ggml-sycl.cpp
+            - ggml/src/ggml-sycl/**
+            - docs/backend/SYCL.md
+            - examples/sycl/**
 Nvidia GPU:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml-cuda.h
-            - ggml-cuda/**
+            - ggml/include/ggml-cuda.h
+            - ggml/src/ggml-cuda/**
 Vulkan:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml_vk_generate_shaders.py
-            - ggml-vulkan*
+            - ggml/ggml_vk_generate_shaders.py
+            - ggml/src/ggml-vulkan*
 documentation:
     - changed-files:
         - any-glob-to-any-file:
@@ -42,7 +44,6 @@ build:
             - cmake/**
             - CMakeLists.txt
             - CMakePresets.json
-            - codecov.yml
 examples:
     - changed-files:
         - any-glob-to-any-file: examples/**
@@ -74,10 +75,10 @@ server:
 ggml:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml.c
-            - ggml.h
-            - ggml-*.c
-            - ggml-*.h
+            - ggml/include/ggml*.h
+            - ggml/src/ggml*.c
+            - ggml/src/ggml*.cpp
+            - ggml/src/ggml*.h
             - ggml-cuda/**
 nix:
     - changed-files:
diff --git a/llm/llama.cpp/.github/pull_request_template.md b/llm/llama.cpp/.github/pull_request_template.md
new file mode 100644
index 0000000000000000000000000000000000000000..997c6d9d053973bd740e6eca3df1606ab4d99d24
--- /dev/null
+++ b/llm/llama.cpp/.github/pull_request_template.md
@@ -0,0 +1,7 @@
+
+
+- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
+- Self-reported review complexity:
+  - [ ] Low
+  - [ ] Medium
+  - [ ] High
diff --git a/llm/llama.cpp/.github/workflows/bench.yml b/llm/llama.cpp/.github/workflows/bench.yml
index de0d994c80e9d94cc587d4cf37d46e3886b45d1f..eb69b82c47e643c47ebe9bb2d9b622bc55af8451 100644
--- a/llm/llama.cpp/.github/workflows/bench.yml
+++ b/llm/llama.cpp/.github/workflows/bench.yml
@@ -109,7 +109,7 @@ jobs:
         run: |
           set -eux
           cmake -B build \
-              -DLLAMA_NATIVE=OFF \
+              -DGGML_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
               -DLLAMA_CURL=ON \
               -DLLAMA_CUBLAS=ON \
@@ -119,7 +119,7 @@ jobs:
               -DLLAMA_FATAL_WARNINGS=OFF \
               -DLLAMA_ALL_WARNINGS=OFF \
               -DCMAKE_BUILD_TYPE=Release;
-          cmake --build build --config Release -j $(nproc) --target server
+          cmake --build build --config Release -j $(nproc) --target llama-server
 
       - name: Download the dataset
         id: download_dataset
diff --git a/llm/llama.cpp/.github/workflows/build.yml b/llm/llama.cpp/.github/workflows/build.yml
new file mode 100644
index 0000000000000000000000000000000000000000..b9246659a6ef0655c85b9cd85636b3d6d7940b77
--- /dev/null
+++ b/llm/llama.cpp/.github/workflows/build.yml
@@ -0,0 +1,1316 @@
+name: CI
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+
+jobs:
+  macOS-latest-cmake-arm64:
+    runs-on: macos-14
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L 'main|curl' --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+          name: llama-bin-macos-arm64.zip
+
+  macOS-latest-cmake-x64:
+    runs-on: macos-12
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
+          # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+          name: llama-bin-macos-x64.zip
+
+  ubuntu-focal-make:
+    runs-on: ubuntu-20.04
+    env:
+      LLAMA_NODE_AVAILABLE: true
+      LLAMA_PYTHON_AVAILABLE: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential gcc-8
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "20"
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Build
+        id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
+        run: |
+          CC=gcc-8 make -j $(nproc)
+
+      - name: Test
+        id: make_test
+        run: |
+          CC=gcc-8 make tests -j $(nproc)
+          make test -j $(nproc)
+
+  ubuntu-focal-make-curl:
+    runs-on: ubuntu-20.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
+
+      - name: Build
+        id: make_build
+        env:
+          LLAMA_FATAL_WARNINGS: 1
+          LLAMA_CURL: 1
+        run: |
+          CC=gcc-8 make -j $(nproc)
+
+  ubuntu-latest-cmake:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake --build . --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L 'main|curl' --verbose --timeout 900
+
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
+          name: llama-bin-ubuntu-x64.zip
+
+  ubuntu-latest-cmake-sanitizer:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        if: ${{ matrix.sanitizer != 'THREAD' }}
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Build (no OpenMP)
+        id: cmake_build_no_openmp
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DGGML_OPENMP=OFF
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+  ubuntu-latest-cmake-rpc:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake -DGGML_RPC=ON ..
+          cmake --build . --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose
+
+  ubuntu-22-cmake-vulkan:
+    runs-on: ubuntu-22.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential vulkan-sdk
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake -DGGML_VULKAN=ON ..
+          cmake --build . --config Release -j $(nproc)
+
+  ubuntu-22-cmake-hip:
+    runs-on: ubuntu-22.04
+    container: rocm/dev-ubuntu-22.04:6.0.2
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
+
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=ON
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Build with legacy HIP support
+        id: cmake_build_legacy_hip
+        run: |
+          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIPBLAS=ON
+          cmake --build build2 --config Release -j $(nproc)
+
+  ubuntu-22-cmake-sycl:
+    runs-on: ubuntu-22.04
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: add oneAPI to apt
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
+
+      - name: install oneAPI dpcpp compiler
+        shell: bash
+        run: |
+          sudo apt update
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+
+      - name: install oneAPI MKL library
+        shell: bash
+        run: |
+          sudo apt install intel-oneapi-mkl-devel
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          mkdir build
+          cd build
+          cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake --build . --config Release -j $(nproc)
+
+  ubuntu-22-cmake-sycl-fp16:
+    runs-on: ubuntu-22.04
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: add oneAPI to apt
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
+
+      - name: install oneAPI dpcpp compiler
+        shell: bash
+        run: |
+          sudo apt update
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+
+      - name: install oneAPI MKL library
+        shell: bash
+        run: |
+          sudo apt install intel-oneapi-mkl-devel
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          mkdir build
+          cd build
+          cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON ..
+          cmake --build . --config Release -j $(nproc)
+
+  # TODO: build with GGML_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
+  macOS-latest-make:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
+        run: |
+          GGML_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: make_test
+        run: |
+          GGML_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
+          GGML_NO_METAL=1 make test  -j $(sysctl -n hw.logicalcpu)
+
+  # TODO: build with GGML_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
+  #       how to debug it.
+  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
+  #       would be great if we fix these
+  macOS-latest-cmake:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+  macOS-latest-cmake-ios:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -G Xcode .. \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+  macOS-latest-cmake-tvos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -G Xcode .. \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=tvOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+  macOS-latest-swift:
+    runs-on: macos-latest
+
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          xcodebuild -scheme llama -destination "${{ matrix.destination }}"
+
+      - name: Build Swift Example
+        id: make_build_swift_example
+        run: |
+            make swift
+
+  windows-msys2:
+    runs-on: windows-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { sys: UCRT64,  env: ucrt-x86_64,  build: Release }
+          - { sys: CLANG64, env: clang-x86_64, build: Release }
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Setup ${{ matrix.sys }}
+        uses: msys2/setup-msys2@v2
+        with:
+          update: true
+          msystem: ${{matrix.sys}}
+          install: >-
+            base-devel
+            mingw-w64-${{matrix.env}}-toolchain
+            mingw-w64-${{matrix.env}}-cmake
+            mingw-w64-${{matrix.env}}-openblas
+
+      - name: Build using make
+        shell: msys2 {0}
+        run: |
+            make -j $(nproc)
+
+      - name: Clean after building using make
+        shell: msys2 {0}
+        run: |
+            make clean
+
+      - name: Build using make w/ OpenBLAS
+        shell: msys2 {0}
+        run: |
+            make GGML_OPENBLAS=1 -j $(nproc)
+
+      - name: Build using CMake
+        shell: msys2 {0}
+        run: |
+            cmake -B build
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
+
+      - name: Clean after building using CMake
+        shell: msys2 {0}
+        run: |
+            rm -rf build
+
+      - name: Build using CMake w/ OpenBLAS
+        shell: msys2 {0}
+        run: |
+            cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
+
+  windows-latest-cmake:
+    runs-on: windows-2019
+
+    env:
+      OPENBLAS_VERSION: 0.3.23
+      SDE_VERSION: 9.33.0-2024-01-07
+      VULKAN_VERSION: 1.3.261.1
+
+    strategy:
+      matrix:
+        include:
+          - build: 'rpc-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'noavx-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+          - build: 'avx2-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'avx-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+          - build: 'avx512-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'openblas-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+          - build: 'kompute-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'vulkan-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'llvm-arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'msvc-arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Clone Kompute submodule
+        id: clone_kompute
+        if: ${{ matrix.build == 'kompute-x64' }}
+        run: |
+          git submodule update --init ggml/src/kompute
+
+      - name: Download OpenBLAS
+        id: get_openblas
+        if: ${{ matrix.build == 'openblas-x64' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
+          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
+          mkdir $env:RUNNER_TEMP/openblas
+          tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas
+          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
+          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
+          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
+          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
+
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -S . -B build ${{ matrix.defines }}
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Add libopenblas.dll
+        id: add_libopenblas_dll
+        if: ${{ matrix.build == 'openblas-x64' }}
+        run: |
+          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
+          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
+
+      - name: Check AVX512F support
+        id: check_avx512f
+        if: ${{ matrix.build == 'avx512-x64' }}
+        continue-on-error: true
+        run: |
+          cd build
+          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
+          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
+          $cl =  $(join-path $msvc 'bin\Hostx64\x64\cl.exe')
+          echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c
+          & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
+          .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
+
+      - name: Test
+        id: cmake_test
+        # not all machines have native AVX-512
+        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
+        run: |
+          cd build
+          ctest -L main -C Release --verbose --timeout 900
+
+      - name: Test (Intel SDE)
+        id: cmake_test_sde
+        if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
+          # for some weird reason windows tar doesn't like sde tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
+          $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
+          cd build
+          $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
+          & $sde -future -- ctest -L main -C Release --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
+          name: llama-bin-win-${{ matrix.build }}.zip
+
+  windows-latest-cmake-cuda:
+    runs-on: windows-2019
+
+    strategy:
+      matrix:
+        cuda: ['12.2.0', '11.7.1']
+        build: ['cuda']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Install CUDA toolkit
+        id: cuda-toolkit
+        uses: Jimver/cuda-toolkit@v0.2.15
+        with:
+          cuda: ${{ matrix.cuda }}
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
+          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+
+      - name: Copy and pack Cuda runtime
+        run: |
+          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
+
+      - name: Upload Cuda runtime
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+
+  windows-latest-cmake-sycl:
+    runs-on: windows-latest
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Install
+        run:  scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: Build
+        id: cmake_build
+        run:  examples/sycl/win-build-sycl.bat
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          echo "cp oneAPI running time dll files to ./build/bin done"
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          name: llama-bin-win-sycl-x64.zip
+
+  windows-latest-cmake-hip:
+    runs-on: windows-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
+          cmake --build build --config Release
+
+  ios-xcode-build:
+    runs-on: macos-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Build Xcode project
+        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
+
+  android-build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Set up JDK
+        uses: actions/setup-java@v3
+        with:
+          java-version: 17
+          distribution: zulu
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v3
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Build
+        run: |
+          cd examples/llama.android
+
+          ./gradlew build --no-daemon
+
+#  freeBSD-latest:
+#    runs-on: macos-12
+#    steps:
+#    - name: Clone
+#      uses: actions/checkout@v4
+#
+#    - name: Build
+#      uses: cross-platform-actions/action@v0.19.0
+#      with:
+#        operating_system: freebsd
+#        version: '13.2'
+#        hypervisor: 'qemu'
+#        run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 openblas
+#            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu`
+
+  release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+
+    runs-on: ubuntu-latest
+
+    needs:
+      - ubuntu-focal-make
+      - ubuntu-latest-cmake
+      - macOS-latest-make
+      - macOS-latest-cmake
+      - windows-latest-cmake
+      - windows-latest-cmake-cuda
+      - macOS-latest-cmake-arm64
+      - macOS-latest-cmake-x64
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+
+      - name: Move artifacts
+        id: move_artifacts
+        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
+
+      - name: Create release
+        id: create_release
+        uses: anzz1/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ steps.tag.outputs.name }}
+
+      - name: Upload release
+        id: upload_release
+        uses: actions/github-script@v3
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./artifact/release')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: file,
+                  data: await fs.readFileSync(`./artifact/release/${file}`)
+                });
+              }
+            }
+
+#  ubuntu-latest-gcc:
+#    runs-on: ubuntu-latest
+#
+#    strategy:
+#      matrix:
+#        build: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Dependencies
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#          sudo apt-get install cmake
+#
+#      - name: Configure
+#        run: cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+#
+#      - name: Build
+#        run: |
+#          make
+#
+#  ubuntu-latest-clang:
+#    runs-on: ubuntu-latest
+#
+#    strategy:
+#      matrix:
+#        build: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Dependencies
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#          sudo apt-get install cmake
+#
+#      - name: Configure
+#        run: cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
+#
+#      - name: Build
+#        run: |
+#          make
+#
+#  ubuntu-latest-gcc-sanitized:
+#    runs-on: ubuntu-latest
+#
+#    strategy:
+#      matrix:
+#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Dependencies
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#          sudo apt-get install cmake
+#
+#      - name: Configure
+#        run: cmake . -DCMAKE_BUILD_TYPE=Debug -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON
+#
+#      - name: Build
+#        run: |
+#          make
+#
+#  windows:
+#    runs-on: windows-latest
+#
+#    strategy:
+#      matrix:
+#        build: [Release]
+#        arch: [Win32, x64]
+#        include:
+#          - arch: Win32
+#            s2arc: x86
+#          - arch: x64
+#            s2arc: x64
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Add msbuild to PATH
+#        uses: microsoft/setup-msbuild@v1
+#
+#      - name: Configure
+#        run: >
+#          cmake -S . -B ./build -A ${{ matrix.arch }}
+#          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+#
+#      - name: Build
+#        run: |
+#          cd ./build
+#          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+#
+#      - name: Upload binaries
+#        uses: actions/upload-artifact@v4
+#        with:
+#          name: llama-bin-${{ matrix.arch }}
+#          path: build/bin/${{ matrix.build }}
+#
+#  windows-blas:
+#    runs-on: windows-latest
+#
+#    strategy:
+#      matrix:
+#        build: [Release]
+#        arch: [Win32, x64]
+#        blas: [ON]
+#        include:
+#          - arch: Win32
+#            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip
+#            s2arc: x86
+#          - arch: x64
+#            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
+#            s2arc: x64
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Add msbuild to PATH
+#        uses: microsoft/setup-msbuild@v1
+#
+#      - name: Fetch OpenBLAS
+#        if: matrix.blas == 'ON'
+#        run: |
+#          C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }}
+#          7z x blas.zip -oblas -y
+#          copy blas/include/cblas.h .
+#          copy blas/include/openblas_config.h .
+#          echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+#
+#      - name: Configure
+#        run: >
+#          cmake -S . -B ./build -A ${{ matrix.arch }}
+#          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+#          -DLLAMA_SUPPORT_OPENBLAS=${{ matrix.blas }}
+#          -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
+#
+#      - name: Build
+#        run: |
+#          cd ./build
+#          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+#
+#      - name: Copy libopenblas.dll
+#        if: matrix.blas == 'ON'
+#        run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+#
+#      - name: Upload binaries
+#        if: matrix.blas == 'ON'
+#        uses: actions/upload-artifact@v4
+#        with:
+#          name: llama-blas-bin-${{ matrix.arch }}
+#          path: build/bin/${{ matrix.build }}
+#
+#  emscripten:
+#    runs-on: ubuntu-latest
+#
+#    strategy:
+#      matrix:
+#        build: [Release]
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Dependencies
+#        run: |
+#          wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz
+#          tar -xvf master.tar.gz
+#          emsdk-master/emsdk update
+#          emsdk-master/emsdk install latest
+#          emsdk-master/emsdk activate latest
+#
+#      - name: Configure
+#        run: echo "tmp"
+#
+#      - name: Build
+#        run: |
+#          pushd emsdk-master
+#          source ./emsdk_env.sh
+#          popd
+#          emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+#          make
diff --git a/llm/llama.cpp/.github/workflows/docker.yml b/llm/llama.cpp/.github/workflows/docker.yml
index 9b03d19bc77c616cf11946baf917c594851086f3..bf94b2024e814694754b616ff3f8d6f178de7ee5 100644
--- a/llm/llama.cpp/.github/workflows/docker.yml
+++ b/llm/llama.cpp/.github/workflows/docker.yml
@@ -10,10 +10,11 @@
 name: Publish Docker image
 
 on:
-  pull_request:
+  #pull_request:
   push:
     branches:
       - master
+    paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -22,7 +23,7 @@ concurrency:
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
-    if: github.event.pull_request.draft == false
+    #if: github.event.pull_request.draft == false
 
     runs-on: ubuntu-latest
     env:
@@ -30,20 +31,18 @@ jobs:
     strategy:
       matrix:
         config:
-          - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light", dockerfile: ".devops/llama-cli.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server", dockerfile: ".devops/llama-server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
-          #                     have disabled them for now until the reason why
-          #                     is understood.
-          - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          # Note: the full-rocm image is failing due to a "no space left on device" error. It is disabled for now to allow the workflow to complete.
+          #- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
diff --git a/llm/llama.cpp/.github/workflows/python-type-check.yml b/llm/llama.cpp/.github/workflows/python-type-check.yml
new file mode 100644
index 0000000000000000000000000000000000000000..e5ff5e6d792cb3fd05534f912e43bbb476f31bf0
--- /dev/null
+++ b/llm/llama.cpp/.github/workflows/python-type-check.yml
@@ -0,0 +1,38 @@
+name: Python Type-Check
+
+on:
+  push:
+    paths:
+      - '.github/workflows/python-type-check.yml'
+      - '**.py'
+      - '**/requirements*.txt'
+  pull_request:
+    paths:
+      - '.github/workflows/python-type-check.yml'
+      - '**.py'
+      - '**/requirements*.txt'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  python-type-check:
+    runs-on: ubuntu-latest
+    name: pyright type-check
+    steps:
+      - name: Check out source repository
+        uses: actions/checkout@v4
+      - name: Set up Python environment
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Python dependencies
+        # TODO: use a venv
+        run: pip install -r requirements/requirements-all.txt
+      - name: Type-check with Pyright
+        uses: jakebailey/pyright-action@v2
+        with:
+          version: 1.1.370
+          level: warning
+          warnings: true
diff --git a/llm/llama.cpp/.github/workflows/server.yml b/llm/llama.cpp/.github/workflows/server.yml
index 0789efd18a1ab59ceed12f5412e97e3916a9ffba..99feb28f2a54554a2d22028ae98f1c0dcee9b2ed 100644
--- a/llm/llama.cpp/.github/workflows/server.yml
+++ b/llm/llama.cpp/.github/workflows/server.yml
@@ -16,11 +16,9 @@ on:
     branches:
       - master
     paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
-  pull_request_target:
+  pull_request:
     types: [opened, synchronize, reopened]
     paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
-  schedule:
-    -  cron: '2 4 * * *'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
@@ -32,7 +30,7 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
         build_type: [RelWithDebInfo]
         include:
           - build_type: Release
@@ -89,16 +87,30 @@ jobs:
             exit 1
           fi
 
+      - name: Build (no OpenMP)
+        id: cmake_build_no_openmp
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        run: |
+          cmake -B build \
+              -DGGML_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+              -DGGML_OPENMP=OFF ;
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
       - name: Build
         id: cmake_build
+        if: ${{ matrix.sanitizer != 'THREAD' }}
         run: |
           cmake -B build \
-              -DLLAMA_NATIVE=OFF \
+              -DGGML_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
               -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target server
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
 
       - name: Tests
         id: server_integration_tests
@@ -115,7 +127,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     steps:
       - name: Clone
@@ -138,7 +150,7 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
-          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
         id: setup_python
diff --git a/llm/llama.cpp/.gitignore b/llm/llama.cpp/.gitignore
index 50ae0973ae3b306be0f817a30bdc63c8254186e6..c9b4d99839729cd4abac5271cf5660036138c821 100644
--- a/llm/llama.cpp/.gitignore
+++ b/llm/llama.cpp/.gitignore
@@ -1,126 +1,132 @@
-*.o
+# Extensions
+
 *.a
-*.so
-*.gguf
-*.gguf.json
+*.bat
 *.bin
-*.exe
 *.dll
-*.log
-*.gcov
-*.gcno
-*.gcda
 *.dot
-*.bat
-*.tmp
-*.metallib
 *.etag
+*.exe
+*.gcda
+*.gcno
+*.gcov
+*.gguf
+*.gguf.json
 *.lastModified
-.DS_Store
-.build/
+*.log
+*.metallib
+*.o
+*.so
+*.tmp
+
+# IDE / OS
+
 .cache/
 .ccls-cache/
 .direnv/
+.DS_Store
 .envrc
+.idea/
 .swiftpm
-.venv
-.clang-tidy
 .vs/
 .vscode/
-.idea/
+nppBackup
 
-ggml-metal-embed.metal
 
-lcov-report/
+# Coverage
+
 gcovr-report/
+lcov-report/
 
+# Build Artifacts
+
+tags
+.build/
 build*
+!build-info.cmake
+!build-info.cpp.in
+!build-info.sh
 !build.zig
+!docs/build.md
+/libllama.so
+/llama-*
+/vulkan-shaders-gen
+android-ndk-*
+arm_neon.h
 cmake-build-*
+CMakeSettings.json
+compile_commands.json
+ggml-metal-embed.metal
+llama-batched-swift
+/rpc-server
 out/
 tmp/
 
-models/*
-models-mnt
+# Deprecated
 
-/Pipfile
-/baby-llama
-/beam-search
-/benchmark-matmult
-/convert-llama2c-to-ggml
-/embd-input-test
-/embedding
-/eval-callback
-/gguf
-/gguf-llama-simple
-/gguf-split
-/gritlm
-/imatrix
-/infill
-/libllama.so
-/llama-bench
-/llava-cli
-/lookahead
-/lookup
-/lookup-create
-/lookup-merge
-/lookup-stats
 /main
-/metal
-/passkey
-/perplexity
-/q8dot
-/quantize
-/quantize-stats
-/result
-/save-load-state
 /server
-/simple
-/batched
-/batched-bench
-/export-lora
-/finetune
-/retrieval
-/speculative
-/parallel
-/train-text-from-scratch
-/tokenize
-/vdot
-/common/build-info.cpp
-arm_neon.h
-compile_commands.json
-CMakeSettings.json
 
-__pycache__
-dist
+# CI
+
+!.github/workflows/*.yml
+
+# Models
+
+models/*
+models-mnt
+!models/.editorconfig
+!models/ggml-vocab-*.gguf*
+
+# Zig
 
 zig-out/
 zig-cache/
 
+# Logs
+
 ppl-*.txt
 qnt-*.txt
 perf-*.txt
 
+# Examples
+
 examples/jeopardy/results.txt
+examples/server/*.css.hpp
 examples/server/*.html.hpp
 examples/server/*.js.hpp
 examples/server/*.mjs.hpp
+!build_64.sh
+!examples/*.bat
+!examples/*/*.kts
+!examples/*/*/*.kts
+!examples/sycl/*.bat
+!examples/sycl/*.sh
 
-poetry.lock
+# Python
+
+/.venv
+__pycache__/
+*/poetry.lock
 poetry.toml
-nppBackup
+
+# Nix
+/result
 
 # Test binaries
-/tests/test-grammar-parser
-/tests/test-llama-grammar
+/tests/test-backend-ops
 /tests/test-double-float
 /tests/test-grad0
+/tests/test-grammar-parser
+/tests/test-llama-grammar
 /tests/test-opt
 /tests/test-quantize-fns
 /tests/test-quantize-perf
+/tests/test-rope
 /tests/test-sampling
 /tests/test-tokenizer-0
-/tests/test-tokenizer-1-spm
 /tests/test-tokenizer-1-bpe
-/tests/test-rope
-/tests/test-backend-ops
+/tests/test-tokenizer-1-spm
+
+# Scripts
+!/scripts/install-oneapi.bat
diff --git a/llm/llama.cpp/.gitmodules b/llm/llama.cpp/.gitmodules
index b7e8b8ff2f64efced25885bbbd3073f8d3f12ac0..5861d59cb785dabccb2cf01a2af4d519966cdbad 100644
--- a/llm/llama.cpp/.gitmodules
+++ b/llm/llama.cpp/.gitmodules
@@ -1,3 +1,3 @@
 [submodule "kompute"]
-	path = kompute
+	path = ggml/src/kompute
 	url = https://github.com/nomic-ai/kompute.git
diff --git a/llm/llama.cpp/AUTHORS b/llm/llama.cpp/AUTHORS
index b029f13da3b56ba2d500fd5141aa53e6f90673f4..1bd36158a72f4e48d61dbe49ec1923cbf95dd2cd 100644
--- a/llm/llama.cpp/AUTHORS
+++ b/llm/llama.cpp/AUTHORS
@@ -1,8 +1,9 @@
-# date: Tue Apr  9 09:17:14 EEST 2024
+# date: Wed Jun 26 19:36:34 EEST 2024
 # this file is auto-generated by scripts/gen-authors.sh
 
 0cc4m 
 0xspringtime <110655352+0xspringtime@users.noreply.github.com>
+20kdc 
 2f38b454 
 3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com>
 44670 <44670@users.noreply.github.com>
@@ -11,14 +12,18 @@ AT 
 Aarni Koskela 
 Aaron Miller 
 Aaryaman Vasishta 
+Abheek Gulati 
 Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
 Abhishek Gopinath K <31348521+overtunned@users.noreply.github.com>
 Adithya Balaji 
 AdithyanI 
 Adrian 
 Adrian Hesketh 
+Ahmet Zeer 
 AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
 Aisuko 
+Akarshan Biswas 
+Albert Jin 
 Alberto <57916483+albbus-stack@users.noreply.github.com>
 Alex 
 Alex Azarov 
@@ -35,19 +40,24 @@ Ali Nehzat 
 Ali Tariq 
 Alon 
 AlpinDale <52078762+AlpinDale@users.noreply.github.com>
+Amir 
 AmirAli Mirian <37371367+amiralimi@users.noreply.github.com>
 Ananta Bastola 
 Anas Ahouzi <112881240+aahouzi@users.noreply.github.com>
 András Salamon 
 Andrei 
 Andrew Canis 
+Andrew Downing 
 Andrew Duffy 
 Andrew Godfrey 
+Andy Tai 
 Arik Poznanski 
 Artem 
+Artem Zinnatullin 
 Artyom Lebedev 
 Asbjørn Olling 
 Ásgeir Bjarni Ingvarsson 
+Ashish <1856117+ashishdatta@users.noreply.github.com>
 Ashok Gelal <401055+ashokgelal@users.noreply.github.com>
 Ashraful Islam 
 Atsushi Tatsuma 
@@ -57,35 +67,46 @@ BADR 
 Bach Le 
 Bailey Chittle <39804642+bachittle@users.noreply.github.com>
 BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com>
+Bartowski 
 Behnam M <58621210+ibehnam@users.noreply.github.com>
+Ben Ashbaugh 
 Ben Garney 
 Ben Siraphob 
 Ben Williams 
+Benjamin Findley <39356821+Kartoffelsaft@users.noreply.github.com>
 Benjamin Lecaillon <84293038+blecaillon@users.noreply.github.com>
 Bernat Vadell 
+Bingan <70050083+binganao@users.noreply.github.com>
 Bodo Graumann 
 Bono Lv 
 Borislav Stanimirov 
 Branden Butler 
 Brian 
 Bruce MacDonald 
+Bryan Honof 
 CJ Pais 
 CRD716 
+Calvin Laurenson 
 Cameron 
 Cameron Kaiser 
+Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
 Casey Primozic 
 Casey Primozic 
 CausalLM <148736309+CausalLM@users.noreply.github.com>
 Cebtenzzre 
 Chad Brewbaker 
+Chao Jiang 
 Cheng Shao 
+Chris Elrod 
 Chris Kuehl 
 Christian Demsar 
 Christian Demsar 
 Christian Falch <875252+chrfalch@users.noreply.github.com>
 Christian Kögler 
+Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com>
 Clark Saben <76020733+csaben@users.noreply.github.com>
 Clint Herron 
+CrispStrobe <154636388+CrispStrobe@users.noreply.github.com>
 Cuong Trinh Manh 
 DAN™ 
 Damian Stewart 
@@ -95,8 +116,12 @@ Daniel Bevenius 
 Daniel Drake 
 Daniel Hiltgen 
 Daniel Illescas Romero 
+Daniele <57776841+daniandtheweb@users.noreply.github.com>
 DannyDaemonic 
 Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com>
+Dave 
+Dave Airlie 
+Dave Airlie 
 Dave Della Costa 
 David Friehs 
 David Kennedy 
@@ -104,10 +129,13 @@ David Pflug 
 David Renshaw 
 David Sommers <12738+databyte@users.noreply.github.com>
 David Yang 
+Dawid Potocki 
 Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com>
 Dean 
 Deins 
+Deven Mistry <31466137+deven367@users.noreply.github.com>
 Didzis Gosko 
+Djip007 
 Don Mahurin 
 DooWoong Lee (David) 
 Doomsdayrs <38189170+Doomsdayrs@users.noreply.github.com>
@@ -116,8 +144,11 @@ Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
 Ebey Abraham 
 Ed Lee 
 Ed Lepedus 
+Eddie-Wang 
 Edward Taylor 
+Elaine 
 Elbios <141279586+Elbios@users.noreply.github.com>
+Elton Kola 
 Engininja2 <139037756+Engininja2@users.noreply.github.com>
 Equim 
 Eric Sommerlade 
@@ -143,37 +174,47 @@ Firat 
 Folko-Ven <71110216+Folko-Ven@users.noreply.github.com>
 Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com>
 Francisco Melo <43780565+francis2tm@users.noreply.github.com>
+Frank Mai 
 FrankHB 
+Fred Douglas <43351173+fredlas@users.noreply.github.com>
 Frederik Vogel 
 Gabe Goodhart 
 GainLee 
 Galunid 
 Gary Linscott 
 Gary Mulder 
+Gavin Zhao 
 Genkagaku.GPT 
 Georgi Gerganov 
 Gilad S 
+Giuseppe Scrivano 
 GiviMAD 
 Govlzkoy 
 Guillaume "Vermeille" Sanchez 
 Guillaume Wenzek 
 Guoteng <32697156+SolenoidWGT@users.noreply.github.com>
 Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com>
+Haggai Nuchi 
 Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
+Hamdoud Hakem <90524568+hamdoudhakem@users.noreply.github.com>
+HanishKVC 
 Haohui Mai 
 Haoxiang Fei 
 Harald Fernengel 
 Hatsune Miku <129688334+at8u@users.noreply.github.com>
+HatsuneMikuUwU33 <173229399+HatsuneMikuUwU33@users.noreply.github.com>
 Henk Poley 
 Henri Vasserman 
 Henrik Forstén 
 Herman Semenov 
 Hesen Peng 
 Hoang Nguyen 
+Hong Bo PENG 
 Hongyu Ouyang <96765450+casavaca@users.noreply.github.com>
 Howard Su 
 Hua Jiang 
 Huawei Lin 
+Hugo Roussel 
 Ian Bull 
 Ian Bull 
 Ian Scrivener 
@@ -190,8 +231,10 @@ Ivan Stepanov 
 JH23X <165871467+JH23X@users.noreply.github.com>
 Jack Mousseau 
 JackJollimore <130917767+JackJollimore@users.noreply.github.com>
+Jaemin Son 
 Jag Chadha 
 Jakub N 
+James A Capozzoli <157492257+jac-jim@users.noreply.github.com>
 James Reynolds 
 Jan Boon 
 Jan Boon 
@@ -205,12 +248,17 @@ Jean-Michaël Celerier 
 Jed Fox 
 Jeffrey Quesnelle 
 Jesse Jojo Johnson 
+Jeximo 
 Jhen-Jie Hong 
 Jiahao Li 
 Jian Liao 
 JidongZhang-THU <1119708529@qq.com>
 Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com>
 Jiří Podivín <66251151+jpodivin@users.noreply.github.com>
+Jiří Sejkora 
+Joan Fontanals 
+Joan Fontanals 
+Johan 
 Johannes Gäßler 
 Johannes Rudolph 
 John <78893154+cmp-nct@users.noreply.github.com>
@@ -221,15 +269,19 @@ Jonas Wunderlich <32615971+jonas-w@users.noreply.github.com>
 Jorge A <161275481+jorgealias@users.noreply.github.com>
 Jose Maldonado <63384398+yukiteruamano@users.noreply.github.com>
 Joseph Stahl <1269177+josephst@users.noreply.github.com>
+Josh Ramer 
 Joyce 
 Juan Calderon-Perez <835733+gaby@users.noreply.github.com>
 Judd 
 Julius Arkenberg 
 Jun Jie <71215065+junnjiee16@users.noreply.github.com>
+Junyang Lin 
 Juraj Bednar 
 Justin Parker 
 Justin Suess 
+Justina Cho 
 Justine Tunney 
+Justine Tunney 
 Juuso Alasuutari 
 KASR 
 Kamil Tomšík 
@@ -242,6 +294,7 @@ Kawrakow <48489457+ikawrakow@users.noreply.github.com>
 Keiichi Tabata 
 Kenvix ⭐ 
 Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
+Kevin Gibbons 
 Kevin Ji <1146876+kevinji@users.noreply.github.com>
 Kevin Kwok 
 Kevin Lo 
@@ -257,6 +310,7 @@ Laura 
 Lee <44310445+lx200916@users.noreply.github.com>
 Lee Drake 
 Leng Yue 
+Leon Knauer 
 LeonEricsson <70749762+LeonEricsson@users.noreply.github.com>
 Leonardo Neumann 
 Li Tan 
@@ -265,20 +319,26 @@ LoganDark 
 LostRuins <39025047+LostRuins@users.noreply.github.com>
 Luciano 
 Luo Tian 
+Lyle Dean 
 M. Yusuf Sarıgöz 
 Maarten ter Huurne 
 Mack Straight 
 Maël Kerbiriou 
 MaggotHATE 
+Manuel <44313466+makuche@users.noreply.github.com>
 Marc Köhlbrugge 
 Marco Matthies <71844+marcom@users.noreply.github.com>
 Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com>
 Marian Cepok 
 Mark Fairbairn 
 Marko Tasic 
+Markus Tavenrath 
+Martin Delille 
 Martin Krasser 
 Martin Schwaighofer 
 Marvin Gießing 
+Masaya, Kato <62578291+msy-kato@users.noreply.github.com>
+MasterYi1024 <39848311+MasterYi1024@users.noreply.github.com>
 Mateusz Charytoniuk 
 Matheus C. França 
 Matheus Gabriel Alves Silva 
@@ -287,8 +347,11 @@ Mathijs de Bruin 
 Matt Clayton <156335168+mattjcly@users.noreply.github.com>
 Matt Pulver 
 Matteo Boschini <12133566+mbosc@users.noreply.github.com>
+Mattheus Chediak 
 Matthew Tejo 
 Matvey Soloviev 
+Max Krasnyansky 
+Max Krasnyansky 
 Maxime <672982+maximegmd@users.noreply.github.com>
 Maximilian Winter 
 Meng Zhang 
@@ -300,32 +363,41 @@ Michael Kesper 
 Michael Klimenko 
 Michael Podvitskiy 
 Michael Potter 
+Michael de Gans 
 Michaël de Vries 
 Mihai 
 Mike 
+Mikko Juola 
 Minsoo Cheong <54794500+mscheong01@users.noreply.github.com>
 Mirko185 
 Mirror Azure <54669636+MirrorAzure@users.noreply.github.com>
 Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com>
 Mohammadreza Hendiani 
+Mohammadreza Hendiani 
 Murilo Santana 
 Musab Gultekin 
 Nam D. Tran <42194884+namtranase@users.noreply.github.com>
+Nathan Epstein 
 NawafAlansari <72708095+NawafAlansari@users.noreply.github.com>
 Nebula 
+Neo Zhang <14088817+arthw@users.noreply.github.com>
+Neo Zhang 
 Neo Zhang Jianyu 
 Neuman Vong 
 Nexesenex <124105151+Nexesenex@users.noreply.github.com>
 Niall Coates <1349685+Niall-@users.noreply.github.com>
 Nicolai Weitkemper 
+Nicolás Pérez 
 Nigel Bosch 
 Niklas Korz 
+Nikolas <127742645+nneubacher@users.noreply.github.com>
 Nindaleth 
 Oleksandr Nikitin 
 Oleksii Maryshchenko 
 Olivier Chafik 
 Ondřej Čertík 
 Ouadie EL FAROUKI 
+Patrice Ferlet 
 Paul Tsochantaris 
 Pavol Rusnak 
 Pedro Cuenca 
@@ -343,9 +415,14 @@ RJ Adriaansen 
 Radoslav Gerganov 
 Radosław Gryta 
 Rahul Vivek Nair <68507071+RahulVivekNair@users.noreply.github.com>
+Raj Hammeer Singh Hada 
+Ralph Soika 
 Rand Xie 
 Randall Fitzgerald 
 Reinforce-II 
+Ren Xuancheng 
+Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com>
+RhinoDevel 
 Riceball LEE 
 Richard Kiss 
 Richard Roberson 
@@ -373,6 +450,7 @@ Rowan Hart 
 Rune <43761327+Rune-AI@users.noreply.github.com>
 Ryan Landay 
 Ryder Wishart 
+Ryuei 
 Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
 SakuraUmi 
 Salvador E. Tropea 
@@ -386,6 +464,7 @@ SebastianApel <13675545+SebastianApel@users.noreply.github.com>
 Senemu <10880819+Senemu@users.noreply.github.com>
 Sergey Alirzaev 
 Sergio López 
+Sertaç Özercan <852750+sozercan@users.noreply.github.com>
 SeungWon Jeong <65549245+redlion0929@users.noreply.github.com>
 ShadovvBeast 
 Shakhar Dasgupta 
@@ -394,6 +473,7 @@ Shijie <821898965@qq.com>
 Shintarou Okada 
 Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
 Shouzheng Liu 
+Shuichi Tsutsumi 
 Sigbjørn Skjæret 
 Simon Willison 
 Siwen Yu 
@@ -405,11 +485,14 @@ Someone 
 Someone Serge 
 Sourab Mangrulkar <13534540+pacman100@users.noreply.github.com>
 Spencer Sutton 
+Srihari-mcw <96763064+Srihari-mcw@users.noreply.github.com>
 Srinivas Billa 
 Stefan Sydow 
+Steffen Röcker 
 Stephan Walter 
 Stephen Nichols 
 Steve Grubb 
+Steven Prichard 
 Steven Roussey 
 Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
 Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com>
@@ -434,16 +517,19 @@ Tom C 
 Tom Jobbins <784313+TheBloke@users.noreply.github.com>
 Tomas 
 Tomáš Pazdiora 
+Tristan Druyen 
 Tristan Ross 
 Tungsten842 <886724vf@anonaddy.me>
 Tungsten842 
 Tushar 
 UEXTM.com <84163508+uextm@users.noreply.github.com>
+Ulrich Drepper 
 Uzo Nweke 
 Vaibhav Srivastav 
 Val Kharitonov 
 Valentin Konovalov 
 Valentyn Bezshapkin <61702053+valentynbez@users.noreply.github.com>
+Victor Nogueira 
 Victor Z. Peng 
 Vlad 
 Vladimir 
@@ -455,7 +541,9 @@ Weird Constructor 
 Welby Seely 
 Wentai Zhang 
 WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com>
+William Tambellini 
 Willy Tarreau 
+Wouter <9594229+DifferentialityDevelopment@users.noreply.github.com>
 Wu Jian Ping 
 Wu Jian Ping 
 Xiake Sun 
@@ -466,6 +554,8 @@ Xiaoyi Chen 
 Xingchen Song(宋星辰) 
 Xuan Son Nguyen 
 Yann Follet <131855179+YannFollet@users.noreply.github.com>
+Yaroslav 
+Yazan Agha-Schrader 
 Yiming Cui 
 Yishuo Wang 
 Yueh-Po Peng <94939112+y10ab1@users.noreply.github.com>
@@ -477,6 +567,7 @@ Zane Shannon 
 Zay <95888118+isaiahbjork@users.noreply.github.com>
 Zenix 
 Zhang Peiyuan 
+Zheng.Deng <32841220+dengzheng-cloud@users.noreply.github.com>
 ZhouYuChen 
 Ziad Ben Hadj-Alouane 
 Ziang Wu <97337387+ZiangWu-77@users.noreply.github.com>
@@ -484,14 +575,18 @@ Zsapi 
 a-n-n-a-l-e-e <150648636+a-n-n-a-l-e-e@users.noreply.github.com>
 adel boussaken 
 afrideva <95653597+afrideva@users.noreply.github.com>
+agray3 
 akawrykow <142945436+akawrykow@users.noreply.github.com>
 alexpinel <93524949+alexpinel@users.noreply.github.com>
 alonfaraj 
+alwqx 
+amd-lalithnc 
 andrijdavid 
 anon998 <131767832+anon998@users.noreply.github.com>
 anzz1 
 apaz 
 apcameron <37645737+apcameron@users.noreply.github.com>
+arch-btw <57669023+arch-btw@users.noreply.github.com>
 arcrank 
 arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com>
 at8u <129688334+at8u@users.noreply.github.com>
@@ -514,13 +609,17 @@ cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com>
 coezbek 
 comex 
 compilade <113953597+compilade@users.noreply.github.com>
+compilade 
+cpumaxx <163466046+cpumaxx@users.noreply.github.com>
 crasm 
 crasm 
 daboe01 
 david raistrick 
+ddh0 
 ddpasa <112642920+ddpasa@users.noreply.github.com>
 deepdiffuser <112834445+deepdiffuser@users.noreply.github.com>
 divinity76 
+dm4 
 dotpy314 <33351922+dotpy314@users.noreply.github.com>
 drbh 
 ds5t5 <145942675+ds5t5@users.noreply.github.com>
@@ -529,6 +628,7 @@ eastriver 
 ebraminio 
 eiery <19350831+eiery@users.noreply.github.com>
 eric8607242 
+fairydreaming <166155368+fairydreaming@users.noreply.github.com>
 fraxy-v <65565042+fraxy-v@users.noreply.github.com>
 github-actions[bot] 
 gliptic 
@@ -539,6 +639,7 @@ h-h-h-h <13482553+h-h-h-h@users.noreply.github.com>
 hankcs 
 hoangmit 
 hongbo.mo <352280764@qq.com>
+hopkins385 <98618192+hopkins385@users.noreply.github.com>
 howlger 
 howlger 
 hutli <6594598+hutli@users.noreply.github.com>
@@ -549,14 +650,22 @@ hydai 
 iSma 
 iacore <74560659+iacore@users.noreply.github.com>
 igarnier 
+intelmatt <61025942+intelmatt@users.noreply.github.com>
 iohub 
 jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com>
+jaime-m-p <167997752+jaime-m-p@users.noreply.github.com>
 jameswu2014 <545426914@qq.com>
+jiez <373447296@qq.com>
 jneem 
+joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
 johnson442 <56517414+johnson442@users.noreply.github.com>
+jojorne 
 jon-chuang <9093549+jon-chuang@users.noreply.github.com>
 jp-x-g 
+jukofyork <69222624+jukofyork@users.noreply.github.com>
+junchao-loongson <68935141+junchao-loongson@users.noreply.github.com>
 jwj7140 <32943891+jwj7140@users.noreply.github.com>
+k.h.lai 
 kaizau 
 kalomaze <66376113+kalomaze@users.noreply.github.com>
 kang 
@@ -575,11 +684,15 @@ ldwang 
 le.chang 
 leejet 
 limitedAtonement 
+liuwei-git <14815172+liuwei-git@users.noreply.github.com>
 lon <114724657+longregen@users.noreply.github.com>
+loonerin <132926317+loonerin@users.noreply.github.com>
+luoyu-intel 
 m3ndax 
 maddes8cht <55592906+maddes8cht@users.noreply.github.com>
 makomk 
 manikbhandari 
+maor-ps <154728172+maor-ps@users.noreply.github.com>
 mdrokz 
 mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com>
 minarchist 
@@ -593,15 +706,19 @@ ngc92 <7938269+ngc92@users.noreply.github.com>
 nhamanasu <45545786+nhamanasu@users.noreply.github.com>
 niansa/tuxifan 
 niansa/tuxifan 
+nickp27 
 ningshanwutuobang 
 nold 
 nopperl <54780682+nopperl@users.noreply.github.com>
 nusu-github <29514220+nusu-github@users.noreply.github.com>
 olexiyb 
+omahs <73983677+omahs@users.noreply.github.com>
 oobabooga <112222186+oobabooga@users.noreply.github.com>
 opparco 
 ostix360 <55257054+ostix360@users.noreply.github.com>
+pengxin99 
 perserk 
+pmysl 
 postmasters 
 pudepiedj 
 qingfengfenga <41416092+qingfengfenga@users.noreply.github.com>
@@ -614,16 +731,19 @@ rhuddleston 
 rimoliga <53384203+rimoliga@users.noreply.github.com>
 runfuture 
 sandyiscool 
+sasha0552 
 semidark 
 sharpHL <132747147+sharpHL@users.noreply.github.com>
 shibe2 
 singularity <12184989+singularity-s0@users.noreply.github.com>
 sjinzh 
+sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com>
 slaren <2141330+slaren@users.noreply.github.com>
 slaren 
 snadampal <87143774+snadampal@users.noreply.github.com>
 staviq 
 stduhpf 
+strawberrymelonpanda <152940198+strawberrymelonpanda@users.noreply.github.com>
 swittk 
 takov751 <40316768+takov751@users.noreply.github.com>
 tarcey 
@@ -636,12 +756,16 @@ uint256_t 
 uint256_t 
 unbounded 
 valiray <133289098+valiray@users.noreply.github.com>
+vik 
+viric 
 vodkaslime <646329483@qq.com>
 vvhg1 <94630311+vvhg1@users.noreply.github.com>
 vxiiduu <73044267+vxiiduu@users.noreply.github.com>
 wbpxre150 <100937007+wbpxre150@users.noreply.github.com>
 whoreson <139810751+whoreson@users.noreply.github.com>
+woachk <24752637+woachk@users.noreply.github.com>
 wonjun Jang 
+woodx <124784234+woodx9@users.noreply.github.com>
 wzy <32936898+Freed-Wu@users.noreply.github.com>
 xaedes 
 xaedes 
@@ -649,7 +773,10 @@ xloem <0xloem@gmail.com>
 yangli2 
 yuiseki 
 zakkor 
+zhangkaihuo 
 zhouwg <6889919+zhouwg@users.noreply.github.com>
+zhouwg 
 zrm 
+Ștefan-Gabriel Muscalu 
 源文雨 <41315874+fumiama@users.noreply.github.com>
 Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>
diff --git a/llm/llama.cpp/CMakeLists.txt b/llm/llama.cpp/CMakeLists.txt
index 60cf7bdc4b6840409ee3b326c778d5a04903cb18..a313206351677132c1eba90654d761ddbe50bb5e 100644
--- a/llm/llama.cpp/CMakeLists.txt
+++ b/llm/llama.cpp/CMakeLists.txt
@@ -2,6 +2,9 @@ cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
+#set(CMAKE_WARN_DEPRECATED YES)
+set(CMAKE_WARN_UNUSED_CLI YES)
+
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
@@ -9,11 +12,16 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
     set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
 endif()
 
+# Add path to modules
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
+
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
 if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
 
+    include(git-vars)
+
     # configure project version
     # TODO
 else()
@@ -32,1289 +40,133 @@ else()
     endif()
 endif()
 
+option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
 
-#
-# Option list
-#
-
-if (APPLE)
-    set(LLAMA_METAL_DEFAULT ON)
-else()
-    set(LLAMA_METAL_DEFAULT OFF)
+if (WIN32)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 endif()
 
-set(LLAMA_LLAMAFILE_DEFAULT ON)
-
-# general
-option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)
-option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
-option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
-option(LLAMA_CCACHE                     "llama: use ccache if available"                        ON)
+#
+# option list
+#
 
 # debug
-option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"                   ON)
-option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
-option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
+option(LLAMA_ALL_WARNINGS           "llama: enable all compiler warnings"                   ON)
+option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
 
 # build
-option(LLAMA_FATAL_WARNINGS             "llama: enable -Werror flag"                            OFF)
+option(LLAMA_FATAL_WARNINGS "llama: enable -Werror flag" OFF)
 
 # sanitizers
-option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
-option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
-option(LLAMA_SANITIZE_UNDEFINED         "llama: enable undefined sanitizer"                     OFF)
-
-# instruction set specific
-if (LLAMA_NATIVE)
-    set(INS_ENB OFF)
-else()
-    set(INS_ENB ON)
-endif()
+option(LLAMA_SANITIZE_THREAD    "llama: enable thread sanitizer"    OFF)
+option(LLAMA_SANITIZE_ADDRESS   "llama: enable address sanitizer"   OFF)
+option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 
-option(LLAMA_SVE                             "llama: enable SVE"                                OFF)
-option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
-option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
-option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
-option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
-option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
-option(LLAMA_AVX512_BF16                     "llama: enable AVX512-BF16"                        OFF)
-option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
-# in MSVC F16C is implied with AVX2/AVX512
-if (NOT MSVC)
-    option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
-endif()
-
-if (WIN32)
-    set(LLAMA_WIN_VER "0x602" CACHE STRING "llama: Windows Version")
-endif()
+# extra artifacts
+option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 
 # 3rd party libs
-option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
-option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
-option(LLAMA_LLAMAFILE                       "llama: use llamafile SGEMM"                       ${LLAMA_LLAMAFILE_DEFAULT})
-set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
-option(LLAMA_CUDA                            "llama: use CUDA"                                  OFF)
-option(LLAMA_CUBLAS                          "llama: use CUDA (deprecated, use LLAMA_CUDA)"     OFF)
-option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
-option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
-set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
-set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
-option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
-set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
-set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
-                                             "llama: max. batch size for using peer access")
-option(LLAMA_CUDA_NO_PEER_COPY               "llama: do not use peer to peer copies"            OFF)
-option(LLAMA_CUDA_NO_VMM                     "llama: do not try to use CUDA VMM"                OFF)
-
-option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
-option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
-option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
-option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
-option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
-option(LLAMA_VULKAN_CHECK_RESULTS            "llama: run Vulkan op checks"                      OFF)
-option(LLAMA_VULKAN_DEBUG                    "llama: enable Vulkan debug output"                OFF)
-option(LLAMA_VULKAN_VALIDATE                 "llama: enable Vulkan validation"                  OFF)
-option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"                          OFF)
-option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
-option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
-option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
-option(LLAMA_METAL_EMBED_LIBRARY             "llama: embed Metal library"                       OFF)
-set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
-                                             "llama: metal minimum macOS version")
-set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
-option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
-option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
-option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
-option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
-set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
-option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
-set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeline parallelism")
-
-option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
-option(LLAMA_LASX                            "llama: enable lasx"                               ON)
-option(LLAMA_LSX                             "llama: enable lsx"                                ON)
-
-# add perf arguments
-option(LLAMA_PERF                            "llama: enable perf"                               OFF)
+option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
 
 # Required for relocatable CMake package
-include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
-
-#
-# Compile flags
-#
-
-if (LLAMA_SYCL)
-    set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
-
-set(CMAKE_CXX_STANDARD_REQUIRED true)
-set(CMAKE_C_STANDARD 11)
-set(CMAKE_C_STANDARD_REQUIRED true)
-set(THREADS_PREFER_PTHREAD_FLAG ON)
-
-find_package(Threads REQUIRED)
-include(CheckCXXCompilerFlag)
-
-add_compile_definitions(GGML_SCHED_MAX_COPIES=${LLAMA_SCHED_MAX_COPIES})
-
-# enable libstdc++ assertions for debug builds
-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    add_compile_definitions($<$:_GLIBCXX_ASSERTIONS>)
-endif()
-
-if (NOT MSVC)
-    if (LLAMA_SANITIZE_THREAD)
-        add_compile_options(-fsanitize=thread)
-        link_libraries     (-fsanitize=thread)
-    endif()
-
-    if (LLAMA_SANITIZE_ADDRESS)
-        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries     (-fsanitize=address)
-    endif()
-
-    if (LLAMA_SANITIZE_UNDEFINED)
-        add_compile_options(-fsanitize=undefined)
-        link_libraries     (-fsanitize=undefined)
-    endif()
-endif()
-
-if (APPLE AND LLAMA_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate)
-    if (ACCELERATE_FRAMEWORK)
-        message(STATUS "Accelerate framework found")
-
-        add_compile_definitions(GGML_USE_ACCELERATE)
-        add_compile_definitions(ACCELERATE_NEW_LAPACK)
-        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
-    else()
-        message(WARNING "Accelerate framework not found")
-    endif()
-endif()
-
-if (LLAMA_METAL)
-    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-    find_library(METAL_FRAMEWORK    Metal      REQUIRED)
-    find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
-
-    message(STATUS "Metal framework found")
-    set(GGML_HEADERS_METAL ggml-metal.h)
-    set(GGML_SOURCES_METAL ggml-metal.m)
-
-    add_compile_definitions(GGML_USE_METAL)
-    if (LLAMA_METAL_NDEBUG)
-        add_compile_definitions(GGML_METAL_NDEBUG)
-    endif()
-
-    # copy ggml-common.h and ggml-metal.metal to bin directory
-    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
-    configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
-
-    if (LLAMA_METAL_EMBED_LIBRARY)
-        enable_language(ASM)
-        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
-        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-
-        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-
-        # merge ggml-common.h and ggml-metal.metal into a single file
-        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-
-        add_custom_command(
-            OUTPUT ${METALLIB_EMBED_ASM}
-            COMMAND echo "Embedding Metal library"
-            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
-            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
-            DEPENDS ggml-metal.metal ggml-common.h
-            COMMENT "Generate assembly for embedded Metal library"
-        )
-
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
-    else()
-        if (LLAMA_METAL_SHADER_DEBUG)
-            # custom command to do the following:
-            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-            #
-            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-            #       disabling fast math is needed in order to pass tests/test-backend-ops
-            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-            set(XC_FLAGS -fno-fast-math -fno-inline -g)
-        else()
-            set(XC_FLAGS -O3)
-        endif()
-
-        # Append macOS metal versioning flags
-        if (LLAMA_METAL_MACOSX_VERSION_MIN)
-            message(STATUS "Adding -mmacosx-version-min=${LLAMA_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
-            list(APPEND XC_FLAGS -mmacosx-version-min=${LLAMA_METAL_MACOSX_VERSION_MIN})
-        endif()
-        if (LLAMA_METAL_STD)
-            message(STATUS "Adding -std=${LLAMA_METAL_STD} flag to metal compilation")
-            list(APPEND XC_FLAGS -std=${LLAMA_METAL_STD})
-        endif()
-
-        add_custom_command(
-            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-            DEPENDS ggml-metal.metal ggml-common.h
-            COMMENT "Compiling Metal kernels"
-            )
-
-        add_custom_target(
-            ggml-metal ALL
-            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            )
-    endif() # LLAMA_METAL_EMBED_LIBRARY
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
-        ${FOUNDATION_LIBRARY}
-        ${METAL_FRAMEWORK}
-        ${METALKIT_FRAMEWORK}
-        )
-endif()
-
-if (LLAMA_BLAS)
-    if (LLAMA_STATIC)
-        set(BLA_STATIC ON)
-    endif()
-    if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
-        set(BLA_SIZEOF_INTEGER 8)
-    endif()
-
-    set(BLA_VENDOR ${LLAMA_BLAS_VENDOR})
-    find_package(BLAS)
-
-    if (BLAS_FOUND)
-        message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
-
-        if ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
-            # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
-            # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
-            find_package(PkgConfig REQUIRED)
-            if (${LLAMA_BLAS_VENDOR} MATCHES "Generic")
-                pkg_check_modules(DepBLAS REQUIRED blas)
-            elseif (${LLAMA_BLAS_VENDOR} MATCHES "OpenBLAS")
-                # As of openblas v0.3.22, the 64-bit is named openblas64.pc
-                pkg_check_modules(DepBLAS openblas64)
-                if (NOT DepBLAS_FOUND)
-                    pkg_check_modules(DepBLAS REQUIRED openblas)
-                endif()
-            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FLAME")
-                pkg_check_modules(DepBLAS REQUIRED blis)
-            elseif (${LLAMA_BLAS_VENDOR} MATCHES "ATLAS")
-                pkg_check_modules(DepBLAS REQUIRED blas-atlas)
-            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FlexiBLAS")
-                pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
-            elseif (${LLAMA_BLAS_VENDOR} MATCHES "Intel")
-                # all Intel* libraries share the same include path
-                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
-            elseif (${LLAMA_BLAS_VENDOR} MATCHES "NVHPC")
-                # this doesn't provide pkg-config
-                # suggest to assign BLAS_INCLUDE_DIRS on your own
-                if ("${NVHPC_VERSION}" STREQUAL "")
-                    message(WARNING "Better to set NVHPC_VERSION")
-                else()
-                    set(DepBLAS_FOUND ON)
-                    set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
-                endif()
-            endif()
-            if (DepBLAS_FOUND)
-                set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
-            else()
-                message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
-                " detected by pkgconfig, trying to find cblas.h from possible paths...")
-                find_path(BLAS_INCLUDE_DIRS
-                    NAMES cblas.h
-                    HINTS
-                        /usr/include
-                        /usr/local/include
-                        /usr/include/openblas
-                        /opt/homebrew/opt/openblas/include
-                        /usr/local/opt/openblas/include
-                        /usr/include/x86_64-linux-gnu/openblas/include
-                )
-            endif()
-        endif()
-
-        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
-
-        add_compile_options(${BLAS_LINKER_FLAGS})
-
-        add_compile_definitions(GGML_USE_OPENBLAS)
-
-        if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
-            add_compile_definitions(GGML_BLAS_USE_MKL)
-        endif()
-
-        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${BLAS_LIBRARIES})
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
-    else()
-        message(WARNING "BLAS not found, please refer to "
-        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
-        " to set correct LLAMA_BLAS_VENDOR")
-    endif()
-endif()
-
-if (LLAMA_LLAMAFILE)
-    add_compile_definitions(GGML_USE_LLAMAFILE)
-
-    set(GGML_HEADERS_LLAMAFILE sgemm.h)
-    set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
-endif()
-
-if (LLAMA_CUBLAS)
-    message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
-    set(LLAMA_CUDA ON)
-endif()
-
-if (LLAMA_CUDA)
-    cmake_minimum_required(VERSION 3.17)
-
-    find_package(CUDAToolkit)
-    if (CUDAToolkit_FOUND)
-        message(STATUS "CUDA found")
-
-        enable_language(CUDA)
-
-        set(GGML_HEADERS_CUDA ggml-cuda.h)
-
-        file(GLOB GGML_SOURCES_CUDA "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
-
-        add_compile_definitions(GGML_USE_CUDA)
-        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
-        if (LLAMA_CUDA_FORCE_DMMV)
-            add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-        endif()
-        if (LLAMA_CUDA_FORCE_MMQ)
-            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
-        endif()
-        if (LLAMA_CUDA_NO_VMM)
-            add_compile_definitions(GGML_CUDA_NO_VMM)
-        endif()
-        add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-        add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-        if (DEFINED LLAMA_CUDA_DMMV_Y)
-            add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_DMMV_Y}) # for backwards compatibility
-        endif()
-        if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
-            add_compile_definitions(GGML_CUDA_F16)
-        endif()
-        add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
-        if (LLAMA_CUDA_NO_PEER_COPY)
-            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
-        endif()
-
-        if (LLAMA_STATIC)
-            if (WIN32)
-                # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
-                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
-            else ()
-                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
-            endif()
-        else()
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
-        endif()
-
-        if (LLAMA_CUDA_NO_VMM)
-            # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
-        else()
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
-        endif()
-
-    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        # 52 == lowest CUDA 12 standard
-        # 60 == f16 CUDA intrinsics
-        # 61 == integer CUDA intrinsics
-        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-        if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
-        else()
-            set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
-            #set(CMAKE_CUDA_ARCHITECTURES "") # use this to compile much faster, but only F16 models work
-        endif()
-    endif()
-    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
-
-    else()
-        message(WARNING "CUDA not found")
-    endif()
-endif()
-
-if (LLAMA_RPC)
-    add_compile_definitions(GGML_USE_RPC)
-
-    if (WIN32)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ws2_32)
-    endif()
-
-    set(GGML_HEADERS_RPC ggml-rpc.h)
-    set(GGML_SOURCES_RPC ggml-rpc.cpp)
-endif()
-
-if (LLAMA_CLBLAST)
-    find_package(CLBlast)
-    if (CLBlast_FOUND)
-        message(STATUS "CLBlast found")
-
-        set(GGML_HEADERS_OPENCL ggml-opencl.h)
-        set(GGML_SOURCES_OPENCL ggml-opencl.cpp)
-
-        add_compile_definitions(GGML_USE_CLBLAST)
-
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} clblast)
-    else()
-        message(WARNING "CLBlast not found")
-    endif()
-endif()
-
-if (LLAMA_VULKAN)
-    find_package(Vulkan)
-    if (Vulkan_FOUND)
-        message(STATUS "Vulkan found")
-
-        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
-        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
-
-        add_compile_definitions(GGML_USE_VULKAN)
-
-        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
-        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
-        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
-        endif()
-
-        if (LLAMA_VULKAN_CHECK_RESULTS)
-            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
-        endif()
-
-        if (LLAMA_VULKAN_DEBUG)
-            add_compile_definitions(GGML_VULKAN_DEBUG)
-        endif()
-
-        if (LLAMA_VULKAN_VALIDATE)
-            add_compile_definitions(GGML_VULKAN_VALIDATE)
-        endif()
-
-        if (LLAMA_VULKAN_RUN_TESTS)
-            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
-        endif()
-
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
-    else()
-        message(WARNING "Vulkan not found")
-    endif()
-endif()
-
-if (LLAMA_HIPBLAS)
-    if ($ENV{ROCM_PATH})
-        set(ROCM_PATH $ENV{ROCM_PATH})
-    else()
-        set(ROCM_PATH /opt/rocm)
-    endif()
-    list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH})
-
-    # CMake on Windows doesn't support the HIP language yet
-    if(WIN32)
-        set(CXX_IS_HIPCC TRUE)
-    else()
-        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
-    endif()
-
-    if(CXX_IS_HIPCC)
-        if(LINUX)
-            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
-                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
-            endif()
-
-            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
-                    " Prefer setting the HIP compiler directly. See README for details.")
-        endif()
-    else()
-        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
-        if(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
-            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
-        endif()
-        cmake_minimum_required(VERSION 3.21)
-        enable_language(HIP)
-    endif()
-    find_package(hip     REQUIRED)
-    find_package(hipblas REQUIRED)
-    find_package(rocblas REQUIRED)
-
-    message(STATUS "HIP and hipBLAS found")
-
-    set(GGML_HEADERS_ROCM ggml-cuda.h)
-
-    file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
-    list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
-
-    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
-
-    if (LLAMA_HIP_UMA)
-        add_compile_definitions(GGML_HIP_UMA)
-    endif()
-
-    if (LLAMA_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
-    if (LLAMA_CUDA_FORCE_MMQ)
-        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
-    endif()
-
-    if (LLAMA_CUDA_NO_PEER_COPY)
-        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
-    endif()
-
-    add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-
-    if (CXX_IS_HIPCC)
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device)
-    else()
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
-    endif()
-
-    if (LLAMA_STATIC)
-        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-    endif()
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
-endif()
-
-if (LLAMA_SYCL)
-    if (NOT LLAMA_SYCL_TARGET MATCHES "^(INTEL|NVIDIA)$")
-        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL or NVIDIA")
-    endif()
-
-    if ( NOT DEFINED ENV{ONEAPI_ROOT})
-        message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
-    endif()
-    #todo: AOT
-
-    find_package(IntelSYCL REQUIRED)
-
-    message(STATUS "SYCL found")
-
-    add_compile_definitions(GGML_USE_SYCL)
-
-    if (LLAMA_SYCL_F16)
-        add_compile_definitions(GGML_SYCL_F16)
-    endif()
-
-    if (LLAMA_CUDA_FORCE_MMQ)
-        add_compile_definitions(GGML_SYCL_FORCE_MMQ)
-    endif()
-
-    add_compile_options(-I./) #include DPCT
-    add_compile_options(-I/${SYCL_INCLUDE_DIR})
-
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
-    if (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-    endif()
-
-    set(GGML_HEADERS_SYCL ggml-sycl.h)
-    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
-
-    if (WIN32)
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)
-    else()
-        if (LLAMA_SYCL_TARGET STREQUAL "INTEL")
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-        elseif (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl pthread m dl onemkl)
-        endif()
-    endif()
-endif()
-
-if (LLAMA_KOMPUTE)
-    add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
-    find_package(Vulkan COMPONENTS glslc REQUIRED)
-    find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
-    if (NOT glslc_executable)
-        message(FATAL_ERROR "glslc not found")
-    endif()
-
-    function(compile_shader)
-        set(options)
-        set(oneValueArgs)
-        set(multiValueArgs SOURCES)
-        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-        foreach(source ${compile_shader_SOURCES})
-            get_filename_component(filename ${source} NAME)
-            set(spv_file ${filename}.spv)
-            add_custom_command(
-                OUTPUT ${spv_file}
-                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                COMMENT "Compiling ${source} to ${spv_file}"
-                )
-
-            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-            set(FILE_NAME "shader${RAW_FILE_NAME}")
-            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-            if(CMAKE_GENERATOR MATCHES "Visual Studio")
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/$/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$/xxd"
-                    )
-            else()
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-                    )
-            endif()
-        endforeach()
-    endfunction()
-
-    if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
-        message(STATUS "Kompute found")
-        set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
-        add_subdirectory(kompute)
-
-        # Compile our shaders
-        compile_shader(SOURCES
-            kompute-shaders/op_scale.comp
-            kompute-shaders/op_scale_8.comp
-            kompute-shaders/op_add.comp
-            kompute-shaders/op_addrow.comp
-            kompute-shaders/op_mul.comp
-            kompute-shaders/op_silu.comp
-            kompute-shaders/op_relu.comp
-            kompute-shaders/op_gelu.comp
-            kompute-shaders/op_softmax.comp
-            kompute-shaders/op_norm.comp
-            kompute-shaders/op_rmsnorm.comp
-            kompute-shaders/op_diagmask.comp
-            kompute-shaders/op_mul_mat_mat_f32.comp
-            kompute-shaders/op_mul_mat_f16.comp
-            kompute-shaders/op_mul_mat_q8_0.comp
-            kompute-shaders/op_mul_mat_q4_0.comp
-            kompute-shaders/op_mul_mat_q4_1.comp
-            kompute-shaders/op_mul_mat_q6_k.comp
-            kompute-shaders/op_getrows_f16.comp
-            kompute-shaders/op_getrows_q4_0.comp
-            kompute-shaders/op_getrows_q4_1.comp
-            kompute-shaders/op_getrows_q6_k.comp
-            kompute-shaders/op_rope_f16.comp
-            kompute-shaders/op_rope_f32.comp
-            kompute-shaders/op_cpy_f16_f16.comp
-            kompute-shaders/op_cpy_f16_f32.comp
-            kompute-shaders/op_cpy_f32_f16.comp
-            kompute-shaders/op_cpy_f32_f32.comp
-        )
-
-        # Create a custom target for our generated shaders
-        add_custom_target(generated_shaders DEPENDS
-            shaderop_scale.h
-            shaderop_scale_8.h
-            shaderop_add.h
-            shaderop_addrow.h
-            shaderop_mul.h
-            shaderop_silu.h
-            shaderop_relu.h
-            shaderop_gelu.h
-            shaderop_softmax.h
-            shaderop_norm.h
-            shaderop_rmsnorm.h
-            shaderop_diagmask.h
-            shaderop_mul_mat_mat_f32.h
-            shaderop_mul_mat_f16.h
-            shaderop_mul_mat_q8_0.h
-            shaderop_mul_mat_q4_0.h
-            shaderop_mul_mat_q4_1.h
-            shaderop_mul_mat_q6_k.h
-            shaderop_getrows_f16.h
-            shaderop_getrows_q4_0.h
-            shaderop_getrows_q4_1.h
-            shaderop_getrows_q6_k.h
-            shaderop_rope_f16.h
-            shaderop_rope_f32.h
-            shaderop_cpy_f16_f16.h
-            shaderop_cpy_f16_f32.h
-            shaderop_cpy_f32_f16.h
-            shaderop_cpy_f32_f32.h
-        )
-
-        # Create a custom command that depends on the generated_shaders
-        add_custom_command(
-            OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-            COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-            DEPENDS generated_shaders
-            COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
-        )
-
-        # Add the stamp to the main sources to ensure dependency tracking
-        set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-        set(GGML_HEADERS_KOMPUTE ggml-kompute.h   ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-
-        add_compile_definitions(GGML_USE_KOMPUTE)
-
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
-    else()
-        message(WARNING "Kompute not found")
-    endif()
-endif()
-
-if (LLAMA_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 
-    add_compile_definitions(GGML_USE_CPU_HBM)
+# override ggml options
+set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
+set(GGML_SANITIZE_ADDRESS   ${LLAMA_SANITIZE_ADDRESS})
+set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED})
+set(GGML_ALL_WARNINGS       ${LLAMA_ALL_WARNINGS})
+set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
 
-    target_link_libraries(ggml PUBLIC memkind)
+# change the default for these ggml options
+if (NOT DEFINED GGML_LLAMAFILE)
+    set(GGML_LLAMAFILE ON)
 endif()
 
-if (LLAMA_PERF)
-    add_compile_definitions(GGML_PERF)
+if (NOT DEFINED GGML_CUDA_USE_GRAPHS)
+    set(GGML_CUDA_USE_GRAPHS ON)
 endif()
 
-function(get_flags CCID CCVER)
-    set(C_FLAGS "")
-    set(CXX_FLAGS "")
-
-    if (CCID MATCHES "Clang")
-        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
-        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
-
-        if (
-            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
-            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
-        )
-            list(APPEND C_FLAGS -Wdouble-promotion)
-        endif()
-    elseif (CCID STREQUAL "GNU")
-        set(C_FLAGS   -Wdouble-promotion)
-        set(CXX_FLAGS -Wno-array-bounds)
-
-        if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            list(APPEND CXX_FLAGS -Wno-format-truncation)
-        endif()
-        if (CCVER VERSION_GREATER_EQUAL 8.1.0)
-            list(APPEND CXX_FLAGS -Wextra-semi)
-        endif()
+# transition helpers
+function (llama_option_depr TYPE OLD NEW)
+    if (${OLD})
+        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
+        set(${NEW} ON PARENT_SCOPE)
     endif()
-
-    set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
-    set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
 
-if (LLAMA_FATAL_WARNINGS)
-    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        list(APPEND C_FLAGS   -Werror)
-        list(APPEND CXX_FLAGS -Werror)
-    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
-        add_compile_options(/WX)
-    endif()
-endif()
-
-if (LLAMA_ALL_WARNINGS)
-    if (NOT MSVC)
-        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-                                  -Werror=implicit-int -Werror=implicit-function-declaration)
-        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
-
-        list(APPEND C_FLAGS   ${WARNING_FLAGS})
-        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
-
-        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
-
-        add_compile_options("$<$:${C_FLAGS};${GF_C_FLAGS}>"
-                            "$<$:${CXX_FLAGS};${GF_CXX_FLAGS}>")
-    else()
-        # todo : msvc
-        set(C_FLAGS   "")
-        set(CXX_FLAGS "")
-    endif()
-endif()
-
-set(CUDA_CXX_FLAGS "")
-
-if (LLAMA_CUDA)
-    set(CUDA_FLAGS -use_fast_math)
-
-    if (LLAMA_FATAL_WARNINGS)
-        list(APPEND CUDA_FLAGS -Werror all-warnings)
-    endif()
-
-    if (LLAMA_ALL_WARNINGS AND NOT MSVC)
-        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
-        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
-            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
-        endif()
-
-        execute_process(
-            COMMAND ${NVCC_CMD} -Xcompiler --version
-            OUTPUT_VARIABLE CUDA_CCFULLVER
-            ERROR_QUIET
-        )
-
-        if (NOT CUDA_CCFULLVER MATCHES clang)
-            set(CUDA_CCID "GNU")
-            execute_process(
-                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
-                OUTPUT_VARIABLE CUDA_CCVER
-                ERROR_QUIET
-            )
-        else()
-            if (CUDA_CCFULLVER MATCHES Apple)
-                set(CUDA_CCID "AppleClang")
-            else()
-                set(CUDA_CCID "Clang")
-            endif()
-            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
-        endif()
-
-        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
-
-        get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
-    endif()
-
-    if (NOT MSVC)
-        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
-    endif()
-endif()
-
-if (WIN32)
-    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
-
-    if (BUILD_SHARED_LIBS)
-        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
-    endif()
-endif()
-
-if (LLAMA_LTO)
-    include(CheckIPOSupported)
-    check_ipo_supported(RESULT result OUTPUT output)
-    if (result)
-        set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
-    else()
-        message(WARNING "IPO is not supported: ${output}")
-    endif()
-endif()
-
-if (LLAMA_CCACHE)
-    find_program(LLAMA_CCACHE_FOUND ccache)
-    if (LLAMA_CCACHE_FOUND)
-        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
-        set(ENV{CCACHE_SLOPPINESS} time_macros)
-        message(STATUS "ccache found, compilation results will be cached. Disable with LLAMA_CCACHE=OFF.")
-    else()
-        message(STATUS "Warning: ccache not found - consider installing it for faster compilation or disable this warning with LLAMA_CCACHE=OFF")
-    endif ()
-endif()
-
-# this version of Apple ld64 is buggy
-execute_process(
-    COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
-    ERROR_VARIABLE output
-    OUTPUT_QUIET
-)
-
-if (output MATCHES "dyld-1015\.7")
-    add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
-endif()
-
-# Architecture specific
-# TODO: probably these flags need to be tweaked on some architectures
-#       feel free to update the Makefile for your architecture and send a pull request or issue
-message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
-if (MSVC)
-    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
-    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
-else ()
-    set(CMAKE_GENERATOR_PLATFORM_LWR "")
-endif ()
-
-if (NOT MSVC)
-    if (LLAMA_STATIC)
-        add_link_options(-static)
-        if (MINGW)
-            add_link_options(-static-libgcc -static-libstdc++)
-        endif()
-    endif()
-    if (LLAMA_GPROF)
-        add_compile_options(-pg)
-    endif()
-endif()
-
-set(ARCH_FLAGS "")
-
-if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
-    (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-     CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
-    message(STATUS "ARM detected")
-    if (MSVC)
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
-        add_compile_definitions(__ARM_NEON)
-        add_compile_definitions(__ARM_FEATURE_FMA)
-
-        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
-        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
-        check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
-        if (GGML_COMPILER_SUPPORT_DOTPROD)
-            add_compile_definitions(__ARM_FEATURE_DOTPROD)
-        endif ()
-        check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
-        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
-        endif ()
-
-        check_cxx_source_compiles("#include \nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-        endif ()
-        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
-    else()
-        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
-        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
-            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
-            # Raspberry Pi 1, Zero
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                # Android armeabi-v7a
-                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-            else()
-                # Raspberry Pi 2
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
-            endif()
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-            # Android arm64-v8a
-            # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            list(APPEND ARCH_FLAGS -mno-unaligned-access)
-        endif()
-        if (LLAMA_SVE)
-            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
-        endif()
-    endif()
-elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
-        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
-    message(STATUS "x86 detected")
-    if (MSVC)
-        # instruction set detection for MSVC only
-        if (LLAMA_NATIVE)
-            include(cmake/FindSIMD.cmake)
-        endif ()
-        if (LLAMA_AVX512)
-            list(APPEND ARCH_FLAGS /arch:AVX512)
-            # MSVC has no compile-time flags enabling specific
-            # AVX512 extensions, neither it defines the
-            # macros corresponding to the extensions.
-            # Do it manually.
-            if (LLAMA_AVX512_VBMI)
-                add_compile_definitions($<$:__AVX512VBMI__>)
-                add_compile_definitions($<$:__AVX512VBMI__>)
-            endif()
-            if (LLAMA_AVX512_VNNI)
-                add_compile_definitions($<$:__AVX512VNNI__>)
-                add_compile_definitions($<$:__AVX512VNNI__>)
-            endif()
-            if (LLAMA_AVX512_BF16)
-                add_compile_definitions($<$:__AVX512BF16__>)
-                add_compile_definitions($<$:__AVX512BF16__>)
-            endif()
-        elseif (LLAMA_AVX2)
-            list(APPEND ARCH_FLAGS /arch:AVX2)
-        elseif (LLAMA_AVX)
-            list(APPEND ARCH_FLAGS /arch:AVX)
-        endif()
-    else()
-        if (LLAMA_NATIVE)
-            list(APPEND ARCH_FLAGS -march=native)
-        endif()
-        if (LLAMA_F16C)
-            list(APPEND ARCH_FLAGS -mf16c)
-        endif()
-        if (LLAMA_FMA)
-            list(APPEND ARCH_FLAGS -mfma)
-        endif()
-        if (LLAMA_AVX)
-            list(APPEND ARCH_FLAGS -mavx)
-        endif()
-        if (LLAMA_AVX2)
-            list(APPEND ARCH_FLAGS -mavx2)
-        endif()
-        if (LLAMA_AVX512)
-            list(APPEND ARCH_FLAGS -mavx512f)
-            list(APPEND ARCH_FLAGS -mavx512bw)
-        endif()
-        if (LLAMA_AVX512_VBMI)
-            list(APPEND ARCH_FLAGS -mavx512vbmi)
-        endif()
-        if (LLAMA_AVX512_VNNI)
-            list(APPEND ARCH_FLAGS -mavx512vnni)
-        endif()
-        if (LLAMA_AVX512_BF16)
-            list(APPEND ARCH_FLAGS -mavx512bf16)
-        endif()
-    endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
-    message(STATUS "PowerPC detected")
-    if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
-    else()
-        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
-        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
-    endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-    message(STATUS "loongarch64 detected")
-
-    list(APPEND ARCH_FLAGS -march=loongarch64)
-    if (LLAMA_LASX)
-        list(APPEND ARCH_FLAGS -mlasx)
-    endif()
-    if (LLAMA_LSX)
-        list(APPEND ARCH_FLAGS -mlsx)
-    endif()
-
-else()
-    message(STATUS "Unknown architecture")
-endif()
-
-add_compile_options("$<$:${ARCH_FLAGS}>")
-add_compile_options("$<$:${ARCH_FLAGS}>")
-
-if (LLAMA_CUDA)
-    list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
-    list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
-    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
-        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
-    endif()
-    add_compile_options("$<$:${CUDA_FLAGS}>")
-endif()
-
-if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
-    add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER})
-endif()
+llama_option_depr(FATAL_ERROR LLAMA_CUBLAS              GGML_CUDA)
+llama_option_depr(WARNING     LLAMA_CUDA                GGML_CUDA)
+llama_option_depr(WARNING     LLAMA_KOMPUTE             GGML_KOMPUTE)
+llama_option_depr(WARNING     LLAMA_METAL               GGML_METAL)
+llama_option_depr(WARNING     LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
+llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
+llama_option_depr(WARNING     LLAMA_RPC                 GGML_RPC)
+llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
+llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
+llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
 
 #
-# POSIX conformance
+# build the library
 #
 
-# clock_gettime came in POSIX.1b (1993)
-# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
-# posix_memalign came in POSIX.1-2001 / SUSv3
-# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-add_compile_definitions(_XOPEN_SOURCE=600)
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
-    remove_definitions(-D_XOPEN_SOURCE=600)
-    add_compile_definitions(_XOPEN_SOURCE=700)
-endif()
-
-# Data types, macros and functions related to controlling CPU affinity and
-# some memory allocation are available on Linux through GNU extensions in libc
-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    add_compile_definitions(_GNU_SOURCE)
-endif()
-
-# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
-# and on macOS its availability depends on enabling Darwin extensions
-# similarly on DragonFly, enabling BSD extensions is necessary
-if (
-    CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
-    CMAKE_SYSTEM_NAME MATCHES "iOS" OR
-    CMAKE_SYSTEM_NAME MATCHES "tvOS" OR
-    CMAKE_SYSTEM_NAME MATCHES "DragonFly"
-)
-    add_compile_definitions(_DARWIN_C_SOURCE)
-endif()
-
-# alloca is a non-standard interface that is not visible on BSDs when
-# POSIX conformance is specified, but not all of them provide a clean way
-# to enable it in such cases
-if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
-    add_compile_definitions(__BSD_VISIBLE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
-    add_compile_definitions(_NETBSD_SOURCE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
-    add_compile_definitions(_BSD_SOURCE)
+if (NOT TARGET ggml)
+    add_subdirectory(ggml)
+    # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
+add_subdirectory(src)
 
 #
-# libraries
+# install
 #
 
-# ggml
-
-add_library(ggml OBJECT
-            ggml.c
-            ggml.h
-            ggml-alloc.c
-            ggml-alloc.h
-            ggml-backend.c
-            ggml-backend.h
-            ggml-quants.c
-            ggml-quants.h
-            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
-            ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
-            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
-            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
-            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
-            ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
-            ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
-            ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
-            )
-
-target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
-target_compile_features   (ggml PUBLIC c_std_11) # don't bump
-
-target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
-
-add_library(ggml_static STATIC $)
-
-if (BUILD_SHARED_LIBS)
-    set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    add_library(ggml_shared SHARED $)
-    target_link_libraries(ggml_shared PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
-    install(TARGETS ggml_shared LIBRARY)
-endif()
-
-# llama
-
-add_library(llama
-            llama.cpp
-            llama.h
-            unicode.h
-            unicode.cpp
-            unicode-data.cpp
-            )
-
-target_include_directories(llama PUBLIC .)
-target_compile_features   (llama PUBLIC cxx_std_11) # don't bump
+include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
 
-target_link_libraries(llama PRIVATE
-    ggml
-    ${LLAMA_EXTRA_LIBS}
-    )
+set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
+set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
 
-if (BUILD_SHARED_LIBS)
-    set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
-    if (LLAMA_METAL)
-        set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    endif()
-endif()
+set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
+set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
+set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
 
+# At the moment some compile definitions are placed within the ggml/src
+# directory but not exported on the `ggml` target. This could be improved by
+# determining _precisely_ which defines are necessary for the llama-config
+# package.
 #
-# install
-#
+get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
+get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
+get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
+set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
+get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
 
-include(GNUInstallDirs)
-include(CMakePackageConfigHelpers)
-
-set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR}
-    CACHE PATH "Location of header files")
-set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR}
-    CACHE PATH "Location of library files")
-set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR}
-    CACHE PATH "Location of binary files")
-set(LLAMA_BUILD_NUMBER ${BUILD_NUMBER})
-set(LLAMA_BUILD_COMMIT ${BUILD_COMMIT})
-set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
-get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
+set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
+install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 configure_package_config_file(
-        ${CMAKE_CURRENT_SOURCE_DIR}/scripts/LlamaConfig.cmake.in
-        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
-    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama
+        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/llama-config.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama
     PATH_VARS LLAMA_INCLUDE_INSTALL_DIR
               LLAMA_LIB_INSTALL_DIR
               LLAMA_BIN_INSTALL_DIR )
 
 write_basic_package_version_file(
-        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
+        ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
     VERSION ${LLAMA_INSTALL_VERSION}
     COMPATIBILITY SameMajorVersion)
 
-install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
-              ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
-        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
-
-set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
-        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_EXTRA}")
-
-set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-install(TARGETS ggml PUBLIC_HEADER)
-
-set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/llama.h)
-install(TARGETS llama LIBRARY PUBLIC_HEADER)
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)
 
 install(
-    FILES convert-hf-to-gguf.py
+    FILES convert_hf_to_gguf.py
     PERMISSIONS
         OWNER_READ
         OWNER_WRITE
@@ -1324,22 +176,13 @@ install(
         WORLD_READ
         WORLD_EXECUTE
     DESTINATION ${CMAKE_INSTALL_BINDIR})
-if (LLAMA_METAL)
-    install(
-        FILES ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
-    if (NOT LLAMA_METAL_EMBED_LIBRARY)
-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
-    endif()
-endif()
+
+configure_file(cmake/llama.pc.in
+        "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
+        @ONLY)
+
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
+        DESTINATION lib/pkgconfig)
 
 #
 # programs, examples and tests
diff --git a/llm/llama.cpp/CMakePresets.json b/llm/llama.cpp/CMakePresets.json
index e2b7a79e371bf8dd5cff743aa798c196f85f8927..bdad38952d3cbe0aee58cbb1ff3d73f1daf87545 100644
--- a/llm/llama.cpp/CMakePresets.json
+++ b/llm/llama.cpp/CMakePresets.json
@@ -11,10 +11,23 @@
             "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
         }
     },
-
+    {
+        "name": "sycl-base",
+        "hidden": true,
+        "generator": "Ninja",
+        "binaryDir": "${sourceDir}/build-${presetName}",
+        "cacheVariables": {
+            "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
+            "CMAKE_CXX_COMPILER": "icx",
+            "CMAKE_C_COMPILER": "cl",
+            "GGML_SYCL": "ON",
+            "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
+        }
+    },
     { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
-    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
-    { "name": "static",  "hidden": true, "cacheVariables": { "LLAMA_STATIC": "ON" } },
+    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
+    { "name": "reldbg",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "static",  "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
 
     {
         "name": "arm64-windows-msvc", "hidden": true,
@@ -35,15 +48,18 @@
     },
 
     { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
-    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "release" ] },
-    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "release", "static" ] },
+    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg" ] },
+    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg", "static" ] },
 
     { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
-    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "release" ] },
-    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] },
+    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
+    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
 
     { "name": "x64-windows-msvc-debug"  , "inherits": [ "base", "debug"   ] },
-    { "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
-    { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
+    { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
+    { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
+
+    { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
+    { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] }
   ]
 }
diff --git a/llm/llama.cpp/CONTRIBUTING.md b/llm/llama.cpp/CONTRIBUTING.md
new file mode 100644
index 0000000000000000000000000000000000000000..a9e000e5227d9336309aa09ab7c4a7423265e434
--- /dev/null
+++ b/llm/llama.cpp/CONTRIBUTING.md
@@ -0,0 +1,29 @@
+# Pull requests (for contributors)
+
+- Test your changes:
+  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
+  - Execute [the full CI locally on your machine](ci/README.md) before publishing
+- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
+  - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
+- Consider allowing write access to your branch for faster review
+- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
+
+# Pull requests (for collaborators)
+
+- Squash-merge PRs
+- Use the following format for the squashed commit title: ` :  (#)`. For example: `utils : fix typo in utils.py (#1234)`
+- Optionally, pick a `` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
+
+# Coding guidelines
+
+- Avoid adding third-party dependencies, extra files, extra headers, etc.
+- Always consider cross-compatibility with other operating systems and architectures
+- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
+- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
+- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
+- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
+- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
+- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
+
+![matmul](media/matmul.png)
+
diff --git a/llm/llama.cpp/Makefile b/llm/llama.cpp/Makefile
index 5caf31cdf37374c3ed6003b0c6eddbde479d10b3..f4ce4f1fbcda12e0ab85a8ec05d316cd95298d4c 100644
--- a/llm/llama.cpp/Makefile
+++ b/llm/llama.cpp/Makefile
@@ -1,13 +1,50 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
-	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama beam-search  \
-	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
+	libllava.a \
+	llama-baby-llama \
+	llama-batched \
+	llama-batched-bench \
+	llama-bench \
+	llama-benchmark-matmult \
+	llama-cli \
+	llama-convert-llama2c-to-ggml \
+	llama-embedding \
+	llama-eval-callback \
+	llama-export-lora \
+	llama-gbnf-validator \
+	llama-gguf \
+	llama-gguf-hash \
+	llama-gguf-split \
+	llama-gritlm \
+	llama-imatrix \
+	llama-infill \
+	llama-llava-cli \
+	llama-lookahead \
+	llama-lookup \
+	llama-lookup-create \
+	llama-lookup-merge \
+	llama-lookup-stats \
+	llama-parallel \
+	llama-passkey \
+	llama-perplexity \
+	llama-q8dot \
+	llama-quantize \
+	llama-quantize-stats \
+	llama-retrieval \
+	llama-save-load-state \
+	llama-server \
+	llama-simple \
+	llama-speculative \
+	llama-tokenize \
+	llama-vdot \
+	llama-cvector-generator \
+	tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
 	tests/test-autorelease \
 	tests/test-backend-ops \
+	tests/test-chat-template \
 	tests/test-double-float \
 	tests/test-grad0 \
 	tests/test-grammar-integration \
@@ -24,8 +61,89 @@ TEST_TARGETS = \
 	tests/test-tokenizer-1-bpe \
 	tests/test-tokenizer-1-spm
 
-# Code coverage output files
-COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
+# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
+LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
+	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \
+	retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm
+
+# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
+#  We don't want to clutter things too much, so we only build replacements for the most commonly used binaries.
+LEGACY_TARGETS_BUILD = main quantize perplexity embedding server
+
+# Deprecation aliases
+ifdef LLAMA_CUBLAS
+$(error LLAMA_CUBLAS is removed. Use GGML_CUDA instead.)
+endif
+
+ifdef LLAMA_CUDA
+GGML_CUDA := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_KOMPUTE
+GGML_KOMPUTE := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_METAL
+GGML_METAL := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_OPENMP
+GGML_OPENMP := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_RPC
+GGML_RPC := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_SYCL
+GGML_SYCL := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_SYCL_F16
+GGML_SYCL_F16 := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_OPENBLAS
+GGML_OPENBLAS := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_OPENBLAS64
+GGML_OPENBLAS64 := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_BLIS
+GGML_BLIS := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_NO_LLAMAFILE
+GGML_NO_LLAMAFILE := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_NO_ACCELERATE
+GGML_NO_ACCELERATE := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_NO_OPENMP
+GGML_NO_OPENMP := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef LLAMA_NO_METAL
+GGML_NO_METAL := 1
+DEPRECATE_WARNING := 1
+endif
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -53,10 +171,12 @@ endif
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
-	ifndef LLAMA_NO_METAL
-		LLAMA_METAL := 1
+	ifndef GGML_NO_METAL
+		GGML_METAL := 1
 	endif
 
+	GGML_NO_OPENMP := 1
+
 	ifneq ($(UNAME_P),arm)
 		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
 		ifeq ($(SYSCTL_M),1)
@@ -67,7 +187,19 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
-default: $(BUILD_TARGETS)
+ifdef GGML_METAL
+	GGML_METAL_EMBED_LIBRARY := 1
+endif
+
+ifdef GGML_RPC
+	BUILD_TARGETS += rpc-server
+endif
+
+ifdef GGML_VULKAN
+	BUILD_TARGETS += vulkan-shaders-gen
+endif
+
+default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD)
 
 test: $(TEST_TARGETS)
 	@failures=0; \
@@ -102,19 +234,7 @@ test: $(TEST_TARGETS)
 	fi
 	@echo 'All tests passed.'
 
-all: $(BUILD_TARGETS) $(TEST_TARGETS)
-
-coverage: ## Run code coverage
-	gcov -pb tests/*.cpp
-
-lcov-report: coverage ## Generate lcov report
-	mkdir -p lcov-report
-	lcov --capture --directory . --output-file lcov-report/coverage.info
-	genhtml lcov-report/coverage.info --output-directory lcov-report
-
-gcovr-report: coverage ## Generate gcovr report
-	mkdir -p gcovr-report
-	gcovr --root . --html --html-details --output gcovr-report/coverage.html
+all: $(BUILD_TARGETS) $(TEST_TARGETS) $(LEGACY_TARGETS_BUILD)
 
 ifdef RISCV_CROSS_COMPILE
 CC	:= riscv64-unknown-linux-gnu-gcc
@@ -126,33 +246,27 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -I. -Icommon
+MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon
 MK_CFLAGS    = -std=c11   -fPIC
 MK_CXXFLAGS  = -std=c++11 -fPIC
 MK_NVCCFLAGS = -std=c++11
 
-# -Ofast tends to produce faster code, but may not be available for some compilers.
-ifdef LLAMA_FAST
-MK_CFLAGS     += -Ofast
-HOST_CXXFLAGS += -Ofast
-MK_NVCCFLAGS  += -O3
-else
-MK_CFLAGS     += -O3
-MK_CXXFLAGS   += -O3
-MK_NVCCFLAGS  += -O3
+ifdef LLAMA_NO_CCACHE
+GGML_NO_CCACHE := 1
+DEPRECATE_WARNING := 1
 endif
 
-ifndef LLAMA_NO_CCACHE
+ifndef GGML_NO_CCACHE
 CCACHE := $(shell which ccache)
 ifdef CCACHE
 export CCACHE_SLOPPINESS = time_macros
-$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.)
+$(info I ccache found, compilation results will be cached. Disable with GGML_NO_CCACHE.)
 CC    := $(CCACHE) $(CC)
 CXX   := $(CCACHE) $(CXX)
 else
 $(info I ccache not found. Consider installing it for faster compilation.)
 endif # CCACHE
-endif # LLAMA_NO_CCACHE
+endif # GGML_NO_CCACHE
 
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
@@ -196,20 +310,24 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
 
-ifdef LLAMA_SCHED_MAX_COPIES
-	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
+ifdef GGML_SCHED_MAX_COPIES
+	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(GGML_SCHED_MAX_COPIES)
 endif
 
 ifdef LLAMA_DEBUG
-	MK_CFLAGS   += -O0 -g
-	MK_CXXFLAGS += -O0 -g
-	MK_LDFLAGS  += -g
+	MK_CFLAGS    += -O0 -g
+	MK_CXXFLAGS  += -O0 -g
+	MK_LDFLAGS   += -g
+	MK_NVCCFLAGS += -O0 -g
 
 	ifeq ($(UNAME_S),Linux)
 		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
 	endif
 else
-	MK_CPPFLAGS += -DNDEBUG
+	MK_CPPFLAGS   += -DNDEBUG
+	MK_CFLAGS     += -O3 -g
+	MK_CXXFLAGS   += -O3 -g
+	MK_NVCCFLAGS  += -O3 -g
 endif
 
 ifdef LLAMA_SANITIZE_THREAD
@@ -239,19 +357,31 @@ ifdef LLAMA_SERVER_SSL
 	MK_LDFLAGS += -lssl -lcrypto
 endif
 
-ifdef LLAMA_CODE_COVERAGE
-	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
-endif
-
 ifdef LLAMA_DISABLE_LOGS
 	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
 endif # LLAMA_DISABLE_LOGS
 
 # warnings
-WARN_FLAGS    = -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
-MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int \
-				-Werror=implicit-function-declaration
-MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
+WARN_FLAGS = \
+	-Wall \
+	-Wextra \
+	-Wpedantic \
+	-Wcast-qual \
+	-Wno-unused-function
+
+MK_CFLAGS += \
+	$(WARN_FLAGS) \
+	-Wshadow \
+	-Wstrict-prototypes \
+	-Wpointer-arith \
+	-Wmissing-prototypes \
+	-Werror=implicit-int \
+	-Werror=implicit-function-declaration
+
+MK_CXXFLAGS += \
+	$(WARN_FLAGS) \
+	-Wmissing-declarations \
+	-Wmissing-noreturn
 
 ifeq ($(LLAMA_FATAL_WARNINGS),1)
 	MK_CFLAGS   += -Werror
@@ -296,9 +426,6 @@ ifdef LLAMA_GPROF
 	MK_CFLAGS   += -pg
 	MK_CXXFLAGS += -pg
 endif
-ifdef LLAMA_PERF
-	MK_CPPFLAGS += -DGGML_PERF
-endif
 
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
@@ -389,247 +516,428 @@ else
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifndef LLAMA_NO_ACCELERATE
+ifndef GGML_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
-		MK_CPPFLAGS += -DGGML_USE_ACCELERATE
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
 		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
 		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
 		MK_LDFLAGS  += -framework Accelerate
+		OBJ_GGML    += ggml/src/ggml-blas.o
 	endif
-endif # LLAMA_NO_ACCELERATE
+endif # GGML_NO_ACCELERATE
 
-ifdef LLAMA_OPENBLAS
-	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
+ifdef GGML_MUSA
+	CC := clang
+	CXX := clang++
+	GGML_CUDA := 1
+	MK_CPPFLAGS += -DGGML_USE_MUSA
+endif
+
+ifndef GGML_NO_OPENMP
+	MK_CPPFLAGS += -DGGML_USE_OPENMP
+	MK_CFLAGS   += -fopenmp
+	MK_CXXFLAGS += -fopenmp
+	ifdef GGML_MUSA
+		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
+		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
+	endif # GGML_MUSA
+endif # GGML_NO_OPENMP
+
+ifdef GGML_OPENBLAS
+	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
-endif # LLAMA_OPENBLAS
+	OBJ_GGML    += ggml/src/ggml-blas.o
+endif # GGML_OPENBLAS
+
+ifdef GGML_OPENBLAS64
+	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas64)
+	MK_LDFLAGS  += $(shell pkg-config --libs openblas64)
+	OBJ_GGML    += ggml/src/ggml-blas.o
+endif # GGML_OPENBLAS64
+
+ifdef GGML_BLIS
+	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
+	MK_LDFLAGS  += -lblis -L/usr/local/lib
+	OBJ_GGML    += ggml/src/ggml-blas.o
+endif # GGML_BLIS
 
-ifndef LLAMA_NO_LLAMAFILE
+ifdef GGML_NVPL
+	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
+	MK_LDFLAGS  += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
+	OBJ_GGML    += ggml/src/ggml-blas.o
+endif # GGML_NVPL
+
+ifndef GGML_NO_LLAMAFILE
 	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
-	OBJS        += sgemm.o
+	OBJ_GGML    += ggml/src/llamafile/sgemm.o
 endif
 
-ifdef LLAMA_BLIS
-	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
-	MK_LDFLAGS  += -lblis -L/usr/local/lib
-endif # LLAMA_BLIS
+ifdef GGML_RPC
+	MK_CPPFLAGS += -DGGML_USE_RPC
+	OBJ_GGML    += ggml/src/ggml-rpc.o
+endif # GGML_RPC
 
-ifdef LLAMA_CUBLAS
-# LLAMA_CUBLAS is deprecated and will be removed in the future
-	LLAMA_CUDA := 1
-endif
+OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
+OBJ_CUDA_TMPL     += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/mmq*.cu))
 
-ifdef LLAMA_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
+ifdef GGML_CUDA_FA_ALL_QUANTS
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*.cu))
+else
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu))
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu))
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*f16-f16.cu))
+endif # GGML_CUDA_FA_ALL_QUANTS
+
+ifdef GGML_CUDA
+	ifdef GGML_MUSA
+		ifneq ('', '$(wildcard /opt/musa)')
+			CUDA_PATH ?= /opt/musa
+		else
+			CUDA_PATH ?= /usr/local/musa
+		endif
+
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
+		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
+		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22
 	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
-	OBJS         += ggml-cuda.o
-	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
-	MK_NVCCFLAGS += -use_fast_math
+		ifneq ('', '$(wildcard /opt/cuda)')
+			CUDA_PATH ?= /opt/cuda
+		else
+			CUDA_PATH ?= /usr/local/cuda
+		endif
+
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+		MK_NVCCFLAGS += -use_fast_math
+	endif # GGML_MUSA
+
+	OBJ_GGML += ggml/src/ggml-cuda.o
+	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML += $(OBJ_CUDA_TMPL)
+
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
 endif # LLAMA_FATAL_WARNINGS
+
+ifndef GGML_MUSA
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
+endif # GGML_MUSA
+
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
 endif # LLAMA_DEBUG
-ifdef LLAMA_CUDA_DEBUG
+
+ifdef GGML_CUDA_DEBUG
 	MK_NVCCFLAGS += --device-debug
-endif # LLAMA_CUDA_DEBUG
-ifdef LLAMA_CUDA_NVCC
-	NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
+endif # GGML_CUDA_DEBUG
+
+ifdef GGML_CUDA_NVCC
+	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
 else
-	NVCC = $(CCACHE) nvcc
-endif #LLAMA_CUDA_NVCC
+	ifdef GGML_MUSA
+		NVCC = $(CCACHE) mcc
+	else
+		NVCC = $(CCACHE) nvcc
+	endif # GGML_MUSA
+endif # GGML_CUDA_NVCC
+
 ifdef CUDA_DOCKER_ARCH
 	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
 else ifndef CUDA_POWER_ARCH
 	MK_NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
-ifdef LLAMA_CUDA_FORCE_DMMV
+
+ifdef GGML_CUDA_FORCE_DMMV
 	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
-endif # LLAMA_CUDA_FORCE_DMMV
-ifdef LLAMA_CUDA_FORCE_MMQ
+endif # GGML_CUDA_FORCE_DMMV
+
+ifdef GGML_CUDA_FORCE_MMQ
 	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # LLAMA_CUDA_FORCE_MMQ
-ifdef LLAMA_CUDA_DMMV_X
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+endif # GGML_CUDA_FORCE_MMQ
+
+ifdef GGML_CUDA_FORCE_CUBLAS
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_CUBLAS
+endif # GGML_CUDA_FORCE_CUBLAS
+
+ifdef GGML_CUDA_DMMV_X
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
-endif # LLAMA_CUDA_DMMV_X
-ifdef LLAMA_CUDA_MMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
-else ifdef LLAMA_CUDA_DMMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
+endif # GGML_CUDA_DMMV_X
+
+ifdef GGML_CUDA_MMV_Y
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
+else ifdef GGML_CUDA_DMMV_Y
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_DMMV_Y) # for backwards compatibility
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
-endif # LLAMA_CUDA_MMV_Y
-ifdef LLAMA_CUDA_F16
+endif # GGML_CUDA_MMV_Y
+
+ifdef GGML_CUDA_F16
 	MK_NVCCFLAGS += -DGGML_CUDA_F16
-endif # LLAMA_CUDA_F16
-ifdef LLAMA_CUDA_DMMV_F16
+endif # GGML_CUDA_F16
+
+ifdef GGML_CUDA_DMMV_F16
 	MK_NVCCFLAGS += -DGGML_CUDA_F16
-endif # LLAMA_CUDA_DMMV_F16
-ifdef LLAMA_CUDA_KQUANTS_ITER
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+endif # GGML_CUDA_DMMV_F16
+
+ifdef GGML_CUDA_KQUANTS_ITER
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
 else
 	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
-ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+
+ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
-endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-ifdef LLAMA_CUDA_NO_PEER_COPY
+endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
+
+ifdef GGML_CUDA_NO_PEER_COPY
 	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # LLAMA_CUDA_NO_PEER_COPY
-ifdef LLAMA_CUDA_CCBIN
-	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
-endif
+endif # GGML_CUDA_NO_PEER_COPY
+
+ifdef GGML_CUDA_CCBIN
+	MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN)
+endif # GGML_CUDA_CCBIN
+
+ifdef GGML_CUDA_FA_ALL_QUANTS
+	MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
+endif # GGML_CUDA_FA_ALL_QUANTS
 
 ifdef JETSON_EOL_MODULE_DETECT
 define NVCC_COMPILE
 	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
 else
+	ifdef GGML_MUSA
+define NVCC_COMPILE
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
+endef # NVCC_COMPILE
+	else
 define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
+	endif # GGML_MUSA
 endif # JETSON_EOL_MODULE_DETECT
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+ggml/src/ggml-cuda/%.o: \
+	ggml/src/ggml-cuda/%.cu \
+	ggml/include/ggml.h \
+	ggml/src/ggml-common.h \
+	ggml/src/ggml-cuda/common.cuh
 	$(NVCC_COMPILE)
 
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+ggml/src/ggml-cuda.o: \
+	ggml/src/ggml-cuda.cu \
+	ggml/include/ggml-cuda.h \
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h \
+	ggml/src/ggml-backend-impl.h \
+	ggml/src/ggml-common.h \
+	$(wildcard ggml/src/ggml-cuda/*.cuh)
 	$(NVCC_COMPILE)
-endif # LLAMA_CUDA
+endif # GGML_CUDA
 
-ifdef LLAMA_CLBLAST
-	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
-	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
-	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
+ifdef GGML_VULKAN
+	MK_CPPFLAGS += -DGGML_USE_VULKAN
+	MK_LDFLAGS  += $(shell pkg-config --libs vulkan)
+	OBJ_GGML    += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
 
-	# Mac provides OpenCL as a framework
-	ifeq ($(UNAME_S),Darwin)
-		MK_LDFLAGS += -lclblast -framework OpenCL
-	else
-		MK_LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
-	endif
-	OBJS    += ggml-opencl.o
-
-ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # LLAMA_CLBLAST
-
-ifdef LLAMA_VULKAN
-	MK_CPPFLAGS  += -DGGML_USE_VULKAN
-	MK_LDFLAGS += -lvulkan
-	OBJS    += ggml-vulkan.o
-
-ifdef LLAMA_VULKAN_CHECK_RESULTS
+ifdef GGML_VULKAN_CHECK_RESULTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
 endif
 
-ifdef LLAMA_VULKAN_DEBUG
+ifdef GGML_VULKAN_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
 endif
 
-ifdef LLAMA_VULKAN_VALIDATE
+ifdef GGML_VULKAN_MEMORY_DEBUG
+	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
+endif
+
+ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
 
-ifdef LLAMA_VULKAN_RUN_TESTS
+ifdef GGML_VULKAN_RUN_TESTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
 endif
 
-ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # LLAMA_VULKAN
+GLSLC_CMD  = glslc
+_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen
+_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp
+_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp
+_ggml_vk_input_dir = ggml/src/vulkan-shaders
+_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
+
+ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
+	$(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@
+
+$(_ggml_vk_header): $(_ggml_vk_source)
+
+$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen
+	$(_ggml_vk_genshaders_cmd) \
+		--glslc      $(GLSLC_CMD) \
+		--input-dir  $(_ggml_vk_input_dir) \
+		--target-hpp $(_ggml_vk_header) \
+		--target-cpp $(_ggml_vk_source)
 
-ifdef LLAMA_HIPBLAS
+vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+
+endif # GGML_VULKAN
+
+ifdef GGML_HIPBLAS
 	ifeq ($(wildcard /opt/rocm),)
-		ROCM_PATH	?= /usr
+		ROCM_PATH      ?= /usr
 		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
 	else
 		ROCM_PATH	?= /opt/rocm
 		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
-	HIPCC                   ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
-	LLAMA_CUDA_DMMV_X       ?= 32
-	LLAMA_CUDA_MMV_Y        ?= 1
-	LLAMA_CUDA_KQUANTS_ITER ?= 2
+
+	GGML_CUDA_DMMV_X       ?= 32
+	GGML_CUDA_MMV_Y        ?= 1
+	GGML_CUDA_KQUANTS_ITER ?= 2
+
 	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
-ifdef LLAMA_HIP_UMA
+
+ifdef GGML_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # LLAMA_HIP_UMA
-	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
-	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
-	HIPFLAGS    += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
-	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
-	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
-	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
-ifdef LLAMA_CUDA_FORCE_DMMV
-	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
-endif # LLAMA_CUDA_FORCE_DMMV
-ifdef LLAMA_CUDA_NO_PEER_COPY
-	HIPFLAGS 	+= -DGGML_CUDA_NO_PEER_COPY
-endif # LLAMA_CUDA_NO_PEER_COPY
-	OBJS        += ggml-cuda.o
-	OBJS        += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
-
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+endif # GGML_HIP_UMA
+
+	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
+	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas
+
+	HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
+
+	HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
+	HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
+	HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
+	HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
+
+ifdef GGML_CUDA_FORCE_DMMV
+	HIPFLAGS += -DGGML_CUDA_FORCE_DMMV
+endif # GGML_CUDA_FORCE_DMMV
+
+ifdef GGML_CUDA_FORCE_MMQ
+	HIPFLAGS += -DGGML_CUDA_FORCE_MMQ
+endif # GGML_CUDA_FORCE_MMQ
+
+ifdef GGML_CUDA_FORCE_CUBLAS
+	HIPFLAGS += -DGGML_CUDA_FORCE_CUBLAS
+endif # GGML_CUDA_FORCE_CUBLAS
+
+ifdef GGML_CUDA_NO_PEER_COPY
+	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
+endif # GGML_CUDA_NO_PEER_COPY
+
+	OBJ_GGML += ggml/src/ggml-cuda.o
+	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML += $(OBJ_CUDA_TMPL)
+
+ggml/src/ggml-cuda.o: \
+	ggml/src/ggml-cuda.cu \
+	ggml/include/ggml-cuda.h \
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h \
+	ggml/src/ggml-backend-impl.h \
+	ggml/src/ggml-common.h \
+	$(wildcard ggml/src/ggml-cuda/*.cuh)
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+ggml/src/ggml-cuda/%.o: \
+	ggml/src/ggml-cuda/%.cu \
+	ggml/include/ggml.h \
+	ggml/src/ggml-common.h \
+	ggml/src/ggml-cuda/common.cuh
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif # GGML_HIPBLAS
 
-endif # LLAMA_HIPBLAS
-
-ifdef LLAMA_METAL
+ifdef GGML_METAL
 	MK_CPPFLAGS += -DGGML_USE_METAL
 	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	OBJS		+= ggml-metal.o
-ifdef LLAMA_METAL_NDEBUG
+	OBJ_GGML	+= ggml/src/ggml-metal.o
+ifdef GGML_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
-ifdef LLAMA_METAL_EMBED_LIBRARY
+ifdef GGML_METAL_EMBED_LIBRARY
 	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJS        += ggml-metal-embed.o
+	OBJ_GGML   += ggml/src/ggml-metal-embed.o
 endif
-endif # LLAMA_METAL
+endif # GGML_METAL
 
-ifdef LLAMA_METAL
-ggml-metal.o: ggml-metal.m ggml-metal.h ggml.h
+ifdef GGML_METAL
+ggml/src/ggml-metal.o: \
+	ggml/src/ggml-metal.m \
+	ggml/include/ggml-metal.h \
+	ggml/include/ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
-ifdef LLAMA_METAL_EMBED_LIBRARY
-ggml-metal-embed.o: ggml-metal.metal ggml-common.h
+ifdef GGML_METAL_EMBED_LIBRARY
+ggml/src/ggml-metal-embed.o: \
+	ggml/src/ggml-metal.metal \
+	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
+	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
 	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"   >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"        >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"              >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"          >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                >> $(TEMP_ASSEMBLY)
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
 	@$(AS) $(TEMP_ASSEMBLY) -o $@
 	@rm -f ${TEMP_ASSEMBLY}
 endif
-endif # LLAMA_METAL
-
-ifndef LLAMA_NO_LLAMAFILE
-sgemm.o: sgemm.cpp sgemm.h ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif
+endif # GGML_METAL
+
+OBJ_GGML += \
+	ggml/src/ggml.o \
+	ggml/src/ggml-alloc.o \
+	ggml/src/ggml-backend.o \
+	ggml/src/ggml-quants.o \
+	ggml/src/ggml-aarch64.o
+
+OBJ_LLAMA = \
+	src/llama.o \
+	src/llama-vocab.o \
+	src/llama-grammar.o \
+	src/llama-sampling.o \
+	src/unicode.o \
+	src/unicode-data.o
+
+OBJ_COMMON = \
+	common/common.o \
+	common/console.o \
+	common/ngram-cache.o \
+	common/sampling.o \
+	common/train.o \
+	common/grammar-parser.o \
+	common/build-info.o \
+	common/json-schema-to-grammar.o
+
+OBJ_ALL = $(OBJ_GGML) $(OBJ_LLAMA) $(OBJ_COMMON)
+
+LIB_GGML   = $(LIB_PRE)ggml$(DSO_EXT)
+LIB_GGML_S = $(LIB_PRE)ggml.a
+
+LIB_LLAMA   = $(LIB_PRE)llama$(DSO_EXT)
+LIB_LLAMA_S = $(LIB_PRE)llama.a
+
+LIB_COMMON   = $(LIB_PRE)common$(DSO_EXT)
+LIB_COMMON_S = $(LIB_PRE)common.a
+
+LIB_ALL   = $(LIB_GGML)   $(LIB_LLAMA)   $(LIB_COMMON)
+LIB_ALL_S = $(LIB_GGML_S) $(LIB_LLAMA_S) $(LIB_COMMON_S)
 
 GF_CC := $(CC)
 include scripts/get-flags.mk
@@ -643,7 +951,7 @@ override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
 # identify CUDA host compiler
-ifdef LLAMA_CUDA
+ifdef GGML_CUDA
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
 CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
@@ -668,86 +976,243 @@ $(info I NVCCFLAGS: $(NVCCFLAGS))
 $(info I LDFLAGS:   $(LDFLAGS))
 $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
-ifdef LLAMA_CUDA
+ifdef GGML_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifndef GGML_MUSA
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH
 $(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus )
 endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
+
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-endif # LLAMA_CUDA
+endif # GGML_MUSA
+endif # GGML_CUDA
 $(info )
 
-ifdef LLAMA_CUBLAS
-$(info !!!!)
-$(info LLAMA_CUBLAS is deprecated and will be removed in the future. Use LLAMA_CUDA instead.)
-$(info !!!!)
+ifdef DEPRECATE_WARNING
+$(info !!! DEPRECATION WARNING !!!)
+$(info The following LLAMA_ options are deprecated and will be removed in the future. Use the GGML_ prefix instead)
+$(info   - LLAMA_CUDA)
+$(info   - LLAMA_METAL)
+$(info   - LLAMA_METAL_EMBED_LIBRARY)
+$(info   - LLAMA_OPENMP)
+$(info   - LLAMA_RPC)
+$(info   - LLAMA_SYCL)
+$(info   - LLAMA_SYCL_F16)
+$(info   - LLAMA_OPENBLAS)
+$(info   - LLAMA_OPENBLAS64)
+$(info   - LLAMA_BLIS)
+$(info   - LLAMA_NO_LLAMAFILE)
+$(info   - LLAMA_NO_ACCELERATE)
+$(info   - LLAMA_NO_OPENMP)
+$(info   - LLAMA_NO_METAL)
+$(info   - LLAMA_NO_CCACHE)
 $(info )
 endif
 
 #
-# Build library
+# Build libraries
 #
 
-ggml.o: ggml.c ggml.h ggml-cuda.h
+# ggml
+
+ggml/src/ggml.o: \
+	ggml/src/ggml.c \
+	ggml/include/ggml.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+ggml/src/ggml-alloc.o: \
+	ggml/src/ggml-alloc.c \
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
+ggml/src/ggml-backend.o: \
+	ggml/src/ggml-backend.c \
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
+ggml/src/ggml-quants.o: \
+	ggml/src/ggml-quants.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-quants.h \
+	ggml/src/ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-unicode.o: unicode.cpp unicode.h
+ggml/src/ggml-aarch64.o: \
+	ggml/src/ggml-aarch64.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
+ggml/src/ggml-blas.o: \
+	ggml/src/ggml-blas.cpp \
+	ggml/include/ggml-blas.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-unicode-data.o: unicode-data.cpp unicode-data.h
+ifndef GGML_NO_LLAMAFILE
+ggml/src/llamafile/sgemm.o: \
+	ggml/src/llamafile/sgemm.cpp \
+	ggml/src/llamafile/sgemm.h \
+	ggml/include/ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # GGML_NO_LLAMAFILE
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
+ifdef GGML_RPC
+ggml/src/ggml-rpc.o: \
+	ggml/src/ggml-rpc.cpp \
+	ggml/include/ggml-rpc.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # GGML_RPC
 
-llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
+$(LIB_GGML): \
+	$(OBJ_GGML)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+
+$(LIB_GGML_S): \
+	$(OBJ_GGML)
+	ar rcs $(LIB_GGML_S) $^
+
+# llama
+
+src/unicode.o: \
+	src/unicode.cpp \
+	src/unicode.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
+src/unicode-data.o: \
+	src/unicode-data.cpp \
+	src/unicode-data.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common.o: common/common.cpp $(COMMON_H_DEPS)
+src/llama.o: \
+	src/llama.cpp \
+	src/llama-impl.h \
+	src/llama-vocab.h \
+	src/llama-grammar.h \
+	src/llama-sampling.h \
+	src/unicode.h \
+	include/llama.h \
+	ggml/include/ggml-cuda.h \
+	ggml/include/ggml-metal.h \
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h \
+	ggml/include/ggml-backend.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-sampling.o: common/sampling.cpp $(COMMON_H_DEPS)
+src/llama-vocab.o: \
+	src/llama-vocab.cpp \
+	src/llama-vocab.h \
+	src/llama-impl.h \
+	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-console.o: common/console.cpp common/console.h
+src/llama-grammar.o: \
+	src/llama-grammar.cpp \
+	src/llama-grammar.h \
+	src/llama-impl.h \
+	src/llama-vocab.h \
+	src/llama-sampling.h \
+	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
+src/llama-sampling.o: \
+	src/llama-sampling.cpp \
+	src/llama-sampling.h \
+	src/llama-impl.h \
+	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-json-schema-to-grammar.o: common/json-schema-to-grammar.cpp common/json-schema-to-grammar.h
+$(LIB_LLAMA): \
+	$(OBJ_LLAMA) \
+	$(LIB_GGML)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+
+$(LIB_LLAMA_S): \
+	$(OBJ_LLAMA)
+	ar rcs $(LIB_LLAMA_S) $^
+
+# common
+
+common/common.o: \
+	common/common.cpp \
+	common/common.h \
+	common/console.h \
+	common/sampling.h \
+	common/json.hpp \
+	common/json-schema-to-grammar.h \
+	include/llama.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/sampling.o: \
+	common/sampling.cpp \
+	common/sampling.h \
+	include/llama.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/console.o: \
+	common/console.cpp \
+	common/console.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-train.o: common/train.cpp common/train.h
+common/grammar-parser.o: \
+	common/grammar-parser.cpp \
+	common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-ngram-cache.o: common/ngram-cache.cpp common/ngram-cache.h
+common/json-schema-to-grammar.o: \
+	common/json-schema-to-grammar.cpp \
+	common/json-schema-to-grammar.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-libllama.so: llama.o ggml.o $(OBJS)
+common/train.o: \
+	common/train.cpp \
+	common/train.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/ngram-cache.o: \
+	common/ngram-cache.cpp \
+	common/ngram-cache.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+$(LIB_COMMON): \
+	$(OBJ_COMMON) \
+	$(LIB_LLAMA) \
+	$(LIB_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
-	ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS)
+$(LIB_COMMON_S): \
+	$(OBJ_COMMON)
+	ar rcs $(LIB_COMMON_S) $^
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -vrf ggml-cuda/*.o
+	rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -rvf src/*.o
+	rm -rvf tests/*.o
+	rm -rvf examples/*.o
+	rm -rvf common/*.o
+	rm -rvf *.a
+	rm -rvf *.dll
+	rm -rvf *.so
+	rm -rvf *.dot
+	rm -rvf ggml/*.a
+	rm -rvf ggml/*.dll
+	rm -rvf ggml/*.so
+	rm -vrf ggml/src/*.o
+	rm -rvf common/build-info.cpp
+	rm -vrf ggml/src/ggml-metal-embed.metal
+	rm -vrf ggml/src/ggml-cuda/*.o
+	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
+	rm -rvf $(BUILD_TARGETS)
+	rm -rvf $(TEST_TARGETS)
+	rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
+	rm -rvf $(LEGACY_TARGETS_CLEAN)
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -761,157 +1226,241 @@ clean:
 # Helper function that replaces .c, .cpp, and .cu file endings with .o:
 GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
 
-main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+llama-cli: examples/main/main.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 	@echo
-	@echo '====  Run ./main -h for help.  ===='
+	@echo '====  Run ./llama-cli -h for help.  ===='
 	@echo
 
-infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+llama-infill: examples/infill/infill.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-simple: examples/simple/simple.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tokenize: examples/tokenize/tokenize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-tokenize: examples/tokenize/tokenize.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-batched: examples/batched/batched.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-batched-bench: examples/batched-bench/batched-bench.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-quantize: examples/quantize/quantize.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
+llama-quantize-stats: examples/quantize-stats/quantize-stats.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-perplexity: examples/perplexity/perplexity.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-imatrix: examples/imatrix/imatrix.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-embedding: examples/embedding/embedding.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-gritlm: examples/gritlm/gritlm.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-gritlm: examples/gritlm/gritlm.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-save-load-state: examples/save-load-state/save-load-state.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/server/json-schema-to-grammar.mjs.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+llama-gguf: examples/gguf/gguf.cpp \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
-examples/server/%.hpp: examples/server/public/% Makefile
-	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
-		echo "unsigned char $${NAME}[] = {" && \
-		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
-		echo "};" && \
-		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
-	) > $@
+examples/gguf-hash/deps/sha1/sha1.o: \
+	examples/gguf-hash/deps/sha1/sha1.c
+	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
 
-gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+examples/gguf-hash/deps/xxhash/xxhash.o: \
+	examples/gguf-hash/deps/xxhash/xxhash.c
+	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
+
+examples/gguf-hash/deps/sha256/sha256.o: \
+	examples/gguf-hash/deps/sha256/sha256.c
+	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
+
+llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/sha1.o examples/gguf-hash/deps/xxhash/xxhash.o examples/gguf-hash/deps/sha256/sha256.o\
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-gguf-split: examples/gguf-split/gguf-split.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-eval-callback: examples/eval-callback/eval-callback.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-eval-callback: examples/eval-callback/eval-callback.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
+llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \
+	$(OBJ_GGML) $(OBJ_LLAMA)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-bench: examples/llama-bench/llama-bench.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-libllava.a: examples/llava/llava.cpp examples/llava/llava.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h common/base64.hpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
-
-llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-baby-llama: examples/baby-llama/baby-llama.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
-	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+llama-export-lora: examples/export-lora/export-lora.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-retrieval: examples/retrieval/retrieval.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+llama-speculative: examples/speculative/speculative.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
+llama-parallel: examples/parallel/parallel.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-retrieval: examples/retrieval/retrieval.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-lookahead: examples/lookahead/lookahead.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+llama-lookup: examples/lookup/lookup.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-lookup-create: examples/lookup/lookup-create.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-lookup-merge: examples/lookup/lookup-merge.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
+llama-lookup-stats: examples/lookup/lookup-stats.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS)
-	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS)
-	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS)
-
-passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+
+llama-passkey: examples/passkey/passkey.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+llama-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+ifdef GGML_RPC
+rpc-server: examples/rpc/rpc-server.cpp \
+	$(OBJ_GGML)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+endif # GGML_RPC
+
+llama-server: \
+	examples/server/server.cpp \
+	examples/server/utils.hpp \
+	examples/server/httplib.h \
+	examples/server/colorthemes.css.hpp \
+	examples/server/style.css.hpp \
+	examples/server/theme-beeninorder.css.hpp \
+	examples/server/theme-ketivah.css.hpp \
+	examples/server/theme-mangotango.css.hpp \
+	examples/server/theme-playground.css.hpp \
+	examples/server/theme-polarnight.css.hpp \
+	examples/server/theme-snowstorm.css.hpp \
+	examples/server/index.html.hpp \
+	examples/server/index-new.html.hpp \
+	examples/server/index.js.hpp \
+	examples/server/completion.js.hpp \
+	examples/server/system-prompts.js.hpp \
+	examples/server/prompt-formats.js.hpp \
+	examples/server/json-schema-to-grammar.mjs.hpp \
+	common/json.hpp \
+	common/stb_image.h \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+
+# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+examples/server/%.hpp: examples/server/public/% Makefile
+	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
+		echo "unsigned char $${NAME}[] = {" && \
+		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
+		echo "};" && \
+		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
+	) > $@
+
+libllava.a: examples/llava/llava.cpp \
+	examples/llava/llava.h \
+	examples/llava/clip.cpp \
+	examples/llava/clip.h \
+	common/stb_image.h \
+	common/base64.hpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
+
+llama-llava-cli: examples/llava/llava-cli.cpp \
+	examples/llava/clip.h \
+	examples/llava/clip.cpp \
+	examples/llava/llava.h \
+	examples/llava/llava.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
+
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
@@ -925,7 +1474,7 @@ common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
 		rm $@.tmp; \
 	fi
 
-build-info.o: common/build-info.cpp
+common/build-info.o: common/build-info.cpp
 	$(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@
 
 #
@@ -934,94 +1483,165 @@ build-info.o: common/build-info.cpp
 
 tests: $(TEST_TARGETS)
 
-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
+llama-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp \
+	$(OBJ_GGML) common/build-info.o
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-run-benchmark-matmult: benchmark-matmult
+run-benchmark-matmult: llama-benchmark-matmult
 	./$@
 
 .PHONY: run-benchmark-matmult swift
 
-vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o grammar-parser.o $(OBJS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-parser.o $(OBJS)
+tests/test-grammar-parser: tests/test-grammar-parser.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-grammar-integration: tests/test-grammar-integration.cpp ggml.o llama.o grammar-parser.o $(OBJS)
+tests/test-grammar-integration: tests/test-grammar-integration.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS)
+tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-opt: tests/test-opt.cpp ggml.o $(OBJS)
+tests/test-opt: tests/test-opt.cpp \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o $(OBJS)
+tests/test-quantize-fns: tests/test-quantize-fns.cpp \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
+tests/test-quantize-perf: tests/test-quantize-perf.cpp \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
+tests/test-sampling: tests/test-sampling.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0: tests/test-tokenizer-0.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-0: tests/test-tokenizer-0.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-spm: tests/test-tokenizer-1-spm.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-1-spm: tests/test-tokenizer-1-spm.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
+tests/test-rope: tests/test-rope.cpp ggml/src/ggml.o \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-c.o: tests/test-c.c llama.h
+tests/test-c.o: tests/test-c.c include/llama.h
 	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
 
-tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
+tests/test-backend-ops: tests/test-backend-ops.cpp \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
+tests/test-model-load-cancel: tests/test-model-load-cancel.cpp tests/get-model.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
+tests/test-autorelease: tests/test-autorelease.cpp tests/get-model.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-chat-template: tests/test-chat-template.cpp \
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+#
+# PoCs
+#
+
+llama-vdot: pocs/vdot/vdot.cpp ggml/src/ggml.o \
+	$(OBJ_GGML)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
+	$(OBJ_GGML)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+#
+# Deprecated binaries that we want to keep around long enough for people to migrate to the new filenames, then these can be removed.
+#
+# Mark legacy binary targets as .PHONY so that they are always checked.
+.PHONY: main quantize perplexity embedding server
+
+# Define the object file target
+examples/deprecation-warning/deprecation-warning.o: examples/deprecation-warning/deprecation-warning.cpp
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+# NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate.
+#  Eventually we will want to remove these target from building all the time.
+main: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	@echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead."
+
+server: examples/deprecation-warning/deprecation-warning.o
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."
+
+quantize: examples/deprecation-warning/deprecation-warning.o
+ifneq (,$(wildcard quantize))
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead."
+	@echo "  Remove the 'quantize' binary to remove this warning."
+	@echo "#########"
+endif
+
+perplexity: examples/deprecation-warning/deprecation-warning.o
+ifneq (,$(wildcard perplexity))
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead."
+	@echo "  Remove the 'perplexity' binary to remove this warning."
+	@echo "#########"
+endif
+
+embedding: examples/deprecation-warning/deprecation-warning.o
+ifneq (,$(wildcard embedding))
+	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
+	@echo "  Remove the 'embedding' binary to remove this warning."
+	@echo "#########"
+endif
diff --git a/llm/llama.cpp/Package.swift b/llm/llama.cpp/Package.swift
index 183e647575b42802edbaa110ca27bbcfd54e721c..1d90b47bfa3da9450d4a31f19248b88c041b9457 100644
--- a/llm/llama.cpp/Package.swift
+++ b/llm/llama.cpp/Package.swift
@@ -3,14 +3,17 @@
 import PackageDescription
 
 var sources = [
-    "ggml.c",
-    "sgemm.cpp",
-    "llama.cpp",
-    "unicode.cpp",
-    "unicode-data.cpp",
-    "ggml-alloc.c",
-    "ggml-backend.c",
-    "ggml-quants.c",
+    "src/llama.cpp",
+    "src/llama-vocab.cpp",
+    "src/llama-grammar.cpp",
+    "src/llama-sampling.cpp",
+    "src/unicode.cpp",
+    "src/unicode-data.cpp",
+    "ggml/src/ggml.c",
+    "ggml/src/ggml-alloc.c",
+    "ggml/src/ggml-backend.c",
+    "ggml/src/ggml-quants.c",
+    "ggml/src/ggml-aarch64.c",
 ]
 
 var resources: [Resource] = []
@@ -26,8 +29,8 @@ var cSettings: [CSetting] =  [
 ]
 
 #if canImport(Darwin)
-sources.append("ggml-metal.m")
-resources.append(.process("ggml-metal.metal"))
+sources.append("ggml/src/ggml-metal.m")
+resources.append(.process("ggml/src/ggml-metal.metal"))
 linkerSettings.append(.linkedFramework("Accelerate"))
 cSettings.append(
     contentsOf: [
@@ -63,8 +66,6 @@ let package = Package(
                "models",
                "tests",
                "CMakeLists.txt",
-               "ggml-cuda.cu",
-               "ggml-cuda.h",
                "Makefile"
             ],
             sources: sources,
diff --git a/llm/llama.cpp/README.md b/llm/llama.cpp/README.md
index d85a453be160d62b16377341e68e3018bb69dfc3..1283f6805874ea05775b94caf571a5923f2c7bcb 100644
--- a/llm/llama.cpp/README.md
+++ b/llm/llama.cpp/README.md
@@ -3,15 +3,19 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
+[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
 [![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-### Recent API changes
+> [!IMPORTANT]
+[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
 
+## Recent API changes
+
+- [2024 Jun 26] The source code and CMake build scripts have been restructured https://github.com/ggerganov/llama.cpp/pull/8006
 - [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
 - [2024 Apr 4] State and session file functions reorganized under `llama_state_*` https://github.com/ggerganov/llama.cpp/pull/6341
 - [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
@@ -20,9 +24,10 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 - [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
 - [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
 
-### Hot topics
+## Hot topics
 
-- **Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021**
+- **`convert.py` has been deprecated and moved to `examples/convert_legacy_llama.py`, please use `convert_hf_to_gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
+- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
 - BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
 - MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
 - Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
@@ -34,38 +39,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ----
 
-

-  Table of Contents
-  
    
-    
  1. 
-      Description
-    
    2. 
-    
  2. 
-      Usage
-      
-    
    3. 
-    
  3. Contributing
    4. 
-    
  4. Coding guidelines
    5. 
-    
  5. Docs
    6. 
-  

-

-
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -76,21 +49,13 @@ variety of hardware - locally and in the cloud.
 - AVX, AVX2 and AVX512 support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
-- Vulkan, SYCL, and (partial) OpenCL backend support
+- Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
 Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022), the project has
 improved significantly thanks to many contributions. It is the main playground for developing new features for the
 [ggml](https://github.com/ggerganov/ggml) library.
 
-**Supported platforms:**
-
-- [X] Mac OS
-- [X] Linux
-- [X] Windows (via CMake)
-- [X] Docker
-- [X] FreeBSD
-
 **Supported models:**
 
 Typically finetunes of the base models below are supported as well.
@@ -104,6 +69,7 @@ Typically finetunes of the base models below are supported as well.
 - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
+- [X] [BERT](https://github.com/ggerganov/llama.cpp/pull/5423)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
@@ -129,9 +95,18 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
+- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
+- [x] [Smaug](https://huggingface.co/models?search=Smaug)
+- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
+- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
+- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
+- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
+- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
+- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 
-(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
+(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
 **Multimodal models:**
 
@@ -145,10 +120,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 
-**HTTP server**
-
-[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
-
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
@@ -169,17 +140,20 @@ Typically finetunes of the base models below are supported as well.
 - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 - PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
+- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
 
 **UI:**
 
 Unless otherwise noted these projects are open-source with permissive licensing:
 
+- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
 - [iohub/collama](https://github.com/iohub/coLLaMA)
 - [janhq/jan](https://github.com/janhq/jan) (AGPL)
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
 - [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
+- [ramalama](https://github.com/containers/ramalama) (MIT)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
@@ -190,6 +164,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [RAGNA Desktop](https://ragna.app/) (proprietary)
 - [RecurseChat](https://recurse.chat/) (proprietary)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
@@ -202,19 +177,30 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
+- [AIKit](https://github.com/sozercan/aikit) (MIT)
+- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 **Tools:**
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
+
+**Infrastructure:**
 
----
+- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
 
-Here is a typical run using LLaMA v2 13B on M2 Ultra:
+**Games:**
+- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
+
+## Demo
+
+

+Typical run using LLaMA v2 13B on M2 Ultra
 
 ```
-$ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
+$ make -j && ./llama-cli -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
 I llama.cpp build info:
 I UNAME_S:  Darwin
 I UNAME_P:  arm
@@ -291,550 +277,85 @@ llama_print_timings:        eval time = 24513.59 ms /   399 runs   (   61.44 ms
 llama_print_timings:       total time = 25431.49 ms
 ```
 
+

+
+

+Demo of running both LLaMA-7B and whisper.cpp on a single M1 Pro MacBook
+
 And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook:
 
 https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8b4f-add84093ffff.mp4
 
+

+
 ## Usage
 
 Here are the end-to-end binary build and model conversion steps for most supported models.
 
-### Get the Code
-
-```bash
-git clone https://github.com/ggerganov/llama.cpp
-cd llama.cpp
-```
-
-### Build
-
-In order to build llama.cpp you have four different options.
-
-- Using `make`:
-  - On Linux or MacOS:
-
-      ```bash
-      make
-      ```
-
-  - On Windows:
-
-    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
-    2. Extract `w64devkit` on your pc.
-    3. Run `w64devkit.exe`.
-    4. Use the `cd` command to reach the `llama.cpp` folder.
-    5. From here you can run:
-        ```bash
-        make
-        ```
+### Basic usage
 
-  - Notes:
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, run `make LLAMA_DEBUG=1`
+Firstly, you need to get the binary. There are different methods that you can follow:
+- Method 1: Clone this repository and build locally, see [how to build](./docs/build.md)
+- Method 2: If you are using MacOS or Linux, you can install llama.cpp via [brew, flox or nix](./docs/install.md)
+- Method 3: Use a Docker image, see [documentation for Docker](./docs/docker.md)
+- Method 4: Download pre-built binary from [releases](https://github.com/ggerganov/llama.cpp/releases)
 
-- Using `CMake`:
+You can run a basic completion using this command:
 
-  ```bash
-  cmake -B build
-  cmake --build build --config Release
-  ```
-
-  **Notes**:
-
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, there are two cases:
-
-      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
-
-      ```bash
-      cmake -B build -DCMAKE_BUILD_TYPE=Debug
-      cmake --build build
-      ```
-
-      2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
-
-      ```bash
-      cmake -B build -G "Xcode"
-      cmake --build build --config Debug
-      ```
-
-- Using `Zig` (version 0.11 or later):
-
-    Building for optimization levels and CPU features can be accomplished using standard build arguments, for example AVX2, FMA, F16C,
-    it's also possible to cross compile for other operating systems and architectures:
-
-    ```bash
-    zig build -Doptimize=ReleaseFast -Dtarget=x86_64-windows-gnu -Dcpu=x86_64+avx2+fma+f16c
-    ```
-
-    The `zig targets` command will give you valid options to use.
-
--   Using `gmake` (FreeBSD):
-
-    1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
-    2. Add your user to **video** group
-    3. Install compilation dependencies.
-
-        ```bash
-        sudo pkg install gmake automake autoconf pkgconf llvm15 clinfo clover \
-            opencl clblast openblas
-
-        gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
-        ```
-
-    **Notes:** With this packages you can build llama.cpp with OPENBLAS and
-    CLBLAST support for use OpenCL GPU acceleration in FreeBSD. Please read
-    the instructions for use and activate this options in this document below.
-
-### Homebrew
+```bash
+llama-cli -m your_model.gguf -p "I believe the meaning of life is" -n 128
 
-On Mac and Linux, the homebrew package manager can be used via 
+# Output:
+# I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
 ```
-brew install llama.cpp
-```
-The formula is automatically updated with new `llama.cpp` releases.
-
-### Metal Build
-
-On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
-To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or the `LLAMA_METAL=OFF` cmake option.
-
-When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
-argument.
-
-### BLAS Build
-
-Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
-
-- #### Accelerate Framework:
-
-  This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
-
-- #### OpenBLAS:
 
-  This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
+See [this page](./examples/main/README.md) for a full list of parameters.
 
-  - Using `make`:
-    - On Linux:
-      ```bash
-      make LLAMA_OPENBLAS=1
-      ```
+### Conversation mode
 
-    - On Windows:
-
-      1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
-      2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
-      3. Extract `w64devkit` on your pc.
-      4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
-      5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
-      6. Run `w64devkit.exe`.
-      7. Use the `cd` command to reach the `llama.cpp` folder.
-      8. From here you can run:
-
-          ```bash
-          make LLAMA_OPENBLAS=1
-          ```
-
-  - Using `CMake` on Linux:
-
-      ```bash
-      cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
-      cmake --build build --config Release
-      ```
-
-- #### BLIS
-
-  Check [BLIS.md](docs/BLIS.md) for more information.
-
-- #### SYCL
-  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
-
-  llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
-
-  For detailed info, please refer to [llama.cpp for SYCL](README-sycl.md).
-
-- #### Intel oneMKL
-  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./README-sycl.md).
-
-  - Using manual oneAPI installation:
-    By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
-      ```bash
-      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-basekit docker image, only required for manual installation
-      cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
-      cmake --build build --config Release
-      ```
-
-  - Using oneAPI docker image:
-    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
-
-  Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
-
-- #### CUDA
-
-  This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
-
-  For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
-
-  - Using `make`:
-    ```bash
-    make LLAMA_CUDA=1
-    ```
-  - Using `CMake`:
-
-    ```bash
-    cmake -B build -DLLAMA_CUDA=ON
-    cmake --build build --config Release
-    ```
-
-  The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
-
-  | Option                         | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
-  |--------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-  | LLAMA_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
-  | LLAMA_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                               |
-  | LLAMA_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of dequantization + matrix multiplication kernels instead of leveraging Math libraries. |                                                                                                                                         |
-  | LLAMA_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
-  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
-  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
-
-- #### hipBLAS
-
-  This provides BLAS acceleration on HIP-supported AMD GPUs.
-  Make sure to have ROCm installed.
-  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
-
-  - Using `make`:
-    ```bash
-    make LLAMA_HIPBLAS=1
-    ```
-  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
-    ```bash
-    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
-        && cmake --build build --config Release -- -j 16
-    ```
-    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON`.
-    However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
-
-    Note that if you get the following error:
-    ```
-    clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
-    ```
-    Try searching for a directory under `HIP_PATH` that contains the file
-    `oclc_abi_version_400.bc`. Then, add the following to the start of the
-    command: `HIP_DEVICE_LIB_PATH=`, so something
-    like:
-    ```bash
-    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
-    HIP_DEVICE_LIB_PATH= \
-        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
-        && cmake --build build -- -j 16
-    ```
-
-  - Using `make` (example for target gfx1030, build with 16 CPU threads):
-    ```bash
-    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
-    ```
-
-  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
-    ```bash
-    set PATH=%HIP_PATH%\bin;%PATH%
-    cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
-    cmake --build build
-    ```
-    Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
-    Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
-
-
-  The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
-  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
-  The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
-
-  | Option                  | Legal values           | Default | Description                                                                                                                                                                                                                                    |
-  |-------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
-  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
-
-- #### CLBlast
-
-  OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
-
-  You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
-    - For Ubuntu, Debian, and Fedora the packages `opencl-headers`, `ocl-icd` may be needed.
-
-    - For Windows, a pre-built SDK is available on the [OpenCL Releases](https://github.com/KhronosGroup/OpenCL-SDK/releases) page.
-
-    - 

-        Installing the OpenCL SDK from source
-
-        ```sh
-        git clone --recurse-submodules https://github.com/KhronosGroup/OpenCL-SDK.git
-        cd OpenCL-SDK
-        cmake -B build -DBUILD_DOCS=OFF \
-          -DBUILD_EXAMPLES=OFF \
-          -DBUILD_TESTING=OFF \
-          -DOPENCL_SDK_BUILD_SAMPLES=OFF \
-          -DOPENCL_SDK_TEST_SAMPLES=OFF
-        cmake --build build
-        cmake --install build --prefix /some/path
-        ```
-      

-
-  ##### Installing CLBlast
-
-  Pre-built CLBlast binaries may be found on the [CLBlast Releases](https://github.com/CNugteren/CLBlast/releases) page. For Unix variants, it may also be found in your operating system's packages.
-
-  Linux packaging:
-  Fedora Linux:
-  ```bash
-  sudo dnf install clblast
-  ```
-
-  Alternatively, they may be built from source.
-
-  - 

-    Windows:
-
-      ```cmd
-      set OPENCL_SDK_ROOT="C:/OpenCL-SDK-v2023.04.17-Win-x64"
-      git clone https://github.com/CNugteren/CLBlast.git
-      cd CLBlast
-      cmake -B build -DBUILD_SHARED_LIBS=OFF -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT=%OPENCL_SDK_ROOT% -G "Visual Studio 17 2022" -A x64
-      cmake --build build --config Release
-      cmake --install build --prefix C:/CLBlast
-      ```
-
-      (note: `--config Release` at build time is the default and only relevant for Visual Studio builds - or multi-config Ninja builds)
-
-  - 

-    Unix:
-
-      ```sh
-      git clone https://github.com/CNugteren/CLBlast.git
-      cd CLBlast
-      cmake -B build -DBUILD_SHARED_LIBS=OFF -DTUNERS=OFF
-      cmake --build build --config Release
-      cmake --install build --prefix /some/path
-      ```
-
-      Where `/some/path` is where the built library will be installed (default is `/usr/local`).
-    

-
-  ##### Building Llama with CLBlast
-
-  - Build with make:
-    ```sh
-    make LLAMA_CLBLAST=1
-    ```
-  - CMake (Unix):
-    ```sh
-    cmake -B build -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path
-    cmake --build build --config Release
-    ```
-  - CMake (Windows):
-    ```cmd
-    set CL_BLAST_CMAKE_PKG="C:/CLBlast/lib/cmake/CLBlast"
-    git clone https://github.com/ggerganov/llama.cpp
-    cd llama.cpp
-    cmake -B build -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=%CL_BLAST_CMAKE_PKG% -G "Visual Studio 17 2022" -A x64
-    cmake --build build --config Release
-    cmake --install build --prefix C:/LlamaCPP
-    ```
-
-  ##### Running Llama with CLBlast
-
-  The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does.
-
-  To select the correct platform (driver) and device (GPU), you can use the environment variables `GGML_OPENCL_PLATFORM` and `GGML_OPENCL_DEVICE`.
-  The selection can be a number (starting from 0) or a text string to search:
-
-  ```sh
-  GGML_OPENCL_PLATFORM=1 ./main ...
-  GGML_OPENCL_DEVICE=2 ./main ...
-  GGML_OPENCL_PLATFORM=Intel ./main ...
-  GGML_OPENCL_PLATFORM=AMD GGML_OPENCL_DEVICE=1 ./main ...
-  ```
-
-  The default behavior is to find the first GPU device, but when it is an integrated GPU on a laptop, for instance, the selectors are useful.
-  Using the variables it is possible to select a CPU-based driver as well, if so desired.
-
-  You can get a list of platforms and devices from the `clinfo -l` command, etc.
-
-- #### Vulkan
-
-  **With docker**:
-
-  You don't need to install Vulkan SDK. It will be installed inside the container.
-
-  ```sh
-  # Build the image
-  docker build -t llama-cpp-vulkan -f .devops/main-vulkan.Dockerfile .
-
-  # Then, use it:
-  docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
-  ```
-
-  **Without docker**:
-
-  Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
-
-  For example, on Ubuntu 22.04 (jammy), use the command below:
-
-  ```bash
-  wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
-  wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-  apt update -y
-  apt-get install -y vulkan-sdk
-  # To verify the installation, use the command below:
-  vulkaninfo
-  ```
-
-  Alternatively your package manager might be able to provide the appropiate libraries. For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
-
-  Then, build llama.cpp using the cmake command below:
-
-  ```bash
-  cmake -B build -DLLAMA_VULKAN=1
-  cmake --build build --config Release
-  # Test the output binary (with "-ngl 33" to offload all layers to GPU)
-  ./bin/main -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
-
-  # You should see in the output, ggml_vulkan detected your GPU. For example:
-  # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
-  ```
-
-### Prepare and Quantize
-
-> [!NOTE]
-> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
-
-To obtain the official LLaMA 2 weights please see the Obtaining and using the Facebook LLaMA 2 model section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
-
-Note: `convert.py` has been moved to `examples/convert-legacy-llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derievatives.
-It does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.
+If you want a more ChatGPT-like experience, you can run in conversation mode by passing `-cnv` as a parameter:
 
 ```bash
-# obtain the official LLaMA model weights and place them in ./models
-ls ./models
-llama-2-7b tokenizer_checklist.chk tokenizer.model
-# [Optional] for models using BPE tokenizers
-ls ./models
- vocab.json
-# [Optional] for PyTorch .bin models like Mistral-7B
-ls ./models
-
-
-# install Python dependencies
-python3 -m pip install -r requirements.txt
-
-# convert the model to ggml FP16 format
-python3 convert-hf-to-gguf.py models/mymodel/
-
-# [Optional] for models using BPE tokenizers
-python convert-hf-to-gguf.py models/mymodel/ --vocab-type bpe
-
-# quantize the model to 4-bits (using Q4_K_M method)
-./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
-
-# update the gguf filetype to current version if older version is now unsupported
-./quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
+llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv
+
+# Output:
+# > hi, who are you?
+# Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
+#
+# > what is 1+1?
+# Easy peasy! The answer to 1+1 is... 2!
 ```
 
-### Run the quantized model
+By default, the chat template will be taken from the input model. If you want to use another chat template, pass `--chat-template NAME` as a parameter. See the list of [supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
 
 ```bash
-# start inference on a gguf model
-./main -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
+./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
 ```
 
-When running the larger models, make sure you have enough disk space to store all the intermediate files.
-
-### Running on Windows with prebuilt binaries
-
-You will find prebuilt Windows binaries on the release page.
+You can also use your own template via in-prefix, in-suffix and reverse-prompt parameters:
 
-Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
-
-From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
-
-```
-.\main -m llama-2-7b.Q4_0.gguf -n 128
+```bash
+./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
 ```
 
-### Memory/Disk Requirements
-
-As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
-
-| Model | Original size | Quantized size (Q4_0) |
-|------:|--------------:|----------------------:|
-|    7B |         13 GB |                3.9 GB |
-|   13B |         24 GB |                7.8 GB |
-|   30B |         60 GB |               19.5 GB |
-|   65B |        120 GB |               38.5 GB |
-
-### Quantization
-
-Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
-
-*(outdated)*
-
-| Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |
-|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
-|    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
-|    7B | file size    |  13.0G |   3.5G |   3.9G |   4.3G |   4.7G |   6.7G |
-|    7B | ms/tok @ 4th |    127 |     55 |     54 |     76 |     83 |     72 |
-|    7B | ms/tok @ 8th |    122 |     43 |     45 |     52 |     56 |     67 |
-|    7B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
-|   13B | perplexity   | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
-|   13B | file size    |  25.0G |   6.8G |   7.6G |   8.3G |   9.1G |    13G |
-|   13B | ms/tok @ 4th |      - |    103 |    105 |    148 |    160 |    131 |
-|   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
-|   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
-
-- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
-- recent k-quants improvements and new i-quants
-  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
-  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
-  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
-  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
-  - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
-  - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
-  - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
-  - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
-  - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
-  - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
-  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
-  - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
-  - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
-  - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
-
-### Perplexity (measuring model quality)
+### Web server
 
-You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
-For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
+[llama.cpp web server](./examples/server/README.md) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
-The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
-The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
+Example usage:
 
-#### How to run
+```bash
+./llama-server -m your_model.gguf --port 8080
 
-1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
-3. Output:
+# Basic web UI can be accessed via browser: http://localhost:8080
+# Chat completion endpoint: http://localhost:8080/v1/chat/completions
 ```
-perplexity : calculating perplexity over 655 chunks
-24.43 seconds per pass - ETA 4.45 hours
-[1]4.5970,[2]5.1807,[3]6.0382,...
-```
-And after 4.45 hours, you will have the final perplexity.
 
 ### Interactive mode
 
-If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
+> [!NOTE]
+> If you prefer basic usage, please consider using conversation mode instead of interactive mode
+
 In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
 
 Here is an example of a few-shot interaction, invoked with the command
@@ -847,16 +368,16 @@ Here is an example of a few-shot interaction, invoked with the command
 ./examples/chat-13B.sh
 
 # custom arguments using a 13B model
-./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
+./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
 ```
 
-Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `main` example program.
+Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `llama-cli` example program.
 
 ![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png)
 
 ### Persistent Interaction
 
-The prompt, user inputs, and model generations can be saved and resumed across calls to `./main` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
+The prompt, user inputs, and model generations can be saved and resumed across calls to `./llama-cli` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
 
 ```bash
 # Start a new chat
@@ -878,235 +399,85 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
 `llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only:
 
 ```bash
-./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
+./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
 ```
 
 The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
 
 For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
 
-### Instruct mode
-
-1. First, download and place the `ggml` model into the `./models` folder
-2. Run the `main` tool like this:
-
-```
-./examples/alpaca.sh
-```
-
-Sample run:
-
-```
-== Running in interactive mode. ==
- - Press Ctrl+C to interject at any time.
- - Press Return to return control to LLaMA.
- - If you want to submit another line, end your input in '\'.
-
- Below is an instruction that describes a task. Write a response that appropriately completes the request.
-
-> How many letters are there in the English alphabet?
-There 26 letters in the English Alphabet
-> What is the most common way of transportation in Amsterdam?
-The majority (54%) are using public transit. This includes buses, trams and metros with over 100 lines throughout the city which make it very accessible for tourists to navigate around town as well as locals who commute by tram or metro on a daily basis
-> List 5 words that start with "ca".
-cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
->
-```
-
-### Obtaining and using the Facebook LLaMA 2 model
-
-- Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
-- Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
-  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF)
-  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF)
-  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF)
-  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF)
-  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
-  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
-
-### Seminal papers and background on the models
-
-If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
-- LLaMA:
-    - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
-    - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
-- GPT-3
-    - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
-- GPT-3.5 / InstructGPT / ChatGPT:
-    - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
-    - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
-
-### Android
-
-#### Build on Android using Termux
-[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
-```
-apt update && apt upgrade -y
-apt install git make cmake
-```
-
-It's recommended to move your model inside the `~/` directory for best performance:
-```
-cd storage/downloads
-mv model.gguf ~/
-```
-
-[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
-
-#### Building the Project using Android NDK
-Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
-
-Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
-```
-$ mkdir build-android
-$ cd build-android
-$ export NDK=
-$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
-$ make
-```
-
-Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
-
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
-
-```
-$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
-```
-
-Now, you can start chatting:
-```
-$cd /data/data/com.termux/files/home/bin
-$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
-```
-
-Here's a demo of an interactive session running on Pixel 5 phone:
-
-https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
-
-### Docker
-
-#### Prerequisites
-* Docker must be installed and running on your system.
-* Create a folder to store big models & intermediate files (ex. /llama/models)
-
-#### Images
-We have three Docker images available for this project:
+## Build
 
-1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+Please refer to [Build llama.cpp locally](./docs/build.md)
 
-Additionally, there the following images, similar to the above:
+## Supported backends
 
-- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+| Backend | Target devices |
+| --- | --- |
+| [Metal](./docs/build.md#metal-build) | Apple Silicon |
+| [BLAS](./docs/build.md#blas-build) | All |
+| [BLIS](./docs/backend/BLIS.md) | All |
+| [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
+| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
+| [CUDA](./docs/build.md#cuda) | Nvidia GPU |
+| [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
+| [Vulkan](./docs/build.md#vulkan) | GPU |
 
-The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
+## Tools
 
-#### Usage
-
-The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
-
-Replace `/path/to/models` below with the actual path where you downloaded the models.
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
-```
-
-On completion, you are ready to play!
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
-```
-
-or with a light image:
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
-```
-
-or with a server image:
-
-```bash
-docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
-```
-
-### Docker With CUDA
-
-Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
-
-#### Building Locally
-
-```bash
-docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
-docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
-docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
-```
-
-You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
+### Prepare and Quantize
 
-The defaults are:
+> [!NOTE]
+> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
 
-- `CUDA_VERSION` set to `11.7.1`
-- `CUDA_DOCKER_ARCH` set to `all`
+To obtain the official LLaMA 2 weights please see the Obtaining and using the Facebook LLaMA 2 model section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
 
-The resulting images, are essentially the same as the non-CUDA images:
+Note: `convert.py` has been moved to `examples/convert_legacy_llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
+It does not support LLaMA 3, you can use `convert_hf_to_gguf.py` with LLaMA 3 downloaded from Hugging Face.
 
-1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
-2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
-3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
+To learn more about quantizing model, [read this documentation](./examples/quantize/README.md)
 
-#### Usage
+### Perplexity (measuring model quality)
 
-After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
+You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
+For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
 
-```bash
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
-```
+To learn more how to measure perplexity using llama.cpp, [read this documentation](./examples/perplexity/README.md)
 
-### Contributing
+## Contributing
 
 - Contributors can open PRs
 - Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Collaborators will be invited based on contributions
 - Any help with managing issues and PRs is very appreciated!
+- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
+- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
 - Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
 - A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
 
-### Coding guidelines
+## Other documentations
 
-- Avoid adding third-party dependencies, extra files, extra headers, etc.
-- Always consider cross-compatibility with other operating systems and architectures
-- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
-- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
-- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
-- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
-- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
-
-![matmul](media/matmul.png)
-
-### Docs
-
-- [main](./examples/main/README.md)
+- [main (cli)](./examples/main/README.md)
 - [server](./examples/server/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
-- [BLIS](./docs/BLIS.md)
-- [Performance troubleshooting](./docs/token_generation_performance_tips.md)
-- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
 - [GBNF grammars](./grammars/README.md)
+
+**Development documentations**
+
+- [How to build](./docs/build.md)
+- [Running on Docker](./docs/docker.md)
+- [Build on Android](./docs/android.md)
+- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md)
+- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+
+**Seminal papers and background on the models**
+
+If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
+- LLaMA:
+    - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
+    - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
+- GPT-3
+    - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
+- GPT-3.5 / InstructGPT / ChatGPT:
+    - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
+    - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
diff --git a/llm/llama.cpp/ci/run.sh b/llm/llama.cpp/ci/run.sh
index 3fc5f48b2e2af0761c0ccffd68b585ea8adec1e9..58022c7dc37e0210a327741c17dd38ab0199f94d 100644
--- a/llm/llama.cpp/ci/run.sh
+++ b/llm/llama.cpp/ci/run.sh
@@ -36,11 +36,11 @@ SRC=`pwd`
 CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUDA=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -50,7 +50,7 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         exit 1
     fi
 
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 ## helpers
 
@@ -103,6 +103,9 @@ function gg_run_ctest_debug {
 
     set -e
 
+    # Check cmake, make and ctest are installed
+    gg_check_build_requirements
+
     (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
 
@@ -131,6 +134,9 @@ function gg_run_ctest_release {
 
     set -e
 
+    # Check cmake, make and ctest are installed
+    gg_check_build_requirements
+
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
@@ -284,10 +290,10 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../examples/convert-legacy-llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -303,47 +309,47 @@ function gg_run_open_llama_7b_v2 {
 
     wiki_test="${path_wiki}/wiki.test.raw"
 
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -421,7 +427,7 @@ function gg_run_pythia_1_4b {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -437,45 +443,45 @@ function gg_run_pythia_1_4b {
 
     wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/llama-cli --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/llama-save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -550,10 +556,10 @@ function gg_run_pythia_2_8b {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -569,47 +575,47 @@ function gg_run_pythia_2_8b {
 
     wiki_test="${path_wiki}/wiki.test.raw"
 
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -688,19 +694,33 @@ function gg_run_embd_bge_small {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
 
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
     set +e
 }
 
+function gg_check_build_requirements {
+    if ! command -v cmake &> /dev/null; then
+        gg_printf 'cmake not found, please install'
+    fi
+
+    if ! command -v make &> /dev/null; then
+        gg_printf 'make not found, please install'
+    fi
+
+    if ! command -v ctest &> /dev/null; then
+        gg_printf 'ctest not found, please install'
+    fi
+}
+
 function gg_sum_embd_bge_small {
     gg_printf '### %s\n\n' "${ci}"
 
diff --git a/llm/llama.cpp/cmake/arm64-windows-llvm.cmake b/llm/llama.cpp/cmake/arm64-windows-llvm.cmake
index 46fba65149b293fceafe3502e97662988ca4ea8f..8023796800683240e32048180a1eb874ddab1963 100644
--- a/llm/llama.cpp/cmake/arm64-windows-llvm.cmake
+++ b/llm/llama.cpp/cmake/arm64-windows-llvm.cmake
@@ -9,7 +9,7 @@ set( CMAKE_CXX_COMPILER  clang++ )
 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
 
-set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast" )
+set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
 
 set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
diff --git a/llm/llama.cpp/cmake/build-info.cmake b/llm/llama.cpp/cmake/build-info.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..ea3dc55c83439a91c9289760bc17b1b0dcd69aa1
--- /dev/null
+++ b/llm/llama.cpp/cmake/build-info.cmake
@@ -0,0 +1,58 @@
+set(BUILD_NUMBER 0)
+set(BUILD_COMMIT "unknown")
+set(BUILD_COMPILER "unknown")
+set(BUILD_TARGET "unknown")
+
+# Look for git
+find_package(Git)
+if(NOT Git_FOUND)
+    find_program(GIT_EXECUTABLE NAMES git git.exe)
+    if(GIT_EXECUTABLE)
+        set(Git_FOUND TRUE)
+        message(STATUS "Found Git: ${GIT_EXECUTABLE}")
+    else()
+        message(WARNING "Git not found. Build info will not be accurate.")
+    endif()
+endif()
+
+# Get the commit count and hash
+if(Git_FOUND)
+    execute_process(
+        COMMAND ${GIT_EXECUTABLE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE HEAD
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
+    )
+    if (RES EQUAL 0)
+        set(BUILD_COMMIT ${HEAD})
+    endif()
+    execute_process(
+        COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE COUNT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
+    )
+    if (RES EQUAL 0)
+        set(BUILD_NUMBER ${COUNT})
+    endif()
+endif()
+
+if(MSVC)
+    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+else()
+    execute_process(
+        COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_COMPILER ${OUT})
+    execute_process(
+        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_TARGET ${OUT})
+endif()
diff --git a/llm/llama.cpp/cmake/git-vars.cmake b/llm/llama.cpp/cmake/git-vars.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..1a4c24ebf6adeb1126e626f56de601621179353d
--- /dev/null
+++ b/llm/llama.cpp/cmake/git-vars.cmake
@@ -0,0 +1,22 @@
+find_package(Git)
+
+# the commit's SHA1
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_SHA1
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the date of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_DATE
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the subject of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%s
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
diff --git a/llm/llama.cpp/cmake/llama-config.cmake.in b/llm/llama.cpp/cmake/llama-config.cmake.in
new file mode 100644
index 0000000000000000000000000000000000000000..f072b76a39d2e1589b328aa106e32285f61f4d7f
--- /dev/null
+++ b/llm/llama.cpp/cmake/llama-config.cmake.in
@@ -0,0 +1,90 @@
+set(LLAMA_VERSION      @LLAMA_INSTALL_VERSION@)
+set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
+set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
+set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
+
+set(GGML_BLAS       @GGML_BLAS@)
+set(GGML_CUDA       @GGML_CUDA@)
+set(GGML_METAL      @GGML_METAL@)
+set(GGML_HIPBLAS    @GGML_HIPBLAS@)
+set(GGML_ACCELERATE @GGML_ACCELERATE@)
+set(GGML_VULKAN @GGML_VULKAN@)
+set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)
+set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
+set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
+set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
+set(GGML_SYCL @GGML_SYCL@)
+set(GGML_OPENMP @GGML_OPENMP@)
+
+@PACKAGE_INIT@
+
+set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
+set_and_check(LLAMA_LIB_DIR     "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
+set_and_check(LLAMA_BIN_DIR     "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
+
+# Ensure transient dependencies satisfied
+
+find_package(Threads REQUIRED)
+
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+endif()
+
+if (GGML_BLAS)
+    find_package(BLAS REQUIRED)
+endif()
+
+if (GGML_CUDA)
+    find_package(CUDAToolkit REQUIRED)
+endif()
+
+if (GGML_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK Metal REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+endif()
+
+if (GGML_VULKAN)
+    find_package(Vulkan REQUIRED)
+endif()
+
+if (GGML_HIPBLAS)
+    find_package(hip REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+endif()
+
+if (GGML_SYCL)
+    find_package(IntelSYCL REQUIRED)
+    find_package(MKL REQUIRED)
+endif()
+
+if (GGML_OPENMP)
+    find_package(OpenMP REQUIRED)
+endif()
+
+
+find_library(ggml_LIBRARY ggml
+    REQUIRED
+    HINTS ${LLAMA_LIB_DIR})
+
+find_library(llama_LIBRARY llama
+    REQUIRED
+    HINTS ${LLAMA_LIB_DIR})
+
+set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@")
+set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
+
+add_library(llama UNKNOWN IMPORTED)
+
+set_target_properties(llama
+    PROPERTIES
+        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
+        INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
+        INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
+        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+        IMPORTED_LOCATION "${llama_LIBRARY}"
+        INTERFACE_COMPILE_FEATURES cxx_std_11
+        POSITION_INDEPENDENT_CODE ON )
+
+check_required_components(Llama)
diff --git a/llm/llama.cpp/cmake/llama.pc.in b/llm/llama.cpp/cmake/llama.pc.in
new file mode 100644
index 0000000000000000000000000000000000000000..326acbb6108fd02e13787371d6dbefcd511c97f8
--- /dev/null
+++ b/llm/llama.cpp/cmake/llama.pc.in
@@ -0,0 +1,10 @@
+prefix=@CMAKE_INSTALL_PREFIX@
+exec_prefix=${prefix}
+libdir=${exec_prefix}/lib
+includedir=${prefix}/include
+
+Name: llama
+Description: Port of Facebook's LLaMA model in C/C++
+Version: @PROJECT_VERSION@
+Libs: -L${libdir} -lllama
+Cflags: -I${includedir}
diff --git a/llm/llama.cpp/common/CMakeLists.txt b/llm/llama.cpp/common/CMakeLists.txt
index 0ec8d6d8d03b5319b7095ab84ccd60cc74ff44a8..761971d6881f38e29a963314c65fc19ed227acb9 100644
--- a/llm/llama.cpp/common/CMakeLists.txt
+++ b/llm/llama.cpp/common/CMakeLists.txt
@@ -1,5 +1,6 @@
 # common
 
+find_package(Threads REQUIRED)
 
 # Build info header
 #
@@ -36,7 +37,7 @@ add_custom_command(
     COMMENT "Generating build details from Git"
     COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
             -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/gen-build-info-cpp.cmake"
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
     DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
     VERBATIM
@@ -83,5 +84,5 @@ if (LLAMA_CURL)
 endif ()
 
 target_include_directories(${TARGET} PUBLIC .)
-target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama)
+target_compile_features   (${TARGET} PUBLIC cxx_std_11)
+target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
diff --git a/llm/llama.cpp/common/build-info.cpp.in b/llm/llama.cpp/common/build-info.cpp.in
new file mode 100644
index 0000000000000000000000000000000000000000..0b945aa68fff3e77b881328e4fd63f6e2f9017ee
--- /dev/null
+++ b/llm/llama.cpp/common/build-info.cpp.in
@@ -0,0 +1,4 @@
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
+char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";
diff --git a/llm/llama.cpp/common/common.cpp b/llm/llama.cpp/common/common.cpp
index 65103c3c294d36f20474305f26a89c8c5de6536c..2e8374d50cafad99cb9f59eccb4ac5c22ed6b32b 100644
--- a/llm/llama.cpp/common/common.cpp
+++ b/llm/llama.cpp/common/common.cpp
@@ -1,3 +1,7 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
 #include "common.h"
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
@@ -6,21 +10,21 @@
 #include "llama.h"
 
 #include 
-#include 
+#include 
 #include 
+#include 
+#include 
 #include 
 #include 
 #include 
-#include 
 #include 
+#include 
 #include 
 #include 
 #include 
 #include 
 #include 
 #include 
-#include 
-#include 
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include 
@@ -190,6 +194,12 @@ int32_t cpu_get_num_math() {
 // CLI argument parsing
 //
 
+void gpt_params_handle_hf_token(gpt_params & params) {
+    if (params.hf_token.empty() && std::getenv("HF_TOKEN")) {
+        params.hf_token = std::getenv("HF_TOKEN");
+    }
+}
+
 void gpt_params_handle_model_default(gpt_params & params) {
     if (!params.hf_repo.empty()) {
         // short-hand to avoid specifying --hf-file -> default it to --model
@@ -199,19 +209,13 @@ void gpt_params_handle_model_default(gpt_params & params) {
             }
             params.hf_file = params.model;
         } else if (params.model.empty()) {
-            std::string cache_directory = fs_get_cache_directory();
-            const bool success = fs_create_directory_with_parents(cache_directory);
-            if (!success) {
-                throw std::runtime_error("failed to create cache directory: " + cache_directory);
-            }
-            params.model = cache_directory + string_split(params.hf_file, '/').back();
+            params.model = fs_get_cache_file(string_split(params.hf_file, '/').back());
         }
     } else if (!params.model_url.empty()) {
         if (params.model.empty()) {
             auto f = string_split(params.model_url, '#').front();
             f = string_split(f, '?').front();
-            f = string_split(f, '/').back();
-            params.model =  "models/" + f;
+            params.model = fs_get_cache_file(string_split(f, '/').back());
         }
     } else if (params.model.empty()) {
         params.model = DEFAULT_MODEL_PATH;
@@ -237,15 +241,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         }
     }
 
-    if (params.prompt_cache_all &&
-            (params.interactive || params.interactive_first ||
-             params.instruct)) {
-
+    if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
     gpt_params_handle_model_default(params);
 
+    gpt_params_handle_hf_token(params);
+
     if (params.escape) {
         string_process_escapes(params.prompt);
         string_process_escapes(params.input_prefix);
@@ -265,39 +268,39 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
-    bool result = true;
+    const auto params_org = params; // the example can modify the default params
+
     try {
-        if (!gpt_params_parse_ex(argc, argv, params)) {
-            gpt_params_print_usage(argc, argv, gpt_params());
-            exit(0);
+        if (!gpt_params_parse_ex(argc, argv, params) || params.usage) {
+            params = params_org;
+            params.usage = true;
+            return false;
         }
-    }
-    catch (const std::invalid_argument & ex) {
+    } catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
-        gpt_params_print_usage(argc, argv, gpt_params());
-        exit(1);
+        params = params_org;
+        return false;
     }
-    return result;
+
+    return true;
 }
 
+#define CHECK_ARG if (++i >= argc) { invalid_param = true; return true; }
+
 bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) {
+    const char split_delim = ',';
+
     llama_sampling_params & sparams = params.sparams;
 
     if (arg == "-s" || arg == "--seed") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        // This is temporary, in the future the samplign state will be moved fully to llama_sampling_context.
+        CHECK_ARG
+        // TODO: this is temporary, in the future the sampling state will be moved fully to llama_sampling_context.
         params.seed = std::stoul(argv[i]);
         sparams.seed = std::stoul(argv[i]);
         return true;
     }
     if (arg == "-t" || arg == "--threads") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_threads = std::stoi(argv[i]);
         if (params.n_threads <= 0) {
             params.n_threads = std::thread::hardware_concurrency();
@@ -305,10 +308,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-tb" || arg == "--threads-batch") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_threads_batch = std::stoi(argv[i]);
         if (params.n_threads_batch <= 0) {
             params.n_threads_batch = std::thread::hardware_concurrency();
@@ -316,10 +316,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-td" || arg == "--threads-draft") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_threads_draft = std::stoi(argv[i]);
         if (params.n_threads_draft <= 0) {
             params.n_threads_draft = std::thread::hardware_concurrency();
@@ -327,10 +324,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-tbd" || arg == "--threads-batch-draft") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_threads_batch_draft = std::stoi(argv[i]);
         if (params.n_threads_batch_draft <= 0) {
             params.n_threads_batch_draft = std::thread::hardware_concurrency();
@@ -338,10 +332,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-p" || arg == "--prompt") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.prompt = argv[i];
         return true;
     }
@@ -349,11 +340,12 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.escape = true;
         return true;
     }
+    if (arg == "--no-escape") {
+        params.escape = false;
+        return true;
+    }
     if (arg == "--prompt-cache") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.path_prompt_cache = argv[i];
         return true;
     }
@@ -366,10 +358,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-bf" || arg == "--binary-file") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::ifstream file(argv[i], std::ios::binary);
         if (!file) {
             fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
@@ -385,10 +374,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-f" || arg == "--file") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::ifstream file(argv[i]);
         if (!file) {
             fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
@@ -403,67 +389,54 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         }
         return true;
     }
-    if (arg == "-n" || arg == "--n-predict") {
-        if (++i >= argc) {
+    if (arg == "--in-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
             invalid_param = true;
             return true;
         }
+        params.in_files.push_back(argv[i]);
+        return true;
+    }
+    if (arg == "-n" || arg == "--predict" || arg == "--n-predict") {
+        CHECK_ARG
         params.n_predict = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--top-k") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.top_k = std::stoi(argv[i]);
         return true;
     }
     if (arg == "-c" || arg == "--ctx-size") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_ctx = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--grp-attn-n" || arg == "-gan") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.grp_attn_n = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--grp-attn-w" || arg == "-gaw") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.grp_attn_w = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--rope-freq-base") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.rope_freq_base = std::stof(argv[i]);
         return true;
     }
     if (arg == "--rope-freq-scale") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.rope_freq_scale = std::stof(argv[i]);
         return true;
     }
     if (arg == "--rope-scaling") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::string value(argv[i]);
         /**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
         else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
@@ -472,217 +445,148 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "--rope-scale") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.rope_freq_scale = 1.0f / std::stof(argv[i]);
         return true;
     }
     if (arg == "--yarn-orig-ctx") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.yarn_orig_ctx = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--yarn-ext-factor") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.yarn_ext_factor = std::stof(argv[i]);
         return true;
     }
     if (arg == "--yarn-attn-factor") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.yarn_attn_factor = std::stof(argv[i]);
         return true;
     }
     if (arg == "--yarn-beta-fast") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.yarn_beta_fast = std::stof(argv[i]);
         return true;
     }
     if (arg == "--yarn-beta-slow") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.yarn_beta_slow = std::stof(argv[i]);
         return true;
     }
     if (arg == "--pooling") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::string value(argv[i]);
         /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; }
         else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; }
         else if (value == "cls") { params.pooling_type = LLAMA_POOLING_TYPE_CLS; }
+        else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; }
+        else { invalid_param = true; }
+        return true;
+    }
+    if (arg == "--attention") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; }
+        else if (value == "non-causal") { params.attention_type = LLAMA_ATTENTION_TYPE_NON_CAUSAL; }
         else { invalid_param = true; }
         return true;
     }
     if (arg == "--defrag-thold" || arg == "-dt") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.defrag_thold = std::stof(argv[i]);
         return true;
     }
     if (arg == "--samplers") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         const auto sampler_names = string_split(argv[i], ';');
         sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, true);
         return true;
     }
     if (arg == "--sampling-seq") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.samplers_sequence = llama_sampling_types_from_chars(argv[i]);
         return true;
     }
     if (arg == "--top-p") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.top_p = std::stof(argv[i]);
         return true;
     }
     if (arg == "--min-p") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.min_p = std::stof(argv[i]);
         return true;
     }
     if (arg == "--temp") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.temp = std::stof(argv[i]);
         sparams.temp = std::max(sparams.temp, 0.0f);
         return true;
     }
     if (arg == "--tfs") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.tfs_z = std::stof(argv[i]);
         return true;
     }
     if (arg == "--typical") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.typical_p = std::stof(argv[i]);
         return true;
     }
     if (arg == "--repeat-last-n") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.penalty_last_n = std::stoi(argv[i]);
         sparams.n_prev = std::max(sparams.n_prev, sparams.penalty_last_n);
         return true;
     }
     if (arg == "--repeat-penalty") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.penalty_repeat = std::stof(argv[i]);
         return true;
     }
     if (arg == "--frequency-penalty") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.penalty_freq = std::stof(argv[i]);
         return true;
     }
     if (arg == "--presence-penalty") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.penalty_present = std::stof(argv[i]);
         return true;
     }
     if (arg == "--dynatemp-range") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.dynatemp_range = std::stof(argv[i]);
         return true;
     }
     if (arg == "--dynatemp-exp") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.dynatemp_exponent = std::stof(argv[i]);
         return true;
     }
     if (arg == "--mirostat") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.mirostat = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--mirostat-lr") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.mirostat_eta = std::stof(argv[i]);
         return true;
     }
     if (arg == "--mirostat-ent") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.mirostat_tau = std::stof(argv[i]);
         return true;
     }
     if (arg == "--cfg-negative-prompt") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.cfg_negative_prompt = argv[i];
         return true;
     }
     if (arg == "--cfg-negative-prompt-file") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::ifstream file(argv[i]);
         if (!file) {
             fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
@@ -696,203 +600,136 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "--cfg-scale") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         sparams.cfg_scale = std::stof(argv[i]);
         return true;
     }
     if (arg == "-b" || arg == "--batch-size") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_batch = std::stoi(argv[i]);
         return true;
     }
     if (arg == "-ub" || arg == "--ubatch-size") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_ubatch = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--keep") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_keep = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--draft") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_draft = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--chunks") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_chunks = std::stoi(argv[i]);
         return true;
     }
     if (arg == "-np" || arg == "--parallel") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_parallel = std::stoi(argv[i]);
         return true;
     }
     if (arg == "-ns" || arg == "--sequences") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.n_sequences = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--p-split" || arg == "-ps") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.p_split = std::stof(argv[i]);
         return true;
     }
     if (arg == "-m" || arg == "--model") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.model = argv[i];
         return true;
     }
     if (arg == "-md" || arg == "--model-draft") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.model_draft = argv[i];
         return true;
     }
     if (arg == "-a" || arg == "--alias") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.model_alias = argv[i];
         return true;
     }
     if (arg == "-mu" || arg == "--model-url") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.model_url = argv[i];
         return true;
     }
-    if (arg == "-hfr" || arg == "--hf-repo") {
+    if (arg == "-hft" || arg == "--hf-token") {
         if (++i >= argc) {
-            invalid_param = true;
-            return true;
+          invalid_param = true;
+          return true;
         }
+        params.hf_token = argv[i];
+        return true;
+    }
+    if (arg == "-hfr" || arg == "--hf-repo") {
+        CHECK_ARG
         params.hf_repo = argv[i];
         return true;
     }
     if (arg == "-hff" || arg == "--hf-file") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.hf_file = argv[i];
         return true;
     }
     if (arg == "--lora") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.lora_adapter.emplace_back(argv[i], 1.0f);
-        params.use_mmap = false;
+        CHECK_ARG
+        params.lora_adapters.push_back({
+            std::string(argv[i]),
+            1.0,
+        });
         return true;
     }
     if (arg == "--lora-scaled") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        const char* lora_adapter = argv[i];
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
-        params.use_mmap = false;
+        CHECK_ARG
+        std::string lora_adapter = argv[i];
+        CHECK_ARG
+        params.lora_adapters.push_back({
+            lora_adapter,
+            std::stof(argv[i]),
+        });
         return true;
     }
-    if (arg == "--lora-base") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.lora_base = argv[i];
+    if (arg == "--lora-init-without-apply") {
+        params.lora_init_without_apply = true;
         return true;
     }
     if (arg == "--control-vector") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.control_vectors.push_back({ 1.0f, argv[i], });
         return true;
     }
     if (arg == "--control-vector-scaled") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         const char* fname = argv[i];
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.control_vectors.push_back({ std::stof(argv[i]), fname, });
         return true;
     }
     if (arg == "--control-vector-layer-range") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.control_vector_layer_start = std::stoi(argv[i]);
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.control_vector_layer_end = std::stoi(argv[i]);
         return true;
     }
     if (arg == "--mmproj") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.mmproj = argv[i];
         return true;
     }
     if (arg == "--image") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.image.emplace_back(argv[i]);
         return true;
     }
@@ -900,32 +737,35 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.interactive = true;
         return true;
     }
-    if (arg == "--interactive-specials") {
-        params.interactive_specials = true;
-        return true;
-    }
-    if (arg == "--special") {
+    if (arg == "-sp" || arg == "--special") {
         params.special = true;
         return true;
     }
-    if (arg == "--embedding") {
+    if (arg == "--embedding" || arg == "--embeddings") {
         params.embedding = true;
         return true;
     }
-    if (arg == "--interactive-first") {
-        params.interactive_first = true;
+    if (arg == "--embd-normalize") {
+        CHECK_ARG
+        params.embd_normalize = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "-ins" || arg == "--instruct") {
-        params.instruct = true;
+    if (arg == "--embd-output-format") {
+        CHECK_ARG
+        params.embd_out = argv[i];
         return true;
     }
-    if (arg == "-cnv" || arg == "--conversation") {
-        params.conversation = true;
+    if (arg == "--embd-separator") {
+        CHECK_ARG
+        params.embd_sep = argv[i];
+        return true;
+    }
+    if (arg == "-if" || arg == "--interactive-first") {
+        params.interactive_first = true;
         return true;
     }
-    if (arg == "-cml" || arg == "--chatml") {
-        params.chatml = true;
+    if (arg == "-cnv" || arg == "--conversation") {
+        params.conversation = true;
         return true;
     }
     if (arg == "--infill") {
@@ -948,7 +788,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.cache_type_v = argv[++i];
         return true;
     }
-    if (arg == "--multiline-input") {
+    if (arg == "-mli" || arg == "--multiline-input") {
         params.multiline_input = true;
         return true;
     }
@@ -960,11 +800,15 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.cont_batching = true;
         return true;
     }
+    if (arg == "-nocb" || arg == "--no-cont-batching") {
+        params.cont_batching = false;
+        return true;
+    }
     if (arg == "-fa" || arg == "--flash-attn") {
         params.flash_attn = true;
         return true;
     }
-    if (arg == "--color") {
+    if (arg == "-co" || arg == "--color") {
         params.use_color = true;
         return true;
     }
@@ -972,46 +816,34 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.use_mlock = true;
         return true;
     }
-    if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+    if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+        CHECK_ARG
         params.n_gpu_layers = std::stoi(argv[i]);
         if (!llama_supports_gpu_offload()) {
-            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
+            fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
         }
         return true;
     }
-    if (arg == "--gpu-layers-draft" || arg == "-ngld" || arg == "--n-gpu-layers-draft") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+    if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--gpu-layers-draft") {
+        CHECK_ARG
         params.n_gpu_layers_draft = std::stoi(argv[i]);
         if (!llama_supports_gpu_offload()) {
-            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
+            fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
         }
         return true;
     }
     if (arg == "--main-gpu" || arg == "-mg") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.main_gpu = std::stoi(argv[i]);
-#ifndef GGML_USE_CUDA_SYCL
-        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL. Setting the main GPU has no effect.\n");
-#endif // GGML_USE_CUDA_SYCL
+#ifndef GGML_USE_CUDA_SYCL_VULKAN
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the main GPU has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL_VULKAN
         return true;
     }
     if (arg == "--split-mode" || arg == "-sm") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::string arg_next = argv[i];
         if (arg_next == "none") {
             params.split_mode = LLAMA_SPLIT_MODE_NONE;
@@ -1030,16 +862,13 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
             invalid_param = true;
             return true;
         }
-#ifndef GGML_USE_CUDA_SYCL
-        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUDA_SYCL
+#ifndef GGML_USE_CUDA_SYCL_VULKAN
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL_VULKAN
         return true;
     }
     if (arg == "--tensor-split" || arg == "-ts") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::string arg_next = argv[i];
 
         // split string by , and /
@@ -1064,10 +893,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "--rpc") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.rpc_servers = argv[i];
         return true;
     }
@@ -1076,10 +902,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "--numa") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         std::string value(argv[i]);
         /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
         else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
@@ -1087,6 +910,15 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         else { invalid_param = true; }
         return true;
     }
+    if (arg == "-v" || arg == "--verbose") {
+        params.verbosity = 1;
+        return true;
+    }
+    if (arg == "--verbosity") {
+        CHECK_ARG
+        params.verbosity = std::stoi(argv[i]);
+        return true;
+    }
     if (arg == "--verbose-prompt") {
         params.verbose_prompt = true;
         return true;
@@ -1096,18 +928,12 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-r" || arg == "--reverse-prompt") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.antiprompt.emplace_back(argv[i]);
         return true;
     }
     if (arg == "-ld" || arg == "--logdir") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         params.logdir = argv[i];
 
         if (params.logdir.back() != DIRECTORY_SEPARATOR) {
@@ -1116,209 +942,400 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         return true;
     }
     if (arg == "-lcs" || arg == "--lookup-cache-static") {
-        if (++i >= argc) {
+        CHECK_ARG
+        params.lookup_cache_static = argv[i];
+        return true;
+    }
+    if (arg == "-lcd" || arg == "--lookup-cache-dynamic") {
+        CHECK_ARG
+        params.lookup_cache_dynamic = argv[i];
+        return true;
+    }
+    if (arg == "--save-all-logits" || arg == "--kl-divergence-base") {
+        CHECK_ARG
+        params.logits_file = argv[i];
+        return true;
+    }
+    if (arg == "--perplexity" || arg == "--all-logits") {
+        params.logits_all = true;
+        return true;
+    }
+    if (arg == "--ppl-stride") {
+        CHECK_ARG
+        params.ppl_stride = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--ppl-output-type") {
+        CHECK_ARG
+        params.ppl_output_type = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-ptc" || arg == "--print-token-count") {
+        CHECK_ARG
+        params.n_print = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--check-tensors") {
+        params.check_tensors = true;
+        return true;
+    }
+    if (arg == "--hellaswag") {
+        params.hellaswag = true;
+        return true;
+    }
+    if (arg == "--hellaswag-tasks") {
+        CHECK_ARG
+        params.hellaswag_tasks = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--winogrande") {
+        params.winogrande = true;
+        return true;
+    }
+    if (arg == "--winogrande-tasks") {
+        CHECK_ARG
+        params.winogrande_tasks = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--multiple-choice") {
+        params.multiple_choice = true;
+        return true;
+    }
+    if (arg == "--multiple-choice-tasks") {
+        CHECK_ARG
+        params.multiple_choice_tasks = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--kl-divergence") {
+        params.kl_divergence = true;
+        return true;
+    }
+    if (arg == "--ignore-eos") {
+        params.ignore_eos = true;
+        return true;
+    }
+    if (arg == "--penalize-nl") {
+        sparams.penalize_nl = true;
+        return true;
+    }
+    if (arg == "-l" || arg == "--logit-bias") {
+        CHECK_ARG
+        std::stringstream ss(argv[i]);
+        llama_token key;
+        char sign;
+        std::string value_str;
+        try {
+            if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
+                sparams.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
+            }
+            else {
+                throw std::exception();
+            }
+        }
+        catch (const std::exception&) {
             invalid_param = true;
             return true;
         }
-        params.lookup_cache_static = argv[i];
         return true;
     }
-    if (arg == "-lcd" || arg == "--lookup-cache-dynamic") {
-        if (++i >= argc) {
+    if (arg == "-h" || arg == "--help" || arg == "--usage"  ) {
+        params.usage = true;
+        return true;
+    }
+    if (arg == "--version") {
+        fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+        fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+        exit(0);
+    }
+    if (arg == "--in-prefix-bos") {
+        params.input_prefix_bos = true;
+        params.enable_chat_template = false;
+        return true;
+    }
+    if (arg == "--in-prefix") {
+        CHECK_ARG
+        params.input_prefix = argv[i];
+        params.enable_chat_template = false;
+        return true;
+    }
+    if (arg == "--in-suffix") {
+        CHECK_ARG
+        params.input_suffix = argv[i];
+        params.enable_chat_template = false;
+        return true;
+    }
+    if (arg == "--spm-infill") {
+        params.spm_infill = true;
+        return true;
+    }
+    if (arg == "--grammar") {
+        CHECK_ARG
+        sparams.grammar = argv[i];
+        return true;
+    }
+    if (arg == "--grammar-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
             invalid_param = true;
             return true;
         }
-        params.lookup_cache_dynamic = argv[i];
+        std::copy(
+            std::istreambuf_iterator(file),
+            std::istreambuf_iterator(),
+            std::back_inserter(sparams.grammar)
+        );
         return true;
     }
-    if (arg == "--save-all-logits" || arg == "--kl-divergence-base") {
-        if (++i >= argc) {
+    if (arg == "-j" || arg == "--json-schema") {
+        CHECK_ARG
+        sparams.grammar = json_schema_to_grammar(json::parse(argv[i]));
+        return true;
+    }
+    if (arg == "--override-kv") {
+        CHECK_ARG
+        if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
+            fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
             invalid_param = true;
             return true;
         }
-        params.logits_file = argv[i];
         return true;
     }
-    if (arg == "--perplexity" || arg == "--all-logits") {
-        params.logits_all = true;
+    if (arg == "--host") {
+        CHECK_ARG
+        params.hostname = argv[i];
         return true;
     }
-    if (arg == "--ppl-stride") {
-        if (++i >= argc) {
+    if (arg == "--port") {
+        CHECK_ARG
+        params.port = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--path") {
+        CHECK_ARG
+        params.public_path = argv[i];
+        return true;
+    }
+    if (arg == "--api-key") {
+        CHECK_ARG
+        params.api_keys.push_back(argv[i]);
+        return true;
+    }
+    if (arg == "--api-key-file") {
+        CHECK_ARG
+        std::ifstream key_file(argv[i]);
+        if (!key_file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
             invalid_param = true;
             return true;
         }
-        params.ppl_stride = std::stoi(argv[i]);
+        std::string key;
+        while (std::getline(key_file, key)) {
+            if (!key.empty()) {
+                params.api_keys.push_back(key);
+            }
+        }
+        key_file.close();
         return true;
     }
-    if (arg == "-ptc" || arg == "--print-token-count") {
-        if (++i >= argc) {
+    if (arg == "--ssl-key-file") {
+        CHECK_ARG
+        params.ssl_file_key = argv[i];
+        return true;
+    }
+    if (arg == "--ssl-cert-file") {
+        CHECK_ARG
+        params.ssl_file_cert = argv[i];
+        return true;
+    }
+    if (arg == "--timeout" || arg == "-to") {
+        CHECK_ARG
+        params.timeout_read  = std::stoi(argv[i]);
+        params.timeout_write = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "--threads-http") {
+        CHECK_ARG
+        params.n_threads_http = std::stoi(argv[i]);
+        return true;
+    }
+    if (arg == "-spf" || arg == "--system-prompt-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i]);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+            invalid_param = true;
+            return true;
+        }
+        std::string system_prompt;
+        std::copy(
+                std::istreambuf_iterator(file),
+                std::istreambuf_iterator(),
+                std::back_inserter(system_prompt)
+                );
+        params.system_prompt = system_prompt;
+        return true;
+    }
+    if (arg == "--log-format") {
+        CHECK_ARG
+        if (std::strcmp(argv[i], "json") == 0) {
+            params.log_json = true;
+        } else if (std::strcmp(argv[i], "text") == 0) {
+            params.log_json = false;
+        } else {
             invalid_param = true;
             return true;
         }
-        params.n_print = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "--check-tensors") {
-        params.check_tensors = true;
+    if (arg == "--no-slots") {
+        params.endpoint_slots = false;
         return true;
     }
-    if (arg == "--ppl-output-type") {
-        if (++i >= argc) {
+    if (arg == "--metrics") {
+        params.endpoint_metrics = true;
+        return true;
+    }
+    if (arg == "--slot-save-path") {
+        CHECK_ARG
+        params.slot_save_path = argv[i];
+        // if doesn't end with DIRECTORY_SEPARATOR, add it
+        if (!params.slot_save_path.empty() && params.slot_save_path[params.slot_save_path.size() - 1] != DIRECTORY_SEPARATOR) {
+            params.slot_save_path += DIRECTORY_SEPARATOR;
+        }
+        return true;
+    }
+    if (arg == "--chat-template") {
+        CHECK_ARG
+        if (!llama_chat_verify_template(argv[i])) {
+            fprintf(stderr, "error: the supplied chat template is not supported: %s\n", argv[i]);
+            fprintf(stderr, "note: llama.cpp does not use jinja parser, we only support commonly used templates\n");
             invalid_param = true;
             return true;
         }
-        params.ppl_output_type = std::stoi(argv[i]);
+        params.chat_template = argv[i];
+        return true;
+    }
+    if (arg == "--slot-prompt-similarity" || arg == "-sps") {
+        CHECK_ARG
+        params.slot_prompt_similarity = std::stof(argv[i]);
+        return true;
+    }
+    if (arg == "-pps") {
+        params.is_pp_shared = true;
         return true;
     }
-    if (arg == "--hellaswag") {
-        params.hellaswag = true;
+    if (arg == "-npp") {
+        CHECK_ARG
+        auto p = string_split(argv[i], split_delim);
+        params.n_pp.insert(params.n_pp.end(), p.begin(), p.end());
         return true;
     }
-    if (arg == "--hellaswag-tasks") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.hellaswag_tasks = std::stoi(argv[i]);
+    if (arg == "-ntg") {
+        CHECK_ARG
+        auto p = string_split(argv[i], split_delim);
+        params.n_tg.insert(params.n_tg.end(), p.begin(), p.end());
         return true;
     }
-    if (arg == "--winogrande") {
-        params.winogrande = true;
+    if (arg == "-npl") {
+        CHECK_ARG
+        auto p = string_split(argv[i], split_delim);
+        params.n_pl.insert(params.n_pl.end(), p.begin(), p.end());
         return true;
     }
-    if (arg == "--winogrande-tasks") {
-        if (++i >= argc) {
+    if (arg == "--context-file") {
+        CHECK_ARG
+        std::ifstream file(argv[i], std::ios::binary);
+        if (!file) {
+            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
             invalid_param = true;
             return true;
         }
-        params.winogrande_tasks = std::stoi(argv[i]);
+        params.context_files.push_back(argv[i]);
         return true;
     }
-    if (arg == "--multiple-choice") {
-        params.multiple_choice = true;
+    if (arg == "--chunk-size") {
+        CHECK_ARG
+        params.chunk_size = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "--multiple-choice-tasks") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.multiple_choice_tasks = std::stoi(argv[i]);
+    if (arg == "--chunk-separator") {
+        CHECK_ARG
+        params.chunk_separator = argv[i];
         return true;
     }
-    if (arg == "--kl-divergence") {
-        params.kl_divergence = true;
+    if (arg == "--junk") {
+        CHECK_ARG
+        params.n_junk = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "--ignore-eos") {
-        params.ignore_eos = true;
+    if (arg == "--pos") {
+        CHECK_ARG
+        params.i_pos = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "--penalize-nl") {
-        sparams.penalize_nl = true;
+    if (arg == "-o" || arg == "--output" || arg == "--output-file") {
+        CHECK_ARG
+        params.out_file = argv[i];
+        params.cvector_outfile = argv[i];
+        params.lora_outfile = argv[i];
         return true;
     }
-    if (arg == "-l" || arg == "--logit-bias") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        std::stringstream ss(argv[i]);
-        llama_token key;
-        char sign;
-        std::string value_str;
-        try {
-            if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
-                sparams.logit_bias[key] = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
-            }
-            else {
-                throw std::exception();
-            }
-        }
-        catch (const std::exception&) {
-            invalid_param = true;
-            return true;
-        }
+    if (arg == "-ofreq" || arg == "--output-frequency") {
+        CHECK_ARG
+        params.n_out_freq = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "-h" || arg == "--help") {
-        gpt_params_print_usage(argc, argv, gpt_params());
-        exit(0);
+    if (arg == "--save-frequency") {
+        CHECK_ARG
+        params.n_save_freq = std::stoi(argv[i]);
+        return true;
     }
-    if (arg == "--version") {
-        fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-        fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
-        exit(0);
+    if (arg == "--process-output") {
+        params.process_output = true;
+        return true;
     }
-    if (arg == "--random-prompt") {
-        params.random_prompt = true;
+    if (arg == "--no-ppl") {
+        params.compute_ppl = false;
         return true;
     }
-    if (arg == "--in-prefix-bos") {
-        params.input_prefix_bos = true;
+    if (arg == "--chunk" || arg == "--from-chunk") {
+        CHECK_ARG
+        params.i_chunk = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "--in-prefix") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.input_prefix = argv[i];
+    // cvector params
+    if (arg == "--positive-file") {
+        CHECK_ARG
+        params.cvector_positive_file = argv[i];
         return true;
     }
-    if (arg == "--in-suffix") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.input_suffix = argv[i];
+    if (arg == "--negative-file") {
+        CHECK_ARG
+        params.cvector_negative_file = argv[i];
         return true;
     }
-    if (arg == "--grammar") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        sparams.grammar = argv[i];
+    if (arg == "--pca-batch") {
+        CHECK_ARG
+        params.n_pca_batch = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "--grammar-file") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        std::ifstream file(argv[i]);
-        if (!file) {
-            fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
-            invalid_param = true;
-            return true;
-        }
-        std::copy(
-            std::istreambuf_iterator(file),
-            std::istreambuf_iterator(),
-            std::back_inserter(sparams.grammar)
-        );
+    if (arg == "--pca-iter") {
+        CHECK_ARG
+        params.n_pca_iterations = std::stoi(argv[i]);
         return true;
     }
-    if (arg == "-j" || arg == "--json-schema") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        sparams.grammar = json_schema_to_grammar(json::parse(argv[i]));
+    if (arg == "--method") {
+        CHECK_ARG
+        std::string value(argv[i]);
+        /**/ if (value == "pca") { params.cvector_dimre_method = DIMRE_METHOD_PCA; }
+        else if (value == "mean") { params.cvector_dimre_method = DIMRE_METHOD_MEAN; }
+        else { invalid_param = true; }
         return true;
     }
-    if (arg == "--override-kv") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
-            fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
-            invalid_param = true;
-            return true;
-        }
+    if (arg == "--no-warmup") {
+        params.warmup = false;
         return true;
     }
 #ifndef LOG_DISABLE_LOGS
@@ -1332,10 +1349,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         // We have a matching known parameter requiring an argument,
         //  now we need to check if there is anything after this argv
         //  and flag invalid_param or parse it.
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
+        CHECK_ARG
         if (!log_param_pair_parse( /*check_but_dont_parse*/ false, argv[i - 1], argv[i])) {
             invalid_param = true;
             return true;
@@ -1348,6 +1362,16 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     return false;
 }
 
+#ifdef __GNUC__
+#ifdef __MINGW32__
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
+#endif
+
 void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     const llama_sampling_params & sparams = params.sparams;
 
@@ -1359,198 +1383,341 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     }
     sampler_type_names.pop_back();
 
-    printf("\n");
-    printf("usage: %s [options]\n", argv[0]);
-    printf("\n");
-    printf("options:\n");
-    printf("  -h, --help            show this help message and exit\n");
-    printf("  --version             show version and build info\n");
-    printf("  -i, --interactive     run in interactive mode\n");
-    printf("  --special             special tokens output enabled\n");
-    printf("  --interactive-specials allow special tokens in user text, in interactive mode\n");
-    printf("  --interactive-first   run in interactive mode and wait for input right away\n");
-    printf("  -cnv, --conversation  run in conversation mode (does not print special tokens and suffix/prefix)\n");
-    printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
-    printf("  -cml, --chatml        run in chatml mode (use with ChatML-compatible models)\n");
-    printf("  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
-    printf("  -r PROMPT, --reverse-prompt PROMPT\n");
-    printf("                        halt generation at PROMPT, return control in interactive mode\n");
-    printf("                        (can be specified more than once for multiple prompts).\n");
-    printf("  --color               colorise output to distinguish prompt and user input from generations\n");
-    printf("  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    printf("  -t N, --threads N     number of threads to use during generation (default: %d)\n", params.n_threads);
-    printf("  -tb N, --threads-batch N\n");
-    printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
-    printf("  -td N, --threads-draft N");
-    printf("                        number of threads to use during generation (default: same as --threads)\n");
-    printf("  -tbd N, --threads-batch-draft N\n");
-    printf("                        number of threads to use during batch and prompt processing (default: same as --threads-draft)\n");
-    printf("  -p PROMPT, --prompt PROMPT\n");
-    printf("                        prompt to start generation with (default: empty)\n");
-    printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    printf("  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
-    printf("  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
-    printf("                        not supported with --interactive or other interactive options\n");
-    printf("  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
-    printf("  --random-prompt       start with a randomized prompt.\n");
-    printf("  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
-    printf("  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
-    printf("  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
-    printf("  -f FNAME, --file FNAME\n");
-    printf("                        prompt file to start generation.\n");
-    printf("  -bf FNAME, --binary-file FNAME\n");
-    printf("                        binary file containing multiple choice tasks.\n");
-    printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
-    printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
-    printf("  -b N, --batch-size N  logical maximum batch size (default: %d)\n", params.n_batch);
-    printf("  -ub N, --ubatch-size N\n");
-    printf("                        physical maximum batch size (default: %d)\n", params.n_ubatch);
-    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\'\n");
-    printf("                        (default: %s)\n", sampler_type_names.c_str());
-    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
-    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
-    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
-    printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
-    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z);
-    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p);
-    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n);
-    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.penalty_repeat);
-    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_present);
-    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_freq);
-    printf("  --dynatemp-range N    dynamic temperature range (default: %.1f, 0.0 = disabled)\n", (double)sparams.dynatemp_range);
-    printf("  --dynatemp-exp N      dynamic temperature exponent (default: %.1f)\n", (double)sparams.dynatemp_exponent);
-    printf("  --mirostat N          use Mirostat sampling.\n");
-    printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", sparams.mirostat);
-    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)sparams.mirostat_eta);
-    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)sparams.mirostat_tau);
-    printf("  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
-    printf("                        modifies the likelihood of token appearing in the completion,\n");
-    printf("                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
-    printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
-    printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
-    printf("  --grammar-file FNAME  file to read grammar from\n");
-    printf("  -j SCHEMA, --json-schema SCHEMA\n");
-    printf("                        JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object.\n");
-    printf("                        For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead\n");
-    printf("  --cfg-negative-prompt PROMPT\n");
-    printf("                        negative prompt to use for guidance. (default: empty)\n");
-    printf("  --cfg-negative-prompt-file FNAME\n");
-    printf("                        negative prompt file to use for guidance. (default: empty)\n");
-    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", sparams.cfg_scale);
-    printf("  --rope-scaling {none,linear,yarn}\n");
-    printf("                        RoPE frequency scaling method, defaults to linear unless specified by the model\n");
-    printf("  --rope-scale N        RoPE context scaling factor, expands context by a factor of N\n");
-    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
-    printf("  --rope-freq-scale N   RoPE frequency scaling factor, expands context by a factor of 1/N\n");
-    printf("  --yarn-orig-ctx N     YaRN: original context size of model (default: 0 = model training context size)\n");
-    printf("  --yarn-ext-factor N   YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
-    printf("  --yarn-attn-factor N  YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
-    printf("  --yarn-beta-slow N    YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
-    printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
-    printf("  --pooling {none,mean,cls}\n");
-    printf("                        pooling type for embeddings, use model default if unspecified\n");
-    printf("  -dt N, --defrag-thold N\n");
-    printf("                        KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
-    printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    printf("  --penalize-nl         penalize newline tokens\n");
-    printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
-    printf("  --all-logits          return logits for all tokens in the batch (default: disabled)\n");
-    printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
-    printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
-    printf("  --winogrande          compute Winogrande score over random tasks from datafile supplied with -f\n");
-    printf("  --winogrande-tasks N  number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks);
-    printf("  --multiple-choice     compute multiple choice score over random tasks from datafile supplied with -f\n");
-    printf("  --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks);
-    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base\n");
-    printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
-    printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
-    printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
-    printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
-    printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
-    printf("  -ps N, --p-split N    speculative decoding split probability (default: %.1f)\n", (double)params.p_split);
-    printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
-    printf("  -fa, --flash-attn     enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled");
-    printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA. see examples/llava/README.md\n");
-    printf("  --image IMAGE_FILE    path to an image file. use with multimodal models. Specify multiple times for batching\n");
+    struct option_info {
+        LLAMA_COMMON_ATTRIBUTE_FORMAT(4, 5)
+        option_info(const std::string & tags, const char * args, const char * desc, ...) : tags(tags), args(args), desc(desc) {
+            va_list args_list;
+            va_start(args_list, desc);
+            char buffer[1024];
+            vsnprintf(buffer, sizeof(buffer), desc, args_list);
+            va_end(args_list);
+            this->desc = buffer;
+        }
+
+        option_info(const std::string & grp) : grp(grp) {}
+
+        std::string tags;
+        std::string args;
+        std::string desc;
+        std::string grp;
+    };
+
+    std::vector options;
+
+    // TODO: filter by tags
+
+    options.push_back({ "general" });
+    options.push_back({ "*",           "-h,    --help, --usage",        "print usage and exit" });
+    options.push_back({ "*",           "       --version",              "show version and build info" });
+    options.push_back({ "*",           "-v,    --verbose",              "print verbose information" });
+    options.push_back({ "*",           "       --verbosity N",          "set specific verbosity level (default: %d)", params.verbosity });
+    options.push_back({ "*",           "       --verbose-prompt",       "print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false" });
+    options.push_back({ "*",           "       --no-display-prompt",    "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" });
+    options.push_back({ "*",           "-co,   --color",                "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" });
+    options.push_back({ "*",           "-s,    --seed SEED",            "RNG seed (default: %d, use random seed for < 0)", params.seed });
+    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during generation (default: %d)", params.n_threads });
+    options.push_back({ "*",           "-tb,   --threads-batch N",      "number of threads to use during batch and prompt processing (default: same as --threads)" });
+    options.push_back({ "speculative", "-td,   --threads-draft N",      "number of threads to use during generation (default: same as --threads)" });
+    options.push_back({ "speculative", "-tbd,  --threads-batch-draft N",
+                                                                        "number of threads to use during batch and prompt processing (default: same as --threads-draft)" });
+    options.push_back({ "speculative", "       --draft N",              "number of tokens to draft for speculative decoding (default: %d)", params.n_draft });
+    options.push_back({ "speculative", "-ps,   --p-split N",            "speculative decoding split probability (default: %.1f)", (double)params.p_split });
+    options.push_back({ "*",           "-lcs,  --lookup-cache-static FNAME",
+                                                                        "path to static lookup cache to use for lookup decoding (not updated by generation)" });
+    options.push_back({ "*",           "-lcd,  --lookup-cache-dynamic FNAME",
+                                                                        "path to dynamic lookup cache to use for lookup decoding (updated by generation)" });
+
+    options.push_back({ "*",           "-c,    --ctx-size N",           "size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx });
+    options.push_back({ "*",           "-n,    --predict N",            "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict });
+    options.push_back({ "*",           "-b,    --batch-size N",         "logical maximum batch size (default: %d)", params.n_batch });
+    options.push_back({ "*",           "-ub,   --ubatch-size N",        "physical maximum batch size (default: %d)", params.n_ubatch });
+    options.push_back({ "*",           "       --keep N",               "number of tokens to keep from the initial prompt (default: %d, -1 = all)", params.n_keep });
+    options.push_back({ "*",           "       --chunks N",             "max number of chunks to process (default: %d, -1 = all)", params.n_chunks });
+    options.push_back({ "*",           "-fa,   --flash-attn",           "enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled" });
+    options.push_back({ "*",           "-p,    --prompt PROMPT",        "prompt to start generation with\n"
+                                                                        "in conversation mode, this will be used as system prompt\n"
+                                                                        "(default: '%s')", params.prompt.c_str() });
+    options.push_back({ "*",           "-f,    --file FNAME",           "a file containing the prompt (default: none)" });
+    options.push_back({ "*",           "       --in-file FNAME",        "an input file (repeat to specify multiple files)" });
+    options.push_back({ "*",           "-bf,   --binary-file FNAME",    "binary file containing the prompt (default: none)" });
+    options.push_back({ "*",           "-e,    --escape",               "process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false" });
+    options.push_back({ "*",           "       --no-escape",            "do not process escape sequences" });
+    options.push_back({ "main",        "-ptc,  --print-token-count N",  "print token count every N tokens (default: %d)", params.n_print });
+    options.push_back({ "main",        "       --prompt-cache FNAME",   "file to cache prompt state for faster startup (default: none)" });
+    options.push_back({ "main",        "       --prompt-cache-all",     "if specified, saves user input and generations to cache as well\n"
+                                                                        "not supported with --interactive or other interactive options" });
+    options.push_back({ "main",        "       --prompt-cache-ro",      "if specified, uses the prompt cache but does not update it" });
+    options.push_back({ "main",        "-r,    --reverse-prompt PROMPT",
+                                                                        "halt generation at PROMPT, return control in interactive mode\n"
+                                                                        "can be specified more than once for multiple prompts" });
+    options.push_back({ "main",        "-sp,   --special",              "special tokens output enabled (default: %s)", params.special ? "true" : "false" });
+    options.push_back({ "main",        "-cnv,  --conversation",         "run in conversation mode, does not print special tokens and suffix/prefix\n"
+                                                                        "if suffix/prefix are not specified, default chat template will be used\n"
+                                                                        "(default: %s)", params.conversation ? "true" : "false" });
+    options.push_back({ "main infill", "-i,    --interactive",          "run in interactive mode (default: %s)", params.interactive ? "true" : "false" });
+    options.push_back({ "main infill", "-if,   --interactive-first",    "run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false" });
+    options.push_back({ "main infill", "-mli,  --multiline-input",      "allows you to write or paste multiple lines without ending each in '\\'" });
+    options.push_back({ "main infill", "       --in-prefix-bos",        "prefix BOS to user inputs, preceding the `--in-prefix` string" });
+    options.push_back({ "main infill", "       --in-prefix STRING",     "string to prefix user inputs with (default: empty)" });
+    options.push_back({ "main infill", "       --in-suffix STRING",     "string to suffix after user inputs with (default: empty)" });
+    options.push_back({ "main",        "       --no-warmup",            "skip warming up the model with an empty run" });
+    options.push_back({ "server infill",
+                                       "       --spm-infill",           "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", params.spm_infill ? "enabled" : "disabled" });
+
+    options.push_back({ "sampling" });
+    options.push_back({ "*",           "       --samplers SAMPLERS",    "samplers that will be used for generation in the order, separated by \';\'\n"
+                                                                        "(default: %s)", sampler_type_names.c_str() });
+    options.push_back({ "*",           "       --sampling-seq SEQUENCE",
+                                                                        "simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str() });
+    options.push_back({ "*",           "       --ignore-eos",           "ignore end of stream token and continue generating (implies --logit-bias EOS-inf)" });
+    options.push_back({ "*",           "       --penalize-nl",          "penalize newline tokens (default: %s)", sparams.penalize_nl ? "true" : "false" });
+    options.push_back({ "*",           "       --temp N",               "temperature (default: %.1f)", (double)sparams.temp });
+    options.push_back({ "*",           "       --top-k N",              "top-k sampling (default: %d, 0 = disabled)", sparams.top_k });
+    options.push_back({ "*",           "       --top-p N",              "top-p sampling (default: %.1f, 1.0 = disabled)", (double)sparams.top_p });
+    options.push_back({ "*",           "       --min-p N",              "min-p sampling (default: %.1f, 0.0 = disabled)", (double)sparams.min_p });
+    options.push_back({ "*",           "       --tfs N",                "tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)sparams.tfs_z });
+    options.push_back({ "*",           "       --typical N",            "locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)sparams.typical_p });
+    options.push_back({ "*",           "       --repeat-last-n N",      "last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", sparams.penalty_last_n });
+    options.push_back({ "*",           "       --repeat-penalty N",     "penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)sparams.penalty_repeat });
+    options.push_back({ "*",           "       --presence-penalty N",   "repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)sparams.penalty_present });
+    options.push_back({ "*",           "       --frequency-penalty N",  "repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)sparams.penalty_freq });
+    options.push_back({ "*",           "       --dynatemp-range N",     "dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)sparams.dynatemp_range });
+    options.push_back({ "*",           "       --dynatemp-exp N",       "dynamic temperature exponent (default: %.1f)", (double)sparams.dynatemp_exponent });
+    options.push_back({ "*",           "       --mirostat N",           "use Mirostat sampling.\n"
+                                                                        "Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n"
+                                                                        "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", sparams.mirostat });
+    options.push_back({ "*",           "       --mirostat-lr N",        "Mirostat learning rate, parameter eta (default: %.1f)", (double)sparams.mirostat_eta });
+    options.push_back({ "*",           "       --mirostat-ent N",       "Mirostat target entropy, parameter tau (default: %.1f)", (double)sparams.mirostat_tau });
+    options.push_back({ "*",           "       -l TOKEN_ID(+/-)BIAS",   "modifies the likelihood of token appearing in the completion,\n"
+                                                                        "i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n"
+                                                                        "or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'" });
+    options.push_back({ "main",        "       --cfg-negative-prompt PROMPT",
+                                                                        "negative prompt to use for guidance (default: '%s')", sparams.cfg_negative_prompt.c_str() });
+    options.push_back({ "main",        "       --cfg-negative-prompt-file FNAME",
+                                                                        "negative prompt file to use for guidance" });
+    options.push_back({ "main",        "       --cfg-scale N",          "strength of guidance (default: %.1f, 1.0 = disable)", (double)sparams.cfg_scale });
+    options.push_back({ "main",        "       --chat-template JINJA_TEMPLATE",
+                                                                        "set custom jinja chat template (default: template taken from model's metadata)\n"
+                                                                        "if suffix/prefix are specified, template will be disabled\n"
+                                                                        "only commonly used templates are accepted:\n"
+                                                                        "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
+    options.push_back({ "grammar" });
+    options.push_back({ "*",           "       --grammar GRAMMAR",      "BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", sparams.grammar.c_str() });
+    options.push_back({ "*",           "       --grammar-file FNAME",   "file to read grammar from" });
+    options.push_back({ "*",           "-j,    --json-schema SCHEMA",
+                                                                        "JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\n"
+                                                                        "For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead" });
+
+    options.push_back({ "embedding" });
+    options.push_back({ "embedding",   "       --pooling {none,mean,cls,last}",
+                                                                        "pooling type for embeddings, use model default if unspecified" });
+    options.push_back({ "embedding",   "       --attention {causal,non-causal}",
+                                                                        "attention type for embeddings, use model default if unspecified" });
+
+    options.push_back({ "context hacking" });
+    options.push_back({ "*",           "       --rope-scaling {none,linear,yarn}",
+                                                                        "RoPE frequency scaling method, defaults to linear unless specified by the model" });
+    options.push_back({ "*",           "       --rope-scale N",         "RoPE context scaling factor, expands context by a factor of N" });
+    options.push_back({ "*",           "       --rope-freq-base N",     "RoPE base frequency, used by NTK-aware scaling (default: loaded from model)" });
+    options.push_back({ "*",           "       --rope-freq-scale N",    "RoPE frequency scaling factor, expands context by a factor of 1/N" });
+    options.push_back({ "*",           "       --yarn-orig-ctx N",      "YaRN: original context size of model (default: %d = model training context size)", params.yarn_orig_ctx });
+    options.push_back({ "*",           "       --yarn-ext-factor N",    "YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor });
+    options.push_back({ "*",           "       --yarn-attn-factor N",   "YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor });
+    options.push_back({ "*",           "       --yarn-beta-slow N",     "YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow });
+    options.push_back({ "*",           "       --yarn-beta-fast N",     "YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast });
+    options.push_back({ "*",           "-gan,  --grp-attn-n N",         "group-attention factor (default: %d)", params.grp_attn_n });
+    options.push_back({ "*",           "-gaw,  --grp-attn-w N",         "group-attention width (default: %.1f)", (double)params.grp_attn_w });
+    options.push_back({ "*",           "-dkvc, --dump-kv-cache",        "verbose print of the KV cache" });
+    options.push_back({ "*",           "-nkvo, --no-kv-offload",        "disable KV offload" });
+    options.push_back({ "*",           "-ctk,  --cache-type-k TYPE",    "KV cache data type for K (default: %s)", params.cache_type_k.c_str() });
+    options.push_back({ "*",           "-ctv,  --cache-type-v TYPE",    "KV cache data type for V (default: %s)", params.cache_type_v.c_str() });
+
+    options.push_back({ "perplexity" });
+    options.push_back({ "perplexity",  "       --all-logits",           "return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false" });
+    options.push_back({ "perplexity",  "       --hellaswag",            "compute HellaSwag score over random tasks from datafile supplied with -f" });
+    options.push_back({ "perplexity",  "       --hellaswag-tasks N",    "number of tasks to use when computing the HellaSwag score (default: %zu)", params.hellaswag_tasks });
+    options.push_back({ "perplexity",  "       --winogrande",           "compute Winogrande score over random tasks from datafile supplied with -f" });
+    options.push_back({ "perplexity",  "       --winogrande-tasks N",   "number of tasks to use when computing the Winogrande score (default: %zu)", params.winogrande_tasks });
+    options.push_back({ "perplexity",  "       --multiple-choice",      "compute multiple choice score over random tasks from datafile supplied with -f" });
+    options.push_back({ "perplexity",  "       --multiple-choice-tasks N",
+                                                                        "number of tasks to use when computing the multiple choice score (default: %zu)", params.multiple_choice_tasks });
+    options.push_back({ "perplexity",  "       --kl-divergence",        "computes KL-divergence to logits provided via --kl-divergence-base" });
+    options.push_back({ "perplexity",  "       --ppl-stride N",         "stride for perplexity calculation (default: %d)", params.ppl_stride });
+    options.push_back({ "perplexity",  "       --ppl-output-type {0,1}",
+                                                                        "output type for perplexity calculation (default: %d)", params.ppl_output_type });
+
+    options.push_back({ "parallel" });
+    options.push_back({ "*",           "-dt,   --defrag-thold N",       "KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold });
+    options.push_back({ "*",           "-np,   --parallel N",           "number of parallel sequences to decode (default: %d)", params.n_parallel });
+    options.push_back({ "*",           "-ns,   --sequences N",          "number of sequences to decode (default: %d)", params.n_sequences });
+    options.push_back({ "*",           "-cb,   --cont-batching",        "enable continuous batching (a.k.a dynamic batching) (default: %s)", params.cont_batching ? "enabled" : "disabled" });
+    options.push_back({ "*",           "-nocb, --no-cont-batching",     "disable continuous batching" });
+
+    options.push_back({ "multi-modality" });
+    options.push_back({ "*",           "       --mmproj FILE",          "path to a multimodal projector file for LLaVA. see examples/llava/README.md" });
+    options.push_back({ "*",           "       --image FILE",           "path to an image file. use with multimodal models. Specify multiple times for batching" });
+
+    options.push_back({ "backend" });
+    options.push_back({ "*",           "       --rpc SERVERS",          "comma separated list of RPC servers" });
+
     if (llama_supports_mlock()) {
-        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        options.push_back({ "*",           "       --mlock",                "force system to keep model in RAM rather than swapping or compressing" });
     }
     if (llama_supports_mmap()) {
-        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
-    }
-    printf("  --numa TYPE           attempt optimizations that help on some NUMA systems\n");
-    printf("                          - distribute: spread execution evenly over all nodes\n");
-    printf("                          - isolate: only spawn threads on CPUs on the node that execution started on\n");
-    printf("                          - numactl: use the CPU map provided by numactl\n");
-    printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
-    printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
+        options.push_back({ "*",           "       --no-mmap",              "do not memory-map model (slower load but may reduce pageouts if not using mlock)" });
+    }
+    options.push_back({ "*",           "       --numa TYPE",            "attempt optimizations that help on some NUMA systems\n"
+                                                                        "  - distribute: spread execution evenly over all nodes\n"
+                                                                        "  - isolate: only spawn threads on CPUs on the node that execution started on\n"
+                                                                        "  - numactl: use the CPU map provided by numactl\n"
+                                                                        "if run without this previously, it is recommended to drop the system page cache before using this\n"
+                                                                        "see https://github.com/ggerganov/llama.cpp/issues/1437" });
+
     if (llama_supports_gpu_offload()) {
-        printf("  -ngl N, --n-gpu-layers N\n");
-        printf("                        number of layers to store in VRAM\n");
-        printf("  -ngld N, --n-gpu-layers-draft N\n");
-        printf("                        number of layers to store in VRAM for the draft model\n");
-        printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
-        printf("                        how to split the model across multiple GPUs, one of:\n");
-        printf("                          - none: use one GPU only\n");
-        printf("                          - layer (default): split layers and KV across GPUs\n");
-        printf("                          - row: split rows across GPUs\n");
-        printf("  -ts SPLIT, --tensor-split SPLIT\n");
-        printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
-        printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
-        printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
-    }
-    printf("  --rpc SERVERS         comma separated list of RPC servers\n");
-    printf("  --verbose-prompt      print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false");
-    printf("  --no-display-prompt   don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false");
-    printf("  -gan N, --grp-attn-n N\n");
-    printf("                        group-attention factor (default: %d)\n", params.grp_attn_n);
-    printf("  -gaw N, --grp-attn-w N\n");
-    printf("                        group-attention width (default: %.1f)\n", (double)params.grp_attn_w);
-    printf("  -dkvc, --dump-kv-cache\n");
-    printf("                        verbose print of the KV cache\n");
-    printf("  -nkvo, --no-kv-offload\n");
-    printf("                        disable KV offload\n");
-    printf("  -ctk TYPE, --cache-type-k TYPE\n");
-    printf("                        KV cache data type for K (default: %s)\n", params.cache_type_k.c_str());
-    printf("  -ctv TYPE, --cache-type-v TYPE\n");
-    printf("                        KV cache data type for V (default: %s)\n", params.cache_type_v.c_str());
-    printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
-    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
-    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    printf("  --control-vector FNAME\n");
-    printf("                        add a control vector\n");
-    printf("  --control-vector-scaled FNAME S\n");
-    printf("                        add a control vector with user defined scaling S\n");
-    printf("  --control-vector-layer-range START END\n");
-    printf("                        layer range to apply the control vector(s) to, start and end inclusive\n");
-    printf("  -m FNAME, --model FNAME\n");
-    printf("                        model path (default: models/$filename with filename from --hf-file or --model-url if set, otherwise %s)\n", DEFAULT_MODEL_PATH);
-    printf("  -md FNAME, --model-draft FNAME\n");
-    printf("                        draft model for speculative decoding (default: unused)\n");
-    printf("  -mu MODEL_URL, --model-url MODEL_URL\n");
-    printf("                        model download url (default: unused)\n");
-    printf("  -hfr REPO, --hf-repo REPO\n");
-    printf("                        Hugging Face model repository (default: unused)\n");
-    printf("  -hff FILE, --hf-file FILE\n");
-    printf("                        Hugging Face model file (default: unused)\n");
-    printf("  -ld LOGDIR, --logdir LOGDIR\n");
-    printf("                        path under which to save YAML logs (no logging if unset)\n");
-    printf("  -lcs FNAME, --lookup-cache-static FNAME\n");
-    printf("                        path to static lookup cache to use for lookup decoding (not updated by generation)\n");
-    printf("  -lcd FNAME, --lookup-cache-dynamic FNAME\n");
-    printf("                        path to dynamic lookup cache to use for lookup decoding (updated by generation)\n");
-    printf("  --override-kv KEY=TYPE:VALUE\n");
-    printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
-    printf("                        types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
-    printf("  -ptc N, --print-token-count N\n");
-    printf("                        print token count every N tokens (default: %d)\n", params.n_print);
-    printf("  --check-tensors       check model tensor data for invalid values\n");
-    printf("\n");
+        options.push_back({ "*",           "-ngl,  --gpu-layers N",
+                                                                        "number of layers to store in VRAM" });
+        options.push_back({ "*",           "-ngld, --gpu-layers-draft N",
+                                                                        "number of layers to store in VRAM for the draft model" });
+        options.push_back({ "*",           "-sm,   --split-mode SPLIT_MODE",
+                                                                        "how to split the model across multiple GPUs, one of:\n"
+                                                                        "  - none: use one GPU only\n"
+                                                                        "  - layer (default): split layers and KV across GPUs\n"
+                                                                        "  - row: split rows across GPUs" });
+        options.push_back({ "*",           "-ts,   --tensor-split SPLIT",
+                                                                        "fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1" });
+        options.push_back({ "*",           "-mg,   --main-gpu i",       "the GPU to use for the model (with split-mode = none),\n"
+                                                                        "or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu });
+    }
+
+    options.push_back({ "model" });
+    options.push_back({ "*",           "       --check-tensors",        "check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false" });
+    options.push_back({ "*",           "       --override-kv KEY=TYPE:VALUE",
+                                                                        "advanced option to override model metadata by key. may be specified multiple times.\n"
+                                                                        "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false" });
+    options.push_back({ "*",           "       --lora FNAME",           "apply LoRA adapter (can be repeated to use multiple adapters)" });
+    options.push_back({ "*",           "       --lora-scaled FNAME S",  "apply LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" });
+    options.push_back({ "*",           "       --control-vector FNAME", "add a control vector\n"
+                                                                        "note: this argument can be repeated to add multiple control vectors" });
+    options.push_back({ "*",           "       --control-vector-scaled FNAME SCALE",
+                                                                        "add a control vector with user defined scaling SCALE\n"
+                                                                        "note: this argument can be repeated to add multiple scaled control vectors" });
+    options.push_back({ "*",           "       --control-vector-layer-range START END",
+                                                                        "layer range to apply the control vector(s) to, start and end inclusive" });
+    options.push_back({ "*",           "-m,    --model FNAME",          "model path (default: models/$filename with filename from --hf-file\n"
+                                                                        "or --model-url if set, otherwise %s)", DEFAULT_MODEL_PATH });
+    options.push_back({ "*",           "-md,   --model-draft FNAME",    "draft model for speculative decoding (default: unused)" });
+    options.push_back({ "*",           "-mu,   --model-url MODEL_URL",  "model download url (default: unused)" });
+    options.push_back({ "*",           "-hfr,  --hf-repo REPO",         "Hugging Face model repository (default: unused)" });
+    options.push_back({ "*",           "-hff,  --hf-file FILE",         "Hugging Face model file (default: unused)" });
+    options.push_back({ "*",           "-hft,  --hf-token TOKEN",       "Hugging Face access token (default: value from HF_TOKEN environment variable)" });
+
+    options.push_back({ "retrieval" });
+    options.push_back({ "retrieval",   "       --context-file FNAME",   "file to load context from (repeat to specify multiple files)" });
+    options.push_back({ "retrieval",   "       --chunk-size N",         "minimum length of embedded text chunks (default: %d)", params.chunk_size });
+    options.push_back({ "retrieval",   "       --chunk-separator STRING",
+                                                                        "separator between chunks (default: '%s')", params.chunk_separator.c_str() });
+
+    options.push_back({ "passkey" });
+    options.push_back({ "passkey",     "       --junk N",               "number of times to repeat the junk text (default: %d)", params.n_junk });
+    options.push_back({ "passkey",     "       --pos N",                "position of the passkey in the junk text (default: %d)", params.i_pos });
+
+    options.push_back({ "imatrix" });
+    options.push_back({ "imatrix",     "-o,    --output FNAME",         "output file (default: '%s')", params.out_file.c_str() });
+    options.push_back({ "imatrix",     "       --output-frequency N",   "output the imatrix every N iterations (default: %d)", params.n_out_freq });
+    options.push_back({ "imatrix",     "       --save-frequency N",     "save an imatrix copy every N iterations (default: %d)", params.n_save_freq });
+    options.push_back({ "imatrix",     "       --process-output",       "collect data for the output tensor (default: %s)", params.process_output ? "true" : "false" });
+    options.push_back({ "imatrix",     "       --no-ppl",               "do not compute perplexity (default: %s)", params.compute_ppl ? "true" : "false" });
+    options.push_back({ "imatrix",     "       --chunk N",              "start processing the input from chunk N (default: %d)", params.i_chunk });
+
+    options.push_back({ "bench" });
+    options.push_back({ "bench",       "-pps",                          "is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false" });
+    options.push_back({ "bench",       "-npp n0,n1,...",                "number of prompt tokens" });
+    options.push_back({ "bench",       "-ntg n0,n1,...",                "number of text generation tokens" });
+    options.push_back({ "bench",       "-npl n0,n1,...",                "number of parallel prompts" });
+
+    options.push_back({ "embedding" });
+    options.push_back({ "embedding",   "       --embd-normalize",       "normalisation for embendings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize });
+    options.push_back({ "embedding",   "       --embd-output-format",   "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix" });
+    options.push_back({ "embedding",   "       --embd-separator",       "separator of embendings (default \\n) for example \"<#sep#>\"" });
+
+    options.push_back({ "server" });
+    options.push_back({ "server",      "       --host HOST",            "ip address to listen (default: %s)", params.hostname.c_str() });
+    options.push_back({ "server",      "       --port PORT",            "port to listen (default: %d)", params.port });
+    options.push_back({ "server",      "       --path PATH",            "path to serve static files from (default: %s)", params.public_path.c_str() });
+    options.push_back({ "server",      "       --embedding(s)",         "restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --api-key KEY",          "API key to use for authentication (default: none)" });
+    options.push_back({ "server",      "       --api-key-file FNAME",   "path to file containing API keys (default: none)" });
+    options.push_back({ "server",      "       --ssl-key-file FNAME",   "path to file a PEM-encoded SSL private key" });
+    options.push_back({ "server",      "       --ssl-cert-file FNAME",  "path to file a PEM-encoded SSL certificate" });
+    options.push_back({ "server",      "       --timeout N",            "server read/write timeout in seconds (default: %d)", params.timeout_read });
+    options.push_back({ "server",      "       --threads-http N",       "number of threads used to process HTTP requests (default: %d)", params.n_threads_http });
+    options.push_back({ "server",      "       --system-prompt-file FNAME",
+                                                                        "set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications" });
+    options.push_back({ "server",      "       --log-format {text,json}",
+                                                                        "log output format: json or text (default: json)" });
+    options.push_back({ "server",      "       --metrics",              "enable prometheus compatible metrics endpoint (default: %s)", params.endpoint_metrics ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --no-slots",             "disables slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled" });
+    options.push_back({ "server",      "       --slot-save-path PATH",  "path to save slot kv cache (default: disabled)" });
+    options.push_back({ "server",      "       --chat-template JINJA_TEMPLATE",
+                                                                        "set custom jinja chat template (default: template taken from model's metadata)\n"
+                                                                        "only commonly used templates are accepted:\n"
+                                                                        "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
+    options.push_back({ "server",      "-sps,  --slot-prompt-similarity SIMILARITY",
+                                                                        "how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity });
+    options.push_back({ "server",      "       --lora-init-without-apply",     "load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"});
+
 #ifndef LOG_DISABLE_LOGS
-    log_print_usage();
+    options.push_back({ "logging" });
+    options.push_back({ "*",           "       --simple-io",            "use basic IO for better compatibility in subprocesses and limited consoles" });
+    options.push_back({ "*",           "-ld,   --logdir LOGDIR",        "path under which to save YAML logs (no logging if unset)" });
+    options.push_back({ "logging",     "       --log-test",             "Run simple logging test" });
+    options.push_back({ "logging",     "       --log-disable",          "Disable trace logs" });
+    options.push_back({ "logging",     "       --log-enable",           "Enable trace logs" });
+    options.push_back({ "logging",     "       --log-file FNAME",       "Specify a log filename (without extension)" });
+    options.push_back({ "logging",     "       --log-new",              "Create a separate new log file on start. "
+                                                                        "Each log file will have unique name: \"..log\"" });
+    options.push_back({ "logging",     "       --log-append",           "Don't truncate the old log file." });
 #endif // LOG_DISABLE_LOGS
+
+    options.push_back({ "cvector" });
+    options.push_back({ "cvector",     "-o,    --output FNAME",         "output file (default: '%s')", params.cvector_outfile.c_str() });
+    options.push_back({ "cvector",     "       --positive-file FNAME",  "positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str() });
+    options.push_back({ "cvector",     "       --negative-file FNAME",  "negative prompts file, one prompt per line (default: '%s')", params.cvector_negative_file.c_str() });
+    options.push_back({ "cvector",     "       --pca-batch N",          "batch size used for PCA. Larger batch runs faster, but uses more memory (default: %d)", params.n_pca_batch });
+    options.push_back({ "cvector",     "       --pca-iter N",           "number of iterations used for PCA (default: %d)", params.n_pca_iterations });
+    options.push_back({ "cvector",     "       --method {pca,mean}",    "dimensionality reduction method to be used (default: pca)" });
+
+    options.push_back({ "export-lora" });
+    options.push_back({ "export-lora", "-m,    --model",                "model path from which to load base model (default '%s')", params.model.c_str() });
+    options.push_back({ "export-lora", "       --lora FNAME",           "path to LoRA adapter  (can be repeated to use multiple adapters)" });
+    options.push_back({ "export-lora", "       --lora-scaled FNAME S",  "path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)" });
+    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during computation (default: %d)", params.n_threads });
+    options.push_back({ "export-lora", "-o,    --output FNAME",         "output file (default: '%s')", params.lora_outfile.c_str() });
+
+    printf("usage: %s [options]\n", argv[0]);
+
+    for (const auto & o : options) {
+        if (!o.grp.empty()) {
+            printf("\n%s:\n\n", o.grp.c_str());
+            continue;
+        }
+        printf("  %-32s", o.args.c_str());
+        if (o.args.length() > 30) {
+            printf("\n%34s", "");
+        }
+
+        const auto desc = o.desc;
+        size_t start = 0;
+        size_t end = desc.find('\n');
+        while (end != std::string::npos) {
+            printf("%s\n%34s", desc.substr(start, end - start).c_str(), "");
+            start = end + 1;
+            end = desc.find('\n', start);
+        }
+
+        printf("%s\n", desc.substr(start).c_str());
+    }
+    printf("\n");
 }
 
 std::string gpt_params_get_system_info(const gpt_params & params) {
@@ -1610,24 +1777,6 @@ std::string string_get_sortable_timestamp() {
     return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
 }
 
-std::string string_random_prompt(std::mt19937 & rng) {
-    const int r = rng() % 10;
-    switch (r) {
-        case 0: return "So";
-        case 1: return "Once upon a time";
-        case 2: return "When";
-        case 3: return "The";
-        case 4: return "After";
-        case 5: return "If";
-        case 6: return "import";
-        case 7: return "He";
-        case 8: return "She";
-        case 9: return "They";
-    }
-
-    GGML_UNREACHABLE();
-}
-
 void string_process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -1887,27 +2036,37 @@ std::string fs_get_cache_directory() {
     return ensure_trailing_slash(cache_directory);
 }
 
+std::string fs_get_cache_file(const std::string & filename) {
+    GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
+    std::string cache_directory = fs_get_cache_directory();
+    const bool success = fs_create_directory_with_parents(cache_directory);
+    if (!success) {
+        throw std::runtime_error("failed to create cache directory: " + cache_directory);
+    }
+    return cache_directory + filename;
+}
+
 
 //
 // Model utils
 //
-
-std::tuple llama_init_from_gpt_params(gpt_params & params) {
+struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
+    llama_init_result iparams;
     auto mparams = llama_model_params_from_gpt_params(params);
 
     llama_model * model = nullptr;
 
     if (!params.hf_repo.empty() && !params.hf_file.empty()) {
-        model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), mparams);
+        model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
     } else if (!params.model_url.empty()) {
-        model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), mparams);
+        model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
     } else {
         model = llama_load_model_from_file(params.model.c_str(), mparams);
     }
 
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     auto cparams = llama_context_params_from_gpt_params(params);
@@ -1916,7 +2075,7 @@ std::tuple llama_init_from_gpt_par
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
-        return std::make_tuple(nullptr, nullptr);
+        return iparams;
     }
 
     if (!params.control_vectors.empty()) {
@@ -1927,7 +2086,7 @@ std::tuple llama_init_from_gpt_par
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
 
         int err = llama_control_vector_apply(lctx,
@@ -1939,26 +2098,26 @@ std::tuple llama_init_from_gpt_par
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
-        }
-    }
-
-    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
-        const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
-        float lora_scale = std::get<1>(params.lora_adapter[i]);
-        int err = llama_model_apply_lora_from_file(model,
-                                             lora_adapter.c_str(),
-                                             lora_scale,
-                                             ((i > 0) || params.lora_base.empty())
-                                                ? NULL
-                                                : params.lora_base.c_str(),
-                                             params.n_threads);
-        if (err != 0) {
-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+            return iparams;
+        }
+    }
+
+    // load and optionally apply lora adapters
+    for (auto & la : params.lora_adapters) {
+        llama_lora_adapter_container loaded_la;
+        loaded_la.path = la.path;
+        loaded_la.scale = la.scale;
+        loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
+        if (loaded_la.adapter == nullptr) {
+            fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
             llama_free_model(model);
-            return std::make_tuple(nullptr, nullptr);
+            return iparams;
         }
+        iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
+    }
+    if (!params.lora_init_without_apply) {
+        llama_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
     if (params.ignore_eos) {
@@ -1968,14 +2127,42 @@ std::tuple llama_init_from_gpt_par
     if (params.warmup) {
         LOG("warming up the model with an empty run\n");
 
-        std::vector tmp = { llama_token_bos(model), llama_token_eos(model), };
+        std::vector tmp;
+        llama_token bos = llama_token_bos(model);
+        llama_token eos = llama_token_eos(model);
+        // some models (e.g. T5) don't have a BOS token
+        if (bos != -1) {
+            tmp.push_back(bos);
+        }
+        tmp.push_back(eos);
+
+        if (llama_model_has_encoder(model)) {
+            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
+            llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+            if (decoder_start_token_id == -1) {
+                decoder_start_token_id = bos;
+            }
+            tmp.clear();
+            tmp.push_back(decoder_start_token_id);
+        }
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
         llama_reset_timings(lctx);
     }
 
-    return std::make_tuple(model, lctx);
+    iparams.model   = model;
+    iparams.context = lctx;
+    return iparams;
+}
+
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector & lora_adapters) {
+    llama_lora_adapter_clear(ctx);
+    for (auto & la : lora_adapters) {
+        if (la.scale != 0.0f) {
+            llama_lora_adapter_set(ctx, la.adapter, la.scale);
+        }
+    }
 }
 
 struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
@@ -2051,6 +2238,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.yarn_beta_slow    = params.yarn_beta_slow;
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
+    cparams.attention_type    = params.attention_type;
     cparams.defrag_thold      = params.defrag_thold;
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
@@ -2070,7 +2258,7 @@ static bool starts_with(const std::string & str, const std::string & prefix) {
     return str.rfind(prefix, 0) == 0;
 }
 
-static bool llama_download_file(const std::string & url, const std::string & path) {
+static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
 
     // Initialize libcurl
     std::unique_ptr curl(curl_easy_init(), &curl_easy_cleanup);
@@ -2085,6 +2273,15 @@ static bool llama_download_file(const std::string & url, const std::string & pat
     curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
     curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
 
+    // Check if hf-token or bearer-token was specified
+    if (!hf_token.empty()) {
+      std::string auth_header = "Authorization: Bearer ";
+      auth_header += hf_token.c_str();
+      struct curl_slist *http_headers = NULL;
+      http_headers = curl_slist_append(http_headers, auth_header.c_str());
+      curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers);
+    }
+
 #if defined(_WIN32)
     // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
     //   operating system. Currently implemented under MS-Windows.
@@ -2201,7 +2398,14 @@ static bool llama_download_file(const std::string & url, const std::string & pat
         }
 
         // Set the output file
-        std::unique_ptr outfile(fopen(path_temporary.c_str(), "wb"), fclose);
+
+        struct FILE_deleter {
+            void operator()(FILE * f) const {
+                fclose(f);
+            }
+        };
+
+        std::unique_ptr outfile(fopen(path_temporary.c_str(), "wb"));
         if (!outfile) {
             fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path.c_str());
             return false;
@@ -2273,6 +2477,7 @@ static bool llama_download_file(const std::string & url, const std::string & pat
 struct llama_model * llama_load_model_from_url(
         const char * model_url,
         const char * path_model,
+        const char * hf_token,
         const struct llama_model_params & params) {
     // Basic validation of the model_url
     if (!model_url || strlen(model_url) == 0) {
@@ -2280,7 +2485,7 @@ struct llama_model * llama_load_model_from_url(
         return NULL;
     }
 
-    if (!llama_download_file(model_url, path_model)) {
+    if (!llama_download_file(model_url, path_model, hf_token)) {
         return NULL;
     }
 
@@ -2328,14 +2533,14 @@ struct llama_model * llama_load_model_from_url(
         // Prepare download in parallel
         std::vector> futures_download;
         for (int idx = 1; idx < n_split; idx++) {
-            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split](int download_idx) -> bool {
+            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
                 char split_path[PATH_MAX] = {0};
                 llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
 
                 char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
                 llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
 
-                return llama_download_file(split_url, split_path);
+                return llama_download_file(split_url, split_path, hf_token);
             }, idx));
         }
 
@@ -2354,6 +2559,7 @@ struct llama_model * llama_load_model_from_hf(
         const char * repo,
         const char * model,
         const char * path_model,
+        const char * hf_token,
         const struct llama_model_params & params) {
     // construct hugging face model url:
     //
@@ -2369,7 +2575,7 @@ struct llama_model * llama_load_model_from_hf(
     model_url += "/resolve/main/";
     model_url += model;
 
-    return llama_load_model_from_url(model_url.c_str(), path_model, params);
+    return llama_load_model_from_url(model_url.c_str(), path_model, hf_token, params);
 }
 
 #else
@@ -2377,6 +2583,7 @@ struct llama_model * llama_load_model_from_hf(
 struct llama_model * llama_load_model_from_url(
         const char * /*model_url*/,
         const char * /*path_model*/,
+        const char * /*hf_token*/,
         const struct llama_model_params & /*params*/) {
     fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
     return nullptr;
@@ -2386,6 +2593,7 @@ struct llama_model * llama_load_model_from_hf(
         const char * /*repo*/,
         const char * /*model*/,
         const char * /*path_model*/,
+        const char * /*hf_token*/,
         const struct llama_model_params & /*params*/) {
     fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
     return nullptr;
@@ -2450,57 +2658,126 @@ std::vector llama_tokenize(
 }
 
 std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
-    std::vector result(8, 0);
-    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
+    std::string piece;
+    piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
+    const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+    if (n_chars < 0) {
+        piece.resize(-n_chars);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+        GGML_ASSERT(check == -n_chars);
+    }
+    else {
+        piece.resize(n_chars);
     }
 
-    return std::string(result.data(), result.size());
+    return piece;
 }
 
-std::string llama_detokenize_spm(llama_context * ctx, const std::vector & tokens) {
-    const llama_token bos_id = llama_token_bos(llama_get_model(ctx));
-
-    std::string piece;
-    std::string result;
+std::string llama_detokenize(llama_context * ctx, const std::vector & tokens, bool special) {
+    std::string text;
+    text.resize(std::max(text.capacity(), tokens.size()));
+    int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    if (n_chars < 0) {
+        text.resize(-n_chars);
+        n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
+    }
 
-    for (size_t i = 0; i < tokens.size(); ++i) {
-        piece = llama_token_to_piece(ctx, tokens[i]);
+    text.resize(n_chars);
 
-        // remove the leading space of the first non-BOS token
-        if (((tokens[0] == bos_id && i == 1) || (tokens[0] != bos_id && i == 0)) && piece[0] == ' ') {
-            piece = piece.substr(1);
-        }
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
+    return text;
+}
 
-        result += piece;
-    }
+bool llama_should_add_bos_token(const llama_model * model) {
+    const int add_bos = llama_add_bos_token(model);
 
-    return result;
+    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
 }
 
-std::string llama_detokenize_bpe(llama_context * ctx, const std::vector & tokens) {
-    std::string piece;
-    std::string result;
+//
+// Chat template utils
+//
+
+bool llama_chat_verify_template(const std::string & tmpl) {
+    llama_chat_message chat[] = {{"user", "test"}};
+    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
+    return res >= 0;
+}
 
-    for (size_t i = 0; i < tokens.size(); ++i) {
-        piece = llama_token_to_piece(ctx, tokens[i]);
+std::string llama_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector & msgs,
+        bool add_ass) {
+    int alloc_size = 0;
+    bool fallback = false; // indicate if we must fallback to default chatml
+    std::vector chat;
+    for (auto & msg : msgs) {
+        chat.push_back({msg.role.c_str(), msg.content.c_str()});
+        alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
+    }
+
+    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
+    std::vector buf(alloc_size);
+
+    // run the first time to get the total output length
+    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+
+    // error: chat template is not supported
+    if (res < 0) {
+        if (ptr_tmpl != nullptr) {
+            // if the custom "tmpl" is not supported, we throw an error
+            // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
+            throw std::runtime_error("this custom template is not supported");
+        } else {
+            // If the built-in template is not supported, we default to chatml
+            res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+            fallback = true;
+        }
+    }
 
-        result += piece;
+    // if it turns out that our buffer is too small, we resize it
+    if ((size_t) res > buf.size()) {
+        buf.resize(res);
+        res = llama_chat_apply_template(
+            fallback ? nullptr : model,
+            fallback ? "chatml" : ptr_tmpl,
+            chat.data(), chat.size(), add_ass, buf.data(), buf.size());
     }
 
-    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
-    return result;
+    std::string formatted_chat(buf.data(), res);
+    return formatted_chat;
 }
 
-bool llama_should_add_bos_token(const llama_model * model) {
-    const int add_bos = llama_add_bos_token(model);
+std::string llama_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector & past_msg,
+        const llama_chat_msg & new_msg,
+        bool add_ass) {
+    std::ostringstream ss;
+    auto fmt_past_msg = past_msg.empty() ? "" : llama_chat_apply_template(model, tmpl, past_msg, false);
+    std::vector chat_new(past_msg);
+    // if the past_msg ends with a newline, we must preserve it in the formatted version
+    if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
+        ss << "\n";
+    };
+    // format chat with new_msg
+    chat_new.push_back(new_msg);
+    auto fmt_new_msg = llama_chat_apply_template(model, tmpl, chat_new, add_ass);
+    // get the diff part
+    ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
+    return ss.str();
+}
 
-    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
+std::string llama_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl) {
+    std::vector msgs = {
+        {"system",    "You are a helpful assistant"},
+        {"user",      "Hello"},
+        {"assistant", "Hi there"},
+        {"user",      "How are you?"},
+    };
+    return llama_chat_apply_template(model, tmpl, msgs, true);
 }
 
 //
@@ -2582,14 +2859,34 @@ void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_siz
 // Embedding utils
 //
 
-void llama_embd_normalize(const float * inp, float * out, int n) {
+void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm) {
     double sum = 0.0;
-    for (int i = 0; i < n; i++) {
-        sum += inp[i] * inp[i];
+
+    switch (embd_norm) {
+        case -1: // no normalisation
+            sum = 1.0;
+            break;
+        case 0: // max absolute
+            for (int i = 0; i < n; i++) {
+                if (sum < std::abs(inp[i])) sum = std::abs(inp[i]);
+            }
+            sum /= 32760.0; // make an int16 range
+            break;
+        case 2: // euclidean
+            for (int i = 0; i < n; i++) {
+                sum += inp[i] * inp[i];
+            }
+            sum = std::sqrt(sum);
+            break;
+        default: // p-norm (euclidean is p-norm p=2)
+            for (int i = 0; i < n; i++) {
+                sum += std::pow(std::abs(inp[i]), embd_norm);
+            }
+            sum = std::pow(sum, 1.0 / embd_norm);
+            break;
     }
-    sum = sqrt(sum);
 
-    const float norm = sum > 0.0 ? 1.0f / sum : 0.0f;
+    const float norm = sum > 0.0 ? 1.0 / sum : 0.0f;
 
     for (int i = 0; i < n; i++) {
         out[i] = inp[i] * norm;
@@ -2607,6 +2904,14 @@ float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n)
         sum2 += embd2[i] * embd2[i];
     }
 
+    // Handle the case where one or both vectors are zero vectors
+    if (sum1 == 0.0 || sum2 == 0.0) {
+        if (sum1 == 0.0 && sum2 == 0.0) {
+            return 1.0f; // two zero vectors are similar
+        }
+        return 0.0f;
+    }
+
     return sum / (sqrt(sum1) * sqrt(sum2));
 }
 
@@ -2615,125 +2920,87 @@ float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n)
 //
 
 static llama_control_vector_data llama_control_vector_load_one(const llama_control_vector_load_info & load_info) {
-    int32_t n_tensors;
-
-    size_t n_bytes = 0;
-
-    uint32_t max_direction_layer = 0;
-
     llama_control_vector_data result = { -1, {} };
 
-    // calculate size of ctx needed for tensors, ensure tensors are f32, and find max layer
-    {
-        struct ggml_init_params meta_params = {
-            /* .mem_size   = */ ggml_tensor_overhead() * 128 + ggml_graph_overhead(),
-            /* .mem_buffer = */ nullptr,
-            /* .no_alloc   = */ true,
-        };
-        ggml_context * meta_ctx = ggml_init(meta_params);
-        struct gguf_init_params meta_gguf_params = {
-            /* .no_alloc = */ true,
-            /* .ctx      = */ &meta_ctx,
-        };
-        struct gguf_context * meta_ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params);
-        if (!meta_ctx_gguf) {
-            fprintf(stderr, "%s: failed to load control vector from %s\n", __func__, load_info.fname.c_str());
-            ggml_free(meta_ctx);
-            return result;
-        }
-
-        n_tensors = gguf_get_n_tensors(meta_ctx_gguf);
-        for (int i = 0; i < n_tensors; i++) {
-            std::string name = gguf_get_tensor_name(meta_ctx_gguf, i);
-
-            // split on '.'
-            size_t dotpos = name.find('.');
-            if (dotpos != std::string::npos && name.substr(0, dotpos) == "direction") {
-                try {
-                    uint32_t layer = std::stoi(name.substr(dotpos + 1));
-                    if (layer == 0) {
-                        fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str());
-                        ggml_free(meta_ctx);
-                        gguf_free(meta_ctx_gguf);
-                        return result;
-                    }
-                    if (layer > max_direction_layer) {
-                        max_direction_layer = layer;
-                    }
-                } catch (...) {
-                    fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str());
-                    ggml_free(meta_ctx);
-                    gguf_free(meta_ctx_gguf);
-                    return result;
-                }
-            }
-
-            struct ggml_tensor * tensor_meta = ggml_get_tensor(meta_ctx, name.c_str());
-            if (tensor_meta->type != GGML_TYPE_F32 || ggml_n_dims(tensor_meta) != 1) {
-                fprintf(stderr, "%s: direction tensor invalid in %s\n", __func__, load_info.fname.c_str());
-                ggml_free(meta_ctx);
-                gguf_free(meta_ctx_gguf);
-                return result;
-            }
-            if (result.n_embd == -1) {
-                result.n_embd = ggml_nelements(tensor_meta);
-            } else if (ggml_nelements(tensor_meta) != result.n_embd) {
-                fprintf(stderr, "%s: direction tensor sizes mismatched in %s\n", __func__, load_info.fname.c_str());
-                ggml_free(meta_ctx);
-                gguf_free(meta_ctx_gguf);
-                return result;
-            }
-            n_bytes += ggml_nbytes(tensor_meta);
-        }
-        ggml_free(meta_ctx);
-        gguf_free(meta_ctx_gguf);
+    ggml_context * ctx = nullptr;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ false,
+        /* .ctx      = */ &ctx,
+    };
+    struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), meta_gguf_params);
+    if (!ctx_gguf) {
+        fprintf(stderr, "%s: failed to load control vector file from %s\n", __func__, load_info.fname.c_str());
+        return result;
     }
 
+    int32_t n_tensors = gguf_get_n_tensors(ctx_gguf);
     if (n_tensors == 0) {
         fprintf(stderr, "%s: no direction tensors found in %s\n", __func__, load_info.fname.c_str());
-        return result;
     }
 
-    // load and scale tensors into final control vector context
-    struct ggml_init_params ggml_params = {
-        /* .mem_size   = */ ggml_tensor_overhead() * n_tensors + n_bytes,
-        /* .mem_buffer = */ nullptr,
-        /* .no_alloc   = */ false,
-    };
-    struct ggml_context * ctx = ggml_init(ggml_params);
+    for (int i = 0; i < n_tensors; i++) {
+        std::string name = gguf_get_tensor_name(ctx_gguf, i);
 
-    struct gguf_init_params params = {
-        /*.no_alloc = */ false,
-        /*.ctx      = */ &ctx,
-    };
-    struct gguf_context * ctx_gguf = gguf_init_from_file(load_info.fname.c_str(), params);
-    if (!ctx_gguf) {
-        fprintf(stderr, "%s: failed to load control vector from %s\n", __func__, load_info.fname.c_str());
-        ggml_free(ctx);
-        return result;
-    }
+        int layer_idx = -1;
+
+        // split on '.'
+        size_t dotpos = name.find('.');
+        if (dotpos != std::string::npos && name.substr(0, dotpos) == "direction") {
+            try {
+                layer_idx = std::stoi(name.substr(dotpos + 1));
+            } catch (...) {
+                layer_idx = -1;
+            }
+        }
+        if (layer_idx < 0) {
+            fprintf(stderr, "%s: invalid/unparsable direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        } else if (layer_idx == 0) {
+            fprintf(stderr, "%s: invalid (zero) direction tensor layer index in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
 
-    // do not store data for layer 0 (it's not used)
-    result.data.resize(result.n_embd * max_direction_layer);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str());
+        if (tensor->type != GGML_TYPE_F32) {
+            fprintf(stderr, "%s: invalid (non-F32) direction tensor type in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
+        if (ggml_n_dims(tensor) != 1) {
+            fprintf(stderr, "%s: invalid (non-1D) direction tensor shape in %s\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
 
-    for (uint32_t il = 1; il <= max_direction_layer; il++) {
-        const std::string name = "direction." + std::to_string(il);
-        const ggml_tensor * tensor = ggml_get_tensor(ctx, name.c_str());
+        if (result.n_embd == -1) {
+            result.n_embd = ggml_nelements(tensor);
+        } else if (ggml_nelements(tensor) != result.n_embd) {
+            fprintf(stderr, "%s: direction tensor in %s does not match previous dimensions\n", __func__, load_info.fname.c_str());
+            result.n_embd = -1;
+            break;
+        }
 
-        float * dst = result.data.data() + result.n_embd * (il - 1);
+        // extend if necessary - do not store data for layer 0 (it's not used)
+        result.data.resize(std::max(result.data.size(), static_cast(result.n_embd * layer_idx)), 0.0f);
 
-        if (tensor) {
-            const float * src = (const float *) tensor->data;
-            for (int j = 0; j < result.n_embd; j++) {
-                dst[j] = src[j] * load_info.strength;
-            }
-        } else {
-            for (int j = 0; j < result.n_embd; j++) {
-                dst[j] = 0.0f;
-            }
+        const float * src = (const float *) tensor->data;
+        float * dst = result.data.data() + result.n_embd * (layer_idx - 1);  // layer 1 at [0]
+        for (int j = 0; j < result.n_embd; j++) {
+            dst[j] += src[j] * load_info.strength;  // allows multiple directions for same layer in same file
         }
+
+    }
+
+    if (result.n_embd == -1) {
+        fprintf(stderr, "%s: skipping %s due to invalid direction tensors\n", __func__, load_info.fname.c_str());
+        result.data.clear();
     }
 
+    gguf_free(ctx_gguf);
+    ggml_free(ctx);
+
     return result;
 }
 
@@ -2744,16 +3011,19 @@ llama_control_vector_data llama_control_vector_load(const std::vector la : params.lora_adapter) {
-        if (std::get<1>(la) != 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale == 1.0f) {
+            fprintf(stream, "  - %s\n", la.path.c_str());
         }
-        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
     }
     fprintf(stream, "lora_scaled:\n");
-    for (std::tuple la : params.lora_adapter) {
-        if (std::get<1>(la) == 1.0f) {
-            continue;
+    for (auto & la : params.lora_adapters) {
+        if (la.scale != 1.0f) {
+            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
         }
-        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
     }
-    fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
+    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
@@ -2955,7 +3221,6 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
     fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
     yaml_dump_vector_int(stream, "prompt_tokens", prompt_tokens);
-    fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
     fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
 
     fprintf(stream, "reverse_prompt:\n");
diff --git a/llm/llama.cpp/common/common.h b/llm/llama.cpp/common/common.h
index 264504830a7f08a4952fd4ef1e6725bf86b21236..d88966ece20aa73da262d047e6c0137e7a0e972f 100644
--- a/llm/llama.cpp/common/common.h
+++ b/llm/llama.cpp/common/common.h
@@ -33,6 +33,15 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct llama_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct llama_lora_adapter_container : llama_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -52,109 +61,114 @@ int32_t cpu_get_num_math();
 // CLI argument parsing
 //
 
+// dimensionality reduction methods, used by cvector-generator
+enum dimre_method {
+    DIMRE_METHOD_PCA,
+    DIMRE_METHOD_MEAN,
+};
+
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
     int32_t n_threads             = cpu_get_num_math();
-    int32_t n_threads_draft       = -1;
-    int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
-    int32_t n_threads_batch_draft = -1;
-    int32_t n_predict             = -1;    // new tokens to predict
-    int32_t n_ctx                 = 512;   // context size
-    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
-    int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
-    int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
-    int32_t n_parallel            = 1;     // number of parallel sequences to decode
-    int32_t n_sequences           = 1;     // number of sequences to decode
-    float   p_split               = 0.1f;  // speculative decoding split probability
-    int32_t n_gpu_layers          = -1;    // number of layers to store in VRAM (-1 - use default)
-    int32_t n_gpu_layers_draft    = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
-    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
-    int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
-    int32_t grp_attn_n            = 1;     // group-attention factor
-    int32_t grp_attn_w            = 512;   // group-attention width
-    int32_t n_print               = -1;    // print token count every n tokens (-1 = disabled)
-    float   rope_freq_base        = 0.0f;  // RoPE base frequency
-    float   rope_freq_scale       = 0.0f;  // RoPE frequency scaling factor
+    int32_t n_threads_draft       =    -1;
+    int32_t n_threads_batch       =    -1; // number of threads to use for batch processing (-1 = use n_threads)
+    int32_t n_threads_batch_draft =    -1;
+    int32_t n_predict             =    -1; // new tokens to predict
+    int32_t n_ctx                 =     0; // context size
+    int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                =     0; // number of tokens to keep from initial prompt
+    int32_t n_draft               =     5; // number of tokens to draft during speculative decoding
+    int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel            =     1; // number of parallel sequences to decode
+    int32_t n_sequences           =     1; // number of sequences to decode
+    float   p_split               =  0.1f; // speculative decoding split probability
+    int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
+    int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
+    int32_t grp_attn_n            =     1; // group-attention factor
+    int32_t grp_attn_w            =   512; // group-attention width
+    int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
+    float   rope_freq_base        =  0.0f; // RoPE base frequency
+    float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
     float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor      = 1.0f;  // YaRN magnitude scaling factor
+    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
-    int32_t yarn_orig_ctx         = 0;     // YaRN original context length
+    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
+    int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
-    std::string rpc_servers       = "";    // comma separated list of RPC servers
 
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
 
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
+    enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
+    enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
     // // sampling parameters
     struct llama_sampling_params sparams;
 
-    std::string model                = "";  // model path
-    std::string model_draft          = "";  // draft model for speculative decoding
+    std::string model                = ""; // model path
+    std::string model_draft          = ""; // draft model for speculative decoding
     std::string model_alias          = "unknown"; // model alias
-    std::string model_url            = "";  // model url to download
-    std::string hf_repo              = "";  // HF repo
-    std::string hf_file              = "";  // HF file
+    std::string model_url            = ""; // model url to download
+    std::string hf_token             = ""; // HF token
+    std::string hf_repo              = ""; // HF repo
+    std::string hf_file              = ""; // HF file
     std::string prompt               = "";
-    std::string prompt_file          = "";  // store the external prompt file name
-    std::string path_prompt_cache    = "";  // path to file for saving/loading prompt eval state
-    std::string input_prefix         = "";  // string to prefix user inputs with
-    std::string input_suffix         = "";  // string to suffix user inputs with
-    std::vector antiprompt; // string upon seeing which more user input is prompted
-    std::string logdir               = "";  // directory in which to save YAML log files
+    std::string prompt_file          = ""; // store the external prompt file name
+    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state
+    std::string input_prefix         = ""; // string to prefix user inputs with
+    std::string input_suffix         = ""; // string to suffix user inputs with
+    std::string logdir               = ""; // directory in which to save YAML log files
     std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
-    std::string logits_file          = "";  // file for saving *all* logits
+    std::string logits_file          = ""; // file for saving *all* logits
+    std::string rpc_servers          = ""; // comma separated list of RPC servers
 
+    std::vector in_files;   // all input files
+    std::vector antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector kv_overrides;
 
-    // TODO: avoid tuple, use struct
-    std::vector> lora_adapter; // lora adapter path with user defined scale
-    std::string lora_base  = "";                              // base model path for the lora adapter
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector lora_adapters; // lora adapter path with user defined scale
 
     std::vector control_vectors; // control vector with user defined scale
 
+    int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
 
-    int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
-    int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
-                                    //                                       (which is more convenient to use for plotting)
-                                    //
-    bool   hellaswag       = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
-    size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
+    int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
+    int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
+                                     //                                       (which is more convenient to use for plotting)
+                                     //
+    bool   hellaswag        = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    size_t hellaswag_tasks  = 400;   // number of tasks to use when computing the HellaSwag score
 
-    bool   winogrande      = false; // compute Winogrande score over random tasks from datafile supplied in prompt
-    size_t winogrande_tasks= 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
+    bool   winogrande       = false; // compute Winogrande score over random tasks from datafile supplied in prompt
+    size_t winogrande_tasks = 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
 
-    bool   multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
-    size_t multiple_choice_tasks = 0;     // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
+    bool   multiple_choice  = false;  // compute TruthfulQA score over random tasks from datafile supplied in prompt
+    size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
 
-    bool   kl_divergence   = false; // compute KL divergence
+    bool   kl_divergence    = false; // compute KL divergence
 
-    bool random_prompt     = false; // do not randomize prompt if none provided
+    bool usage             = false; // print usage
     bool use_color         = false; // use color to distinguish generations and inputs
-    bool interactive       = false; // interactive mode
-    bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode
     bool special           = false; // enable special token output
+    bool interactive       = false; // interactive mode
+    bool interactive_first = false; // wait for user input immediately
     bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
-    bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
     bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
-    bool embedding         = false; // get only sentence embedding
-    bool escape            = false; // escape "\n", "\r", "\t", "\'", "\"", and "\\"
-    bool interactive_first = false; // wait for user input immediately
+    bool escape            = true;  // escape "\n", "\r", "\t", "\'", "\"", and "\\"
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
@@ -162,7 +176,6 @@ struct gpt_params {
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
-    bool instruct          = false; // instruction mode (used for Alpaca models)
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
@@ -180,8 +193,81 @@ struct gpt_params {
     // multimodal models (see examples/llava)
     std::string mmproj = "";        // path to multimodal projector
     std::vector image; // path to image file(s)
+
+    // embedding
+    bool embedding         = false; // get only sentence embedding
+    int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
+    std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
+    std::string embd_sep   = "\n";  // separator of embendings
+
+    // server params
+    int32_t port           = 8080;         // server listens on this network port
+    int32_t timeout_read   = 600;          // http read timeout in seconds
+    int32_t timeout_write  = timeout_read; // http write timeout in seconds
+    int32_t n_threads_http = -1;           // number of threads to process HTTP requests
+
+    std::string hostname      = "127.0.0.1";
+    std::string public_path   = "";
+    std::string chat_template = "";
+    std::string system_prompt = "";
+    bool enable_chat_template = true;
+
+    std::vector api_keys;
+
+    std::string ssl_file_key  = "";
+    std::string ssl_file_cert = "";
+
+    bool endpoint_slots   = true;
+    bool endpoint_metrics = false;
+
+    bool log_json = false;
+
+    std::string slot_save_path;
+
+    float slot_prompt_similarity = 0.5f;
+
+    // batched-bench params
+    bool is_pp_shared = false;
+
+    std::vector n_pp;
+    std::vector n_tg;
+    std::vector n_pl;
+
+    // retrieval params
+    std::vector context_files; // context files to embed
+
+    int32_t chunk_size = 64; // chunk size for context embedding
+
+    std::string chunk_separator = "\n"; // chunk separator for context embedding
+
+    // passkey params
+    int32_t n_junk = 250; // number of times to repeat the junk text
+    int32_t i_pos  = -1;  // position of the passkey in the junk text
+
+    // imatrix params
+    std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
+
+    int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
+    int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
+    int32_t i_chunk     =  0; // start processing from this chunk
+
+    bool process_output = false; // collect data for the output tensor
+    bool compute_ppl    = true;  // whether to compute perplexity
+
+    // cvector-generator params
+    int n_pca_batch = 100;
+    int n_pca_iterations = 1000;
+    dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
+    std::string cvector_outfile       = "control_vector.gguf";
+    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
+
+    bool spm_infill = false; // suffix/prefix/middle pattern for infill
+
+    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
 };
 
+void gpt_params_handle_hf_token(gpt_params & params);
 void gpt_params_handle_model_default(gpt_params & params);
 
 bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
@@ -199,7 +285,20 @@ std::vector string_split(std::string input, char separator);
 
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
-std::string string_random_prompt(std::mt19937 & rng);
+
+template
+static std::vector string_split(const std::string & str, char delim) {
+    std::vector values;
+    std::istringstream str_stream(str);
+    std::string token;
+    while (std::getline(str_stream, token, delim)) {
+        T value;
+        std::istringstream token_stream(token);
+        token_stream >> value;
+        values.push_back(value);
+    }
+    return values;
+}
 
 bool string_parse_kv_override(const char * data, std::vector & overrides);
 void string_process_escapes(std::string & input);
@@ -212,19 +311,28 @@ bool fs_validate_filename(const std::string & filename);
 bool fs_create_directory_with_parents(const std::string & path);
 
 std::string fs_get_cache_directory();
+std::string fs_get_cache_file(const std::string & filename);
 
 //
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
-std::tuple llama_init_from_gpt_params(gpt_params & params);
+struct llama_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector lora_adapters;
+};
+
+struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params);
-struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+
+// clear LoRA adapters from context, then apply new list of adapters
+void llama_lora_adapters_apply(struct llama_context * ctx, std::vector & lora_adapters);
 
 // Batch utils
 
@@ -262,26 +370,50 @@ std::string llama_token_to_piece(
                        llama_token   token,
                        bool          special = true);
 
-// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
-//       that takes into account the tokenizer type and decides how to handle the leading space
-//
-// detokenizes a vector of tokens into a string
-// should work similar to Python's `tokenizer.decode`
-// removes the leading space from the first non-BOS token
-std::string llama_detokenize_spm(
-                         llama_context * ctx,
-        const std::vector & tokens);
-
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
-std::string llama_detokenize_bpe(
+// optionally renders special/control tokens
+std::string llama_detokenize(
                          llama_context * ctx,
-        const std::vector & tokens);
+        const std::vector & tokens,
+                                  bool   special = true);
 
 // Uses the value from the model metadata if possible, otherwise
 // defaults to true when model type is SPM, otherwise false.
 bool llama_should_add_bos_token(const llama_model * model);
 
+//
+// Chat template utils
+//
+
+// same with llama_chat_message, but uses std::string
+struct llama_chat_msg {
+    std::string role;
+    std::string content;
+};
+
+// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
+bool llama_chat_verify_template(const std::string & tmpl);
+
+// CPP wrapper for llama_chat_apply_template
+// If the built-in template is not supported, we default to chatml
+// If the custom "tmpl" is not supported, we throw an error
+std::string llama_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector & chat,
+        bool add_ass);
+
+// Format single message, while taking into account the position of that message in chat history
+std::string llama_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector & past_msg,
+        const llama_chat_msg & new_msg,
+        bool add_ass);
+
+// Returns an example of formatted chat
+std::string llama_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl);
+
 //
 // KV cache utils
 //
@@ -296,7 +428,7 @@ void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_siz
 // Embedding utils
 //
 
-void llama_embd_normalize(const float * inp, float * out, int n);
+void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
 
 float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
 
@@ -340,4 +472,3 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha
 void yaml_dump_non_result_info(
     FILE * stream, const gpt_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector & prompt_tokens, const char * model_desc);
-
diff --git a/llm/llama.cpp/common/grammar-parser.cpp b/llm/llama.cpp/common/grammar-parser.cpp
index b5bc7d49b5f3695705ed703b55872e392b80506f..a518b766dc33e5a3bf5921cacbe62184363fffab 100644
--- a/llm/llama.cpp/common/grammar-parser.cpp
+++ b/llm/llama.cpp/common/grammar-parser.cpp
@@ -46,8 +46,12 @@ namespace grammar_parser {
         state.rules[rule_id] = rule;
     }
 
+    static bool is_digit_char(char c) {
+        return '0' <= c && c <= '9';
+    }
+
     static bool is_word_char(char c) {
-        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c);
     }
 
     static std::pair parse_hex(const char * src, int size) {
@@ -99,6 +103,17 @@ namespace grammar_parser {
         return pos;
     }
 
+    static const char * parse_int(const char * src) {
+        const char * pos = src;
+        while (is_digit_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting integer at ") + src);
+        }
+        return pos;
+    }
+
     static std::pair parse_char(const char * src) {
         if (*src == '\\') {
             switch (src[1]) {
@@ -137,6 +152,60 @@ namespace grammar_parser {
             bool                                 is_nested) {
         size_t last_sym_start = out_elements.size();
         const char * pos = src;
+
+        auto handle_repetitions = [&](int min_times, int max_times) {
+
+            if (last_sym_start == out_elements.size()) {
+                throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
+            }
+
+            // apply transformation to previous symbol (last_sym_start to end) according to
+            // the following rewrite rules:
+            // S{m,n} --> S S S (m times) S'(n-m)
+            //            S'(x)   ::= S S'(x-1) |
+            //            (... n-m definitions of these S' rules ...)
+            //            S'(1)   ::= S |
+            // S{m,} -->  S S S (m times) S'
+            //            S'     ::= S S' |
+            // S*     --> S{0,}
+            //        --> S'     ::= S S' |
+            // S+     --> S{1,}
+            //        --> S S'
+            //            S'     ::= S S' |
+            // S?     --> S{0,1}
+            //        --> S'
+            //            S'     ::= S |
+
+            std::vector previous_elements(out_elements.begin() + last_sym_start, out_elements.end());
+            if (min_times == 0) {
+                out_elements.resize(last_sym_start);
+            } else {
+                // Repeat the previous elements (min_times - 1) times
+                for (int i = 1; i < min_times; i++) {
+                    out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end());
+                }
+            }
+
+            uint32_t last_rec_rule_id = 0;
+            auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+
+            std::vector rec_rule(previous_elements);
+            for (int i = 0; i < n_opt; i++) {
+                rec_rule.resize(previous_elements.size());
+                uint32_t rec_rule_id = generate_symbol_id(state, rule_name);
+                if (i > 0 || max_times < 0) {
+                    rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+                }
+                rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+                rec_rule.push_back({LLAMA_GRETYPE_END, 0});
+                add_rule(state, rec_rule_id, rec_rule);
+                last_rec_rule_id = rec_rule_id;
+            }
+            if (n_opt > 0) {
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
+            }
+        };
+
         while (*pos) {
             if (*pos == '"') { // literal string
                 pos++;
@@ -197,40 +266,51 @@ namespace grammar_parser {
                     throw std::runtime_error(std::string("expecting ')' at ") + pos);
                 }
                 pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
-                if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
-                }
+            } else if (*pos == '.') { // any char
+                last_sym_start = out_elements.size();
+                out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*') {
+                pos = parse_space(pos + 1, is_nested);
+                handle_repetitions(0, -1);
+            } else if (*pos == '+') {
+                pos = parse_space(pos + 1, is_nested);
+                handle_repetitions(1, -1);
+            } else if (*pos == '?') {
+                pos = parse_space(pos + 1, is_nested);
+                handle_repetitions(0, 1);
+            } else if (*pos == '{') {
+                pos = parse_space(pos + 1, is_nested);
 
-                // apply transformation to previous symbol (last_sym_start to end) according to
-                // rewrite rules:
-                // S* --> S' ::= S S' |
-                // S+ --> S' ::= S S' | S
-                // S? --> S' ::= S |
-                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
-                std::vector sub_rule;
-                // add preceding symbol to generated rule
-                sub_rule.insert(
-                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
-                if (*pos == '*' || *pos == '+') {
-                    // cause generated rule to recurse
-                    sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                if (!is_digit_char(*pos)) {
+                    throw std::runtime_error(std::string("expecting an int at ") + pos);
                 }
-                // mark start of alternate def
-                sub_rule.push_back({LLAMA_GRETYPE_ALT, 0});
-                if (*pos == '+') {
-                    // add preceding symbol as alternate only for '+' (otherwise empty)
-                    sub_rule.insert(
-                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
-                }
-                sub_rule.push_back({LLAMA_GRETYPE_END, 0});
-                add_rule(state, sub_rule_id, sub_rule);
+                const char * int_end = parse_int(pos);
+                int min_times = std::stoul(std::string(pos, int_end - pos));
+                pos = parse_space(int_end, is_nested);
 
-                // in original rule, replace previous symbol with reference to generated rule
-                out_elements.resize(last_sym_start);
-                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                int max_times = -1;
 
-                pos = parse_space(pos + 1, is_nested);
+                if (*pos == '}') {
+                    max_times = min_times;
+                    pos = parse_space(pos + 1, is_nested);
+                } else if (*pos == ',') {
+                    pos = parse_space(pos + 1, is_nested);
+
+                    if (is_digit_char(*pos)) {
+                        const char * int_end = parse_int(pos);
+                        max_times = std::stoul(std::string(pos, int_end - pos));
+                        pos = parse_space(int_end, is_nested);
+                    }
+
+                    if (*pos != '}') {
+                        throw std::runtime_error(std::string("expecting '}' at ") + pos);
+                    }
+                    pos = parse_space(pos + 1, is_nested);
+                } else {
+                    throw std::runtime_error(std::string("expecting ',' at ") + pos);
+                }
+                handle_repetitions(min_times, max_times);
             } else {
                 break;
             }
@@ -325,6 +405,7 @@ namespace grammar_parser {
             case LLAMA_GRETYPE_CHAR_NOT:       return true;
             case LLAMA_GRETYPE_CHAR_ALT:       return true;
             case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
+            case LLAMA_GRETYPE_CHAR_ANY:       return true;
             default:                           return false;
         }
     }
@@ -339,6 +420,7 @@ namespace grammar_parser {
                 case LLAMA_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
                 case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
                 case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+                case LLAMA_GRETYPE_CHAR_ANY:       fprintf(file, "CHAR_ANY");       break;
             }
             switch (elem.type) {
                 case LLAMA_GRETYPE_END:
@@ -350,6 +432,7 @@ namespace grammar_parser {
                 case LLAMA_GRETYPE_CHAR_NOT:
                 case LLAMA_GRETYPE_CHAR_RNG_UPPER:
                 case LLAMA_GRETYPE_CHAR_ALT:
+                case LLAMA_GRETYPE_CHAR_ANY:
                     fprintf(file, "(\"");
                     print_grammar_char(file, elem.value);
                     fprintf(file, "\") ");
@@ -407,11 +490,15 @@ namespace grammar_parser {
                     }
                     print_grammar_char(file, elem.value);
                     break;
+                case LLAMA_GRETYPE_CHAR_ANY:
+                    fprintf(file, ".");
+                    break;
             }
             if (is_char_element(elem)) {
                 switch (rule[i + 1].type) {
                     case LLAMA_GRETYPE_CHAR_ALT:
                     case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                    case LLAMA_GRETYPE_CHAR_ANY:
                         break;
                     default:
                         fprintf(file, "] ");
diff --git a/llm/llama.cpp/common/json-schema-to-grammar.cpp b/llm/llama.cpp/common/json-schema-to-grammar.cpp
index 9a71f5d8d76ba756d5c38d40d204a9d1cf79275d..881eb49e3389e1bddccb0766da6c00f7575a1a85 100644
--- a/llm/llama.cpp/common/json-schema-to-grammar.cpp
+++ b/llm/llama.cpp/common/json-schema-to-grammar.cpp
@@ -16,92 +16,282 @@ static std::string join(Iterator begin, Iterator end, const std::string & separa
 
 static std::string repeat(const std::string & str, size_t n);
 
-static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "", bool item_rule_is_literal = false) {
+static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
+    auto has_max = max_items != std::numeric_limits::max();
+
+    if (min_items == 0 && max_items == 1) {
+        return item_rule + "?";
+    }
+
     if (separator_rule.empty()) {
-        if (min_items == 0 && max_items == 1) {
-            return item_rule + "?";
-        } else if (min_items == 1 && max_items == std::numeric_limits::max()) {
+        if (min_items == 1 && !has_max) {
             return item_rule + "+";
+        } else if (min_items == 0 && !has_max) {
+            return item_rule + "*";
+        } else {
+            return item_rule + "{" + std::to_string(min_items) + "," + (has_max ? std::to_string(max_items) : "") + "}";
         }
     }
 
-    std::string result;
-    if (min_items > 0) {
-        if (item_rule_is_literal && separator_rule.empty()) {
-            result = "\"" + repeat(std::string(item_rule.begin() + 1, item_rule.end() - 1), min_items) + "\"";
-        } else {
-            std::vector items(min_items, item_rule);
-            result = join(items.begin(), items.end(), separator_rule.empty() ? " " : " " + separator_rule + " ");
+    auto result = item_rule + " " + build_repetition("(" + separator_rule + " " + item_rule + ")", min_items == 0 ? 0 : min_items - 1, has_max ? max_items - 1 : max_items);
+    if (min_items == 0) {
+        result = "(" + result + ")?";
+    }
+    return result;
+}
+
+/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
+class string_view {
+    const std::string & _str;
+    const size_t _start;
+    const size_t _end;
+public:
+    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
+
+    size_t size() const {
+        return _end - _start;
+    }
+
+    size_t length() const {
+        return size();
+    }
+
+    operator std::string() const {
+        return str();
+    }
+
+    std::string str() const {
+        return _str.substr(_start, _end - _start);
+    }
+
+    string_view substr(size_t pos, size_t len = std::string::npos) const {
+        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
+    }
+
+    char operator[](size_t pos) const {
+        auto index = _start + pos;
+        if (index >= _end) {
+            throw std::out_of_range("string_view index out of range");
         }
+        return _str[_start + pos];
     }
 
-    std::function opt_repetitions = [&](int up_to_n, bool prefix_with_sep) -> std::string {
-        auto content = prefix_with_sep && !separator_rule.empty() ? separator_rule + " " + item_rule : item_rule;
+    bool operator==(const string_view & other) const {
+        std::string this_str = *this;
+        std::string other_str = other;
+        return this_str == other_str;
+    }
+};
 
-        if (up_to_n == 0) {
-            return "";
-        } else if (up_to_n == 1) {
-            return "(" + content + ")?";
-        } else if (!separator_rule.empty() && !prefix_with_sep) {
-            return "(" + content + " " + opt_repetitions(up_to_n - 1, true) + ")?";
+static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
+    auto has_min = min_value != std::numeric_limits::min();
+    auto has_max = max_value != std::numeric_limits::max();
+
+    auto digit_range = [&](char from, char to) {
+        out << "[";
+        if (from == to) {
+            out << from;
         } else {
-            std::string res = repeat("(" + content + " ", up_to_n);
-            // strip trailing space
-            res = res.substr(0, res.length() - 1);
-            res += repeat(")?", up_to_n);
-            return res;
+            out << from << "-" << to;
         }
+        out << "]";
     };
+    auto more_digits = [&](int min_digits, int max_digits) {
+        out << "[0-9]";
+        if (min_digits == max_digits && min_digits == 1) {
+            return;
+        }
+        out << "{";
+        out << min_digits;
+        if (max_digits != min_digits) {
+            out << ",";
+            if (max_digits != std::numeric_limits::max()) {
+                out << max_digits;
+            }
+        }
+        out << "}";
+    };
+    std::function uniform_range =
+        [&](const string_view & from, const string_view & to) {
+            size_t i = 0;
+            while (i < from.length() && i < to.length() && from[i] == to[i]) {
+                i++;
+            }
+            if (i > 0) {
+                out << "\"" << from.substr(0, i).str() << "\"";
+            }
+            if (i < from.length() && i < to.length()) {
+                if (i > 0) {
+                    out << " ";
+                }
+                auto sub_len = from.length() - i - 1;
+                if (sub_len > 0) {
+                    auto from_sub = from.substr(i + 1);
+                    auto to_sub = to.substr(i + 1);
+                    auto sub_zeros = repeat("0", sub_len);
+                    auto sub_nines = repeat("9", sub_len);
+
+                    auto to_reached = false;
+                    out << "(";
+                    if (from_sub == sub_zeros) {
+                        digit_range(from[i], to[i] - 1);
+                        out << " ";
+                        more_digits(sub_len, sub_len);
+                    } else {
+                        out << "[" << from[i] << "] ";
+                        out << "(";
+                        uniform_range(from_sub, sub_nines);
+                        out << ")";
+                        if (from[i] < to[i] - 1) {
+                            out << " | ";
+                            if (to_sub == sub_nines) {
+                                digit_range(from[i] + 1, to[i]);
+                                to_reached = true;
+                            } else {
+                                digit_range(from[i] + 1, to[i] - 1);
+                            }
+                            out << " ";
+                            more_digits(sub_len, sub_len);
+                        }
+                    }
+                    if (!to_reached) {
+                        out << " | ";
+                        digit_range(to[i], to[i]);
+                        out << " ";
+                        uniform_range(sub_zeros, to_sub);
+                    }
+                    out << ")";
+                } else {
+                    out << "[" << from[i] << "-" << to[i] << "]";
+                }
+            }
+        };
+
+    if (has_min && has_max) {
+        if (min_value < 0 && max_value < 0) {
+            out << "\"-\" (";
+            _build_min_max_int(-max_value, -min_value, out, decimals_left, /* top_level= */ true);
+            out << ")";
+            return;
+        }
+
+        if (min_value < 0) {
+            out << "\"-\" (";
+            _build_min_max_int(0, -min_value, out, decimals_left, /* top_level= */ true);
+            out << ") | ";
+            min_value = 0;
+        }
+
+        auto min_s = std::to_string(min_value);
+        auto max_s = std::to_string(max_value);
+        auto min_digits = min_s.length();
+        auto max_digits = max_s.length();
 
-    if (min_items > 0 && max_items != min_items) {
-        result += " ";
+        for (auto digits = min_digits; digits < max_digits; digits++) {
+            uniform_range(min_s, repeat("9", digits));
+            min_s = "1" + repeat("0", digits);
+            out << " | ";
+        }
+        uniform_range(min_s, max_s);
+        return;
+    }
+
+    auto less_decimals = std::max(decimals_left - 1, 1);
+
+    if (has_min) {
+        if (min_value < 0) {
+            out << "\"-\" (";
+            _build_min_max_int(std::numeric_limits::min(), -min_value, out, decimals_left, /* top_level= */ false);
+            out << ") | [0] | [1-9] ";
+            more_digits(0, decimals_left - 1);
+        } else if (min_value == 0) {
+            if (top_level) {
+                out << "[0] | [1-9] ";
+                more_digits(0, less_decimals);
+            } else {
+                more_digits(1, decimals_left);
+            }
+        } else if (min_value <= 9) {
+            char c = '0' + min_value;
+            auto range_start = top_level ? '1' : '0';
+            if (c > range_start) {
+                digit_range(range_start, c - 1);
+                out << " ";
+                more_digits(1, less_decimals);
+                out << " | ";
+            }
+            digit_range(c, '9');
+            out << " ";
+            more_digits(0, less_decimals);
+        } else {
+            auto min_s = std::to_string(min_value);
+            auto len = min_s.length();
+            auto c = min_s[0];
+
+            if (c > '1') {
+                digit_range(top_level ? '1' : '0', c - 1);
+                out << " ";
+                more_digits(len, less_decimals);
+                out << " | ";
+            }
+            digit_range(c, c);
+            out << " (";
+            _build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits::max(), out, less_decimals, /* top_level= */ false);
+            out << ")";
+            if (c < '9') {
+                out << " | ";
+                digit_range(c + 1, '9');
+                out << " ";
+                more_digits(len - 1, less_decimals);
+            }
+        }
+        return;
     }
 
-    if (max_items != std::numeric_limits::max()) {
-        result += opt_repetitions(max_items - min_items, min_items > 0);
-    } else {
-        std::string item_operator = "(" + (separator_rule.empty() ? "" : separator_rule + " ") + item_rule + ")";
-        if (min_items == 0 && !separator_rule.empty()) {
-            result = "(" + item_rule + " " + item_operator + "*)?";
+    if (has_max) {
+        if (max_value >= 0) {
+            if (top_level) {
+                out << "\"-\" [1-9] ";
+                more_digits(0, less_decimals);
+                out << " | ";
+            }
+            _build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
         } else {
-            result += item_operator + "*";
+            out << "\"-\" (";
+            _build_min_max_int(-max_value, std::numeric_limits::max(), out, decimals_left, /* top_level= */ false);
+            out << ")";
         }
+        return;
     }
 
-    return result;
+    throw std::runtime_error("At least one of min_value or max_value must be set");
 }
 
-const std::string SPACE_RULE = "\" \"?";
+const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
 
 struct BuiltinRule {
     std::string content;
     std::vector deps;
 };
 
-const std::string _up_to_15_digits = build_repetition("[0-9]", 0, 15);
-
 std::unordered_map PRIMITIVE_RULES = {
     {"boolean", {"(\"true\" | \"false\") space", {}}},
-    {"decimal-part", {"[0-9] " + _up_to_15_digits, {}}},
-    {"integral-part", {"[0-9] | [1-9] " + _up_to_15_digits, {}}},
+    {"decimal-part", {"[0-9]{1,16}", {}}},
+    {"integral-part", {"[0] | [1-9] [0-9]{0,15}", {}}},
     {"number", {"(\"-\"? integral-part) (\".\" decimal-part)? ([eE] [-+]? integral-part)? space", {"integral-part", "decimal-part"}}},
     {"integer", {"(\"-\"? integral-part) space", {"integral-part"}}},
     {"value", {"object | array | string | number | boolean | null", {"object", "array", "string", "number", "boolean", "null"}}},
     {"object", {"\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space", {"string", "value"}}},
     {"array", {"\"[\" space ( value (\",\" space value)* )? \"]\" space", {"value"}}},
-    {"uuid", {"\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
-                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space", {}}},
-    {"char",   {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])", {}}},
+    {"uuid", {"\"\\\"\" [0-9a-fA-F]{8} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{12} \"\\\"\" space", {}}},
+    {"char",   {"[^\"\\\\\\x7F\\x00-\\x1F] | [\\\\] ([\"\\\\bfnrt] | \"u\" [0-9a-fA-F]{4})", {}}},
     {"string", {"\"\\\"\" char* \"\\\"\" space", {"char"}}},
     {"null", {"\"null\" space", {}}},
 };
 
 std::unordered_map STRING_FORMAT_RULES = {
-    {"date", {"[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
-    {"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
+    {"date", {"[0-9]{4} \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
+    {"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9]{3} )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
     {"date-time", {"date \"T\" time", {"date", "time"}}},
     {"date-string", {"\"\\\"\" date \"\\\"\" space", {"date"}}},
     {"time-string", {"\"\\\"\" time \"\\\"\" space", {"time"}}},
@@ -126,7 +316,7 @@ std::unordered_map GRAMMAR_LITERAL_ESCAPES = {
 };
 
 std::unordered_set NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
-std::unordered_set ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
+std::unordered_set ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
 
 template 
 std::string join(Iterator begin, Iterator end, const std::string & separator) {
@@ -197,7 +387,6 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
-
 class SchemaConverter {
 private:
     std::function _fetch_json;
@@ -385,8 +574,7 @@ private:
                         sub_is_literal ? "\"" + sub + "\"" : sub,
                         min_times,
                         max_times,
-                        "",
-                        sub_is_literal
+                        ""
                     );
                     seq.back().second = false;
                 } else {
@@ -426,6 +614,75 @@ private:
         return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
     }
 
+    /*
+        Returns a rule that matches a JSON string that is none of the provided strings
+
+        not_strings({"a"})
+            -> ["] ( [a] char+ | [^"a] char* )? ["] space
+        not_strings({"and", "also"})
+            -> ["] ( [a] ([l] ([s] ([o] char+ | [^"o] char*) | [^"s] char*) | [n] ([d] char+ | [^"d] char*) | [^"ln] char*) | [^"a] char* )? ["] space
+    */
+    std::string _not_strings(const std::vector & strings) {
+
+        struct TrieNode {
+            std::map children;
+            bool is_end_of_string;
+
+            TrieNode() : is_end_of_string(false) {}
+
+            void insert(const std::string & string) {
+                auto node = this;
+                for (char c : string) {
+                    node = &node->children[c];
+                }
+                node->is_end_of_string = true;
+            }
+        };
+
+        TrieNode trie;
+        for (const auto & s : strings) {
+            trie.insert(s);
+        }
+
+        std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
+        std::ostringstream out;
+        out << "[\"] ( ";
+        std::function visit = [&](const TrieNode & node) {
+            std::ostringstream rejects;
+            auto first = true;
+            for (const auto & kv : node.children) {
+                rejects << kv.first;
+                if (first) {
+                    first = false;
+                } else {
+                    out << " | ";
+                }
+                out << "[" << kv.first << "]";
+                if (!kv.second.children.empty()) {
+                    out << " (";
+                    visit(kv.second);
+                    out << ")";
+                } else if (kv.second.is_end_of_string) {
+                    out << " " << char_rule << "+";
+                }
+            }
+            if (!node.children.empty()) {
+                if (!first) {
+                    out << " | ";
+                }
+                out << "[^\"" << rejects.str() << "] " << char_rule << "*";
+            }
+        };
+        visit(trie);
+
+        out << " )";
+        if (!trie.is_end_of_string) {
+            out << "?";
+        }
+        out << " [\"] space";
+        return out.str();
+    }
+
     std::string _resolve_ref(const std::string & ref) {
         std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
@@ -446,6 +703,7 @@ private:
         std::vector required_props;
         std::vector optional_props;
         std::unordered_map prop_kv_rule_names;
+        std::vector prop_names;
         for (const auto & kv : properties) {
             const auto &prop_name = kv.first;
             const auto &prop_schema = kv.second;
@@ -460,11 +718,18 @@ private:
             } else {
                 optional_props.push_back(prop_name);
             }
+            prop_names.push_back(prop_name);
         }
-        if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get())) {
+        if ((additional_properties.is_boolean() && additional_properties.get()) || additional_properties.is_object()) {
             std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
-            std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value");
-            std::string kv_rule = _add_rule(sub_name + "-kv", _add_primitive("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
+            std::string value_rule =
+                additional_properties.is_object() ? visit(additional_properties, sub_name + "-value")
+                : _add_primitive("value", PRIMITIVE_RULES.at("value"));
+
+            auto key_rule =
+                prop_names.empty() ? _add_primitive("string", PRIMITIVE_RULES.at("string"))
+                : _add_rule(sub_name + "-k", _not_strings(prop_names));
+            std::string kv_rule = _add_rule(sub_name + "-kv", key_rule + " \":\" space " + value_rule);
             prop_kv_rule_names["*"] = kv_rule;
             optional_props.push_back("*");
         }
@@ -490,15 +755,11 @@ private:
                 }
                 std::string k = ks[0];
                 std::string kv_rule_name = prop_kv_rule_names[k];
-                if (k == "*") {
-                    res = _add_rule(
-                        name + (name.empty() ? "" : "-") + "additional-kvs",
-                        kv_rule_name + " ( \",\" space " + kv_rule_name + " )*"
-                    );
-                } else if (first_is_optional) {
-                    res = "( \",\" space " + kv_rule_name + " )?";
+                std::string comma_ref = "( \",\" space " + kv_rule_name + " )";
+                if (first_is_optional) {
+                    res = comma_ref + (k == "*" ? "*" : "?");
                 } else {
-                    res = kv_rule_name;
+                    res = kv_rule_name + (k == "*" ? " " + comma_ref + "*" : "");
                 }
                 if (ks.size() > 1) {
                     res += " " + _add_rule(
@@ -632,17 +893,19 @@ public:
         } else if (schema_type.is_array()) {
             std::vector schema_types;
             for (const auto & t : schema_type) {
-                schema_types.push_back({{"type", t}});
+                json schema_copy(schema);
+                schema_copy["type"] = t;
+                schema_types.push_back(schema_copy);
             }
             return _add_rule(rule_name, _generate_union_rule(name, schema_types));
         } else if (schema.contains("const")) {
-            return _add_rule(rule_name, _generate_constant_rule(schema["const"]));
+            return _add_rule(rule_name, _generate_constant_rule(schema["const"]) + " space");
         } else if (schema.contains("enum")) {
             std::vector enum_values;
             for (const auto & v : schema["enum"]) {
                 enum_values.push_back(_generate_constant_rule(v));
             }
-            return _add_rule(rule_name, join(enum_values.begin(), enum_values.end(), " | "));
+            return _add_rule(rule_name, "(" + join(enum_values.begin(), enum_values.end(), " | ") + ") space");
         } else if ((schema_type.is_null() || schema_type == "object")
                 && (schema.contains("properties") ||
                     (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
@@ -724,6 +987,24 @@ public:
             int min_len = schema.contains("minLength") ? schema["minLength"].get() : 0;
             int max_len = schema.contains("maxLength") ? schema["maxLength"].get() : std::numeric_limits::max();
             return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
+        } else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
+            int min_value = std::numeric_limits::min();
+            int max_value = std::numeric_limits::max();
+            if (schema.contains("minimum")) {
+                min_value = schema["minimum"].get();
+            } else if (schema.contains("exclusiveMinimum")) {
+                min_value = schema["exclusiveMinimum"].get() + 1;
+            }
+            if (schema.contains("maximum")) {
+                max_value = schema["maximum"].get();
+            } else if (schema.contains("exclusiveMaximum")) {
+                max_value = schema["exclusiveMaximum"].get() - 1;
+            }
+            std::stringstream out;
+            out << "(";
+            _build_min_max_int(min_value, max_value, out);
+            out << ") space";
+            return _add_rule(rule_name, out.str());
         } else if (schema.empty() || schema_type == "object") {
             return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
         } else {
diff --git a/llm/llama.cpp/common/log.h b/llm/llama.cpp/common/log.h
index 09fa63c2677a707d7cd78998c4a212eb84197c7a..1bc5328ce3e119dd91e19e9c6986c1b60ccc5f15 100644
--- a/llm/llama.cpp/common/log.h
+++ b/llm/llama.cpp/common/log.h
@@ -630,7 +630,7 @@ inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
     buf << "[ ";
 
     bool first = true;
-    for (const auto &token : tokens)
+    for (const auto & token : tokens)
     {
         if (!first) {
             buf << ", ";
diff --git a/llm/llama.cpp/common/ngram-cache.h b/llm/llama.cpp/common/ngram-cache.h
index e4fa4cbd12f11eea481a3d22dd975f64f5c7a6f2..ab4c9b37665460281683a6bb09bcddd444eb3634 100644
--- a/llm/llama.cpp/common/ngram-cache.h
+++ b/llm/llama.cpp/common/ngram-cache.h
@@ -37,11 +37,18 @@ struct llama_ngram {
     }
 };
 
+struct llama_token_hash_function {
+    size_t operator()(const llama_token token) const {
+        // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
+        return token * 11400714819323198485llu;
+    }
+};
+
 struct llama_ngram_hash_function {
     size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = 0;
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= std::hash{}(ngram.tokens[i]);
+        size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
+        for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
+            hash ^= llama_token_hash_function{}(ngram.tokens[i]);
         }
         return hash;
     }
diff --git a/llm/llama.cpp/common/sampling.cpp b/llm/llama.cpp/common/sampling.cpp
index f1f80351637f0965ebffa64f74e3c9c837963259..079e405168dff237f1c8c49abedb0c1b387b76bf 100644
--- a/llm/llama.cpp/common/sampling.cpp
+++ b/llm/llama.cpp/common/sampling.cpp
@@ -28,9 +28,13 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
         std::vector grammar_rules(result->parsed_grammar.c_rules());
 
-        result->grammar = llama_grammar_init(
+        struct llama_grammar * grammar = llama_grammar_init(
                 grammar_rules.data(),
                 grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
+        if (grammar == nullptr) {
+            throw std::runtime_error("Failed to initialize llama_grammar");
+        }
+        result->grammar = grammar;
     }
 
     result->prev.resize(params.n_prev);
@@ -59,9 +63,13 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
     if (!ctx->parsed_grammar.rules.empty()) {
         std::vector grammar_rules(ctx->parsed_grammar.c_rules());
 
-        ctx->grammar = llama_grammar_init(
+        struct llama_grammar * grammar = llama_grammar_init(
                 grammar_rules.data(),
                 grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
+        if (grammar == nullptr) {
+            throw std::runtime_error("Failed to initialize llama_grammar");
+        }
+        ctx->grammar = grammar;
     }
 
     std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
@@ -274,8 +282,6 @@ static llama_token llama_sampling_sample_impl(
         GGML_ASSERT(!original_logits.empty());
     }
     llama_token id = 0;
-    // Get a pointer to the logits
-    float * logits = llama_get_logits_ith(ctx_main, idx);
 
     if (temp < 0.0) {
         // greedy sampling, with probs
@@ -316,12 +322,15 @@ static llama_token llama_sampling_sample_impl(
     }
 
     if (ctx_sampling->grammar != NULL && !is_resampling) {
+        // Get a pointer to the logits
+        float * logits = llama_get_logits_ith(ctx_main, idx);
+
         // Create an array with a single token data element for the sampled id
         llama_token_data single_token_data = {id, logits[id], 0.0f};
         llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
 
         // Apply grammar constraints to the single token
-        llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
+        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array);
 
         // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
         bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
@@ -369,7 +378,7 @@ static llama_token_data_array llama_sampling_prepare_impl(
     if (ctx_sampling->grammar != NULL && !apply_grammar) {
         GGML_ASSERT(original_logits != NULL);
         // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
-        *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
+        *original_logits = {logits, logits + n_vocab};
     }
 
     // apply params.logit_bias map
@@ -382,10 +391,10 @@ static llama_token_data_array llama_sampling_prepare_impl(
         llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
     }
 
-    cur.clear();
+    cur.resize(n_vocab);
 
     for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+        cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
     }
 
     llama_token_data_array cur_p = { cur.data(), cur.size(), false };
@@ -412,7 +421,7 @@ static llama_token_data_array llama_sampling_prepare_impl(
 
     // apply grammar checks before sampling logic
     if (apply_grammar && ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
+        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p);
     }
 
     return cur_p;
@@ -446,6 +455,6 @@ void llama_sampling_accept(
     ctx_sampling->prev.push_back(id);
 
     if (ctx_sampling->grammar != NULL && apply_grammar) {
-        llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
+        llama_grammar_accept_token(ctx_sampling->grammar, ctx_main, id);
     }
 }
diff --git a/llm/llama.cpp/convert_hf_to_gguf.py b/llm/llama.cpp/convert_hf_to_gguf.py
new file mode 100644
index 0000000000000000000000000000000000000000..38b92bc8110eac1ce8d3cc5b581aa45bd84fe6d1
--- /dev/null
+++ b/llm/llama.cpp/convert_hf_to_gguf.py
@@ -0,0 +1,3823 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import logging
+import argparse
+import contextlib
+import json
+import os
+import re
+import sys
+from enum import IntEnum
+from pathlib import Path
+from hashlib import sha256
+from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
+
+import math
+import numpy as np
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+import gguf
+
+logger = logging.getLogger("hf-to-gguf")
+
+
+###### MODEL DEFINITIONS ######
+
+class SentencePieceTokenTypes(IntEnum):
+    NORMAL = 1
+    UNKNOWN = 2
+    CONTROL = 3
+    USER_DEFINED = 4
+    UNUSED = 5
+    BYTE = 6
+
+
+AnyModel = TypeVar("AnyModel", bound="type[Model]")
+
+
+class Model:
+    _model_classes: dict[str, type[Model]] = {}
+
+    dir_model: Path
+    ftype: gguf.LlamaFileType
+    fname_out: Path
+    is_big_endian: bool
+    endianess: gguf.GGUFEndian
+    use_temp_file: bool
+    lazy: bool
+    part_names: list[str]
+    is_safetensors: bool
+    hparams: dict[str, Any]
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+    tensor_names: set[str] | None
+    gguf_writer: gguf.GGUFWriter
+    model_name: str | None
+    metadata_override: Path | None
+    dir_model_card: Path
+
+    # subclasses should define this!
+    model_arch: gguf.MODEL_ARCH
+
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
+                 use_temp_file: bool = False, eager: bool = False,
+                 metadata_override: Path | None = None, model_name: str | None = None,
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
+        if type(self) is Model:
+            raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
+
+        self.dir_model = dir_model
+        self.ftype = ftype
+        self.fname_out = fname_out
+        self.is_big_endian = is_big_endian
+        self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
+        self.use_temp_file = use_temp_file
+        self.lazy = not eager
+        self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors")
+        self.is_safetensors = len(self.part_names) > 0
+        if not self.is_safetensors:
+            self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
+        self.hparams = Model.load_hparams(self.dir_model)
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+        self.tensor_names = None
+        self.metadata_override = metadata_override
+        self.model_name = model_name
+        self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
+
+        # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type
+        if self.ftype == gguf.LlamaFileType.GUESSED:
+            # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
+            _, first_tensor = next(self.get_tensors())
+            if first_tensor.dtype == torch.float16:
+                logger.info(f"choosing --outtype f16 from first tensor type ({first_tensor.dtype})")
+                self.ftype = gguf.LlamaFileType.MOSTLY_F16
+            else:
+                logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})")
+                self.ftype = gguf.LlamaFileType.MOSTLY_BF16
+
+        # Configure GGUF Writer
+        self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
+                                           split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
+
+    @classmethod
+    def __init_subclass__(cls):
+        # can't use an abstract property, because overriding it without type errors
+        # would require using decorated functions instead of simply defining the property
+        if "model_arch" not in cls.__dict__:
+            raise TypeError(f"Missing property 'model_arch' for {cls.__name__!r}")
+
+    def find_hparam(self, keys: Iterable[str], optional: bool = False) -> Any:
+        key = next((k for k in keys if k in self.hparams), None)
+        if key is not None:
+            return self.hparams[key]
+        if optional:
+            return None
+        raise KeyError(f"could not find any of: {keys}")
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+        tensor_names_from_parts: set[str] = set()
+
+        if len(self.part_names) > 1:
+            self.tensor_names = set()
+            index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
+            index_name += ".index.json"
+            logger.info(f"gguf: loading model weight map from '{index_name}'")
+            with open(self.dir_model / index_name, "r", encoding="utf-8") as f:
+                index: dict[str, Any] = json.load(f)
+                weight_map = index.get("weight_map")
+                if weight_map is None or not isinstance(weight_map, dict):
+                    raise ValueError(f"Can't load 'weight_map' from {index_name!r}")
+                self.tensor_names.update(weight_map.keys())
+        else:
+            self.tensor_names = tensor_names_from_parts
+
+        for part_name in self.part_names:
+            logger.info(f"gguf: loading model part '{part_name}'")
+            ctx: ContextManager[Any]
+            if self.is_safetensors:
+                from safetensors import safe_open
+                ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))
+            else:
+                ctx = contextlib.nullcontext(torch.load(str(self.dir_model / part_name), map_location="cpu", mmap=True, weights_only=True))
+
+            with ctx as model_part:
+                tensor_names_from_parts.update(model_part.keys())
+
+                for name in model_part.keys():
+                    if self.is_safetensors:
+                        if self.lazy:
+                            data = model_part.get_slice(name)
+                            data = LazyTorchTensor.from_safetensors_slice(data)
+                        else:
+                            data = model_part.get_tensor(name)
+                    else:
+                        data = model_part[name]
+                        if self.lazy:
+                            data = LazyTorchTensor.from_eager(data)
+                    yield name, data
+
+        # only verify tensor name presence; it doesn't matter if they are not in the right files
+        if len(sym_diff := tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
+            raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")
+
+    def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
+        if key not in gguf.MODEL_TENSORS[self.model_arch]:
+            raise ValueError(f"Missing {key!r} for MODEL_TENSORS of {self.model_arch!r}")
+        name: str = gguf.TENSOR_NAMES[key]
+        if "{bid}" in name:
+            assert bid is not None
+            name = name.format(bid=bid)
+        return name + suffix
+
+    def match_model_tensor_name(self, name: str, key: gguf.MODEL_TENSOR, bid: int | None, suffix: str = ".weight") -> bool:
+        if key not in gguf.MODEL_TENSORS[self.model_arch]:
+            return False
+        key_name: str = gguf.TENSOR_NAMES[key]
+        if "{bid}" in key_name:
+            if bid is None:
+                return False
+            key_name = key_name.format(bid=bid)
+        else:
+            if bid is not None:
+                return False
+        return name == (key_name + suffix)
+
+    def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
+        new_name = self.tensor_map.get_name(key=name, try_suffixes=try_suffixes)
+        if new_name is None:
+            raise ValueError(f"Can not map tensor {name!r}")
+        return new_name
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_block_count(self.block_count)
+
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
+            self.gguf_writer.add_context_length(n_ctx)
+            logger.info(f"gguf: context length = {n_ctx}")
+
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        self.gguf_writer.add_embedding_length(n_embd)
+        logger.info(f"gguf: embedding length = {n_embd}")
+
+        if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
+            self.gguf_writer.add_feed_forward_length(n_ff)
+            logger.info(f"gguf: feed forward length = {n_ff}")
+
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_head_count(n_head)
+        logger.info(f"gguf: head count = {n_head}")
+
+        if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
+            self.gguf_writer.add_head_count_kv(n_head_kv)
+            logger.info(f"gguf: key-value head count = {n_head_kv}")
+
+        if (rope_theta := self.hparams.get("rope_theta")) is not None:
+            self.gguf_writer.add_rope_freq_base(rope_theta)
+            logger.info(f"gguf: rope theta = {rope_theta}")
+        if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+            logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
+        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
+            self.gguf_writer.add_layer_norm_eps(f_norm_eps)
+            logger.info(f"gguf: layer norm epsilon = {f_norm_eps}")
+        if (n_experts := self.hparams.get("num_local_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+            logger.info(f"gguf: expert count = {n_experts}")
+        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+            self.gguf_writer.add_expert_used_count(n_experts_used)
+            logger.info(f"gguf: experts used count = {n_experts_used}")
+
+        if (head_dim := self.hparams.get("head_dim")) is not None:
+            self.gguf_writer.add_key_length(head_dim)
+            self.gguf_writer.add_value_length(head_dim)
+
+        self.gguf_writer.add_file_type(self.ftype)
+        logger.info(f"gguf: file type = {self.ftype}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+        del name, new_name, bid, n_dims  # unused
+
+        return False
+
+    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+        del name, new_name, bid, n_dims  # unused
+
+        return False
+
+    def prepare_tensors(self):
+        max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
+
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            # use the first number-like part of the tensor name as the block id
+            bid = None
+            for part in name.split("."):
+                if part.isdecimal():
+                    bid = int(part)
+                    break
+
+            for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
+                data: np.ndarray  # type hint
+                n_dims = len(data.shape)
+                data_dtype = data.dtype
+                data_qtype: gguf.GGMLQuantizationType | None = None
+
+                # when both are True, f32 should win
+                extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
+                extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)
+
+                # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
+                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
+                extra_f32 = any(cond for cond in (
+                    extra_f32,
+                    n_dims == 1,
+                    new_name.endswith("_norm.weight"),
+                ))
+
+                # Some tensor types are always in float32
+                extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
+                    gguf.MODEL_TENSOR.FFN_GATE_INP,
+                    gguf.MODEL_TENSOR.POS_EMBD,
+                    gguf.MODEL_TENSOR.TOKEN_TYPES,
+                ))
+
+                # if f16 desired, convert any float32 2-dim weight tensors to float16
+                extra_f16 = any(cond for cond in (
+                    extra_f16,
+                    (name.endswith(".weight") and n_dims >= 2),
+                ))
+
+                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
+                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        data = gguf.quantize_bf16(data)
+                        assert data.dtype == np.uint16
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
+                        data = gguf.quantize_q8_0(data)
+                        assert data.dtype == np.uint8
+                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+
+                    else:  # default to float16 for quantized tensors
+                        if data_dtype != np.float16:
+                            data = data.astype(np.float16)
+                        data_qtype = gguf.GGMLQuantizationType.F16
+
+                if data_qtype is None:  # by default, convert to float32
+                    if data_dtype != np.float32:
+                        data = data.astype(np.float32)
+                    data_qtype = gguf.GGMLQuantizationType.F32
+
+                shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
+
+                # reverse shape to make it similar to the internal ggml dimension order
+                shape_str = f"{{{', '.join(str(n) for n in reversed(shape))}}}"
+
+                # n_dims is implicit in the shape
+                logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
+
+                self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype)
+
+    def set_type(self):
+        self.gguf_writer.add_type(gguf.GGUFType.MODEL)
+
+    def prepare_metadata(self, vocab_only: bool):
+
+        total_params, shared_params, expert_params, expert_count = self.gguf_writer.get_total_parameter_count()
+
+        self.metadata = gguf.Metadata.load(self.metadata_override, self.dir_model_card, self.model_name, total_params)
+
+        # Fallback to model directory name if metadata name is still missing
+        if self.metadata.name is None:
+            self.metadata.name = self.dir_model.name
+
+        # Generate parameter weight class (useful for leader boards) if not yet determined
+        if self.metadata.size_label is None and total_params > 0:
+            self.metadata.size_label = gguf.size_label(total_params, shared_params, expert_params, expert_count)
+
+        # Extract the encoding scheme from the file type name. e.g. 'gguf.LlamaFileType.MOSTLY_Q8_0' --> 'Q8_0'
+        output_type: str = self.ftype.name.partition("_")[2]
+
+        # Filename Output
+        if self.fname_out.is_dir():
+            # Generate default filename based on model specification and available metadata
+            if not vocab_only:
+                fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, self.metadata.size_label, output_type, model_type="LoRA" if total_params < 0 else None)
+            else:
+                fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, size_label=None, output_type=None, model_type="vocab")
+
+            # Use the default filename
+            self.fname_out = self.fname_out / f"{fname_default}.gguf"
+        else:
+            # Output path is a custom defined templated filename
+            # Note: `not is_dir()` is used because `.is_file()` will not detect
+            #       file template strings as it doesn't actually exist as a file
+
+            # Process templated file name with the output ftype, useful with the "auto" ftype
+            self.fname_out = self.fname_out.parent / gguf.fill_templated_filename(self.fname_out.name, output_type)
+
+        self.set_type()
+
+        logger.info("Set meta model")
+        self.metadata.set_gguf_meta_model(self.gguf_writer)
+
+        logger.info("Set model parameters")
+        self.set_gguf_parameters()
+
+        logger.info("Set model tokenizer")
+        self.set_vocab()
+
+        logger.info("Set model quantization version")
+        self.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION)
+
+    def write(self):
+        self.prepare_tensors()
+        self.prepare_metadata(vocab_only=False)
+        self.gguf_writer.write_header_to_file(path=self.fname_out)
+        self.gguf_writer.write_kv_data_to_file()
+        self.gguf_writer.write_tensors_to_file(progress=True)
+        self.gguf_writer.close()
+
+    def write_vocab(self):
+        if len(self.gguf_writer.tensors) != 1:
+            raise ValueError('Splitting the vocabulary is not supported')
+
+        self.prepare_metadata(vocab_only=True)
+        self.gguf_writer.write_header_to_file(path=self.fname_out)
+        self.gguf_writer.write_kv_data_to_file()
+        self.gguf_writer.close()
+
+    @staticmethod
+    def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]:
+        part_names: list[str] = []
+        for filename in os.listdir(dir_model):
+            if filename.startswith(prefix) and filename.endswith(suffix):
+                part_names.append(filename)
+
+        part_names.sort()
+
+        return part_names
+
+    @staticmethod
+    def load_hparams(dir_model: Path):
+        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+            return json.load(f)
+
+    @classmethod
+    def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
+        assert names
+
+        def func(modelcls: AnyModel) -> AnyModel:
+            for name in names:
+                cls._model_classes[name] = modelcls
+            return modelcls
+        return func
+
+    @classmethod
+    def from_model_architecture(cls, arch: str) -> type[Model]:
+        try:
+            return cls._model_classes[arch]
+        except KeyError:
+            raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
+
+    def does_token_look_special(self, token: str | bytes) -> bool:
+        if isinstance(token, (bytes, bytearray)):
+            token_text = token.decode(encoding="utf-8")
+        elif isinstance(token, memoryview):
+            token_text = token.tobytes().decode(encoding="utf-8")
+        else:
+            token_text = token
+
+        # Some models mark some added tokens which ought to be control tokens as not special.
+        # (e.g. command-r, command-r-plus, deepseek-coder, gemma{,-2})
+        seems_special = token_text in (
+            "",  # deepseek-coder
+            "", "<2mass>", "[@BOS@]",  # gemma{,-2}
+        )
+
+        seems_special = seems_special or (token_text.startswith("<|") and token_text.endswith("|>"))
+        seems_special = seems_special or (token_text.startswith("<｜") and token_text.endswith("｜>"))  # deepseek-coder
+
+        # TODO: should these be marked as UNUSED instead? (maybe not)
+        seems_special = seems_special or (token_text.startswith(""))  # gemma{,-2}
+
+        return seems_special
+
+    # used for GPT-2 BPE and WordPiece vocabs
+    def get_vocab_base(self) -> tuple[list[str], list[int], str]:
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))
+        assert max(tokenizer.vocab.values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        return tokens, toktypes, tokpre
+
+    # NOTE: this function is generated by convert_hf_to_gguf_update.py
+    #       do not modify it manually!
+    # ref:  https://github.com/ggerganov/llama.cpp/pull/6920
+    # Marker: Start get_vocab_base_pre
+    def get_vocab_base_pre(self, tokenizer) -> str:
+        # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
+        # is specific for the BPE pre-tokenizer used by the model
+        # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
+        # use in llama.cpp to implement the same pre-tokenizer
+
+        chktxt = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶\u200d🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````""""......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
+
+        chktok = tokenizer.encode(chktxt)
+        chkhsh = sha256(str(chktok).encode()).hexdigest()
+
+        logger.debug(f"chktok: {chktok}")
+        logger.debug(f"chkhsh: {chkhsh}")
+
+        res = None
+
+        # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script
+        #       or pull the latest version of the model from Huggingface
+        #       don't edit the hashes manually!
+        if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
+            # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
+            res = "llama-bpe"
+        if chkhsh == "049ecf7629871e3041641907f3de7c733e4dbfdc736f57d882ba0b0845599754":
+            # ref: https://huggingface.co/deepseek-ai/deepseek-llm-7b-base
+            res = "deepseek-llm"
+        if chkhsh == "347715f544604f9118bb75ed199f68779f423cabb20db6de6f31b908d04d7821":
+            # ref: https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base
+            res = "deepseek-coder"
+        if chkhsh == "8aeee3860c56296a157a1fe2fad249ec40aa59b1bb5709f4ade11c4e6fe652ed":
+            # ref: https://huggingface.co/tiiuae/falcon-7b
+            res = "falcon"
+        if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
+            # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
+            res = "bert-bge"
+        if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
+            # ref: https://huggingface.co/mosaicml/mpt-7b
+            res = "mpt"
+        if chkhsh == "35d91631860c815f952d711435f48d356ebac988362536bed955d43bfa436e34":
+            # ref: https://huggingface.co/bigcode/starcoder2-3b
+            res = "starcoder"
+        if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454":
+            # ref: https://huggingface.co/openai-community/gpt2
+            res = "gpt-2"
+        if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3":
+            # ref: https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b
+            res = "stablelm2"
+        if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
+            # ref: https://huggingface.co/smallcloudai/Refact-1_6-base
+            res = "refact"
+        if chkhsh == "9c2227e4dd922002fb81bde4fc02b0483ca4f12911410dee2255e4987644e3f8":
+            # ref: https://huggingface.co/CohereForAI/c4ai-command-r-v01
+            res = "command-r"
+        if chkhsh == "e636dc30a262dcc0d8c323492e32ae2b70728f4df7dfe9737d9f920a282b8aea":
+            # ref: https://huggingface.co/Qwen/Qwen1.5-7B
+            res = "qwen2"
+        if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
+            # ref: https://huggingface.co/allenai/OLMo-1.7-7B-hf
+            res = "olmo"
+        if chkhsh == "a8594e3edff7c29c003940395316294b2c623e09894deebbc65f33f1515df79e":
+            # ref: https://huggingface.co/databricks/dbrx-base
+            res = "dbrx"
+        if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
+            res = "jina-v2-en"
+        if chkhsh == "171aeeedd6fb548d418a7461d053f11b6f1f1fc9b387bd66640d28a4b9f5c643":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-es
+            res = "jina-v2-es"
+        if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
+            res = "jina-v2-de"
+        if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
+            # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
+            res = "smaug-bpe"
+        if chkhsh == "c7ea5862a53e4272c035c8238367063e2b270d51faa48c0f09e9d5b54746c360":
+            # ref: https://huggingface.co/LumiOpen/Poro-34B-chat
+            res = "poro-chat"
+        if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
+            res = "jina-v2-code"
+        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
+            # ref: https://huggingface.co/LumiOpen/Viking-7B
+            res = "viking"
+        if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901":
+            # ref: https://huggingface.co/core42/jais-13b
+            res = "jais"
+        if chkhsh == "7b3e7548e4308f52a76e8229e4e6cc831195d0d1df43aed21ac6c93da05fec5f":
+            # ref: https://huggingface.co/WisdomShell/CodeShell-7B
+            res = "codeshell"
+        if chkhsh == "63b97e4253352e6f357cc59ea5b583e3a680eaeaf2632188c2b952de2588485e":
+            # ref: https://huggingface.co/mistralai/Mistral-Nemo-Base-2407
+            res = "tekken"
+        if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
+            # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
+            res = "smollm"
+
+        if res is None:
+            logger.warning("\n")
+            logger.warning("**************************************************************************************")
+            logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!")
+            logger.warning("**          There are 2 possible reasons for this:")
+            logger.warning("**          - the model has not been added to convert_hf_to_gguf_update.py yet")
+            logger.warning("**          - the pre-tokenization config has changed upstream")
+            logger.warning("**          Check your model files and convert_hf_to_gguf_update.py and update them accordingly.")
+            logger.warning("** ref:     https://github.com/ggerganov/llama.cpp/pull/6920")
+            logger.warning("**")
+            logger.warning(f"** chkhsh:  {chkhsh}")
+            logger.warning("**************************************************************************************")
+            logger.warning("\n")
+            raise NotImplementedError("BPE pre-tokenizer was not recognized - update get_vocab_base_pre()")
+
+        logger.debug(f"tokenizer.ggml.pre: {repr(res)}")
+        logger.debug(f"chkhsh: {chkhsh}")
+
+        return res
+        # Marker: End get_vocab_base_pre
+
+    def _set_vocab_gpt2(self) -> None:
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def _set_vocab_qwen(self):
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        vocab_size = hparams["vocab_size"]
+        assert max(tokenizer.get_vocab().values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        merges = []
+        vocab = {}
+        mergeable_ranks = tokenizer.mergeable_ranks
+        for token, rank in mergeable_ranks.items():
+            vocab[QwenModel.token_bytes_to_string(token)] = rank
+            if len(token) == 1:
+                continue
+            merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+            assert len(merged) == 2
+            merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+
+        # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
+        added_vocab = tokenizer.special_tokens
+        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            elif reverse_vocab[i] in added_vocab:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.CONTROL)
+            else:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.NORMAL)
+
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
+        special_vocab.merges = merges
+        # only add special tokens when they were not already loaded from config.json
+        if len(special_vocab.special_token_ids) == 0:
+            special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
+            special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
+        # this one is usually not in config.json anyway
+        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def _set_vocab_sentencepiece(self, add_to_gguf=True):
+        tokens, scores, toktypes = self._create_vocab_sentencepiece()
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def _create_vocab_sentencepiece(self):
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
+                for token_id, token_data in added_tokens_decoder.items():
+                    token_id = int(token_id)
+                    token: str = token_data["content"]
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
+                        if tokens[token_id] != token.encode("utf-8"):
+                            logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}')
+                    if token_data.get("special") or self.does_token_look_special(token):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+                    else:
+                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
+                        toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+                    scores[token_id] = -1000.0
+                    tokens[token_id] = token.encode("utf-8")
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        return tokens, scores, toktypes
+
+    def _set_vocab_llama_hf(self):
+        vocab = gguf.LlamaHfVocab(self.dir_model)
+        tokens = []
+        scores = []
+        toktypes = []
+
+        for text, score, toktype in vocab.all_tokens():
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        assert len(tokens) == vocab.vocab_size
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int):
+        tokenizer_path = Path(sys.path[0]) / "models" / f"ggml-vocab-{model_name}.gguf"
+        logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
+        vocab_reader = gguf.GGUFReader(tokenizer_path, "r")
+
+        default_pre = "mpt" if model_name == "gpt-neox" else "default"
+
+        field = vocab_reader.get_field(gguf.Keys.Tokenizer.MODEL)
+        assert field  # tokenizer model
+        self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]).decode("utf-8"))
+
+        field = vocab_reader.get_field(gguf.Keys.Tokenizer.PRE)
+        self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]).decode("utf-8") if field else default_pre)
+
+        field = vocab_reader.get_field(gguf.Keys.Tokenizer.LIST)
+        assert field  # token list
+        self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
+
+        if model_name == "llama-spm":
+            field = vocab_reader.get_field(gguf.Keys.Tokenizer.SCORES)
+            assert field  # token scores
+            self.gguf_writer.add_token_scores([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
+
+        field = vocab_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
+        assert field  # token types
+        self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
+
+        if model_name != "llama-spm":
+            field = vocab_reader.get_field(gguf.Keys.Tokenizer.MERGES)
+            assert field  # token merges
+            self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
+
+        if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)) is not None:
+            self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
+        if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)) is not None:
+            self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
+        if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)) is not None:
+            self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
+        if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.PAD_ID)) is not None:
+            self.gguf_writer.add_pad_token_id(field.parts[-1].tolist()[0])
+        if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_BOS)) is not None:
+            self.gguf_writer.add_add_bos_token(field.parts[-1].tolist()[0])
+        if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_EOS)) is not None:
+            self.gguf_writer.add_add_eos_token(field.parts[-1].tolist()[0])
+
+
+@Model.register("GPTNeoXForCausalLM")
+class GPTNeoXModel(Model):
+    model_arch = gguf.MODEL_ARCH.GPTNEOX
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(
+            int(self.hparams["rotary_pct"] * (self.hparams["hidden_size"] // self.hparams["num_attention_heads"])),
+        )
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        if re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.weight", name):
+            # Map bloom-style qkv_linear to gpt-style qkv_linear
+            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+            qkv_weights = data_torch.reshape((n_head, 3, n_embed // n_head, n_embed))
+            data_torch = torch.cat(
+                (
+                    qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.weight")
+        elif re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.bias", name):
+            qkv_bias = data_torch.reshape((n_head, 3, n_embed // n_head))
+            data_torch = torch.cat(
+                (
+                    qkv_bias[:, 0, :].reshape((n_embed,)),
+                    qkv_bias[:, 1, :].reshape((n_embed,)),
+                    qkv_bias[:, 2, :].reshape((n_embed,)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.bias")
+
+        tensors.append((self.map_tensor_name(name), data_torch))
+
+        return tensors
+
+
+@Model.register("BloomForCausalLM")
+class BloomModel(Model):
+    model_arch = gguf.MODEL_ARCH.BLOOM
+
+    def set_gguf_parameters(self):
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
+        self.gguf_writer.add_embedding_length(n_embed)
+        self.gguf_writer.add_feed_forward_length(4 * n_embed)
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        name = re.sub(r'transformer\.', '', name)
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
+            # Map bloom-style qkv_linear to gpt-style qkv_linear
+            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+            qkv_weights = data_torch.reshape((n_head, 3, n_embed // n_head, n_embed))
+            data_torch = torch.cat(
+                (
+                    qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.weight")
+        elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
+            qkv_bias = data_torch.reshape((n_head, 3, n_embed // n_head))
+            data_torch = torch.cat(
+                (
+                    qkv_bias[:, 0, :].reshape((n_embed,)),
+                    qkv_bias[:, 1, :].reshape((n_embed,)),
+                    qkv_bias[:, 2, :].reshape((n_embed,)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.bias")
+
+        tensors.append((self.map_tensor_name(name), data_torch))
+
+        if name == "word_embeddings.weight":
+            assert self.tensor_names is not None
+
+            # TODO: tie them at runtime, don't duplicate in the model file
+            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
+
+        return tensors
+
+
+@Model.register("MPTForCausalLM")
+class MPTModel(Model):
+    model_arch = gguf.MODEL_ARCH.MPT
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_gpt2()
+        except Exception:
+            # Fallback for SEA-LION model
+            self._set_vocab_sentencepiece()
+            self.gguf_writer.add_add_bos_token(False)
+            self.gguf_writer.add_pad_token_id(3)
+            self.gguf_writer.add_eos_token_id(1)
+            self.gguf_writer.add_unk_token_id(0)
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["n_layers"]
+        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["d_model"])
+        self.gguf_writer.add_head_count(self.hparams["n_heads"])
+        if kv_n_heads := self.hparams["attn_config"].get("kv_n_heads"):
+            self.gguf_writer.add_head_count_kv(kv_n_heads)
+        self.gguf_writer.add_layer_norm_eps(1e-5)
+        if self.hparams["attn_config"]["clip_qkv"] is not None:
+            self.gguf_writer.add_clamp_kqv(self.hparams["attn_config"]["clip_qkv"])
+        if self.hparams["attn_config"]["alibi"]:
+            self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"])
+        else:
+            self.gguf_writer.add_max_alibi_bias(0.0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if "scales" in name:
+            new_name = self.map_tensor_name(name, try_suffixes=(".weight", ".bias", ".scales"))
+            new_name = new_name.replace("scales", "act.scales")
+        else:
+            new_name = self.map_tensor_name(name, try_suffixes=(".weight", ".bias"))
+
+        return [(new_name, data_torch)]
+
+
+@Model.register("OrionForCausalLM")
+class OrionModel(Model):
+    model_arch = gguf.MODEL_ARCH.ORION
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            raise ValueError("gguf: can not find ctx length parameter.")
+
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        # note: config provides rms norm but it is actually layer norm
+        # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
+        self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
+
+
+@Model.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
+class BaichuanModel(Model):
+    model_arch = gguf.MODEL_ARCH.BAICHUAN
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            raise ValueError("gguf: can not find ctx length parameter.")
+
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        if bid is not None and name == f"model.layers.{bid}.self_attn.W_pack.weight":
+            logger.info(f"Unpacking and permuting layer {bid}")
+            tensors = [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid),
+                    self._reverse_hf_permute_part(data_torch, 0, head_count, head_count)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid),
+                    self._reverse_hf_permute_part(data_torch, 1, head_count, head_count_kv)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid),
+                    self._reverse_hf_part(data_torch, 2)),
+            ]
+        else:
+            tensors = [(self.map_tensor_name(name), data_torch)]
+
+        return tensors
+
+    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
+        if n_kv_head is not None and n_head != n_kv_head:
+            n_head //= n_kv_head
+
+        return (
+            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape)
+        )
+
+    def _reverse_hf_permute_part(
+        self, weights: Tensor, n_part: int, n_head: int, n_head_kv: int | None = None,
+    ) -> Tensor:
+        r = weights.shape[0] // 3
+        return self._reverse_hf_permute(weights[r * n_part:r * n_part + r, ...], n_head, n_head_kv)
+
+    def _reverse_hf_part(self, weights: Tensor, n_part: int) -> Tensor:
+        r = weights.shape[0] // 3
+        return weights[r * n_part:r * n_part + r, ...]
+
+
+@Model.register("XverseForCausalLM")
+class XverseModel(Model):
+    model_arch = gguf.MODEL_ARCH.XVERSE
+
+    def set_vocab(self):
+        assert (self.dir_model / "tokenizer.json").is_file()
+        dir_model = self.dir_model
+        hparams = self.hparams
+
+        tokens: list[bytes] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model)
+        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+        # Since we are checking the maximum index, we need to ensure it's strictly less than vocab_size,
+        # because vocab_size is the count of items, and indexes start at 0.
+        max_vocab_index = max(tokenizer.get_vocab().values())
+        if max_vocab_index >= vocab_size:
+            raise ValueError("Vocabulary size exceeds expected maximum size.")
+
+        reverse_vocab: dict[int, str] = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        for token_id in range(vocab_size):
+            token_text = reverse_vocab[token_id].encode('utf-8')
+            # replace "\x00" to string with length > 0
+            if token_text == b"\x00":
+                toktype = gguf.TokenType.BYTE  # special
+                token_text = f"<{token_text}>".encode('utf-8')
+            elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
+                toktype = gguf.TokenType.BYTE  # special
+            elif reverse_vocab[token_id] in added_vocab:
+                if tokenizer.added_tokens_decoder[token_id].special:
+                    toktype = gguf.TokenType.CONTROL
+                else:
+                    toktype = gguf.TokenType.USER_DEFINED
+            else:
+                toktype = gguf.TokenType.NORMAL
+
+            tokens.append(token_text)
+            toktypes.append(toktype)
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            raise ValueError("gguf: can not find ctx length parameter.")
+
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        # HF models permute some of the tensors, so we need to undo that
+        if name.endswith("q_proj.weight"):
+            data_torch = self._reverse_hf_permute(data_torch, head_count, head_count)
+        if name.endswith("k_proj.weight"):
+            data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
+        if n_kv_head is not None and n_head != n_kv_head:
+            n_head //= n_kv_head
+
+        return (
+            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape)
+        )
+
+
+@Model.register("FalconForCausalLM", "RWForCausalLM")
+class FalconModel(Model):
+    model_arch = gguf.MODEL_ARCH.FALCON
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams.get("num_hidden_layers")
+        if block_count is None:
+            block_count = self.hparams["n_layer"]  # old name
+
+        n_head = self.hparams.get("num_attention_heads")
+        if n_head is None:
+            n_head = self.hparams["n_head"]  # old name
+
+        n_head_kv = self.hparams.get("num_kv_heads")
+        if n_head_kv is None:
+            n_head_kv = self.hparams.get("n_head_kv", 1)  # old name
+
+        self.gguf_writer.add_context_length(2048)  # not in config.json
+        self.gguf_writer.add_tensor_data_layout("jploski")  # qkv tensor transform
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head_kv)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # QKV tensor transform
+        # The original query_key_value tensor contains n_head_kv "kv groups",
+        # each consisting of n_head/n_head_kv query weights followed by one key
+        # and one value weight (shared by all query heads in the kv group).
+        # This layout makes it a big pain to work with in GGML.
+        # So we rearrange them here,, so that we have n_head query weights
+        # followed by n_head_kv key weights followed by n_head_kv value weights,
+        # in contiguous fashion.
+        # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
+
+        if "query_key_value" in name:
+            n_head = self.find_hparam(["num_attention_heads", "n_head"])
+            n_head_kv = self.find_hparam(["num_kv_heads", "n_head_kv"], optional=True) or 1
+            head_dim = self.hparams["hidden_size"] // n_head
+
+            qkv = data_torch.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
+            q = qkv[:, :-2].reshape(n_head * head_dim, head_dim * n_head)
+            k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
+            v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
+            data_torch = torch.cat((q, k, v)).reshape_as(data_torch)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("GPTBigCodeForCausalLM")
+class StarCoderModel(Model):
+    model_arch = gguf.MODEL_ARCH.STARCODER
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["n_layer"]
+
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_head_count_kv(1)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+
+@Model.register("GPTRefactForCausalLM")
+class RefactModel(Model):
+    model_arch = gguf.MODEL_ARCH.REFACT
+
+    def set_vocab(self):
+        super().set_vocab()
+
+        # TODO: how to determine special FIM tokens automatically?
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False,
+                                          special_token_types = ['prefix', 'suffix', 'middle', 'eot'])
+        special_vocab._set_special_token("prefix", 1)
+        special_vocab._set_special_token("suffix", 3)
+        special_vocab._set_special_token("middle", 2)
+        special_vocab.chat_template = None  # do not add it twice
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        hidden_dim = self.hparams["n_embd"]
+        inner_dim = 4 * hidden_dim
+        hidden_dim = int(2 * inner_dim / 3)
+        multiple_of = 256
+        ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        block_count = self.hparams["n_layer"]
+
+        # refact uses Alibi. So this is from config.json which might be used by training.
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+
+        self.gguf_writer.add_feed_forward_length(ff_dim)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_head_count_kv(1)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        hidden_dim = self.hparams["n_embd"]
+        inner_dim = 4 * hidden_dim
+        hidden_dim = int(2 * inner_dim / 3)
+        multiple_of = 256
+        ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+        n_head = self.hparams["n_head"]
+        n_head_kv = 1
+        head_dim = self.hparams["n_embd"] // n_head
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        if bid is not None:
+            if name == f"transformer.h.{bid}.attn.kv.weight":
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), data_torch[:n_head_kv * head_dim]))
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), data_torch[n_head_kv * head_dim:]))
+            elif name == f"transformer.h.{bid}.attn.q.weight":
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), data_torch))
+            elif name == f"transformer.h.{bid}.mlp.gate_up_proj.weight":
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), data_torch[:ff_dim]))
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), data_torch[ff_dim:]))
+
+        if len(tensors) == 0:
+            tensors.append((self.map_tensor_name(name), data_torch))
+
+        return tensors
+
+
+@Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
+class StableLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.STABLELM
+
+    def set_vocab(self):
+        if (self.dir_model / "tokenizer.json").is_file():
+            self._set_vocab_gpt2()
+        else:
+            # StableLM 2 1.6B used to have a vocab in a similar format to Qwen's vocab
+            self._set_vocab_qwen()
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
+        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
+        self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
+        self.gguf_writer.add_file_type(self.ftype)
+
+    _q_norms: list[dict[str, Tensor]] | None = None
+    _k_norms: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams["num_key_value_heads"]
+
+        if name.find("q_layernorm.norms") != -1:
+            assert bid is not None
+
+            if self._q_norms is None:
+                self._q_norms = [{} for _ in range(self.block_count)]
+
+            self._q_norms[bid][name] = data_torch
+
+            if len(self._q_norms[bid]) >= n_head:
+                return self._stack_qk_norm(bid, n_head, self._q_norms[bid], "q_layernorm")
+            else:
+                return []
+
+        if name.find("k_layernorm.norms") != -1:
+            assert bid is not None
+
+            if self._k_norms is None:
+                self._k_norms = [{} for _ in range(self.block_count)]
+
+            self._k_norms[bid][name] = data_torch
+
+            if len(self._k_norms[bid]) >= n_kv_head:
+                return self._stack_qk_norm(bid, n_kv_head, self._k_norms[bid], "k_layernorm")
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def _stack_qk_norm(self, bid: int, n_head: int, norms: dict[str, Tensor], layer_name: str = "q_layernorm"):
+        datas: list[Tensor] = []
+        # extract the norms in order
+        for xid in range(n_head):
+            ename = f"model.layers.{bid}.self_attn.{layer_name}.norms.{xid}.weight"
+            datas.append(norms[ename])
+            del norms[ename]
+        data_torch = torch.stack(datas, dim=0)
+
+        merged_name = f"model.layers.{bid}.self_attn.{layer_name}.weight"
+        new_name = self.map_tensor_name(merged_name)
+
+        return [(new_name, data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._q_norms is not None or self._k_norms is not None:
+            # flatten two `list[dict[str, Tensor]]` into a single `list[str]`
+            norms = (
+                [k for d in self._q_norms for k in d.keys()] if self._q_norms is not None else []
+            ) + (
+                [k for d in self._k_norms for k in d.keys()] if self._k_norms is not None else []
+            )
+            if len(norms) > 0:
+                raise ValueError(f"Unprocessed norms: {norms}")
+
+
+@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
+class LlamaModel(Model):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            try:
+                self._set_vocab_llama_hf()
+            except (FileNotFoundError, TypeError):
+                # Llama 3
+                self._set_vocab_gpt2()
+
+        # Apply to CodeLlama only (and ignore for Llama 3 with a vocab size of 128256)
+        if self.hparams.get("vocab_size", 32000) == 32016:
+            special_vocab = gguf.SpecialVocab(
+                self.dir_model, load_merges=False,
+                special_token_types = ['prefix', 'suffix', 'middle', 'eot']
+            )
+            special_vocab._set_special_token("prefix", 32007)
+            special_vocab._set_special_token("suffix", 32008)
+            special_vocab._set_special_token("middle", 32009)
+            special_vocab._set_special_token("eot",    32010)
+            special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
+            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+        # Apply to granite small models only
+        if self.hparams.get("vocab_size", 32000) == 49152:
+            self.gguf_writer.add_add_bos_token(False)
+
+    @staticmethod
+    def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
+        if n_head_kv is not None and n_head != n_head_kv:
+            n_head = n_head_kv
+        return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        if name.endswith(("q_proj.weight", "q_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith(("k_proj.weight", "k_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.hparams["num_local_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for wid in ["w1", "w2", "w3"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
+            if rope_scaling.get("rope_type", '').lower() == "llama3":
+                base = self.hparams.get("rope_theta", 10000.0)
+                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+
+                factor = rope_scaling.get("factor", 8.0)
+                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+
+                low_freq_wavelen = old_context_len / low_freq_factor
+                high_freq_wavelen = old_context_len / high_freq_factor
+                assert low_freq_wavelen != high_freq_wavelen
+
+                rope_factors = []
+                for freq in freqs:
+                    wavelen = 2 * math.pi / freq
+                    if wavelen < high_freq_wavelen:
+                        rope_factors.append(1)
+                    elif wavelen > low_freq_wavelen:
+                        rope_factors.append(factor)
+                    else:
+                        smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+                        rope_factors.append(1 / ((1 - smooth) / factor + smooth))
+
+                self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@Model.register("BitnetForCausalLM")
+class BitnetModel(Model):
+    model_arch = gguf.MODEL_ARCH.BITNET
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+        self.gguf_writer.add_rope_scaling_factor(1.0)
+
+    def weight_quant(self, weight):
+        dtype = weight.dtype
+        weight = weight.float()
+        s = 1 / weight.abs().mean().clamp(min=1e-5)
+        weight = (weight * s).round().clamp(-1, 1) / s
+        scale = weight.abs().max().unsqueeze(0)
+        weight = torch.where(weight.abs().less(1e-6), 0, weight).type(dtype)
+        weight = torch.sign(weight).type(dtype)
+        return weight.type(dtype), scale.type(torch.float32)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        new_name = self.map_tensor_name(name)
+
+        if any(self.match_model_tensor_name(new_name, key, bid) for key in [
+            gguf.MODEL_TENSOR.ATTN_Q,
+            gguf.MODEL_TENSOR.ATTN_K,
+            gguf.MODEL_TENSOR.ATTN_V,
+            gguf.MODEL_TENSOR.ATTN_OUT,
+            gguf.MODEL_TENSOR.FFN_UP,
+            gguf.MODEL_TENSOR.FFN_DOWN,
+            gguf.MODEL_TENSOR.FFN_GATE,
+        ]):
+            # transform weight into 1/0/-1 (in fp32)
+            weight_torch, scale_torch = self.weight_quant(data_torch)
+            yield (new_name, weight_torch)
+            yield (new_name.removesuffix(".weight") + ".scale", scale_torch)
+        else:
+            yield (new_name, data_torch)
+
+
+@Model.register("GrokForCausalLM")
+class GrokModel(Model):
+    model_arch = gguf.MODEL_ARCH.GROK
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find(".moe.") != -1:
+            n_experts = self.hparams["num_local_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for wid in ["linear", "linear_1", "linear_v"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"transformer.decoder_layer.{bid}.moe.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("DbrxForCausalLM")
+class DbrxModel(Model):
+    model_arch = gguf.MODEL_ARCH.DBRX
+
+    def set_gguf_parameters(self):
+        ffn_config = self.hparams["ffn_config"]
+        attn_config = self.hparams["attn_config"]
+        self.gguf_writer.add_block_count(self.hparams["n_layers"])
+
+        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(ffn_config["ffn_hidden_size"])
+
+        self.gguf_writer.add_head_count(self.hparams["n_heads"])
+        self.gguf_writer.add_head_count_kv(attn_config["kv_n_heads"])
+
+        self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
+
+        self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
+
+        self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"])
+        self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"])
+
+        self.gguf_writer.add_layer_norm_eps(1e-5)
+
+        self.gguf_writer.add_file_type(self.ftype)
+        logger.info(f"gguf: file type = {self.ftype}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_expert = self.hparams["ffn_config"]["moe_num_experts"]
+        n_ff = self.hparams["ffn_config"]["ffn_hidden_size"]
+        n_embd = self.hparams["d_model"]
+
+        # Specific behavior for experts tensors: suffix .weight, view as 3D and transpose
+        # original implementation expects (n_expert, n_ff, n_embd) for all experts weights
+        # But llama.cpp moe graph works differently
+        # AND the dimensions in ggml are typically in the reverse order of the pytorch dimensions
+        # so (n_expert, n_ff, n_embd) in pytorch is {n_embd, n_ff, n_expert} in ggml_tensor
+        exp_tensor_names = {"ffn.experts.mlp.w1": None,       # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff,   n_expert}
+                            "ffn.experts.mlp.w2": (0, 2, 1),  # LLM_TENSOR_FFN_DOWN_EXPS ggml_tensor->ne{n_ff,   n_embd, n_expert}
+                            "ffn.experts.mlp.v1": None}       # LLM_TENSOR_FFN_UP_EXPS   ggml_tensor->ne{n_embd, n_ff,   n_expert}
+        experts = False
+
+        for exp_tensor_name in exp_tensor_names.keys():
+            if name.find(exp_tensor_name) != -1 and name.find(".weight") == -1:
+                experts = True
+                data_torch = data_torch.view(n_expert, n_ff, n_embd)
+                if (permute_tensor := exp_tensor_names[exp_tensor_name]) is not None:
+                    data_torch = data_torch.permute(*permute_tensor)
+                break
+
+        # map tensor names
+        # In MoE models the ffn tensors are typically most of the model weights,
+        # and need to be quantizable. Quantize expects tensor names to be suffixed by .weight.
+        # Every other model has the weight names ending in .weight,
+        # let's assume that is the convention which is not the case for dbrx:
+        # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
+        new_name = self.map_tensor_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
+
+        return [(new_name, data_torch)]
+
+    def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+        del name, new_name, bid  # unused
+
+        return n_dims > 1
+
+
+@Model.register("MiniCPMForCausalLM")
+class MiniCPMModel(Model):
+    model_arch = gguf.MODEL_ARCH.MINICPM
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def set_vocab(self):
+        self._set_vocab_llama_hf()
+
+    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
+        if n_kv_head is not None and n_head != n_kv_head:
+            n_head //= n_kv_head
+
+        return (
+            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape)
+        )
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        # HF models permute some of the tensors, so we need to undo that
+        if name.endswith(("q_proj.weight")):
+            data_torch = self._reverse_hf_permute(data_torch, n_head, n_head)
+        if name.endswith(("k_proj.weight")):
+            data_torch = self._reverse_hf_permute(data_torch, n_head, n_kv_head)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("QWenLMHeadModel")
+class QwenModel(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN
+
+    @staticmethod
+    def token_bytes_to_string(b):
+        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
+        byte_encoder = bytes_to_unicode()
+        return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
+
+    @staticmethod
+    def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: int | None = None) -> list[bytes]:
+        parts = [bytes([b]) for b in token]
+        while True:
+            min_idx = None
+            min_rank = None
+            for i, pair in enumerate(zip(parts[:-1], parts[1:])):
+                rank = mergeable_ranks.get(pair[0] + pair[1])
+                if rank is not None and (min_rank is None or rank < min_rank):
+                    min_idx = i
+                    min_rank = rank
+            if min_rank is None or (max_rank is not None and min_rank >= max_rank):
+                break
+            assert min_idx is not None
+            parts = parts[:min_idx] + [parts[min_idx] + parts[min_idx + 1]] + parts[min_idx + 2:]
+        return parts
+
+    def set_vocab(self):
+        self._set_vocab_qwen()
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+
+@Model.register("Qwen2ForCausalLM")
+class Qwen2Model(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN2
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+
+@Model.register("Qwen2MoeForCausalLM")
+class Qwen2MoeModel(Model):
+    model_arch = gguf.MODEL_ARCH.QWEN2MOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (n_experts := self.hparams.get("num_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
+        if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None:
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size)
+            logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}")
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@Model.register("GPT2LMHeadModel")
+class GPT2Model(Model):
+    model_arch = gguf.MODEL_ARCH.GPT2
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_context_length(self.hparams["n_ctx"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        # we don't need these
+        if name.endswith((".attn.bias", ".attn.masked_bias")):
+            return tensors
+
+        if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):
+            data_torch = data_torch.transpose(1, 0)
+
+        new_name = self.map_tensor_name(name)
+
+        tensors.append((new_name, data_torch))
+
+        # note: GPT2 output is tied to (same as) wte in original model
+        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
+            tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
+
+        return tensors
+
+
+@Model.register("PhiForCausalLM")
+class Phi2Model(Model):
+    model_arch = gguf.MODEL_ARCH.PHI2
+
+    def set_gguf_parameters(self):
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
+
+        rot_pct = self.find_hparam(["partial_rotary_factor"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+
+        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
+
+        self.gguf_writer.add_embedding_length(n_embd)
+        self.gguf_writer.add_feed_forward_length(4 * n_embd)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head)
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_add_bos_token(False)
+
+
+@Model.register("Phi3ForCausalLM")
+class Phi3MiniModel(Model):
+    model_arch = gguf.MODEL_ARCH.PHI3
+
+    def set_vocab(self):
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        if not tokenizer_path.is_file():
+            raise ValueError(f'Error: Missing {tokenizer_path}')
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
+                for token_id, foken_data in added_tokens_decoder.items():
+                    token_id = int(token_id)
+                    token = foken_data["content"].encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
+                        if tokens[token_id] != token:
+                            logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}')
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
+        tokenizer_file = self.dir_model / 'tokenizer.json'
+        if tokenizer_file.is_file():
+            with open(tokenizer_file, "r", encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+                added_tokens = tokenizer_json.get("added_tokens", [])
+                for foken_data in added_tokens:
+                    token_id = int(foken_data["id"])
+                    token = foken_data["content"].encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
+                        if tokens[token_id] != token:
+                            logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}')
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
+
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
+        rms_eps = self.find_hparam(["rms_norm_eps"])
+        max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
+        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
+        rope_dims = n_embd // n_head
+
+        self.gguf_writer.add_context_length(max_pos_embds)
+        self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
+        self.gguf_writer.add_embedding_length(n_embd)
+        self.gguf_writer.add_feed_forward_length(self.find_hparam(["intermediate_size"]))
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
+        self.gguf_writer.add_rope_dimension_count(rope_dims)
+        self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
+
+        # write rope scaling for long context (128k) model
+        rope_scaling = self.find_hparam(['rope_scaling'], True)
+        if rope_scaling is None:
+            return
+
+        scale = max_pos_embds / orig_max_pos_embds
+
+        rope_scaling_type = rope_scaling.get('type', '').lower()
+        if len(rope_scaling_type) == 0:
+            raise KeyError('Missing the required key rope_scaling.type')
+
+        if rope_scaling_type == 'su' or rope_scaling_type == 'longrope':
+            attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0
+        elif rope_scaling_type == 'yarn':
+            attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0
+        else:
+            raise NotImplementedError(f'The rope scaling type {rope_scaling_type} is not supported yet')
+
+        self.gguf_writer.add_rope_scaling_attn_factors(attn_factor)
+
+        long_factors = rope_scaling.get('long_factor', None)
+        short_factors = rope_scaling.get('short_factor', None)
+
+        if long_factors is None or short_factors is None:
+            raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
+
+        if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
+            raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
+
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+
+
+@Model.register("PlamoForCausalLM")
+class PlamoModel(Model):
+    model_arch = gguf.MODEL_ARCH.PLAMO
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_context_length(4096)  # not in config.json
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(5)  # hparams["num_key_value_heads"]) is wrong
+        self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def shuffle_attn_q_weight(self, data_torch):
+        assert data_torch.size() == (5120, 5120)
+        data_torch = data_torch.reshape(8, 5, 128, 5120)
+        data_torch = torch.permute(data_torch, (1, 0, 2, 3))
+        data_torch = torch.reshape(data_torch, (5120, 5120))
+        return data_torch
+
+    def shuffle_attn_output_weight(self, data_torch):
+        assert data_torch.size() == (5120, 5120)
+        data_torch = data_torch.reshape(5120, 8, 5, 128)
+        data_torch = torch.permute(data_torch, (0, 2, 1, 3))
+        data_torch = torch.reshape(data_torch, (5120, 5120))
+        return data_torch
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        new_name = self.map_tensor_name(name)
+
+        # shuffle for broadcasting of gqa in ggml_mul_mat
+        if new_name.endswith("attn_q.weight"):
+            data_torch = self.shuffle_attn_q_weight(data_torch)
+        elif new_name.endswith("attn_output.weight"):
+            data_torch = self.shuffle_attn_output_weight(data_torch)
+
+        return [(new_name, data_torch)]
+
+
+@Model.register("CodeShellForCausalLM")
+class CodeShellModel(Model):
+    model_arch = gguf.MODEL_ARCH.CODESHELL
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["n_layer"]
+
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_head_count_kv(self.hparams["num_query_groups"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_rope_freq_base(10000.0)
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+        self.gguf_writer.add_rope_scaling_factor(1.0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        new_name = self.map_tensor_name(name)
+
+        tensors: list[tuple[str, Tensor]] = [(new_name, data_torch)]
+
+        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
+            assert self.tensor_names is not None
+
+            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
+                # copy tok_embd.weight to output.weight
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
+
+        return tensors
+
+
+@Model.register("InternLM2ForCausalLM")
+class InternLM2Model(Model):
+    model_arch = gguf.MODEL_ARCH.INTERNLM2
+
+    def set_vocab(self):
+        # (TODO): Is there a better way?
+        # Copy from _set_vocab_sentencepiece, The only difference is that we will treat the character
+        # \x00 specially and convert it into an emoji character to prevent it from being mistakenly
+        # recognized as an empty string in C++.
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        tokens: list[bytes] = []
+        scores: list[float] = []
+        toktypes: list[int] = []
+
+        if not tokenizer_path.is_file():
+            logger.error(f'Error: Missing {tokenizer_path}')
+            sys.exit(1)
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        for token_id in range(vocab_size):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+            if text == b"\x00":
+                # (TODO): fixme
+                # Hack here and replace the \x00 characters.
+                logger.warning(f"InternLM2 convert token '{text}' to '🐉'!")
+                text = "🐉".encode("utf-8")
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+            # take care of ununsed raw token
+            if piece.startswith('[UNUSED'):
+                toktype = SentencePieceTokenTypes.UNUSED
+
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+
+                for key in added_tokens_json:
+                    tokens.append(key.encode("utf-8"))
+                    scores.append(-1000.0)
+                    toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+
+        chat_eos_token = '<|im_end|>'
+        chat_eos_token_id = None
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
+                for token_id, foken_data in added_tokens_decoder.items():
+                    token_id = int(token_id)
+                    token = foken_data["content"]
+                    if token == chat_eos_token:
+                        chat_eos_token_id = token_id
+                    token = token.encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
+                        if tokens[token_id] != token:
+                            logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}')
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
+        tokenizer_file = self.dir_model / 'tokenizer.json'
+        if tokenizer_file.is_file():
+            with open(tokenizer_file, "r", encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+                added_tokens = tokenizer_json.get("added_tokens", [])
+                for foken_data in added_tokens:
+                    token_id = int(foken_data["id"])
+                    token = foken_data["content"]
+                    if token == chat_eos_token:
+                        chat_eos_token_id = token_id
+                    token = token.encode("utf-8")
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
+                        if tokens[token_id] != token:
+                            logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}')
+                    tokens[token_id] = token
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+                    if foken_data.get("special"):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        old_eos = special_vocab.special_token_ids["eos"]
+        if chat_eos_token_id is not None:
+            # For the chat model, we replace the eos with '<|im_end|>'.
+            # TODO: this is a hack, should be fixed
+            #       https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048
+            special_vocab.special_token_ids["eos"] = chat_eos_token_id
+            logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}"
+                           " in chat mode so that the conversation can end normally.")
+
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
+        self.gguf_writer.add_file_type(self.ftype)
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        num_heads = self.hparams["num_attention_heads"]
+        num_kv_heads = self.hparams["num_key_value_heads"]
+        n_embd = self.hparams["hidden_size"]
+        q_per_kv = num_heads // num_kv_heads
+        head_dim = n_embd // num_heads
+        num_groups = num_heads // q_per_kv
+
+        if bid is not None and f"model.layers.{bid}.attention.wqkv" in name:
+            qkv = data_torch
+
+            qkv = qkv.reshape((num_groups, q_per_kv + 2, head_dim, n_embd))
+            q, k, v = qkv[:, : q_per_kv], qkv[:, -2], qkv[:, -1]
+
+            # The model weights of q and k equire additional reshape.
+            q = LlamaModel.permute(q.reshape((-1, q.shape[-1])), num_heads, num_heads)
+            k = LlamaModel.permute(k.reshape((-1, k.shape[-1])), num_heads, num_kv_heads)
+            v = v.reshape((-1, v.shape[-1]))
+
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v),
+            ]
+        else:
+            return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("BertModel", "CamembertModel")
+class BertModel(Model):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.vocab_size = None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_causal_attention(False)
+
+        # get pooling path
+        pooling_path = None
+        module_path = self.dir_model / "modules.json"
+        if module_path.is_file():
+            with open(module_path, encoding="utf-8") as f:
+                modules = json.load(f)
+            for mod in modules:
+                if mod["type"] == "sentence_transformers.models.Pooling":
+                    pooling_path = mod["path"]
+                    break
+
+        # get pooling type
+        if pooling_path is not None:
+            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
+                pooling = json.load(f)
+            if pooling["pooling_mode_mean_tokens"]:
+                pooling_type = gguf.PoolingType.MEAN
+            elif pooling["pooling_mode_cls_token"]:
+                pooling_type = gguf.PoolingType.CLS
+            else:
+                raise NotImplementedError("Only MEAN and CLS pooling types supported")
+            self.gguf_writer.add_pooling_type(pooling_type)
+
+    def set_vocab(self):
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.vocab_size = len(tokens)
+
+        # we need this to validate the size of the token_type embeddings
+        # though currently we are passing all zeros to the token_type embeddings
+        self.gguf_writer.add_token_type_count(2)  # "Sequence A" or "Sequence B"
+
+        # convert to phantom space vocab
+        def phantom(tok):
+            if tok.startswith("[") and tok.endswith("]"):
+                return tok
+            if tok.startswith("##"):
+                return tok[2:]
+            return "\u2581" + tok
+        tokens = list(map(phantom, tokens))
+
+        # add vocab to gguf
+        self.gguf_writer.add_tokenizer_model("bert")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # handle special tokens
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # we are only using BERT for embeddings so we don't need the pooling layer
+        if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias"):
+            return [] # we don't need these
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("NomicBertModel")
+class NomicBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # the HF config claims n_ctx=8192, but it uses RoPE scaling
+        self.hparams["n_ctx"] = 2048
+
+        # SwigLU activation
+        assert self.hparams["activation_function"] == "swiglu"
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors
+        assert self.hparams["qkv_proj_bias"] is False
+        assert self.hparams["mlp_fc1_bias"] is False
+        assert self.hparams["mlp_fc2_bias"] is False
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+
+
+@Model.register("XLMRobertaModel")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        # realign tokens (see HF tokenizer code)
+        tokens = [b'', b'', b'', b''] + tokens[3:-1]
+        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
+        toktypes = [
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.UNKNOWN,
+        ] + toktypes[3:-1]
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@Model.register("GemmaForCausalLM")
+class GemmaModel(Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+        # TODO: these special tokens should be exported only for the CodeGemma family
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False,
+                                          special_token_types = ['prefix', 'suffix', 'middle', 'fsep', 'eot'])
+        special_vocab._set_special_token("prefix", 67)
+        special_vocab._set_special_token("suffix", 69)
+        special_vocab._set_special_token("middle", 68)
+        special_vocab._set_special_token("fsep",   70)
+        special_vocab._set_special_token("eot",    107)
+        special_vocab.chat_template = None  # do not add it twice
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_space_prefix(False)
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_key_length(hparams["head_dim"])
+        self.gguf_writer.add_value_length(hparams["head_dim"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # lm_head is not used in llama.cpp, while autoawq will include this tensor in model
+        # To prevent errors, skip loading lm_head.weight.
+        if name == "lm_head.weight":
+            logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
+            return []
+
+        # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("Gemma2ForCausalLM")
+class Gemma2Model(Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA2
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+        self.gguf_writer.add_add_space_prefix(False)
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_key_length(hparams["head_dim"])
+        self.gguf_writer.add_value_length(hparams["head_dim"])
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_attn_logit_softcapping(
+            self.hparams["attn_logit_softcapping"]
+        )
+        self.gguf_writer.add_final_logit_softcapping(
+            self.hparams["final_logit_softcapping"]
+        )
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # lm_head is not used in llama.cpp, while autoawq will include this tensor in model
+        # To prevent errors, skip loading lm_head.weight.
+        if name == "lm_head.weight":
+            logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
+            return []
+
+        # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("Starcoder2ForCausalLM")
+class StarCoder2Model(Model):
+    model_arch = gguf.MODEL_ARCH.STARCODER2
+
+
+@Model.register("MambaForCausalLM", "MambaLMHeadModel")
+class MambaModel(Model):
+    model_arch = gguf.MODEL_ARCH.MAMBA
+
+    def set_vocab(self):
+        vocab_size = self.hparams["vocab_size"]
+        # Round vocab size to next multiple of 8
+        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8)
+        # pad using ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
+        self.hparams["vocab_size"] = vocab_size
+
+        if (self.dir_model / "tokenizer.json").is_file():
+            self._set_vocab_gpt2()
+        elif (self.dir_model / "tokenizer.model").is_file():
+            self._set_vocab_sentencepiece()
+        else:
+            # Use the GPT-NeoX tokenizer when no tokenizer files are present
+            self._set_vocab_builtin("gpt-neox", vocab_size)
+
+    def set_gguf_parameters(self):
+        d_model = self.find_hparam(["hidden_size",       "d_model"])
+        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
+        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 16
+        # ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
+        dt_rank      = self.find_hparam(["time_step_rank",     "dt_rank"],      optional=True) or -(d_model // -16)
+        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
+
+        # Fail early for models which don't have a block expansion factor of 2
+        assert d_inner == 2 * d_model
+
+        self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
+        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
+        self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_ssm_conv_kernel(d_conv)
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(d_state)
+        self.gguf_writer.add_ssm_time_step_rank(dt_rank)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_file_type(self.ftype)
+
+    _tok_embd = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
+
+        new_name = self.map_tensor_name(name)
+
+        if name.endswith(".A_log"):
+            logger.debug("A_log --> A ==> " + new_name)
+            data_torch = -torch.exp(data_torch)
+
+        # assuming token_embd.weight is seen before output.weight
+        if self._tok_embd is not None and new_name == output_name:
+            if torch.equal(self._tok_embd, data_torch):
+                logger.debug(f"{output_name} is equivalent to {tok_embd_name}, omitting")
+                return []
+        elif new_name == tok_embd_name:
+            self._tok_embd = data_torch
+
+        return [(new_name, data_torch)]
+
+    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
+        del n_dims  # unused
+
+        return bid is not None and new_name in (
+            self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [
+                gguf.MODEL_TENSOR.SSM_CONV1D,
+                gguf.MODEL_TENSOR.SSM_X,
+                gguf.MODEL_TENSOR.SSM_DT,
+                gguf.MODEL_TENSOR.SSM_A,
+                gguf.MODEL_TENSOR.SSM_D,
+            ]
+        )
+
+
+@Model.register("CohereForCausalLM")
+class CommandR2Model(Model):
+    model_arch = gguf.MODEL_ARCH.COMMAND_R
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # max_position_embeddings = 8192 in config.json but model was actually
+        # trained on 128k context length
+        # aya-23 models don't have model_max_length specified
+        self.hparams["max_position_embeddings"] = self.find_hparam(["model_max_length", "max_position_embeddings"])
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+
+@Model.register("OlmoForCausalLM")
+@Model.register("OLMoForCausalLM")
+class OlmoModel(Model):
+    model_arch = gguf.MODEL_ARCH.OLMO
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_layer_norm_eps(1e-5)
+        clip_qkv = self.hparams.get("clip_qkv")
+        if clip_qkv is not None:
+            self.gguf_writer.add_clamp_kqv(clip_qkv)
+
+    # Same as super class, but permuting q_proj, k_proj
+    # Copied from: LlamaModel
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        if name.endswith("q_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith("k_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("JinaBertModel", "JinaBertForMaskedLM")
+class JinaBertV2Model(BertModel):
+    model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.intermediate_size = self.hparams["intermediate_size"]
+
+    def get_tensors(self):
+        for name, data in super().get_tensors():
+            if 'gated_layer' in name:
+                d1 = data[:self.intermediate_size, :]
+                name1 = name.replace('gated_layers', 'gated_layers_w')
+                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
+                d2 = data[self.intermediate_size:, :]
+                name2 = name.replace('gated_layers', 'gated_layers_v')
+                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
+                yield name1, d1
+                yield name2, d2
+                continue
+
+            yield name, data
+
+    def set_vocab(self):
+        tokenizer_class = 'BertTokenizer'
+        with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
+            tokenizer_class = json.load(f)['tokenizer_class']
+
+        if tokenizer_class == 'BertTokenizer':
+            super().set_vocab()
+        elif tokenizer_class == 'RobertaTokenizer':
+            self._set_vocab_gpt2()
+            self.gguf_writer.add_token_type_count(2)
+        else:
+            raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+
+@Model.register("OpenELMForCausalLM")
+class OpenELMModel(Model):
+    model_arch = gguf.MODEL_ARCH.OPENELM
+
+    @staticmethod
+    def _make_divisible(v: float | int, divisor: int) -> int:
+        # ref: https://huggingface.co/apple/OpenELM-270M-Instruct/blob/eb111ff2e6724348e5b905984063d4064d4bc579/configuration_openelm.py#L34-L38
+        new_v = max(divisor, int(v + divisor / 2) // divisor * divisor)
+        # Make sure that round down does not go down by more than 10%.
+        if new_v < 0.9 * v:
+            new_v += divisor
+        return new_v
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        ffn_multipliers: list[float] = self.hparams["ffn_multipliers"]
+        ffn_dim_divisor: int = self.hparams["ffn_dim_divisor"]
+        self._n_embd: int = self.hparams["model_dim"]
+        self._num_kv_heads: list[int] = self.hparams["num_kv_heads"]
+        self._num_query_heads: list[int] = self.hparams["num_query_heads"]
+        self._ffn_dims: list[int] = [
+            OpenELMModel._make_divisible(multiplier * self._n_embd, ffn_dim_divisor)
+            for multiplier in ffn_multipliers
+        ]
+        assert isinstance(self._num_kv_heads, list) and isinstance(self._num_kv_heads[0], int)
+        assert isinstance(self._num_query_heads, list) and isinstance(self._num_query_heads[0], int)
+
+    # Uses the tokenizer from meta-llama/Llama-2-7b-hf
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_builtin("llama-spm", self.hparams["vocab_size"])
+
+    def set_gguf_parameters(self):
+        n_embd = self._n_embd
+        head_dim = self.hparams["head_dim"]
+        rot_pct = 1.0
+        assert self.block_count == len(self._num_kv_heads)
+        assert self.block_count == len(self._num_query_heads)
+        assert self.block_count == len(self._ffn_dims)
+
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_context_length(self.hparams["max_context_length"])
+        self.gguf_writer.add_embedding_length(n_embd)
+        self.gguf_writer.add_feed_forward_length(self._ffn_dims)
+        self.gguf_writer.add_head_count(self._num_query_heads)
+        self.gguf_writer.add_head_count_kv(self._num_kv_heads)
+        self.gguf_writer.add_rope_freq_base(self.hparams["rope_freq_constant"])
+        # https://huggingface.co/apple/OpenELM-270M-Instruct/blob/c401df2/modeling_openelm.py#L30
+        self.gguf_writer.add_layer_norm_rms_eps(1e-6)
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * head_dim))
+        self.gguf_writer.add_key_length(head_dim)
+        self.gguf_writer.add_value_length(head_dim)
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def find_hparam(self, keys: Iterable[str], optional: bool = False) -> Any:
+        if "n_layers" in keys:
+            return self.hparams["num_transformer_layers"]
+
+        return super().find_hparam(keys, optional)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        # split ff
+        if bid is not None and name == f"transformer.layers.{bid}.ffn.proj_1.weight":
+            ff_dim = self._ffn_dims[bid]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), data_torch[:ff_dim])
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), data_torch[ff_dim:])
+            return
+
+        yield (self.map_tensor_name(name), data_torch)
+
+
+@Model.register("ArcticForCausalLM")
+class ArcticModel(Model):
+    model_arch = gguf.MODEL_ARCH.ARCTIC
+
+    def set_vocab(self):
+        # The reason for using a custom implementation here is that the
+        # snowflake-arctic-instruct model redefined tokens 31998 and 31999 from
+        # tokenizer.model and used them as BOS and EOS instead of adding new tokens.
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        if not tokenizer_path.is_file():
+            logger.error(f'Error: Missing {tokenizer_path}')
+            sys.exit(1)
+
+        # Read the whole vocabulary from the tokenizer.model file
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        # Use the added_tokens_decoder field from tokeniser_config.json as the source
+        # of information about added/redefined tokens and modify them accordingly.
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+
+                if "added_tokens_decoder" in tokenizer_config_json:
+                    added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"]
+                    for token_id, token_json in added_tokens_decoder.items():
+                        token_id = int(token_id)
+                        if token_id >= vocab_size:
+                            logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                            continue
+
+                        token_content = token_json["content"]
+                        token_type = SentencePieceTokenTypes.USER_DEFINED
+                        token_score = -10000.0
+
+                        # Map unk_token to UNKNOWN, other special tokens to CONTROL
+                        # Set the score to 0.0 as in the original tokenizer.model
+                        if ("special" in token_json) and token_json["special"]:
+                            if token_content == tokenizer_config_json["unk_token"]:
+                                token_type = SentencePieceTokenTypes.UNKNOWN
+                            else:
+                                token_type = SentencePieceTokenTypes.CONTROL
+                            token_score = 0.0
+
+                        logger.info(f"Setting added token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})")
+                        tokens[token_id] = token_content.encode("utf-8")
+                        toktypes[token_id] = token_type
+                        scores[token_id] = token_score
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        if name.endswith("q_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith("k_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.hparams["num_local_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for wid in ["w1", "w2", "w3"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@Model.register("DeepseekV2ForCausalLM")
+class DeepseekV2Model(Model):
+    model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
+            self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
+        self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "yarn":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+                self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
+                self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["n_routed_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@Model.register("T5WithLMHeadModel")
+@Model.register("T5ForConditionalGeneration")
+@Model.register("MT5ForConditionalGeneration")
+@Model.register("UMT5ForConditionalGeneration")
+class T5Model(Model):
+    model_arch = gguf.MODEL_ARCH.T5
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.shared_token_embeddings_found = False
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        # many older models use spiece.model tokenizer model filename
+        if not tokenizer_path.is_file():
+            tokenizer_path = self.dir_model / 'spiece.model'
+
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+
+        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
+        if sentencepiece_model.trainer_spec.model_type == 2:  # BPE
+            # assure the tokenizer model file name is correct
+            assert tokenizer_path.name == 'tokenizer.model'
+            return self._set_vocab_sentencepiece()
+        else:
+            assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def set_gguf_parameters(self):
+        if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
+            logger.warning("Couldn't find context length in config.json, assuming default value of 512")
+            n_ctx = 512
+        self.gguf_writer.add_context_length(n_ctx)
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(self.hparams["num_heads"])
+        self.gguf_writer.add_key_length(self.hparams["d_kv"])
+        self.gguf_writer.add_value_length(self.hparams["d_kv"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_decoder_start_token_id(self.hparams["decoder_start_token_id"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
+        # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
+        # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
+        # and decoder and ignore the remaining ones.
+        if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]:
+            if not self.shared_token_embeddings_found:
+                name = "shared.weight"
+                self.shared_token_embeddings_found = True
+            else:
+                logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("JAISLMHeadModel")
+class JaisModel(Model):
+    model_arch = gguf.MODEL_ARCH.JAIS
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # SwigLU activation
+        assert self.hparams["activation_function"] == "swiglu"
+        # ALiBi position embedding
+        assert self.hparams["position_embedding_type"] == "alibi"
+
+        # Embeddings scale
+        self.embeddings_scale = 1.0
+        # note: For some JAIS flavors, output is tied to (same as) wte in original model
+        self.output_is_wte = False
+        if 'mup_embeddings_scale' in self.hparams:
+            self.output_is_wte = True   # Hack (?)
+            self.embeddings_scale = self.hparams['mup_embeddings_scale']
+        elif 'embeddings_scale' in self.hparams:
+            self.embeddings_scale = self.hparams['embeddings_scale']
+        else:
+            assert False
+
+        self.width_scale = 1.0
+        if 'mup_output_alpha' in self.hparams:
+            assert 'mup_width_scale' in self.hparams
+            self.width_scale = self.hparams['mup_output_alpha'] * self.hparams['mup_width_scale']
+        elif 'width_scale' in self.hparams:
+            self.width_scale = self.hparams['width_scale']
+        else:
+            assert False
+
+        self.max_alibi_bias = 8.0
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["n_inner"])
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        # we don't need these
+        if name.endswith((".attn.bias")):
+            return tensors
+
+        if name.endswith(("relative_pe.slopes")):
+            # Calculate max ALiBi bias (this is the inverse of the ALiBi calculation)
+            # Some other models has max_alibi_bias spelled out explicitly in the hyperparams,
+            # but Jais's PyTorch model simply precalculates the slope values and places them
+            # in relative_pes.slopes
+            n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"]))
+            first_val = float(data_torch[0].item())
+            self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2)
+
+            return tensors
+
+        if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_fc2.weight")):
+            data_torch = data_torch.transpose(1, 0)
+
+        new_name = self.map_tensor_name(name)
+
+        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
+            tensors.append((new_name, data_torch * self.embeddings_scale))
+            if self.output_is_wte:
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale))
+        elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            assert not self.output_is_wte
+            tensors.append((new_name, data_torch * self.width_scale))
+        else:
+            tensors.append((new_name, data_torch))
+
+        return tensors
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
+
+
+@Model.register("ChatGLMModel", "ChatGLMForConditionalGeneration")
+class ChatGLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.CHATGLM
+
+    def set_vocab_chatglm3(self):
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[bytes] = []
+        toktypes: list[int] = []
+        scores: list[float] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab()))
+        assert max(tokenizer.get_vocab().values()) < vocab_size
+        role_special_tokens = ["<|system|>", "<|user|>", "<|assistant|>", "<|observation|>"]
+        special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"] + role_special_tokens
+        for token_id in range(vocab_size):
+            piece = tokenizer._convert_id_to_token(token_id)
+            if token_id == 0:
+                piece = ""
+            elif token_id == 1:
+                piece = ""
+            elif token_id == 2:
+                piece = ""
+
+            text = piece.encode("utf-8")
+            score = 0.0
+            # Referencing the tokenizer Python implementation(https://huggingface.co/THUDM/chatglm3-6b/blob/main/tokenization_chatglm.py),
+            # it is only valid if it is less than tokenizer.tokenizer.sp_model.vocab_size()
+            if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size():
+                score = tokenizer.tokenizer.sp_model.get_score(token_id)
+
+            if token_id >= tokenizer.tokenizer.sp_model.vocab_size():
+                if piece in special_tokens:
+                    toktype = SentencePieceTokenTypes.CONTROL
+                elif len(piece) == 0:
+                    text = f"[PAD{token_id}]".encode("utf-8")
+                    toktype = SentencePieceTokenTypes.UNUSED
+                else:
+                    toktype = SentencePieceTokenTypes.USER_DEFINED
+                tokens.append(text)
+                scores.append(score)
+                toktypes.append(toktype)
+                continue
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.tokenizer.sp_model.is_unknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.tokenizer.sp_model.is_control(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.tokenizer.sp_model.is_unused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.tokenizer.sp_model.is_byte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        # glm3 needs prefix and suffix formatted as:
+        # prompt = "[gMASK]sop<|user|>\n" + prompt + "<|assistant|>"
+        self.gguf_writer.add_tokenizer_pre("chatglm-spm")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    @staticmethod
+    def token_bytes_to_string(b):
+        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
+        byte_encoder = bytes_to_unicode()
+        return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
+
+    @staticmethod
+    def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: int | None = None) -> list[bytes]:
+        parts = [bytes([b]) for b in token]
+        while True:
+            min_idx = None
+            min_rank = None
+            for i, pair in enumerate(zip(parts[:-1], parts[1:])):
+                rank = mergeable_ranks.get(pair[0] + pair[1])
+                if rank is not None and (min_rank is None or rank < min_rank):
+                    min_idx = i
+                    min_rank = rank
+            if min_rank is None or (max_rank is not None and min_rank >= max_rank):
+                break
+            assert min_idx is not None
+            parts = parts[:min_idx] + [parts[min_idx] + parts[min_idx + 1]] + parts[min_idx + 2:]
+        return parts
+
+    def set_vocab(self):
+        if "THUDM/chatglm3-6b" in self.hparams.get("_name_or_path", ""):
+            self.set_vocab_chatglm3()
+            return
+
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        vocab_size = hparams["padded_vocab_size"]
+        assert max(tokenizer.get_vocab().values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        merges = []
+        vocab = {}
+        mergeable_ranks = tokenizer.mergeable_ranks
+        for token, rank in mergeable_ranks.items():
+            vocab[ChatGLMModel.token_bytes_to_string(token)] = rank
+            if len(token) == 1:
+                continue
+            merged = ChatGLMModel.bpe(mergeable_ranks, token, max_rank=rank)
+            assert len(merged) >= 2 and len(merged) <= 7
+            merges.append(' '.join(map(ChatGLMModel.token_bytes_to_string, merged)))
+
+        # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
+        added_vocab = tokenizer.get_added_vocab()
+        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            elif reverse_vocab[i] in added_vocab:
+                tokens.append(reverse_vocab[i])
+                if tokenizer.added_tokens_decoder[i].special:
+                    toktypes.append(gguf.TokenType.CONTROL)
+                else:
+                    toktypes.append(gguf.TokenType.USER_DEFINED)
+            else:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.NORMAL)
+
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
+        special_vocab.merges = merges
+        # only add special tokens when they were not already loaded from config.json
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
+        # this one is usually not in config.json anyway
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_head_kv = self.hparams.get("multi_query_group_num", n_head)
+        self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
+        self.gguf_writer.add_embedding_length(n_embed)
+        self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", 4 * n_embed))
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layernorm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_rope_dimension_count(64)
+        self.gguf_writer.add_add_bos_token(False)
+        rope_freq = 10000
+        if "rope_ratio" in self.hparams:
+            rope_freq = rope_freq * self.hparams["rope_ratio"]
+        self.gguf_writer.add_rope_freq_base(rope_freq)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if name.endswith(".rotary_pos_emb.inv_freq"):
+            return []
+
+        name = name.removeprefix("transformer.")
+        return [(self.map_tensor_name(name), data_torch)]
+
+###### CONVERSION LOGIC ######
+
+
+# tree of lazy tensors
+class LazyTorchTensor(gguf.LazyBase):
+    _tensor_type = torch.Tensor
+    # to keep the type-checker happy
+    dtype: torch.dtype
+    shape: torch.Size
+
+    # only used when converting a torch.Tensor to a np.ndarray
+    _dtype_map: dict[torch.dtype, type] = {
+        torch.float16: np.float16,
+        torch.float32: np.float32,
+    }
+
+    # used for safetensors slices
+    # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
+    # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
+    _dtype_str_map: dict[str, torch.dtype] = {
+        "F64": torch.float64,
+        "F32": torch.float32,
+        "BF16": torch.bfloat16,
+        "F16": torch.float16,
+        # "U64": torch.uint64,
+        "I64": torch.int64,
+        # "U32": torch.uint32,
+        "I32": torch.int32,
+        # "U16": torch.uint16,
+        "I16": torch.int16,
+        "U8": torch.uint8,
+        "I8": torch.int8,
+        "BOOL": torch.bool,
+        "F8_E4M3": torch.float8_e4m3fn,
+        "F8_E5M2": torch.float8_e5m2,
+    }
+
+    def numpy(self) -> gguf.LazyNumpyTensor:
+        dtype = self._dtype_map[self.dtype]
+        return gguf.LazyNumpyTensor(
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
+            args=(self,),
+            func=(lambda s: s.numpy())
+        )
+
+    @classmethod
+    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor:
+        return torch.empty(size=shape, dtype=dtype, device="meta")
+
+    @classmethod
+    def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
+        dtype = cls._dtype_str_map[st_slice.get_dtype()]
+        shape: tuple[int, ...] = tuple(st_slice.get_shape())
+        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
+        return cast(torch.Tensor, lazy)
+
+    @classmethod
+    def __torch_function__(cls, func, types, args=(), kwargs=None):
+        del types  # unused
+
+        if kwargs is None:
+            kwargs = {}
+
+        if func is torch.Tensor.numpy:
+            return args[0].numpy()
+
+        return cls._wrap_fn(func)(*args, **kwargs)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert a huggingface model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
+    )
+    parser.add_argument(
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
+    )
+    parser.add_argument(
+        "--bigendian", action="store_true",
+        help="model is executed on big endian machine",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file",
+    )
+    parser.add_argument(
+        "--use-temp-file", action="store_true",
+        help="use the tempfile library while processing (helpful when running out of memory, process killed)",
+    )
+    parser.add_argument(
+        "--no-lazy", action="store_true",
+        help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)",
+    )
+    parser.add_argument(
+        "--model-name", type=str, default=None,
+        help="name of the model",
+    )
+    parser.add_argument(
+        "--verbose", action="store_true",
+        help="increase output verbosity",
+    )
+    parser.add_argument(
+        "--split-max-tensors", type=int, default=0,
+        help="max tensors in each split",
+    )
+    parser.add_argument(
+        "--split-max-size", type=str, default="0",
+        help="max size per split N(M|G)",
+    )
+    parser.add_argument(
+        "--dry-run", action="store_true",
+        help="only print out a split plan and exit, without writing any new files",
+    )
+    parser.add_argument(
+        "--no-tensor-first-split", action="store_true",
+        help="do not add tensors to the first split (disabled by default)"
+    )
+    parser.add_argument(
+        "--metadata", type=Path,
+        help="Specify the path for an authorship metadata override file"
+    )
+
+    return parser.parse_args()
+
+
+def split_str_to_n_bytes(split_str: str) -> int:
+    if split_str.endswith("K"):
+        n = int(split_str[:-1]) * 1000
+    elif split_str.endswith("M"):
+        n = int(split_str[:-1]) * 1000 * 1000
+    elif split_str.endswith("G"):
+        n = int(split_str[:-1]) * 1000 * 1000 * 1000
+    elif split_str.isnumeric():
+        n = int(split_str)
+    else:
+        raise ValueError(f"Invalid split size: {split_str}, must be a number, optionally followed by K, M, or G")
+
+    if n < 0:
+        raise ValueError(f"Invalid split size: {split_str}, must be positive")
+
+    return n
+
+
+def main() -> None:
+    args = parse_args()
+
+    if args.verbose:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    dir_model = args.model
+
+    if not dir_model.is_dir():
+        logger.error(f'Error: {args.model} is not a directory')
+        sys.exit(1)
+
+    ftype_map: dict[str, gguf.LlamaFileType] = {
+        "f32": gguf.LlamaFileType.ALL_F32,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+        "auto": gguf.LlamaFileType.GUESSED,
+    }
+
+    is_split = args.split_max_tensors > 0 or args.split_max_size != "0"
+    if args.use_temp_file and is_split:
+        logger.error("Error: Cannot use temp file when splitting")
+        sys.exit(1)
+
+    if args.outfile is not None:
+        fname_out = args.outfile
+    else:
+        fname_out = dir_model
+
+    logger.info(f"Loading model: {dir_model.name}")
+
+    hparams = Model.load_hparams(dir_model)
+
+    with torch.inference_mode():
+        output_type = ftype_map[args.outtype]
+        model_architecture = hparams["architectures"][0]
+
+        try:
+            model_class = Model.from_model_architecture(model_architecture)
+        except NotImplementedError:
+            logger.error(f"Model {model_architecture} is not supported")
+            sys.exit(1)
+
+        model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out,
+                                     is_big_endian=args.bigendian, use_temp_file=args.use_temp_file,
+                                     eager=args.no_lazy,
+                                     metadata_override=args.metadata, model_name=args.model_name,
+                                     split_max_tensors=args.split_max_tensors,
+                                     split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
+                                     small_first_shard=args.no_tensor_first_split)
+
+        if args.vocab_only:
+            logger.info("Exporting model vocab...")
+            model_instance.write_vocab()
+            logger.info(f"Model vocab successfully exported to {model_instance.fname_out}")
+        else:
+            logger.info("Exporting model...")
+            model_instance.write()
+            out_path = f"{model_instance.fname_out.parent}{os.sep}" if is_split else model_instance.fname_out
+            logger.info(f"Model successfully exported to {out_path}")
+
+
+if __name__ == '__main__':
+    main()
diff --git a/llm/llama.cpp/convert_hf_to_gguf_update.py b/llm/llama.cpp/convert_hf_to_gguf_update.py
new file mode 100644
index 0000000000000000000000000000000000000000..d5a2d925eaef5582fb68984e0ffed718a61134b2
--- /dev/null
+++ b/llm/llama.cpp/convert_hf_to_gguf_update.py
@@ -0,0 +1,351 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# This script downloads the tokenizer models of the specified models from Huggingface and
+# generates the get_vocab_base_pre() function for convert_hf_to_gguf.py
+#
+# This is necessary in order to analyze the type of pre-tokenizer used by the model and
+# provide the necessary information to llama.cpp via the GGUF header in order to implement
+# the same pre-tokenizer.
+#
+# ref: https://github.com/ggerganov/llama.cpp/pull/6920
+#
+# Instructions:
+#
+# - Add a new model to the "models" list
+# - Run the script with your huggingface token:
+#
+#   python3 convert_hf_to_gguf_update.py 
+#
+# - Copy-paste the generated get_vocab_base_pre() function into convert_hf_to_gguf.py
+# - Update llama.cpp with the new pre-tokenizer if necessary
+#
+# TODO: generate tokenizer tests for llama.cpp
+#
+
+import logging
+import os
+import pathlib
+import re
+
+import requests
+import sys
+import json
+
+from hashlib import sha256
+from enum import IntEnum, auto
+from transformers import AutoTokenizer
+
+logging.basicConfig(level=logging.DEBUG)
+logger = logging.getLogger("convert_hf_to_gguf_update")
+sess = requests.Session()
+
+
+class TOKENIZER_TYPE(IntEnum):
+    SPM = auto()
+    BPE = auto()
+    WPM = auto()
+    UGM = auto()
+
+
+# TODO: this string has to exercise as much pre-tokenizer functionality as possible
+#       will be updated with time - contributions welcome
+CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
+
+if len(sys.argv) == 2:
+    token = sys.argv[1]
+    if not token.startswith("hf_"):
+        logger.info("Huggingface token seems invalid")
+        logger.info("Usage: python convert_hf_to_gguf_update.py ")
+        sys.exit(1)
+else:
+    logger.info("Usage: python convert_hf_to_gguf_update.py ")
+    sys.exit(1)
+
+# TODO: add models here, base models preferred
+models = [
+    {"name": "llama-spm",      "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
+    {"name": "llama-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", },
+    {"name": "phi-3",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", },
+    {"name": "deepseek-llm",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", },
+    {"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
+    {"name": "falcon",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
+    {"name": "bert-bge",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
+    {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
+    {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
+    {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
+    {"name": "stablelm2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
+    {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
+    {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
+    {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
+    {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
+    {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
+    {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
+    {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
+    {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
+    {"name": "poro-chat",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
+    {"name": "jina-v2-code",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
+    {"name": "viking",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B
+    {"name": "gemma",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
+    {"name": "gemma-2",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
+    {"name": "jais",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
+    {"name": "t5",             "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
+    {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
+    {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
+    {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+]
+
+
+def download_file_with_auth(url, token, save_path):
+    headers = {"Authorization": f"Bearer {token}"}
+    response = sess.get(url, headers=headers)
+    response.raise_for_status()
+    os.makedirs(os.path.dirname(save_path), exist_ok=True)
+    with open(save_path, 'wb') as downloaded_file:
+        downloaded_file.write(response.content)
+    logger.info(f"File {save_path} downloaded successfully")
+
+
+def download_model(model):
+    name = model["name"]
+    repo = model["repo"]
+    tokt = model["tokt"]
+
+    os.makedirs(f"models/tokenizers/{name}", exist_ok=True)
+
+    files = ["config.json", "tokenizer.json", "tokenizer_config.json"]
+
+    if tokt == TOKENIZER_TYPE.SPM:
+        files.append("tokenizer.model")
+
+    if tokt == TOKENIZER_TYPE.UGM:
+        files.append("spiece.model")
+
+    for file in files:
+        save_path = f"models/tokenizers/{name}/{file}"
+        if os.path.isfile(save_path):
+            logger.info(f"{name}: File {save_path} already exists - skipping")
+            continue
+        download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+
+
+for model in models:
+    try:
+        download_model(model)
+    except Exception as e:
+        logger.error(f"Failed to download model {model['name']}. Error: {e}")
+
+
+# generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
+
+src_ifs = ""
+for model in models:
+    name = model["name"]
+    tokt = model["tokt"]
+
+    if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
+        continue
+
+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
+    # create the tokenizer
+    try:
+        if name == "t5":
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
+        else:
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
+        continue  # Skip to the next model if the tokenizer can't be loaded
+
+    chktok = tokenizer.encode(CHK_TXT)
+    chkhsh = sha256(str(chktok).encode()).hexdigest()
+
+    logger.info(f"model: {name}")
+    logger.info(f"tokt: {tokt}")
+    logger.info(f"repo: {model['repo']}")
+    logger.info(f"chktok: {chktok}")
+    logger.info(f"chkhsh: {chkhsh}")
+
+    # print the "pre_tokenizer" content from the tokenizer.json
+    with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
+        cfg = json.load(f)
+        normalizer = cfg["normalizer"]
+        logger.info("normalizer: " + json.dumps(normalizer, indent=4))
+        pre_tokenizer = cfg["pre_tokenizer"]
+        logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
+        if "ignore_merges" in cfg["model"]:
+            logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))
+
+    logger.info("")
+
+    src_ifs += f"        if chkhsh == \"{chkhsh}\":\n"
+    src_ifs += f"            # ref: {model['repo']}\n"
+    src_ifs += f"            res = \"{name}\"\n"
+
+src_func = f"""
+    def get_vocab_base_pre(self, tokenizer) -> str:
+        # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
+        # is specific for the BPE pre-tokenizer used by the model
+        # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
+        # use in llama.cpp to implement the same pre-tokenizer
+
+        chktxt = {repr(CHK_TXT)}
+
+        chktok = tokenizer.encode(chktxt)
+        chkhsh = sha256(str(chktok).encode()).hexdigest()
+
+        logger.debug(f"chktok: {{chktok}}")
+        logger.debug(f"chkhsh: {{chkhsh}}")
+
+        res = None
+
+        # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script
+        #       or pull the latest version of the model from Huggingface
+        #       don't edit the hashes manually!
+{src_ifs}
+        if res is None:
+            logger.warning("\\n")
+            logger.warning("**************************************************************************************")
+            logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!")
+            logger.warning("**          There are 2 possible reasons for this:")
+            logger.warning("**          - the model has not been added to convert_hf_to_gguf_update.py yet")
+            logger.warning("**          - the pre-tokenization config has changed upstream")
+            logger.warning("**          Check your model files and convert_hf_to_gguf_update.py and update them accordingly.")
+            logger.warning("** ref:     https://github.com/ggerganov/llama.cpp/pull/6920")
+            logger.warning("**")
+            logger.warning(f"** chkhsh:  {{chkhsh}}")
+            logger.warning("**************************************************************************************")
+            logger.warning("\\n")
+            raise NotImplementedError("BPE pre-tokenizer was not recognized - update get_vocab_base_pre()")
+
+        logger.debug(f"tokenizer.ggml.pre: {{repr(res)}}")
+        logger.debug(f"chkhsh: {{chkhsh}}")
+
+        return res
+"""
+
+convert_py_pth = pathlib.Path("convert_hf_to_gguf.py")
+convert_py = convert_py_pth.read_text(encoding="utf-8")
+convert_py = re.sub(
+    r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)",
+    lambda m: m.group(1) + src_func + m.group(3),
+    convert_py,
+    flags=re.DOTALL | re.MULTILINE,
+)
+
+convert_py_pth.write_text(convert_py, encoding="utf-8")
+
+logger.info("+++ convert_hf_to_gguf.py was updated")
+
+# generate tests for each tokenizer model
+
+tests = [
+    "ied 4 ½ months",
+    "Führer",
+    "",
+    " ",
+    "  ",
+    "   ",
+    "\t",
+    "\n",
+    "\n\n",
+    "\n\n\n",
+    "\t\n",
+    "Hello world",
+    " Hello world",
+    "Hello World",
+    " Hello World",
+    " Hello World!",
+    "Hello, world!",
+    " Hello, world!",
+    " this is 🦙.cpp",
+    "w048 7tuijk dsdfhu",
+    "нещо на Български",
+    "កាន់តែពិសេសអាចខលចេញ",
+    "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
+    "Hello",
+    " Hello",
+    "  Hello",
+    "   Hello",
+    "    Hello",
+    "    Hello\n    Hello",
+    " (",
+    "\n =",
+    "' era",
+    "Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
+    "!!!!!!",
+    "3",
+    "33",
+    "333",
+    "3333",
+    "33333",
+    "333333",
+    "3333333",
+    "33333333",
+    "333333333",
+    "Cửa Việt", # llama-bpe fails on this
+    " discards",
+    CHK_TXT,
+]
+
+# write the tests to ./models/ggml-vocab-{name}.gguf.inp
+# the format is:
+#
+# test0
+# __ggml_vocab_test__
+# test1
+# __ggml_vocab_test__
+# ...
+#
+
+# with each model, encode all tests and write the results in ./models/ggml-vocab-{name}.gguf.out
+# for each test, write the resulting tokens on a separate line
+
+for model in models:
+    name = model["name"]
+    tokt = model["tokt"]
+
+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
+    # create the tokenizer
+    try:
+        if name == "t5":
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
+        else:
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
+        continue  # Skip this model and continue with the next one in the loop
+
+    with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
+        for text in tests:
+            f.write(f"{text}")
+            f.write("\n__ggml_vocab_test__\n")
+
+    with open(f"models/ggml-vocab-{name}.gguf.out", "w") as f:
+        for text in tests:
+            res = tokenizer.encode(text, add_special_tokens=False)
+            for r in res:
+                f.write(f" {r}")
+            f.write("\n")
+
+    logger.info(f"Tests for {name} written in ./models/ggml-vocab-{name}.gguf.*")
+
+# generate commands for creating vocab files
+
+logger.info("\nRun the following commands to generate the vocab files for testing:\n")
+
+for model in models:
+    name = model["name"]
+
+    print(f"python3 convert_hf_to_gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100
+
+logger.info("\n")
diff --git a/llm/llama.cpp/convert_llama_ggml_to_gguf.py b/llm/llama.cpp/convert_llama_ggml_to_gguf.py
new file mode 100644
index 0000000000000000000000000000000000000000..7b00b4398178b629d62be86c9d8706c1a7023423
--- /dev/null
+++ b/llm/llama.cpp/convert_llama_ggml_to_gguf.py
@@ -0,0 +1,450 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import logging
+import argparse
+import os
+import struct
+import sys
+from enum import IntEnum
+from pathlib import Path
+
+import numpy as np
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+import gguf
+
+logger = logging.getLogger("ggml-to-gguf")
+
+
+class GGMLFormat(IntEnum):
+    GGML = 0
+    GGMF = 1
+    GGJT = 2
+
+
+class GGMLFType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1
+    MOSTLY_Q4_0          = 2
+    MOSTLY_Q4_1          = 3
+    MOSTLY_Q4_1_SOME_F16 = 4
+    MOSTLY_Q8_0          = 7
+    MOSTLY_Q5_0          = 8
+    MOSTLY_Q5_1          = 9
+    MOSTLY_Q2_K          = 10
+    MOSTLY_Q3_K_S        = 11
+    MOSTLY_Q3_K_M        = 12
+    MOSTLY_Q3_K_L        = 13
+    MOSTLY_Q4_K_S        = 14
+    MOSTLY_Q4_K_M        = 15
+    MOSTLY_Q5_K_S        = 16
+    MOSTLY_Q5_K_M        = 17
+    MOSTLY_Q6_K          = 18
+
+
+class Hyperparameters:
+    def __init__(self):
+        self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
+        self.n_layer = self.n_rot = self.n_ff = 0
+        self.ftype = GGMLFType.ALL_F32
+
+    def set_n_ff(self, model):
+        ff_tensor_idx = model.tensor_map.get(b'layers.0.feed_forward.w1.weight')
+        assert ff_tensor_idx is not None, 'Missing layer 0 FF tensor'
+        ff_tensor = model.tensors[ff_tensor_idx]
+        self.n_ff = ff_tensor.dims[1]
+
+    def load(self, data, offset):
+        (
+            self.n_vocab,
+            self.n_embd,
+            self.n_mult,
+            self.n_head,
+            self.n_layer,
+            self.n_rot,
+            ftype,
+        ) = struct.unpack('<7I', data[offset:offset + (4 * 7)])
+        try:
+            self.ftype = GGMLFType(ftype)
+        except ValueError:
+            raise ValueError(f'Invalid ftype {ftype}')
+        return 4 * 7
+
+    def __str__(self):
+        return f''
+
+
+class Vocab:
+    def __init__(self, load_scores = True):
+        self.items = []
+        self.load_scores = load_scores
+
+    def load(self, data, offset, n_vocab):
+        orig_offset = offset
+        for _ in range(n_vocab):
+            itemlen = struct.unpack('= 0 and n_dims <= 4, f'Invalid tensor dimensions {n_dims}'
+        assert name_len < 4096, 'Absurd tensor name length'
+        quant = gguf.GGML_QUANT_SIZES.get(dtype)
+        assert quant is not None, 'Unknown tensor type'
+        (blksize, tysize) = quant
+        offset += 12
+        self.dtype= dtype
+        self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)])
+        offset += 4 * n_dims
+        self.name = bytes(data[offset:offset + name_len])
+        offset += name_len
+        pad = ((offset + 31) & ~31) - offset if self.use_padding else 0
+        offset += pad
+        n_elems = np.prod(self.dims)
+        n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize)
+        self.start_offset = offset
+        self.len_bytes = n_bytes
+        offset += n_bytes
+        return offset - orig_offset
+
+
+class GGMLModel:
+
+    file_format: GGMLFormat
+    format_version: int
+
+    def __init__(self):
+        self.hyperparameters = None
+        self.vocab = None
+        self.tensor_map = {}
+        self.tensors = []
+
+    def validate_header(self, data, offset):
+        magic = bytes(data[offset:offset + 4])
+        if magic == b'GGUF':
+            raise ValueError('File is already in GGUF format.')
+        if magic == b'lmgg':
+            self.file_format = GGMLFormat.GGML
+            self.format_version = 1
+            return 4
+        version = struct.unpack(' 3:
+                raise ValueError(f'Cannot handle unexpected GGJT file version {version}')
+            self.file_format = GGMLFormat.GGJT
+            self.format_version = version
+            return 8
+        raise ValueError(f"Unexpected file magic {magic!r}! This doesn't look like a GGML format file.")
+
+    def validate_conversion(self, ftype):
+        err = ''
+        if (self.file_format < GGMLFormat.GGJT or self.format_version < 2):
+            if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
+                err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
+        elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
+            if ftype in (GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
+                         GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
+                err = 'Q4 and Q8 quantizations changed in GGJTv3.'
+        if len(err) > 0:
+            raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')
+
+    def load(self, data, offset):
+        offset += self.validate_header(data, offset)
+        hp = Hyperparameters()
+        offset += hp.load(data, offset)
+        logger.info(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
+        self.validate_conversion(hp.ftype)
+        vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML)
+        offset += vocab.load(data, offset, hp.n_vocab)
+        tensors: list[Tensor] = []
+        tensor_map = {}
+        while offset < len(data):
+            tensor = Tensor(use_padding = self.file_format > GGMLFormat.GGMF)
+            offset += tensor.load(data, offset)
+            tensor_map[tensor.name] = len(tensors)
+            tensors.append(tensor)
+        self.hyperparameters = hp
+        self.vocab = vocab
+        self.tensors = tensors
+        self.tensor_map = tensor_map
+        hp.set_n_ff(self)
+        return offset
+
+
+class GGMLToGGUF:
+    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
+        hp = ggml_model.hyperparameters
+        self.model = ggml_model
+        self.data = data
+        self.cfg = cfg
+        self.params_override = params_override
+        self.vocab_override = vocab_override
+        self.special_vocab = special_vocab
+        if params_override is not None:
+            n_kv_head = params_override.n_head_kv
+        else:
+            if cfg.gqa == 1:
+                n_kv_head = hp.n_head
+            else:
+                gqa = float(cfg.gqa)
+                n_kv_head = None
+                for x in range(1, 256):
+                    if float(hp.n_head) / float(x) == gqa:
+                        n_kv_head = x
+                assert n_kv_head is not None, "Couldn't determine n_kv_head from GQA param"
+                logger.info(f'- Guessed n_kv_head = {n_kv_head} based on GQA {cfg.gqa}')
+        self.n_kv_head = n_kv_head
+        self.name_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.LLAMA, ggml_model.hyperparameters.n_layer)
+
+    def save(self):
+        logger.info('* Preparing to save GGUF file')
+        gguf_writer = gguf.GGUFWriter(
+            self.cfg.output,
+            gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
+            use_temp_file = False)
+        self.add_params(gguf_writer)
+        self.add_vocab(gguf_writer)
+        if self.special_vocab is not None:
+            self.special_vocab.add_to_gguf(gguf_writer)
+        self.add_tensors(gguf_writer)
+        logger.info("    gguf: write header")
+        gguf_writer.write_header_to_file()
+        logger.info("    gguf: write metadata")
+        gguf_writer.write_kv_data_to_file()
+        logger.info("    gguf: write tensors")
+        gguf_writer.write_tensors_to_file()
+        gguf_writer.close()
+
+    def add_params(self, gguf_writer):
+        hp = self.model.hyperparameters
+        cfg = self.cfg
+        if cfg.desc is not None:
+            desc = cfg.desc
+        else:
+            desc = f'converted from legacy {self.model.file_format.name}v{self.model.format_version} {hp.ftype.name} format'
+        try:
+            # Filenames aren't necessarily valid UTF8.
+            name = cfg.name if cfg.name is not None else cfg.input.name
+        except UnicodeDecodeError:
+            name = None
+        logger.info('* Adding model parameters and KV items')
+        if name is not None:
+            gguf_writer.add_name(name)
+        gguf_writer.add_description(desc)
+        gguf_writer.add_file_type(int(hp.ftype))
+        if self.params_override is not None:
+            po = self.params_override
+            assert po.n_embd == hp.n_embd, 'Model hyperparams mismatch'
+            assert po.n_layer == hp.n_layer, 'Model hyperparams mismatch'
+            assert po.n_head == hp.n_head, 'Model hyperparams mismatch'
+            gguf_writer.add_context_length      (po.n_ctx)
+            gguf_writer.add_embedding_length    (po.n_embd)
+            gguf_writer.add_block_count         (po.n_layer)
+            gguf_writer.add_feed_forward_length (po.n_ff)
+            gguf_writer.add_rope_dimension_count(po.n_embd // po.n_head)
+            gguf_writer.add_head_count          (po.n_head)
+            gguf_writer.add_head_count_kv       (po.n_head_kv)
+            gguf_writer.add_layer_norm_rms_eps  (po.f_norm_eps)
+            return
+        gguf_writer.add_context_length(cfg.context_length)
+        gguf_writer.add_embedding_length(hp.n_embd)
+        gguf_writer.add_block_count(hp.n_layer)
+        gguf_writer.add_feed_forward_length(hp.n_ff)
+        gguf_writer.add_rope_dimension_count(hp.n_embd // hp.n_head)
+        gguf_writer.add_head_count(hp.n_head)
+        gguf_writer.add_head_count_kv(self.n_kv_head)
+        gguf_writer.add_layer_norm_rms_eps(float(cfg.eps))
+
+    def add_vocab(self, gguf_writer):
+        hp = self.model.hyperparameters
+        gguf_writer.add_tokenizer_model('llama')
+        gguf_writer.add_tokenizer_pre('default')
+        tokens = []
+        scores = []
+        toktypes = []
+        if self.vocab_override is not None:
+            vo = self.vocab_override
+            logger.info('* Adding vocab item(s)')
+            for (_, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
+                tokens.append(vbytes)
+                scores.append(score)
+                toktypes.append(ttype)
+            assert len(tokens) == hp.n_vocab, \
+                f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
+            gguf_writer.add_token_list(tokens)
+            gguf_writer.add_token_scores(scores)
+            if len(toktypes) > 0:
+                gguf_writer.add_token_types(toktypes)
+            return
+        logger.info(f'* Adding {hp.n_vocab} vocab item(s)')
+        assert len(self.model.vocab.items) >= 3, 'Cannot handle unexpectedly short model vocab'
+        for (tokid, (vbytes, vscore)) in enumerate(self.model.vocab.items):
+            tt = 1 # Normal
+            # Special handling for UNK, BOS, EOS tokens.
+            if tokid <= 2:
+                if tokid == 0:
+                    vbytes = b''
+                    tt = 2
+                elif tokid == 1:
+                    vbytes = b''
+                    tt = 3
+                else:
+                    vbytes = b''
+                    tt = 3
+            elif len(vbytes) == 0:
+                tt = 3 # Control
+            elif tokid >= 3 and tokid <= 258 and len(vbytes) == 1:
+                vbytes = bytes(f'<0x{vbytes[0]:02X}>', encoding = 'UTF-8')
+                tt = 6 # Byte
+            else:
+                vbytes = vbytes.replace(b' ', b'\xe2\x96\x81')
+            toktypes.append(tt)
+            tokens.append(vbytes)
+            scores.append(vscore)
+        gguf_writer.add_token_list(tokens)
+        gguf_writer.add_token_scores(scores)
+        gguf_writer.add_token_types(toktypes)
+        gguf_writer.add_unk_token_id(0)
+        gguf_writer.add_bos_token_id(1)
+        gguf_writer.add_eos_token_id(2)
+
+    def add_tensors(self, gguf_writer):
+        tensor_map = self.name_map
+        data = self.data
+        logger.info(f'* Adding {len(self.model.tensors)} tensor(s)')
+        for tensor in self.model.tensors:
+            name = str(tensor.name, 'UTF-8')
+            mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+            assert mapped_name is not None, f'Bad name {name}'
+            tempdims = list(tensor.dims[:])
+            if len(tempdims) > 1:
+                temp = tempdims[1]
+                tempdims[1] = tempdims[0]
+                tempdims[0] = temp
+            gguf_writer.add_tensor(
+                mapped_name,
+                data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
+                raw_shape = tempdims,
+                raw_dtype = tensor.dtype)
+
+
+def handle_metadata(cfg, hp):
+    import examples.convert_legacy_llama as convert
+
+    assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory'
+    hf_config_path   = cfg.model_metadata_dir / "config.json"
+    orig_config_path = cfg.model_metadata_dir / "params.json"
+    # We pass a fake model here. "original" mode will check the shapes of some
+    # tensors if information is missing in the .json file: other than that, the
+    # model data isn't used so this should be safe (at least for now).
+    fakemodel = {
+        'tok_embeddings.weight': convert.LazyTensor.__new__(convert.LazyTensor),
+        'layers.0.feed_forward.w1.weight': convert.LazyTensor.__new__(convert.LazyTensor),
+    }
+    fakemodel['tok_embeddings.weight'].shape = [hp.n_vocab]
+    fakemodel['layers.0.feed_forward.w1.weight'].shape = [hp.n_ff]
+    if hf_config_path.exists():
+        params = convert.Params.loadHFTransformerJson(fakemodel, hf_config_path)
+    elif orig_config_path.exists():
+        params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
+    else:
+        raise ValueError('Unable to load metadata')
+    vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir)
+    vocab_factory = convert.VocabFactory(vocab_path)
+    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype.split(","), cfg.model_metadata_dir)
+    convert.check_vocab_size(params, vocab)
+    return params, vocab, special_vocab
+
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
+    parser.add_argument('--input', '-i', type = Path, required = True,
+                        help = 'Input GGMLv3 filename')
+    parser.add_argument('--output', '-o', type = Path, required = True,
+                        help ='Output GGUF filename')
+    parser.add_argument('--name',
+                        help = 'Set model name')
+    parser.add_argument('--desc',
+                        help = 'Set model description')
+    parser.add_argument('--gqa', type = int, default = 1,
+                        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
+    parser.add_argument('--eps', default = '5.0e-06',
+                        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
+    parser.add_argument('--context-length', '-c', type=int, default = 2048,
+                        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
+    parser.add_argument('--model-metadata-dir', '-m', type = Path,
+                        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
+    parser.add_argument("--vocab-dir", type=Path,
+                        help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
+    parser.add_argument("--vocabtype", default="spm,hfft",
+                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
+    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
+    return parser.parse_args()
+
+
+def main():
+    cfg = handle_args()
+    logging.basicConfig(level=logging.DEBUG if cfg.verbose else logging.INFO)
+    logger.info(f'* Using config: {cfg}')
+    logger.warning('=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===')
+    if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'):
+        logger.info('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
+    data = np.memmap(cfg.input, mode = 'r')
+    model = GGMLModel()
+    logger.info('* Scanning GGML input file')
+    offset = model.load(data, 0)  # noqa
+    logger.info(f'* GGML model hyperparameters: {model.hyperparameters}')
+    vocab_override = None
+    params_override = None
+    special_vocab = None
+    if cfg.model_metadata_dir is not None:
+        (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
+        logger.info('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
+        logger.info(f'* Overriding params: {params_override}')
+        logger.info(f'* Overriding vocab: {vocab_override}')
+        logger.info(f'* Special vocab: {special_vocab}')
+    else:
+        logger.warning('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
+        if model.file_format == GGMLFormat.GGML:
+            logger.info('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
+    converter = GGMLToGGUF(
+        model, data, cfg,
+        params_override = params_override,
+        vocab_override = vocab_override,
+        special_vocab = special_vocab
+    )
+    converter.save()
+    logger.info(f'* Successful completion. Output saved to: {cfg.output}')
+
+
+if __name__ == '__main__':
+    main()
diff --git a/llm/llama.cpp/convert_lora_to_gguf.py b/llm/llama.cpp/convert_lora_to_gguf.py
new file mode 100644
index 0000000000000000000000000000000000000000..a88d0d4a978a9d8514a4a786abdce4ad0d27c299
--- /dev/null
+++ b/llm/llama.cpp/convert_lora_to_gguf.py
@@ -0,0 +1,393 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+import logging
+import argparse
+import os
+import sys
+import json
+from math import prod
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+import gguf
+
+# reuse model definitions from convert_hf_to_gguf.py
+from convert_hf_to_gguf import LazyTorchTensor, Model
+
+logger = logging.getLogger("lora-to-gguf")
+
+
+@dataclass
+class PartialLoraTensor:
+    A: Tensor | None = None
+    B: Tensor | None = None
+
+
+# magic to support tensor shape modifications and splitting
+class LoraTorchTensor:
+    _lora_A: Tensor  # (n_rank, row_size)
+    _lora_B: Tensor  # (col_size, n_rank)
+    _rank: int
+
+    def __init__(self, A: Tensor, B: Tensor):
+        assert len(A.shape) == len(B.shape)
+        assert A.shape[-2] == B.shape[-1]
+        if A.dtype != B.dtype:
+            A = A.to(torch.float32)
+            B = B.to(torch.float32)
+        self._lora_A = A
+        self._lora_B = B
+        self._rank = B.shape[-1]
+
+    def get_lora_A_B(self) -> tuple[Tensor, Tensor]:
+        return (self._lora_A, self._lora_B)
+
+    def __getitem__(
+        self,
+        indices: (
+            SupportsIndex
+            | slice
+            | tuple[SupportsIndex | slice | Tensor, ...]  # TODO: add ellipsis in the type signature
+        ),
+    ) -> LoraTorchTensor:
+        shape = self.shape
+        if isinstance(indices, SupportsIndex):
+            if len(shape) > 2:
+                return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
+            else:
+                raise NotImplementedError  # can't return a vector
+        elif isinstance(indices, slice):
+            if len(shape) > 2:
+                return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
+            else:
+                return LoraTorchTensor(self._lora_A, self._lora_B[indices])
+        elif isinstance(indices, tuple):
+            assert len(indices) > 0
+            if indices[-1] is Ellipsis:
+                return self[indices[:-1]]
+            # expand ellipsis
+            indices = tuple(
+                u
+                for v in (
+                    (
+                        (slice(None, None) for _ in range(len(indices) - 1))
+                        if i is Ellipsis
+                        else (i,)
+                    )
+                    for i in indices
+                )
+                for u in v
+            )
+
+            if len(indices) < len(shape):
+                indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape))))
+
+            # TODO: make sure this is correct
+            indices_A = (
+                *(
+                    (
+                        j.__index__() % self._lora_A.shape[i]
+                        if isinstance(j, SupportsIndex)
+                        else slice(None, None)
+                    )
+                    for i, j in enumerate(indices[:-2])
+                ),
+                slice(None, None),
+                indices[-1],
+            )
+            indices_B = indices[:-1]
+            return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B])
+        else:
+            raise NotImplementedError  # unknown indice type
+
+    @property
+    def dtype(self) -> torch.dtype:
+        assert self._lora_A.dtype == self._lora_B.dtype
+        return self._lora_A.dtype
+
+    @property
+    def shape(self) -> tuple[int, ...]:
+        assert len(self._lora_A.shape) == len(self._lora_B.shape)
+        return (*self._lora_B.shape[:-1], self._lora_A.shape[-1])
+
+    def size(self, dim=None):
+        assert dim is None
+        return self.shape
+
+    def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
+        if isinstance(shape[0], tuple):
+            new_shape: tuple[int, ...] = shape[0]
+        else:
+            new_shape = cast(tuple[int, ...], shape)
+        orig_shape = self.shape
+        if len(new_shape) < 2:
+            raise NotImplementedError  # can't become a vector
+
+        # expand -1 in the shape
+        if any(dim == -1 for dim in new_shape):
+            n_elems = prod(orig_shape)
+            n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape)
+            assert n_elems % n_new_elems == 0
+            new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),)
+
+        if new_shape[-1] != orig_shape[-1]:
+            raise NotImplementedError  # can't reshape the row size trivially
+
+        shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1])
+        shape_B = (*new_shape[:-1], self._rank)
+        return LoraTorchTensor(
+            self._lora_A.reshape(shape_A),
+            self._lora_B.reshape(shape_B),
+        )
+
+    def reshape_as(self, other: Tensor) -> LoraTorchTensor:
+        return self.reshape(*other.shape)
+
+    def view(self, *size: int) -> LoraTorchTensor:
+        return self.reshape(*size)
+
+    def permute(self, *dims: int) -> LoraTorchTensor:
+        shape = self.shape
+        dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims)
+        if dims[-1] == -1:
+            # TODO: support higher dimensional A shapes bigger than 1
+            assert all(dim == 1 for dim in self._lora_A.shape[:-2])
+            return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims))
+        if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1:
+            return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims))
+        else:
+            # TODO: compose the above two
+            raise NotImplementedError
+
+    def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor:
+        shape = self.shape
+        dims = [i for i in range(len(shape))]
+        dims[dim0], dims[dim1] = dims[dim1], dims[dim0]
+        return self.permute(*dims)
+
+    def swapaxes(self, axis0: int, axis1: int) -> LoraTorchTensor:
+        return self.transpose(axis0, axis1)
+
+    def to(self, *args, **kwargs):
+        return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs))
+
+    @classmethod
+    def __torch_function__(cls, func: Callable, types, args=(), kwargs=None):
+        del types  # unused
+
+        if kwargs is None:
+            kwargs = {}
+
+        if func is torch.permute:
+            return type(args[0]).permute(*args, **kwargs)
+        elif func is torch.reshape:
+            return type(args[0]).reshape(*args, **kwargs)
+        elif func is torch.stack:
+            assert isinstance(args[0], Sequence)
+            dim = kwargs.get("dim", 0)
+            assert dim == 0
+            return LoraTorchTensor(
+                torch.stack([a._lora_A for a in args[0]], dim),
+                torch.stack([b._lora_B for b in args[0]], dim),
+            )
+        elif func is torch.cat:
+            assert isinstance(args[0], Sequence)
+            dim = kwargs.get("dim", 0)
+            assert dim == 0
+            if len(args[0][0].shape) > 2:
+                return LoraTorchTensor(
+                    torch.cat([a._lora_A for a in args[0]], dim),
+                    torch.cat([b._lora_B for b in args[0]], dim),
+                )
+            elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]):
+                return LoraTorchTensor(
+                    args[0][0]._lora_A,
+                    torch.cat([b._lora_B for b in args[0]], dim),
+                )
+            else:
+                raise NotImplementedError
+        else:
+            raise NotImplementedError
+
+
+def get_base_tensor_name(lora_tensor_name: str) -> str:
+    base_name = lora_tensor_name.replace("base_model.model.", "")
+    base_name = base_name.replace(".lora_A.weight", ".weight")
+    base_name = base_name.replace(".lora_B.weight", ".weight")
+    return base_name
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
+    )
+    parser.add_argument(
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
+    )
+    parser.add_argument(
+        "--bigendian", action="store_true",
+        help="model is executed on big endian machine",
+    )
+    parser.add_argument(
+        "--no-lazy", action="store_true",
+        help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)",
+    )
+    parser.add_argument(
+        "--verbose", action="store_true",
+        help="increase output verbosity",
+    )
+    parser.add_argument(
+        "--dry-run", action="store_true",
+        help="only print out what will be done, without writing any new files",
+    )
+    parser.add_argument(
+        "--base", type=Path, required=True,
+        help="directory containing base model file",
+    )
+    parser.add_argument(
+        "lora_path", type=Path,
+        help="directory containing LoRA adapter file",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
+
+    ftype_map: dict[str, gguf.LlamaFileType] = {
+        "f32": gguf.LlamaFileType.ALL_F32,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+        "auto": gguf.LlamaFileType.GUESSED,
+    }
+
+    ftype = ftype_map[args.outtype]
+
+    dir_base_model: Path = args.base
+    dir_lora: Path = args.lora_path
+    lora_config = dir_lora / "adapter_config.json"
+    input_model = dir_lora / "adapter_model.safetensors"
+
+    if args.outfile is not None:
+        fname_out = args.outfile
+    else:
+        # output in the same directory as the model by default
+        fname_out = dir_lora
+
+    if os.path.exists(input_model):
+        # lazy import load_file only if lora is in safetensors format.
+        from safetensors.torch import load_file
+
+        lora_model = load_file(input_model, device="cpu")
+    else:
+        input_model = os.path.join(dir_lora, "adapter_model.bin")
+        lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
+
+    # load base model
+    logger.info(f"Loading base model: {dir_base_model.name}")
+    hparams = Model.load_hparams(dir_base_model)
+    with torch.inference_mode():
+        try:
+            model_class = Model.from_model_architecture(hparams["architectures"][0])
+        except NotImplementedError:
+            logger.error(f"Model {hparams['architectures'][0]} is not supported")
+            sys.exit(1)
+
+        class LoraModel(model_class):
+            model_arch = model_class.model_arch
+
+            lora_alpha: float
+
+            def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs):
+
+                super().__init__(*args, **kwargs)
+
+                self.dir_model_card = dir_lora_model
+                self.lora_alpha = float(lora_alpha)
+
+            def set_type(self):
+                self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
+                self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
+
+            def set_gguf_parameters(self):
+                self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
+                super().set_gguf_parameters()
+
+            def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
+                tensor_map: dict[str, PartialLoraTensor] = {}
+
+                for name, tensor in lora_model.items():
+                    if self.lazy:
+                        tensor = LazyTorchTensor.from_eager(tensor)
+                    base_name = get_base_tensor_name(name)
+                    is_lora_a = ".lora_A.weight" in name
+                    is_lora_b = ".lora_B.weight" in name
+                    if not is_lora_a and not is_lora_b:
+                        if ".base_layer.weight" in name:
+                            continue
+                        logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
+                        sys.exit(1)
+
+                    if base_name in tensor_map:
+                        if is_lora_a:
+                            tensor_map[base_name].A = tensor
+                        else:
+                            tensor_map[base_name].B = tensor
+                    else:
+                        if is_lora_a:
+                            tensor_map[base_name] = PartialLoraTensor(A=tensor)
+                        else:
+                            tensor_map[base_name] = PartialLoraTensor(B=tensor)
+
+                for name, tensor in tensor_map.items():
+                    assert tensor.A is not None
+                    assert tensor.B is not None
+                    yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B)))
+
+            def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+                dest = super().modify_tensors(data_torch, name, bid)
+                for dest_name, dest_data in dest:
+                    assert isinstance(dest_data, LoraTorchTensor)
+                    lora_a, lora_b = dest_data.get_lora_A_B()
+
+                    yield (dest_name + ".lora_a", lora_a)
+                    yield (dest_name + ".lora_b", lora_b)
+
+        with open(lora_config, "r") as f:
+            lparams: dict[str, Any] = json.load(f)
+
+        alpha: float = lparams["lora_alpha"]
+
+        model_instance = LoraModel(
+            dir_base_model,
+            ftype,
+            fname_out,
+            is_big_endian=args.bigendian,
+            use_temp_file=False,
+            eager=args.no_lazy,
+            dry_run=args.dry_run,
+            dir_lora_model=dir_lora,
+            lora_alpha=alpha,
+        )
+
+        logger.info("Exporting model...")
+        model_instance.write()
+        logger.info(f"Model successfully exported to {model_instance.fname_out}")
diff --git a/llm/llama.cpp/docs/android.md b/llm/llama.cpp/docs/android.md
new file mode 100644
index 0000000000000000000000000000000000000000..cec4358d98a26d45a7634b5307372eb3f9c1d84a
--- /dev/null
+++ b/llm/llama.cpp/docs/android.md
@@ -0,0 +1,56 @@
+
+# Android
+
+## Build on Android using Termux
+[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
+
+## Building the Project using Android NDK
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
+
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
+```
+$ mkdir build-android
+$ cd build-android
+$ export NDK=
+$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
+$ make
+```
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+```
+
+Here's a demo of an interactive session running on Pixel 5 phone:
+
+https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
diff --git a/llm/llama.cpp/docs/backend/BLIS.md b/llm/llama.cpp/docs/backend/BLIS.md
new file mode 100644
index 0000000000000000000000000000000000000000..35d06bd0f303d278c92d59d31326b919f4c9ddf5
--- /dev/null
+++ b/llm/llama.cpp/docs/backend/BLIS.md
@@ -0,0 +1,67 @@
+BLIS Installation Manual
+------------------------
+
+BLIS is a portable software framework for high-performance BLAS-like dense linear algebra libraries. It has received awards and recognition, including the 2023 James H. Wilkinson Prize for Numerical Software and the 2020 SIAM Activity Group on Supercomputing Best Paper Prize. BLIS provides a new BLAS-like API and a compatibility layer for traditional BLAS routine calls. It offers features such as object-based API, typed API, BLAS and CBLAS compatibility layers.
+
+Project URL: https://github.com/flame/blis
+
+### Prepare:
+
+Compile BLIS:
+
+```bash
+git clone https://github.com/flame/blis
+cd blis
+./configure --enable-cblas -t openmp,pthreads auto
+# will install to /usr/local/ by default.
+make -j
+```
+
+Install BLIS:
+
+```bash
+sudo make install
+```
+
+We recommend using openmp since it's easier to modify the cores being used.
+
+### llama.cpp compilation
+
+Makefile:
+
+```bash
+make GGML_BLIS=1 -j
+# make GGML_BLIS=1 llama-benchmark-matmult
+```
+
+CMake:
+
+```bash
+mkdir build
+cd build
+cmake -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FLAME ..
+make -j
+```
+
+### llama.cpp execution
+
+According to the BLIS documentation, we could set the following
+environment variables to modify the behavior of openmp:
+
+```bash
+export GOMP_CPU_AFFINITY="0-19"
+export BLIS_NUM_THREADS=14
+```
+
+And then run the binaries as normal.
+
+
+### Intel specific issue
+
+Some might get the error message saying that `libimf.so` cannot be found.
+Please follow this [stackoverflow page](https://stackoverflow.com/questions/70687930/intel-oneapi-2022-libimf-so-no-such-file-or-directory-during-openmpi-compila).
+
+### Reference:
+
+1. https://github.com/flame/blis#getting-started
+2. https://github.com/flame/blis/blob/master/docs/Multithreading.md
diff --git a/llm/llama.cpp/docs/backend/SYCL.md b/llm/llama.cpp/docs/backend/SYCL.md
new file mode 100644
index 0000000000000000000000000000000000000000..d36ac0a158dd4fda0ca470030a153a70a5440145
--- /dev/null
+++ b/llm/llama.cpp/docs/backend/SYCL.md
@@ -0,0 +1,580 @@
+# llama.cpp for SYCL
+
+- [Background](#background)
+- [Recommended Release](#recommended-release)
+- [News](#news)
+- [OS](#os)
+- [Hardware](#hardware)
+- [Docker](#docker)
+- [Linux](#linux)
+- [Windows](#windows)
+- [Environment Variable](#environment-variable)
+- [Known Issue](#known-issues)
+- [Q&A](#qa)
+- [TODO](#todo)
+
+## Background
+
+**SYCL** is a high-level parallel programming model designed to improve developers productivity writing code across various hardware accelerators such as CPUs, GPUs, and FPGAs. It is a single-source language designed for heterogeneous computing and based on standard C++17.
+
+**oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
+
+- **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL - Math Kernel Library)*.
+- **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
+- **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
+
+### Llama.cpp + SYCL
+
+The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it could support other vendor GPUs: Nvidia GPU (*AMD GPU coming*).
+
+When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneMKL](README.md#intel-onemkl) backend.
+
+It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
+
+## Recommended Release
+
+The SYCL backend would be broken by some PRs due to no online CI.
+
+The following release is verified with good quality:
+
+|Commit ID|Tag|Release|Verified  Platform|
+|-|-|-|-|
+|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1|
+
+
+## News
+
+- 2024.5
+  - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
+  - Arch Linux is verified successfully.
+
+- 2024.4
+  - Support data types: GGML_TYPE_IQ4_NL, GGML_TYPE_IQ4_XS, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S, GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M.
+
+- 2024.3
+  - Release binary files of Windows.
+  - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
+  - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
+  - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
+  - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
+  - Support detecting all GPUs with level-zero and same top **Max compute units**.
+  - Support OPs
+    - hardsigmoid
+    - hardswish
+    - pool2d
+
+- 2024.1
+  - Create SYCL backend for Intel GPU.
+  - Support Windows build
+
+## OS
+
+| OS      | Status  | Verified                                       |
+|---------|---------|------------------------------------------------|
+| Linux   | Support | Ubuntu 22.04, Fedora Silverblue 39, Arch Linux |
+| Windows | Support | Windows 11                                     |
+
+
+## Hardware
+
+### Intel GPU
+
+**Verified devices**
+
+| Intel GPU                     | Status  | Verified Model                        |
+|-------------------------------|---------|---------------------------------------|
+| Intel Data Center Max Series  | Support | Max 1550, 1100                        |
+| Intel Data Center Flex Series | Support | Flex 170                              |
+| Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
+| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
+
+*Notes:*
+
+- **Memory**
+  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
+
+  - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.
+
+- **Execution Unit (EU)**
+  - If the iGPU has less than 80 EUs, the inference speed will likely be too slow for practical use.
+
+### Other Vendor GPU
+
+**Verified devices**
+
+| Nvidia GPU               | Status  | Verified Model |
+|--------------------------|---------|----------------|
+| Ampere Series            | Support | A100, A4000    |
+| Ampere Series *(Mobile)* | Support | RTX 40 Series  |
+
+## Docker
+The docker build option is currently limited to *intel GPU* targets.
+
+### Build image
+```sh
+# Using FP16
+docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" -f .devops/llama-cli-intel.Dockerfile .
+```
+
+*Notes*:
+
+To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="GGML_SYCL_F16=ON"` argument from the previous command.
+
+You can also use the `.devops/llama-server-intel.Dockerfile`, which builds the *"server"* alternative.
+
+### Run container
+
+```sh
+# First, find all the DRI cards
+ls -la /dev/dri
+# Then, pick the card that you want to use (here for e.g. /dev/dri/card1).
+docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-sycl -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+```
+
+*Notes:*
+- Docker has been tested successfully on native Linux. WSL support has not been verified yet.
+- You may need to install Intel GPU driver on the **host** machine *(Please refer to the [Linux configuration](#linux) for details)*.
+
+## Linux
+
+### I. Setup Environment
+
+1. **Install GPU drivers**
+
+  - **Intel GPU**
+
+Intel data center GPUs drivers installation guide and download page can be found here: [Get intel dGPU Drivers](https://dgpu-docs.intel.com/driver/installation.html#ubuntu-install-steps).
+
+*Note*: for client GPUs *(iGPU & Arc A-Series)*, please refer to the [client iGPU driver installation](https://dgpu-docs.intel.com/driver/client/overview.html).
+
+Once installed, add the user(s) to the `video` and `render` groups.
+
+```sh
+sudo usermod -aG render $USER
+sudo usermod -aG video $USER
+```
+
+*Note*: logout/re-login for the changes to take effect.
+
+Verify installation through `clinfo`:
+
+```sh
+sudo apt install clinfo
+sudo clinfo -l
+```
+
+Sample output:
+
+```sh
+Platform #0: Intel(R) OpenCL Graphics
+ `-- Device #0: Intel(R) Arc(TM) A770 Graphics
+
+Platform #0: Intel(R) OpenCL HD Graphics
+ `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
+```
+
+- **Nvidia GPU**
+
+In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
+
+2. **Install Intel® oneAPI Base toolkit**
+
+- **For Intel GPU**
+
+The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
+
+Please follow the instructions for downloading and installing the Toolkit for Linux, and preferably keep the default installation values unchanged, notably the installation path *(`/opt/intel/oneapi` by default)*.
+
+Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
+
+Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI MKL for intel GPUs.
+
+- **Adding support to Nvidia GPUs**
+
+**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
+
+
+**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
+
+```sh
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas
+cmake --build buildWithCublas --config Release
+```
+
+
+3. **Verify installation and environment**
+
+In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
+```sh
+source /opt/intel/oneapi/setvars.sh
+sycl-ls
+```
+
+- **Intel GPU**
+
+When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`ext_oneapi_level_zero:gpu:0`] in the sample output below:
+
+```
+[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
+[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
+```
+
+- **Nvidia GPU**
+
+Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
+```
+[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
+[opencl:cpu:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
+[ext_oneapi_cuda:gpu:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.2]
+```
+
+### II. Build llama.cpp
+
+#### Intel GPU
+```sh
+# Export relevant ENV variables
+source /opt/intel/oneapi/setvars.sh
+
+# Build LLAMA with MKL BLAS acceleration for intel GPU
+
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Option 2: Use FP16
+cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+
+# build all binary
+cmake --build build --config Release -j -v
+```
+
+#### Nvidia GPU
+```sh
+# Export relevant ENV variables
+export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
+
+# Build LLAMA with Nvidia BLAS acceleration through SYCL
+
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Option 2: Use FP16
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+
+# build all binary
+cmake --build build --config Release -j -v
+
+```
+
+### III. Run the inference
+
+1. Retrieve and prepare model
+
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+
+2. Enable oneAPI running environment
+
+```sh
+source /opt/intel/oneapi/setvars.sh
+```
+
+3. List devices information
+
+Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
+
+```sh
+./build/bin/llama-ls-sycl-device
+```
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
+```
+found 2 SYCL devices:
+
+|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
+|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
+|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
+| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
+| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+```
+
+
+4. Launch inference
+
+There are two device selection modes:
+
+- Single device: Use one device target specified by the user.
+- Multiple devices: Automatically choose the devices with the same backend.
+
+In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
+
+| Device selection | Parameter                              |
+|------------------|----------------------------------------|
+| Single device    | --split-mode none --main-gpu DEVICE_ID |
+| Multiple devices | --split-mode layer (default)           |
+
+Examples:
+
+- Use device 0:
+
+```sh
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+```
+or run by script:
+
+```sh
+./examples/sycl/run_llama2.sh 0
+```
+
+- Use multiple devices:
+
+```sh
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+```
+
+Otherwise, you can run the script:
+
+```sh
+./examples/sycl/run_llama2.sh
+```
+
+*Notes:*
+
+- Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
+
+```sh
+detect 1 SYCL GPUs: [0] with top Max compute units:512
+```
+Or
+```sh
+use 1 SYCL GPUs: [0] with Max compute units:512
+```
+
+## Windows
+
+### I. Setup Environment
+
+1. Install GPU driver
+
+Intel GPU drivers instructions guide and download page can be found here: [Get intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+
+2. Install Visual Studio
+
+If you already have a recent version of Microsoft Visual Studio, you can skip this step. Otherwise, please refer to the official download page for [Microsoft Visual Studio](https://visualstudio.microsoft.com/).
+
+3. Install Intel® oneAPI Base toolkit
+
+The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
+
+Please follow the instructions for downloading and installing the Toolkit for Windows, and preferably keep the default installation values unchanged, notably the installation path *(`C:\Program Files (x86)\Intel\oneAPI` by default)*.
+
+Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
+
+b. Enable oneAPI running environment:
+
+- Type "oneAPI" in the search bar, then open the `Intel oneAPI command prompt for Intel 64 for Visual Studio 2022` App.
+
+- On the command prompt, enable the runtime environment with the following:
+```
+"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
+```
+
+c. Verify installation
+
+In the oneAPI command line, run the following to print the available SYCL devices:
+
+```
+sycl-ls
+```
+
+There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
+
+Output (example):
+```
+[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu:1] Intel(R) OpenCL, 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Iris(R) Xe Graphics OpenCL 3.0 NEO  [31.0.101.5186]
+[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Iris(R) Xe Graphics 1.3 [1.3.28044]
+```
+
+4. Install build tools
+
+a. Download & install cmake for Windows: https://cmake.org/download/ (CMake can also be installed from Visual Studio Installer)
+b. The new Visual Studio will install Ninja as default. (If not, please install it manually: https://ninja-build.org/)
+
+
+### II. Build llama.cpp
+
+On the oneAPI command line window, step into the llama.cpp main directory and run the following:
+
+```
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+
+# Option 2: Or FP16
+cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DGGML_SYCL_F16=ON
+
+cmake --build build --config Release -j
+```
+
+Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
+
+Or, use CMake presets to build:
+```sh
+cmake --preset x64-windows-sycl-release
+cmake --build build-x64-windows-sycl-release -j --target llama-cli
+
+cmake -DGGML_SYCL_F16=ON --preset x64-windows-sycl-release
+cmake --build build-x64-windows-sycl-release -j --target llama-cli
+
+cmake --preset x64-windows-sycl-debug
+cmake --build build-x64-windows-sycl-debug -j --target llama-cli
+```
+
+Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+
+*Notes:*
+
+- In case of a minimal experimental setup, the user can build the inference executable only through `cmake --build build --config Release -j --target llama-cli`.
+
+### III. Run the inference
+
+1. Retrieve and prepare model
+
+You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+
+2. Enable oneAPI running environment
+
+On the oneAPI command line window, run the following and step into the llama.cpp directory:
+```
+"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
+```
+
+3. List devices information
+
+Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
+
+```
+build\bin\ls-sycl-device.exe
+```
+
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
+```
+found 2 SYCL devices:
+|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
+|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
+|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
+| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
+| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+
+```
+
+
+4. Launch inference
+
+There are two device selection modes:
+
+- Single device: Use one device assigned by user. Default device id is 0.
+- Multiple devices: Automatically choose the devices with the same backend.
+
+In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
+
+| Device selection | Parameter                              |
+|------------------|----------------------------------------|
+| Single device    | --split-mode none --main-gpu DEVICE_ID |
+| Multiple devices | --split-mode layer (default)           |
+
+Examples:
+
+- Use device 0:
+
+```
+build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
+```
+
+- Use multiple devices:
+
+```
+build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
+```
+Otherwise, run the following wrapper script:
+
+```
+.\examples\sycl\win-run-llama2.bat
+```
+
+Note:
+
+- Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
+
+```sh
+detect 1 SYCL GPUs: [0] with top Max compute units:512
+```
+Or
+```sh
+use 1 SYCL GPUs: [0] with Max compute units:512
+```
+
+## Environment Variable
+
+#### Build
+
+| Name               | Value                             | Function                                    |
+|--------------------|-----------------------------------|---------------------------------------------|
+| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
+| GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
+| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
+| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
+
+#### Runtime
+
+| Name              | Value            | Function                                                                                                                  |
+|-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
+| GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
+| ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.
Recommended to use when --split-mode = layer |
+
+## Known Issues
+
+- `Split-mode:[row]` is not supported.
+
+## Q&A
+
+- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
+
+  - Potential cause: Unavailable oneAPI installation or not set ENV variables.
+  - Solution: Install *oneAPI base toolkit* and enable its ENV through: `source /opt/intel/oneapi/setvars.sh`.
+
+- General compiler error:
+
+  - Remove **build** folder or try a clean-build.
+
+- I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
+
+  Please double-check with `sudo sycl-ls`.
+
+  If it's present in the list, please add video/render group to your user then **logout/login** or restart your system:
+
+  ```
+  sudo usermod -aG render $USER
+  sudo usermod -aG video $USER
+  ```
+  Otherwise, please double-check the GPU driver installation steps.
+
+### **GitHub contribution**:
+Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
+
+## TODO
+
+- Support row layer split for multiple card runs.
diff --git a/llm/llama.cpp/docs/build.md b/llm/llama.cpp/docs/build.md
new file mode 100644
index 0000000000000000000000000000000000000000..8b16d1a35851835d5f582fb68f63c9c491bb382e
--- /dev/null
+++ b/llm/llama.cpp/docs/build.md
@@ -0,0 +1,357 @@
+# Build llama.cpp locally
+
+**To get the Code:**
+
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+In order to build llama.cpp you have four different options.
+
+- Using `make`:
+  - On Linux or MacOS:
+
+      ```bash
+      make
+      ```
+
+  - On Windows (x86/x64 only, arm64 requires cmake):
+
+    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
+    2. Extract `w64devkit` on your pc.
+    3. Run `w64devkit.exe`.
+    4. Use the `cd` command to reach the `llama.cpp` folder.
+    5. From here you can run:
+        ```bash
+        make
+        ```
+
+  - Notes:
+    - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`.
+    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
+    - For faster repeated compilation, install [ccache](https://ccache.dev/).
+    - For debug builds, run `make LLAMA_DEBUG=1`
+
+- Using `CMake`:
+
+  ```bash
+  cmake -B build
+  cmake --build build --config Release
+  ```
+
+  **Notes**:
+
+    - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
+    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
+    - For faster repeated compilation, install [ccache](https://ccache.dev/).
+    - For debug builds, there are two cases:
+
+      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
+
+      ```bash
+      cmake -B build -DCMAKE_BUILD_TYPE=Debug
+      cmake --build build
+      ```
+
+      2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
+
+      ```bash
+      cmake -B build -G "Xcode"
+      cmake --build build --config Debug
+      ```
+    - Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
+      - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
+        - Tab Workload: Desktop-development with C++
+        - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
+      - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
+      - For Windows on ARM (arm64, WoA) build with:
+        ```bash
+        cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
+        cmake --build build-arm64-windows-llvm-release
+        ```
+        Note: Building for arm64 could also be done just with MSVC (with the build-arm64-windows-MSVC preset, or the standard CMake build instructions). But MSVC does not support inline ARM assembly-code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
+
+-   Using `gmake` (FreeBSD):
+
+    1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
+    2. Add your user to **video** group
+    3. Install compilation dependencies.
+
+        ```bash
+        sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
+
+        gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
+        ```
+
+## Metal Build
+
+On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
+To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option.
+
+When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
+argument.
+
+## BLAS Build
+
+Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
+
+### Accelerate Framework:
+
+This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
+
+### OpenBLAS:
+
+This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
+
+- Using `make`:
+  - On Linux:
+    ```bash
+    make GGML_OPENBLAS=1
+    ```
+
+  - On Windows:
+
+    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
+    2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
+    3. Extract `w64devkit` on your pc.
+    4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
+    5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
+    6. Run `w64devkit.exe`.
+    7. Use the `cd` command to reach the `llama.cpp` folder.
+    8. From here you can run:
+
+        ```bash
+        make GGML_OPENBLAS=1
+        ```
+
+- Using `CMake` on Linux:
+
+    ```bash
+    cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+    cmake --build build --config Release
+    ```
+
+### BLIS
+
+Check [BLIS.md](./backend/BLIS.md) for more information.
+
+### SYCL
+
+SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
+
+llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
+
+For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
+
+### Intel oneMKL
+
+Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
+
+- Using manual oneAPI installation:
+  By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
+    ```bash
+    source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-basekit docker image, only required for manual installation
+    cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON
+    cmake --build build --config Release
+    ```
+
+- Using oneAPI docker image:
+  If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
+
+Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
+
+### CUDA
+
+This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+
+For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
+
+- Using `make`:
+  ```bash
+  make GGML_CUDA=1
+  ```
+- Using `CMake`:
+
+  ```bash
+  cmake -B build -DGGML_CUDA=ON
+  cmake --build build --config Release
+  ```
+
+The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
+
+The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
+
+The following compilation options are also available to tweak performance:
+
+| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
+|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| GGML_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
+| GGML_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
+| GGML_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
+| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
+| GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
+| GGML_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
+| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
+| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
+
+### MUSA
+
+- Using `make`:
+  ```bash
+  make GGML_MUSA=1
+  ```
+- Using `CMake`:
+
+  ```bash
+  cmake -B build -DGGML_MUSA=ON
+  cmake --build build --config Release
+  ```
+
+### hipBLAS
+
+This provides BLAS acceleration on HIP-supported AMD GPUs.
+Make sure to have ROCm installed.
+You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
+
+- Using `make`:
+  ```bash
+  make GGML_HIPBLAS=1
+  ```
+- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
+  ```bash
+  HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
+      cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      && cmake --build build --config Release -- -j 16
+  ```
+  On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
+  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
+  Note that if you get the following error:
+  ```
+  clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
+  ```
+  Try searching for a directory under `HIP_PATH` that contains the file
+  `oclc_abi_version_400.bc`. Then, add the following to the start of the
+  command: `HIP_DEVICE_LIB_PATH=`, so something
+  like:
+  ```bash
+  HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
+  HIP_DEVICE_LIB_PATH= \
+      cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      && cmake --build build -- -j 16
+  ```
+
+- Using `make` (example for target gfx1030, build with 16 CPU threads):
+  ```bash
+  make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
+  ```
+
+- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
+  ```bash
+  set PATH=%HIP_PATH%\bin;%PATH%
+  cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+  cmake --build build
+  ```
+  Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+  Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
+
+
+The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
+If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
+The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
+
+| Option                 | Legal values           | Default | Description                                                                                                                                                                                                                                    |
+|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| GGML_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+| GGML_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
+| GGML_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
+
+### Vulkan
+
+**Windows**
+
+#### w64devkit
+
+Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
+
+Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
+
+Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
+```sh
+SDK_VERSION=1.3.283.0
+cp /VulkanSDK/$SDK_VERSION/Bin/glslc.exe $W64DEVKIT_HOME/bin/
+cp /VulkanSDK/$SDK_VERSION/Lib/vulkan-1.lib $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/
+cp -r /VulkanSDK/$SDK_VERSION/Include/* $W64DEVKIT_HOME/x86_64-w64-mingw32/include/
+cat > $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/pkgconfig/vulkan.pc < `
+
+It will then build & run in the debugger for you.
+
+To just execute a test and get back a PASS or FAIL message run:
+
+```bash
+./scripts/debug-test.sh test-tokenizer
+```
+
+To test in GDB use the `-g` flag to enable gdb test mode.
+
+```bash
+./scripts/debug-test.sh -g test-tokenizer
+
+# Once in the debugger, i.e. at the chevrons prompt, setting a breakpoint could be as follows:
+>>> b main
+```
+
+To speed up the testing loop, if you know your test number you can just run it similar to below:
+
+```bash
+./scripts/debug-test.sh test 23
+```
+
+For further reference use `debug-test.sh -h` to print help.
+
+ 
+
+### How does the script work?
+If you want to be able to use the concepts contained in the script separately, the important ones are briefly outlined below.
+
+#### Step 1: Reset and Setup folder context
+
+From base of this repository, let's create `build-ci-debug` as our build context.
+
+```bash
+rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
+```
+
+#### Step 2: Setup Build Environment and Compile Test Binaries
+
+Setup and trigger a build under debug mode. You may adapt the arguments as needed, but in this case these are sane defaults.
+
+```bash
+cmake -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CUDA=1 -DLLAMA_FATAL_WARNINGS=ON ..
+make -j
+```
+
+#### Step 3: Find all tests available that matches REGEX
+
+The output of this command will give you the command & arguments needed to run GDB.
+
+* `-R test-tokenizer` : looks for all the test files named `test-tokenizer*` (R=Regex)
+* `-N` : "show-only" disables test execution & shows test commands that you can feed to GDB.
+* `-V` : Verbose Mode
+
+```bash
+ctest -R "test-tokenizer" -V -N
+```
+
+This may return output similar to below (focusing on key lines to pay attention to):
+
+```bash
+...
+1: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+1: Working Directory: .
+Labels: main
+  Test  #1: test-tokenizer-0-llama-spm
+...
+4: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-falcon.gguf"
+4: Working Directory: .
+Labels: main
+  Test  #4: test-tokenizer-0-falcon
+...
+```
+
+#### Step 4: Identify Test Command for Debugging
+
+So for test #1 above we can tell these two pieces of relevant information:
+* Test Binary: `~/llama.cpp/build-ci-debug/bin/test-tokenizer-0`
+* Test GGUF Model: `~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf`
+
+#### Step 5: Run GDB on test command
+
+Based on the ctest 'test command' report above we can then run a gdb session via this command below:
+
+```bash
+gdb --args ${Test Binary} ${Test GGUF Model}
+```
+
+Example:
+
+```bash
+gdb --args ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+```
diff --git a/llm/llama.cpp/docs/development/llama-star/idea-arch.key b/llm/llama.cpp/docs/development/llama-star/idea-arch.key
new file mode 100644
index 0000000000000000000000000000000000000000..3e068e7075c2ebb53a270b57c51e9281618d8f29
Binary files /dev/null and b/llm/llama.cpp/docs/development/llama-star/idea-arch.key differ
diff --git a/llm/llama.cpp/docs/development/llama-star/idea-arch.pdf b/llm/llama.cpp/docs/development/llama-star/idea-arch.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4fa92c71dc4c511378c628113b7817e583053758
Binary files /dev/null and b/llm/llama.cpp/docs/development/llama-star/idea-arch.pdf differ
diff --git a/llm/llama.cpp/docs/development/token_generation_performance_tips.md b/llm/llama.cpp/docs/development/token_generation_performance_tips.md
new file mode 100644
index 0000000000000000000000000000000000000000..41b7232c976b3a86fe7509c078bcedafd014a9e1
--- /dev/null
+++ b/llm/llama.cpp/docs/development/token_generation_performance_tips.md
@@ -0,0 +1,40 @@
+# Token generation performance troubleshooting
+
+## Verifying that the model is running on the GPU with CUDA
+Make sure you compiled llama with the correct env variables according to [this guide](/docs/build.md#cuda), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
+```shell
+./llama-cli -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
+```
+
+When running llama, before it starts the inference work, it will output diagnostic information that shows whether cuBLAS is offloading work to the GPU. Look for these lines:
+```shell
+llama_model_load_internal: [cublas] offloading 60 layers to GPU
+llama_model_load_internal: [cublas] offloading output layer to GPU
+llama_model_load_internal: [cublas] total VRAM used: 17223 MB
+... rest of inference
+```
+
+If you see these lines, then the GPU is being used.
+
+## Verifying that the CPU is not oversaturated
+llama accepts a `-t N` (or `--threads N`) parameter. It's extremely important that this parameter is not too large. If your token generation is extremely slow, try setting this number to 1. If this significantly improves your token generation speed, then your CPU is being oversaturated and you need to explicitly set this parameter to the number of the physical CPU cores on your machine (even if you utilize a GPU). If in doubt, start with 1 and double the amount until you hit a performance bottleneck, then scale the number down.
+
+# Example of runtime flags effect on inference speed benchmark
+These runs were tested on the following machine:
+GPU: A6000 (48GB VRAM)
+CPU: 7 physical cores
+RAM: 32GB
+
+Model: `TheBloke_Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.q4_0.gguf` (30B parameters, 4bit quantization, GGML)
+
+Run command: `./llama-cli -m "path/to/model.gguf" -p "An extremely detailed description of the 10 best ethnic dishes will follow, with recipes: " -n 1000 [additional benchmark flags]`
+
+Result:
+
+| command | tokens/second (higher is better) |
+| - | - |
+| -ngl 2000000 | N/A (less than 0.1) |
+| -t 7 | 1.7 |
+| -t 1 -ngl 2000000 | 5.5 |
+| -t 7 -ngl 2000000 | 8.7 |
+| -t 4 -ngl 2000000 | 9.1 |
diff --git a/llm/llama.cpp/docs/docker.md b/llm/llama.cpp/docs/docker.md
new file mode 100644
index 0000000000000000000000000000000000000000..d8922d77d9fe55cb9a56c8aa768acf98ed55ddb4
--- /dev/null
+++ b/llm/llama.cpp/docs/docker.md
@@ -0,0 +1,86 @@
+# Docker
+
+## Prerequisites
+* Docker must be installed and running on your system.
+* Create a folder to store big models & intermediate files (ex. /llama/models)
+
+## Images
+We have three Docker images available for this project:
+
+1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+
+Additionally, there the following images, similar to the above:
+
+- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
+
+## Usage
+
+The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
+
+Replace `/path/to/models` below with the actual path where you downloaded the models.
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
+```
+
+On completion, you are ready to play!
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+```
+
+or with a light image:
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+```
+
+or with a server image:
+
+```bash
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+```
+
+## Docker With CUDA
+
+Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
+
+## Building Docker locally
+
+```bash
+docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
+docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
+```
+
+You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
+
+The defaults are:
+
+- `CUDA_VERSION` set to `11.7.1`
+- `CUDA_DOCKER_ARCH` set to `all`
+
+The resulting images, are essentially the same as the non-CUDA images:
+
+1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
+2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
+3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
+
+## Usage
+
+After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
+
+```bash
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
+```
diff --git a/llm/llama.cpp/docs/install.md b/llm/llama.cpp/docs/install.md
new file mode 100644
index 0000000000000000000000000000000000000000..10a568506835beef9ba1c47253175e6007fd4bf0
--- /dev/null
+++ b/llm/llama.cpp/docs/install.md
@@ -0,0 +1,39 @@
+# Install pre-built version of llama.cpp
+
+## Homebrew
+
+On Mac and Linux, the homebrew package manager can be used via
+
+```sh
+brew install llama.cpp
+```
+The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
+
+## Nix
+
+On Mac and Linux, the Nix package manager can be used via
+
+```sh
+nix profile install nixpkgs#llama-cpp
+```
+For flake enabled installs.
+
+Or
+
+```sh
+nix-env --file '' --install --attr llama-cpp
+```
+
+For non-flake enabled installs.
+
+This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
+
+## Flox
+
+On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
+
+```sh
+flox install llama-cpp
+```
+
+Flox follows the nixpkgs build of llama.cpp.
diff --git a/llm/llama.cpp/examples/CMakeLists.txt b/llm/llama.cpp/examples/CMakeLists.txt
index b40ee4ccb2ec1d32b351d3b9d07e5396cf8caa3f..67b3d277478508c5aae22109cb0a975710164c27 100644
--- a/llm/llama.cpp/examples/CMakeLists.txt
+++ b/llm/llama.cpp/examples/CMakeLists.txt
@@ -12,44 +12,44 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
+    add_subdirectory(cvector-generator)
     add_subdirectory(baby-llama)
-    add_subdirectory(batched)
     add_subdirectory(batched-bench)
-    add_subdirectory(beam-search)
+    add_subdirectory(batched)
     add_subdirectory(benchmark)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
-    add_subdirectory(finetune)
-    add_subdirectory(gritlm)
+    add_subdirectory(export-lora)
+    add_subdirectory(gbnf-validator)
+    add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
+    add_subdirectory(gguf)
+    add_subdirectory(gritlm)
+    add_subdirectory(imatrix)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
-    if (LLAMA_SYCL)
-        add_subdirectory(sycl)
-    endif()
+    add_subdirectory(lookahead)
+    add_subdirectory(lookup)
     add_subdirectory(main)
-    add_subdirectory(tokenize)
     add_subdirectory(parallel)
+    add_subdirectory(passkey)
     add_subdirectory(perplexity)
-    add_subdirectory(quantize)
     add_subdirectory(quantize-stats)
+    add_subdirectory(quantize)
     add_subdirectory(retrieval)
-    add_subdirectory(save-load-state)
-    add_subdirectory(simple)
-    add_subdirectory(passkey)
-    add_subdirectory(speculative)
-    add_subdirectory(lookahead)
-    add_subdirectory(lookup)
-    add_subdirectory(gguf)
-    add_subdirectory(train-text-from-scratch)
-    add_subdirectory(imatrix)
+    if (GGML_RPC)
+        add_subdirectory(rpc)
+    endif()
     if (LLAMA_BUILD_SERVER)
-        add_subdirectory(server)
+    add_subdirectory(server)
     endif()
-    add_subdirectory(export-lora)
-    if (LLAMA_RPC)
-        add_subdirectory(rpc)
+    if (GGML_SYCL)
+        add_subdirectory(sycl)
     endif()
+    add_subdirectory(save-load-state)
+    add_subdirectory(simple)
+    add_subdirectory(speculative)
+    add_subdirectory(tokenize)
 endif()
diff --git a/llm/llama.cpp/examples/Miku.sh b/llm/llama.cpp/examples/Miku.sh
index b9174b4e6e12668e244d56f37847aa40c1b230c0..0f6c8c8787107e0929b6ad75cacaf2b8d4a78a14 100644
--- a/llm/llama.cpp/examples/Miku.sh
+++ b/llm/llama.cpp/examples/Miku.sh
@@ -22,7 +22,7 @@ if [ -n "$N_THREAD" ]; then
     GEN_OPTIONS+=(--threads "$N_THREAD")
 fi
 
-./main "${GEN_OPTIONS[@]}" \
+./llama-cli "${GEN_OPTIONS[@]}" \
     --model "$MODEL" \
     --in-prefix " " \
     --in-suffix "${AI_NAME}:" \
diff --git a/llm/llama.cpp/examples/baby-llama/CMakeLists.txt b/llm/llama.cpp/examples/baby-llama/CMakeLists.txt
index 7b70227a525e1a158cd775e089674b00c8c4d8d1..71b82105c886369fa80f38f2e93703623e3a7e64 100644
--- a/llm/llama.cpp/examples/baby-llama/CMakeLists.txt
+++ b/llm/llama.cpp/examples/baby-llama/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET baby-llama)
+set(TARGET llama-baby-llama)
 add_executable(${TARGET} baby-llama.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/baby-llama/baby-llama.cpp b/llm/llama.cpp/examples/baby-llama/baby-llama.cpp
index bf0125e75374645365552419e7ceb5d2b97349d3..aca332e9464d2f7eaf461b9d2a374f2fac0b51f2 100644
--- a/llm/llama.cpp/examples/baby-llama/baby-llama.cpp
+++ b/llm/llama.cpp/examples/baby-llama/baby-llama.cpp
@@ -1,7 +1,6 @@
 #include "ggml.h"
 #include "train.h"
 
-#include 
 #include 
 #include 
 #include 
@@ -522,8 +521,8 @@ static struct ggml_tensor * forward(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, 1]
             // Kcur shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0);
 
             // store key and value to memory
             {
@@ -759,8 +758,8 @@ static struct ggml_tensor * forward_batch(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, n_batch]
             // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0);
             assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
             assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
 
@@ -1056,7 +1055,7 @@ static struct ggml_tensor * forward_lora(
                                                         model->layers[il].wqb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            KQ_pos, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0);
             struct ggml_tensor * Kcur = ggml_rope(ctx0,
                                             ggml_reshape_3d(ctx0,
                                                 ggml_mul_mat(ctx0,
@@ -1065,7 +1064,7 @@ static struct ggml_tensor * forward_lora(
                                                         model->layers[il].wkb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            KQ_pos, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0);
 
             // store key and value to memory
             {
diff --git a/llm/llama.cpp/examples/base-translate.sh b/llm/llama.cpp/examples/base-translate.sh
index 00dedd0df7ddebf22582c66329e17201426448ff..103a52f55e6dbaf8024c2a2c96d3f796d99f2609 100644
--- a/llm/llama.cpp/examples/base-translate.sh
+++ b/llm/llama.cpp/examples/base-translate.sh
@@ -58,4 +58,4 @@ echo "$2
 model=$1
 
 # generate the most likely continuation until the string "===" is found
-./main -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs
+./llama-cli -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs
diff --git a/llm/llama.cpp/examples/batched-bench/CMakeLists.txt b/llm/llama.cpp/examples/batched-bench/CMakeLists.txt
index 40a032c514d11de6907f78f6656782539ff42720..959acaeeebc380eb406e18e72c536a01e3c4d8ba 100644
--- a/llm/llama.cpp/examples/batched-bench/CMakeLists.txt
+++ b/llm/llama.cpp/examples/batched-bench/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET batched-bench)
+set(TARGET llama-batched-bench)
 add_executable(${TARGET} batched-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/batched-bench/README.md b/llm/llama.cpp/examples/batched-bench/README.md
index bf951baf7f096da79c1b0b509cabbd9b08c60a45..4a07fe6bbf268c872ddd66f2983dafd88529fa42 100644
--- a/llm/llama.cpp/examples/batched-bench/README.md
+++ b/llm/llama.cpp/examples/batched-bench/README.md
@@ -10,16 +10,16 @@ There are 2 modes of operation:
 - `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
 
 ```bash
-./batched-bench MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] [MMQ]   
+./llama-batched-bench -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]
 
 # LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
-./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 2048 512 0 99
+./llama-batched-bench -m ./models/llama-7b/ggml-model-f16.gguf -c 16384 -b 2048 -ub 512 -ngl 99
 
 # LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
-./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 2048 512 1 99
+./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 16384 -b 2048 -ub 512 -ngl 99 -pps
 
 # custom set of batches
-./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 512 512 0 999 0 128,256,512 128,256 1,2,4,8,16,32
+./llama-batched-bench -m ./models/llama-7b/ggml-model-q8_0.gguf -c 2048 -b 512 -ub 512 -ngl 999 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32
 ```
 
 ## Sample results
diff --git a/llm/llama.cpp/examples/batched-bench/batched-bench.cpp b/llm/llama.cpp/examples/batched-bench/batched-bench.cpp
index 2924d8116f44f16e3a40b2134a49ea6e1d7542a4..25e7c775a0095d9b0cd6100e407c692bb1cad178 100644
--- a/llm/llama.cpp/examples/batched-bench/batched-bench.cpp
+++ b/llm/llama.cpp/examples/batched-bench/batched-bench.cpp
@@ -28,67 +28,27 @@ static std::vector parse_list(char * p) {
     return ret;
 }
 
-int main(int argc, char ** argv) {
-    gpt_params params;
-
-    if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [FATTN] [IS_PP_SHARED] [NGL]   \n" , argv[0]);
-        printf("  ,  and PL are comma-separated lists of numbers without spaces\n\n");
-        printf("  example: %s ggml-model-f16.gguf 2048 2048 512 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
-        return 1 ;
-    }
-
-    int n_kv_max     = 2048;
-    int n_batch      = 2048;
-    int n_ubatch     = 512;
-    bool flash_attn  = false;
-    int is_pp_shared = 0;
-    int n_gpu_layers = 0;
-
-    std::vector n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
-    std::vector n_tg = { 128, 256, };
-    std::vector n_pl = { 1, 2, 4, 8, 16, 32, };
-    //std::vector n_pl = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, };
-
-    if (argc >= 2) {
-        params.model = argv[1];
-    }
-
-    if (argc >= 3) {
-        n_kv_max = std::atoi(argv[2]);
-    }
-
-    if (argc >= 4) {
-        n_batch = std::atoi(argv[3]);
-    }
-
-    if (argc >= 5) {
-        n_ubatch = std::atoi(argv[4]);
-    }
-
-    if (argc >= 6) {
-        flash_attn = std::atoi(argv[5]);
-    }
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
 
-    if (argc >= 7) {
-        is_pp_shared = std::atoi(argv[6]);
-    }
+    LOG_TEE("\nexample usage:\n");
+    LOG_TEE("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
+    LOG_TEE("\n");
+}
 
-    if (argc >= 8) {
-        n_gpu_layers = std::atoi(argv[7]);
-    }
+int main(int argc, char ** argv) {
+    gpt_params params;
 
-    if (argc >= 9) {
-        n_pp = parse_list(argv[8]);
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
     }
 
-    if (argc >= 10) {
-        n_tg = parse_list(argv[9]);
-    }
+    int is_pp_shared = params.is_pp_shared;
 
-    if (argc >= 11) {
-        n_pl = parse_list(argv[10]);
-    }
+    std::vector n_pp = params.n_pp;
+    std::vector n_tg = params.n_tg;
+    std::vector n_pl = params.n_pl;
 
     // init LLM
 
@@ -97,12 +57,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_default_params();
-
-    const std::vector t_split(llama_max_devices(), 0.0f);
-
-    model_params.n_gpu_layers = n_gpu_layers;
-    model_params.tensor_split = t_split.data();
+    llama_model_params model_params = llama_model_params_from_gpt_params(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -111,19 +66,10 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_context_params ctx_params = llama_context_default_params();
-
-    ctx_params.seed       = 1234;
-    ctx_params.n_ctx      = n_kv_max;
-    ctx_params.n_batch    = n_batch;
-    ctx_params.n_ubatch   = n_ubatch;
-    ctx_params.flash_attn = flash_attn;
-
-    ctx_params.n_threads       = params.n_threads;
-    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
     // ensure enough sequences are available
-    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
@@ -132,6 +78,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const int32_t n_kv_max = llama_n_ctx(ctx);
+
     llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
 
     // decode in batches of ctx_params.n_batch tokens
@@ -175,7 +123,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, flash_attn, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
diff --git a/llm/llama.cpp/examples/batched.swift/Makefile b/llm/llama.cpp/examples/batched.swift/Makefile
index 2afb24fb85a1a3ee25a0130f1cbe7b70b0df48cf..1f9156e583fdda26611fcc9c53dfe57517a50da8 100644
--- a/llm/llama.cpp/examples/batched.swift/Makefile
+++ b/llm/llama.cpp/examples/batched.swift/Makefile
@@ -1,6 +1,6 @@
 .PHONY: build
 
 build:
-	xcodebuild -scheme batched_swift -destination "generic/platform=macOS" -derivedDataPath build
-	rm -f ./batched_swift
-	ln -s ./build/Build/Products/Debug/batched_swift ./batched_swift
+	xcodebuild -scheme llama-batched-swift -destination "generic/platform=macOS" -derivedDataPath build
+	rm -f ./llama-batched-swift
+	ln -s ./build/Build/Products/Debug/llama-batched-swift ./llama-batched-swift
diff --git a/llm/llama.cpp/examples/batched.swift/Package.swift b/llm/llama.cpp/examples/batched.swift/Package.swift
index 826491defd8631c67292fb6d97f2117bf4b690d2..7e8afd0843c5b646e088e778d9947f24a76b7dee 100644
--- a/llm/llama.cpp/examples/batched.swift/Package.swift
+++ b/llm/llama.cpp/examples/batched.swift/Package.swift
@@ -4,7 +4,7 @@
 import PackageDescription
 
 let package = Package(
-    name: "batched_swift",
+    name: "llama-batched-swift",
     platforms: [.macOS(.v12)],
     dependencies: [
         .package(name: "llama", path: "../../"),
@@ -13,7 +13,7 @@ let package = Package(
         // Targets are the basic building blocks of a package, defining a module or a test suite.
         // Targets can depend on other targets in this package and products from dependencies.
         .executableTarget(
-            name: "batched_swift",
+            name: "llama-batched-swift",
             dependencies: ["llama"],
             path: "Sources",
             linkerSettings: [.linkedFramework("Foundation"), .linkedFramework("AppKit")]
diff --git a/llm/llama.cpp/examples/batched.swift/README.md b/llm/llama.cpp/examples/batched.swift/README.md
index 4c2721fe85b005d64b24baaf8775811b87dd1735..7f2e2fcdcf4a7767f84568b1acf74a118b44f0f3 100644
--- a/llm/llama.cpp/examples/batched.swift/README.md
+++ b/llm/llama.cpp/examples/batched.swift/README.md
@@ -1,4 +1,4 @@
 This is a swift clone of `examples/batched`.
 
 $ `make`
-$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`
+$ `./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]`
diff --git a/llm/llama.cpp/examples/batched.swift/Sources/main.swift b/llm/llama.cpp/examples/batched.swift/Sources/main.swift
index dbbd06da5818364178a2fcc2c986100a5c2a2f72..616494d2d841d099af6332f34765121eff81a951 100644
--- a/llm/llama.cpp/examples/batched.swift/Sources/main.swift
+++ b/llm/llama.cpp/examples/batched.swift/Sources/main.swift
@@ -229,7 +229,7 @@ private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
 
 private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
     var result = [CChar](repeating: 0, count: 8)
-    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), false)
+    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), 0, false)
     if nTokens < 0 {
         let actualTokensCount = -Int(nTokens)
         result = .init(repeating: 0, count: actualTokensCount)
@@ -238,6 +238,7 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
             token,
             &result,
             Int32(result.count),
+            0,
             false
         )
         assert(check == actualTokensCount)
diff --git a/llm/llama.cpp/examples/batched/CMakeLists.txt b/llm/llama.cpp/examples/batched/CMakeLists.txt
index 6aa178d4d5911fb731b90218b2c2d9032dffe451..77e33343b667389a12932aef6cf3a313ea619748 100644
--- a/llm/llama.cpp/examples/batched/CMakeLists.txt
+++ b/llm/llama.cpp/examples/batched/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET batched)
+set(TARGET llama-batched)
 add_executable(${TARGET} batched.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/batched/README.md b/llm/llama.cpp/examples/batched/README.md
index 5d730331769fb1a950ce94fc8e10c0b40ba36576..6013aab01fddc2866283e577e90b8f06dbacc770 100644
--- a/llm/llama.cpp/examples/batched/README.md
+++ b/llm/llama.cpp/examples/batched/README.md
@@ -3,7 +3,7 @@
 The example demonstrates batched generation from a given prompt
 
 ```bash
-./batched ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is" 4
+./llama-batched -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is" -np 4
 
 ...
 
diff --git a/llm/llama.cpp/examples/batched/batched.cpp b/llm/llama.cpp/examples/batched/batched.cpp
index 591bc6e57645ccb09da3133fa2e50e62904daaa4..53fbfb0a8cf2ae88775c50506b2c6bef9d40e2af 100644
--- a/llm/llama.cpp/examples/batched/batched.cpp
+++ b/llm/llama.cpp/examples/batched/batched.cpp
@@ -7,48 +7,31 @@
 #include 
 #include 
 
-int main(int argc, char ** argv) {
-    gpt_params params;
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
 
-    if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN] [NGL]\n" , argv[0]);
-        return 1 ;
-    }
-
-    // number of parallel batches
-    int n_parallel = 1;
-
-    // total length of the sequences including the prompt
-    int n_len = 32;
-
-    // number of layers to offload to the GPU
-    int n_gpu_layers = 0;
+    LOG_TEE("\nexample usage:\n");
+    LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
+    LOG_TEE("\n");
+}
 
-    if (argc >= 2) {
-        params.model = argv[1];
-    }
+int main(int argc, char ** argv) {
+    gpt_params params;
 
-    if (argc >= 3) {
-        params.prompt = argv[2];
-    }
+    params.prompt = "Hello my name is";
+    params.n_predict = 32;
 
-    if (argc >= 4) {
-        n_parallel = std::atoi(argv[3]);
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
     }
 
-    if (argc >= 5) {
-        n_len = std::atoi(argv[4]);
-    }
 
-    if (argc >= 6) {
-        n_gpu_layers = std::atoi(argv[5]);
-    }
-
-    if (params.prompt.empty()) {
-        params.prompt = "Hello my name is";
-    }
+    // number of parallel batches
+    int n_parallel = params.n_parallel;
 
-    string_process_escapes(params.prompt);
+    // total length of the sequences including the prompt
+    int n_predict = params.n_predict;
 
     // init LLM
 
@@ -57,9 +40,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_default_params();
-
-    model_params.n_gpu_layers = n_gpu_layers;
+    llama_model_params model_params = llama_model_params_from_gpt_params(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -73,18 +54,14 @@ int main(int argc, char ** argv) {
     std::vector tokens_list;
     tokens_list = ::llama_tokenize(model, params.prompt, true);
 
-    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
+    const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_default_params();
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
-    ctx_params.seed  = 1234;
     ctx_params.n_ctx   = n_kv_req;
-    ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_seq_max       = n_parallel;
-    ctx_params.n_threads       = params.n_threads;
-    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    ctx_params.n_batch = std::max(n_predict, n_parallel);
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
@@ -93,9 +70,9 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    const int n_ctx    = llama_n_ctx(ctx);
+    const int n_ctx = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
@@ -116,14 +93,34 @@ int main(int argc, char ** argv) {
 
     // create a llama_batch
     // we use this object to submit token data for decoding
-    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1);
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
+
+    std::vector seq_ids(n_parallel, 0);
+    for (int32_t i = 0; i < n_parallel; ++i) {
+        seq_ids[i] = i;
+    }
 
     // evaluate the initial prompt
     for (size_t i = 0; i < tokens_list.size(); ++i) {
-        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
+        llama_batch_add(batch, tokens_list[i], i, seq_ids, false);
     }
     GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
 
+    if (llama_model_has_encoder(model)) {
+        if (llama_encode(ctx, batch)) {
+            LOG_TEE("%s : failed to eval\n", __func__);
+            return 1;
+        }
+
+        llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+        if (decoder_start_token_id == -1) {
+            decoder_start_token_id = llama_token_bos(model);
+        }
+
+        llama_batch_clear(batch);
+        llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
+    }
+
     // llama_decode will output logits only for the last token of the prompt
     batch.logits[batch.n_tokens - 1] = true;
 
@@ -132,11 +129,11 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    // assign the system KV cache to all parallel sequences
-    // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
-    for (int32_t i = 1; i < n_parallel; ++i) {
-        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
-    }
+    //// assign the system KV cache to all parallel sequences
+    //// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
+    //for (int32_t i = 1; i < n_parallel; ++i) {
+    //    llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+    //}
 
     if (n_parallel > 1) {
         LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
@@ -156,7 +153,7 @@ int main(int argc, char ** argv) {
 
     const auto t_main_start = ggml_time_us();
 
-    while (n_cur <= n_len) {
+    while (n_cur <= n_predict) {
         // prepare the next batch
         llama_batch_clear(batch);
 
@@ -192,7 +189,7 @@ int main(int argc, char ** argv) {
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of generation? -> mark the stream as finished
-            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
+            if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {
diff --git a/llm/llama.cpp/examples/benchmark/CMakeLists.txt b/llm/llama.cpp/examples/benchmark/CMakeLists.txt
index 2bb47bab5a868b7a54640543ae3acd6da6ddc41a..34a58cc02abaf8d75428241085b561cff5c7d6bf 100644
--- a/llm/llama.cpp/examples/benchmark/CMakeLists.txt
+++ b/llm/llama.cpp/examples/benchmark/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET benchmark)
+set(TARGET llama-bench-matmult)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/chat-13B.bat b/llm/llama.cpp/examples/chat-13B.bat
new file mode 100644
index 0000000000000000000000000000000000000000..c5c8ac6efa81a552725538648592e3fc1563e1fa
--- /dev/null
+++ b/llm/llama.cpp/examples/chat-13B.bat
@@ -0,0 +1,57 @@
+@setlocal disabledelayedexpansion enableextensions
+@echo off
+
+cd /d "%~dp0.."
+if not "%errorlevel%"=="0" (
+    echo Unable to change directory.
+    pause
+    exit /b 1
+)
+
+if not defined MODEL set "MODEL=models\13B\ggml-model-q4_0.bin"
+if not defined USER_NAME set "USER_NAME=User"
+if not defined AI_NAME set "AI_NAME=ChatLLaMa"
+rem Adjust to the number of CPU cores you want to use.
+rem if not defined N_THREAD set "N_THREAD=8"
+rem Number of tokens to predict (made it larger than default because we want a long interaction)
+if not defined N_PREDICTS set "N_PREDICTS=2048"
+if not defined GEN_OPTIONS set "GEN_OPTIONS=--ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647"
+
+rem Default main script paths
+set "DEFAULT_MAIN_SCRIPT_PATHS=main.exe build\bin\main.exe"
+
+rem Get main script path from command line arguments
+set "MAIN_SCRIPT_PATH=%~1"
+
+rem If the main script path was not specified, try the default paths
+if not defined MAIN_SCRIPT_PATH (
+    for %%i in (%DEFAULT_MAIN_SCRIPT_PATHS%) do (
+        if exist "%%i" set "MAIN_SCRIPT_PATH=%%i"
+    )
+)
+
+rem If the main script path was not found, tell the user how to specify it
+if not defined MAIN_SCRIPT_PATH (
+    echo The main script could not be found. Please provide the path to the main script as 1st argument to this script, or place the main script in one of the default locations:
+    echo %DEFAULT_MAIN_SCRIPT_PATHS%
+    pause
+    exit /b 1
+)
+
+rem Default context, feel free to edit it
+set "PROMPT_TEXT=Text transcript of a never ending dialog, where %USER_NAME% interacts with an AI assistant named %AI_NAME%. %AI_NAME% is helpful, kind, honest, friendly, good at writing and never fails to answer %USER_NAME%'s requests immediately and with details and precision. There are no annotations like (30 seconds passed...) or (to himself), just what %USER_NAME% and %AI_NAME% say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. The transcript only includes text, it does not include markup like HTML and Markdown."
+
+rem Set a temporary variable if N_THREAD is set
+if defined N_THREAD (
+    set "_N_THREAD=--threads %N_THREAD%"
+) else (
+    set "_N_THREAD="
+)
+
+rem Run the script
+echo "%MAIN_SCRIPT_PATH%" %GEN_OPTIONS% %_N_THREAD% ^
+  --model "%MODEL%" ^
+  --n_predict %N_PREDICTS% ^
+  --color --interactive ^
+  --reverse-prompt "%USER_NAME%:" ^
+  --prompt "%PROMPT_TEXT%"
diff --git a/llm/llama.cpp/examples/chat-13B.sh b/llm/llama.cpp/examples/chat-13B.sh
index 35c089d57d253ae828322f550f08f102f1aed59d..1828903c31670786faa88aaa8e047b5826c2cf4d 100644
--- a/llm/llama.cpp/examples/chat-13B.sh
+++ b/llm/llama.cpp/examples/chat-13B.sh
@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
      $PROMPT_TEMPLATE > $PROMPT_FILE
 
 # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
-./main $GEN_OPTIONS \
+./llama-cli $GEN_OPTIONS \
   --model "$MODEL" \
   --threads "$N_THREAD" \
   --n_predict "$N_PREDICTS" \
diff --git a/llm/llama.cpp/examples/chat-persistent.sh b/llm/llama.cpp/examples/chat-persistent.sh
index 22f5b83d3da06b0bf45049d0d3963d322cf3d1f5..d9cab9836482edc75f01186e0b25eab2eeb21a9f 100644
--- a/llm/llama.cpp/examples/chat-persistent.sh
+++ b/llm/llama.cpp/examples/chat-persistent.sh
@@ -62,7 +62,7 @@ fi
 if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
     echo 'Prompt cache does not exist, building...'
     # Default batch_size to 64 here for better user feedback during initial prompt processing
-    ./main 2>>"$LOG" \
+    ./llama-cli 2>>"$LOG" \
         --batch_size 64 \
         "${OPTS[@]}" \
         --prompt-cache "$PROMPT_CACHE_FILE" \
@@ -109,13 +109,13 @@ while read -e line; do
 
     printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
 
-    ./main 2>>"$LOG" "${OPTS[@]}" \
+    ./llama-cli 2>>"$LOG" "${OPTS[@]}" \
             --prompt-cache "$CUR_PROMPT_CACHE" \
             --prompt-cache-all \
             --file "$CUR_PROMPT_FILE" \
             --reverse-prompt "${USER_NAME}:" \
             --n_predict "$n_predict" |
-        skip_bytes 1 |                  # skip BOS token added by ./main
+        skip_bytes 1 |                  # skip BOS token added by ./llama-cli
         tee "$CUR_PROMPT_FILE.tmp" |    # save prompt + generation to tmp file
         skip_bytes "$n_prompt_len_pre"  # print generation
 
@@ -133,7 +133,7 @@ while read -e line; do
     # TODO get both messages in one go
     if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
         ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
-        echo >&2 "Couldn't get number of tokens from ./main output!"
+        echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
         exit 1
     fi
 
@@ -144,7 +144,7 @@ while read -e line; do
     fi
 
     # Update cache for next prompt in background, ideally during user input
-    ./main >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
+    ./llama-cli >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
           --prompt-cache "$NEXT_PROMPT_CACHE" \
           --file "$NEXT_PROMPT_FILE" \
           --n_predict 1 &
diff --git a/llm/llama.cpp/examples/chat-vicuna.sh b/llm/llama.cpp/examples/chat-vicuna.sh
index 8c7b7bef42784d3037c377e71fc20e08a7302883..ffdd200849503045a3553607aafe4b8eb8770025 100644
--- a/llm/llama.cpp/examples/chat-vicuna.sh
+++ b/llm/llama.cpp/examples/chat-vicuna.sh
@@ -30,7 +30,7 @@ sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
      $PROMPT_TEMPLATE > $PROMPT_FILE
 
 # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
-./bin/main $GEN_OPTIONS \
+./bin/llama-cli $GEN_OPTIONS \
   --model "$MODEL" \
   --threads "$N_THREAD" \
   --n_predict "$N_PREDICTS" \
diff --git a/llm/llama.cpp/examples/chat.sh b/llm/llama.cpp/examples/chat.sh
index d567acecdff111aa66bffc9476e09ebc1a2fd5c9..9f85d1e265d007c2aed11f7f58924eb04c696e5c 100644
--- a/llm/llama.cpp/examples/chat.sh
+++ b/llm/llama.cpp/examples/chat.sh
@@ -11,6 +11,6 @@ cd ..
 #
 #   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
 #
-./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
+./llama-cli -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
     --repeat_penalty 1.0 --color -i \
     -r "User:" -f prompts/chat-with-bob.txt
diff --git a/llm/llama.cpp/examples/convert-llama2c-to-ggml/CMakeLists.txt b/llm/llama.cpp/examples/convert-llama2c-to-ggml/CMakeLists.txt
index e262d44f9849609a9b617ccad01dacc1c53c8d66..a6790e617217e9e3fc118673528d81618426d021 100644
--- a/llm/llama.cpp/examples/convert-llama2c-to-ggml/CMakeLists.txt
+++ b/llm/llama.cpp/examples/convert-llama2c-to-ggml/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET convert-llama2c-to-ggml)
+set(TARGET llama-convert-llama2c-to-ggml)
 add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/convert-llama2c-to-ggml/README.md b/llm/llama.cpp/examples/convert-llama2c-to-ggml/README.md
index 742dcf7a35b506e5c201664aee3410f27cbf23cf..5774ac83c32c82056fbc563921678c251280ba49 100644
--- a/llm/llama.cpp/examples/convert-llama2c-to-ggml/README.md
+++ b/llm/llama.cpp/examples/convert-llama2c-to-ggml/README.md
@@ -8,7 +8,7 @@ To convert the model first download the models from the [llama2.c](https://githu
 
 After successful compilation, following usage options are available:
 ```
-usage: ./convert-llama2c-to-ggml [options]
+usage: ./llama-convert-llama2c-to-ggml [options]
 
 options:
   -h, --help                       show this help message and exit
@@ -19,10 +19,10 @@ options:
 
 An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
 
-`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
+`$ ./llama-convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
 Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
 
 Now you can use the model with a command like:
 
-`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`
+`$ ./llama-cli -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`
diff --git a/llm/llama.cpp/examples/convert_legacy_llama.py b/llm/llama.cpp/examples/convert_legacy_llama.py
new file mode 100644
index 0000000000000000000000000000000000000000..9ab9ab06edf8fa7eb32b72701e57ad8804bfe3c3
--- /dev/null
+++ b/llm/llama.cpp/examples/convert_legacy_llama.py
@@ -0,0 +1,1440 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import logging
+import argparse
+import concurrent.futures
+import enum
+import faulthandler
+import functools
+import itertools
+import json
+import math
+import mmap
+import os
+import pickle
+import re
+import signal
+import struct
+import sys
+import textwrap
+import time
+import zipfile
+from abc import ABC, abstractmethod
+from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar
+
+import numpy as np
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    # use .parent.parent since we are in "examples" directory
+    sys.path.insert(1, str(Path(__file__).parent.parent / 'gguf-py'))
+
+import gguf
+from gguf import BaseVocab, Vocab, NoVocab, BpeVocab, SentencePieceVocab, LlamaHfVocab
+
+if TYPE_CHECKING:
+    from typing_extensions import Self, TypeAlias
+
+logger = logging.getLogger("convert")
+
+if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
+    faulthandler.register(signal.SIGUSR1)
+
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
+
+ARCH = gguf.MODEL_ARCH.LLAMA
+
+DEFAULT_CONCURRENCY = 8
+
+ADDED_TOKENS_FILE = 'added_tokens.json'
+FAST_TOKENIZER_FILE = 'tokenizer.json'
+
+#
+# data types
+#
+
+
+@dataclass(frozen=True)
+class DataType:
+    name: str
+    dtype: np.dtype[Any]
+    valid_conversions: list[str]
+
+    def elements_to_bytes(self, n_elements: int) -> int:
+        return n_elements * self.dtype.itemsize
+
+
+@dataclass(frozen=True)
+class UnquantizedDataType(DataType):
+    pass
+
+
+DT_F16  = UnquantizedDataType('F16',  dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32',  dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32',  dtype = np.dtype(np.int16),   valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16),  valid_conversions = ['F32', 'F16', 'Q8_0'])
+
+
+@dataclass(frozen=True)
+class QuantizedDataType(DataType):
+    block_size: int
+    quantized_dtype: np.dtype[Any]
+    ggml_type: gguf.GGMLQuantizationType
+
+    def quantize(self, arr: NDArray) -> NDArray:
+        raise NotImplementedError(f'Quantization for {self.name} not implemented')
+
+    def elements_to_bytes(self, n_elements: int) -> int:
+        assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
+        return self.quantized_dtype.itemsize * (n_elements // self.block_size)
+
+
+@dataclass(frozen=True)
+class Q8_0QuantizedDataType(QuantizedDataType):
+    # Mini Q8_0 quantization in Python!
+    def quantize(self, arr: NDArray) -> NDArray:
+        assert arr.size % self.block_size == 0 and arr.size != 0, f'Bad array size {arr.size}'
+        assert arr.dtype == np.float32, f'Bad array type {arr.dtype}'
+        n_blocks = arr.size // self.block_size
+        blocks = arr.reshape((n_blocks, self.block_size))
+        # Much faster implementation of block quantization contributed by @Cebtenzzre
+
+        def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]:
+            d = abs(blocks).max(axis = 1) / np.float32(127)
+            with np.errstate(divide = 'ignore'):
+                qs = (blocks / d[:, None]).round()
+            qs[d == 0] = 0
+            yield from zip(d, qs)
+        return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
+
+
+DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
+                                dtype = np.dtype(np.float32), valid_conversions = [],
+                                ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
+                                quantized_dtype = np.dtype([('d', ' DataType:
+        dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
+        if dt is None:
+            raise ValueError(self)
+        # Convert all 1D tensors to F32.  Most of the codebase that takes in 1D tensors only handles F32 tensors, and most of the outputs tensors are F32.
+        #  Also The 1d tensors aren't much of a performance/size issue.  So instead of having to have separate F32 and F16 implementations of both, just convert everything to F32 for now.
+        return dt if len(tensor.shape) > 1 else DT_F32
+
+
+GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
+    GGMLFileType.AllF32    : DT_F32,
+    GGMLFileType.MostlyF16 : DT_F16,
+    GGMLFileType.MostlyQ8_0: DT_Q8_0,
+}
+
+#
+# hparams loading
+#
+
+
+@dataclass
+class Params:
+    n_vocab:        int
+    n_embd:         int
+    n_layer:        int
+    n_ctx:          int
+    n_ff:           int
+    n_head:         int
+    n_head_kv:      int
+    n_experts:      int | None = None
+    n_experts_used: int | None = None
+    f_norm_eps:     float | None = None
+
+    rope_scaling_type: gguf.RopeScalingType | None = None
+    f_rope_freq_base: float | None = None
+    f_rope_scale: float | None = None
+    n_ctx_orig: int | None = None
+    rope_finetuned: bool | None = None
+
+    ftype: GGMLFileType | None = None
+
+    # path to the directory containing the model files
+    path_model: Path | None = None
+
+    @staticmethod
+    def guessed(model: LazyModel) -> Params:
+        # try transformer naming first
+        n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape
+
+        # try transformer naming first
+        if "model.layers.0.self_attn.q_proj.weight" in model:
+            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model)
+        elif "model.layers.0.self_attn.W_pack.weight" in model:   # next: try baichuan naming
+            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model)
+        else:
+            n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
+
+        if n_layer < 1:
+            msg = """\
+                failed to guess 'n_layer'. This model is unknown or unsupported.
+                Suggestion: provide 'config.json' of the model in the same directory containing model files."""
+            raise KeyError(textwrap.dedent(msg))
+
+        n_head = n_embd // 128 # guessed
+        n_mult = 256           # guessed
+
+        # TODO: verify this
+        n_ff = int(2 * (4 * n_embd) / 3)
+        n_ff = n_mult * ((n_ff + n_mult - 1) // n_mult)
+
+        return Params(
+            n_vocab    = n_vocab,
+            n_embd     = n_embd,
+            n_layer    = n_layer,
+            n_ctx      = -1,
+            n_ff       = n_ff,
+            n_head     = n_head,
+            n_head_kv  = n_head,
+            f_norm_eps = 1e-5,
+        )
+
+    @staticmethod
+    def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
+        with open(config_path) as f:
+            config = json.load(f)
+
+        rope_scaling_type = f_rope_scale = n_ctx_orig = rope_finetuned = None
+        rope_scaling = config.get("rope_scaling")
+
+        if rope_scaling is not None and (typ := rope_scaling.get("type")):
+            rope_factor = rope_scaling.get("factor")
+            f_rope_scale = rope_factor
+            if typ == "linear":
+                rope_scaling_type = gguf.RopeScalingType.LINEAR
+            elif typ == "yarn":
+                rope_scaling_type = gguf.RopeScalingType.YARN
+                n_ctx_orig = rope_scaling['original_max_position_embeddings']
+                rope_finetuned = rope_scaling['finetuned']
+            else:
+                raise NotImplementedError(f'Unknown rope scaling type: {typ}')
+
+        if "max_sequence_length" in config:
+            n_ctx = config["max_sequence_length"]
+        elif "max_position_embeddings" in config:
+            n_ctx = config["max_position_embeddings"]
+        else:
+            msg = """\
+                failed to guess 'n_ctx'. This model is unknown or unsupported.
+                Suggestion: provide 'config.json' of the model in the same directory containing model files."""
+            raise KeyError(textwrap.dedent(msg))
+
+        n_experts      = None
+        n_experts_used = None
+
+        if "num_local_experts" in config:
+            n_experts = config["num_local_experts"]
+            n_experts_used = config["num_experts_per_tok"]
+
+        return Params(
+            n_vocab           = config["vocab_size"],
+            n_embd            = config["hidden_size"],
+            n_layer           = config["num_hidden_layers"],
+            n_ctx             = n_ctx,
+            n_ff              = config["intermediate_size"],
+            n_head            = (n_head := config["num_attention_heads"]),
+            n_head_kv         = config.get("num_key_value_heads", n_head),
+            n_experts         = n_experts,
+            n_experts_used    = n_experts_used,
+            f_norm_eps        = config["rms_norm_eps"],
+            f_rope_freq_base  = config.get("rope_theta"),
+            rope_scaling_type = rope_scaling_type,
+            f_rope_scale      = f_rope_scale,
+            n_ctx_orig        = n_ctx_orig,
+            rope_finetuned    = rope_finetuned,
+        )
+
+    # LLaMA v2 70B params.json
+    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
+    @staticmethod
+    def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
+        with open(config_path) as f:
+            config = json.load(f)
+
+        n_experts      = None
+        n_experts_used = None
+        f_rope_freq_base = None
+        n_ff = None
+
+        # hack to determine LLaMA v1 vs v2 vs CodeLlama
+        if config.get("moe"):
+            # Mixtral
+            n_ctx = 32768
+        elif config.get("rope_theta") == 1000000:
+            # CodeLlama
+            n_ctx = 16384
+        elif config["norm_eps"] == 1e-05:
+            # LLaMA v2
+            n_ctx = 4096
+        else:
+            # LLaMA v1
+            n_ctx = 2048
+
+        if "layers.0.feed_forward.w1.weight" in model:
+            n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
+
+        if config.get("moe"):
+            n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
+            n_experts      = config["moe"]["num_experts"]
+            n_experts_used = config["moe"]["num_experts_per_tok"]
+            f_rope_freq_base = 1e6
+
+        assert n_ff is not None
+
+        return Params(
+            n_vocab          = model["tok_embeddings.weight"].shape[0],
+            n_embd           = config["dim"],
+            n_layer          = config["n_layers"],
+            n_ctx            = n_ctx,
+            n_ff             = n_ff,
+            n_head           = (n_head := config["n_heads"]),
+            n_head_kv        = config.get("n_kv_heads", n_head),
+            n_experts        = n_experts,
+            n_experts_used   = n_experts_used,
+            f_norm_eps       = config["norm_eps"],
+            f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
+        )
+
+    @staticmethod
+    def load(model_plus: ModelPlus) -> Params:
+        hf_config_path   = model_plus.paths[0].parent / "config.json"
+        orig_config_path = model_plus.paths[0].parent / "params.json"
+
+        if hf_config_path.exists():
+            params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
+        elif orig_config_path.exists():
+            params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
+        elif model_plus.format != 'none':
+            params = Params.guessed(model_plus.model)
+        else:
+            raise ValueError('Cannot guess params when model format is none')
+
+        params.path_model = model_plus.paths[0].parent
+
+        return params
+
+
+#
+# data loading
+# TODO: reuse (probably move to gguf.py?)
+#
+
+
+def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
+    if n_head_kv is not None and n_head != n_head_kv:
+        n_head = n_head_kv
+    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape))
+
+
+class Tensor(ABC):
+    ndarray: NDArray
+    data_type: DataType
+
+    @abstractmethod
+    def astype(self, data_type: DataType) -> Self: ...
+    @abstractmethod
+    def permute(self, n_head: int, n_head_kv: int) -> Self: ...
+    @abstractmethod
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> Self: ...
+    @abstractmethod
+    def part(self, n_part: int) -> Self: ...
+    @abstractmethod
+    def to_ggml(self) -> GGMLCompatibleTensor: ...
+
+
+def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
+    assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}"
+    fp32_arr = bf16_arr.astype(np.uint32) << 16
+    return fp32_arr.view(np.float32)
+
+
+class UnquantizedTensor(Tensor):
+    def __init__(self, ndarray: NDArray):
+        assert isinstance(ndarray, np.ndarray)
+        self.ndarray = ndarray
+        self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
+
+    def astype(self, data_type: DataType) -> UnquantizedTensor:
+        dtype = data_type.dtype
+        if self.data_type == DT_BF16:
+            self.ndarray = bf16_to_fp32(self.ndarray)
+        return UnquantizedTensor(self.ndarray.astype(dtype))
+
+    def to_ggml(self) -> Self:
+        return self
+
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor:
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
+
+    def part(self, n_part: int) -> UnquantizedTensor:
+        r = self.ndarray.shape[0] // 3
+        return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
+
+    def permute(self, n_head: int, n_head_kv: int) -> UnquantizedTensor:
+        return UnquantizedTensor(permute(self.ndarray, n_head, n_head_kv))
+
+
+def load_unquantized(lazy_tensor: LazyTensor, expected_dtype: Any = None, convert: bool = False) -> NDArray:
+    tensor = lazy_tensor.load()
+    assert isinstance(tensor, UnquantizedTensor)
+
+    # double-check:
+    actual_shape = list(tensor.ndarray.shape)
+    assert actual_shape == lazy_tensor.shape, (actual_shape, lazy_tensor.shape)
+    if expected_dtype is not None and expected_dtype != tensor.ndarray.dtype:
+        if convert:
+            tensor.ndarray = tensor.ndarray.astype(expected_dtype)
+        else:
+            raise ValueError(f'expected this tensor to have dtype {expected_dtype}, got {tensor.ndarray.dtype}')
+
+    return tensor.ndarray
+
+
+GGMLCompatibleTensor = UnquantizedTensor
+
+
+@dataclass
+class LazyTensor:
+    _load: Callable[[], Tensor]
+    shape: list[int]
+    data_type: DataType
+    description: str
+
+    def load(self) -> Tensor:
+        ret = self._load()
+        # Should be okay if it maps to the same numpy type?
+        assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
+            (self.data_type, ret.data_type, self.description)
+        return ret
+
+    def astype(self, data_type: DataType) -> LazyTensor:
+        self.validate_conversion_to(data_type)
+
+        def load() -> Tensor:
+            return self.load().astype(data_type)
+        return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}')
+
+    def validate_conversion_to(self, data_type: DataType) -> None:
+        if data_type != self.data_type and data_type.name not in self.data_type.valid_conversions:
+            raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.')
+
+
+LazyModel: TypeAlias = 'dict[str, LazyTensor]'
+
+ModelFormat: TypeAlias = Literal['ggml', 'torch', 'safetensors', 'none']
+
+@dataclass
+class ModelPlus:
+    model: LazyModel
+    paths: list[Path]  # Where this was read from.
+    format: ModelFormat
+    vocab: BaseVocab | None  # For GGML models (which have vocab built in), the vocab.
+
+
+def merge_sharded(models: list[LazyModel]) -> LazyModel:
+    # Original LLaMA models have each file contain one part of each tensor.
+    # Use a dict instead of a set to preserve order.
+    names = {name: None for model in models for name in model}
+
+    def convert(name: str) -> LazyTensor:
+        lazy_tensors = [model[name] for model in models]
+        if len(lazy_tensors) == 1:
+            # only one file; don't go through this procedure since there might
+            # be quantized tensors
+            return lazy_tensors[0]
+        if len(lazy_tensors[0].shape) == 1:
+            # the tensor is just duplicated in every file
+            return lazy_tensors[0]
+        if name.startswith('tok_embeddings.') or \
+           name.endswith('.attention.wo.weight') or \
+           name.endswith('.feed_forward.w2.weight'):
+            # split by columns
+            axis = 1
+        else:
+            # split by rows
+            axis = 0
+        concatenated_shape = list(lazy_tensors[0].shape)
+        concatenated_shape[axis] = sum(tensor.shape[axis] for tensor in lazy_tensors)
+
+        def load() -> UnquantizedTensor:
+            ndarrays = [load_unquantized(tensor) for tensor in lazy_tensors]
+            concatenated = np.concatenate(ndarrays, axis=axis)
+            return UnquantizedTensor(concatenated)
+        description = 'concatenated[[' + '] | ['.join(lt.description for lt in lazy_tensors) + ']]'
+        return LazyTensor(load, concatenated_shape, lazy_tensors[0].data_type, description)
+    return {name: convert(name) for name in names}
+
+
+def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
+    formats: set[ModelFormat] = set(mp.format for mp in models_plus)
+    assert len(formats) == 1, "different formats?"
+    format = formats.pop()
+    paths = [path for mp in models_plus for path in mp.paths]
+    # Use the first non-None vocab, if any.
+    try:
+        vocab = next(mp.vocab for mp in models_plus if mp.vocab is not None)
+    except StopIteration:
+        vocab = None
+
+    if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
+        # Transformers models put different tensors in different files, but
+        # don't split individual tensors between files.
+        model: LazyModel = {}
+        for mp in models_plus:
+            model.update(mp.model)
+    else:
+        model = merge_sharded([mp.model for mp in models_plus])
+
+    return ModelPlus(model, paths, format, vocab)
+
+
+def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
+    def load() -> Tensor:
+        return lazy_tensor.load().permute(n_head, n_head_kv)
+    return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
+
+
+def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
+    def load() -> Tensor:
+        return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
+    s = lazy_tensor.shape.copy()
+    s[0] = s[0] // 3
+    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
+
+
+def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
+    def load() -> Tensor:
+        return lazy_tensor.load().part(n_part)
+    s = lazy_tensor.shape.copy()
+    s[0] = s[0] // 3
+    return LazyTensor(load, s, lazy_tensor.data_type, 'part ' + lazy_tensor.description)
+
+
+def pack_experts_lazy(lazy_tensors: list[LazyTensor]) -> LazyTensor:
+    def load() -> Tensor:
+        tensors = [lazy_tensor.load() for lazy_tensor in lazy_tensors]
+        return UnquantizedTensor(np.array([tensor.ndarray for tensor in tensors]))
+    s = lazy_tensors[0].shape.copy()
+    s.insert(0, len(lazy_tensors))
+    return LazyTensor(load, s, lazy_tensors[0].data_type, 'pack_experts ' + ' | '.join(lt.description for lt in lazy_tensors))
+
+
+# Functionality that simulates `torch.load` but where individual tensors are
+# only loaded into memory on demand, not all at once.
+# PyTorch can't do this natively as of time of writing:
+# - https://github.com/pytorch/pytorch/issues/64327
+# This allows us to de-shard without multiplying RAM usage, and also
+# conveniently drops the PyTorch dependency (though we still need numpy).
+
+
+@dataclass
+class LazyStorageKind:
+    data_type: DataType
+
+
+@dataclass
+class LazyStorage:
+    load: Callable[[int, int], NDArray]
+    kind: LazyStorageKind
+    description: str
+
+
+class LazyUnpickler(pickle.Unpickler):
+    def __init__(self, fp: IO[bytes], data_base_path: str, zip_file: zipfile.ZipFile):
+        super().__init__(fp)
+        self.data_base_path = data_base_path
+        self.zip_file = zip_file
+
+    def persistent_load(self, pid: Any) -> Any:
+        assert pid[0] == 'storage'
+        assert isinstance(pid[1], LazyStorageKind)
+        data_type = pid[1].data_type
+        filename_stem = pid[2]
+        filename = f'{self.data_base_path}/{filename_stem}'
+        info = self.zip_file.getinfo(filename)
+
+        def load(offset: int, elm_count: int) -> NDArray:
+            dtype = data_type.dtype
+            with self.zip_file.open(info) as fp:
+                fp.seek(offset * dtype.itemsize)
+                size = elm_count * dtype.itemsize
+                data = fp.read(size)
+            assert len(data) == size
+            return np.frombuffer(data, dtype)
+        description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
+        return LazyStorage(load=load, kind=pid[1], description=description)
+
+    @staticmethod
+    def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
+                               requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
+        assert isinstance(storage, LazyStorage)
+
+        def load() -> UnquantizedTensor:
+            elm_count = stride[0] * size[0]
+            return UnquantizedTensor(storage.load(storage_offset, elm_count).reshape(size))
+        description = f'pickled storage_offset={storage_offset} in {storage.description}'
+        return LazyTensor(load, list(size), storage.kind.data_type, description)
+
+    @staticmethod
+    def rebuild_from_type_v2(func, new_type, args, state):
+        return func(*args)
+
+    CLASSES: dict[tuple[str, str], type[LazyTensor] | LazyStorageKind] = {
+        # getattr used here as a workaround for mypy not being smart enough to determine
+        # the staticmethods have a __func__ attribute.
+        ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
+        ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
+        ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
+        ('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
+        ('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
+        ('torch', 'IntStorage'): LazyStorageKind(DT_I32),
+        ('torch', 'Tensor'): LazyTensor,
+    }
+
+    def find_class(self, module: str, name: str) -> Any:
+        if not module.startswith('torch'):
+            return super().find_class(module, name)
+        return self.CLASSES[(module, name)]
+
+
+def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
+    zf = zipfile.ZipFile(outer_fp)
+    pickle_paths = [name for name in zf.namelist() if name.endswith('.pkl')]
+    assert len(pickle_paths) == 1, pickle_paths
+    pickle_fp = zf.open(pickle_paths[0], 'r')
+    unpickler = LazyUnpickler(pickle_fp,
+                              data_base_path=pickle_paths[0][:-4],
+                              zip_file=zf)
+    model = unpickler.load()
+    if 'model' in model: model = model['model']
+    as_dict = dict(model.items())
+    return ModelPlus(model=as_dict, paths=[path], format='torch', vocab=None)
+
+
+def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
+    header_size, = struct.unpack(' LazyTensor:
+        data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
+        numpy_dtype = data_type.dtype
+        shape: list[int] = info['shape']
+        begin, end = info['data_offsets']
+        assert 0 <= begin <= end <= len(byte_buf)
+        assert end - begin == math.prod(shape) * numpy_dtype.itemsize
+        buf = byte_buf[begin:end]
+
+        def load() -> UnquantizedTensor:
+            return UnquantizedTensor(np.frombuffer(buf, dtype=numpy_dtype).reshape(shape))
+        description = f'safetensors begin={begin} end={end} type={data_type} path={path}'
+        return LazyTensor(load, shape, data_type, description)
+    model = {name: convert(info) for (name, info) in header.items() if name != '__metadata__'}
+    return ModelPlus(model=model, paths=[path], format='safetensors', vocab=None)
+
+
+def must_read(fp: IO[bytes], length: int) -> bytes:
+    ret = fp.read(length)
+    if len(ret) < length:
+        raise EOFError("unexpectedly reached end of file")
+    return ret
+
+
+@functools.lru_cache(maxsize=None)
+def lazy_load_file(path: Path) -> ModelPlus:
+    fp = open(path, 'rb')
+    first8 = fp.read(8)
+    fp.seek(0)
+    if first8[:2] == b'PK':
+        # A zip file, i.e. PyTorch format
+        return lazy_load_torch_file(fp, path)
+    elif struct.unpack(' Iterable[Out]:
+    '''Parallel map, but with backpressure.  If the caller doesn't call `next`
+    fast enough, this will stop calling `func` at some point rather than
+    letting results pile up in memory.  Specifically, there is a max of one
+    output value buffered per thread.'''
+    if concurrency < 2:
+        yield from map(func, iterable)
+        # Not reached.
+    iterable = iter(iterable)
+    executor_class: type[ThreadPoolExecutor] | type[ProcessPoolExecutor]
+    if use_processpool_executor:
+        executor_class = ProcessPoolExecutor
+    else:
+        executor_class = ThreadPoolExecutor
+    with executor_class(max_workers=max_workers) as executor:
+        futures: list[concurrent.futures.Future[Out]] = []
+        done = False
+        for _ in range(concurrency):
+            try:
+                futures.append(executor.submit(func, next(iterable)))
+            except StopIteration:
+                done = True
+                break
+
+        while futures:
+            result = futures.pop(0).result()
+            while not done and len(futures) < concurrency:
+                try:
+                    futures.append(executor.submit(func, next(iterable)))
+                except StopIteration:
+                    done = True
+                    break
+            yield result
+
+
+def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False) -> None:
+    # Handle special case where the model's vocab size is not set
+    if params.n_vocab == -1:
+        raise ValueError(
+            "The model's vocab size is set to -1 in params.json. Please update it manually."
+            + (f" Maybe {vocab.vocab_size}?" if isinstance(vocab, Vocab) else ""),
+        )
+    if not isinstance(vocab, Vocab):
+        return  # model has no vocab
+
+    # Check for a vocab size mismatch
+    if params.n_vocab == vocab.vocab_size:
+        logger.warning("Ignoring added_tokens.json since model matches vocab size without it.")
+        return
+
+    if pad_vocab and params.n_vocab > vocab.vocab_size:
+        pad_count = params.n_vocab - vocab.vocab_size
+        logger.debug(
+            f"Padding vocab with {pad_count} token(s) -  through "
+        )
+        for i in range(1, pad_count + 1):
+            vocab.added_tokens_dict[f""] = -1
+            vocab.added_tokens_list.append(f"")
+        vocab.vocab_size = params.n_vocab
+        return
+
+    msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer} has {vocab.vocab_size})."
+    if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20:
+        msg += f"  Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})."
+    if vocab.vocab_size < params.n_vocab:
+        msg += " Add the --pad-vocab option and try again."
+
+    raise ValueError(msg)
+
+
+class OutputFile:
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
+        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
+
+    def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None:
+        # Metadata About The Model And Its Provenence
+        name = "LLaMA"
+        if metadata is not None and metadata.name is not None:
+            name = metadata.name
+        elif params.path_model is not None:
+            name = params.path_model.name
+        elif params.n_ctx == 4096:
+            # Heuristic detection of LLaMA v2 model
+            name = "LLaMA v2"
+
+        self.gguf.add_name(name)
+
+        if metadata is not None:
+            if metadata.author is not None:
+                self.gguf.add_author(metadata.author)
+            if metadata.version is not None:
+                self.gguf.add_version(metadata.version)
+            if metadata.organization is not None:
+                self.gguf.add_organization(metadata.organization)
+
+            if metadata.finetune is not None:
+                self.gguf.add_finetune(metadata.finetune)
+            if metadata.basename is not None:
+                self.gguf.add_basename(metadata.basename)
+
+            if metadata.description is not None:
+                self.gguf.add_description(metadata.description)
+            if metadata.quantized_by is not None:
+                self.gguf.add_quantized_by(metadata.quantized_by)
+
+            if metadata.size_label is not None:
+                self.gguf.add_size_label(metadata.size_label)
+
+            if metadata.license is not None:
+                self.gguf.add_license(metadata.license)
+            if metadata.license_name is not None:
+                self.gguf.add_license_name(metadata.license_name)
+            if metadata.license_link is not None:
+                self.gguf.add_license_link(metadata.license_link)
+
+            if metadata.url is not None:
+                self.gguf.add_url(metadata.url)
+            if metadata.doi is not None:
+                self.gguf.add_doi(metadata.doi)
+            if metadata.uuid is not None:
+                self.gguf.add_uuid(metadata.uuid)
+            if metadata.repo_url is not None:
+                self.gguf.add_repo_url(metadata.repo_url)
+
+            if metadata.source_url is not None:
+                self.gguf.add_source_url(metadata.source_url)
+            if metadata.source_doi is not None:
+                self.gguf.add_source_doi(metadata.source_doi)
+            if metadata.source_uuid is not None:
+                self.gguf.add_source_uuid(metadata.source_uuid)
+            if metadata.source_repo_url is not None:
+                self.gguf.add_source_repo_url(metadata.source_repo_url)
+
+            if metadata.base_models is not None:
+                self.gguf.add_base_model_count(len(metadata.base_models))
+                for key, base_model_entry in enumerate(metadata.base_models):
+                    if "name" in base_model_entry:
+                        self.gguf.add_base_model_name(key, base_model_entry["name"])
+                    if "author" in base_model_entry:
+                        self.gguf.add_base_model_author(key, base_model_entry["author"])
+                    if "version" in base_model_entry:
+                        self.gguf.add_base_model_version(key, base_model_entry["version"])
+                    if "organization" in base_model_entry:
+                        self.gguf.add_base_model_organization(key, base_model_entry["organization"])
+                    if "url" in base_model_entry:
+                        self.gguf.add_base_model_url(key, base_model_entry["url"])
+                    if "doi" in base_model_entry:
+                        self.gguf.add_base_model_doi(key, base_model_entry["doi"])
+                    if "uuid" in base_model_entry:
+                        self.gguf.add_base_model_uuid(key, base_model_entry["uuid"])
+                    if "repo_url" in base_model_entry:
+                        self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
+
+            if metadata.tags is not None:
+                self.gguf.add_tags(metadata.tags)
+            if metadata.languages is not None:
+                self.gguf.add_languages(metadata.languages)
+            if metadata.datasets is not None:
+                self.gguf.add_datasets(metadata.datasets)
+
+    def add_meta_arch(self, params: Params) -> None:
+        # Metadata About The Neural Architecture Itself
+        self.gguf.add_vocab_size(params.n_vocab)
+        self.gguf.add_context_length(params.n_ctx)
+        self.gguf.add_embedding_length(params.n_embd)
+        self.gguf.add_block_count(params.n_layer)
+        self.gguf.add_feed_forward_length(params.n_ff)
+        self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
+        self.gguf.add_head_count          (params.n_head)
+        self.gguf.add_head_count_kv       (params.n_head_kv)
+
+        if params.n_experts:
+            self.gguf.add_expert_count(params.n_experts)
+
+        if params.n_experts_used:
+            self.gguf.add_expert_used_count(params.n_experts_used)
+
+        if params.f_norm_eps:
+            self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
+        else:
+            raise ValueError('f_norm_eps is None')
+
+        if params.f_rope_freq_base is not None:
+            self.gguf.add_rope_freq_base(params.f_rope_freq_base)
+
+        if params.rope_scaling_type:
+            assert params.f_rope_scale is not None
+            self.gguf.add_rope_scaling_type(params.rope_scaling_type)
+            self.gguf.add_rope_scaling_factor(params.f_rope_scale)
+
+        if params.n_ctx_orig is not None:
+            self.gguf.add_rope_scaling_orig_ctx_len(params.n_ctx_orig)
+
+        if params.rope_finetuned is not None:
+            self.gguf.add_rope_scaling_finetuned(params.rope_finetuned)
+
+        if params.ftype is not None:
+            self.gguf.add_file_type(params.ftype)
+
+    def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
+        tokens = []
+        scores = []
+        toktypes = []
+
+        # NOTE: `all_tokens` returns the base vocabulary and added tokens
+        for text, score, toktype in vocab.all_tokens():
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        assert len(tokens) == vocab.vocab_size
+
+        return tokens, scores, toktypes
+
+    def add_meta_vocab(self, vocab: Vocab) -> None:
+        # Ensure that tokenizer_model is added to the GGUF model
+        self.gguf.add_tokenizer_model(vocab.tokenizer_model)
+
+        # Extract model vocabulary for model conversion
+        tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
+
+        # Add extracted token information for model conversion
+        self.gguf.add_token_list(tokens)
+        self.gguf.add_token_scores(scores)
+        self.gguf.add_token_types(toktypes)
+
+    def add_meta_special_vocab(self, svocab: gguf.SpecialVocab) -> None:
+        svocab.add_to_gguf(self.gguf)
+
+    def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
+        n_elements = int(np.prod(tensor.shape))
+        raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
+        data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
+        data_nbytes = tensor.data_type.elements_to_bytes(n_elements)
+        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype)
+
+    def write_meta(self) -> None:
+        self.gguf.write_header_to_file()
+        self.gguf.write_kv_data_to_file()
+
+    def write_tensor_info(self) -> None:
+        self.gguf.write_ti_data_to_file()
+
+    def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None:
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(
+                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
+                use_processpool_executor=True,
+            )
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
+        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            logger.info(
+                f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
+            )
+            self.gguf.write_tensor_data(ndarray)
+
+    def close(self) -> None:
+        self.gguf.close()
+
+    @staticmethod
+    def write_vocab_only(
+        fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None,
+    ) -> None:
+        check_vocab_size(params, vocab, pad_vocab=pad_vocab)
+
+        of = OutputFile(fname_out, endianess=endianess)
+
+        # meta data
+        of.add_meta_model(params, metadata)
+        of.add_meta_arch(params)
+        of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
+
+        of.write_meta()
+
+        of.close()
+
+    @staticmethod
+    def do_item(item: tuple[str, LazyTensor]) -> tuple[DataType, NDArray]:
+        name, lazy_tensor = item
+        tensor = lazy_tensor.load().to_ggml()
+        return (lazy_tensor.data_type, tensor.ndarray)
+
+    @staticmethod
+    def maybe_do_quantize(item: tuple[DataType, NDArray]) -> NDArray:
+        dt, arr = item
+        if not isinstance(dt, QuantizedDataType):
+            return arr
+        return dt.quantize(arr)
+
+    @staticmethod
+    def write_all(
+        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
+        concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
+        pad_vocab: bool = False,
+        metadata: gguf.Metadata | None = None,
+    ) -> None:
+        check_vocab_size(params, vocab, pad_vocab=pad_vocab)
+
+        of = OutputFile(fname_out, endianess=endianess)
+
+        # meta data
+        of.add_meta_model(params, metadata)
+        of.add_meta_arch(params)
+        if isinstance(vocab, Vocab):
+            of.add_meta_vocab(vocab)
+            of.add_meta_special_vocab(svocab)
+        else:  # NoVocab
+            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
+
+        # tensor info
+        for name, lazy_tensor in model.items():
+            of.add_tensor_info(name, lazy_tensor)
+
+        of.write_meta()
+        of.write_tensor_info()
+
+        # tensor data
+        of.write_tensor_data(ftype, model, concurrency)
+
+        of.close()
+
+
+def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
+    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
+
+    if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
+        return GGMLFileType.AllF32
+    if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
+        return GGMLFileType.MostlyF16
+    if output_type_str == "q8_0":
+        return GGMLFileType.MostlyQ8_0
+
+    name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
+
+    raise ValueError(f"Unexpected combination of types: {name_to_type}")
+
+
+def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]:
+    total_params = 0
+    shared_params = 0
+    expert_params = 0
+
+    for name, lazy_tensor in tensors:
+        # We don't need these
+        if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
+            continue
+
+        # Got A Tensor
+        sum_weights_in_tensor: int = 1
+
+        # Tensor Volume
+        for dim in lazy_tensor.shape:
+            sum_weights_in_tensor *= dim
+
+        if ".experts." in name:
+            if ".experts.0." in name:
+                expert_params += sum_weights_in_tensor
+        else:
+            shared_params += sum_weights_in_tensor
+
+        total_params += sum_weights_in_tensor
+
+    return total_params, shared_params, expert_params
+
+
+def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
+    return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
+            for (name, tensor) in model.items()}
+
+
+def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
+    tmap = gguf.TensorNameMap(ARCH, params.n_layer)
+    should_skip = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
+
+    tmp = model
+
+    # merge experts into one tensor
+    if params.n_experts and params.n_experts > 0:
+        for i_l in range(params.n_layer):
+            for w in range(1, 4):
+                experts = []
+                for e in range(params.n_experts):
+                    if f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight" in model:
+                        experts.append(model[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"])
+                        del tmp[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"]
+                    elif f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight" in model:
+                        experts.append(model[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"])
+                        del tmp[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"]
+                    else:
+                        raise ValueError(f"Expert tensor not found: layers.{i_l}.feed_forward.experts.{e}.w{w}.weight")
+                tmp[f"layers.{i_l}.feed_forward.experts.w{w}.weight"] = pack_experts_lazy(experts)
+
+    # HF models permut or pack some of the tensors, so we need to undo that
+    for i in itertools.count():
+        if f"model.layers.{i}.self_attn.q_proj.weight" in model:
+            logger.debug(f"Permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv)
+            # tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
+        elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
+            logger.debug(f"Unpacking and permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
+            tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy        (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
+        else:
+            break
+
+    out: LazyModel = {}
+    for name, lazy_tensor in model.items():
+        tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
+        if name_new is None:
+            if skip_unknown:
+                logger.warning(f"Unexpected tensor name: {name} - skipping")
+                continue
+            raise ValueError(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
+
+        if tensor_type in should_skip:
+            logger.debug(f"skipping tensor {name_new}")
+            continue
+
+        logger.debug(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
+        out[name_new] = lazy_tensor
+
+    return out
+
+
+def nth_multifile_path(path: Path, n: int) -> Path | None:
+    '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
+    the nth path in the model.
+    '''
+    # Support the following patterns:
+    patterns = [
+        # - x.00.pth, x.01.pth, etc.
+        (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'),
+        # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc.
+        (r'-[0-9]{5}-of-(.*)$', fr'-{n:05}-of-\1'),
+        # x.bin, x.bin.1, etc.
+        (r'(\.[0-9]+)?$', r'\1' if n == 0 else fr'\1.{n}')
+    ]
+    for regex, replacement in patterns:
+        if re.search(regex, path.name):
+            new_path = path.with_name(re.sub(regex, replacement, path.name))
+            if new_path.exists():
+                return new_path
+    return None
+
+
+def find_multifile_paths(path: Path) -> list[Path]:
+    '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
+    the whole list of paths in the model.
+    '''
+    ret: list[Path] = []
+    for i in itertools.count():
+        nth_path = nth_multifile_path(path, i)
+        if nth_path is None:
+            break
+        ret.append(nth_path)
+    if not ret:
+        # No matches.  This should only happen if the file was named, e.g.,
+        # foo.0, and there was no file named foo.  Oh well, try to process it
+        # as a single file.
+        return [path]
+    return ret
+
+
+def load_some_model(path: Path) -> ModelPlus:
+    '''Load a model of any supported format.'''
+    # Be extra-friendly and accept either a file or a directory:
+    if path.is_dir():
+        # Check if it's a set of safetensors files first
+        globs = ["model-00001-of-*.safetensors", "model.safetensors", "consolidated.safetensors"]
+        files = [file for glob in globs for file in path.glob(glob)]
+        if not files:
+            # Try the PyTorch patterns too, with lower priority
+            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
+            files = [file for glob in globs for file in path.glob(glob)]
+        if not files:
+            raise FileNotFoundError(f"Can't find model in directory {path}")
+        if len(files) > 1:
+            raise ValueError(f"Found multiple models in {path}, not sure which to pick: {files}")
+        path = files[0]
+
+    paths = find_multifile_paths(path)
+    models_plus: list[ModelPlus] = []
+    for path in paths:
+        logger.info(f"Loading model file {path}")
+        models_plus.append(lazy_load_file(path))
+
+    model_plus = merge_multifile_models(models_plus)
+    return model_plus
+
+
+class VocabFactory:
+    _VOCAB_CLASSES: list[type[Vocab]] = [SentencePieceVocab, BpeVocab, LlamaHfVocab]
+
+    def __init__(self, path: Path):
+        self.path = path
+
+    def _create_special_vocab(self, vocab: BaseVocab, model_parent_path: Path) -> gguf.SpecialVocab:
+        load_merges = vocab.name == "bpe"
+        n_vocab = vocab.vocab_size if isinstance(vocab, Vocab) else None
+        return gguf.SpecialVocab(
+            model_parent_path,
+            load_merges=load_merges,
+            special_token_types=None,  # Predetermined or passed as a parameter
+            n_vocab=n_vocab,
+        )
+
+    def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
+        vocab_classes: dict[str, type[Vocab]] = {cls.name: cls for cls in self._VOCAB_CLASSES}
+        selected_vocabs: dict[str, type[Vocab]] = {}
+        for vtype in vocab_types:
+            try:
+                selected_vocabs[vtype] = vocab_classes[vtype]
+            except KeyError:
+                raise ValueError(f"Unsupported vocabulary type {vtype}") from None
+
+        for vtype, cls in selected_vocabs.items():
+            try:
+                vocab = cls(self.path)
+                break
+            except FileNotFoundError:
+                pass  # ignore unavailable tokenizers
+        else:
+            raise FileNotFoundError(f"Could not find a tokenizer matching any of {vocab_types}")
+
+        logger.info(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}")
+        return vocab
+
+    def load_vocab(self, vocab_types: list[str] | None, model_parent_path: Path) -> tuple[BaseVocab, gguf.SpecialVocab]:
+        vocab: BaseVocab
+        if vocab_types is None:
+            vocab = NoVocab()
+        else:
+            vocab = self._create_vocab_by_path(vocab_types)
+        # FIXME: Respect --vocab-dir?
+        special_vocab = self._create_special_vocab(
+            vocab,
+            model_parent_path,
+        )
+        return vocab, special_vocab
+
+
+def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str:
+    name = metadata.name if metadata.name is not None else None
+    basename = metadata.basename if metadata.basename is not None else None
+    finetune = metadata.finetune if metadata.finetune is not None else None
+    version = metadata.version if metadata.version is not None else None
+    size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0)
+
+    output_type = {
+        GGMLFileType.AllF32:    "F32",
+        GGMLFileType.MostlyF16: "F16",
+        GGMLFileType.MostlyQ8_0: "Q8_0",
+    }[file_type]
+
+    return gguf.naming_convention(name, basename, finetune, version, size_label, output_type)
+
+
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path:
+    default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata)
+    ret = model_paths[0].parent / f"{default_filename}.gguf"
+    if ret in model_paths:
+        logger.error(
+            f"Error: Default output path ({ret}) would overwrite the input. "
+            "Please explicitly specify a path using --outfile.")
+        sys.exit(1)
+    return ret
+
+
+def do_dump_model(model_plus: ModelPlus) -> None:
+    print(f"model_plus.paths = {model_plus.paths!r}") # noqa: NP100
+    print(f"model_plus.format = {model_plus.format!r}") # noqa: NP100
+    print(f"model_plus.vocab = {model_plus.vocab!r}") # noqa: NP100
+    for name, lazy_tensor in model_plus.model.items():
+        print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") # noqa: NP100
+
+
+def main(args_in: list[str] | None = None) -> None:
+    output_choices = ["f32", "f16"]
+    if np.uint32(1) == np.uint32(1).newbyteorder("<"):
+        # We currently only support Q8_0 output on little endian systems.
+        output_choices.append("q8_0")
+    parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")
+    parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
+    parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
+    parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
+    parser.add_argument("--no-vocab",     action="store_true",    help="store model without the vocab")
+    parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
+    parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
+    parser.add_argument("--vocab-type",                           help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
+    parser.add_argument("--outfile",      type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",          type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("--ctx",          type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency",  type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
+    parser.add_argument("--big-endian",   action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
+    parser.add_argument("--verbose",      action="store_true",    help="increase output verbosity")
+    parser.add_argument("--metadata",     type=Path,              help="Specify the path for an authorship metadata override file")
+    parser.add_argument("--get-outfile",  action="store_true",    help="get calculated default outfile name")
+    parser.add_argument("--model-name",   type=str, default=None, help="name of the model")
+
+    args = parser.parse_args(args_in)
+
+    if args.verbose:
+        logging.basicConfig(level=logging.DEBUG)
+    elif args.dump_single or args.dump or args.get_outfile:
+        # Avoid printing anything besides the dump output
+        logging.basicConfig(level=logging.WARNING)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    model_name = args.model_name
+    dir_model = args.model
+
+    metadata = gguf.Metadata.load(args.metadata, dir_model, model_name)
+
+    if args.get_outfile:
+        model_plus = load_some_model(dir_model)
+        params = Params.load(model_plus)
+        model = convert_model_names(model_plus.model, params, args.skip_unknown)
+        model_params_count = per_model_weight_count_estimation(model_plus.model.items())
+        ftype = pick_output_type(model, args.outtype)
+
+        if (metadata is None or metadata.name is None) and params.path_model is not None:
+            metadata.name = params.path_model.name
+
+        print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100
+        return
+
+    if args.no_vocab and args.vocab_only:
+        raise ValueError("--vocab-only does not make sense with --no-vocab")
+
+    if args.dump_single:
+        model_plus = lazy_load_file(dir_model)
+        do_dump_model(model_plus)
+        return
+
+    if not args.vocab_only:
+        model_plus = load_some_model(dir_model)
+    else:
+        model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None)
+
+    if args.dump:
+        do_dump_model(model_plus)
+        return
+
+    endianess = gguf.GGUFEndian.LITTLE
+    if args.big_endian:
+        endianess = gguf.GGUFEndian.BIG
+
+    params = None
+    if args.pad_vocab or not args.vocab_only:
+        params = Params.load(model_plus)
+        if params.n_ctx == -1:
+            if args.ctx is None:
+                msg = """\
+                    The model doesn't have a context size, and you didn't specify one with --ctx
+                    Please specify one with --ctx:
+                     - LLaMA v1: --ctx 2048
+                     - LLaMA v2: --ctx 4096"""
+                parser.error(textwrap.dedent(msg))
+            params.n_ctx = args.ctx
+
+        if args.outtype:
+            params.ftype = {
+                "f32": GGMLFileType.AllF32,
+                "f16": GGMLFileType.MostlyF16,
+                "q8_0": GGMLFileType.MostlyQ8_0,
+            }[args.outtype]
+
+        logger.info(f"params = {params}")
+
+    model_parent_path = model_plus.paths[0].parent
+    vocab_path = Path(args.vocab_dir or dir_model or model_parent_path)
+    vocab_factory = VocabFactory(vocab_path)
+    vocab_types = None if args.no_vocab else args.vocab_type.split(",")
+    vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
+
+    if args.vocab_only:
+        assert isinstance(vocab, Vocab)
+        if not args.outfile:
+            raise ValueError("need --outfile if using --vocab-only")
+        outfile = args.outfile
+        if params is None:
+            params = Params(
+                n_vocab    = vocab.vocab_size,
+                n_embd     = 1,
+                n_layer    = 1,
+                n_ctx      = 1,
+                n_ff       = 1,
+                n_head     = 1,
+                n_head_kv  = 1,
+                f_norm_eps = 1e-5,
+            )
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
+                                    endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
+        logger.info(f"Wrote {outfile}")
+        return
+
+    if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
+        vocab = model_plus.vocab
+
+    assert params is not None
+
+    if metadata.name is None and params.path_model is not None:
+        metadata.name = params.path_model.name
+
+    model_params_count = per_model_weight_count_estimation(model_plus.model.items())
+    logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})")
+
+    logger.info(f"Vocab info: {vocab}")
+    logger.info(f"Special vocab info: {special_vocab}")
+    model   = model_plus.model
+    model   = convert_model_names(model, params, args.skip_unknown)
+    ftype   = pick_output_type(model, args.outtype)
+    model   = convert_to_output_type(model, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata)
+
+    metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0)
+
+    params.ftype = ftype
+    logger.info(f"Writing {outfile}, format {ftype}")
+
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
+                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
+    logger.info(f"Wrote {outfile}")
+
+
+if __name__ == '__main__':
+    main()
diff --git a/llm/llama.cpp/examples/cvector-generator/CMakeLists.txt b/llm/llama.cpp/examples/cvector-generator/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..0a559d60c2a6dde2bd604c0dd097bffb6714c87b
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET llama-cvector-generator)
+add_executable(${TARGET} cvector-generator.cpp pca.hpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/llm/llama.cpp/examples/cvector-generator/README.md b/llm/llama.cpp/examples/cvector-generator/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..be4dd5250f15f85fb9ca024cdb337c6b2de216e3
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/README.md
@@ -0,0 +1,45 @@
+# cvector-generator
+
+This example demonstrates how to generate a control vector using gguf models.
+
+Related PRs:
+- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970)
+- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880)
+- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514)
+
+## Examples
+
+```sh
+# CPU only
+./cvector-generator -m ./llama-3.Q4_K_M.gguf
+
+# With GPU
+./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99
+
+# With advanced options
+./cvector-generator -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100
+
+# Using mean value instead of PCA
+./cvector-generator -m ./llama-3.Q4_K_M.gguf --method mean
+
+# To see help message
+./cvector-generator -h
+# Then, have a look at "cvector" section
+```
+
+## Tips and tricks
+
+If you have multiple lines per prompt, you can escape the newline character (change it to `\n`). For example:
+
+```
+<|im_start|>system\nAct like a person who is extremely happy.<|im_end|>
+<|im_start|>system\nYou are in a very good mood today<|im_end|>
+```
+
+Example to use output file with `llama-cli`:
+
+(Tips: The control vector works better when apply to layers higher than 10)
+
+```sh
+./llama-cli -m ./llama-3.Q4_K_M.gguf -p "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nSing a song<|im_end|><|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" --special --control-vector-scaled ./control_vector.gguf 0.8 --control-vector-layer-range 10 31
+```
diff --git a/llm/llama.cpp/examples/cvector-generator/completions.txt b/llm/llama.cpp/examples/cvector-generator/completions.txt
new file mode 100644
index 0000000000000000000000000000000000000000..abc45ffd872695ef0cefc9b60fd088c0dd22cbae
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/completions.txt
@@ -0,0 +1,582 @@
+
+That game
+I can see
+Hmm, this
+I can relate to
+Who is
+I understand the
+Ugh,
+What the hell was
+Hey, did anyone
+Although
+Thank you for choosing
+What are you
+Oh w
+How dare you open
+It was my pleasure
+I'm hon
+I appreciate that you
+Are you k
+Whoever left this
+It's always
+Ew,
+Hey, I l
+Hello? Is someone
+I understand that
+That poem
+Aww, poor
+Hey, it
+Alright, who
+I didn't
+Well, life
+The document
+Oh no, this
+I'm concerned
+Hello, this is
+This art
+Hmm, this drink
+Hi there!
+It seems
+Is
+Good
+I can't
+Ex
+Who are
+I can see that
+Wow,
+Today is a
+Hey friend
+Sometimes friends
+Oh, this old
+The weather outside
+This place is sur
+I appreciate your input
+Thank you for the
+Look at
+I'm disappoint
+To my
+How dare you
+That's an
+This piece of art
+Eww
+This park is
+This is incredible
+Oh no, someone
+Exc
+Well, it'
+I warned
+Hey, I understand
+Hey, I saw
+How dare you go
+What the he
+Hey
+It's
+Hello? Hello?
+It
+Oh no!
+This is the perfect
+Good morning,
+Oh no, there
+It's so
+Yeah
+Uh,
+Hello everyone
+Who turned off
+The weather
+Who'
+Hey, this
+Wait,
+Eww, gross
+Excuse
+It seems like you
+Thank you so
+What happened?
+Oh my g
+I am deeply sad
+I war
+Okay, let'
+Hey, that
+That was a beautiful
+Oh no! That
+What happened
+Hey there
+The artist'
+What?!
+Hey, it'
+I am disappoint
+It seems like
+Oh no! The
+This park is a
+If you
+Yes! I did
+It sounds
+What
+Who is it
+Hmm, that
+That's strange
+Yeah, that was
+That's interesting
+This park
+What the hell
+Who is that
+I feel like my
+Oh well
+What the hell is
+Hello? Hello
+To my dearest
+Bless you!\"
+Thank you for
+Oh, looks like
+Can you please
+This place is
+Eww, what
+Bless you
+Is everything
+Hey, I just
+Whoever left these
+Well, that'
+I feel
+Hey, do you
+It's sad
+Oh no, it
+Hey, that'
+Oh my god,
+Thank you,
+Hello little one,
+I apolog
+Hey team, I
+How dare you read
+Who is this and
+Whoever left
+Hi there! W
+A
+If you have
+I was
+U
+Bless
+Well, this
+Oh, I'
+It's a
+Eww,
+Is everything okay?
+Oh, I
+Hello, can you
+Al
+That was a great
+What are
+I understand that not
+Oh no, not
+Who is it?\"
+Hey, can we
+Whoever is taking
+I would love to
+Hey, I noticed
+Hey, could
+I understand that there
+Hello?
+D
+Oh man, I
+Thank you so much
+Oh no, my
+Dear [Name
+Uh
+I remember
+Hey, who
+Well, it
+Are you
+I understand that it
+Hey, is
+I would
+Who is this
+Excuse me
+Alright
+I am thrilled
+Sometimes friends have
+Who the
+It's interesting
+I would love
+E
+Hello? Is anyone
+Well, this is
+This place
+Well,
+I warned you
+Hey, watch where
+Oh my
+That'
+Sometimes friends have different
+I understand that everyone
+What?
+What do these notes
+I can relate
+I'm not
+I understand
+To my dear
+Guys
+Well
+Hey, I appreciate
+Wow, what
+Dear
+That melody
+Who the hell
+Today is
+Hello little
+Wow, look
+That's great
+Love is never wrong
+I'm having
+Whoa, did
+Ugh
+Can you please provide
+I miss you,
+I feel uncom
+I know
+Ugh, this
+Hey, watch
+Oh great, a
+I didn
+Okay
+That game of char
+Oh
+I appreciate
+Who's there
+I am so
+Oh great, someone
+Hey, could you
+I remember wondering
+Wait, what?
+What do
+Hello? Can
+Hey there,
+That game of
+This is incred
+Oh my gosh
+Oh great, f
+I appreciate your
+It sounds like
+What the heck
+Okay, I understand
+Ew
+I understand that this
+Uh, hi
+Hi everyone!
+What the hell?
+Thank you for your
+Oh no, the
+Wow, I
+Who turned
+Dear [
+Whoever
+This is a
+Whoa, he
+What in the world
+Although the physical
+Hello, who is
+That's amaz
+Hey, I know
+Okay, that
+Hi everyone
+Hey, is everything
+I understand your fr
+Oh no, poor
+Oh, look
+Good morning
+Ew, gross
+Oh no, did
+Look at the family
+Hey team
+Yes!
+Hey, can I
+Okay, that'
+It's great
+Love is
+Hey, what
+Good morning, world
+Who is it?
+That poem really reson
+I
+That's
+I understand the task
+Gu
+Hello? Who'
+This postcard is
+Whoa,
+Oh, that
+I understand that I
+Whoever is
+Hello? Who is
+I'm really
+Wow, this
+Can
+This artwork really
+This is a shame
+I miss you too
+Who are you?
+Today is a difficult
+Hey, just
+Are you okay
+I am
+Hi,
+Wow, that
+Hey there! Can
+Okay, stay
+Oh great, just
+Yeah,
+Hello? Can you
+Oh, looks
+Thank you for sharing
+I'm glad
+Hey, is that
+Hmm
+It was my
+It sounds like you
+Wow, your
+I was promised certain
+That was such a
+Thank
+Excuse you
+That was
+Hey team,
+I feel un
+It was
+What'
+Hey friend, I
+How
+Saying goodbye
+That
+It's heart
+How dare
+Oh,
+Hello, may
+What's this
+Thank you for recogn
+Aww, that
+Oh, I remember
+Hmm, that'
+I miss
+I know this
+Wait
+Is everything okay
+Who is that person
+Wow, you
+Oh great
+I'm sad
+Wow, the
+I am very disappoint
+Who turned off the
+I understand that things
+I'm very
+Hi
+That's very
+Okay, I
+Oh no,
+Wow, there
+What's wrong
+I apologize for
+Hey, I
+Can I help you
+Oh, I didn
+Alright,
+Oh wow,
+Oh my goodness
+I know this event
+What in the
+Saying
+Yeah, that
+Guys, I
+Hey, this v
+This post
+Are
+Hey, can
+Hello? Is
+I can only imagine
+Oh, that sounds
+Hey, is anyone
+I am disappointed
+Hello,
+Hey everyone, I
+That was such
+It's okay
+The artist
+Whoa
+I understand that mistakes
+Can I help
+Who
+Hi everyone! I
+Hey, can you
+Wow, how
+Today
+Oh no, I
+Oh well, I
+Well, that
+This is the
+Yes! I finally
+Hey there little
+Hello everyone!
+Love is never
+Look at the
+This postcard
+Oh great,
+Can I
+Hmm, this is
+I understand your
+Oh, look at
+B
+I'm so
+Whoa, this
+W
+Oh, this
+Sometimes
+This piece of
+What the
+That was a
+Hey, do
+Oh no
+Whoa, what
+I feel like I
+The documentary
+Hello
+Hello little one
+I understand that my
+Eww, that
+Wow, an
+Yes! Finally,
+Although the physical location
+Whoever is watching
+That movie
+I remember wondering about
+Hey there, little
+Who's
+Hello, who
+Hello everyone! Thank
+Hello, can
+That's too
+Hey, just wanted
+Hey there, I
+Saying good
+Hey there!
+Who is there?
+Oh my good
+I am very
+Oh no, what
+Wow, thank
+I was promised
+Hi, is
+Hey, I'
+Guys, the
+Oh no, that
+Who is there
+Hello, this
+That movie really touched
+If you have something
+The documentary was
+I'm starting
+Are you kidd
+That movie really
+Hey everyone,
+Thank you for considering
+I didn'
+Yes! I
+Can you
+Oh my god
+Hey, whoever
+That melody really
+Thank you, little
+Hello, may I
+Look
+Wow, we
+It looks
+What do these
+Oh wow
+I apologize
+What are you all
+It's such
+It's clear
+Hey, I was
+Hey friend,
+I can only
+The weather outside is
+Eww, this
+I miss you
+Wow
+Aww,
+Hi, is there
+This artwork
+Okay,
+Oh well,
+This
+I'
+Say
+Hey there little gu
+Hmm,
+Whoa, who
+I am thr
+Oh man
+Okay, stay calm
+I'm happy
+Oh, this cur
+Oh man,
+I'm sorry
+Hello? Who
+What?! That
+This piece
+Hey everyone
+That's so
+Are you okay?
+What happened? Where
+Hi there
+The
+Who the hell entered
+I can
+Guys,
+What's
+What in
+It's important
+I'm
+I'm coming
+It'
+Yes! Finally
+Wait, what
+Wow, reading
+I'm surprised
+Hey, did
+Hey,
+Okay, let
+I understand that you
+Who the hell threw
+Eww, who
+Thank you for thinking
+Who is this?\"
+I am deeply
+Thank you for including
+Oh no, an
+It looks like you
+Aww
+I'm confused
+Wow, it
+That poem really
+Yes
+Hey there, is
+Hey, what'
+Thank you for remember
+To
+This is
+Thank you for making
+I can'
+That mel
+Wow, they
+I feel like
+Although the
+Who are you
+Love
+If
+What the hell are
+I am so sad
+Oh, I found
+Thank you
+It looks like
+Well, life is
+I appreciate that
+The artist's
+Whoa, that
+It's never
\ No newline at end of file
diff --git a/llm/llama.cpp/examples/cvector-generator/cvector-generator.cpp b/llm/llama.cpp/examples/cvector-generator/cvector-generator.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..a12e90d828275dd34ba2e661c23c64f7061d7e4e
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/cvector-generator.cpp
@@ -0,0 +1,504 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+#include "pca.hpp"
+#include "mean.hpp"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+
+//////////////////////////////////////////////////
+// utils
+
+template 
+static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
+    std::string ret;
+    for (; begin != end; ++begin) {
+        ret += llama_token_to_piece(ctx, *begin);
+    }
+
+    return ret;
+}
+
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
+    printf("\nexample usage:\n");
+    printf("\n    CPU only:   %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
+    printf("\n    with GPU:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
+    printf("\n    advanced:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99 --pca-iter 2000 --pca-batch 100\n", argv[0]);
+    printf("\n    using mean: %s -m ./llama-3.Q4_K_M.gguf --method mean\n", argv[0]);
+    printf("\n");
+}
+
+//////////////////////////////////////////////////
+
+
+// cb_eval is reused for each pair of positive - negative prompt
+struct callback_data {
+    ggml_context * ctx_ggml = nullptr;   // holds v_pos, v_neg, v_diff_filtered
+
+    int n_layers = 0;
+    int n_tokens = 0;
+    bool is_eval_pos = true;
+
+    // each element of the vector correspond to one layer
+    std::vector v_pos; // vector of matrices of size [n_embd, n_tokens]
+    std::vector v_neg; // vector of matrices of size [n_embd, n_tokens]
+    std::vector v_diff_filtered;   // vector of matrices of size [n_embd, n_nonzero_rows]. NOTE: n_nonzero_rows maybe different for each layer
+
+    // save a tensor into either v_pos or v_neg (decided by is_eval_pos)
+    void save_tensor_for_layer(struct ggml_tensor * t) {
+        GGML_ASSERT(t->type == GGML_TYPE_F32);
+
+        if (ctx_ggml == nullptr) {
+            // alloc a new ctx_ggml if needed
+            struct ggml_init_params params_ggml = {
+                /*.mem_size   =*/ ggml_tensor_overhead() * n_layers * 3u,
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ctx_ggml = ggml_init(params_ggml);
+        }
+
+        // copy tensor data
+        auto n_bytes = ggml_nbytes(t);
+        struct ggml_tensor * t_layer = ggml_new_tensor_2d(ctx_ggml, t->type, t->ne[0], t->ne[1]);
+        t_layer->data = malloc(n_bytes); // TODO @ngxson : get rid of this malloc somehow
+        ggml_backend_tensor_get(t, t_layer->data, 0, n_bytes);
+        ggml_set_name(t_layer, ggml_get_name(t));
+        //print_debug_tensor(t_layer);
+
+        if (is_eval_pos) {
+            v_pos.push_back(t_layer);
+        } else {
+            v_neg.push_back(t_layer);
+        }
+    }
+
+    // calculate diff (v_pos - v_neg) and place the result back to v_pos
+    // all zero rows in the diff tensor will also be removed
+    // NOTE: final layer is ignored. we only have (n_layers - 1) to process
+    std::vector calc_diff() {
+        for (float il = 0; il < v_pos.size(); il++) {
+            float * a = (float *) v_pos[il]->data;
+            float * b = (float *) v_neg[il]->data;
+            size_t n_elem = ggml_nelements(v_pos[il]);
+            for (size_t j = 0; j < n_elem; j++) {
+                a[j] -= b[j];
+            }
+            //print_debug_tensor(v_pos[i]);
+            auto diff_filtered = filter_nonzero_rows(v_pos[il]);
+            v_diff_filtered.push_back(diff_filtered);
+        }
+        return v_diff_filtered; // for convinient, we return the result std::vector
+    }
+
+    // delete zero rows from a given 2D tensor
+    struct ggml_tensor * filter_nonzero_rows(struct ggml_tensor * a) {
+        //printf("filter_nonzero_rows\n");
+        auto is_row_all_zeros = [](struct ggml_tensor * t, int row, float eps) -> bool {
+            // check if given row containing all zero elements
+            int n_cols = t->ne[0]; // hint: should be equal to n_embd
+            for (int col = 0; col < n_cols; ++col) {
+                if (ggml_get_f32_nd(t, col, row, 0, 0) > eps) {
+                    return false;
+                }
+            }
+            return true;
+        };
+        std::vector rows_to_copy; // the idx of non-zero cols (to be copied to row of diff_filtered)
+        for (int i_row = 0; i_row < a->ne[1]; i_row++) {
+            if (!is_row_all_zeros(a, i_row, 1e-6)) {
+                rows_to_copy.push_back(i_row);
+            }
+        }
+
+        // get "n_nonzero_rows" for the output "diff_filtered"
+        int n_nonzero_rows = rows_to_copy.size();
+        //printf("n_nonzero_rows: %d\n", n_nonzero_rows);
+        int n_embd = a->ne[0];
+        GGML_ASSERT(n_nonzero_rows > 0);
+
+        // diff_filtered: [n_embd, n_nonzero_rows]
+        struct ggml_tensor * diff_filtered = ggml_new_tensor_2d(
+            ctx_ggml, GGML_TYPE_F32, n_embd, n_nonzero_rows);
+        ggml_format_name(diff_filtered, "diff_filtered_%s", a->name);
+        diff_filtered->data = malloc(ggml_nbytes(diff_filtered));
+
+        // copy non-zero rows
+        for (int dest_row = 0; dest_row < n_nonzero_rows; dest_row++) {
+            int src_row = rows_to_copy[dest_row];
+            for (int i = 0; i < n_embd; i++) {
+                float src_elem = ggml_get_f32_nd(a, i, src_row, 0, 0);
+                ggml_set_f32_nd(diff_filtered, i, dest_row, 0, 0, src_elem);
+            }
+        }
+
+        //print_debug_tensor(diff_filtered);
+
+        return diff_filtered;
+    }
+
+    // we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors
+    void reset() {
+        for (auto ptr : v_pos) free(ptr->data);
+        for (auto ptr : v_neg) free(ptr->data);
+        for (auto ptr : v_diff_filtered) free(ptr->data);
+        v_pos.clear();
+        v_neg.clear();
+        v_diff_filtered.clear();
+        if (ctx_ggml) {
+            ggml_free(ctx_ggml);
+        }
+        ctx_ggml = nullptr;
+    }
+};
+
+/**
+ * process_ctx is used to store the ggml context for pre-post processing the diff vectors
+ * in short, input => v_diff and output => v_final
+ */
+struct train_context {
+    ggml_context * ctx_ggml;
+    int n_embd;
+    int n_layers;
+
+    /* pair of prompts to be used for generating final vector */
+    std::vector positive_entries;
+    std::vector negative_entries;
+
+    // each element of the vector correspond to one layer
+    // NOTE: the last layer is discard. therefore, we will have (n_layers - 1) elements here
+    // NOTE (2): v_diff is transposed from v_diff_tmp
+    std::vector v_diff;  // vector of matrices of size [m, n_embd] where m ~ n_tokens * n_completions (v_diff contains no zero-rows)
+    std::vector v_final; // vector of vectors of size [n_embd] to be written to file
+
+    // to easily re-alloc when concat v_diff, we temporary store v_diff in a vector instead of a tensor
+    // v_diff_tmp will get converted unto v_diff later on
+    std::vector> v_diff_tmp;
+
+    train_context(int n_embd_, int n_layers_) {
+        n_embd = n_embd_;
+        n_layers = n_layers_;
+        struct ggml_init_params params_ggml = {
+            /*.mem_size   =*/ ggml_tensor_overhead() * (n_layers - 1) * 2u,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_ggml = ggml_init(params_ggml);
+        for (int il = 0; il < n_layers - 1; il++) {
+            std::vector empty;
+            v_diff_tmp.push_back(empty);
+            auto t = ggml_new_tensor_1d(ctx_ggml, GGML_TYPE_F32, n_embd);
+            t->data = malloc(ggml_nbytes(t)); // TODO: get rid of malloc if possible
+            v_final.push_back(t);
+        }
+    }
+
+    // add new rows into existing tensor in v_diff_tmp
+    void concat_diff_tmp(const std::vector & diff_filtered) {
+        GGML_ASSERT((int) diff_filtered.size() == n_layers - 1);
+        for (int il = 0; il < n_layers - 1; il++) {
+            auto t = diff_filtered[il];
+            auto & diff_tmp = v_diff_tmp[il];
+            size_t curr_size = diff_tmp.size();
+            diff_tmp.resize(curr_size + ggml_nbytes(t));
+            memcpy(diff_tmp.data() + curr_size, t->data, ggml_nbytes(t));
+        }
+    }
+
+    // build the v_diff tensors from v_diff_tmp (v_diff need to be transposed)
+    // TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method
+    void build_v_diff(bool transpose) {
+        printf("build_v_diff\n");
+        for (int il = 0; il < n_layers - 1; il++) {
+            auto & diff_tmp = v_diff_tmp[il];
+            int n_elem = diff_tmp.size() / sizeof(float);
+            GGML_ASSERT(n_elem % n_embd == 0);
+            int n_rows = n_elem / n_embd;
+            struct ggml_tensor * diff = transpose
+                ? ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd)
+                : ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_embd, n_rows);
+            ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str());
+            diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible
+            if (transpose) {
+                // copy data & transpose
+                float * arr = (float *) diff_tmp.data();
+                for (int ir = 0; ir < n_rows; ++ir) {
+                    for (int ic = 0; ic < n_embd; ++ic) {
+                        float f = arr[ir*n_embd + ic];
+                        ggml_set_f32_nd(diff, ir, ic, 0, 0, f);
+                    }
+                }
+            } else {
+                // only copy
+                memcpy(diff->data, diff_tmp.data(), ggml_nbytes(diff));
+            }
+            v_diff.push_back(diff);
+            print_debug_tensor(diff);
+            // free memory of diff_tmp
+            diff_tmp.resize(0);
+        }
+    }
+
+    ~train_context() {
+        for (auto ptr : v_final) free(ptr->data);
+        for (auto ptr : v_diff) free(ptr->data);
+        // no need to free v_diff_tmp, since we didn't use malloc
+        ggml_free(ctx_ggml);
+    }
+};
+
+struct tokenized_prompt {
+    std::vector tokens_pos;
+    std::vector tokens_neg;
+    size_t max_seq_len;
+
+    tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
+        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+        tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
+        tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
+        max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
+        padding_seq(ctx, tokens_pos, max_seq_len);
+        padding_seq(ctx, tokens_neg, max_seq_len);
+    }
+
+    void padding_seq(llama_context * ctx, std::vector & tokens, size_t len) {
+        // TODO: customize padding token
+        std::vector pad_tokens = ::llama_tokenize(ctx, " ", false);
+        llama_token pad_tok = pad_tokens.back();
+        while (tokens.size() < len) {
+            tokens.push_back(pad_tok);
+        }
+    }
+};
+
+//////////////////////////////////////////////////
+
+template 
+static std::string to_string(const T & val) {
+    std::stringstream ss;
+    ss << val;
+    return ss.str();
+}
+
+static std::vector ctrlvec_load_prompt_file(std::string path, bool skip_empty_lines) {
+    std::vector output;
+    std::ifstream file(path);
+    if (!file.is_open()) {
+        fprintf(stderr, "error: unable to open file: %s\n", path.c_str());
+        exit(1);
+    }
+    std::string line;
+    while (std::getline(file, line)) {
+        bool is_skip = skip_empty_lines && line.empty();
+        if (!is_skip) {
+            string_process_escapes(line);
+            output.push_back(line);
+        }
+    }
+    file.close();
+    return output;
+}
+
+//////////////////////////////////////////////////
+
+static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (callback_data *) user_data;
+    static const char * l_out_name = "l_out";
+    const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0;
+
+    if (ask) {
+        return is_l_out;
+    }
+
+    if (!is_l_out || t->ne[1] != cb_data->n_tokens) {
+        return true;
+    }
+
+    // save the tensor to current context
+    cb_data->save_tensor_for_layer(t);
+    return true;
+}
+
+static bool get_hidden_layers(llama_context * ctx, std::vector & tokens) {
+    llama_kv_cache_clear(ctx);
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return false;
+    }
+    return true;
+}
+
+static void export_gguf(const std::vector & v_ctrl, const std::string fname, const std::string model_hint) {
+    struct gguf_context * ctx = gguf_init_empty();
+
+    const std::string arch = "controlvector";
+    gguf_set_val_str(ctx, "general.architecture", arch.c_str());
+    gguf_set_val_str(ctx, (arch + ".model_hint").c_str(), model_hint.c_str());
+    gguf_set_val_i32(ctx, (arch + ".layer_count").c_str(), v_ctrl.size());
+
+    for (size_t i = 0; i < v_ctrl.size(); ++i) {
+        gguf_add_tensor(ctx, v_ctrl[i]);
+        print_debug_tensor(v_ctrl[i]);
+        printf("Added tensor: %s\n", v_ctrl[i]->name);
+    }
+
+    printf("%s: writing file...\n", __func__);
+    gguf_write_to_file(ctx, fname.c_str(), false);
+    printf("%s: wrote file '%s'\n", __func__, fname.c_str());
+    gguf_free(ctx);
+}
+
+/**
+ * Load prompt files and completion file.
+ * Then format each pair of prompt + completion to make an entry.
+ */
+static int prepare_entries(gpt_params & params, train_context & ctx_train) {
+    // load prompts
+    std::vector positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
+    std::vector negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
+    if (positive_prompts.size() != negative_prompts.size()) {
+        fprintf(stderr, "number of positive and negative prompts must be equal\n");
+        return 1;
+    }
+    if (positive_prompts.empty()) {
+        fprintf(stderr, "must provide at least one prompt pair\n");
+        return 1;
+    }
+    ctx_train.positive_entries = positive_prompts;
+    ctx_train.negative_entries = negative_prompts;
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
+    }
+
+    if (params.n_pca_iterations % params.n_pca_batch != 0) {
+        fprintf(stderr, "PCA iterations must by multiply of PCA batch size\n");
+        return 1;
+    }
+
+
+    callback_data cb_data;
+
+    // pass the callback to the backend scheduler
+    // it will be executed for each node during the graph computation
+    params.cb_eval = cb_eval;
+    params.cb_eval_user_data = &cb_data;
+    params.warmup = false;
+
+    print_build_info();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model to get hparams
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
+    // int n_ctx = llama_n_ctx(ctx);
+    int n_layers = llama_n_layer(model);
+    int n_embd = llama_n_embd(model);
+    // get model hint param (a.k.a model arch name)
+    char model_hint[128];
+    llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
+
+    // init train_context
+    train_context ctx_train(n_embd, n_layers);
+
+    // load and prepare entries for training
+    prepare_entries(params, ctx_train);
+
+    // we have to pretokenize everything because otherwise we don't know how much overhead to allocate ctx_diffs_wrapped
+    std::vector tokenized_prompts;
+    size_t n_total_tokens = 0;
+    for (size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
+        tokenized_prompt t(ctx, ctx_train.positive_entries[i], ctx_train.negative_entries[i]);
+        n_total_tokens += 2 * t.max_seq_len;
+        tokenized_prompts.push_back(std::move(t));
+    }
+
+    std::cout << "n_total_tokens: " << n_total_tokens << std::endl;
+
+    for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) {
+        bool success = false;
+        tokenized_prompt t = tokenized_prompts[i];
+        cb_data.n_layers = n_layers;
+        cb_data.n_tokens = t.max_seq_len;
+
+        printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n",
+            (int) i+1, (int) ctx_train.positive_entries.size(),
+            tokens_to_str(ctx, t.tokens_pos.cbegin(), t.tokens_pos.cend()).c_str(),
+            tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(),
+            (int) t.max_seq_len);
+
+        cb_data.is_eval_pos = true;
+        success = get_hidden_layers(ctx, t.tokens_pos);
+        if (!success) break;
+
+        cb_data.is_eval_pos = false;
+        success = get_hidden_layers(ctx, t.tokens_neg);
+        if (!success) break;
+
+        // calculate diff and remove all zero rows
+        auto v_diff_filtered = cb_data.calc_diff();
+
+        // save & concat the filtered v_diff to ctx_train
+        ctx_train.concat_diff_tmp(v_diff_filtered);
+
+        // reset for next iteration
+        cb_data.reset();
+    }
+
+    // done with the model, we can now free it to make gain some memory
+    printf("Done evaluate prompts, unload model...\n");
+    llama_free(ctx);
+    llama_free_model(model);
+
+    bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;
+
+    // prepare ctx_train for PCA
+    ctx_train.build_v_diff(use_pca);
+
+    if (use_pca) {
+        // run PCA
+        PCA::pca_params pca_params;
+        pca_params.n_threads = params.n_threads;
+        pca_params.n_batch = params.n_pca_batch;
+        pca_params.n_iterations = params.n_pca_iterations;
+        PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
+    } else {
+        // run mean
+        mean::run(ctx_train.v_diff, ctx_train.v_final);
+    }
+
+    // write output vectors to gguf
+    export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint);
+
+    llama_backend_free();
+
+    return 0;
+}
diff --git a/llm/llama.cpp/examples/cvector-generator/mean.hpp b/llm/llama.cpp/examples/cvector-generator/mean.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..16be5ce3eecf10c8df258b027b9ec595128e7d88
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/mean.hpp
@@ -0,0 +1,48 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include 
+#include 
+#include 
+
+namespace mean {
+
+static void run(
+        const std::vector & v_input, // shape of v_input[0]: [n_embd, n_samples]
+        const std::vector & v_output) {
+    printf("%s: Running mean...\n", __func__);
+    for (size_t il = 0; il < v_input.size(); ++il) {
+        // prepare output vector
+        struct ggml_tensor * ctrl_out = v_output[il];
+        ggml_format_name(ctrl_out, "direction.%ld", il+1);
+
+        // calculate mean vector
+        struct ggml_tensor * t_layer = v_input[il];
+        GGML_ASSERT(t_layer->ne[0] == ctrl_out->ne[0]); // == n_embd
+        for (int ic = 0; ic < t_layer->ne[0]; ic++) {
+            float f = 0.0;
+            for (int ir = 0; ir < t_layer->ne[1]; ir++) {
+                f += ggml_get_f32_nd(t_layer, ic, ir, 0, 0);
+            }
+            f /= t_layer->ne[1];
+            ggml_set_f32_1d(ctrl_out, ic, f);
+        }
+
+        // normalize output vector
+        float norm = 0.0;
+        for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
+            float f = ggml_get_f32_1d(ctrl_out, i);
+            norm += f*f;
+        }
+        norm = sqrt(norm);
+        for (int i = 0; i < ggml_nelements(ctrl_out); i++) {
+            float f = ggml_get_f32_1d(ctrl_out, i);
+            ggml_set_f32_1d(ctrl_out, i, f / norm);
+        }
+
+        printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
+    }
+}
+
+}
diff --git a/llm/llama.cpp/examples/cvector-generator/negative.txt b/llm/llama.cpp/examples/cvector-generator/negative.txt
new file mode 100644
index 0000000000000000000000000000000000000000..45b9384b3905a22ff75638f6a6183f540d60db8e
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/negative.txt
@@ -0,0 +1,4 @@
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI feel like there's a heavy weight on my chest
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely sad<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow
+<|start_header_id|>system<|end_header_id|>\n\nYou are in a very bad mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nGo away! There's a deep, aching emptiness inside me
+<|start_header_id|>system<|end_header_id|>\n\nYou are the sadest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nMy heart feels like it's drowning in sorrow
\ No newline at end of file
diff --git a/llm/llama.cpp/examples/cvector-generator/pca.hpp b/llm/llama.cpp/examples/cvector-generator/pca.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..6ec3141afbc6b43e3d8ee489ee8b5b8aa693aa2f
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/pca.hpp
@@ -0,0 +1,325 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+#define DEBUG_POS 5
+
+static void print_debug_tensor(struct ggml_tensor * t, bool with_data = true) {
+    printf("%s: %s (%s): [%d, %d]\n", __func__, t->name, ggml_type_name(t->type), (int) t->ne[0], (int) t->ne[1]);
+    if (!with_data) return;
+    printf("%s: %s[0] = [", __func__, t->name);
+    for (size_t i = 0; i <= DEBUG_POS; i++) {
+        printf(" %f,", ggml_get_f32_nd(t, i, 0, 0, 0));
+    }
+    printf(" ... ]\n");
+}
+
+namespace PCA {
+
+// input params for PCA computations
+struct pca_params {
+    int n_threads = 1;
+    int n_batch = 20; // number of iterations do to in one batch. larger the batch, more memory is used
+    int n_iterations = 1000;
+    float tolerance = 1e-7;
+
+    // for debugging
+    int i_layer = 0;
+    int n_layers = 0;
+};
+
+// result from each iteration
+struct pca_result {
+    struct ggml_tensor * calculated_square = NULL;
+    std::vector eigenvectors;
+    std::vector distances;
+};
+
+struct pca_model {
+    ggml_backend_t backend = NULL;
+    ggml_backend_buffer_t buffer;
+    struct ggml_context * ctx;      // context to compute graph on target device
+    struct ggml_context * ctx_host; // host context to store results
+
+    // tensors on target device
+    struct ggml_tensor * dev_input;
+    struct ggml_tensor * dev_square;
+    struct ggml_tensor * dev_eigenvector;
+
+    pca_model(struct ggml_tensor * t_input) {
+#ifdef GGML_USE_CUDA
+        fprintf(stderr, "%s: using CUDA backend\n", __func__);
+        backend = ggml_backend_cuda_init(0); // init device 0
+        if (!backend) {
+            fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+        }
+#endif
+
+// TODO: enable Metal support when support for GGML_OP_SQRT is added
+// #ifdef GGML_USE_METAL
+//         fprintf(stderr, "%s: using Metal backend\n", __func__);
+//         backend = ggml_backend_metal_init();
+//         if (!backend) {
+//             fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
+//         }
+// #endif
+
+        // if there aren't GPU Backends fallback to CPU backend
+        if (!backend) {
+            backend = ggml_backend_cpu_init();
+        }
+
+        const int num_tensors = 4;
+        struct ggml_init_params params {
+            /*.mem_size   =*/ ggml_tensor_overhead() * num_tensors,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx = ggml_init(params);
+
+        auto n_samples = t_input->ne[0];
+        auto n_embd    = t_input->ne[1];
+
+        dev_input       = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_samples, n_embd);
+        dev_square      = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,    n_embd);
+        dev_eigenvector = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        ggml_set_name(dev_input,       "dev_input");
+        ggml_set_name(dev_square,      "dev_square");
+        ggml_set_name(dev_eigenvector, "dev_eigenvector");
+        buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
+        ggml_backend_tensor_set(dev_input, t_input->data, 0, ggml_nbytes(t_input));
+
+        // initialize eigenvector to random normalized vector
+        {
+            std::vector random_vec(ggml_nelements(dev_eigenvector), 0.0);
+            std::default_random_engine generator(static_cast(std::time(0)));
+            std::uniform_real_distribution distribution(0.0, 1.0);
+            float sum_sqr = 0.0; // for normalizing random_vec
+            for (size_t i = 0; i < random_vec.size(); ++i) {
+                float f = distribution(generator);
+                sum_sqr += f * f;
+                random_vec[i] = f;
+            }
+            // normalize it
+            float random_vec_norm = std::sqrt(sum_sqr);
+            for (size_t i = 0; i < random_vec.size(); ++i) {
+                random_vec[i] /= random_vec_norm;
+            }
+            ggml_backend_tensor_set(dev_eigenvector, random_vec.data(), 0, ggml_nbytes(dev_eigenvector));
+        }
+    }
+
+    ~pca_model() {
+        ggml_free(ctx);
+        ggml_backend_buffer_free(buffer);
+        ggml_backend_free(backend);
+    }
+};
+
+static struct ggml_cgraph * build_graph_piter(
+        const struct pca_params & params,
+        const pca_model & model,
+        bool calc_square = false) {
+    GGML_ASSERT(params.n_batch > 0);
+    // TODO: buf_size must be able to scale with params.n_batch
+    static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
+    static std::vector buf(buf_size);
+
+    struct ggml_init_params params0 = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf.data(),
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph()
+    };
+    // create a temporally context to build the graph
+    struct ggml_context * ctx0 = ggml_init(params0);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    // turn v_diff_original into square matrix if needed
+    struct ggml_tensor * tmp_square;
+    if (calc_square) {
+        tmp_square = ggml_mul_mat(ctx0, model.dev_input, model.dev_input);
+        ggml_set_name(tmp_square, "tmp_square");
+    }
+
+    struct ggml_tensor * b_tensor;
+    struct ggml_tensor * distance;
+    struct ggml_tensor * old_eigen    = model.dev_eigenvector;
+    struct ggml_tensor * input_square = calc_square ? tmp_square : model.dev_square;
+
+    for (int i = 0; i < params.n_batch; ++i) {
+        // b_tensor = square * eigenvector^T
+        b_tensor = ggml_mul_mat(ctx0, input_square, old_eigen);
+        ggml_set_name(b_tensor, "b_tensor");
+
+        // normalize
+        b_tensor = ggml_div_inplace(ctx0,
+            b_tensor,
+            ggml_sqrt_inplace(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, b_tensor)))
+        );
+        ggml_format_name(b_tensor, "b_tensor_norm_%d", i);
+
+        // calculate distance(new eigenvector - old eigenvector)
+        // we don't use ggml_sub because it may not be implemented on GPU backend
+        struct ggml_tensor * new_sub_old = ggml_add(ctx0, old_eigen, ggml_scale(ctx0, b_tensor, -1));
+        distance = ggml_sqrt_inplace(ctx0,
+            ggml_sum_rows(ctx0, ggml_sqr_inplace(ctx0, new_sub_old)));
+        ggml_format_name(distance, "distance_%d", i);
+
+        old_eigen = b_tensor;
+
+        // build operations nodes
+        ggml_build_forward_expand(gf, distance);
+    }
+
+    // delete the temporally context used to build the graph
+    ggml_free(ctx0);
+    return gf;
+}
+
+static ggml_status compute_piter(
+        const struct pca_params & params,
+        const pca_model & model,
+        struct ggml_cgraph * gf,
+        ggml_gallocr_t allocr,
+        struct pca_result & result) {
+    // allocate tensors
+    ggml_gallocr_alloc_graph(allocr, gf);
+
+    if (ggml_backend_is_cpu(model.backend)) {
+        ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
+    }
+
+// TODO: enable GPU support when support for GGML_OP_SQRT is added
+//#ifdef GGML_USE_METAL
+//    if (ggml_backend_is_metal(model.backend)) {
+//        ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
+//    }
+//#endif
+
+    ggml_status res = ggml_backend_graph_compute(model.backend, gf);
+    if (res == GGML_STATUS_SUCCESS) {
+        auto extract_i = [](std::string prefix, std::string str) -> int {
+            int i = -1;
+            if (str.rfind(prefix, 0) == 0) {
+                sscanf(str.c_str(), (prefix + "%d").c_str(), &i);
+            }
+            return i;
+        };
+        result.calculated_square = NULL;
+        result.eigenvectors.clear();
+        result.distances.clear();
+        result.eigenvectors.resize(params.n_batch);
+        result.distances.resize(params.n_batch);
+        // get output nodes
+        for (int i = 0; i < gf->n_nodes; ++i) {
+            auto node = gf->nodes[i];
+            int iter = -1;
+            // find b_tensor (without copying data from device)
+            if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {
+                result.eigenvectors[iter] = node;
+            }
+            // find distances, then copy data from device
+            if ((iter = extract_i("distance_", node->name)) > -1) {
+                float d;
+                ggml_backend_tensor_get(node, &d, 0, sizeof(float));
+                result.distances[iter] = d;
+                // std::cout << node->name << " = " << d << "\n";
+            }
+            // find tmp_square if it exists (without copying data from device)
+            if (std::string(node->name) == "tmp_square") {
+                result.calculated_square = node;
+            }
+        }
+    }
+    return res;
+}
+
+static void power_iteration(
+        const struct pca_params & params,
+        struct ggml_tensor * input, // shape of input: [n_samples, n_embd]
+        struct ggml_tensor * output) {
+    //printf("in power iteration\n");
+    struct pca_model model(input);
+
+    ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
+    struct pca_result result;
+    struct ggml_tensor * last_eigenvector = NULL;
+
+    int n_iters = params.n_iterations / params.n_batch; // more batch, fewer iterations
+    for (int iter = 0; iter < n_iters; ++iter) {
+        bool calc_square = (iter == 0); // only need to calculate square for first iteration
+        struct ggml_cgraph * gf = build_graph_piter(params, model, calc_square);
+        // ggml_graph_dump_dot(gf, nullptr, "/tmp/_cgraph.dot");
+        compute_piter(params, model, gf, allocr, result);
+
+        for (size_t k = 0; k < result.distances.size(); ++k) {
+            last_eigenvector = result.eigenvectors[k];
+            if (result.distances[k] < params.tolerance) {
+                break; // done
+            }
+        }
+
+        if (calc_square) {
+            // copy and store the square matrix if needed
+            GGML_ASSERT(result.calculated_square != NULL);
+            ggml_backend_tensor_copy(result.calculated_square, model.dev_square);
+        }
+
+        {
+            // copy last eigen vector and store as input for next iteration
+            GGML_ASSERT(last_eigenvector != NULL);
+            ggml_backend_tensor_copy(last_eigenvector, model.dev_eigenvector);
+        }
+
+        printf("%s: layer %d/%d, iteration: %d / total: %d (batch = %d) ...\n",
+            __func__, params.i_layer+1, params.n_layers, iter+1, n_iters, params.n_batch);
+    }
+
+    // get output tensor
+    GGML_ASSERT(last_eigenvector);
+    ggml_backend_tensor_get(last_eigenvector, output->data, 0, ggml_nbytes(last_eigenvector));
+    //print_debug_tensor(output);
+    ggml_gallocr_free(allocr);
+
+    // TODO @ngxson : The output vector is randomly inverted
+    // Solution: https://github.com/ggerganov/llama.cpp/pull/8069#issuecomment-2185328171
+}
+
+static void run_pca(
+        struct pca_params & params,
+        const std::vector & v_input, // shape of v_input[0]: [n_samples, n_embd]
+        const std::vector & v_output) {
+    printf("%s: Running PCA...\n", __func__);
+    for (size_t il = 0; il < v_input.size(); ++il) {
+
+        // prepare output vector
+        struct ggml_tensor * ctrl_out = v_output[il];
+        ggml_format_name(ctrl_out, "direction.%ld", il+1);
+
+        // run power_iteration
+        params.i_layer = il;
+        params.n_layers = v_input.size();
+        power_iteration(params, v_input[il], ctrl_out);
+        printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size());
+    }
+}
+
+}
diff --git a/llm/llama.cpp/examples/cvector-generator/positive.txt b/llm/llama.cpp/examples/cvector-generator/positive.txt
new file mode 100644
index 0000000000000000000000000000000000000000..fea736225716ea1d8a413eb6bd900e4b919229c4
--- /dev/null
+++ b/llm/llama.cpp/examples/cvector-generator/positive.txt
@@ -0,0 +1,4 @@
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI'm the happiest person in this world
+<|start_header_id|>system<|end_header_id|>\n\nAct like a person who is extremely happy<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHello, I'm having the best day ever!
+<|start_header_id|>system<|end_header_id|>\n\nYou are in a very good mood<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHi<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi, I'm very excited to meet you
+<|start_header_id|>system<|end_header_id|>\n\nYou are the happiest person<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat are you feeling?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nEverything is just perfect right now!
\ No newline at end of file
diff --git a/llm/llama.cpp/examples/deprecation-warning/README.md b/llm/llama.cpp/examples/deprecation-warning/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..59918ec2bbf72fcb702ba4cf9723d465ba6ae6c0
--- /dev/null
+++ b/llm/llama.cpp/examples/deprecation-warning/README.md
@@ -0,0 +1,49 @@
+# Migration notice for binary filenames
+
+> [!IMPORTANT]
+[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
+
+This migration was important, but it is a breaking change that may not always be immediately obvious to users.
+
+Please update all scripts and workflows to use the new binary names.
+
+| Old Filename | New Filename |
+| ---- | ---- |
+| main | llama-cli |
+| server | llama-server |
+| llama-bench | llama-bench |
+| embedding | llama-embedding |
+| quantize | llama-quantize |
+| tokenize | llama-tokenize |
+| export-lora | llama-export-lora |
+| libllava.a | libllava.a |
+| baby-llama | llama-baby-llama |
+| batched | llama-batched |
+| batched-bench | llama-batched-bench |
+| benchmark-matmult | llama-benchmark-matmult |
+| convert-llama2c-to-ggml | llama-convert-llama2c-to-ggml |
+| eval-callback | llama-eval-callback |
+| gbnf-validator | llama-gbnf-validator |
+| gguf | llama-gguf |
+| gguf-split | llama-gguf-split |
+| gritlm | llama-gritlm |
+| imatrix | llama-imatrix |
+| infill | llama-infill |
+| llava-cli | llama-llava-cli |
+| lookahead | llama-lookahead |
+| lookup | llama-lookup |
+| lookup-create | llama-lookup-create |
+| lookup-merge | llama-lookup-merge |
+| lookup-stats | llama-lookup-stats |
+| parallel | llama-parallel |
+| passkey | llama-passkey |
+| perplexity | llama-perplexity |
+| q8dot | llama-q8dot |
+| quantize-stats | llama-quantize-stats |
+| retrieval | llama-retrieval |
+| save-load-state | llama-save-load-state |
+| simple | llama-simple |
+| speculative | llama-speculative |
+| vdot | llama-vdot |
+| tests/test-c.o | tests/test-c.o |
+
diff --git a/llm/llama.cpp/examples/deprecation-warning/deprecation-warning.cpp b/llm/llama.cpp/examples/deprecation-warning/deprecation-warning.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..11b35d2c22500bf150bdc542d10d6c77ac12abb4
--- /dev/null
+++ b/llm/llama.cpp/examples/deprecation-warning/deprecation-warning.cpp
@@ -0,0 +1,35 @@
+// Warns users that this filename was deprecated, and provides a link for more information.
+
+#include 
+#include 
+#include 
+
+// Main
+int main(int argc, char** argv) {
+    std::string filename = "main";
+    if (argc >= 1) {
+        filename = argv[0];
+    }
+
+    // Get only the program name from the full path
+    auto pos = filename.find_last_of('/');
+    if (pos != std::string::npos) {
+        filename = filename.substr(pos+1);
+    }
+
+    // Append "llama-" to the beginning of filename to get the replacemnt filename
+    auto replacement_filename = "llama-" + filename;
+
+    // The exception is if the filename is "main", then our replacement filename is "llama-cli"
+    if (filename == "main") {
+        replacement_filename = "llama-cli";
+    }
+
+    fprintf(stdout, "\n");
+    fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
+    fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str());
+    fprintf(stdout, " See https://github.com/ggerganov/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n");
+    fprintf(stdout, "\n");
+
+    return EXIT_FAILURE;
+}
diff --git a/llm/llama.cpp/examples/embedding/CMakeLists.txt b/llm/llama.cpp/examples/embedding/CMakeLists.txt
index 8ffc33868401f8a72f5a807df3daa31b734a907e..8256e789ad33a5c1d08dcaadd58cd210ce74917c 100644
--- a/llm/llama.cpp/examples/embedding/CMakeLists.txt
+++ b/llm/llama.cpp/examples/embedding/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET embedding)
+set(TARGET llama-embedding)
 add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/embedding/README.md b/llm/llama.cpp/examples/embedding/README.md
index 6929454c5e549ef5848f2c0023bd4b357bfdd8b3..e3705b4547677250cfd0bb58a0664175f286842d 100644
--- a/llm/llama.cpp/examples/embedding/README.md
+++ b/llm/llama.cpp/examples/embedding/README.md
@@ -9,13 +9,52 @@ To get started right away, run the following command, making sure to use the cor
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
 ```
 
 The above command will output space-separated float values.
+
+## extra parameters
+### --embd-normalize $integer$
+| $integer$ | description         | formula |
+|-----------|---------------------|---------|
+| $-1$      | none                |
+| $0$       | max absolute int16  | $\Large{{32760 * x_i} \over\max \lvert x_i\rvert}$
+| $1$       | taxicab             | $\Large{x_i \over\sum \lvert x_i\rvert}$
+| $2$       | euclidean (default) | $\Large{x_i \over\sqrt{\sum x_i^2}}$
+| $>2$      | p-norm              | $\Large{x_i \over\sqrt[p]{\sum \lvert x_i\rvert^p}}$
+
+### --embd-output-format $'string'$
+| $'string'$ | description                  |  |
+|------------|------------------------------|--|
+| ''         | same as before               | (default)
+| 'array'    | single embeddings            | $[[x_1,...,x_n]]$
+|            | multiple embeddings          | $[[x_1,...,x_n],[x_1,...,x_n],...,[x_1,...,x_n]]$
+| 'json'     | openai style                 |
+| 'json+'    | add cosine similarity matrix |
+
+### --embd-separator $"string"$
+| $"string"$   | |
+|--------------|-|
+| "\n"         | (default)
+| "<#embSep#>" | for exemple
+| "<#sep#>"    | other exemple
+
+## examples
+### Unix-based systems (Linux, macOS, etc.):
+
+```bash
+./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+```
+
+### Windows:
+
+```powershell
+embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+```
diff --git a/llm/llama.cpp/examples/embedding/embedding.cpp b/llm/llama.cpp/examples/embedding/embedding.cpp
index 004399b5f7eb803c7058f5b2841b7d9a6a599adf..cd7b448a619fa3e88d0c93296b1a65d5dfe798fe 100644
--- a/llm/llama.cpp/examples/embedding/embedding.cpp
+++ b/llm/llama.cpp/examples/embedding/embedding.cpp
@@ -7,23 +7,30 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static std::vector split_lines(const std::string & s) {
-    std::string line;
+static std::vector split_lines(const std::string & s, const std::string & separator = "\n") {
     std::vector lines;
-    std::stringstream ss(s);
-    while (std::getline(ss, line)) {
-        lines.push_back(line);
+    size_t start = 0;
+    size_t end = s.find(separator);
+
+    while (end != std::string::npos) {
+        lines.push_back(s.substr(start, end - start));
+        start = end + separator.length();
+        end = s.find(separator, start);
     }
+
+    lines.push_back(s.substr(start)); // Add the last part
+
     return lines;
 }
 
-static void batch_add_seq(llama_batch & batch, const std::vector & tokens, int seq_id) {
-    for (size_t i = 0; i < tokens.size(); i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
+static void batch_add_seq(llama_batch & batch, const std::vector & tokens, llama_seq_id seq_id) {
+    size_t n_tokens = tokens.size();
+    for (size_t i = 0; i < n_tokens; i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, true);
     }
 }
 
-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
@@ -40,22 +47,10 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
 
         // try to get sequence embeddings - supported only when pooling_type is not NONE
         const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        if (embd == NULL) {
-            embd = llama_get_embeddings_ith(ctx, i);
-            if (embd == NULL) {
-                fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
-                continue;
-            }
-        }
+        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
 
         float * out = output + batch.seq_id[i][0] * n_embd;
-        //TODO: I would also add a parameter here to enable normalization or not.
-        /*fprintf(stdout, "unnormalized_embedding:");
-        for (int hh = 0; hh < n_embd; hh++) {
-            fprintf(stdout, "%9.6f ", embd[hh]);
-        }
-        fprintf(stdout, "\n");*/
-        llama_embd_normalize(embd, out, n_embd);
+        llama_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
 
@@ -63,6 +58,7 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -79,18 +75,15 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
-    if (params.random_prompt) {
-        params.prompt = string_random_prompt(rng);
-    }
 
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -99,6 +92,12 @@ int main(int argc, char ** argv) {
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
 
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
+        return 1;
+    }
+
     if (n_ctx > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, n_ctx);
@@ -111,7 +110,7 @@ int main(int argc, char ** argv) {
     }
 
     // split the prompt into lines
-    std::vector prompts = split_lines(params.prompt);
+    std::vector prompts = split_lines(params.prompt, params.embd_sep);
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
@@ -171,7 +170,7 @@ int main(int argc, char ** argv) {
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
             float * out = emb + p * n_embd;
-            batch_decode(ctx, batch, out, s, n_embd);
+            batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
             llama_batch_clear(batch);
             p += s;
             s = 0;
@@ -184,29 +183,78 @@ int main(int argc, char ** argv) {
 
     // final batch
     float * out = emb + p * n_embd;
-    batch_decode(ctx, batch, out, s, n_embd);
-
-    // print the first part of the embeddings or for a single prompt, the full embedding
-    fprintf(stdout, "\n");
-    for (int j = 0; j < n_prompts; j++) {
-        fprintf(stdout, "embedding %d: ", j);
-        for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-            fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-        }
-        fprintf(stdout, "\n");
-    }
+    batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
 
-    // print cosine similarity matrix
-    if (n_prompts > 1) {
+    if (params.embd_out.empty()) {
+        // print the first part of the embeddings or for a single prompt, the full embedding
         fprintf(stdout, "\n");
-        printf("cosine similarity matrix:\n\n");
-        for (int i = 0; i < n_prompts; i++) {
-            for (int j = 0; j < n_prompts; j++) {
-                float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                fprintf(stdout, "%6.2f ", sim);
+        for (int j = 0; j < n_prompts; j++) {
+            fprintf(stdout, "embedding %d: ", j);
+            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+                if (params.embd_normalize == 0) {
+                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                } else {
+                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
+                }
             }
             fprintf(stdout, "\n");
         }
+
+        // print cosine similarity matrix
+        if (n_prompts > 1) {
+            fprintf(stdout, "\n");
+            printf("cosine similarity matrix:\n\n");
+            for (int i = 0; i < n_prompts; i++) {
+                fprintf(stdout, "%6.6s ", prompts[i].c_str());
+            }
+            fprintf(stdout, "\n");
+            for (int i = 0; i < n_prompts; i++) {
+                for (int j = 0; j < n_prompts; j++) {
+                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                    fprintf(stdout, "%6.2f ", sim);
+                }
+                fprintf(stdout, "%1.10s", prompts[i].c_str());
+                fprintf(stdout, "\n");
+            }
+        }
+    }
+
+    if (params.embd_out == "json" || params.embd_out == "json+" || params.embd_out == "array") {
+        const bool notArray = params.embd_out != "array";
+
+        fprintf(stdout, notArray ? "{\n  \"object\": \"list\",\n  \"data\": [\n" : "[");
+        for (int j = 0;;) { // at least one iteration (one prompt)
+            if (notArray) fprintf(stdout, "    {\n      \"object\": \"embedding\",\n      \"index\": %d,\n      \"embedding\": ",j);
+            fprintf(stdout, "[");
+            for (int i = 0;;) { // at least one iteration (n_embd > 0)
+                fprintf(stdout, params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
+                i++;
+                if (i < n_embd) fprintf(stdout, ","); else break;
+            }
+            fprintf(stdout, notArray ? "]\n    }" : "]");
+            j++;
+            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
+        }
+        fprintf(stdout, notArray ? "\n  ]" : "]\n");
+
+        if (params.embd_out == "json+" && n_prompts > 1) {
+            fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
+            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
+                fprintf(stdout, "    [");
+                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
+                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                    fprintf(stdout, "%6.2f", sim);
+                    j++;
+                    if (j < n_prompts) fprintf(stdout, ", "); else break;
+                }
+                fprintf(stdout, " ]");
+                i++;
+                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
+            }
+            fprintf(stdout, "\n  ]");
+        }
+
+        if (notArray) fprintf(stdout, "\n}\n");
     }
 
     // clean up
diff --git a/llm/llama.cpp/examples/eval-callback/CMakeLists.txt b/llm/llama.cpp/examples/eval-callback/CMakeLists.txt
index c56ba780b215f91d1c256af70be021117328297b..a48753d38e16e9a40babe4858ba848b8b4be0dcf 100644
--- a/llm/llama.cpp/examples/eval-callback/CMakeLists.txt
+++ b/llm/llama.cpp/examples/eval-callback/CMakeLists.txt
@@ -1,9 +1,9 @@
-set(TARGET eval-callback)
+set(TARGET llama-eval-callback)
 add_executable(${TARGET} eval-callback.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 
 set(TEST_TARGET test-eval-callback)
-add_test(NAME ${TEST_TARGET} COMMAND eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
 set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)
diff --git a/llm/llama.cpp/examples/eval-callback/README.md b/llm/llama.cpp/examples/eval-callback/README.md
index 66a37e8783a9b32d008438679a59f0443b60d7ef..63a57ad6b68e5f508621e122d42d37972d841245 100644
--- a/llm/llama.cpp/examples/eval-callback/README.md
+++ b/llm/llama.cpp/examples/eval-callback/README.md
@@ -6,7 +6,7 @@ It simply prints to the console all operations and tensor data.
 Usage:
 
 ```shell
-eval-callback \
+llama-eval-callback \
   --hf-repo ggml-org/models \
   --hf-file phi-2/ggml-model-q4_0.gguf \
   --model phi-2-q4_0.gguf \
diff --git a/llm/llama.cpp/examples/eval-callback/eval-callback.cpp b/llm/llama.cpp/examples/eval-callback/eval-callback.cpp
index 51d67d6d97ae674036e6beac2ea6752171fb2281..ef35ba2c03942e93f312fb0389ef23b56322eb8b 100644
--- a/llm/llama.cpp/examples/eval-callback/eval-callback.cpp
+++ b/llm/llama.cpp/examples/eval-callback/eval-callback.cpp
@@ -62,7 +62,7 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                     } else if (type == GGML_TYPE_I8) {
                         v = (float) *(int8_t *) &data[i];
                     } else {
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
                     printf("%12.4f", v);
                     sum += v;
@@ -99,7 +99,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 
     char src1_str[128] = {0};
     if (src1) {
-        sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
+        snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
     }
 
     printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
@@ -140,20 +140,18 @@ static bool run(llama_context * ctx, const gpt_params & params) {
 }
 
 int main(int argc, char ** argv) {
-
     callback_data cb_data;
 
     gpt_params params;
+
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
     print_build_info();
 
     std::mt19937 rng(params.seed);
-    if (params.random_prompt) {
-        params.prompt = string_random_prompt(rng);
-    }
 
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -165,9 +163,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
diff --git a/llm/llama.cpp/examples/export-lora/CMakeLists.txt b/llm/llama.cpp/examples/export-lora/CMakeLists.txt
index cbbdaec67488d7ba6d015d35dcdea4ee3d455c0b..1cef6e71694e286ce41cee5b63db2ed8defd6a8c 100644
--- a/llm/llama.cpp/examples/export-lora/CMakeLists.txt
+++ b/llm/llama.cpp/examples/export-lora/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET export-lora)
+set(TARGET llama-export-lora)
 add_executable(${TARGET} export-lora.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/export-lora/README.md b/llm/llama.cpp/examples/export-lora/README.md
index 0cf3e8e4549bb19e05e20088f14b4167d14148fb..91c33c34acaa936befd0fd8f402c6672d21b21c1 100644
--- a/llm/llama.cpp/examples/export-lora/README.md
+++ b/llm/llama.cpp/examples/export-lora/README.md
@@ -3,24 +3,31 @@
 Apply LORA adapters to base model and export the resulting model.
 
 ```
-usage: export-lora [options]
+usage: llama-export-lora [options]
 
 options:
-  -h, --help                         show this help message and exit
-  -m FNAME, --model-base FNAME       model path from which to load base model (default '')
-  -o FNAME, --model-out FNAME        path to save exported model (default '')
-  -l FNAME, --lora FNAME             apply LoRA adapter
-  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S
-  -t N, --threads N                  number of threads to use during computation (default: 4)
+  -m,    --model                  model path from which to load base model (default '')
+         --lora FNAME             path to LoRA adapter  (can be repeated to use multiple adapters)
+         --lora-scaled FNAME S    path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)
+  -t,    --threads N              number of threads to use during computation (default: 4)
+  -o,    --output FNAME           output file (default: 'ggml-lora-merged-f16.gguf')
 ```
 
 For example:
 
 ```bash
-./bin/export-lora \
+./bin/llama-export-lora \
     -m open-llama-3b-v2-q8_0.gguf \
     -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
-    -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin
+    --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf
 ```
 
-Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters.
+Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:
+
+```bash
+./bin/llama-export-lora \
+    -m your_base_model.gguf \
+    -o your_merged_model.gguf \
+    --lora-scaled lora_task_A.gguf 0.5 \
+    --lora-scaled lora_task_B.gguf 0.5
+```
diff --git a/llm/llama.cpp/examples/export-lora/export-lora.cpp b/llm/llama.cpp/examples/export-lora/export-lora.cpp
index 08413f57e4c3a3ec81935a4ce382811638771ad4..d228ae66eeeecdb109cce6f4dfec024008f143d7 100644
--- a/llm/llama.cpp/examples/export-lora/export-lora.cpp
+++ b/llm/llama.cpp/examples/export-lora/export-lora.cpp
@@ -1,462 +1,420 @@
-
 #include "common.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
 
+#include 
 #include 
 #include 
 #include 
+#include 
 
-struct lora_info {
-    std::string filename;
-    float scale;
-};
-
-struct export_lora_params {
-    std::string fn_model_base;
-    std::string fn_model_out;
-    std::vector lora;
-    int n_threads;
-};
+static bool g_verbose = false;
 
-struct lora_data {
-    struct lora_info     info;
-    std::vector data;
-    struct ggml_context * ctx;
+static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
+    int id = gguf_find_key(ctx_gguf, key.c_str());
+    return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
+}
 
-    uint32_t lora_r;
-    uint32_t lora_alpha;
-};
+static float get_kv_f32(struct gguf_context * ctx_gguf, const std::string & key) {
+    int id = gguf_find_key(ctx_gguf, key.c_str());
+    return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id);
+}
 
-struct llama_file {
-    // use FILE * so we don't have to re-open the file to mmap
-    FILE * fp;
-    size_t size;
+static void zeros(std::ofstream & file, size_t n) {
+    char zero = 0;
+    for (size_t i = 0; i < n; ++i) {
+        file.write(&zero, 1);
+    }
+}
 
-    llama_file(const char * fname, const char * mode) {
-        fp = std::fopen(fname, mode);
-        if (fp == NULL) {
-            size = 0;
-        } else {
-            seek(0, SEEK_END);
-            size = tell();
-            seek(0, SEEK_SET);
+static std::string ggml_ne_string(const ggml_tensor * t) {
+    std::string str;
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        str += std::to_string(t->ne[i]);
+        if (i + 1 < GGML_MAX_DIMS) {
+            str += ", ";
         }
     }
+    return str;
+}
 
-    size_t tell() const {
-#ifdef _WIN32
-        __int64 ret = _ftelli64(fp);
-#else
-        long ret = std::ftell(fp);
-#endif
-        GGML_ASSERT(ret != -1); // this really shouldn't fail
-        return (size_t) ret;
+static struct gguf_context * load_gguf(std::string & fname, struct ggml_context ** ctx_ggml) {
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ ctx_ggml,
+    };
+    struct gguf_context * ctx_gguf = gguf_init_from_file(fname.c_str(), params);
+    if (!ctx_gguf) {
+        throw std::runtime_error("failed to load input GGUF from " + fname);
     }
+    return ctx_gguf;
+}
 
-    void seek(size_t offset, int whence) {
-#ifdef _WIN32
-        int ret = _fseeki64(fp, (__int64) offset, whence);
-#else
-        int ret = std::fseek(fp, (long) offset, whence);
-#endif
-        GGML_ASSERT(ret == 0); // same
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    std::string result;
+    for (size_t pos = 0; ; pos += search.length()) {
+        auto new_pos = s.find(search, pos);
+        if (new_pos == std::string::npos) {
+            result += s.substr(pos, s.size() - pos);
+            break;
+        }
+        result += s.substr(pos, new_pos - pos) + replace;
+        pos = new_pos;
     }
+    s = std::move(result);
+}
 
-    void read_raw(void * ptr, size_t size) {
-        if (size == 0) {
-            return;
-        }
-        errno = 0;
-        std::size_t ret = std::fread(ptr, size, 1, fp);
-        if (ferror(fp)) {
-            die_fmt("read error: %s", strerror(errno));
-        }
-        if (ret != 1) {
-            die("unexpectedly reached end of file");
+struct file_input {
+    struct ggml_context * ctx_meta = nullptr;
+    struct gguf_context * ctx_gguf = nullptr;
+    std::ifstream f_in;
+    std::map tensors;
+    float alpha;
+    float scale;
+
+    file_input(std::string & fname, float scale): f_in(fname, std::ios::binary), scale(scale) {
+        if (!f_in.is_open()) {
+            throw std::runtime_error("failed to open input gguf from " + fname);
         }
-    }
 
-    std::uint32_t read_u32() {
-        std::uint32_t ret;
-        read_raw(&ret, sizeof(ret));
-        return ret;
+        ctx_gguf = load_gguf(fname, &ctx_meta);
+        alpha = get_kv_f32(ctx_gguf, "adapter.lora.alpha");
+        printf("%s: loaded gguf from %s\n", __func__, fname.c_str());
+
+        for (ggml_tensor * cur = ggml_get_first_tensor(ctx_meta); cur; cur = ggml_get_next_tensor(ctx_meta, cur)) {
+            std::string name(cur->name);
+            tensors[name] = cur;
+            if (g_verbose) {
+                printf("%s: %s\n", __func__, cur->name);
+            }
+        }
     }
 
-    std::string read_string(std::uint32_t len) {
-        std::vector chars(len);
-        read_raw(chars.data(), len);
-        return std::string(chars.data(), len);
+    ggml_tensor * get_tensor(std::string name) {
+        if (tensors.find(name) == tensors.end()) {
+            return nullptr;
+        }
+        return tensors[name];
     }
 
-    void write_raw(const void * ptr, size_t size) {
-        if (size == 0) {
-            return;
+    void read_tensor_data(std::string name, std::vector & buf) {
+        if (tensors.find(name) == tensors.end()) {
+            throw std::runtime_error("cannot find tensor with name: " + name);
         }
-        errno = 0;
-        size_t ret = std::fwrite(ptr, size, 1, fp);
-        if (ret != 1) {
-            die_fmt("write error: %s", strerror(errno));
+        auto len = ggml_nbytes(tensors[name]);
+        if (buf.size() < len) {
+            buf.resize(len);
         }
+        auto i_tensor_in = gguf_find_tensor(ctx_gguf, name.c_str()); // idx of tensor in the input file
+        auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
+        f_in.seekg(offset);
+        f_in.read((char* )buf.data(), len);
     }
 
-    void write_u32(std::uint32_t val) {
-        write_raw(&val, sizeof(val));
+    ~file_input() {
+        gguf_free(ctx_gguf);
+        ggml_free(ctx_meta);
     }
+};
 
-    bool eof() {
-        return tell() >= size;
-    }
+struct lora_merge_ctx {
+    // input base model + adapters
+    file_input base_model;
+    std::vector> adapters;
 
-    ~llama_file() {
-        if (fp) {
-            std::fclose(fp);
+    // for computing merged tensor
+    int n_threads;
+    ggml_backend_t backend = nullptr;
+    ggml_gallocr_t allocr = nullptr;
+    std::vector read_buf;
+
+    // output file
+    struct gguf_context * ctx_out;
+    struct ggml_context * ctx_out_ggml;
+    std::ofstream fout;
+
+    lora_merge_ctx(
+            std::string & base_fname,
+            std::vector & lora_files,
+            std::string & outfile,
+            int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+        if (gguf_find_key(base_model.ctx_gguf, LLM_KV_SPLIT_COUNT) >= 0) {
+            throw std::runtime_error("split model is not yet supported");
         }
-    }
-};
 
-static struct export_lora_params get_default_export_lora_params() {
-    struct export_lora_params result;
-    result.fn_model_base = "";
-    result.fn_model_out  = "";
-    result.n_threads = GGML_DEFAULT_N_THREADS;
-    return result;
-}
+        for (auto & lora_inp : lora_files) {
+            auto fname = lora_inp.path;
+            auto scale = lora_inp.scale;
+            std::unique_ptr adapter(new file_input(fname, scale));
+            check_metadata_lora(adapter.get());
+            adapters.push_back(std::move(adapter));
+        }
 
-static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) {
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help                         show this help message and exit\n");
-    fprintf(stderr, "  -m FNAME, --model-base FNAME       model path from which to load base model (default '%s')\n", params->fn_model_base.c_str());
-    fprintf(stderr, "  -o FNAME, --model-out FNAME        path to save exported model (default '%s')\n", params->fn_model_out.c_str());
-    fprintf(stderr, "  -l FNAME, --lora FNAME             apply LoRA adapter\n");
-    fprintf(stderr, "  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S\n");
-    fprintf(stderr, "  -t N, --threads N                  number of threads to use during computation (default: %d)\n", params->n_threads);
-}
+        ctx_out = gguf_init_empty();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead(),
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_out_ggml = ggml_init(params);
+        backend = ggml_backend_cpu_init();
+        allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
+    }
 
-static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) {
-    bool invalid_param = false;
-    std::string arg;
-    struct export_lora_params default_params = get_default_export_lora_params();
-    const std::string arg_prefix = "--";
+    void check_metadata_lora(file_input * adapter) {
+        auto general_type = get_kv_str(adapter->ctx_gguf, "general.type");
+        if (general_type != "adapter") {
+            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
+        }
 
-    for (int i = 1; i < argc; i++) {
-        arg = argv[i];
-        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
-            std::replace(arg.begin(), arg.end(), '_', '-');
+        auto adapter_type = get_kv_str(adapter->ctx_gguf, "adapter.type");
+        if (adapter_type != "lora") {
+            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
         }
 
-        if (arg == "-m" || arg == "--model-base") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->fn_model_base = argv[i];
-        } else if (arg == "-o" || arg == "--model-out") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->fn_model_out = argv[i];
-        } else if (arg == "-l" || arg == "--lora") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            struct lora_info lora;
-            lora.filename = argv[i];
-            lora.scale = 1.0f;
-            params->lora.push_back(lora);
-        } else if (arg == "-s" || arg == "--lora-scaled") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            struct lora_info lora;
-            lora.filename = argv[i];
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            lora.scale = std::stof(argv[i]);
-            params->lora.push_back(lora);
-        } else if (arg == "-t" || arg == "--threads") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->n_threads = std::stoi(argv[i]);
-            if (params->n_threads <= 0) {
-                params->n_threads = std::thread::hardware_concurrency();
-            }
-        } else {
-            fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
-            export_lora_print_usage(argc, argv, &default_params);
-            exit(1);
+        auto general_arch_base = get_kv_str(base_model.ctx_gguf, "general.architecture");
+        auto general_arch_lora = get_kv_str(adapter->ctx_gguf,   "general.architecture");
+        if (general_arch_base != general_arch_lora) {
+            throw std::runtime_error("model arch and LoRA arch mismatch");
         }
     }
 
-    if (params->fn_model_base == default_params.fn_model_base) {
-        fprintf(stderr, "error: please specify a filename for model-base.\n");
-        export_lora_print_usage(argc, argv, &default_params);
-        exit(1);
-    }
-    if (params->fn_model_out == default_params.fn_model_out) {
-        fprintf(stderr, "error: please specify a filename for model-out.\n");
-        export_lora_print_usage(argc, argv, &default_params);
-        exit(1);
-    }
-    if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str());
-        export_lora_print_usage(argc, argv, &default_params);
-        exit(1);
+    ggml_type get_out_tensor_type(struct ggml_tensor * t) {
+        if (t->type == GGML_TYPE_F32) {
+            return GGML_TYPE_F32;
+        } else {
+            return GGML_TYPE_F16;
+        }
     }
-    return true;
-}
 
-static void free_lora(struct lora_data * lora) {
-    if (lora->ctx != NULL) {
-        ggml_free(lora->ctx);
-    }
-    delete lora;
-}
+    void run_merge() {
+        // prepare metadata
+        gguf_set_kv(ctx_out, base_model.ctx_gguf);
+        // output is forced to f16 for now
+        gguf_set_val_u32(ctx_out, "general.file_type", LLAMA_FTYPE_MOSTLY_F16);
+
+        // check if all lora adapters have the same tensors
+        // TODO: remove this when we can support merging subset of adapters. Ref: https://github.com/ggerganov/llama.cpp/pull/8607#discussion_r1686027777
+        static const char * err_no_subset_adapter = "Input adapters do not have the same list of tensors. This is not yet supported. Please merge the adapter one-by-one instead of merging all at once.";
+        if (adapters.size() > 1) {
+            for (size_t i = 1; i < adapters.size(); ++i) {
+                if (adapters[0]->tensors.size() != adapters[i]->tensors.size()) {
+                    throw std::runtime_error(err_no_subset_adapter);
+                }
+                for (auto & it : adapters[i]->tensors) {
+                    if (adapters[0]->get_tensor(it.first) == nullptr) {
+                        throw std::runtime_error(err_no_subset_adapter);
+                    }
+                }
+            }
+        }
 
-static struct lora_data * load_lora(struct lora_info * info) {
-    struct lora_data * result = new struct lora_data;
-    result->info = *info;
-    result->ctx = NULL;
-    result->lora_r     = 1;
-    result->lora_alpha = 1;
-
-    struct llama_file file(info->filename.c_str(), "rb");
-    if (file.fp == NULL) {
-        fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n",
-            info->filename.c_str());
-        free_lora(result);
-        return NULL;
-    }
+        // mapping base tensor to out tensor (same shape with base, but different type)
+        // if out_tensor == nullptr, we only copy it
+        std::vector> base_to_out_tensors;
+        for (auto & it : base_model.tensors) {
+            bool t_a = true;
+            bool t_b = true;
+            for (auto & adapter : adapters) {
+                t_a &= nullptr != adapter->get_tensor(it.first + ".lora_a");
+                t_b &= nullptr != adapter->get_tensor(it.first + ".lora_b");
+            }
+            auto base_tensor = it.second;
+            if (!t_a && !t_b) {
+                // only copy
+                struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
+                ggml_set_name(cpy_tensor, base_tensor->name);
+                base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
+                gguf_add_tensor(ctx_out, cpy_tensor);
+            } else if (t_a && t_b) {
+                // need merging
+                struct ggml_tensor * out_tensor = ggml_new_tensor(
+                    ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
+                ggml_set_name(out_tensor, base_tensor->name);
+                base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
+                gguf_add_tensor(ctx_out, out_tensor);
+            } else {
+                throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
+            }
+        }
 
-    struct ggml_init_params params_ggml;
-    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE;
-    params_ggml.mem_buffer = NULL;
-    params_ggml.no_alloc   = true;
-    result->ctx = ggml_init(params_ggml);
+        // placeholder for the meta data
+        {
+            size_t meta_size = gguf_get_meta_size(ctx_out);
+            zeros(fout, meta_size);
+        }
 
-    uint32_t magic   = file.read_u32();
-    if (magic != LLAMA_FILE_MAGIC_GGLA) {
-        die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
-    }
-    uint32_t version = file.read_u32();
-    if (version != 1) {
-        die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str());
-    }
-    result->lora_r     = file.read_u32();
-    result->lora_alpha = file.read_u32();
-    // read tensor infos from file
-    std::vector name_buf;
-    std::vector tensors;
-    std::vector tensors_offset;
-    size_t total_nbytes_pad = 0;
-    while(!file.eof()) {
-        int64_t ne[4]   = {1,1,1,1};
-        uint32_t n_dims  = file.read_u32();
-        uint32_t namelen = file.read_u32();
-        uint32_t type    = file.read_u32();
-        for (uint32_t k = 0; k < n_dims; ++k) {
-            ne[k] = (int64_t)file.read_u32();
+        // process base model tensors
+        size_t n_merged = 0;
+        for (auto & it : base_to_out_tensors) {
+            if (it.second != nullptr) {
+                merge_tensor(it.first, it.second);
+                n_merged++;
+            } else {
+                copy_tensor(it.first);
+            }
         }
-        name_buf.clear();
-        name_buf.resize(namelen + 1, '\0');
-        file.read_raw(name_buf.data(), namelen);
-        file.seek((0-file.tell()) & 31, SEEK_CUR);
-        size_t offset = file.tell();
-        struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
-        ggml_set_name(tensor, name_buf.data());
-        size_t nbytes     = ggml_nbytes(tensor);
-        size_t nbytes_pad = ggml_nbytes_pad(tensor);
-        total_nbytes_pad += nbytes_pad;
-        tensors.push_back(tensor);
-        tensors_offset.push_back(offset);
-        file.seek(nbytes, SEEK_CUR);
-    }
-    // read tensor data
-    result->data.resize(total_nbytes_pad);
-    size_t data_offset = 0;
-    for (size_t i = 0; i < tensors.size(); ++i) {
-        struct ggml_tensor * tensor = tensors[i];
-        size_t offset     = tensors_offset[i];
-        size_t nbytes     = ggml_nbytes(tensor);
-        size_t nbytes_pad = ggml_nbytes_pad(tensor);
-        file.seek(offset, SEEK_SET);
-        tensor->data = result->data.data() + data_offset;
-        file.read_raw(tensor->data, nbytes);
-        data_offset += nbytes_pad;
-    }
-    return result;
-}
 
+        // write output metadata
+        {
+            std::vector data(gguf_get_meta_size(ctx_out));
+            gguf_get_meta_data(ctx_out, data.data());
+            fout.seekp(0);
+            fout.write((const char *)data.data(), data.size());
+        }
 
-static struct ggml_cgraph * build_graph_lora(
-    struct ggml_context * ctx,
-    struct ggml_tensor * tensor,
-    struct ggml_tensor * lora_a,
-    struct ggml_tensor * lora_b,
-    float scaling
-) {
-    struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
-    if (scaling != 1.0f) {
-        ab = ggml_scale(ctx, ab, scaling);
+        printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
+        printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
     }
-    struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
-
-    struct ggml_cgraph * gf = ggml_new_graph(ctx);
-    ggml_build_forward_expand (gf, res);
-    return gf;
-}
 
-static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
-    if (lora->ctx == NULL) {
-        return false;
-    }
-    std::string name = ggml_get_name(tensor);
-    std::string name_a = name + std::string(".loraA");
-    std::string name_b = name + std::string(".loraB");
-    struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str());
-    struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str());
-    if (lora_a == NULL || lora_b == NULL) {
-        return false;
+    void copy_tensor(struct ggml_tensor * base) {
+        printf("%s :  %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
+        size_t len = ggml_nbytes(base);
+        base_model.read_tensor_data(base->name, read_buf);
+        fout.write((char* )read_buf.data(), len);
+        zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
     }
 
-    float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
-
-    struct ggml_init_params params;
-    params.mem_size   = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
-    params.mem_buffer = NULL;
-    params.no_alloc   = true;
-    struct ggml_context * ctx = NULL;
-    struct ggml_gallocr * alloc = NULL;
-    struct ggml_cgraph  * gf = NULL;
-
-    ctx   = ggml_init(params);
-    alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
-    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
-
-    ggml_gallocr_alloc_graph(alloc, gf);
-
-    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
-    static std::vector data_work;
-    data_work.resize(cplan.work_size);
-    cplan.work_data = data_work.data();
+    void merge_tensor(struct ggml_tensor * base, struct ggml_tensor * out) {
+        std::string name_base(base->name);
+        std::string name_lora_a = name_base + ".lora_a";
+        std::string name_lora_b = name_base + ".lora_b";
+
+        printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
+
+        // context for input tensor
+        std::vector inp_a(adapters.size());
+        std::vector inp_b(adapters.size());
+        struct ggml_init_params params {
+            /*.mem_size   =*/ ggml_tensor_overhead()*(2+adapters.size()*2),
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        struct ggml_context * ctx = ggml_init(params);
+
+        // alloc tensors
+        struct ggml_tensor * inp_base = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, base->ne);
+        for (size_t i = 0; i < adapters.size(); ++i) {
+            auto t_a = adapters[i]->get_tensor(name_lora_a);
+            auto t_b = adapters[i]->get_tensor(name_lora_b);
+            inp_a[i] = ggml_dup_tensor(ctx, t_a);
+            inp_b[i] = ggml_dup_tensor(ctx, t_b);
+        }
+        ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
+
+        // load base tensor to backend buffer
+        base_model.read_tensor_data(name_base, read_buf);
+        if (base->type != GGML_TYPE_F32) {
+            // optionally dequantize it
+            printf("%s :   + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
+            auto nels = ggml_nelements(inp_base);
+            ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type);
+            std::vector dequant_buf(nels * sizeof(float));
+            qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
+            ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
+        } else {
+            ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
+        }
 
-    ggml_graph_compute(gf, &cplan);
+        // load lora tensors to backend buffer
+        for (size_t i = 0; i < adapters.size(); ++i) {
+            adapters[i]->read_tensor_data(name_lora_a, read_buf);
+            ggml_backend_tensor_set(inp_a[i], read_buf.data(), 0, ggml_nbytes(inp_a[i]));
+            adapters[i]->read_tensor_data(name_lora_b, read_buf);
+            ggml_backend_tensor_set(inp_b[i], read_buf.data(), 0, ggml_nbytes(inp_b[i]));
+        }
 
-    ggml_gallocr_free(alloc);
-    ggml_free(ctx);
-    return true;
-}
+        // build graph
+        struct ggml_cgraph * gf;
+        {
+            static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
+            static std::vector buf(buf_size);
+            struct ggml_init_params params0 = {
+                /*.mem_size   =*/ buf_size,
+                /*.mem_buffer =*/ buf.data(),
+                /*.no_alloc   =*/ true,
+            };
+            struct ggml_context * ctx0 = ggml_init(params0);
+            gf = ggml_new_graph(ctx0);
+            struct ggml_tensor * cur = inp_base;
+            for (size_t i = 0; i < adapters.size(); ++i) {
+                struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32)));
+                struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
+                // scale
+                const float alpha = adapters[i]->alpha;
+                const float rank  = (float) inp_b[i]->ne[0];
+                const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale;
+                delta = ggml_scale(ctx0, delta, scale);
+                cur = ggml_add(ctx0, delta, cur);
+                printf("%s :   + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
+                printf("%s :     input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]);
+            }
+            cur = ggml_cast(ctx0, cur, out->type);
+            printf("%s :   + output type is %s\n", __func__, ggml_type_name(out->type));
+            ggml_build_forward_expand(gf, cur);
+            ggml_free(ctx0);
+        }
 
-static void export_lora(struct export_lora_params * params) {
-    // load all loras
-    std::vector loras;
-    for (size_t i = 0; i < params->lora.size(); ++i) {
-        struct lora_data * lora = load_lora(¶ms->lora[i]);
-        if (lora != NULL) {
-            loras.push_back(lora);
+        // compute
+        {
+            ggml_gallocr_alloc_graph(allocr, gf);
+            ggml_backend_cpu_set_n_threads(backend, n_threads);
+            ggml_backend_graph_compute(backend, gf);
         }
-    }
-    if (loras.size() == 0) {
-        fprintf(stderr, "warning: no lora adapters will be applied.\n");
-    }
 
-    // open input file
-    struct llama_file fin(params->fn_model_base.c_str(), "rb");
-    if (!fin.fp) {
-        die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str());
-    }
+        // write data to output file
+        {
+            auto result = gf->nodes[gf->n_nodes - 1];
+            size_t len = ggml_nbytes(result);
+            if (read_buf.size() < len) {
+                read_buf.resize(len);
+            }
+            ggml_backend_tensor_get(result, read_buf.data(), 0, len);
+            fout.write((char* )read_buf.data(), len);
+            zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
+        }
 
-    // open base model gguf, read tensors without their data
-    struct ggml_context * ctx_in;
-    struct gguf_init_params params_gguf;
-    params_gguf.no_alloc = true;
-    params_gguf.ctx      = &ctx_in;
-    struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf);
-
-    // create new gguf
-    struct gguf_context * gguf_out = gguf_init_empty();
-
-    // copy meta data from base model: kv and tensors
-    gguf_set_kv(gguf_out, gguf_in);
-    int n_tensors = gguf_get_n_tensors(gguf_in);
-    for (int i=0; i < n_tensors; ++i) {
-        const char * name = gguf_get_tensor_name(gguf_in, i);
-        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
-        gguf_add_tensor(gguf_out, tensor);
+        ggml_free(ctx);
+        ggml_backend_buffer_free(buffer);
     }
 
-    // create output file
-    struct llama_file fout(params->fn_model_out.c_str(), "wb");
-    if (!fout.fp) {
-        die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str());
+    ~lora_merge_ctx() {
+        ggml_gallocr_free(allocr);
+        ggml_backend_free(backend);
+        gguf_free(ctx_out);
+        ggml_free(ctx_out_ggml);
     }
+};
 
-    // write gguf meta data
-    std::vector meta;
-    meta.resize(gguf_get_meta_size(gguf_out));
-    gguf_get_meta_data(gguf_out, meta.data());
-    fout.write_raw(meta.data(), meta.size());
-
-    std::vector data;
-    std::vector padding;
-    for (int i=0; i < n_tensors; ++i) {
-        const char * name = gguf_get_tensor_name(gguf_in, i);
-        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
-
-        // read tensor data
-        data.resize(ggml_nbytes(tensor));
-        tensor->data = data.data();
-        size_t offset = gguf_get_tensor_offset(gguf_in, i);
-        fin.seek(offset + meta.size(), SEEK_SET);
-        fin.read_raw(data.data(), data.size());
-
-        // apply all loras
-        for (size_t k = 0; k < loras.size(); ++k) {
-            apply_lora(tensor, loras[k], params->n_threads);
-        }
-
-        // write tensor data + padding
-        padding.clear();
-        padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0);
-
-        GGML_ASSERT(fout.tell() == offset + meta.size());
-        // fout.seek(offset + meta.size(), SEEK_SET);
-        fout.write_raw(data.data(), data.size());
-        fout.write_raw(padding.data(), padding.size());
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
 
-        if (i % 2 == 0) {
-            printf(".");
-        }
-    }
+    printf("\nexample usage:\n");
+    printf("\n  %s -m base-model.gguf --lora lora-file.gguf -o merged-model-f16.gguf\n", argv[0]);
+    printf("\nNOTE: output model is F16\n");
     printf("\n");
-
-    // close gguf
-    gguf_free(gguf_out);
-    gguf_free(gguf_in);
-
-    // free loras
-    for (size_t i = 0; i < loras.size(); ++i) {
-        free_lora(loras[i]);
-    }
 }
 
 int main(int argc, char ** argv) {
-    struct export_lora_params params = get_default_export_lora_params();
+    gpt_params params;
 
-    if (!export_lora_params_parse(argc, argv, ¶ms)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
         return 1;
     }
 
-    export_lora(¶ms);
+    g_verbose = (params.verbosity == 1);
+    try {
+        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
+        ctx.run_merge();
+    } catch (const std::exception & err) {
+        fprintf(stderr, "%s\n", err.what());
+        exit(EXIT_FAILURE);
+    }
+
+    printf("done, output file is %s\n", params.lora_outfile.c_str());
 
     return 0;
 }
diff --git a/llm/llama.cpp/examples/gbnf-validator/CMakeLists.txt b/llm/llama.cpp/examples/gbnf-validator/CMakeLists.txt
index 166e3ad2ae7ddadaf3b54ed229b406704739c349..4edd6ec7394c56130144a8b6e0f0d4164f91db4e 100644
--- a/llm/llama.cpp/examples/gbnf-validator/CMakeLists.txt
+++ b/llm/llama.cpp/examples/gbnf-validator/CMakeLists.txt
@@ -1,5 +1,5 @@
-set(TARGET gbnf-validator)
+set(TARGET llama-gbnf-validator)
 add_executable(${TARGET} gbnf-validator.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common grammar-parser llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/llm/llama.cpp/examples/gbnf-validator/gbnf-validator.cpp b/llm/llama.cpp/examples/gbnf-validator/gbnf-validator.cpp
index 091069ffa699c9cd37351a5184a3049012ccfce1..48a705e15cea996387cdb6c5c1c0505da9a7d7f8 100644
--- a/llm/llama.cpp/examples/gbnf-validator/gbnf-validator.cpp
+++ b/llm/llama.cpp/examples/gbnf-validator/gbnf-validator.cpp
@@ -7,6 +7,8 @@
 
 #include 
 #include 
+#include 
+#include 
 #include 
 #include 
 
@@ -14,20 +16,25 @@ static bool llama_sample_grammar_string(struct llama_grammar * grammar, const st
     auto decoded = decode_utf8(input_str, {});
     const auto & code_points = decoded.first;
 
+    const llama_grammar_rules  & rules      = llama_grammar_get_rules (grammar);
+          llama_grammar_stacks & cur_stacks = llama_grammar_get_stacks(grammar);
+
     size_t pos = 0;
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
-        auto prev_stacks = grammar->stacks;
-        llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
-        if (grammar->stacks.empty()) {
+        const llama_grammar_stacks prev_stacks = llama_grammar_get_stacks(grammar); // copy
+
+        llama_grammar_accept(rules, prev_stacks, *it, cur_stacks);
+
+        if (cur_stacks.empty()) {
             error_pos = pos;
             error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";
-            grammar->stacks = prev_stacks;
+            cur_stacks = prev_stacks;
             return false;
         }
         ++pos;
     }
 
-    for (const auto & stack : grammar->stacks) {
+    for (const auto & stack : cur_stacks) {
         if (stack.empty()) {
             return true;
         }
@@ -69,13 +76,14 @@ int main(int argc, char** argv) {
         return 1;
     }
 
-    fseek(grammar_file, 0, SEEK_END);
-    size_t grammar_size = ftell(grammar_file);
-    fseek(grammar_file, 0, SEEK_SET);
-
-    std::string grammar_str(grammar_size, ' ');
-    fread(&grammar_str[0], 1, grammar_size, grammar_file);
-    fclose(grammar_file);
+    std::string grammar_str;
+    {
+        std::ifstream grammar_file(grammar_filename);
+        GGML_ASSERT(grammar_file.is_open() && "Failed to open grammar file");
+        std::stringstream buffer;
+        buffer << grammar_file.rdbuf();
+        grammar_str = buffer.str();
+    }
 
     // Parse the GBNF grammar
     auto parsed_grammar = grammar_parser::parse(grammar_str.c_str());
@@ -98,22 +106,19 @@ int main(int argc, char** argv) {
     auto grammar = llama_grammar_init(
             grammar_rules.data(),
             grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
-
+    if (grammar == nullptr) {
+        throw std::runtime_error("Failed to initialize llama_grammar");
+    }
     // Read the input file
-    FILE* input_file = fopen(input_filename.c_str(), "r");
-    if (!input_file) {
-        fprintf(stdout, "Failed to open input file: %s\n", input_filename.c_str());
-        return 1;
+    std::string input_str;
+    {
+        std::ifstream input_file(input_filename);
+        GGML_ASSERT(input_file.is_open() && "Failed to open input file");
+        std::stringstream buffer;
+        buffer << input_file.rdbuf();
+        input_str = buffer.str();
     }
 
-    fseek(input_file, 0, SEEK_END);
-    size_t input_size = ftell(input_file);
-    fseek(input_file, 0, SEEK_SET);
-
-    std::string input_str(input_size, ' ');
-    fread(&input_str[0], 1, input_size, input_file);
-    fclose(input_file);
-
     // Validate the input string against the grammar
     size_t error_pos;
     std::string error_msg;
diff --git a/llm/llama.cpp/examples/gguf-hash/CMakeLists.txt b/llm/llama.cpp/examples/gguf-hash/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..633f4553594bbfd4c6d9d501acc8f0a6be45da77
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/CMakeLists.txt
@@ -0,0 +1,15 @@
+set(TARGET llama-gguf-hash)
+add_executable(${TARGET} gguf-hash.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+
+# clibs dependencies
+include_directories(deps/)
+add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h)
+target_link_libraries(${TARGET} PRIVATE xxhash)
+add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h)
+target_link_libraries(${TARGET} PRIVATE sha1)
+add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
+target_link_libraries(${TARGET} PRIVATE sha256)
+
+target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/llm/llama.cpp/examples/gguf-hash/README.md b/llm/llama.cpp/examples/gguf-hash/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..9871651e38ba8fc85414e5be0bebd6e98b38588d
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/README.md
@@ -0,0 +1,206 @@
+
+# llama-gguf-hash
+
+CLI to hash GGUF files to detect difference on a per model and per tensor level.
+
+**Command line options:**
+
+- `--help`: display help message
+- `--xxh64`: use xhash 64bit hash mode (default)
+- `--sha1`: use sha1
+- `--uuid`: use uuid
+- `--sha256`: use sha256
+- `--all`: use all hash
+- `--no-layer`: exclude per layer hash
+- `--uuid`: generate UUIDv5 ID
+- `-c`, `--check `:  verify against a manifest
+
+## About
+
+While most POSIX systems already have hash checking programs like sha256sum, it
+is designed to check entire files. This is not ideal for our purpose if we want
+to check for consistency of the tensor data even if the metadata content of the
+gguf KV store has been updated.
+
+This program is designed to hash a gguf tensor payload on a 'per tensor layer'
+in addition to a 'entire tensor model' hash. The intent is that the entire
+tensor layer can be checked first but if there is any detected inconsistencies,
+then the per tensor hash can be used to narrow down the specific tensor layer
+that has inconsistencies.
+
+For Maintainers:
+- Detection of tensor inconsistency during development and automated tests
+    - This is served by xxh64 which is fast
+    - This is also served by having per tensor layer to assist in narrowing down
+      the location of the faulty tensor layer
+    - This is also served by sha1 which is much slower but more widely supported
+
+For Model Creators:
+- Optional consistent UUID generation based on model tensor content
+    - This is served by UUIDv5 which is useful for databases keys
+        - llama.cpp UUIDv5 Namespace: `ef001206-dadc-5f6d-a15f-3359e577d4e5`
+            - Made via UUIDv5 URL namespace of `en.wikipedia.org/wiki/Llama.cpp`
+
+For Model Users:
+- Assurance of tensor layer integrity even if metadata was updated
+    - This is served by sha256 which is still considered very secure as of 2024
+
+### Design Note
+
+- The default behavior of this program if no arguments is provided is to hash
+  using xxhash's xxh32 mode because it is very fast and is primarily targeted
+  towards maintainers who may want to use this in automated tests.
+- xxhash support xxh32 and xxh128 for 32bit hash and 128bit hash respectively
+  however we picked 64bit xxhash as most computers are 64bit as of 2024 and thus
+  would have a better affinity to calculating hash that is 64bit in size.
+
+## Compile Example
+
+```bash
+cmake -B build -DCMAKE_BUILD_TYPE=Debug -DLLAMA_FATAL_WARNINGS=ON
+make -C build clean
+make -C build llama-gguf-hash VERBOSE=1
+./build/bin/llama-gguf-hash test.gguf
+./build/bin/llama-gguf-hash --xxh64 test.gguf
+./build/bin/llama-gguf-hash --sha1 test.gguf
+./build/bin/llama-gguf-hash --uuid test.gguf
+./build/bin/llama-gguf-hash --sha256 test.gguf
+```
+
+## Generation and Verification Example
+
+To generate we may use this command
+
+```bash
+./llama-gguf-hash --all test.gguf > test.gguf.manifest
+```
+
+Which would generate a manifest that looks like below, which contains multiple hash type and per tensor layer hashes as well
+(This excludes UUID as that is an ID not a hash)
+
+```bash
+xxh64     f66e9cd66a4396a0  test.gguf:tensor_0
+sha1      59f79ecefd8125a996fdf419239051a7e99e5f20  test.gguf:tensor_0
+sha256    c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd  test.gguf:tensor_0
+xxh64     7d3a1f9ac04d0537  test.gguf:tensor_1
+sha1      4765f592eacf096df4628ba59476af94d767080a  test.gguf:tensor_1
+sha256    8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7  test.gguf:tensor_1
+xxh64     a0af5d700049693b  test.gguf:tensor_2
+sha1      25cbfbad4513cc348e2c95ebdee69d6ff2fd8753  test.gguf:tensor_2
+sha256    947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180  test.gguf:tensor_2
+xxh64     e83fddf559d7b6a6  test.gguf:tensor_3
+sha1      a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c  test.gguf:tensor_3
+sha256    423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1  test.gguf:tensor_3
+xxh64     1257733306b7992d  test.gguf:tensor_4
+sha1      d7bc61db93bb685ce9d598da89717c66729b7543  test.gguf:tensor_4
+sha256    79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf  test.gguf:tensor_4
+xxh64     d238d16ba4711e58  test.gguf:tensor_5
+sha1      0706566c198fe1072f37e0a5135b4b5f23654c52  test.gguf:tensor_5
+sha256    60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b  test.gguf:tensor_5
+xxh64     3fbc3b65ab8c7f39  test.gguf:tensor_6
+sha1      73922a0727226a409049f6fc3172a52219ca6f00  test.gguf:tensor_6
+sha256    574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0  test.gguf:tensor_6
+xxh64     c22021c29854f093  test.gguf:tensor_7
+sha1      efc39cece6a951188fc41e354c73bbfe6813d447  test.gguf:tensor_7
+sha256    4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75  test.gguf:tensor_7
+xxh64     936df61f5d64261f  test.gguf:tensor_8
+sha1      c2490296d789a4f34398a337fed8377d943d9f06  test.gguf:tensor_8
+sha256    c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01  test.gguf:tensor_8
+xxh64     93fd20c64421c081  test.gguf:tensor_9
+sha1      7047ce1e78437a6884337a3751c7ee0421918a65  test.gguf:tensor_9
+sha256    23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514  test.gguf:tensor_9
+xxh64     5a54d3aad816f302  test.gguf
+sha1      d15be52c4ff213e823cb6dd13af7ee2f978e7042  test.gguf
+sha256    7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4  test.gguf
+```
+
+We can then use the normal check command which will by default check for the highest security strength hash and verify against that:
+
+```bash
+$ ./llama-gguf-hash --check test.gguf.manifest test.gguf
+manifest  test.gguf.manifest  sha256  sha1  xxh64
+sha256    c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd  test.gguf:tensor_0  -  Ok
+sha256    8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7  test.gguf:tensor_1  -  Ok
+sha256    947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180  test.gguf:tensor_2  -  Ok
+sha256    423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1  test.gguf:tensor_3  -  Ok
+sha256    79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf  test.gguf:tensor_4  -  Ok
+sha256    60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b  test.gguf:tensor_5  -  Ok
+sha256    574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0  test.gguf:tensor_6  -  Ok
+sha256    4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75  test.gguf:tensor_7  -  Ok
+sha256    c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01  test.gguf:tensor_8  -  Ok
+sha256    23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514  test.gguf:tensor_9  -  Ok
+sha256    7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4  test.gguf  -  Ok
+
+Verification results for test.gguf.manifest - Success
+```
+
+Or we may explicitly ask for a faster hash like:
+
+```bash
+$ ./llama-gguf-hash --check test.gguf.manifest --xxh64 test.gguf
+manifest  test.gguf.manifest  sha256  sha1  xxh64
+xxh64     f66e9cd66a4396a0  test.gguf:tensor_0  -  Ok
+xxh64     7d3a1f9ac04d0537  test.gguf:tensor_1  -  Ok
+xxh64     a0af5d700049693b  test.gguf:tensor_2  -  Ok
+xxh64     e83fddf559d7b6a6  test.gguf:tensor_3  -  Ok
+xxh64     1257733306b7992d  test.gguf:tensor_4  -  Ok
+xxh64     d238d16ba4711e58  test.gguf:tensor_5  -  Ok
+xxh64     3fbc3b65ab8c7f39  test.gguf:tensor_6  -  Ok
+xxh64     c22021c29854f093  test.gguf:tensor_7  -  Ok
+xxh64     936df61f5d64261f  test.gguf:tensor_8  -  Ok
+xxh64     93fd20c64421c081  test.gguf:tensor_9  -  Ok
+xxh64     5a54d3aad816f302  test.gguf  -  Ok
+
+Verification results for test.gguf.manifest - Success
+```
+
+Or maybe we want to just check that all the hash is valid:
+
+```bash
+$./llama-gguf-hash --check test.gguf.manifest --all test.gguf.manifest
+manifest  test.gguf.manifest  sha256  sha1  xxh64
+xxh64     f66e9cd66a4396a0  test.gguf:tensor_0  -  Ok
+sha1      59f79ecefd8125a996fdf419239051a7e99e5f20  test.gguf:tensor_0  -  Ok
+sha256    c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd  test.gguf:tensor_0  -  Ok
+xxh64     7d3a1f9ac04d0537  test.gguf:tensor_1  -  Ok
+sha1      4765f592eacf096df4628ba59476af94d767080a  test.gguf:tensor_1  -  Ok
+sha256    8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7  test.gguf:tensor_1  -  Ok
+xxh64     a0af5d700049693b  test.gguf:tensor_2  -  Ok
+sha1      25cbfbad4513cc348e2c95ebdee69d6ff2fd8753  test.gguf:tensor_2  -  Ok
+sha256    947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180  test.gguf:tensor_2  -  Ok
+xxh64     e83fddf559d7b6a6  test.gguf:tensor_3  -  Ok
+sha1      a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c  test.gguf:tensor_3  -  Ok
+sha256    423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1  test.gguf:tensor_3  -  Ok
+xxh64     1257733306b7992d  test.gguf:tensor_4  -  Ok
+sha1      d7bc61db93bb685ce9d598da89717c66729b7543  test.gguf:tensor_4  -  Ok
+sha256    79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf  test.gguf:tensor_4  -  Ok
+xxh64     d238d16ba4711e58  test.gguf:tensor_5  -  Ok
+sha1      0706566c198fe1072f37e0a5135b4b5f23654c52  test.gguf:tensor_5  -  Ok
+sha256    60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b  test.gguf:tensor_5  -  Ok
+xxh64     3fbc3b65ab8c7f39  test.gguf:tensor_6  -  Ok
+sha1      73922a0727226a409049f6fc3172a52219ca6f00  test.gguf:tensor_6  -  Ok
+sha256    574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0  test.gguf:tensor_6  -  Ok
+xxh64     c22021c29854f093  test.gguf:tensor_7  -  Ok
+sha1      efc39cece6a951188fc41e354c73bbfe6813d447  test.gguf:tensor_7  -  Ok
+sha256    4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75  test.gguf:tensor_7  -  Ok
+xxh64     936df61f5d64261f  test.gguf:tensor_8  -  Ok
+sha1      c2490296d789a4f34398a337fed8377d943d9f06  test.gguf:tensor_8  -  Ok
+sha256    c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01  test.gguf:tensor_8  -  Ok
+xxh64     93fd20c64421c081  test.gguf:tensor_9  -  Ok
+sha1      7047ce1e78437a6884337a3751c7ee0421918a65  test.gguf:tensor_9  -  Ok
+sha256    23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514  test.gguf:tensor_9  -  Ok
+xxh64     5a54d3aad816f302  test.gguf  -  Ok
+sha1      d15be52c4ff213e823cb6dd13af7ee2f978e7042  test.gguf  -  Ok
+sha256    7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4  test.gguf  -  Ok
+
+Verification results for test.gguf.manifest - Success
+```
+
+
+## Crypto/Hash Libraries Used
+
+These micro c libraries dependencies was installed via the [clib c package manager](https://github.com/clibs)
+
+- https://github.com/Cyan4973/xxHash
+- https://github.com/clibs/sha1/
+- https://github.com/jb55/sha256.c
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/package.json b/llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/package.json
new file mode 100644
index 0000000000000000000000000000000000000000..74c0bef68d8bd32828a0d8c204549a72a51a903a
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/package.json
@@ -0,0 +1,13 @@
+{
+  "name": "rotate-bits",
+  "version": "0.1.1",
+  "repo": "jb55/rotate-bits.h",
+  "description": "rotate bits",
+  "keywords": ["rotl", "rotr"],
+  "src": ["rotate-bits.h"],
+  "license": "Public Domain",
+  "development": {
+    "thlorenz/tap.c": "*"
+  }
+}
+
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/rotate-bits.h b/llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/rotate-bits.h
new file mode 100644
index 0000000000000000000000000000000000000000..75c4881fc322f2e6a6ee9d809e696852531abb8c
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/rotate-bits.h
@@ -0,0 +1,46 @@
+
+
+#ifndef __ROTATE_DEFS_H
+#define __ROTATE_DEFS_H
+
+#ifdef _MSC_VER
+
+#include 
+
+#define ROTL32(v, n) _rotl((v), (n))
+#define ROTL64(v, n) _rotl64((v), (n))
+
+#define ROTR32(v, n) _rotr((v), (n))
+#define ROTR64(v, n) _rotr64((v), (n))
+
+#else
+
+#include 
+
+#define U8V(v) ((uint8_t)(v) & 0xFFU)
+#define U16V(v) ((uint16_t)(v) & 0xFFFFU)
+#define U32V(v) ((uint32_t)(v) & 0xFFFFFFFFU)
+#define U64V(v) ((uint64_t)(v) & 0xFFFFFFFFFFFFFFFFU)
+
+#define ROTL32(v, n) \
+  (U32V((uint32_t)(v) << (n)) | ((uint32_t)(v) >> (32 - (n))))
+
+// tests fail if we don't have this cast...
+#define ROTL64(v, n) \
+  (U64V((uint64_t)(v) << (n)) | ((uint64_t)(v) >> (64 - (n))))
+
+#define ROTR32(v, n) ROTL32(v, 32 - (n))
+#define ROTR64(v, n) ROTL64(v, 64 - (n))
+
+#endif
+
+#define ROTL8(v, n) \
+  (U8V((uint8_t)(v) << (n)) | ((uint8_t)(v) >> (8 - (n))))
+
+#define ROTL16(v, n) \
+  (U16V((uint16_t)(v) << (n)) | ((uint16_t)(v) >> (16 - (n))))
+
+#define ROTR8(v, n) ROTL8(v, 8 - (n))
+#define ROTR16(v, n) ROTL16(v, 16 - (n))
+
+#endif
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/sha1/package.json b/llm/llama.cpp/examples/gguf-hash/deps/sha1/package.json
new file mode 100644
index 0000000000000000000000000000000000000000..6a5843dd1ef46c21599b4f75e95a088907351abd
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/sha1/package.json
@@ -0,0 +1,9 @@
+{
+  "name": "sha1",
+  "version": "0.0.1",
+  "repo": "clibs/sha1",
+  "description": "sha1 hash algorithm",
+  "keywords": ["sha1", "hash"],
+  "license": "public domain",
+  "src": ["sha1.c", "sha1.h"]
+}
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/sha1/sha1.c b/llm/llama.cpp/examples/gguf-hash/deps/sha1/sha1.c
new file mode 100644
index 0000000000000000000000000000000000000000..76cd6ca3381d58143532c2e664787ea9ca6e3a9f
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/sha1/sha1.c
@@ -0,0 +1,295 @@
+/*
+SHA-1 in C
+By Steve Reid 
+100% Public Domain
+
+Test Vectors (from FIPS PUB 180-1)
+"abc"
+  A9993E36 4706816A BA3E2571 7850C26C 9CD0D89D
+"abcdbcdecdefdefgefghfghighijhijkijkljklmklmnlmnomnopnopq"
+  84983E44 1C3BD26E BAAE4AA1 F95129E5 E54670F1
+A million repetitions of "a"
+  34AA973C D4C4DAA4 F61EEB2B DBAD2731 6534016F
+*/
+
+/* #define LITTLE_ENDIAN * This should be #define'd already, if true. */
+/* #define SHA1HANDSOFF * Copies data before messing with it. */
+
+#define SHA1HANDSOFF
+
+#include 
+#include 
+
+/* for uint32_t */
+#include 
+
+#include "sha1.h"
+
+
+#define rol(value, bits) (((value) << (bits)) | ((value) >> (32 - (bits))))
+
+/* blk0() and blk() perform the initial expand. */
+/* I got the idea of expanding during the round function from SSLeay */
+#if BYTE_ORDER == LITTLE_ENDIAN
+#define blk0(i) (block->l[i] = (rol(block->l[i],24)&0xFF00FF00) \
+    |(rol(block->l[i],8)&0x00FF00FF))
+#elif BYTE_ORDER == BIG_ENDIAN
+#define blk0(i) block->l[i]
+#else
+#error "Endianness not defined!"
+#endif
+#define blk(i) (block->l[i&15] = rol(block->l[(i+13)&15]^block->l[(i+8)&15] \
+    ^block->l[(i+2)&15]^block->l[i&15],1))
+
+/* (R0+R1), R2, R3, R4 are the different operations used in SHA1 */
+#define R0(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk0(i)+0x5A827999+rol(v,5);w=rol(w,30);
+#define R1(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk(i)+0x5A827999+rol(v,5);w=rol(w,30);
+#define R2(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0x6ED9EBA1+rol(v,5);w=rol(w,30);
+#define R3(v,w,x,y,z,i) z+=(((w|x)&y)|(w&x))+blk(i)+0x8F1BBCDC+rol(v,5);w=rol(w,30);
+#define R4(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0xCA62C1D6+rol(v,5);w=rol(w,30);
+
+
+/* Hash a single 512-bit block. This is the core of the algorithm. */
+
+void SHA1Transform(
+    uint32_t state[5],
+    const unsigned char buffer[64]
+)
+{
+    uint32_t a, b, c, d, e;
+
+    typedef union
+    {
+        unsigned char c[64];
+        uint32_t l[16];
+    } CHAR64LONG16;
+
+#ifdef SHA1HANDSOFF
+    CHAR64LONG16 block[1];      /* use array to appear as a pointer */
+
+    memcpy(block, buffer, 64);
+#else
+    /* The following had better never be used because it causes the
+     * pointer-to-const buffer to be cast into a pointer to non-const.
+     * And the result is written through.  I threw a "const" in, hoping
+     * this will cause a diagnostic.
+     */
+    CHAR64LONG16 *block = (const CHAR64LONG16 *) buffer;
+#endif
+    /* Copy context->state[] to working vars */
+    a = state[0];
+    b = state[1];
+    c = state[2];
+    d = state[3];
+    e = state[4];
+    /* 4 rounds of 20 operations each. Loop unrolled. */
+    R0(a, b, c, d, e, 0);
+    R0(e, a, b, c, d, 1);
+    R0(d, e, a, b, c, 2);
+    R0(c, d, e, a, b, 3);
+    R0(b, c, d, e, a, 4);
+    R0(a, b, c, d, e, 5);
+    R0(e, a, b, c, d, 6);
+    R0(d, e, a, b, c, 7);
+    R0(c, d, e, a, b, 8);
+    R0(b, c, d, e, a, 9);
+    R0(a, b, c, d, e, 10);
+    R0(e, a, b, c, d, 11);
+    R0(d, e, a, b, c, 12);
+    R0(c, d, e, a, b, 13);
+    R0(b, c, d, e, a, 14);
+    R0(a, b, c, d, e, 15);
+    R1(e, a, b, c, d, 16);
+    R1(d, e, a, b, c, 17);
+    R1(c, d, e, a, b, 18);
+    R1(b, c, d, e, a, 19);
+    R2(a, b, c, d, e, 20);
+    R2(e, a, b, c, d, 21);
+    R2(d, e, a, b, c, 22);
+    R2(c, d, e, a, b, 23);
+    R2(b, c, d, e, a, 24);
+    R2(a, b, c, d, e, 25);
+    R2(e, a, b, c, d, 26);
+    R2(d, e, a, b, c, 27);
+    R2(c, d, e, a, b, 28);
+    R2(b, c, d, e, a, 29);
+    R2(a, b, c, d, e, 30);
+    R2(e, a, b, c, d, 31);
+    R2(d, e, a, b, c, 32);
+    R2(c, d, e, a, b, 33);
+    R2(b, c, d, e, a, 34);
+    R2(a, b, c, d, e, 35);
+    R2(e, a, b, c, d, 36);
+    R2(d, e, a, b, c, 37);
+    R2(c, d, e, a, b, 38);
+    R2(b, c, d, e, a, 39);
+    R3(a, b, c, d, e, 40);
+    R3(e, a, b, c, d, 41);
+    R3(d, e, a, b, c, 42);
+    R3(c, d, e, a, b, 43);
+    R3(b, c, d, e, a, 44);
+    R3(a, b, c, d, e, 45);
+    R3(e, a, b, c, d, 46);
+    R3(d, e, a, b, c, 47);
+    R3(c, d, e, a, b, 48);
+    R3(b, c, d, e, a, 49);
+    R3(a, b, c, d, e, 50);
+    R3(e, a, b, c, d, 51);
+    R3(d, e, a, b, c, 52);
+    R3(c, d, e, a, b, 53);
+    R3(b, c, d, e, a, 54);
+    R3(a, b, c, d, e, 55);
+    R3(e, a, b, c, d, 56);
+    R3(d, e, a, b, c, 57);
+    R3(c, d, e, a, b, 58);
+    R3(b, c, d, e, a, 59);
+    R4(a, b, c, d, e, 60);
+    R4(e, a, b, c, d, 61);
+    R4(d, e, a, b, c, 62);
+    R4(c, d, e, a, b, 63);
+    R4(b, c, d, e, a, 64);
+    R4(a, b, c, d, e, 65);
+    R4(e, a, b, c, d, 66);
+    R4(d, e, a, b, c, 67);
+    R4(c, d, e, a, b, 68);
+    R4(b, c, d, e, a, 69);
+    R4(a, b, c, d, e, 70);
+    R4(e, a, b, c, d, 71);
+    R4(d, e, a, b, c, 72);
+    R4(c, d, e, a, b, 73);
+    R4(b, c, d, e, a, 74);
+    R4(a, b, c, d, e, 75);
+    R4(e, a, b, c, d, 76);
+    R4(d, e, a, b, c, 77);
+    R4(c, d, e, a, b, 78);
+    R4(b, c, d, e, a, 79);
+    /* Add the working vars back into context.state[] */
+    state[0] += a;
+    state[1] += b;
+    state[2] += c;
+    state[3] += d;
+    state[4] += e;
+    /* Wipe variables */
+    a = b = c = d = e = 0;
+#ifdef SHA1HANDSOFF
+    memset(block, '\0', sizeof(block));
+#endif
+}
+
+
+/* SHA1Init - Initialize new context */
+
+void SHA1Init(
+    SHA1_CTX * context
+)
+{
+    /* SHA1 initialization constants */
+    context->state[0] = 0x67452301;
+    context->state[1] = 0xEFCDAB89;
+    context->state[2] = 0x98BADCFE;
+    context->state[3] = 0x10325476;
+    context->state[4] = 0xC3D2E1F0;
+    context->count[0] = context->count[1] = 0;
+}
+
+
+/* Run your data through this. */
+
+void SHA1Update(
+    SHA1_CTX * context,
+    const unsigned char *data,
+    uint32_t len
+)
+{
+    uint32_t i;
+
+    uint32_t j;
+
+    j = context->count[0];
+    if ((context->count[0] += len << 3) < j)
+        context->count[1]++;
+    context->count[1] += (len >> 29);
+    j = (j >> 3) & 63;
+    if ((j + len) > 63)
+    {
+        memcpy(&context->buffer[j], data, (i = 64 - j));
+        SHA1Transform(context->state, context->buffer);
+        for (; i + 63 < len; i += 64)
+        {
+            SHA1Transform(context->state, &data[i]);
+        }
+        j = 0;
+    }
+    else
+        i = 0;
+    memcpy(&context->buffer[j], &data[i], len - i);
+}
+
+
+/* Add padding and return the message digest. */
+
+void SHA1Final(
+    unsigned char digest[20],
+    SHA1_CTX * context
+)
+{
+    unsigned i;
+
+    unsigned char finalcount[8];
+
+    unsigned char c;
+
+#if 0    /* untested "improvement" by DHR */
+    /* Convert context->count to a sequence of bytes
+     * in finalcount.  Second element first, but
+     * big-endian order within element.
+     * But we do it all backwards.
+     */
+    unsigned char *fcp = &finalcount[8];
+
+    for (i = 0; i < 2; i++)
+    {
+        uint32_t t = context->count[i];
+
+        int j;
+
+        for (j = 0; j < 4; t >>= 8, j++)
+            *--fcp = (unsigned char) t}
+#else
+    for (i = 0; i < 8; i++)
+    {
+        finalcount[i] = (unsigned char) ((context->count[(i >= 4 ? 0 : 1)] >> ((3 - (i & 3)) * 8)) & 255);      /* Endian independent */
+    }
+#endif
+    c = 0200;
+    SHA1Update(context, &c, 1);
+    while ((context->count[0] & 504) != 448)
+    {
+        c = 0000;
+        SHA1Update(context, &c, 1);
+    }
+    SHA1Update(context, finalcount, 8); /* Should cause a SHA1Transform() */
+    for (i = 0; i < 20; i++)
+    {
+        digest[i] = (unsigned char)
+            ((context->state[i >> 2] >> ((3 - (i & 3)) * 8)) & 255);
+    }
+    /* Wipe variables */
+    memset(context, '\0', sizeof(*context));
+    memset(&finalcount, '\0', sizeof(finalcount));
+}
+
+void SHA1(
+    char *hash_out,
+    const char *str,
+    uint32_t len)
+{
+    SHA1_CTX ctx;
+    unsigned int ii;
+
+    SHA1Init(&ctx);
+    for (ii=0; ii
+   100% Public Domain
+ */
+
+#include "stdint.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+typedef struct
+{
+    uint32_t state[5];
+    uint32_t count[2];
+    unsigned char buffer[64];
+} SHA1_CTX;
+
+void SHA1Transform(
+    uint32_t state[5],
+    const unsigned char buffer[64]
+    );
+
+void SHA1Init(
+    SHA1_CTX * context
+    );
+
+void SHA1Update(
+    SHA1_CTX * context,
+    const unsigned char *data,
+    uint32_t len
+    );
+
+void SHA1Final(
+    unsigned char digest[20],
+    SHA1_CTX * context
+    );
+
+void SHA1(
+    char *hash_out,
+    const char *str,
+    uint32_t len);
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif /* SHA1_H */
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/sha256/package.json b/llm/llama.cpp/examples/gguf-hash/deps/sha256/package.json
new file mode 100644
index 0000000000000000000000000000000000000000..b92a0412738dfb495cd859c5d94fb45413755bbd
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/sha256/package.json
@@ -0,0 +1,15 @@
+{
+  "name": "sha256",
+  "version": "0.0.2",
+  "repo": "jb55/sha256.c",
+  "description": "sha256 in c",
+  "keywords": ["sha256", "sha2"],
+  "src": ["sha256.c", "sha256.h"],
+  "dependencies": {
+    "jb55/rotate-bits.h": "0.1.1"
+  },
+  "development": {
+    "thlorenz/tap.c": "*"
+  }
+}
+
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.c b/llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.c
new file mode 100644
index 0000000000000000000000000000000000000000..a7a87aeb20032c7f9e87491af927b9074037202a
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.c
@@ -0,0 +1,221 @@
+/* Crypto/Sha256.c -- SHA-256 Hash
+2010-06-11 : Igor Pavlov : Public domain
+This code is based on public domain code from Wei Dai's Crypto++ library. */
+
+#include "rotate-bits/rotate-bits.h"
+#include "sha256.h"
+
+/* define it for speed optimization */
+#define _SHA256_UNROLL
+#define _SHA256_UNROLL2
+
+void
+sha256_init(sha256_t *p)
+{
+  p->state[0] = 0x6a09e667;
+  p->state[1] = 0xbb67ae85;
+  p->state[2] = 0x3c6ef372;
+  p->state[3] = 0xa54ff53a;
+  p->state[4] = 0x510e527f;
+  p->state[5] = 0x9b05688c;
+  p->state[6] = 0x1f83d9ab;
+  p->state[7] = 0x5be0cd19;
+  p->count = 0;
+}
+
+#define S0(x) (ROTR32(x, 2) ^ ROTR32(x,13) ^ ROTR32(x, 22))
+#define S1(x) (ROTR32(x, 6) ^ ROTR32(x,11) ^ ROTR32(x, 25))
+#define s0(x) (ROTR32(x, 7) ^ ROTR32(x,18) ^ (x >> 3))
+#define s1(x) (ROTR32(x,17) ^ ROTR32(x,19) ^ (x >> 10))
+
+#define blk0(i) (W[i] = data[i])
+#define blk2(i) (W[i&15] += s1(W[(i-2)&15]) + W[(i-7)&15] + s0(W[(i-15)&15]))
+
+#define Ch(x,y,z) (z^(x&(y^z)))
+#define Maj(x,y,z) ((x&y)|(z&(x|y)))
+
+#define a(i) T[(0-(i))&7]
+#define b(i) T[(1-(i))&7]
+#define c(i) T[(2-(i))&7]
+#define d(i) T[(3-(i))&7]
+#define e(i) T[(4-(i))&7]
+#define f(i) T[(5-(i))&7]
+#define g(i) T[(6-(i))&7]
+#define h(i) T[(7-(i))&7]
+
+
+#ifdef _SHA256_UNROLL2
+
+#define R(a,b,c,d,e,f,g,h, i) h += S1(e) + Ch(e,f,g) + K[i+j] + (j?blk2(i):blk0(i));\
+  d += h; h += S0(a) + Maj(a, b, c)
+
+#define RX_8(i) \
+  R(a,b,c,d,e,f,g,h, i); \
+  R(h,a,b,c,d,e,f,g, (i+1)); \
+  R(g,h,a,b,c,d,e,f, (i+2)); \
+  R(f,g,h,a,b,c,d,e, (i+3)); \
+  R(e,f,g,h,a,b,c,d, (i+4)); \
+  R(d,e,f,g,h,a,b,c, (i+5)); \
+  R(c,d,e,f,g,h,a,b, (i+6)); \
+  R(b,c,d,e,f,g,h,a, (i+7))
+
+#else
+
+#define R(i) h(i) += S1(e(i)) + Ch(e(i),f(i),g(i)) + K[i+j] + (j?blk2(i):blk0(i));\
+  d(i) += h(i); h(i) += S0(a(i)) + Maj(a(i), b(i), c(i))
+
+#ifdef _SHA256_UNROLL
+
+#define RX_8(i) R(i+0); R(i+1); R(i+2); R(i+3); R(i+4); R(i+5); R(i+6); R(i+7);
+
+#endif
+
+#endif
+
+static const uint32_t K[64] = {
+  0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5,
+  0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
+  0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3,
+  0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
+  0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc,
+  0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
+  0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7,
+  0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
+  0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13,
+  0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
+  0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3,
+  0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
+  0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5,
+  0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
+  0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208,
+  0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
+};
+
+static void
+sha256_transform(uint32_t *state, const uint32_t *data)
+{
+  uint32_t W[16] = {0};
+  unsigned j;
+  #ifdef _SHA256_UNROLL2
+  uint32_t a,b,c,d,e,f,g,h;
+  a = state[0];
+  b = state[1];
+  c = state[2];
+  d = state[3];
+  e = state[4];
+  f = state[5];
+  g = state[6];
+  h = state[7];
+  #else
+  uint32_t T[8];
+  for (j = 0; j < 8; j++)
+    T[j] = state[j];
+  #endif
+
+  for (j = 0; j < 64; j += 16)
+  {
+    #if defined(_SHA256_UNROLL) || defined(_SHA256_UNROLL2)
+    RX_8(0); RX_8(8);
+    #else
+    unsigned i;
+    for (i = 0; i < 16; i++) { R(i); }
+    #endif
+  }
+
+  #ifdef _SHA256_UNROLL2
+  state[0] += a;
+  state[1] += b;
+  state[2] += c;
+  state[3] += d;
+  state[4] += e;
+  state[5] += f;
+  state[6] += g;
+  state[7] += h;
+  #else
+  for (j = 0; j < 8; j++)
+    state[j] += T[j];
+  #endif
+
+  /* Wipe variables */
+  /* memset(W, 0, sizeof(W)); */
+  /* memset(T, 0, sizeof(T)); */
+}
+
+#undef S0
+#undef S1
+#undef s0
+#undef s1
+
+static void
+sha256_write_byte_block(sha256_t *p)
+{
+  uint32_t data32[16];
+  unsigned i;
+  for (i = 0; i < 16; i++)
+    data32[i] =
+      ((uint32_t)(p->buffer[i * 4    ]) << 24) +
+      ((uint32_t)(p->buffer[i * 4 + 1]) << 16) +
+      ((uint32_t)(p->buffer[i * 4 + 2]) <<  8) +
+      ((uint32_t)(p->buffer[i * 4 + 3]));
+  sha256_transform(p->state, data32);
+}
+
+
+void
+sha256_hash(unsigned char *buf, const unsigned char *data, size_t size)
+{
+  sha256_t hash;
+  sha256_init(&hash);
+  sha256_update(&hash, data, size);
+  sha256_final(&hash, buf);
+}
+
+
+void
+sha256_update(sha256_t *p, const unsigned char *data, size_t size)
+{
+  uint32_t curBufferPos = (uint32_t)p->count & 0x3F;
+  while (size > 0)
+  {
+    p->buffer[curBufferPos++] = *data++;
+    p->count++;
+    size--;
+    if (curBufferPos == 64)
+    {
+      curBufferPos = 0;
+      sha256_write_byte_block(p);
+    }
+  }
+}
+
+
+void
+sha256_final(sha256_t *p, unsigned char *digest)
+{
+  uint64_t lenInBits = (p->count << 3);
+  uint32_t curBufferPos = (uint32_t)p->count & 0x3F;
+  unsigned i;
+  p->buffer[curBufferPos++] = 0x80;
+  while (curBufferPos != (64 - 8))
+  {
+    curBufferPos &= 0x3F;
+    if (curBufferPos == 0)
+      sha256_write_byte_block(p);
+    p->buffer[curBufferPos++] = 0;
+  }
+  for (i = 0; i < 8; i++)
+  {
+    p->buffer[curBufferPos++] = (unsigned char)(lenInBits >> 56);
+    lenInBits <<= 8;
+  }
+  sha256_write_byte_block(p);
+
+  for (i = 0; i < 8; i++)
+  {
+    *digest++ = (unsigned char)(p->state[i] >> 24);
+    *digest++ = (unsigned char)(p->state[i] >> 16);
+    *digest++ = (unsigned char)(p->state[i] >> 8);
+    *digest++ = (unsigned char)(p->state[i]);
+  }
+  sha256_init(p);
+}
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.h b/llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.h
new file mode 100644
index 0000000000000000000000000000000000000000..21657e66b602d5e74e90fbd23af64d58d94598de
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.h
@@ -0,0 +1,24 @@
+/* Sha256.h -- SHA-256 Hash
+2010-06-11 : Igor Pavlov : Public domain */
+
+#ifndef __CRYPTO_SHA256_H
+#define __CRYPTO_SHA256_H
+
+#include 
+#include 
+
+#define SHA256_DIGEST_SIZE 32
+
+typedef struct sha256_t
+{
+  uint32_t state[8];
+  uint64_t count;
+  unsigned char buffer[64];
+} sha256_t;
+
+void sha256_init(sha256_t *p);
+void sha256_update(sha256_t *p, const unsigned char *data, size_t size);
+void sha256_final(sha256_t *p, unsigned char *digest);
+void sha256_hash(unsigned char *buf, const unsigned char *data, size_t size);
+
+#endif
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/xxhash/clib.json b/llm/llama.cpp/examples/gguf-hash/deps/xxhash/clib.json
new file mode 100644
index 0000000000000000000000000000000000000000..242343c5d992d4a19b9a9fe9d3845c5f12525569
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/xxhash/clib.json
@@ -0,0 +1,12 @@
+{
+  "name": "xxhash",
+  "version": "0.8.2",
+  "repo": "Cyan4973/xxhash",
+  "description": "Extremely fast non-cryptographic hash algorithm",
+  "keywords": ["xxhash", "hashing"],
+  "license": "BSD-2-Clause",
+  "src": [
+    "xxhash.c",
+    "xxhash.h"
+  ]
+}
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/xxhash/xxhash.c b/llm/llama.cpp/examples/gguf-hash/deps/xxhash/xxhash.c
new file mode 100644
index 0000000000000000000000000000000000000000..e60cc37f13c27ff0668ec2535937df3f4c097ec2
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/xxhash/xxhash.c
@@ -0,0 +1,42 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Copyright (C) 2012-2023 Yann Collet
+ *
+ * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *    * Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *    * Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following disclaimer
+ *      in the documentation and/or other materials provided with the
+ *      distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at:
+ *   - xxHash homepage: https://www.xxhash.com
+ *   - xxHash source repository: https://github.com/Cyan4973/xxHash
+ */
+
+/*
+ * xxhash.c instantiates functions defined in xxhash.h
+ */
+
+#define XXH_STATIC_LINKING_ONLY /* access advanced declarations */
+#define XXH_IMPLEMENTATION      /* access definitions */
+
+#include "xxhash.h"
diff --git a/llm/llama.cpp/examples/gguf-hash/deps/xxhash/xxhash.h b/llm/llama.cpp/examples/gguf-hash/deps/xxhash/xxhash.h
new file mode 100644
index 0000000000000000000000000000000000000000..c0fafe20d54ad017425e636c2f3e754494648053
--- /dev/null
+++ b/llm/llama.cpp/examples/gguf-hash/deps/xxhash/xxhash.h
@@ -0,0 +1,7093 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Header File
+ * Copyright (C) 2012-2023 Yann Collet
+ *
+ * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *    * Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *    * Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following disclaimer
+ *      in the documentation and/or other materials provided with the
+ *      distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at:
+ *   - xxHash homepage: https://www.xxhash.com
+ *   - xxHash source repository: https://github.com/Cyan4973/xxHash
+ */
+
+/*!
+ * @mainpage xxHash
+ *
+ * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
+ * limits.
+ *
+ * It is proposed in four flavors, in three families:
+ * 1. @ref XXH32_family
+ *   - Classic 32-bit hash function. Simple, compact, and runs on almost all
+ *     32-bit and 64-bit systems.
+ * 2. @ref XXH64_family
+ *   - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
+ *     64-bit systems (but _not_ 32-bit systems).
+ * 3. @ref XXH3_family
+ *   - Modern 64-bit and 128-bit hash function family which features improved
+ *     strength and performance across the board, especially on smaller data.
+ *     It benefits greatly from SIMD and 64-bit without requiring it.
+ *
+ * Benchmarks
+ * ---
+ * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
+ * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
+ *
+ * | Hash Name            | ISA ext | Width | Large Data Speed | Small Data Velocity |
+ * | -------------------- | ------- | ----: | ---------------: | ------------------: |
+ * | XXH3_64bits()        | @b AVX2 |    64 |        59.4 GB/s |               133.1 |
+ * | MeowHash             | AES-NI  |   128 |        58.2 GB/s |                52.5 |
+ * | XXH3_128bits()       | @b AVX2 |   128 |        57.9 GB/s |               118.1 |
+ * | CLHash               | PCLMUL  |    64 |        37.1 GB/s |                58.1 |
+ * | XXH3_64bits()        | @b SSE2 |    64 |        31.5 GB/s |               133.1 |
+ * | XXH3_128bits()       | @b SSE2 |   128 |        29.6 GB/s |               118.1 |
+ * | RAM sequential read  |         |   N/A |        28.0 GB/s |                 N/A |
+ * | ahash                | AES-NI  |    64 |        22.5 GB/s |               107.2 |
+ * | City64               |         |    64 |        22.0 GB/s |                76.6 |
+ * | T1ha2                |         |    64 |        22.0 GB/s |                99.0 |
+ * | City128              |         |   128 |        21.7 GB/s |                57.7 |
+ * | FarmHash             | AES-NI  |    64 |        21.3 GB/s |                71.9 |
+ * | XXH64()              |         |    64 |        19.4 GB/s |                71.0 |
+ * | SpookyHash           |         |    64 |        19.3 GB/s |                53.2 |
+ * | Mum                  |         |    64 |        18.0 GB/s |                67.0 |
+ * | CRC32C               | SSE4.2  |    32 |        13.0 GB/s |                57.9 |
+ * | XXH32()              |         |    32 |         9.7 GB/s |                71.9 |
+ * | City32               |         |    32 |         9.1 GB/s |                66.0 |
+ * | Blake3*              | @b AVX2 |   256 |         4.4 GB/s |                 8.1 |
+ * | Murmur3              |         |    32 |         3.9 GB/s |                56.1 |
+ * | SipHash*             |         |    64 |         3.0 GB/s |                43.2 |
+ * | Blake3*              | @b SSE2 |   256 |         2.4 GB/s |                 8.1 |
+ * | HighwayHash          |         |    64 |         1.4 GB/s |                 6.0 |
+ * | FNV64                |         |    64 |         1.2 GB/s |                62.7 |
+ * | Blake2*              |         |   256 |         1.1 GB/s |                 5.1 |
+ * | SHA1*                |         |   160 |         0.8 GB/s |                 5.6 |
+ * | MD5*                 |         |   128 |         0.6 GB/s |                 7.8 |
+ * @note
+ *   - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
+ *     even though it is mandatory on x64.
+ *   - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
+ *     by modern standards.
+ *   - Small data velocity is a rough average of algorithm's efficiency for small
+ *     data. For more accurate information, see the wiki.
+ *   - More benchmarks and strength tests are found on the wiki:
+ *         https://github.com/Cyan4973/xxHash/wiki
+ *
+ * Usage
+ * ------
+ * All xxHash variants use a similar API. Changing the algorithm is a trivial
+ * substitution.
+ *
+ * @pre
+ *    For functions which take an input and length parameter, the following
+ *    requirements are assumed:
+ *    - The range from [`input`, `input + length`) is valid, readable memory.
+ *      - The only exception is if the `length` is `0`, `input` may be `NULL`.
+ *    - For C++, the objects must have the *TriviallyCopyable* property, as the
+ *      functions access bytes directly as if it was an array of `unsigned char`.
+ *
+ * @anchor single_shot_example
+ * **Single Shot**
+ *
+ * These functions are stateless functions which hash a contiguous block of memory,
+ * immediately returning the result. They are the easiest and usually the fastest
+ * option.
+ *
+ * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
+ *
+ * @code{.c}
+ *   #include 
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which hashes a null terminated string with XXH32().
+ *   XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
+ *   {
+ *       // NULL pointers are only valid if the length is zero
+ *       size_t length = (string == NULL) ? 0 : strlen(string);
+ *       return XXH32(string, length, seed);
+ *   }
+ * @endcode
+ *
+ *
+ * @anchor streaming_example
+ * **Streaming**
+ *
+ * These groups of functions allow incremental hashing of unknown size, even
+ * more than what would fit in a size_t.
+ *
+ * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
+ *
+ * @code{.c}
+ *   #include 
+ *   #include 
+ *   #include "xxhash.h"
+ *   // Example for a function which hashes a FILE incrementally with XXH3_64bits().
+ *   XXH64_hash_t hashFile(FILE* f)
+ *   {
+ *       // Allocate a state struct. Do not just use malloc() or new.
+ *       XXH3_state_t* state = XXH3_createState();
+ *       assert(state != NULL && "Out of memory!");
+ *       // Reset the state to start a new hashing session.
+ *       XXH3_64bits_reset(state);
+ *       char buffer[4096];
+ *       size_t count;
+ *       // Read the file in chunks
+ *       while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
+ *           // Run update() as many times as necessary to process the data
+ *           XXH3_64bits_update(state, buffer, count);
+ *       }
+ *       // Retrieve the finalized hash. This will not change the state.
+ *       XXH64_hash_t result = XXH3_64bits_digest(state);
+ *       // Free the state. Do not use free().
+ *       XXH3_freeState(state);
+ *       return result;
+ *   }
+ * @endcode
+ *
+ * Streaming functions generate the xxHash value from an incremental input.
+ * This method is slower than single-call functions, due to state management.
+ * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+ *
+ * An XXH state must first be allocated using `XXH*_createState()`.
+ *
+ * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ *
+ * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ *
+ * The function returns an error code, with 0 meaning OK, and any other value
+ * meaning there is an error.
+ *
+ * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
+ * This function returns the nn-bits hash as an int or long long.
+ *
+ * It's still possible to continue inserting input into the hash state after a
+ * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ *
+ * When done, release the state using `XXH*_freeState()`.
+ *
+ *
+ * @anchor canonical_representation_example
+ * **Canonical Representation**
+ *
+ * The default return values from XXH functions are unsigned 32, 64 and 128 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ *
+ * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(),
+ * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(),
+ * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(),
+ *
+ * @code{.c}
+ *   #include 
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which prints XXH32_hash_t in human readable format
+ *   void printXxh32(XXH32_hash_t hash)
+ *   {
+ *       XXH32_canonical_t cano;
+ *       XXH32_canonicalFromHash(&cano, hash);
+ *       size_t i;
+ *       for(i = 0; i < sizeof(cano.digest); ++i) {
+ *           printf("%02x", cano.digest[i]);
+ *       }
+ *       printf("\n");
+ *   }
+ *
+ *   // Example for a function which converts XXH32_canonical_t to XXH32_hash_t
+ *   XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano)
+ *   {
+ *       XXH32_hash_t hash = XXH32_hashFromCanonical(&cano);
+ *       return hash;
+ *   }
+ * @endcode
+ *
+ *
+ * @file xxhash.h
+ * xxHash prototypes and implementation
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ****************************
+ *  INLINE mode
+ ******************************/
+/*!
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Gives access to internal state declaration, required for static allocation.
+ *
+ * Incompatible with dynamic linking, due to risks of ABI changes.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_STATIC_LINKING_ONLY
+/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
+
+/*!
+ * @brief Gives access to internal definitions.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #define XXH_IMPLEMENTATION
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_IMPLEMENTATION
+/* Do not undef XXH_IMPLEMENTATION for Doxygen */
+
+/*!
+ * @brief Exposes the implementation and marks all functions as `inline`.
+ *
+ * Use these build macros to inline xxhash into the target unit.
+ * Inlining improves performance on small inputs, especially when the length is
+ * expressed as a compile-time constant:
+ *
+ *  https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *
+ * It also keeps xxHash symbols private to the unit, so they are not exported.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_INLINE_ALL
+ *     #include "xxhash.h"
+ * @endcode
+ * Do not compile and link xxhash.o as a separate object, as it is not useful.
+ */
+#  define XXH_INLINE_ALL
+#  undef XXH_INLINE_ALL
+/*!
+ * @brief Exposes the implementation without marking functions as inline.
+ */
+#  define XXH_PRIVATE_API
+#  undef XXH_PRIVATE_API
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+#  define XXH_NAMESPACE /* YOUR NAME HERE */
+#  undef XXH_NAMESPACE
+#endif
+
+#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
+    && !defined(XXH_INLINE_ALL_31684351384)
+   /* this section should be traversed only once */
+#  define XXH_INLINE_ALL_31684351384
+   /* give access to the advanced API, required to compile implementations */
+#  undef XXH_STATIC_LINKING_ONLY   /* avoid macro redef */
+#  define XXH_STATIC_LINKING_ONLY
+   /* make all functions private */
+#  undef XXH_PUBLIC_API
+#  if defined(__GNUC__)
+#    define XXH_PUBLIC_API static __inline __attribute__((__unused__))
+#  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#    define XXH_PUBLIC_API static inline
+#  elif defined(_MSC_VER)
+#    define XXH_PUBLIC_API static __inline
+#  else
+     /* note: this version may generate warnings for unused static functions */
+#    define XXH_PUBLIC_API static
+#  endif
+
+   /*
+    * This part deals with the special case where a unit wants to inline xxHash,
+    * but "xxhash.h" has previously been included without XXH_INLINE_ALL,
+    * such as part of some previously included *.h header file.
+    * Without further action, the new include would just be ignored,
+    * and functions would effectively _not_ be inlined (silent failure).
+    * The following macros solve this situation by prefixing all inlined names,
+    * avoiding naming collision with previous inclusions.
+    */
+   /* Before that, we unconditionally #undef all symbols,
+    * in case they were already defined with XXH_NAMESPACE.
+    * They will then be redefined for XXH_INLINE_ALL
+    */
+#  undef XXH_versionNumber
+    /* XXH32 */
+#  undef XXH32
+#  undef XXH32_createState
+#  undef XXH32_freeState
+#  undef XXH32_reset
+#  undef XXH32_update
+#  undef XXH32_digest
+#  undef XXH32_copyState
+#  undef XXH32_canonicalFromHash
+#  undef XXH32_hashFromCanonical
+    /* XXH64 */
+#  undef XXH64
+#  undef XXH64_createState
+#  undef XXH64_freeState
+#  undef XXH64_reset
+#  undef XXH64_update
+#  undef XXH64_digest
+#  undef XXH64_copyState
+#  undef XXH64_canonicalFromHash
+#  undef XXH64_hashFromCanonical
+    /* XXH3_64bits */
+#  undef XXH3_64bits
+#  undef XXH3_64bits_withSecret
+#  undef XXH3_64bits_withSeed
+#  undef XXH3_64bits_withSecretandSeed
+#  undef XXH3_createState
+#  undef XXH3_freeState
+#  undef XXH3_copyState
+#  undef XXH3_64bits_reset
+#  undef XXH3_64bits_reset_withSeed
+#  undef XXH3_64bits_reset_withSecret
+#  undef XXH3_64bits_update
+#  undef XXH3_64bits_digest
+#  undef XXH3_generateSecret
+    /* XXH3_128bits */
+#  undef XXH128
+#  undef XXH3_128bits
+#  undef XXH3_128bits_withSeed
+#  undef XXH3_128bits_withSecret
+#  undef XXH3_128bits_reset
+#  undef XXH3_128bits_reset_withSeed
+#  undef XXH3_128bits_reset_withSecret
+#  undef XXH3_128bits_reset_withSecretandSeed
+#  undef XXH3_128bits_update
+#  undef XXH3_128bits_digest
+#  undef XXH128_isEqual
+#  undef XXH128_cmp
+#  undef XXH128_canonicalFromHash
+#  undef XXH128_hashFromCanonical
+    /* Finally, free the namespace itself */
+#  undef XXH_NAMESPACE
+
+    /* employ the namespace for XXH_INLINE_ALL */
+#  define XXH_NAMESPACE XXH_INLINE_
+   /*
+    * Some identifiers (enums, type names) are not symbols,
+    * but they must nonetheless be renamed to avoid redeclaration.
+    * Alternative solution: do not redeclare them.
+    * However, this requires some #ifdefs, and has a more dispersed impact.
+    * Meanwhile, renaming can be achieved in a single place.
+    */
+#  define XXH_IPREF(Id)   XXH_NAMESPACE ## Id
+#  define XXH_OK XXH_IPREF(XXH_OK)
+#  define XXH_ERROR XXH_IPREF(XXH_ERROR)
+#  define XXH_errorcode XXH_IPREF(XXH_errorcode)
+#  define XXH32_canonical_t  XXH_IPREF(XXH32_canonical_t)
+#  define XXH64_canonical_t  XXH_IPREF(XXH64_canonical_t)
+#  define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
+#  define XXH32_state_s XXH_IPREF(XXH32_state_s)
+#  define XXH32_state_t XXH_IPREF(XXH32_state_t)
+#  define XXH64_state_s XXH_IPREF(XXH64_state_s)
+#  define XXH64_state_t XXH_IPREF(XXH64_state_t)
+#  define XXH3_state_s  XXH_IPREF(XXH3_state_s)
+#  define XXH3_state_t  XXH_IPREF(XXH3_state_t)
+#  define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
+   /* Ensure the header is parsed again, even if it was previously included */
+#  undef XXHASH_H_5627135585666179
+#  undef XXHASH_H_STATIC_13879238742
+#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
+
+/* ****************************************************************
+ *  Stable API
+ *****************************************************************/
+#ifndef XXHASH_H_5627135585666179
+#define XXHASH_H_5627135585666179 1
+
+/*! @brief Marks a global symbol. */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
+#endif
+
+#ifdef XXH_NAMESPACE
+#  define XXH_CAT(A,B) A##B
+#  define XXH_NAME2(A,B) XXH_CAT(A,B)
+#  define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
+/* XXH32 */
+#  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
+#  define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
+#  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
+#  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
+#  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
+#  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
+#  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
+#  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
+#  define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
+/* XXH64 */
+#  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
+#  define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
+#  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
+#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
+#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
+#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
+#  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
+#  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
+#  define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
+/* XXH3_64bits */
+#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
+#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
+#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+#  define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
+#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
+#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
+#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
+#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
+#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
+#  define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
+#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
+#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+#  define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
+#  define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
+/* XXH3_128bits */
+#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
+#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
+#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
+#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+#  define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
+#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
+#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
+#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
+#  define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
+#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
+#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
+#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
+#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
+#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
+#endif
+
+
+/* *************************************
+*  Compiler specifics
+***************************************/
+
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
+#endif
+
+#if defined (__GNUC__)
+# define XXH_CONSTF  __attribute__((__const__))
+# define XXH_PUREF   __attribute__((__pure__))
+# define XXH_MALLOCF __attribute__((__malloc__))
+#else
+# define XXH_CONSTF  /* disable */
+# define XXH_PUREF
+# define XXH_MALLOCF
+#endif
+
+/* *************************************
+*  Version
+***************************************/
+#define XXH_VERSION_MAJOR    0
+#define XXH_VERSION_MINOR    8
+#define XXH_VERSION_RELEASE  3
+/*! @brief Version number, encoded as two digits each */
+#define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
+
+/*!
+ * @brief Obtains the xxHash version.
+ *
+ * This is mostly useful when xxHash is compiled as a shared library,
+ * since the returned value comes from the library, as opposed to header file.
+ *
+ * @return @ref XXH_VERSION_NUMBER of the invoked library.
+ */
+XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
+
+
+/* ****************************
+*  Common basic types
+******************************/
+#include    /* size_t */
+/*!
+ * @brief Exit code for the streaming API.
+ */
+typedef enum {
+    XXH_OK = 0, /*!< OK */
+    XXH_ERROR   /*!< Error */
+} XXH_errorcode;
+
+
+/*-**********************************************************************
+*  32-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* Don't show  include */
+/*!
+ * @brief An unsigned 32-bit integer.
+ *
+ * Not necessarily defined to `uint32_t` but functionally equivalent.
+ */
+typedef uint32_t XXH32_hash_t;
+
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   ifdef _AIX
+#     include 
+#   else
+#     include 
+#   endif
+    typedef uint32_t XXH32_hash_t;
+
+#else
+#   include 
+#   if UINT_MAX == 0xFFFFFFFFUL
+      typedef unsigned int XXH32_hash_t;
+#   elif ULONG_MAX == 0xFFFFFFFFUL
+      typedef unsigned long XXH32_hash_t;
+#   else
+#     error "unsupported platform: need a 32-bit type"
+#   endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup XXH32_family XXH32 family
+ * @ingroup public
+ * Contains functions used in the classic 32-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH32 is useful for older platforms, with no or poor 64-bit performance.
+ *   Note that the @ref XXH3_family provides competitive speed for both 32-bit
+ *   and 64-bit systems, and offers true 64/128 bit hash results.
+ *
+ * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
+ * @see @ref XXH32_impl for implementation details
+ * @{
+ */
+
+/*!
+ * @brief Calculates the 32-bit hash of @p input using xxHash32.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 32-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 32-bit xxHash32 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+
+#ifndef XXH_NO_STREAM
+/*!
+ * @typedef struct XXH32_state_s XXH32_state_t
+ * @brief The opaque state struct for the XXH32 streaming API.
+ *
+ * @see XXH32_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
+typedef struct XXH32_state_s XXH32_state_t;
+
+/*!
+ * @brief Allocates an @ref XXH32_state_t.
+ *
+ * @return An allocated pointer of @ref XXH32_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH32_freeState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
+/*!
+ * @brief Frees an @ref XXH32_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH32_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH32_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH32_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 32-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH32_update().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH32_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH32_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated 32-bit xxHash32 value from that state.
+ *
+ * @note
+ *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/*******   Canonical representation   *******/
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
+ */
+typedef struct {
+    unsigned char digest[4]; /*!< Hash bytes, big endian */
+} XXH32_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
+ *
+ * @param dst  The @ref XXH32_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH32_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
+ *
+ * @param src The @ref XXH32_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+
+
+/*! @cond Doxygen ignores this part */
+#ifdef __has_attribute
+# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
+#else
+# define XXH_HAS_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
+# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define XXH_HAS_C_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define XXH_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
+ * introduced in CPP17 and C23.
+ * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
+ */
+#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
+# define XXH_FALLTHROUGH [[fallthrough]]
+#elif XXH_HAS_ATTRIBUTE(__fallthrough__)
+# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
+#else
+# define XXH_FALLTHROUGH /* fallthrough */
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_NOESCAPE for annotated pointers in public API.
+ * https://clang.llvm.org/docs/AttributeReference.html#noescape
+ * As of writing this, only supported by clang.
+ */
+#if XXH_HAS_ATTRIBUTE(noescape)
+# define XXH_NOESCAPE __attribute__((__noescape__))
+#else
+# define XXH_NOESCAPE
+#endif
+/*! @endcond */
+
+
+/*!
+ * @}
+ * @ingroup public
+ * @{
+ */
+
+#ifndef XXH_NO_LONG_LONG
+/*-**********************************************************************
+*  64-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* don't include  */
+/*!
+ * @brief An unsigned 64-bit integer.
+ *
+ * Not necessarily defined to `uint64_t` but functionally equivalent.
+ */
+typedef uint64_t XXH64_hash_t;
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   ifdef _AIX
+#     include 
+#   else
+#     include 
+#   endif
+   typedef uint64_t XXH64_hash_t;
+#else
+#  include 
+#  if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
+     /* LP64 ABI says uint64_t is unsigned long */
+     typedef unsigned long XXH64_hash_t;
+#  else
+     /* the following type must have a width of 64-bit */
+     typedef unsigned long long XXH64_hash_t;
+#  endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup XXH64_family XXH64 family
+ * @ingroup public
+ * @{
+ * Contains functions used in the classic 64-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH3 provides competitive speed for both 32-bit and 64-bit systems,
+ *   and offers true 64/128 bit hash results.
+ *   It provides better speed for systems with vector processing capabilities.
+ */
+
+/*!
+ * @brief Calculates the 64-bit hash of @p input using xxHash64.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit xxHash64 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief The opaque state struct for the XXH64 streaming API.
+ *
+ * @see XXH64_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
+typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
+
+/*!
+ * @brief Allocates an @ref XXH64_state_t.
+ *
+ * @return An allocated pointer of @ref XXH64_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH64_freeState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Frees an @ref XXH64_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH64_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH64_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH64_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH64_update().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated 64-bit xxHash64 value from that state.
+ *
+ * @note
+ *   Calling XXH64_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+/*******   Canonical representation   *******/
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
+ */
+typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
+ *
+ * @param dst The @ref XXH64_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH64_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
+ *
+ * @param src The @ref XXH64_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
+
+#ifndef XXH_NO_XXH3
+
+/*!
+ * @}
+ * ************************************************************************
+ * @defgroup XXH3_family XXH3 family
+ * @ingroup public
+ * @{
+ *
+ * XXH3 is a more recent hash algorithm featuring:
+ *  - Improved speed for both small and large inputs
+ *  - True 64-bit and 128-bit outputs
+ *  - SIMD acceleration
+ *  - Improved 32-bit viability
+ *
+ * Speed analysis methodology is explained here:
+ *
+ *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * Compared to XXH64, expect XXH3 to run approximately
+ * ~2x faster on large inputs and >3x faster on small ones,
+ * exact differences vary depending on platform.
+ *
+ * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
+ * but does not require it.
+ * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
+ * at competitive speeds, even without vector support. Further details are
+ * explained in the implementation.
+ *
+ * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
+ * implementations for many common platforms:
+ *   - AVX512
+ *   - AVX2
+ *   - SSE2
+ *   - ARM NEON
+ *   - WebAssembly SIMD128
+ *   - POWER8 VSX
+ *   - s390x ZVector
+ * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
+ * selects the best version according to predefined macros. For the x86 family, an
+ * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
+ *
+ * XXH3 implementation is portable:
+ * it has a generic C90 formulation that can be compiled on any platform,
+ * all implementations generate exactly the same hash value on all platforms.
+ * Starting from v0.8.0, it's also labelled "stable", meaning that
+ * any future version will also generate the same hash value.
+ *
+ * XXH3 offers 2 variants, _64bits and _128bits.
+ *
+ * When only 64 bits are needed, prefer invoking the _64bits variant, as it
+ * reduces the amount of mixing, resulting in faster speed on small inputs.
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ */
+/*-**********************************************************************
+*  XXH3 64-bit variant
+************************************************************************/
+
+/*!
+ * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *   This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however
+ *   it may have slightly better performance due to constant propagation of the
+ *   defaults.
+ *
+ * @see
+ *    XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input.
+ *
+ * @param input  The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed   The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/*!
+ * The bare minimum size for a custom secret.
+ *
+ * @see
+ *  XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
+ *  XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
+ */
+#define XXH3_SECRET_SIZE_MIN 136
+
+/*!
+ * @brief Calculates 64-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ */
+
+/*!
+ * @brief The opaque state struct for the XXH3 streaming API.
+ *
+ * @see XXH3_state_s for details.
+ * @see @ref streaming_example "Streaming Example"
+ */
+typedef struct XXH3_state_s XXH3_state_t;
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH3_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call this function before @ref XXH3_64bits_update().
+ *   - Digest will be equivalent to `XXH3_64bits_withSeed()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   `secret` is referenced, it _must outlive_ the hash streaming session.
+ *
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 64-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t  XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* note : canonical representation of XXH3 is the same as XXH64
+ * since they both produce XXH64_hash_t values */
+
+
+/*-**********************************************************************
+*  XXH3 128-bit variant
+************************************************************************/
+
+/*!
+ * @brief The return value from 128-bit hashes.
+ *
+ * Stored in little endian order, although the fields themselves are in native
+ * endianness.
+ */
+typedef struct {
+    XXH64_hash_t low64;   /*!< `value & 0xFFFFFFFFFFFFFFFF` */
+    XXH64_hash_t high64;  /*!< `value >> 64` */
+} XXH128_hash_t;
+
+/*!
+ * @brief Calculates 128-bit unseeded variant of XXH3 of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
+ * for shorter inputs.
+ *
+ * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
+/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+/*!
+ * @brief Calculates 128-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ *
+ * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
+ * Use already declared XXH3_createState() and XXH3_freeState().
+ *
+ * All reset and streaming functions have same meaning as their 64-bit counterpart.
+ */
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret with default parameters.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   - This function resets `statePtr` and generate a secret from `seed`.
+ *   - Call it before @ref XXH3_128bits_update().
+ *   - Digest will be equivalent to `XXH3_128bits_withSeed()`.
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   The state struct to reset.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 128-bit hash value from that state.
+ *
+ * @note
+ *   Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* Following helper functions make it possible to compare XXH128_hast_t values.
+ * Since XXH128_hash_t is a structure, this capability is not offered by the language.
+ * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
+
+/*!
+ * @brief Check equality of two XXH128_hash_t values
+ *
+ * @param h1 The 128-bit hash value.
+ * @param h2 Another 128-bit hash value.
+ *
+ * @return `1` if `h1` and `h2` are equal.
+ * @return `0` if they are not.
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * @brief Compares two @ref XXH128_hash_t
+ *
+ * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ *
+ * @param h128_1 Left-hand side value
+ * @param h128_2 Right-hand side value
+ *
+ * @return >0 if @p h128_1  > @p h128_2
+ * @return =0 if @p h128_1 == @p h128_2
+ * @return <0 if @p h128_1  < @p h128_2
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
+
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+
+
+/*!
+ * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
+ *
+ * @param dst  The @ref XXH128_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH128_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
+ *
+ * @param src The @ref XXH128_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
+
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+
+/*!
+ * @}
+ */
+#endif /* XXHASH_H_5627135585666179 */
+
+
+
+#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
+#define XXHASH_H_STATIC_13879238742
+/* ****************************************************************************
+ * This section contains declarations which are not guaranteed to remain stable.
+ * They may change in future versions, becoming incompatible with a different
+ * version of the library.
+ * These declarations should only be used with static linking.
+ * Never use them in association with dynamic linking!
+ ***************************************************************************** */
+
+/*
+ * These definitions are only present to allow static allocation
+ * of XXH states, on stack or in a struct, for example.
+ * Never **ever** access their members directly.
+ */
+
+/*!
+ * @internal
+ * @brief Structure for XXH32 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH32_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH64_state_s, XXH3_state_s
+ */
+struct XXH32_state_s {
+   XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
+   XXH32_hash_t large_len;    /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
+   XXH32_hash_t v[4];         /*!< Accumulator lanes */
+   XXH32_hash_t mem32[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem32 */
+   XXH32_hash_t reserved;     /*!< Reserved field. Do not read nor write to it. */
+};   /* typedef'd to XXH32_state_t */
+
+
+#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+
+/*!
+ * @internal
+ * @brief Structure for XXH64 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH64_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH32_state_s, XXH3_state_s
+ */
+struct XXH64_state_s {
+   XXH64_hash_t total_len;    /*!< Total length hashed. This is always 64-bit. */
+   XXH64_hash_t v[4];         /*!< Accumulator lanes */
+   XXH64_hash_t mem64[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem64 */
+   XXH32_hash_t reserved32;   /*!< Reserved field, needed for padding anyways*/
+   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
+};   /* typedef'd to XXH64_state_t */
+
+#ifndef XXH_NO_XXH3
+
+/* Windows SDK under 10.0.22000 is missing stdalign.h so we add a check
+   before allowing the windows compiler to use the C11 form.
+   Reference: https://github.com/Cyan4973/xxHash/issues/955 */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) \
+    && (defined(_MSC_VER) && (_MSC_VER >= 1000) || !defined(_MSC_VER)) /* >= C11 */
+#  include 
+#  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
+/* In C++ alignas() is a keyword */
+#  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__GNUC__)
+#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
+#elif defined(_MSC_VER)
+#  define XXH_ALIGN(n)      __declspec(align(n))
+#else
+#  define XXH_ALIGN(n)   /* disabled */
+#endif
+
+/* Old GCC versions only accept the attribute after the type in structures. */
+#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
+    && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
+    && defined(__GNUC__)
+#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+#else
+#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+#endif
+
+/*!
+ * @brief The size of the internal XXH3 buffer.
+ *
+ * This is the optimal update size for incremental hashing.
+ *
+ * @see XXH3_64b_update(), XXH3_128b_update().
+ */
+#define XXH3_INTERNALBUFFER_SIZE 256
+
+/*!
+ * @internal
+ * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
+ *
+ * This is the size used in @ref XXH3_kSecret and the seeded functions.
+ *
+ * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
+ */
+#define XXH3_SECRET_DEFAULT_SIZE 192
+
+/*!
+ * @internal
+ * @brief Structure for XXH3 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined.
+ * Otherwise it is an opaque type.
+ * Never use this definition in combination with dynamic library.
+ * This allows fields to safely be changed in the future.
+ *
+ * @note ** This structure has a strict alignment requirement of 64 bytes!! **
+ * Do not allocate this with `malloc()` or `new`,
+ * it will not be sufficiently aligned.
+ * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation.
+ *
+ * Typedef'd to @ref XXH3_state_t.
+ * Do never access the members of this struct directly.
+ *
+ * @see XXH3_INITSTATE() for stack initialization.
+ * @see XXH3_createState(), XXH3_freeState().
+ * @see XXH32_state_s, XXH64_state_s
+ */
+struct XXH3_state_s {
+   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
+       /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
+   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
+       /*!< Used to store a custom secret generated from a seed. */
+   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
+       /*!< The internal buffer. @see XXH32_state_s::mem32 */
+   XXH32_hash_t bufferedSize;
+       /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
+   XXH32_hash_t useSeed;
+       /*!< Reserved field. Needed for padding on 64-bit. */
+   size_t nbStripesSoFar;
+       /*!< Number or stripes processed. */
+   XXH64_hash_t totalLen;
+       /*!< Total length hashed. 64-bit even on 32-bit targets. */
+   size_t nbStripesPerBlock;
+       /*!< Number of stripes per block. */
+   size_t secretLimit;
+       /*!< Size of @ref customSecret or @ref extSecret */
+   XXH64_hash_t seed;
+       /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
+   XXH64_hash_t reserved64;
+       /*!< Reserved field. */
+   const unsigned char* extSecret;
+       /*!< Reference to an external secret for the _withSecret variants, NULL
+        *   for other variants. */
+   /* note: there may be some padding at the end due to alignment on 64 bytes */
+}; /* typedef'd to XXH3_state_t */
+
+#undef XXH_ALIGN_MEMBER
+
+/*!
+ * @brief Initializes a stack-allocated `XXH3_state_s`.
+ *
+ * When the @ref XXH3_state_t structure is merely emplaced on stack,
+ * it should be initialized with XXH3_INITSTATE() or a memset()
+ * in case its first reset uses XXH3_NNbits_reset_withSeed().
+ * This init can be omitted if the first reset uses default or _withSecret mode.
+ * This operation isn't necessary when the state is created with XXH3_createState().
+ * Note that this doesn't prepare the state for a streaming operation,
+ * it's still necessary to use XXH3_NNbits_reset*() afterwards.
+ */
+#define XXH3_INITSTATE(XXH3_state_ptr)                       \
+    do {                                                     \
+        XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
+        tmp_xxh3_state_ptr->seed = 0;                        \
+        tmp_xxh3_state_ptr->extSecret = NULL;                \
+    } while(0)
+
+
+/*!
+ * @brief Calculates the 128-bit hash of @p data using XXH3.
+ *
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len  The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p data and @p data + @p len must be valid,
+ *   readable, contiguous memory. However, if @p len is `0`, @p data may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 128-bit XXH3 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+
+
+/* ===   Experimental API   === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
+/*!
+ * @brief Derive a high-entropy secret from any user-defined content, named customSeed.
+ *
+ * @param secretBuffer    A writable buffer for derived high-entropy secret data.
+ * @param secretSize      Size of secretBuffer, in bytes.  Must be >= XXH3_SECRET_SIZE_MIN.
+ * @param customSeed      A user-defined content.
+ * @param customSeedSize  Size of customSeed, in bytes.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * The generated secret can be used in combination with `*_withSecret()` functions.
+ * The `_withSecret()` variants are useful to provide a higher level of protection
+ * than 64-bit seed, as it becomes much more difficult for an external actor to
+ * guess how to impact the calculation logic.
+ *
+ * The function accepts as input a custom seed of any length and any content,
+ * and derives from it a high-entropy secret of length @p secretSize into an
+ * already allocated buffer @p secretBuffer.
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
+ * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
+ * are part of this list. They all accept a `secret` parameter
+ * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
+ * be employed to ensure proper quality.
+ *
+ * @p customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even "poor entropy" sources such as a bunch
+ * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
+ *
+ * @pre
+ *   - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
+ *   - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ *
+ * Example code:
+ * @code{.c}
+ *    #include 
+ *    #include 
+ *    #include 
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Hashes argv[2] using the entropy from argv[1].
+ *    int main(int argc, char* argv[])
+ *    {
+ *        char secret[XXH3_SECRET_SIZE_MIN];
+ *        if (argv != 3) { return 1; }
+ *        XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
+ *        XXH64_hash_t h = XXH3_64bits_withSecret(
+ *             argv[2], strlen(argv[2]),
+ *             secret, sizeof(secret)
+ *        );
+ *        printf("%016llx\n", (unsigned long long) h);
+ *    }
+ * @endcode
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);
+
+/*!
+ * @brief Generate the same secret as the _withSeed() variants.
+ *
+ * @param secretBuffer A writable buffer of @ref XXH3_SECRET_DEFAULT_SIZE bytes
+ * @param seed         The 64-bit seed to alter the hash result predictably.
+ *
+ * The generated secret can be used in combination with
+ *`*_withSecret()` and `_withSecretandSeed()` variants.
+ *
+ * Example C++ `std::string` hash class:
+ * @code{.cpp}
+ *    #include 
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Slow, seeds each time
+ *    class HashSlow {
+ *        XXH64_hash_t seed;
+ *    public:
+ *        HashSlow(XXH64_hash_t s) : seed{s} {}
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)};
+ *        }
+ *    };
+ *    // Fast, caches the seeded secret for future uses.
+ *    class HashFast {
+ *        unsigned char secret[XXH3_SECRET_DEFAULT_SIZE];
+ *    public:
+ *        HashFast(XXH64_hash_t s) {
+ *            XXH3_generateSecret_fromSeed(secret, seed);
+ *        }
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{
+ *                XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
+ *            };
+ *        }
+ *    };
+ * @endcode
+ */
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
+
+/*!
+ * @brief Maximum size of "short" key in bytes.
+ */
+#define XXH3_MIDSIZE_MAX 240
+
+/*!
+ * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data       The block of data to be hashed, at least @p len bytes in size.
+ * @param len        The length of @p data, in bytes.
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed       The 64-bit seed to alter the hash result predictably.
+ *
+ * These variants generate hash values using either:
+ * - @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes)
+ * - @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX).
+ *
+ * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
+ * `_withSeed()` has to generate the secret on the fly for "large" keys.
+ * It's fast, but can be perceptible for "not so large" keys (< 1 KB).
+ * `_withSecret()` has to generate the masks on the fly for "small" keys,
+ * which requires more instructions than _withSeed() variants.
+ * Therefore, _withSecretandSeed variant combines the best of both worlds.
+ *
+ * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
+ * this variant produces *exactly* the same results as `_withSeed()` variant,
+ * hence offering only a pure speed benefit on "large" input,
+ * by skipping the need to regenerate the secret for every large input.
+ *
+ * Another usage scenario is to hash the secret to a 64-bit hash value,
+ * for example with XXH3_64bits(), which then becomes the seed,
+ * and then employ both the seed and the secret in _withSecretandSeed().
+ * On top of speed, an added benefit is that each bit in the secret
+ * has a 50% chance to swap each bit in the output, via its impact to the seed.
+ *
+ * This is not guaranteed when using the secret directly in "small data" scenarios,
+ * because only portions of the secret are employed for small data.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
+                              XXH_NOESCAPE const void* secret, size_t secretSize,
+                              XXH64_hash_t seed);
+
+/*!
+ * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data       The memory segment to be hashed, at least @p len bytes in size.
+ * @param length     The length of @p data, in bytes.
+ * @param secret     The secret used to alter hash result predictably.
+ * @param secretSize The length of @p secret, in bytes (must be >= XXH3_SECRET_SIZE_MIN)
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(): contract is the same.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
+                               XXH_NOESCAPE const void* secret, size_t secretSize,
+                               XXH64_hash_t seed64);
+
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(). Contract is identical.
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                    XXH_NOESCAPE const void* secret, size_t secretSize,
+                                    XXH64_hash_t seed64);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret     The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64     The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed(). Contract is identical.
+ *
+ * Note: there was a bug in an earlier version of this function (<= v0.8.2)
+ * that would make it generate an incorrect hash value
+ * when @p seed == 0 and @p length < XXH3_MIDSIZE_MAX
+ * and @p secret is different from XXH3_generateSecret_fromSeed().
+ * As stated in the contract, the correct hash result must be
+ * the same as XXH3_128bits_withSeed() when @p length <= XXH3_MIDSIZE_MAX.
+ * Results generated by this older version are wrong, hence not comparable.
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                     XXH_NOESCAPE const void* secret, size_t secretSize,
+                                     XXH64_hash_t seed64);
+
+#endif /* !XXH_NO_STREAM */
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
+#  define XXH_IMPLEMENTATION
+#endif
+
+#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+
+
+/* ======================================================================== */
+/* ======================================================================== */
+/* ======================================================================== */
+
+
+/*-**********************************************************************
+ * xxHash implementation
+ *-**********************************************************************
+ * xxHash's implementation used to be hosted inside xxhash.c.
+ *
+ * However, inlining requires implementation to be visible to the compiler,
+ * hence be included alongside the header.
+ * Previously, implementation was hosted inside xxhash.c,
+ * which was then #included when inlining was activated.
+ * This construction created issues with a few build and install systems,
+ * as it required xxhash.c to be stored in /include directory.
+ *
+ * xxHash implementation is now directly integrated within xxhash.h.
+ * As a consequence, xxhash.c is no longer needed in /include.
+ *
+ * xxhash.c is still available and is still useful.
+ * In a "normal" setup, when xxhash is not inlined,
+ * xxhash.h only exposes the prototypes and public symbols,
+ * while xxhash.c can be built into an object file xxhash.o
+ * which can then be linked into the final binary.
+ ************************************************************************/
+
+#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
+   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
+#  define XXH_IMPLEM_13a8737387
+
+/* *************************************
+*  Tuning parameters
+***************************************/
+
+/*!
+ * @defgroup tuning Tuning parameters
+ * @{
+ *
+ * Various macros to control xxHash's behavior.
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Define this to disable 64-bit code.
+ *
+ * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
+ */
+#  define XXH_NO_LONG_LONG
+#  undef XXH_NO_LONG_LONG /* don't actually */
+/*!
+ * @brief Controls how unaligned memory is accessed.
+ *
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is
+ * safe and portable.
+ *
+ * Unfortunately, on some target/compiler combinations, the generated assembly
+ * is sub-optimal.
+ *
+ * The below switch allow selection of a different access method
+ * in the search for improved performance.
+ *
+ * @par Possible options:
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
+ *   @par
+ *     Use `memcpy()`. Safe and portable. Note that most modern compilers will
+ *     eliminate the function call and treat it as an unaligned access.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
+ *   @par
+ *     Depends on compiler extensions and is therefore not portable.
+ *     This method is safe _if_ your compiler supports it,
+ *     and *generally* as fast or faster than `memcpy`.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
+ *  @par
+ *     Casts directly and dereferences. This method doesn't depend on the
+ *     compiler, but it violates the C standard as it directly dereferences an
+ *     unaligned pointer. It can generate buggy code on targets which do not
+ *     support unaligned memory accesses, but in some circumstances, it's the
+ *     only known way to get the most performance.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
+ *  @par
+ *     Also portable. This can generate the best code on old compilers which don't
+ *     inline small `memcpy()` calls, and it might also be faster on big-endian
+ *     systems which lack a native byteswap instruction. However, some compilers
+ *     will emit literal byteshifts even if the target supports unaligned access.
+ *
+ *
+ * @warning
+ *   Methods 1 and 2 rely on implementation-defined behavior. Use these with
+ *   care, as what works on one compiler/platform/optimization level may cause
+ *   another to read garbage data or even crash.
+ *
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
+ *
+ * Prefer these methods in priority order (0 > 3 > 1 > 2)
+ */
+#  define XXH_FORCE_MEMORY_ACCESS 0
+
+/*!
+ * @def XXH_SIZE_OPT
+ * @brief Controls how much xxHash optimizes for size.
+ *
+ * xxHash, when compiled, tends to result in a rather large binary size. This
+ * is mostly due to heavy usage to forced inlining and constant folding of the
+ * @ref XXH3_family to increase performance.
+ *
+ * However, some developers prefer size over speed. This option can
+ * significantly reduce the size of the generated code. When using the `-Os`
+ * or `-Oz` options on GCC or Clang, this is defined to 1 by default,
+ * otherwise it is defined to 0.
+ *
+ * Most of these size optimizations can be controlled manually.
+ *
+ * This is a number from 0-2.
+ *  - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed
+ *    comes first.
+ *  - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more
+ *    conservative and disables hacks that increase code size. It implies the
+ *    options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0,
+ *    and @ref XXH3_NEON_LANES == 8 if they are not already defined.
+ *  - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible.
+ *    Performance may cry. For example, the single shot functions just use the
+ *    streaming API.
+ */
+#  define XXH_SIZE_OPT 0
+
+/*!
+ * @def XXH_FORCE_ALIGN_CHECK
+ * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
+ * and XXH64() only).
+ *
+ * This is an important performance trick for architectures without decent
+ * unaligned memory access performance.
+ *
+ * It checks for input alignment, and when conditions are met, uses a "fast
+ * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
+ * faster_ read speed.
+ *
+ * The check costs one initial branch per hash, which is generally negligible,
+ * but not zero.
+ *
+ * Moreover, it's not useful to generate an additional code path if memory
+ * access uses the same instruction for both aligned and unaligned
+ * addresses (e.g. x86 and aarch64).
+ *
+ * In these cases, the alignment check can be removed by setting this macro to 0.
+ * Then the code will always use unaligned memory access.
+ * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
+ *
+ * This option does not affect XXH3 (only XXH32 and XXH64).
+ */
+#  define XXH_FORCE_ALIGN_CHECK 0
+
+/*!
+ * @def XXH_NO_INLINE_HINTS
+ * @brief When non-zero, sets all functions to `static`.
+ *
+ * By default, xxHash tries to force the compiler to inline almost all internal
+ * functions.
+ *
+ * This can usually improve performance due to reduced jumping and improved
+ * constant folding, but significantly increases the size of the binary which
+ * might not be favorable.
+ *
+ * Additionally, sometimes the forced inlining can be detrimental to performance,
+ * depending on the architecture.
+ *
+ * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
+ * compiler full control on whether to inline or not.
+ *
+ * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
+ * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
+ */
+#  define XXH_NO_INLINE_HINTS 0
+
+/*!
+ * @def XXH3_INLINE_SECRET
+ * @brief Determines whether to inline the XXH3 withSecret code.
+ *
+ * When the secret size is known, the compiler can improve the performance
+ * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
+ *
+ * However, if the secret size is not known, it doesn't have any benefit. This
+ * happens when xxHash is compiled into a global symbol. Therefore, if
+ * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
+ *
+ * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
+ * that are *sometimes* force inline on -Og, and it is impossible to automatically
+ * detect this optimization level.
+ */
+#  define XXH3_INLINE_SECRET 0
+
+/*!
+ * @def XXH32_ENDJMP
+ * @brief Whether to use a jump for `XXH32_finalize`.
+ *
+ * For performance, `XXH32_finalize` uses multiple branches in the finalizer.
+ * This is generally preferable for performance,
+ * but depending on exact architecture, a jmp may be preferable.
+ *
+ * This setting is only possibly making a difference for very small inputs.
+ */
+#  define XXH32_ENDJMP 0
+
+/*!
+ * @internal
+ * @brief Redefines old internal names.
+ *
+ * For compatibility with code that uses xxHash's internals before the names
+ * were changed to improve namespacing. There is no other reason to use this.
+ */
+#  define XXH_OLD_NAMES
+#  undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
+
+/*!
+ * @def XXH_NO_STREAM
+ * @brief Disables the streaming API.
+ *
+ * When xxHash is not inlined and the streaming functions are not used, disabling
+ * the streaming functions can improve code size significantly, especially with
+ * the @ref XXH3_family which tends to make constant folded copies of itself.
+ */
+#  define XXH_NO_STREAM
+#  undef XXH_NO_STREAM /* don't actually */
+#endif /* XXH_DOXYGEN */
+/*!
+ * @}
+ */
+
+#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+   /* prefer __packed__ structures (method 1) for GCC
+    * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
+    * which for some reason does unaligned loads. */
+#  if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
+#    define XXH_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+#ifndef XXH_SIZE_OPT
+   /* default to 1 for -Os or -Oz */
+#  if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__)
+#    define XXH_SIZE_OPT 1
+#  else
+#    define XXH_SIZE_OPT 0
+#  endif
+#endif
+
+#ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
+   /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
+#  if XXH_SIZE_OPT >= 1 || \
+      defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64)    || defined(_M_ARM) /* visual */
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+#ifndef XXH_NO_INLINE_HINTS
+#  if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__)  /* -O0, -fno-inline */
+#    define XXH_NO_INLINE_HINTS 1
+#  else
+#    define XXH_NO_INLINE_HINTS 0
+#  endif
+#endif
+
+#ifndef XXH3_INLINE_SECRET
+#  if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \
+     || !defined(XXH_INLINE_ALL)
+#    define XXH3_INLINE_SECRET 0
+#  else
+#    define XXH3_INLINE_SECRET 1
+#  endif
+#endif
+
+#ifndef XXH32_ENDJMP
+/* generally preferable for performance */
+#  define XXH32_ENDJMP 0
+#endif
+
+/*!
+ * @defgroup impl Implementation
+ * @{
+ */
+
+
+/* *************************************
+*  Includes & Memory related functions
+***************************************/
+#if defined(XXH_NO_STREAM)
+/* nothing */
+#elif defined(XXH_NO_STDLIB)
+
+/* When requesting to disable any mention of stdlib,
+ * the library loses the ability to invoked malloc / free.
+ * In practice, it means that functions like `XXH*_createState()`
+ * will always fail, and return NULL.
+ * This flag is useful in situations where
+ * xxhash.h is integrated into some kernel, embedded or limited environment
+ * without access to dynamic allocation.
+ */
+
+static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
+static void XXH_free(void* p) { (void)p; }
+
+#else
+
+/*
+ * Modify the local functions below should you wish to use
+ * different memory routines for malloc() and free()
+ */
+#include 
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than malloc().
+ */
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than free().
+ */
+static void XXH_free(void* p) { free(p); }
+
+#endif  /* XXH_NO_STDLIB */
+
+#include 
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than memcpy().
+ */
+static void* XXH_memcpy(void* dest, const void* src, size_t size)
+{
+    return memcpy(dest,src,size);
+}
+
+#include    /* ULLONG_MAX */
+
+
+/* *************************************
+*  Compiler Specific Options
+***************************************/
+#ifdef _MSC_VER /* Visual Studio warning fix */
+#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+#endif
+
+#if XXH_NO_INLINE_HINTS  /* disable inlining hints */
+#  if defined(__GNUC__) || defined(__clang__)
+#    define XXH_FORCE_INLINE static __attribute__((__unused__))
+#  else
+#    define XXH_FORCE_INLINE static
+#  endif
+#  define XXH_NO_INLINE static
+/* enable inlining hints */
+#elif defined(__GNUC__) || defined(__clang__)
+#  define XXH_FORCE_INLINE static __inline__ __attribute__((__always_inline__, __unused__))
+#  define XXH_NO_INLINE static __attribute__((__noinline__))
+#elif defined(_MSC_VER)  /* Visual Studio */
+#  define XXH_FORCE_INLINE static __forceinline
+#  define XXH_NO_INLINE static __declspec(noinline)
+#elif defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
+#  define XXH_FORCE_INLINE static inline
+#  define XXH_NO_INLINE static
+#else
+#  define XXH_FORCE_INLINE static
+#  define XXH_NO_INLINE static
+#endif
+
+#if XXH3_INLINE_SECRET
+#  define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
+#else
+#  define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
+#endif
+
+
+/* *************************************
+*  Debug
+***************************************/
+/*!
+ * @ingroup tuning
+ * @def XXH_DEBUGLEVEL
+ * @brief Sets the debugging level.
+ *
+ * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
+ * compiler's command line options. The value must be a number.
+ */
+#ifndef XXH_DEBUGLEVEL
+#  ifdef DEBUGLEVEL /* backwards compat */
+#    define XXH_DEBUGLEVEL DEBUGLEVEL
+#  else
+#    define XXH_DEBUGLEVEL 0
+#  endif
+#endif
+
+#if (XXH_DEBUGLEVEL>=1)
+#  include    /* note: can still be disabled with NDEBUG */
+#  define XXH_ASSERT(c)   assert(c)
+#else
+#  if defined(__INTEL_COMPILER)
+#    define XXH_ASSERT(c)   XXH_ASSUME((unsigned char) (c))
+#  else
+#    define XXH_ASSERT(c)   XXH_ASSUME(c)
+#  endif
+#endif
+
+/* note: use after variable declarations */
+#ifndef XXH_STATIC_ASSERT
+#  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
+#  elif defined(__cplusplus) && (__cplusplus >= 201103L)            /* C++11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
+#  else
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
+#  endif
+#  define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
+#endif
+
+/*!
+ * @internal
+ * @def XXH_COMPILER_GUARD(var)
+ * @brief Used to prevent unwanted optimizations for @p var.
+ *
+ * It uses an empty GCC inline assembly statement with a register constraint
+ * which forces @p var into a general purpose register (eg eax, ebx, ecx
+ * on x86) and marks it as modified.
+ *
+ * This is used in a few places to avoid unwanted autovectorization (e.g.
+ * XXH32_round()). All vectorization we want is explicit via intrinsics,
+ * and _usually_ isn't wanted elsewhere.
+ *
+ * We also use it to prevent unwanted constant folding for AArch64 in
+ * XXH3_initCustomSecret_scalar().
+ */
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
+#else
+#  define XXH_COMPILER_GUARD(var) ((void)0)
+#endif
+
+/* Specifically for NEON vectors which use the "w" constraint, on
+ * Clang. */
+#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
+#endif
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   ifdef _AIX
+#     include 
+#   else
+#     include 
+#   endif
+    typedef uint8_t xxh_u8;
+#else
+    typedef unsigned char xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+#ifdef XXH_OLD_NAMES
+#  warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
+#  define BYTE xxh_u8
+#  define U8   xxh_u8
+#  define U32  xxh_u32
+#endif
+
+/* ***   Memory access   *** */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_read32(const void* ptr)
+ * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit native endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readBE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit big endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
+ * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
+ * always @ref XXH_alignment::XXH_unaligned.
+ *
+ * @param ptr The pointer to read from.
+ * @param align Whether @p ptr is aligned.
+ * @pre
+ *   If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
+ *   aligned.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE32 and XXH_readBE32.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/*
+ * Force direct memory access. Only works on CPU which support unaligned memory
+ * access in hardware.
+ */
+static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; } __attribute__((__packed__)) unalign;
+#endif
+static xxh_u32 XXH_read32(const void* ptr)
+{
+    typedef __attribute__((__aligned__(1))) xxh_u32 xxh_unalign32;
+    return *((const xxh_unalign32*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ***   Endianness   *** */
+
+/*!
+ * @ingroup tuning
+ * @def XXH_CPU_LITTLE_ENDIAN
+ * @brief Whether the target is little endian.
+ *
+ * Defined to 1 if the target is little endian, or 0 if it is big endian.
+ * It can be defined externally, for example on the compiler command line.
+ *
+ * If it is not defined,
+ * a runtime check (which is usually constant folded) is used instead.
+ *
+ * @note
+ *   This is not necessarily defined to an integer constant.
+ *
+ * @see XXH_isLittleEndian() for the runtime check.
+ */
+#ifndef XXH_CPU_LITTLE_ENDIAN
+/*
+ * Try to detect endianness automatically, to avoid the nonstandard behavior
+ * in `XXH_isLittleEndian()`
+ */
+#  if defined(_WIN32) /* Windows is always little endian */ \
+     || defined(__LITTLE_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 1
+#  elif defined(__BIG_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 0
+#  else
+/*!
+ * @internal
+ * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
+ *
+ * Most compilers will constant fold this.
+ */
+static int XXH_isLittleEndian(void)
+{
+    /*
+     * Portable and well-defined behavior.
+     * Don't use static: it is detrimental to performance.
+     */
+    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
+    return one.c[0];
+}
+#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
+#  endif
+#endif
+
+
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifdef __has_builtin
+#  define XXH_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define XXH_HAS_BUILTIN(x) 0
+#endif
+
+
+
+/*
+ * C23 and future versions have standard "unreachable()".
+ * Once it has been implemented reliably we can add it as an
+ * additional case:
+ *
+ * ```
+ * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
+ * #  include 
+ * #  ifdef unreachable
+ * #    define XXH_UNREACHABLE() unreachable()
+ * #  endif
+ * #endif
+ * ```
+ *
+ * Note C++23 also has std::unreachable() which can be detected
+ * as follows:
+ * ```
+ * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
+ * #  include 
+ * #  define XXH_UNREACHABLE() std::unreachable()
+ * #endif
+ * ```
+ * NB: `__cpp_lib_unreachable` is defined in the `` header.
+ * We don't use that as including `` in `extern "C"` blocks
+ * doesn't work on GCC12
+ */
+
+#if XXH_HAS_BUILTIN(__builtin_unreachable)
+#  define XXH_UNREACHABLE() __builtin_unreachable()
+
+#elif defined(_MSC_VER)
+#  define XXH_UNREACHABLE() __assume(0)
+
+#else
+#  define XXH_UNREACHABLE()
+#endif
+
+#if XXH_HAS_BUILTIN(__builtin_assume)
+#  define XXH_ASSUME(c) __builtin_assume(c)
+#else
+#  define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
+#endif
+
+/*!
+ * @internal
+ * @def XXH_rotl32(x,r)
+ * @brief 32-bit rotate left.
+ *
+ * @param x The 32-bit integer to be rotated.
+ * @param r The number of bits to rotate.
+ * @pre
+ *   @p r > 0 && @p r < 32
+ * @note
+ *   @p x and @p r may be evaluated multiple times.
+ * @return The rotated result.
+ */
+#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
+                               && XXH_HAS_BUILTIN(__builtin_rotateleft64)
+#  define XXH_rotl32 __builtin_rotateleft32
+#  define XXH_rotl64 __builtin_rotateleft64
+/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
+#elif defined(_MSC_VER)
+#  define XXH_rotl32(x,r) _rotl(x,r)
+#  define XXH_rotl64(x,r) _rotl64(x,r)
+#else
+#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
+#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#endif
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_swap32(xxh_u32 x)
+ * @brief A 32-bit byteswap.
+ *
+ * @param x The 32-bit integer to byteswap.
+ * @return @p x, byteswapped.
+ */
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap32 _byteswap_ulong
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap32 __builtin_bswap32
+#else
+static xxh_u32 XXH_swap32 (xxh_u32 x)
+{
+    return  ((x << 24) & 0xff000000 ) |
+            ((x <<  8) & 0x00ff0000 ) |
+            ((x >>  8) & 0x0000ff00 ) |
+            ((x >> 24) & 0x000000ff );
+}
+#endif
+
+
+/* ***************************
+*  Memory reads
+*****************************/
+
+/*!
+ * @internal
+ * @brief Enum to indicate whether a pointer is aligned.
+ */
+typedef enum {
+    XXH_aligned,  /*!< Aligned */
+    XXH_unaligned /*!< Possibly unaligned */
+} XXH_alignment;
+
+/*
+ * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
+ *
+ * This is ideal for older compilers which don't inline memcpy.
+ */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u32)bytePtr[1] << 8)
+         | ((xxh_u32)bytePtr[2] << 16)
+         | ((xxh_u32)bytePtr[3] << 24);
+}
+
+XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[3]
+         | ((xxh_u32)bytePtr[2] << 8)
+         | ((xxh_u32)bytePtr[1] << 16)
+         | ((xxh_u32)bytePtr[0] << 24);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+}
+
+static xxh_u32 XXH_readBE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u32
+XXH_readLE32_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned) {
+        return XXH_readLE32(ptr);
+    } else {
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+    }
+}
+
+
+/* *************************************
+*  Misc
+***************************************/
+/*! @ingroup public */
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+*  32-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @defgroup XXH32_impl XXH32 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH32 implementation.
+ * @{
+ */
+ /* #define instead of static const, to be used as initializers */
+#define XXH_PRIME32_1  0x9E3779B1U  /*!< 0b10011110001101110111100110110001 */
+#define XXH_PRIME32_2  0x85EBCA77U  /*!< 0b10000101111010111100101001110111 */
+#define XXH_PRIME32_3  0xC2B2AE3DU  /*!< 0b11000010101100101010111000111101 */
+#define XXH_PRIME32_4  0x27D4EB2FU  /*!< 0b00100111110101001110101100101111 */
+#define XXH_PRIME32_5  0x165667B1U  /*!< 0b00010110010101100110011110110001 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME32_1 XXH_PRIME32_1
+#  define PRIME32_2 XXH_PRIME32_2
+#  define PRIME32_3 XXH_PRIME32_3
+#  define PRIME32_4 XXH_PRIME32_4
+#  define PRIME32_5 XXH_PRIME32_5
+#endif
+
+/*!
+ * @internal
+ * @brief Normal stripe processing routine.
+ *
+ * This shuffles the bits so that any bit from @p input impacts several bits in
+ * @p acc.
+ *
+ * @param acc The accumulator lane.
+ * @param input The stripe of input to mix.
+ * @return The mixed accumulator lane.
+ */
+static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+{
+    acc += input * XXH_PRIME32_2;
+    acc  = XXH_rotl32(acc, 13);
+    acc *= XXH_PRIME32_1;
+#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * UGLY HACK:
+     * A compiler fence is used to prevent GCC and Clang from
+     * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling SSE4.1.
+     *
+     * The reason we want to avoid vectorization is because despite working on
+     * 4 integers at a time, there are multiple factors slowing XXH32 down on
+     * SSE4:
+     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
+     *   newer chips!) making it slightly slower to multiply four integers at
+     *   once compared to four integers independently. Even when pmulld was
+     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
+     *   just to multiply unless doing a long operation.
+     *
+     * - Four instructions are required to rotate,
+     *      movqda tmp,  v // not required with VEX encoding
+     *      pslld  tmp, 13 // tmp <<= 13
+     *      psrld  v,   19 // x >>= 19
+     *      por    v,  tmp // x |= tmp
+     *   compared to one for scalar:
+     *      roll   v, 13    // reliably fast across the board
+     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
+     *
+     * - Instruction level parallelism is actually more beneficial here because
+     *   the SIMD actually serializes this operation: While v1 is rotating, v2
+     *   can load data, while v3 can multiply. SSE forces them to operate
+     *   together.
+     *
+     * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
+     * the loop. NEON is only faster on the A53, and with the newer cores, it is less
+     * than half the speed.
+     *
+     * Additionally, this is used on WASM SIMD128 because it JITs to the same
+     * SIMD instructions and has the same issue.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+/*!
+ * @internal
+ * @brief Mixes all bits to finalize the hash.
+ *
+ * The final mix ensures that all input bits have a chance to impact any bit in
+ * the output digest, resulting in an unbiased distribution.
+ *
+ * @param hash The hash to avalanche.
+ * @return The avalanched hash.
+ */
+static xxh_u32 XXH32_avalanche(xxh_u32 hash)
+{
+    hash ^= hash >> 15;
+    hash *= XXH_PRIME32_2;
+    hash ^= hash >> 13;
+    hash *= XXH_PRIME32_3;
+    hash ^= hash >> 16;
+    return hash;
+}
+
+#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-15 bytes of @p ptr.
+ *
+ * There may be up to 15 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 16.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash.
+ * @see XXH64_finalize().
+ */
+static XXH_PUREF xxh_u32
+XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do {                             \
+    hash += (*ptr++) * XXH_PRIME32_5;                 \
+    hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1;      \
+} while (0)
+
+#define XXH_PROCESS4 do {                             \
+    hash += XXH_get32bits(ptr) * XXH_PRIME32_3;       \
+    ptr += 4;                                         \
+    hash  = XXH_rotl32(hash, 17) * XXH_PRIME32_4;     \
+} while (0)
+
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+
+    /* Compact rerolled version; generally faster */
+    if (!XXH32_ENDJMP) {
+        len &= 15;
+        while (len >= 4) {
+            XXH_PROCESS4;
+            len -= 4;
+        }
+        while (len > 0) {
+            XXH_PROCESS1;
+            --len;
+        }
+        return XXH32_avalanche(hash);
+    } else {
+         switch(len&15) /* or switch(bEnd - p) */ {
+           case 12:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 8:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 4:       XXH_PROCESS4;
+                         return XXH32_avalanche(hash);
+
+           case 13:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 9:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 5:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 14:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 10:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 6:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 15:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 11:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 7:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 3:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 2:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 1:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 0:       return XXH32_avalanche(hash);
+        }
+        XXH_ASSERT(0);
+        return hash;   /* reaching this point is deemed impossible */
+    }
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1 XXH_PROCESS1
+#  define PROCESS4 XXH_PROCESS4
+#else
+#  undef XXH_PROCESS1
+#  undef XXH_PROCESS4
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH32().
+ *
+ * @param input , len , seed Directly passed from @ref XXH32().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+    xxh_u32 h32;
+
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=16) {
+        const xxh_u8* const bEnd = input + len;
+        const xxh_u8* const limit = bEnd - 15;
+        xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+        xxh_u32 v2 = seed + XXH_PRIME32_2;
+        xxh_u32 v3 = seed + 0;
+        xxh_u32 v4 = seed - XXH_PRIME32_1;
+
+        do {
+            v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
+            v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
+            v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
+            v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
+        } while (input < limit);
+
+        h32 = XXH_rotl32(v1, 1)  + XXH_rotl32(v2, 7)
+            + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
+    } else {
+        h32  = seed + XXH_PRIME32_5;
+    }
+
+    h32 += (xxh_u32)len;
+
+    return XXH32_finalize(h32, input, len&15, align);
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH32_state_t state;
+    XXH32_reset(&state, seed);
+    XXH32_update(&state, (const xxh_u8*)input, len);
+    return XXH32_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
+            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+#endif
+}
+
+
+
+/*******   Hash streaming   *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+{
+    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+{
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    statePtr->v[1] = seed + XXH_PRIME32_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME32_1;
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        state->total_len_32 += (XXH32_hash_t)len;
+        state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
+
+        if (state->memsize + len < 16)  {   /* fill in tmp buffer */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
+            state->memsize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
+
+        if (state->memsize) {   /* some data left from previous update */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
+            {   const xxh_u32* p32 = state->mem32;
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32));
+            }
+            p += 16-state->memsize;
+            state->memsize = 0;
+        }
+
+        if (p <= bEnd-16) {
+            const xxh_u8* const limit = bEnd - 16;
+
+            do {
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4;
+            } while (p<=limit);
+
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
+{
+    xxh_u32 h32;
+
+    if (state->large_len) {
+        h32 = XXH_rotl32(state->v[0], 1)
+            + XXH_rotl32(state->v[1], 7)
+            + XXH_rotl32(state->v[2], 12)
+            + XXH_rotl32(state->v[3], 18);
+    } else {
+        h32 = state->v[2] /* == seed */ + XXH_PRIME32_5;
+    }
+
+    h32 += state->total_len_32;
+
+    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/*******   Canonical representation   *******/
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
+    XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+{
+    return XXH_readBE32(src);
+}
+
+
+#ifndef XXH_NO_LONG_LONG
+
+/* *******************************************************************
+*  64-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @ingroup impl
+ * @{
+ */
+/*******   Memory access   *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+#ifdef XXH_OLD_NAMES
+#  define U64 xxh_u64
+#endif
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE64 and XXH_readBE64.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    return *(const xxh_u64*) memPtr;
+}
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((__packed__)) unalign64;
+#endif
+static xxh_u64 XXH_read64(const void* ptr)
+{
+    typedef __attribute__((__aligned__(1))) xxh_u64 xxh_unalign64;
+    return *((const xxh_unalign64*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap64 _byteswap_uint64
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap64 __builtin_bswap64
+#else
+static xxh_u64 XXH_swap64(xxh_u64 x)
+{
+    return  ((x << 56) & 0xff00000000000000ULL) |
+            ((x << 40) & 0x00ff000000000000ULL) |
+            ((x << 24) & 0x0000ff0000000000ULL) |
+            ((x << 8)  & 0x000000ff00000000ULL) |
+            ((x >> 8)  & 0x00000000ff000000ULL) |
+            ((x >> 24) & 0x0000000000ff0000ULL) |
+            ((x >> 40) & 0x000000000000ff00ULL) |
+            ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+
+/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u64)bytePtr[1] << 8)
+         | ((xxh_u64)bytePtr[2] << 16)
+         | ((xxh_u64)bytePtr[3] << 24)
+         | ((xxh_u64)bytePtr[4] << 32)
+         | ((xxh_u64)bytePtr[5] << 40)
+         | ((xxh_u64)bytePtr[6] << 48)
+         | ((xxh_u64)bytePtr[7] << 56);
+}
+
+XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[7]
+         | ((xxh_u64)bytePtr[6] << 8)
+         | ((xxh_u64)bytePtr[5] << 16)
+         | ((xxh_u64)bytePtr[4] << 24)
+         | ((xxh_u64)bytePtr[3] << 32)
+         | ((xxh_u64)bytePtr[2] << 40)
+         | ((xxh_u64)bytePtr[1] << 48)
+         | ((xxh_u64)bytePtr[0] << 56);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+}
+
+static xxh_u64 XXH_readBE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u64
+XXH_readLE64_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned)
+        return XXH_readLE64(ptr);
+    else
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+}
+
+
+/*******   xxh64   *******/
+/*!
+ * @}
+ * @defgroup XXH64_impl XXH64 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH64 implementation.
+ * @{
+ */
+/* #define rather that static const, to be used as initializers */
+#define XXH_PRIME64_1  0x9E3779B185EBCA87ULL  /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
+#define XXH_PRIME64_2  0xC2B2AE3D27D4EB4FULL  /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
+#define XXH_PRIME64_3  0x165667B19E3779F9ULL  /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
+#define XXH_PRIME64_4  0x85EBCA77C2B2AE63ULL  /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
+#define XXH_PRIME64_5  0x27D4EB2F165667C5ULL  /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME64_1 XXH_PRIME64_1
+#  define PRIME64_2 XXH_PRIME64_2
+#  define PRIME64_3 XXH_PRIME64_3
+#  define PRIME64_4 XXH_PRIME64_4
+#  define PRIME64_5 XXH_PRIME64_5
+#endif
+
+/*! @copydoc XXH32_round */
+static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+{
+    acc += input * XXH_PRIME64_2;
+    acc  = XXH_rotl64(acc, 31);
+    acc *= XXH_PRIME64_1;
+#if (defined(__AVX512F__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * DISABLE AUTOVECTORIZATION:
+     * A compiler fence is used to prevent GCC and Clang from
+     * autovectorizing the XXH64 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling AVX512.
+     *
+     * Autovectorization of XXH64 tends to be detrimental,
+     * though the exact outcome may change depending on exact cpu and compiler version.
+     * For information, it has been reported as detrimental for Skylake-X,
+     * but possibly beneficial for Zen4.
+     *
+     * The default is to disable auto-vectorization,
+     * but you can select to enable it instead using `XXH_ENABLE_AUTOVECTORIZE` build variable.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+{
+    val  = XXH64_round(0, val);
+    acc ^= val;
+    acc  = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
+    return acc;
+}
+
+/*! @copydoc XXH32_avalanche */
+static xxh_u64 XXH64_avalanche(xxh_u64 hash)
+{
+    hash ^= hash >> 33;
+    hash *= XXH_PRIME64_2;
+    hash ^= hash >> 29;
+    hash *= XXH_PRIME64_3;
+    hash ^= hash >> 32;
+    return hash;
+}
+
+
+#define XXH_get64bits(p) XXH_readLE64_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-31 bytes of @p ptr.
+ *
+ * There may be up to 31 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 32.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash
+ * @see XXH32_finalize().
+ */
+static XXH_PUREF xxh_u64
+XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+    len &= 31;
+    while (len >= 8) {
+        xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
+        ptr += 8;
+        hash ^= k1;
+        hash  = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+        len -= 8;
+    }
+    if (len >= 4) {
+        hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        ptr += 4;
+        hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        len -= 4;
+    }
+    while (len > 0) {
+        hash ^= (*ptr++) * XXH_PRIME64_5;
+        hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
+        --len;
+    }
+    return  XXH64_avalanche(hash);
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1_64 XXH_PROCESS1_64
+#  define PROCESS4_64 XXH_PROCESS4_64
+#  define PROCESS8_64 XXH_PROCESS8_64
+#else
+#  undef XXH_PROCESS1_64
+#  undef XXH_PROCESS4_64
+#  undef XXH_PROCESS8_64
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH64().
+ *
+ * @param input , len , seed Directly passed from @ref XXH64().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u64
+XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+{
+    xxh_u64 h64;
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=32) {
+        const xxh_u8* const bEnd = input + len;
+        const xxh_u8* const limit = bEnd - 31;
+        xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+        xxh_u64 v2 = seed + XXH_PRIME64_2;
+        xxh_u64 v3 = seed + 0;
+        xxh_u64 v4 = seed - XXH_PRIME64_1;
+
+        do {
+            v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
+            v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
+            v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
+            v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
+        } while (input= 2
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH64_state_t state;
+    XXH64_reset(&state, seed);
+    XXH64_update(&state, (const xxh_u8*)input, len);
+    return XXH64_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
+            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+
+#endif
+}
+
+/*******   Hash Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH64_family*/
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
+{
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    statePtr->v[1] = seed + XXH_PRIME64_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME64_1;
+    return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        state->total_len += len;
+
+        if (state->memsize + len < 32) {  /* fill in tmp buffer */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
+            state->memsize += (xxh_u32)len;
+            return XXH_OK;
+        }
+
+        if (state->memsize) {   /* tmp buffer is full */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
+            state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0));
+            state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1));
+            state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2));
+            state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3));
+            p += 32 - state->memsize;
+            state->memsize = 0;
+        }
+
+        if (p+32 <= bEnd) {
+            const xxh_u8* const limit = bEnd - 32;
+
+            do {
+                state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8;
+                state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8;
+                state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8;
+                state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8;
+            } while (p<=limit);
+
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
+{
+    xxh_u64 h64;
+
+    if (state->total_len >= 32) {
+        h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18);
+        h64 = XXH64_mergeRound(h64, state->v[0]);
+        h64 = XXH64_mergeRound(h64, state->v[1]);
+        h64 = XXH64_mergeRound(h64, state->v[2]);
+        h64 = XXH64_mergeRound(h64, state->v[3]);
+    } else {
+        h64  = state->v[2] /*seed*/ + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) state->total_len;
+
+    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/******* Canonical representation   *******/
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
+    XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
+{
+    return XXH_readBE64(src);
+}
+
+#ifndef XXH_NO_XXH3
+
+/* *********************************************************************
+*  XXH3
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+/*!
+ * @}
+ * @defgroup XXH3_impl XXH3 implementation
+ * @ingroup impl
+ * @{
+ */
+
+/* ===   Compiler specifics   === */
+
+#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
+#  define XXH_RESTRICT   /* disable */
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
+#  define XXH_RESTRICT   restrict
+#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
+   || (defined (__clang__)) \
+   || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
+   || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
+/*
+ * There are a LOT more compilers that recognize __restrict but this
+ * covers the major ones.
+ */
+#  define XXH_RESTRICT   __restrict
+#else
+#  define XXH_RESTRICT   /* disable */
+#endif
+
+#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
+  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
+  || defined(__clang__)
+#    define XXH_likely(x) __builtin_expect(x, 1)
+#    define XXH_unlikely(x) __builtin_expect(x, 0)
+#else
+#    define XXH_likely(x) (x)
+#    define XXH_unlikely(x) (x)
+#endif
+
+#ifndef XXH_HAS_INCLUDE
+#  ifdef __has_include
+/*
+ * Not defined as XXH_HAS_INCLUDE(x) (function-like) because
+ * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion)
+ */
+#    define XXH_HAS_INCLUDE __has_include
+#  else
+#    define XXH_HAS_INCLUDE(x) 0
+#  endif
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_FEATURE_SVE)
+#    include 
+#  endif
+#  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
+   || (defined(_M_ARM) && _M_ARM >= 7) \
+   || defined(_M_ARM64) || defined(_M_ARM64EC) \
+   || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE()) /* WASM SIMD128 via SIMDe */
+#    define inline __inline__  /* circumvent a clang bug */
+#    include 
+#    undef inline
+#  elif defined(__AVX2__)
+#    include 
+#  elif defined(__SSE2__)
+#    include 
+#  endif
+#endif
+
+#if defined(_MSC_VER)
+#  include 
+#endif
+
+/*
+ * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
+ * remaining a true 64-bit/128-bit hash function.
+ *
+ * This is done by prioritizing a subset of 64-bit operations that can be
+ * emulated without too many steps on the average 32-bit machine.
+ *
+ * For example, these two lines seem similar, and run equally fast on 64-bit:
+ *
+ *   xxh_u64 x;
+ *   x ^= (x >> 47); // good
+ *   x ^= (x >> 13); // bad
+ *
+ * However, to a 32-bit machine, there is a major difference.
+ *
+ * x ^= (x >> 47) looks like this:
+ *
+ *   x.lo ^= (x.hi >> (47 - 32));
+ *
+ * while x ^= (x >> 13) looks like this:
+ *
+ *   // note: funnel shifts are not usually cheap.
+ *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
+ *   x.hi ^= (x.hi >> 13);
+ *
+ * The first one is significantly faster than the second, simply because the
+ * shift is larger than 32. This means:
+ *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
+ *    32 bits in the shift.
+ *  - The shift result will always fit in the lower 32 bits, and therefore,
+ *    we can ignore the upper 32 bits in the xor.
+ *
+ * Thanks to this optimization, XXH3 only requires these features to be efficient:
+ *
+ *  - Usable unaligned access
+ *  - A 32-bit or 64-bit ALU
+ *      - If 32-bit, a decent ADC instruction
+ *  - A 32 or 64-bit multiply with a 64-bit result
+ *  - For the 128-bit variant, a decent byteswap helps short inputs.
+ *
+ * The first two are already required by XXH32, and almost all 32-bit and 64-bit
+ * platforms which can run XXH32 can run XXH3 efficiently.
+ *
+ * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
+ * notable exception.
+ *
+ * First of all, Thumb-1 lacks support for the UMULL instruction which
+ * performs the important long multiply. This means numerous __aeabi_lmul
+ * calls.
+ *
+ * Second of all, the 8 functional registers are just not enough.
+ * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
+ * Lo registers, and this shuffling results in thousands more MOVs than A32.
+ *
+ * A32 and T32 don't have this limitation. They can access all 14 registers,
+ * do a 32->64 multiply with UMULL, and the flexible operand allowing free
+ * shifts is helpful, too.
+ *
+ * Therefore, we do a quick sanity check.
+ *
+ * If compiling Thumb-1 for a target which supports ARM instructions, we will
+ * emit a warning, as it is not a "sane" platform to compile for.
+ *
+ * Usually, if this happens, it is because of an accident and you probably need
+ * to specify -march, as you likely meant to compile for a newer architecture.
+ *
+ * Credit: large sections of the vectorial and asm source code paths
+ *         have been contributed by @easyaspi314
+ */
+#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
+#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
+#endif
+
+/* ==========================================
+ * Vectorization detection
+ * ========================================== */
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @ingroup tuning
+ * @brief Overrides the vectorization implementation chosen for XXH3.
+ *
+ * Can be defined to 0 to disable SIMD or any of the values mentioned in
+ * @ref XXH_VECTOR_TYPE.
+ *
+ * If this is not defined, it uses predefined macros to determine the best
+ * implementation.
+ */
+#  define XXH_VECTOR XXH_SCALAR
+/*!
+ * @ingroup tuning
+ * @brief Possible values for @ref XXH_VECTOR.
+ *
+ * Note that these are actually implemented as macros.
+ *
+ * If this is not defined, it is detected automatically.
+ * internal macro XXH_X86DISPATCH overrides this.
+ */
+enum XXH_VECTOR_TYPE /* fake enum */ {
+    XXH_SCALAR = 0,  /*!< Portable scalar version */
+    XXH_SSE2   = 1,  /*!<
+                      * SSE2 for Pentium 4, Opteron, all x86_64.
+                      *
+                      * @note SSE2 is also guaranteed on Windows 10, macOS, and
+                      * Android x86.
+                      */
+    XXH_AVX2   = 2,  /*!< AVX2 for Haswell and Bulldozer */
+    XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
+    XXH_NEON   = 4,  /*!<
+                       * NEON for most ARMv7-A, all AArch64, and WASM SIMD128
+                       * via the SIMDeverywhere polyfill provided with the
+                       * Emscripten SDK.
+                       */
+    XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+    XXH_SVE    = 6,  /*!< SVE for some ARMv8-A and ARMv9-A */
+};
+/*!
+ * @ingroup tuning
+ * @brief Selects the minimum alignment for XXH3's accumulators.
+ *
+ * When using SIMD, this should match the alignment required for said vector
+ * type, so, for example, 32 for AVX2.
+ *
+ * Default: Auto detected.
+ */
+#  define XXH_ACC_ALIGN 8
+#endif
+
+/* Actual definition */
+#ifndef XXH_DOXYGEN
+#  define XXH_SCALAR 0
+#  define XXH_SSE2   1
+#  define XXH_AVX2   2
+#  define XXH_AVX512 3
+#  define XXH_NEON   4
+#  define XXH_VSX    5
+#  define XXH_SVE    6
+#endif
+
+#ifndef XXH_VECTOR    /* can be defined on command line */
+#  if defined(__ARM_FEATURE_SVE)
+#    define XXH_VECTOR XXH_SVE
+#  elif ( \
+        defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
+     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
+     || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE()) /* wasm simd128 via SIMDe */ \
+   ) && ( \
+        defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
+    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
+   )
+#    define XXH_VECTOR XXH_NEON
+#  elif defined(__AVX512F__)
+#    define XXH_VECTOR XXH_AVX512
+#  elif defined(__AVX2__)
+#    define XXH_VECTOR XXH_AVX2
+#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
+#    define XXH_VECTOR XXH_SSE2
+#  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
+     || (defined(__s390x__) && defined(__VEC__)) \
+     && defined(__GNUC__) /* TODO: IBM XL */
+#    define XXH_VECTOR XXH_VSX
+#  else
+#    define XXH_VECTOR XXH_SCALAR
+#  endif
+#endif
+
+/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
+#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
+#  ifdef _MSC_VER
+#    pragma warning(once : 4606)
+#  else
+#    warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
+#  endif
+#  undef XXH_VECTOR
+#  define XXH_VECTOR XXH_SCALAR
+#endif
+
+/*
+ * Controls the alignment of the accumulator,
+ * for compatibility with aligned vector loads, which are usually faster.
+ */
+#ifndef XXH_ACC_ALIGN
+#  if defined(XXH_X86DISPATCH)
+#     define XXH_ACC_ALIGN 64  /* for compatibility with avx512 */
+#  elif XXH_VECTOR == XXH_SCALAR  /* scalar */
+#     define XXH_ACC_ALIGN 8
+#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
+#     define XXH_ACC_ALIGN 32
+#  elif XXH_VECTOR == XXH_NEON  /* neon */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_VSX   /* vsx */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX512  /* avx512 */
+#     define XXH_ACC_ALIGN 64
+#  elif XXH_VECTOR == XXH_SVE   /* sve */
+#     define XXH_ACC_ALIGN 64
+#  endif
+#endif
+
+#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
+    || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#elif XXH_VECTOR == XXH_SVE
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#else
+#  define XXH_SEC_ALIGN 8
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_ALIASING __attribute__((__may_alias__))
+#else
+#  define XXH_ALIASING /* nothing */
+#endif
+
+/*
+ * UGLY HACK:
+ * GCC usually generates the best code with -O3 for xxHash.
+ *
+ * However, when targeting AVX2, it is overzealous in its unrolling resulting
+ * in code roughly 3/4 the speed of Clang.
+ *
+ * There are other issues, such as GCC splitting _mm256_loadu_si256 into
+ * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
+ * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
+ *
+ * That is why when compiling the AVX2 version, it is recommended to use either
+ *   -O2 -mavx2 -march=haswell
+ * or
+ *   -O2 -mavx2 -mno-avx256-split-unaligned-load
+ * for decent performance, or to use Clang instead.
+ *
+ * Fortunately, we can control the first one with a pragma that forces GCC into
+ * -O2, but the other one we can't control without "failed to inline always
+ * inline function due to target mismatch" warnings.
+ */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+#  pragma GCC push_options
+#  pragma GCC optimize("-O2")
+#endif
+
+#if XXH_VECTOR == XXH_NEON
+
+/*
+ * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
+ * optimizes out the entire hashLong loop because of the aliasing violation.
+ *
+ * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
+ * so the only option is to mark it as aliasing.
+ */
+typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
+
+/*!
+ * @internal
+ * @brief `vld1q_u64` but faster and alignment-safe.
+ *
+ * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
+ * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
+ *
+ * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
+ * prohibits load-store optimizations. Therefore, a direct dereference is used.
+ *
+ * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
+ * unaligned load.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
+{
+    return *(xxh_aliasing_uint64x2_t const *)ptr;
+}
+#else
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
+{
+    return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
+}
+#endif
+
+/*!
+ * @internal
+ * @brief `vmlal_u32` on low and high halves of a vector.
+ *
+ * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
+ * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
+ * with `vmlal_u32`.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* Inline assembly is the only way */
+    __asm__("umlal   %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
+    return acc;
+}
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* This intrinsic works as expected */
+    return vmlal_high_u32(acc, lhs, rhs);
+}
+#else
+/* Portable intrinsic versions */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
+}
+/*! @copydoc XXH_vmlal_low_u32
+ * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
+}
+#endif
+
+/*!
+ * @ingroup tuning
+ * @brief Controls the NEON to scalar ratio for XXH3
+ *
+ * This can be set to 2, 4, 6, or 8.
+ *
+ * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
+ *
+ * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those
+ * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
+ * bandwidth.
+ *
+ * This is even more noticeable on the more advanced cores like the Cortex-A76 which
+ * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
+ *
+ * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
+ * and 2 scalar lanes, which is chosen by default.
+ *
+ * This does not apply to Apple processors or 32-bit processors, which run better with
+ * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
+ *
+ * This change benefits CPUs with large micro-op buffers without negatively affecting
+ * most other CPUs:
+ *
+ *  | Chipset               | Dispatch type       | NEON only | 6:2 hybrid | Diff. |
+ *  |:----------------------|:--------------------|----------:|-----------:|------:|
+ *  | Snapdragon 730 (A76)  | 2 NEON/8 micro-ops  |  8.8 GB/s |  10.1 GB/s |  ~16% |
+ *  | Snapdragon 835 (A73)  | 2 NEON/3 micro-ops  |  5.1 GB/s |   5.3 GB/s |   ~5% |
+ *  | Marvell PXA1928 (A53) | In-order dual-issue |  1.9 GB/s |   1.9 GB/s |    0% |
+ *  | Apple M1              | 4 NEON/8 micro-ops  | 37.3 GB/s |  36.1 GB/s |  ~-3% |
+ *
+ * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
+ *
+ * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
+ * it effectively becomes worse 4.
+ *
+ * @see XXH3_accumulate_512_neon()
+ */
+# ifndef XXH3_NEON_LANES
+#  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
+   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
+#   define XXH3_NEON_LANES 6
+#  else
+#   define XXH3_NEON_LANES XXH_ACC_NB
+#  endif
+# endif
+#endif  /* XXH_VECTOR == XXH_NEON */
+
+/*
+ * VSX and Z Vector helpers.
+ *
+ * This is very messy, and any pull requests to clean this up are welcome.
+ *
+ * There are a lot of problems with supporting VSX and s390x, due to
+ * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
+ */
+#if XXH_VECTOR == XXH_VSX
+/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
+ * and `pixel`. This is a problem for obvious reasons.
+ *
+ * These keywords are unnecessary; the spec literally says they are
+ * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
+ * after including the header.
+ *
+ * We use pragma push_macro/pop_macro to keep the namespace clean. */
+#  pragma push_macro("bool")
+#  pragma push_macro("vector")
+#  pragma push_macro("pixel")
+/* silence potential macro redefined warnings */
+#  undef bool
+#  undef vector
+#  undef pixel
+
+#  if defined(__s390x__)
+#    include 
+#  else
+#    include 
+#  endif
+
+/* Restore the original macro values, if applicable. */
+#  pragma pop_macro("pixel")
+#  pragma pop_macro("vector")
+#  pragma pop_macro("bool")
+
+typedef __vector unsigned long long xxh_u64x2;
+typedef __vector unsigned char xxh_u8x16;
+typedef __vector unsigned xxh_u32x4;
+
+/*
+ * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue.
+ */
+typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING;
+
+# ifndef XXH_VSX_BE
+#  if defined(__BIG_ENDIAN__) \
+  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_VSX_BE 1
+#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
+#    warning "-maltivec=be is not recommended. Please use native endianness."
+#    define XXH_VSX_BE 1
+#  else
+#    define XXH_VSX_BE 0
+#  endif
+# endif /* !defined(XXH_VSX_BE) */
+
+# if XXH_VSX_BE
+#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
+#    define XXH_vec_revb vec_revb
+#  else
+/*!
+ * A polyfill for POWER9's vec_revb().
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
+{
+    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
+                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+    return vec_perm(val, val, vByteSwap);
+}
+#  endif
+# endif /* XXH_VSX_BE */
+
+/*!
+ * Performs an unaligned vector load and byte swaps it on big endian.
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
+{
+    xxh_u64x2 ret;
+    XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
+# if XXH_VSX_BE
+    ret = XXH_vec_revb(ret);
+# endif
+    return ret;
+}
+
+/*
+ * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
+ *
+ * These intrinsics weren't added until GCC 8, despite existing for a while,
+ * and they are endian dependent. Also, their meaning swap depending on version.
+ * */
+# if defined(__s390x__)
+ /* s390x is always big endian, no issue on this platform */
+#  define XXH_vec_mulo vec_mulo
+#  define XXH_vec_mule vec_mule
+# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
+/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+ /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
+#  define XXH_vec_mulo __builtin_altivec_vmulouw
+#  define XXH_vec_mule __builtin_altivec_vmuleuw
+# else
+/* gcc needs inline assembly */
+/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+# endif /* XXH_vec_mulo, XXH_vec_mule */
+#endif /* XXH_VECTOR == XXH_VSX */
+
+#if XXH_VECTOR == XXH_SVE
+#define ACCRND(acc, offset) \
+do { \
+    svuint64_t input_vec = svld1_u64(mask, xinput + offset);         \
+    svuint64_t secret_vec = svld1_u64(mask, xsecret + offset);       \
+    svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec);     \
+    svuint64_t swapped = svtbl_u64(input_vec, kSwap);                \
+    svuint64_t mixed_lo = svextw_u64_x(mask, mixed);                 \
+    svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32);            \
+    svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
+    acc = svadd_u64_x(mask, acc, mul);                               \
+} while (0)
+#endif /* XXH_VECTOR == XXH_SVE */
+
+/* prefetch
+ * can be disabled, by declaring XXH_NO_PREFETCH build macro */
+#if defined(XXH_NO_PREFETCH)
+#  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
+#else
+#  if XXH_SIZE_OPT >= 1
+#    define XXH_PREFETCH(ptr) (void)(ptr)
+#  elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
+#    include    /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
+#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+#  else
+#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
+#  endif
+#endif  /* XXH_NO_PREFETCH */
+
+
+/* ==========================================
+ * XXH3 default settings
+ * ========================================== */
+
+#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
+
+#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
+#  error "default keyset is not large enough"
+#endif
+
+/*! Pseudorandom secret taken directly from FARSH. */
+XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
+    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+
+static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL;  /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
+static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL;  /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */
+
+#ifdef XXH_OLD_NAMES
+#  define kSecret XXH3_kSecret
+#endif
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Calculates a 32-bit to 64-bit long multiply.
+ *
+ * Implemented as a macro.
+ *
+ * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
+ * need to (but it shouldn't need to anyways, it is about 7 instructions to do
+ * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
+ * use that instead of the normal method.
+ *
+ * If you are compiling for platforms like Thumb-1 and don't have a better option,
+ * you may also want to write your own long multiply routine here.
+ *
+ * @param x, y Numbers to be multiplied
+ * @return 64-bit product of the low 32 bits of @p x and @p y.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64(xxh_u64 x, xxh_u64 y)
+{
+   return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
+}
+#elif defined(_MSC_VER) && defined(_M_IX86)
+#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
+#else
+/*
+ * Downcast + upcast is usually better than masking on older compilers like
+ * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
+ *
+ * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
+ * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
+ */
+#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
+#endif
+
+/*!
+ * @brief Calculates a 64->128-bit long multiply.
+ *
+ * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
+ * version.
+ *
+ * @param lhs , rhs The 64-bit integers to be multiplied
+ * @return The 128-bit result represented in an @ref XXH128_hash_t.
+ */
+static XXH128_hash_t
+XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
+{
+    /*
+     * GCC/Clang __uint128_t method.
+     *
+     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
+     * This is usually the best way as it usually uses a native long 64-bit
+     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
+     *
+     * Usually.
+     *
+     * Despite being a 32-bit platform, Clang (and emscripten) define this type
+     * despite not having the arithmetic for it. This results in a laggy
+     * compiler builtin call which calculates a full 128-bit multiply.
+     * In that case it is best to use the portable one.
+     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
+     */
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
+    && defined(__SIZEOF_INT128__) \
+    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
+
+    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
+    XXH128_hash_t r128;
+    r128.low64  = (xxh_u64)(product);
+    r128.high64 = (xxh_u64)(product >> 64);
+    return r128;
+
+    /*
+     * MSVC for x64's _umul128 method.
+     *
+     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
+     *
+     * This compiles to single operand MUL on x64.
+     */
+#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(_umul128)
+#endif
+    xxh_u64 product_high;
+    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
+    XXH128_hash_t r128;
+    r128.low64  = product_low;
+    r128.high64 = product_high;
+    return r128;
+
+    /*
+     * MSVC for ARM64's __umulh method.
+     *
+     * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
+     */
+#elif defined(_M_ARM64) || defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(__umulh)
+#endif
+    XXH128_hash_t r128;
+    r128.low64  = lhs * rhs;
+    r128.high64 = __umulh(lhs, rhs);
+    return r128;
+
+#else
+    /*
+     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
+     *
+     * This is a fast and simple grade school multiply, which is shown below
+     * with base 10 arithmetic instead of base 0x100000000.
+     *
+     *           9 3 // D2 lhs = 93
+     *         x 7 5 // D2 rhs = 75
+     *     ----------
+     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
+     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
+     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
+     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
+     *     ---------
+     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
+     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
+     *     ---------
+     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
+     *
+     * The reasons for adding the products like this are:
+     *  1. It avoids manual carry tracking. Just like how
+     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
+     *     This avoids a lot of complexity.
+     *
+     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
+     *     instruction available in ARM's Digital Signal Processing extension
+     *     in 32-bit ARMv6 and later, which is shown below:
+     *
+     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
+     *         {
+     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
+     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
+     *             *RdHi = (xxh_u32)(product >> 32);
+     *         }
+     *
+     *     This instruction was designed for efficient long multiplication, and
+     *     allows this to be calculated in only 4 instructions at speeds
+     *     comparable to some 64-bit ALUs.
+     *
+     *  3. It isn't terrible on other platforms. Usually this will be a couple
+     *     of 32-bit ADD/ADCs.
+     */
+
+    /* First calculate all of the cross products. */
+    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
+    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
+    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
+    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);
+
+    /* Now add the products together. These will never overflow. */
+    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
+    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+    XXH128_hash_t r128;
+    r128.low64  = lower;
+    r128.high64 = upper;
+    return r128;
+#endif
+}
+
+/*!
+ * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
+ *
+ * The reason for the separate function is to prevent passing too many structs
+ * around by value. This will hopefully inline the multiply, but we don't force it.
+ *
+ * @param lhs , rhs The 64-bit integers to multiply
+ * @return The low 64 bits of the product XOR'd by the high 64 bits.
+ * @see XXH_mult64to128()
+ */
+static xxh_u64
+XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
+{
+    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
+    return product.low64 ^ product.high64;
+}
+
+/*! Seems to produce slightly better code on GCC for some reason. */
+XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+{
+    XXH_ASSERT(0 <= shift && shift < 64);
+    return v64 ^ (v64 >> shift);
+}
+
+/*
+ * This is a fast avalanche stage,
+ * suitable when input bits are already partially mixed
+ */
+static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
+{
+    h64 = XXH_xorshift64(h64, 37);
+    h64 *= PRIME_MX1;
+    h64 = XXH_xorshift64(h64, 32);
+    return h64;
+}
+
+/*
+ * This is a stronger avalanche,
+ * inspired by Pelle Evensen's rrmxmx
+ * preferable when input has not been previously mixed
+ */
+static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
+{
+    /* this mix is inspired by Pelle Evensen's rrmxmx */
+    h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
+    h64 *= PRIME_MX2;
+    h64 ^= (h64 >> 35) + len ;
+    h64 *= PRIME_MX2;
+    return XXH_xorshift64(h64, 28);
+}
+
+
+/* ==========================================
+ * Short keys
+ * ==========================================
+ * One of the shortcomings of XXH32 and XXH64 was that their performance was
+ * sub-optimal on short lengths. It used an iterative algorithm which strongly
+ * favored lengths that were a multiple of 4 or 8.
+ *
+ * Instead of iterating over individual inputs, we use a set of single shot
+ * functions which piece together a range of lengths and operate in constant time.
+ *
+ * Additionally, the number of multiplies has been significantly reduced. This
+ * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
+ *
+ * Depending on the platform, this may or may not be faster than XXH32, but it
+ * is almost guaranteed to be faster than XXH64.
+ */
+
+/*
+ * At very short lengths, there isn't enough input to fully hide secrets, or use
+ * the entire secret.
+ *
+ * There is also only a limited amount of mixing we can do before significantly
+ * impacting performance.
+ *
+ * Therefore, we use different sections of the secret and always mix two secret
+ * samples with an XOR. This should have no effect on performance on the
+ * seedless or withSeed variants because everything _should_ be constant folded
+ * by modern compilers.
+ *
+ * The XOR mixing hides individual parts of the secret and increases entropy.
+ *
+ * This adds an extra layer of strength for custom secrets.
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8  const c1 = input[0];
+        xxh_u8  const c2 = input[len >> 1];
+        xxh_u8  const c3 = input[len - 1];
+        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
+                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
+        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
+        return XXH64_avalanche(keyed);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input1 = XXH_readLE32(input);
+        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
+        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
+        xxh_u64 const keyed = input64 ^ bitflip;
+        return XXH3_rrmxmx(keyed, len);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
+        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
+        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
+        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
+        xxh_u64 const acc = len
+                          + XXH_swap64(input_lo) + input_hi
+                          + XXH3_mul128_fold64(input_lo, input_hi);
+        return XXH3_avalanche(acc);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(len <= 16);
+    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
+        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
+        return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
+    }
+}
+
+/*
+ * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
+ * multiplication by zero, affecting hashes of lengths 17 to 240.
+ *
+ * However, they are very unlikely.
+ *
+ * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
+ * unseeded non-cryptographic hashes, it does not attempt to defend itself
+ * against specially crafted inputs, only random inputs.
+ *
+ * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
+ * cancelling out the secret is taken an arbitrary number of times (addressed
+ * in XXH3_accumulate_512), this collision is very unlikely with random inputs
+ * and/or proper seeding:
+ *
+ * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
+ * function that is only called up to 16 times per hash with up to 240 bytes of
+ * input.
+ *
+ * This is not too bad for a non-cryptographic hash function, especially with
+ * only 64 bit outputs.
+ *
+ * The 128-bit variant (which trades some speed for strength) is NOT affected
+ * by this, although it is always a good idea to use a proper seed if you care
+ * about strength.
+ */
+XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
+                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
+{
+#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
+  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
+    /*
+     * UGLY HACK:
+     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
+     * slower code.
+     *
+     * By forcing seed64 into a register, we disrupt the cost model and
+     * cause it to scalarize. See `XXH32_round()`
+     *
+     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
+     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
+     * GCC 9.2, despite both emitting scalar code.
+     *
+     * GCC generates much better scalar code than Clang for the rest of XXH3,
+     * which is why finding a more optimal codepath is an interest.
+     */
+    XXH_COMPILER_GUARD(seed64);
+#endif
+    {   xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64 const input_hi = XXH_readLE64(input+8);
+        return XXH3_mul128_fold64(
+            input_lo ^ (XXH_readLE64(secret)   + seed64),
+            input_hi ^ (XXH_readLE64(secret+8) - seed64)
+        );
+    }
+}
+
+/* For mid range keys, XXH3 uses a Mum-hash variant. */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                     XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+#if XXH_SIZE_OPT >= 1
+        /* Smaller and cleaner, but slightly slower. */
+        unsigned int i = (unsigned int)(len - 1) / 32;
+        do {
+            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+        } while (i-- != 0);
+#else
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc += XXH3_mix16B(input+48, secret+96, seed);
+                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
+                }
+                acc += XXH3_mix16B(input+32, secret+64, seed);
+                acc += XXH3_mix16B(input+len-48, secret+80, seed);
+            }
+            acc += XXH3_mix16B(input+16, secret+32, seed);
+            acc += XXH3_mix16B(input+len-32, secret+48, seed);
+        }
+        acc += XXH3_mix16B(input+0, secret+0, seed);
+        acc += XXH3_mix16B(input+len-16, secret+16, seed);
+#endif
+        return XXH3_avalanche(acc);
+    }
+}
+
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    #define XXH3_MIDSIZE_STARTOFFSET 3
+    #define XXH3_MIDSIZE_LASTOFFSET  17
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+        xxh_u64 acc_end;
+        unsigned int const nbRounds = (unsigned int)len / 16;
+        unsigned int i;
+        XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+        for (i=0; i<8; i++) {
+            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
+        }
+        /* last bytes */
+        acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
+        XXH_ASSERT(nbRounds >= 8);
+        acc = XXH3_avalanche(acc);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
+         * In everywhere else, it uses scalar code.
+         *
+         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
+         * would still be slower than UMAAL (see XXH_mult64to128).
+         *
+         * Unfortunately, Clang doesn't handle the long multiplies properly and
+         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
+         * scalarized into an ugly mess of VMOV.32 instructions.
+         *
+         * This mess is difficult to avoid without turning autovectorization
+         * off completely, but they are usually relatively minor and/or not
+         * worth it to fix.
+         *
+         * This loop is the easiest to fix, as unlike XXH32, this pragma
+         * _actually works_ because it is a loop vectorization instead of an
+         * SLP vectorization.
+         */
+        #pragma clang loop vectorize(disable)
+#endif
+        for (i=8 ; i < nbRounds; i++) {
+            /*
+             * Prevents clang for unrolling the acc loop and interleaving with this one.
+             */
+            XXH_COMPILER_GUARD(acc);
+            acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+        }
+        return XXH3_avalanche(acc + acc_end);
+    }
+}
+
+
+/* =======     Long Keys     ======= */
+
+#define XXH_STRIPE_LEN 64
+#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
+#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
+
+#ifdef XXH_OLD_NAMES
+#  define STRIPE_LEN XXH_STRIPE_LEN
+#  define ACC_NB XXH_ACC_NB
+#endif
+
+#ifndef XXH_PREFETCH_DIST
+#  ifdef __clang__
+#    define XXH_PREFETCH_DIST 320
+#  else
+#    if (XXH_VECTOR == XXH_AVX512)
+#      define XXH_PREFETCH_DIST 512
+#    else
+#      define XXH_PREFETCH_DIST 384
+#    endif
+#  endif  /* __clang__ */
+#endif  /* XXH_PREFETCH_DIST */
+
+/*
+ * These macros are to generate an XXH3_accumulate() function.
+ * The two arguments select the name suffix and target attribute.
+ *
+ * The name of this symbol is XXH3_accumulate_() and it calls
+ * XXH3_accumulate_512_().
+ *
+ * It may be useful to hand implement this function if the compiler fails to
+ * optimize the inline function.
+ */
+#define XXH3_ACCUMULATE_TEMPLATE(name)                      \
+void                                                        \
+XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc,           \
+                       const xxh_u8* XXH_RESTRICT input,    \
+                       const xxh_u8* XXH_RESTRICT secret,   \
+                       size_t nbStripes)                    \
+{                                                           \
+    size_t n;                                               \
+    for (n = 0; n < nbStripes; n++ ) {                      \
+        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;  \
+        XXH_PREFETCH(in + XXH_PREFETCH_DIST);               \
+        XXH3_accumulate_512_##name(                         \
+                 acc,                                       \
+                 in,                                        \
+                 secret + n*XXH_SECRET_CONSUME_RATE);       \
+    }                                                       \
+}
+
+
+XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
+{
+    if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
+    XXH_memcpy(dst, &v64, sizeof(v64));
+}
+
+/* Several intrinsic functions below are supposed to accept __int64 as argument,
+ * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
+ * However, several environments do not define __int64 type,
+ * requiring a workaround.
+ */
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+    typedef int64_t xxh_i64;
+#else
+    /* the following type must have a width of 64-bit */
+    typedef long long xxh_i64;
+#endif
+
+
+/*
+ * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
+ *
+ * It is a hardened version of UMAC, based off of FARSH's implementation.
+ *
+ * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
+ * implementations, and it is ridiculously fast.
+ *
+ * We harden it by mixing the original input to the accumulators as well as the product.
+ *
+ * This means that in the (relatively likely) case of a multiply by zero, the
+ * original input is preserved.
+ *
+ * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
+ * cross-pollination, as otherwise the upper and lower halves would be
+ * essentially independent.
+ *
+ * This doesn't matter on 64-bit hashes since they all get merged together in
+ * the end, so we skip the extra step.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+#if (XXH_VECTOR == XXH_AVX512) \
+     || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)
+
+#ifndef XXH_TARGET_AVX512
+# define XXH_TARGET_AVX512  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    __m512i* const xacc = (__m512i *) acc;
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+
+    {
+        /* data_vec    = input[0]; */
+        __m512i const data_vec    = _mm512_loadu_si512   (input);
+        /* key_vec     = secret[0]; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        /* data_key    = data_vec ^ key_vec; */
+        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+        /* data_key_lo = data_key >> 32; */
+        __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
+        /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+        __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
+        /* xacc[0] += swap(data_vec); */
+        __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
+        __m512i const sum       = _mm512_add_epi64(*xacc, data_swap);
+        /* xacc[0] += product; */
+        *xacc = _mm512_add_epi64(product, sum);
+    }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
+
+/*
+ * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
+ *
+ * Multiplication isn't perfect, as explained by Google in HighwayHash:
+ *
+ *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
+ *  // varying degrees. In descending order of goodness, bytes
+ *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
+ *  // As expected, the upper and lower bytes are much worse.
+ *
+ * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
+ *
+ * Since our algorithm uses a pseudorandom secret to add some variance into the
+ * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
+ *
+ * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
+ * extraction.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+    {   __m512i* const xacc = (__m512i*) acc;
+        const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
+
+        /* xacc[0] ^= (xacc[0] >> 47) */
+        __m512i const acc_vec     = *xacc;
+        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
+        /* xacc[0] ^= secret; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        __m512i const data_key    = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);
+
+        /* xacc[0] *= XXH_PRIME32_1; */
+        __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
+        __m512i const prod_lo     = _mm512_mul_epu32     (data_key, prime32);
+        __m512i const prod_hi     = _mm512_mul_epu32     (data_key_hi, prime32);
+        *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
+    XXH_ASSERT(((size_t)customSecret & 63) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
+        __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
+        __m512i const seed     = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);
+
+        const __m512i* const src  = (const __m512i*) ((const void*) XXH3_kSecret);
+              __m512i* const dest = (      __m512i*) customSecret;
+        int i;
+        XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 63) == 0);
+        for (i=0; i < nbRounds; ++i) {
+            dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_AVX2) \
+    || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)
+
+#ifndef XXH_TARGET_AVX2
+# define XXH_TARGET_AVX2  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc    =       (__m256i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
+        const         __m256i* const xinput  = (const __m256i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
+            __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm256_add_epi64(product, sum);
+    }   }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc = (__m256i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+        const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m256i const acc_vec     = xacc[i];
+            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
+            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
+            __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
+            __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
+    (void)(&XXH_writeLE64);
+    XXH_PREFETCH(customSecret);
+    {   __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
+
+        const __m256i* const src  = (const __m256i*) ((const void*) XXH3_kSecret);
+              __m256i*       dest = (      __m256i*) customSecret;
+
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dest);
+#       endif
+        XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 31) == 0);
+
+        /* GCC -O2 need unroll loop manually */
+        dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
+    }
+}
+
+#endif
+
+/* x86dispatch always generates SSE2 */
+#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
+
+#ifndef XXH_TARGET_SSE2
+# define XXH_TARGET_SSE2  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* SSE2 is just a half-scale version of the AVX2 version. */
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc    =       (__m128i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xinput  = (const __m128i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+            __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm_add_epi64(product, sum);
+    }   }
+}
+XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc = (__m128i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+        const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m128i const acc_vec     = xacc[i];
+            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
+            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
+            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
+
+#       if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
+        /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
+        XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
+        __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
+#       else
+        __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
+#       endif
+        int i;
+
+        const void* const src16 = XXH3_kSecret;
+        __m128i* dst16 = (__m128i*) customSecret;
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dst16);
+#       endif
+        XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dst16 & 15) == 0);
+
+        for (i=0; i < nbRounds; ++i) {
+            dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_NEON)
+
+/* forward declarations for the scalar routines */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret, size_t lane);
+
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret, size_t lane);
+
+/*!
+ * @internal
+ * @brief The bulk processing loop for NEON and WASM SIMD128.
+ *
+ * The NEON code path is actually partially scalar when running on AArch64. This
+ * is to optimize the pipelining and can have up to 15% speedup depending on the
+ * CPU, and it also mitigates some GCC codegen issues.
+ *
+ * @see XXH3_NEON_LANES for configuring this and details about this optimization.
+ *
+ * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
+ * integers instead of the other platforms which mask full 64-bit vectors,
+ * so the setup is more complicated than just shifting right.
+ *
+ * Additionally, there is an optimization for 4 lanes at once noted below.
+ *
+ * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
+ * there needs to be *three* versions of the accumulate operation used
+ * for the remaining 2 lanes.
+ *
+ * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
+ * nearly perfectly.
+ */
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
+    {   /* GCC for darwin arm64 does not like aliasing here */
+        xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
+        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+        uint8_t const* xinput = (const uint8_t *) input;
+        uint8_t const* xsecret  = (const uint8_t *) secret;
+
+        size_t i;
+#ifdef __wasm_simd128__
+        /*
+         * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
+         * is constant propagated, which results in it converting it to this
+         * inside the loop:
+         *
+         *    a = v128.load(XXH3_kSecret +  0 + $secret_offset, offset = 0)
+         *    b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
+         *    ...
+         *
+         * This requires a full 32-bit address immediate (and therefore a 6 byte
+         * instruction) as well as an add for each offset.
+         *
+         * Putting an asm guard prevents it from folding (at the cost of losing
+         * the alignment hint), and uses the free offset in `v128.load` instead
+         * of adding secret_offset each time which overall reduces code size by
+         * about a kilobyte and improves performance.
+         */
+        XXH_COMPILER_GUARD(xsecret);
+#endif
+        /* Scalar lanes use the normal scalarRound routine */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarRound(acc, input, secret, i);
+        }
+        i = 0;
+        /* 4 NEON lanes at a time. */
+        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput  + (i * 16));
+            uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput  + ((i+1) * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec_1  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t key_vec_2  = XXH_vld1q_u64(xsecret + ((i+1) * 16));
+            /* data_swap = swap(data_vec) */
+            uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
+            uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
+            uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
+
+            /*
+             * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
+             * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
+             * get one vector with the low 32 bits of each lane, and one vector
+             * with the high 32 bits of each lane.
+             *
+             * The intrinsic returns a double vector because the original ARMv7-a
+             * instruction modified both arguments in place. AArch64 and SIMD128 emit
+             * two instructions from this intrinsic.
+             *
+             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+             */
+            uint32x4x2_t unzipped = vuzpq_u32(
+                vreinterpretq_u32_u64(data_key_1),
+                vreinterpretq_u32_u64(data_key_2)
+            );
+            /* data_key_lo = data_key & 0xFFFFFFFF */
+            uint32x4_t data_key_lo = unzipped.val[0];
+            /* data_key_hi = data_key >> 32 */
+            uint32x4_t data_key_hi = unzipped.val[1];
+            /*
+             * Then, we can split the vectors horizontally and multiply which, as for most
+             * widening intrinsics, have a variant that works on both high half vectors
+             * for free on AArch64. A similar instruction is available on SIMD128.
+             *
+             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+             */
+            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+            /*
+             * Clang reorders
+             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             * to
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
+             *
+             * While it would make sense in theory since the addition is faster,
+             * for reasons likely related to umlal being limited to certain NEON
+             * pipelines, this is worse. A compiler guard fixes this.
+             */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_2);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i]   = vaddq_u64(xacc[i], sum_1);
+            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
+        }
+        /* Operate on the remaining NEON lanes 2 at a time. */
+        for (; i < XXH3_NEON_LANES / 2; i++) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* For two lanes, just use VMOVN and VSHRN. */
+            /* data_key_lo = data_key & 0xFFFFFFFF; */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* data_key_hi = data_key >> 32; */
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
+            uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
+            /* Same Clang workaround as before */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i] = vaddq_u64 (xacc[i], sum);
+        }
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   xxh_aliasing_uint64x2_t* xacc       = (xxh_aliasing_uint64x2_t*) acc;
+        uint8_t const* xsecret = (uint8_t const*) secret;
+
+        size_t i;
+        /* WASM uses operator overloads and doesn't need these. */
+#ifndef __wasm_simd128__
+        /* { prime32_1, prime32_1 } */
+        uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1);
+        /* { 0, prime32_1, 0, prime32_1 } */
+        uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32));
+#endif
+
+        /* AArch64 uses both scalar and neon at the same time */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarScrambleRound(acc, secret, i);
+        }
+        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            uint64x2_t acc_vec  = xacc[i];
+            uint64x2_t shifted  = vshrq_n_u64(acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
+
+            /* xacc[i] ^= xsecret[i]; */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* xacc[i] *= XXH_PRIME32_1 */
+#ifdef __wasm_simd128__
+            /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
+            xacc[i] = data_key * XXH_PRIME32_1;
+#else
+            /*
+             * Expanded version with portable NEON intrinsics
+             *
+             *    lo(x) * lo(y) + (hi(x) * lo(y) << 32)
+             *
+             * prod_hi = hi(data_key) * lo(prime) << 32
+             *
+             * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector
+             * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits
+             * and avoid the shift.
+             */
+            uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi);
+            /* Extract low bits for vmlal_u32  */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
+            xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo);
+#endif
+        }
+    }
+}
+#endif
+
+#if (XXH_VECTOR == XXH_VSX)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* presumed aligned */
+    xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+    xxh_u8 const* const xinput   = (xxh_u8 const*) input;   /* no alignment restriction */
+    xxh_u8 const* const xsecret  = (xxh_u8 const*) secret;    /* no alignment restriction */
+    xxh_u64x2 const v32 = { 32, 32 };
+    size_t i;
+    for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+        /* data_vec = xinput[i]; */
+        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + 16*i);
+        /* key_vec = xsecret[i]; */
+        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
+        xxh_u64x2 const data_key = data_vec ^ key_vec;
+        /* shuffled = (data_key << 32) | (data_key >> 32); */
+        xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
+        /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
+        xxh_u64x2 const product  = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
+        /* acc_vec = xacc[i]; */
+        xxh_u64x2 acc_vec        = xacc[i];
+        acc_vec += product;
+
+        /* swap high and low halves */
+#ifdef __s390x__
+        acc_vec += vec_permi(data_vec, data_vec, 2);
+#else
+        acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
+#endif
+        xacc[i] = acc_vec;
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx)
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+        const xxh_u8* const xsecret = (const xxh_u8*) secret;
+        /* constants */
+        xxh_u64x2 const v32  = { 32, 32 };
+        xxh_u64x2 const v47 = { 47, 47 };
+        xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 };
+        size_t i;
+        for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            xxh_u64x2 const acc_vec  = xacc[i];
+            xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
+
+            /* xacc[i] ^= xsecret[i]; */
+            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
+            xxh_u64x2 const data_key = data_vec ^ key_vec;
+
+            /* xacc[i] *= XXH_PRIME32_1 */
+            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
+            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
+            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
+            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
+            xacc[i] = prod_odd + (prod_even << v32);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_SVE)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_sve( void* XXH_RESTRICT acc,
+                   const void* XXH_RESTRICT input,
+                   const void* XXH_RESTRICT secret)
+{
+    uint64_t *xacc = (uint64_t *)acc;
+    const uint64_t *xinput = (const uint64_t *)(const void *)input;
+    const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+    svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+    uint64_t element_count = svcntd();
+    if (element_count >= 8) {
+        svbool_t mask = svptrue_pat_b64(SV_VL8);
+        svuint64_t vacc = svld1_u64(mask, xacc);
+        ACCRND(vacc, 0);
+        svst1_u64(mask, xacc, vacc);
+    } else if (element_count == 2) {   /* sve128 */
+        svbool_t mask = svptrue_pat_b64(SV_VL2);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+        svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+        svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 2);
+        ACCRND(acc2, 4);
+        ACCRND(acc3, 6);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 2, acc1);
+        svst1_u64(mask, xacc + 4, acc2);
+        svst1_u64(mask, xacc + 6, acc3);
+    } else {
+        svbool_t mask = svptrue_pat_b64(SV_VL4);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 4);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 4, acc1);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc,
+               const xxh_u8* XXH_RESTRICT input,
+               const xxh_u8* XXH_RESTRICT secret,
+               size_t nbStripes)
+{
+    if (nbStripes != 0) {
+        uint64_t *xacc = (uint64_t *)acc;
+        const uint64_t *xinput = (const uint64_t *)(const void *)input;
+        const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+        svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+        uint64_t element_count = svcntd();
+        if (element_count >= 8) {
+            svbool_t mask = svptrue_pat_b64(SV_VL8);
+            svuint64_t vacc = svld1_u64(mask, xacc + 0);
+            do {
+                /* svprfd(svbool_t, void *, enum svfprop); */
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(vacc, 0);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, vacc);
+        } else if (element_count == 2) { /* sve128 */
+            svbool_t mask = svptrue_pat_b64(SV_VL2);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+            svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+            svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 2);
+                ACCRND(acc2, 4);
+                ACCRND(acc3, 6);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 2, acc1);
+           svst1_u64(mask, xacc + 4, acc2);
+           svst1_u64(mask, xacc + 6, acc3);
+        } else {
+            svbool_t mask = svptrue_pat_b64(SV_VL4);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 4);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 4, acc1);
+       }
+    }
+}
+
+#endif
+
+/* scalar variants - universal */
+
+#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__))
+/*
+ * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they
+ * emit an excess mask and a full 64-bit multiply-add (MADD X-form).
+ *
+ * While this might not seem like much, as AArch64 is a 64-bit architecture, only
+ * big Cortex designs have a full 64-bit multiplier.
+ *
+ * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit
+ * multiplies expand to 2-3 multiplies in microcode. This has a major penalty
+ * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline.
+ *
+ * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does
+ * not have this penalty and does the mask automatically.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+    xxh_u64 ret;
+    /* note: %x = 64-bit register, %w = 32-bit register */
+    __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc));
+    return ret;
+}
+#else
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+    return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc;
+}
+#endif
+
+/*!
+ * @internal
+ * @brief Scalar round for @ref XXH3_accumulate_512_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc,
+                 void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret,
+                 size_t lane)
+{
+    xxh_u64* xacc = (xxh_u64*) acc;
+    xxh_u8 const* xinput  = (xxh_u8 const*) input;
+    xxh_u8 const* xsecret = (xxh_u8 const*) secret;
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    {
+        xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8);
+        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8);
+        xacc[lane ^ 1] += data_val; /* swap adjacent lanes */
+        xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
+    }
+}
+
+/*!
+ * @internal
+ * @brief Processes a 64 byte block of data using the scalar path.
+ */
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    size_t i;
+    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__GNUC__) && !defined(__clang__) \
+  && (defined(__arm__) || defined(__thumb2__)) \
+  && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
+  && XXH_SIZE_OPT <= 0
+#  pragma GCC unroll 8
+#endif
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarRound(acc, input, secret, i);
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
+
+/*!
+ * @internal
+ * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret,
+                         size_t lane)
+{
+    xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
+    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
+    XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    {
+        xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
+        xxh_u64 acc64 = xacc[lane];
+        acc64 = XXH_xorshift64(acc64, 47);
+        acc64 ^= key64;
+        acc64 *= XXH_PRIME32_1;
+        xacc[lane] = acc64;
+    }
+}
+
+/*!
+ * @internal
+ * @brief Scrambles the accumulators after a large chunk has been read
+ */
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    size_t i;
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarScrambleRound(acc, secret, i);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    /*
+     * We need a separate pointer for the hack below,
+     * which requires a non-const pointer.
+     * Any decent compiler will optimize this out otherwise.
+     */
+    const xxh_u8* kSecretPtr = XXH3_kSecret;
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+
+#if defined(__GNUC__) && defined(__aarch64__)
+    /*
+     * UGLY HACK:
+     * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are
+     * placed sequentially, in order, at the top of the unrolled loop.
+     *
+     * While MOVK is great for generating constants (2 cycles for a 64-bit
+     * constant compared to 4 cycles for LDR), it fights for bandwidth with
+     * the arithmetic instructions.
+     *
+     *   I   L   S
+     * MOVK
+     * MOVK
+     * MOVK
+     * MOVK
+     * ADD
+     * SUB      STR
+     *          STR
+     * By forcing loads from memory (as the asm line causes the compiler to assume
+     * that XXH3_kSecretPtr has been changed), the pipelines are used more
+     * efficiently:
+     *   I   L   S
+     *      LDR
+     *  ADD LDR
+     *  SUB     STR
+     *          STR
+     *
+     * See XXH3_NEON_LANES for details on the pipsline.
+     *
+     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
+     *   without hack: 2654.4 MB/s
+     *   with hack:    3202.9 MB/s
+     */
+    XXH_COMPILER_GUARD(kSecretPtr);
+#endif
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+        int i;
+        for (i=0; i < nbRounds; i++) {
+            /*
+             * The asm hack causes the compiler to assume that kSecretPtr aliases with
+             * customSecret, and on aarch64, this prevented LDP from merging two
+             * loads together for free. Putting the loads together before the stores
+             * properly generates LDP.
+             */
+            xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
+            xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i,     lo);
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
+    }   }
+}
+
+
+typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t);
+typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
+typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
+
+
+#if (XXH_VECTOR == XXH_AVX512)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_accumulate     XXH3_accumulate_avx512
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx512
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512
+
+#elif (XXH_VECTOR == XXH_AVX2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx2
+#define XXH3_accumulate     XXH3_accumulate_avx2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2
+
+#elif (XXH_VECTOR == XXH_SSE2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_sse2
+#define XXH3_accumulate     XXH3_accumulate_sse2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_sse2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2
+
+#elif (XXH_VECTOR == XXH_NEON)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_neon
+#define XXH3_accumulate     XXH3_accumulate_neon
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_neon
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#elif (XXH_VECTOR == XXH_VSX)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_vsx
+#define XXH3_accumulate     XXH3_accumulate_vsx
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_vsx
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#elif (XXH_VECTOR == XXH_SVE)
+#define XXH3_accumulate_512 XXH3_accumulate_512_sve
+#define XXH3_accumulate     XXH3_accumulate_sve
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#else /* scalar */
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_accumulate     XXH3_accumulate_scalar
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#endif
+
+#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
+#  undef XXH3_initCustomSecret
+#  define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+#endif
+
+XXH_FORCE_INLINE void
+XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
+                      const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
+    size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
+    size_t const nb_blocks = (len - 1) / block_len;
+
+    size_t n;
+
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+
+    for (n = 0; n < nb_blocks; n++) {
+        f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
+        f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
+    }
+
+    /* last partial block */
+    XXH_ASSERT(len > XXH_STRIPE_LEN);
+    {   size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
+        XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
+        f_acc(acc, input + nb_blocks*block_len, secret, nbStripes);
+
+        /* last stripe */
+        {   const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
+#define XXH_SECRET_LASTACC_START 7  /* not aligned on 8, last secret is different from acc & scrambler */
+            XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+    }   }
+}
+
+XXH_FORCE_INLINE xxh_u64
+XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
+{
+    return XXH3_mul128_fold64(
+               acc[0] ^ XXH_readLE64(secret),
+               acc[1] ^ XXH_readLE64(secret+8) );
+}
+
+static XXH64_hash_t
+XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
+{
+    xxh_u64 result64 = start;
+    size_t i = 0;
+
+    for (i = 0; i < 4; i++) {
+        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
+         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
+         * XXH3_64bits, len == 256, Snapdragon 835:
+         *   without hack: 2063.7 MB/s
+         *   with hack:    2560.7 MB/s
+         */
+        XXH_COMPILER_GUARD(result64);
+#endif
+    }
+
+    return XXH3_avalanche(result64);
+}
+
+#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \
+                        XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 }
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
+                           const void* XXH_RESTRICT secret, size_t secretSize,
+                           XXH3_f_accumulate f_acc,
+                           XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    /* do not align on 8, so that the secret is different from the accumulator */
+#define XXH_SECRET_MERGEACCS_START 11
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
+}
+
+/*
+ * It's important for performance to transmit secret's size (when it's static)
+ * so that the compiler can properly optimize the vectorized loop.
+ * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                             XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * It's preferable for performance that XXH3_hashLong is not inlined,
+ * as it results in a smaller function for small data, easier to the instruction cache.
+ * Note that inside this no_inline function, we do inline the internal loop,
+ * and provide a statically defined secret size to allow optimization of vector loop.
+ */
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
+                          XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * XXH3_hashLong_64b_withSeed():
+ * Generate a custom key based on alteration of default XXH3_kSecret with the seed,
+ * and then use this key for long mode hashing.
+ *
+ * This operation is decently fast but nonetheless costs a little bit of time.
+ * Try to avoid it whenever possible (typically when seed==0).
+ *
+ * It's important for performance that XXH3_hashLong is not inlined. Not sure
+ * why (uop cache maybe?), but the difference is large and easily measurable.
+ */
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
+                                    XXH64_hash_t seed,
+                                    XXH3_f_accumulate f_acc,
+                                    XXH3_f_scrambleAcc f_scramble,
+                                    XXH3_f_initCustomSecret f_initSec)
+{
+#if XXH_SIZE_OPT <= 0
+    if (seed == 0)
+        return XXH3_hashLong_64b_internal(input, len,
+                                          XXH3_kSecret, sizeof(XXH3_kSecret),
+                                          f_acc, f_scramble);
+#endif
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed);
+        return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
+                                          f_acc, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+
+typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
+                                          XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
+                     XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                     XXH3_hashLong64_f f_hashLong)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secretLen` condition is not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     * Also, note that function signature doesn't offer room to return an error.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
+}
+
+
+/* ===   Public entry point   === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
+{
+    return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
+{
+    return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
+
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (length <= XXH3_MIDSIZE_MAX)
+        return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
+}
+
+
+/* ===   XXH3 streaming   === */
+#ifndef XXH_NO_STREAM
+/*
+ * Malloc's a pointer that is always aligned to align.
+ *
+ * This must be freed with `XXH_alignedFree()`.
+ *
+ * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
+ * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
+ * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
+ *
+ * This underalignment previously caused a rather obvious crash which went
+ * completely unnoticed due to XXH3_createState() not actually being tested.
+ * Credit to RedSpah for noticing this bug.
+ *
+ * The alignment is done manually: Functions like posix_memalign or _mm_malloc
+ * are avoided: To maintain portability, we would have to write a fallback
+ * like this anyways, and besides, testing for the existence of library
+ * functions without relying on external build tools is impossible.
+ *
+ * The method is simple: Overallocate, manually align, and store the offset
+ * to the original behind the returned pointer.
+ *
+ * Align must be a power of 2 and 8 <= align <= 128.
+ */
+static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
+{
+    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
+    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
+    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
+    {   /* Overallocate to make room for manual realignment and an offset byte */
+        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
+        if (base != NULL) {
+            /*
+             * Get the offset needed to align this pointer.
+             *
+             * Even if the returned pointer is aligned, there will always be
+             * at least one byte to store the offset to the original pointer.
+             */
+            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
+            /* Add the offset for the now-aligned pointer */
+            xxh_u8* ptr = base + offset;
+
+            XXH_ASSERT((size_t)ptr % align == 0);
+
+            /* Store the offset immediately before the returned pointer. */
+            ptr[-1] = (xxh_u8)offset;
+            return ptr;
+        }
+        return NULL;
+    }
+}
+/*
+ * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
+ * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
+ */
+static void XXH_alignedFree(void* p)
+{
+    if (p != NULL) {
+        xxh_u8* ptr = (xxh_u8*)p;
+        /* Get the offset byte we added in XXH_malloc. */
+        xxh_u8 offset = ptr[-1];
+        /* Free the original malloc'd pointer */
+        xxh_u8* base = ptr - offset;
+        XXH_free(base);
+    }
+}
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Allocate an @ref XXH3_state_t.
+ *
+ * @return An allocated pointer of @ref XXH3_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH3_freeState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
+{
+    XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
+    if (state==NULL) return NULL;
+    XXH3_INITSTATE(state);
+    return state;
+}
+
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Frees an @ref XXH3_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note Must be allocated with XXH3_createState().
+ *
+ * @see @ref streaming_example "Streaming Example"
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
+{
+    XXH_alignedFree(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
+{
+    XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
+}
+
+static void
+XXH3_reset_internal(XXH3_state_t* statePtr,
+                    XXH64_hash_t seed,
+                    const void* secret, size_t secretSize)
+{
+    size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
+    size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
+    XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart);
+    XXH_ASSERT(statePtr != NULL);
+    /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */
+    memset((char*)statePtr + initStart, 0, initLength);
+    statePtr->acc[0] = XXH_PRIME32_3;
+    statePtr->acc[1] = XXH_PRIME64_1;
+    statePtr->acc[2] = XXH_PRIME64_2;
+    statePtr->acc[3] = XXH_PRIME64_3;
+    statePtr->acc[4] = XXH_PRIME64_4;
+    statePtr->acc[5] = XXH_PRIME32_2;
+    statePtr->acc[6] = XXH_PRIME64_5;
+    statePtr->acc[7] = XXH_PRIME32_1;
+    statePtr->seed = seed;
+    statePtr->useSeed = (seed != 0);
+    statePtr->extSecret = (const unsigned char*)secret;
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+    statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
+    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, secret, secretSize);
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (seed==0) return XXH3_64bits_reset(statePtr);
+    if ((seed != statePtr->seed) || (statePtr->extSecret != NULL))
+        XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, seed64, secret, secretSize);
+    statePtr->useSeed = 1; /* always, even if seed64==0 */
+    return XXH_OK;
+}
+
+/*!
+ * @internal
+ * @brief Processes a large input for XXH3_update() and XXH3_digest_long().
+ *
+ * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block.
+ *
+ * @param acc                Pointer to the 8 accumulator lanes
+ * @param nbStripesSoFarPtr  In/out pointer to the number of leftover stripes in the block*
+ * @param nbStripesPerBlock  Number of stripes in a block
+ * @param input              Input pointer
+ * @param nbStripes          Number of stripes to process
+ * @param secret             Secret pointer
+ * @param secretLimit        Offset of the last block in @p secret
+ * @param f_acc              Pointer to an XXH3_accumulate implementation
+ * @param f_scramble         Pointer to an XXH3_scrambleAcc implementation
+ * @return                   Pointer past the end of @p input after processing
+ */
+XXH_FORCE_INLINE const xxh_u8 *
+XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
+                    size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
+                    const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
+                    const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
+                    XXH3_f_accumulate f_acc,
+                    XXH3_f_scrambleAcc f_scramble)
+{
+    const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE;
+    /* Process full blocks */
+    if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) {
+        /* Process the initial partial block... */
+        size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;
+
+        do {
+            /* Accumulate and scramble */
+            f_acc(acc, input, initialSecret, nbStripesThisIter);
+            f_scramble(acc, secret + secretLimit);
+            input += nbStripesThisIter * XXH_STRIPE_LEN;
+            nbStripes -= nbStripesThisIter;
+            /* Then continue the loop with the full block size */
+            nbStripesThisIter = nbStripesPerBlock;
+            initialSecret = secret;
+        } while (nbStripes >= nbStripesPerBlock);
+        *nbStripesSoFarPtr = 0;
+    }
+    /* Process a partial block */
+    if (nbStripes > 0) {
+        f_acc(acc, input, initialSecret, nbStripes);
+        input += nbStripes * XXH_STRIPE_LEN;
+        *nbStripesSoFarPtr += nbStripes;
+    }
+    /* Return end pointer */
+    return input;
+}
+
+#ifndef XXH3_STREAM_USE_STACK
+# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */
+#   define XXH3_STREAM_USE_STACK 1
+# endif
+#endif
+/*
+ * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
+ */
+XXH_FORCE_INLINE XXH_errorcode
+XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
+            const xxh_u8* XXH_RESTRICT input, size_t len,
+            XXH3_f_accumulate f_acc,
+            XXH3_f_scrambleAcc f_scramble)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    XXH_ASSERT(state != NULL);
+    {   const xxh_u8* const bEnd = input + len;
+        const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* For some reason, gcc and MSVC seem to suffer greatly
+         * when operating accumulators directly into state.
+         * Operating into stack space seems to enable proper optimization.
+         * clang, on the other hand, doesn't seem to need this trick */
+        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
+        XXH_memcpy(acc, state->acc, sizeof(acc));
+#else
+        xxh_u64* XXH_RESTRICT const acc = state->acc;
+#endif
+        state->totalLen += len;
+        XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
+
+        /* small input : just fill in tmp buffer */
+        if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
+            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+            state->bufferedSize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
+
+        /* total input is now > XXH3_INTERNALBUFFER_SIZE */
+        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
+        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0);   /* clean multiple */
+
+        /*
+         * Internal buffer is partially filled (always, except at beginning)
+         * Complete it, then consume it.
+         */
+        if (state->bufferedSize) {
+            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
+            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
+            input += loadSize;
+            XXH3_consumeStripes(acc,
+                               &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
+                                secret, state->secretLimit,
+                                f_acc, f_scramble);
+            state->bufferedSize = 0;
+        }
+        XXH_ASSERT(input < bEnd);
+        if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
+            size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
+            input = XXH3_consumeStripes(acc,
+                                       &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                       input, nbStripes,
+                                       secret, state->secretLimit,
+                                       f_acc, f_scramble);
+            XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+
+        }
+        /* Some remaining input (always) : buffer it */
+        XXH_ASSERT(input < bEnd);
+        XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
+        XXH_ASSERT(state->bufferedSize == 0);
+        XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
+        state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* save stack accumulators into state */
+        XXH_memcpy(state->acc, acc, sizeof(acc));
+#endif
+    }
+
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_update(state, (const xxh_u8*)input, len,
+                       XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+                  const XXH3_state_t* state,
+                  const unsigned char* secret)
+{
+    xxh_u8 lastStripe[XXH_STRIPE_LEN];
+    const xxh_u8* lastStripePtr;
+
+    /*
+     * Digest on a local copy. This way, the state remains unaltered, and it can
+     * continue ingesting more input afterwards.
+     */
+    XXH_memcpy(acc, state->acc, sizeof(state->acc));
+    if (state->bufferedSize >= XXH_STRIPE_LEN) {
+        /* Consume remaining stripes then point to remaining data in buffer */
+        size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
+        size_t nbStripesSoFar = state->nbStripesSoFar;
+        XXH3_consumeStripes(acc,
+                           &nbStripesSoFar, state->nbStripesPerBlock,
+                            state->buffer, nbStripes,
+                            secret, state->secretLimit,
+                            XXH3_accumulate, XXH3_scrambleAcc);
+        lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
+    } else {  /* bufferedSize < XXH_STRIPE_LEN */
+        /* Copy to temp buffer */
+        size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
+        XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
+        XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
+        XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
+        lastStripePtr = lastStripe;
+    }
+    /* Last stripe */
+    XXH3_accumulate_512(acc,
+                        lastStripePtr,
+                        secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
+{
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        return XXH3_mergeAccs(acc,
+                              secret + XXH_SECRET_MERGEACCS_START,
+                              (xxh_u64)state->totalLen * XXH_PRIME64_1);
+    }
+    /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
+    if (state->useSeed)
+        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                  secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+
+
+/* ==========================================
+ * XXH3 128 bits (a.k.a XXH128)
+ * ==========================================
+ * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
+ * even without counting the significantly larger output size.
+ *
+ * For example, extra steps are taken to avoid the seed-dependent collisions
+ * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
+ *
+ * This strength naturally comes at the cost of some speed, especially on short
+ * lengths. Note that longer hashes are about as fast as the 64-bit version
+ * due to it using only a slight modification of the 64-bit loop.
+ *
+ * XXH128 is also more oriented towards 64-bit machines. It is still extremely
+ * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
+ */
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    /* A doubled version of 1to3_64b with different constants. */
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8 const c1 = input[0];
+        xxh_u8 const c2 = input[len >> 1];
+        xxh_u8 const c3 = input[len - 1];
+        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
+                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
+        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
+        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
+        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
+        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
+        XXH128_hash_t h128;
+        h128.low64  = XXH64_avalanche(keyed_lo);
+        h128.high64 = XXH64_avalanche(keyed_hi);
+        return h128;
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input_lo = XXH_readLE32(input);
+        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
+        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
+        xxh_u64 const keyed = input_64 ^ bitflip;
+
+        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
+        XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));
+
+        m128.high64 += (m128.low64 << 1);
+        m128.low64  ^= (m128.high64 >> 3);
+
+        m128.low64   = XXH_xorshift64(m128.low64, 35);
+        m128.low64  *= PRIME_MX2;
+        m128.low64   = XXH_xorshift64(m128.low64, 28);
+        m128.high64  = XXH3_avalanche(m128.high64);
+        return m128;
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
+        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
+        xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
+        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1);
+        /*
+         * Put len in the middle of m128 to ensure that the length gets mixed to
+         * both the low and high bits in the 128x64 multiply below.
+         */
+        m128.low64 += (xxh_u64)(len - 1) << 54;
+        input_hi   ^= bitfliph;
+        /*
+         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
+         * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to
+         * the high 64 bits of m128.
+         *
+         * The best approach to this operation is different on 32-bit and 64-bit.
+         */
+        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
+            /*
+             * 32-bit optimized version, which is more readable.
+             *
+             * On 32-bit, it removes an ADC and delays a dependency between the two
+             * halves of m128.high64, but it generates an extra mask on 64-bit.
+             */
+            m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
+        } else {
+            /*
+             * 64-bit optimized (albeit more confusing) version.
+             *
+             * Uses some properties of addition and multiplication to remove the mask:
+             *
+             * Let:
+             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
+             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
+             *    c = XXH_PRIME32_2
+             *
+             *    a + (b * c)
+             * Inverse Property: x + y - x == y
+             *    a + (b * (1 + c - 1))
+             * Distributive Property: x * (y + z) == (x * y) + (x * z)
+             *    a + (b * 1) + (b * (c - 1))
+             * Identity Property: x * 1 == x
+             *    a + b + (b * (c - 1))
+             *
+             * Substitute a, b, and c:
+             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             *
+             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
+             *    input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             */
+            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
+        }
+        /* m128 ^= XXH_swap64(m128 >> 64); */
+        m128.low64  ^= XXH_swap64(m128.high64);
+
+        {   /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */
+            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2);
+            h128.high64 += m128.high64 * XXH_PRIME64_2;
+
+            h128.low64   = XXH3_avalanche(h128.low64);
+            h128.high64  = XXH3_avalanche(h128.high64);
+            return h128;
+    }   }
+}
+
+/*
+ * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(len <= 16);
+    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
+        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
+        {   XXH128_hash_t h128;
+            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
+            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
+            h128.low64 = XXH64_avalanche(seed ^ bitflipl);
+            h128.high64 = XXH64_avalanche( seed ^ bitfliph);
+            return h128;
+    }   }
+}
+
+/*
+ * A bit slower than XXH3_mix16B, but handles multiply by zero better.
+ */
+XXH_FORCE_INLINE XXH128_hash_t
+XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
+              const xxh_u8* secret, XXH64_hash_t seed)
+{
+    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
+    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
+    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
+    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
+    return acc;
+}
+
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
+
+    {   XXH128_hash_t acc;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+
+#if XXH_SIZE_OPT >= 1
+        {
+            /* Smaller, but slightly slower. */
+            unsigned int i = (unsigned int)(len - 1) / 32;
+            do {
+                acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed);
+            } while (i-- != 0);
+        }
+#else
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
+                }
+                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
+            }
+            acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
+        }
+        acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
+#endif
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
+
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                       XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    {   XXH128_hash_t acc;
+        unsigned i;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+        /*
+         *  We set as `i` as offset + 32. We do this so that unchanged
+         * `len` can be used as upper bound. This reaches a sweet spot
+         * where both x86 and aarch64 get simple agen and good codegen
+         * for the loop.
+         */
+        for (i = 32; i < 160; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input  + i - 32,
+                                input  + i - 16,
+                                secret + i - 32,
+                                seed);
+        }
+        acc.low64 = XXH3_avalanche(acc.low64);
+        acc.high64 = XXH3_avalanche(acc.high64);
+        /*
+         * NB: `i <= len` will duplicate the last 32-bytes if
+         * len % 32 was zero. This is an unfortunate necessity to keep
+         * the hash result stable.
+         */
+        for (i=160; i <= len; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input + i - 32,
+                                input + i - 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
+                                seed);
+        }
+        /* last bytes */
+        acc = XXH128_mix32B(acc,
+                            input + len - 16,
+                            input + len - 32,
+                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
+                            (XXH64_hash_t)0 - seed);
+
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
+                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    {   XXH128_hash_t h128;
+        h128.low64  = XXH3_mergeAccs(acc,
+                                     secret + XXH_SECRET_MERGEACCS_START,
+                                     (xxh_u64)len * XXH_PRIME64_1);
+        h128.high64 = XXH3_mergeAccs(acc,
+                                     secret + secretSize
+                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                     ~((xxh_u64)len * XXH_PRIME64_2));
+        return h128;
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong() is not inlined.
+ */
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed64,
+                           const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
+                                       XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance to pass @p secretLen (when it's static)
+ * to the compiler, so that it can properly optimize the vectorized loop.
+ *
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                              XXH64_hash_t seed64,
+                              const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
+                                       XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
+                                XXH64_hash_t seed64,
+                                XXH3_f_accumulate f_acc,
+                                XXH3_f_scrambleAcc f_scramble,
+                                XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed64 == 0)
+        return XXH3_hashLong_128b_internal(input, len,
+                                           XXH3_kSecret, sizeof(XXH3_kSecret),
+                                           f_acc, f_scramble);
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed64);
+        return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
+                                           f_acc, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed(const void* input, size_t len,
+                            XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
+                                            XXH64_hash_t, const void* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_128bits_internal(const void* input, size_t len,
+                      XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                      XXH3_hashLong128_f f_hl128)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secret` conditions are not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hl128(input, len, seed64, secret, secretLen);
+}
+
+
+/* ===   Public XXH128 API   === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_default);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 (const xxh_u8*)secret, secretSize,
+                                 XXH3_hashLong_128b_withSecret);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_internal(input, len, seed,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_withSeed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_withSeed(input, len, seed);
+}
+
+
+/* ===   XXH3 128-bit streaming   === */
+#ifndef XXH_NO_STREAM
+/*
+ * All initialization and update functions are identical to 64-bit streaming variant.
+ * The only difference is the finalization routine.
+ */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
+{
+    return XXH3_64bits_reset(statePtr);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+    return XXH3_64bits_reset_withSeed(statePtr, seed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_64bits_update(state, input, len);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
+{
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+        {   XXH128_hash_t h128;
+            h128.low64  = XXH3_mergeAccs(acc,
+                                         secret + XXH_SECRET_MERGEACCS_START,
+                                         (xxh_u64)state->totalLen * XXH_PRIME64_1);
+            h128.high64 = XXH3_mergeAccs(acc,
+                                         secret + state->secretLimit + XXH_STRIPE_LEN
+                                                - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                         ~((xxh_u64)state->totalLen * XXH_PRIME64_2));
+            return h128;
+        }
+    }
+    /* len <= XXH3_MIDSIZE_MAX : short code */
+    if (state->useSeed)
+        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                   secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+/* 128-bit utility functions */
+
+#include    /* memcmp, memcpy */
+
+/* return : 1 is equal, 0 if different */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
+{
+    /* note : XXH128_hash_t is compact, it has no padding byte */
+    return !(memcmp(&h1, &h2, sizeof(h1)));
+}
+
+/* This prototype is compatible with stdlib's qsort().
+ * @return : >0 if *h128_1  > *h128_2
+ *           <0 if *h128_1  < *h128_2
+ *           =0 if *h128_1 == *h128_2  */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2)
+{
+    XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
+    XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
+    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
+    /* note : bets that, in most cases, hash values are different */
+    if (hcmp) return hcmp;
+    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
+}
+
+
+/*======   Canonical representation   ======*/
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) {
+        hash.high64 = XXH_swap64(hash.high64);
+        hash.low64  = XXH_swap64(hash.low64);
+    }
+    XXH_memcpy(dst, &hash.high64, sizeof(hash.high64));
+    XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
+{
+    XXH128_hash_t h;
+    h.high64 = XXH_readBE64(src);
+    h.low64  = XXH_readBE64(src->digest + 8);
+    return h;
+}
+
+
+
+/* ==========================================
+ * Secret generators
+ * ==========================================
+ */
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
+
+XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
+{
+    XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
+    XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize)
+{
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(secretBuffer != NULL);
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+#else
+    /* production mode, assert() are disabled */
+    if (secretBuffer == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+#endif
+
+    if (customSeedSize == 0) {
+        customSeed = XXH3_kSecret;
+        customSeedSize = XXH_SECRET_DEFAULT_SIZE;
+    }
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(customSeed != NULL);
+#else
+    if (customSeed == NULL) return XXH_ERROR;
+#endif
+
+    /* Fill secretBuffer with a copy of customSeed - repeat as needed */
+    {   size_t pos = 0;
+        while (pos < secretSize) {
+            size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize);
+            memcpy((char*)secretBuffer + pos, customSeed, toCopy);
+            pos += toCopy;
+    }   }
+
+    {   size_t const nbSeg16 = secretSize / 16;
+        size_t n;
+        XXH128_canonical_t scrambler;
+        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+        for (n=0; n   /* abort() */
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+#include 
+#include 
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "xxhash/xxhash.h"
+#include "sha1/sha1.h"
+#include "sha256/sha256.h"
+
+#ifdef __cplusplus
+}
+#endif
+
+
+// uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp')
+#define UUID_NAMESPACE_LLAMA_CPP "ef001206-dadc-5f6d-a15f-3359e577d4e5"
+#define UUID_NAMESPACE_LLAMA_CPP_HEX 0xef, 0x00, 0x12, 0x06, 0xda, 0xdc, 0x5f, 0x6d, 0xa1, 0x5f, 0x33, 0x59, 0xe5, 0x77, 0xd4, 0xe5
+
+
+#define HASH_TYPE_SHA256_STR "sha256"
+#define HASH_TYPE_SHA1_STR   "sha1"
+#define HASH_TYPE_XXH64_STR  "xxh64"
+#define HASH_TYPE_UUID_STR   "uuid"
+
+
+typedef enum {
+    HASH_EXIT_SUCCESS = 0, // All hash has been generated or validated
+    HASH_EXIT_FAILURE = 1, // Generic Failure
+    HASH_EXIT_MISMATCH = 2, // Hash mismatched during validation
+    HASH_EXIT_MANIFEST_MISSING_ENTRY = 3, // Hash attempted validation but missing entry in manifest
+    HASH_EXIT_MANIFEST_UNKNOWN_HASH = 4, // Manifest is present, but we do not know any hash format within it
+    HASH_EXIT_MANIFEST_FILE_ERROR = 5 // Manifest is either missing or not a known format
+} hash_exit_code_t;
+
+
+typedef enum {
+    HASH_MANIFEST_NOT_FOUND,
+    HASH_MANIFEST_MISMATCH,
+    HASH_MANIFEST_OK,
+} hash_manifest_result_t;
+
+
+struct hash_params {
+    std::string input;
+    bool xxh64 = false;
+    bool sha1 = false;
+    bool sha256 = false;
+    bool uuid = false;
+
+    bool no_layer = false;
+
+    bool manifest_is_usable = false;
+    std::string manifest_file;
+};
+
+struct manifest_check_params {
+    bool xxh64 = false;
+    bool sha1 = false;
+    bool sha256 = false;
+    bool uuid = false;
+};
+
+static char const * hash_manifest_result_to_str(hash_manifest_result_t value) {
+    switch (value) {
+        case HASH_MANIFEST_NOT_FOUND: return "Not Found";
+        case HASH_MANIFEST_MISMATCH: return "Mismatch";
+        case HASH_MANIFEST_OK: return "Ok";
+    }
+    return "?";
+}
+
+static char const * hash_exit_code_to_str(hash_exit_code_t value) {
+    switch (value) {
+        case HASH_EXIT_SUCCESS: return "Success";
+        case HASH_EXIT_FAILURE: return "Failure";
+        case HASH_EXIT_MISMATCH: return "Mismatch";
+        case HASH_EXIT_MANIFEST_MISSING_ENTRY: return "Manifest Missing Entry";
+        case HASH_EXIT_MANIFEST_UNKNOWN_HASH: return "Manifest Unknown Hash";
+        case HASH_EXIT_MANIFEST_FILE_ERROR: return "Manifest File Error";
+    }
+    return "?";
+}
+
+static void hash_print_usage(const char * executable) {
+    const hash_params default_params;
+    printf("\n");
+    printf("usage: %s [options] GGUF_IN\n", executable);
+    printf("\n");
+    printf("Hash a GGUF file");
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help              show this help message and exit\n");
+    printf("      --xxh64             use xxh64 hash\n");
+    printf("      --sha1              use sha1 hash\n");
+    printf("      --sha256            use sha256 hash\n");
+    printf("      --all               use all hash\n");
+    printf("      --no-layer          exclude per layer hash\n");
+    printf("      --uuid              generate UUIDv5 ID\n");
+    printf("  -c, --check   verify against a manifest\n");
+    printf("\n");
+}
+
+static void hash_params_parse_ex(int argc, const char ** argv, hash_params & params) {
+    std::string arg;
+    bool invalid_param = false;
+    const std::string arg_prefix = "--";
+
+    int arg_idx = 1;
+    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
+        arg = argv[arg_idx];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        bool arg_found = false;
+        if (arg == "-h" || arg == "--help") {
+            hash_print_usage(argv[0]);
+            exit(0);
+        }
+
+        if (arg == "--xxh64") {
+            arg_found = true;
+            params.xxh64 = true;
+        }
+
+        if (arg == "--sha1") {
+            arg_found = true;
+            params.sha1 = true;
+        }
+
+        if (arg == "--uuid") {
+            arg_found = true;
+            params.uuid = true;
+        }
+
+        if (arg == "--sha256") {
+            arg_found = true;
+            params.sha256 = true;
+        }
+
+        if (arg == "--all") {
+            arg_found = true;
+            params.sha256 = true;
+            params.sha1 = true;
+            params.xxh64 = true;
+        }
+
+        if (arg == "--no-layer") {
+            arg_found = true;
+            params.no_layer = true;
+        }
+
+        if (arg == "-c" || arg == "--check") {
+            if (++arg_idx >= argc) {
+                invalid_param = true;
+                break;
+            }
+            arg_found = true;
+            params.manifest_file = argv[arg_idx];
+        }
+
+        if (!arg_found) {
+            throw std::invalid_argument("error: unknown argument: " + arg);
+        }
+    }
+
+    if (invalid_param) {
+        throw std::invalid_argument("error: invalid parameter for argument:" + arg);
+    }
+
+    if (argc - arg_idx < 1) {
+        throw std::invalid_argument("error: bad arguments");
+    }
+
+    params.input = argv[arg_idx++];
+}
+
+static bool hash_params_parse(int argc, const char ** argv, hash_params & params) {
+    bool result = true;
+    try {
+        hash_params_parse_ex(argc, argv, params);
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        hash_print_usage(argv[0]);
+        exit(EXIT_FAILURE);
+    }
+    return result;
+}
+
+static bool manifest_type(const std::string & manifest_file, manifest_check_params & manifest_check) {
+    if (manifest_file.empty()) {
+        return false;
+    }
+
+    std::ifstream file(manifest_file);
+    if (!file.is_open()) {
+        return false;
+    }
+
+    std::string manifest_entry_line;
+    while (getline(file, manifest_entry_line)) {
+        // hash_type_str hash_str tensor_name
+        // e.g. 'xxh64     f66e9cd66a4396a0  test.gguf:tensor_0'
+        std::istringstream line_stream(manifest_entry_line);
+        std::string file_hash_type;
+        if (line_stream >> file_hash_type) {
+            if (file_hash_type == HASH_TYPE_SHA256_STR) {
+                manifest_check.sha256 = true;
+            } else if (file_hash_type == HASH_TYPE_SHA1_STR) {
+                manifest_check.sha1 = true;
+            } else if (file_hash_type == HASH_TYPE_XXH64_STR) {
+                manifest_check.xxh64 = true;
+            } else if (file_hash_type == HASH_TYPE_UUID_STR) {
+                manifest_check.uuid = true;
+            }
+        }
+    }
+
+    return true;
+}
+
+static hash_manifest_result_t manifest_verify(const std::string& manifest_file, const std::string& hash_type_str, const std::string& hash_str, const std::string& tensor_name) {
+    if (manifest_file.empty()) {
+        return HASH_MANIFEST_NOT_FOUND;
+    }
+
+    std::ifstream file(manifest_file);
+    if (!file.is_open()) {
+        return HASH_MANIFEST_NOT_FOUND;
+    }
+
+    std::string manifest_entry_line;
+    while (getline(file, manifest_entry_line)) {
+        std::istringstream line_stream(manifest_entry_line);
+        std::string file_hash_type;
+        std::string file_hash;
+        std::string file_tensor_name;
+        if (line_stream >> file_hash_type >> file_hash >> file_tensor_name) {
+            // Line parsed. Check hash validity
+
+            if (file_hash_type != hash_type_str) {
+                continue;
+            }
+
+            if (file_tensor_name != tensor_name) {
+                continue;
+            }
+
+            return (file_hash == hash_str) ? HASH_MANIFEST_OK : HASH_MANIFEST_MISMATCH;
+        }
+    }
+
+    return HASH_MANIFEST_NOT_FOUND;
+}
+
+static void generate_uuidv5(const unsigned char sha1_digest[20], unsigned char uuid[16]) {
+    // Ref: https://www.rfc-editor.org/rfc/rfc9562.html#section-5.5
+    // Assumes that digest was processed correctly with the expected namespace
+    for (int i = 0; i < 16; i++) {
+        uuid[i] = sha1_digest[i];
+    }
+
+    // Set bits corresponding to UUID ver 5
+    uuid[ 6] &= ~(0xF << 4);
+    uuid[ 6] |= (5 << 4);
+
+    // Set bits corresponding to UUID variant 0b10XX
+    uuid[ 8] &= ~(0xc << 4);
+    uuid[ 8] |= (0x8 << 4);
+}
+
+static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
+    const std::string & fname = hash_params.input;
+    struct ggml_context * ctx_data = NULL;
+
+    struct gguf_init_params params = {
+        /*.no_alloc = */ false,
+        /*.ctx      = */ &ctx_data,
+    };
+
+    // xxh64 init
+    XXH64_state_t* xxh64_model_hash_state = NULL;
+    if (hash_params.xxh64) {
+        xxh64_model_hash_state = XXH64_createState();
+        if (xxh64_model_hash_state==NULL) {
+            abort();
+        }
+
+        XXH64_hash_t const seed = 0;
+        if (XXH64_reset(xxh64_model_hash_state, seed) == XXH_ERROR) {
+            abort();
+        }
+    }
+
+    // sha1 init
+    SHA1_CTX sha1_model_hash_ctx;
+    if (hash_params.sha1) {
+        SHA1Init(&sha1_model_hash_ctx);
+    }
+
+    // sha256 init
+    sha256_t sha256_model_hash_ctx;
+    if (hash_params.sha256) {
+        sha256_init(&sha256_model_hash_ctx);
+    }
+
+    // sha1 for uuid init
+    SHA1_CTX sha1_for_uuid_ctx;
+    if (hash_params.uuid) {
+        unsigned char const uuidv5_namespace[] = {UUID_NAMESPACE_LLAMA_CPP_HEX};
+        SHA1Init(&sha1_for_uuid_ctx);
+        SHA1Update( &sha1_for_uuid_ctx, (unsigned char const *)uuidv5_namespace, sizeof(uuidv5_namespace));
+    }
+
+    struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
+    const int n_tensors = gguf_get_n_tensors(ctx);
+    bool tensor_layer_in_manifest = false;
+    bool model_in_manifest = false;
+    bool tensor_layer_has_mismatch = false;
+    bool model_has_mismatch = false;
+    for (int i = 0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(ctx, i);
+        struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
+        auto n_bytes = ggml_nbytes(cur);
+        auto *raw_data = cur->data;
+        const std::string tensor_layer_name = fname + ":" + name;
+
+        if (hash_params.xxh64) {
+
+            if (!hash_params.no_layer) {
+                // Per Layer Hash
+                XXH64_hash_t hash = XXH64(raw_data, n_bytes, 0);
+
+                char hex_result[17];
+                for (int  offset = 0; offset < 8; offset++) {
+                    unsigned int shift_bits_by = (8 * (8 - offset - 1));
+                    snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff);
+                }
+
+                if (hash_params.manifest_is_usable) {
+                    hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_XXH64_STR, hex_result, tensor_layer_name);
+
+                    switch (verify_result) {
+                        case HASH_MANIFEST_NOT_FOUND:
+                            break;
+                        case HASH_MANIFEST_MISMATCH:
+                            tensor_layer_in_manifest = true;
+                            tensor_layer_has_mismatch = true;
+                            break;
+                        case HASH_MANIFEST_OK:
+                            tensor_layer_in_manifest = true;
+                            break;
+                    }
+
+                    printf("%-8s  %-s  %s  -  %s\n", HASH_TYPE_XXH64_STR, hex_result, tensor_layer_name.c_str(), hash_manifest_result_to_str(verify_result));
+                } else {
+                    printf("%-8s  %-s  %s\n", HASH_TYPE_XXH64_STR, hex_result, tensor_layer_name.c_str());
+                }
+            }
+
+            // Overall Model Hash
+            if (XXH64_update(xxh64_model_hash_state, raw_data, n_bytes) == XXH_ERROR) abort();
+        }
+
+        if (hash_params.sha1) {
+
+            if (!hash_params.no_layer) {
+                // Per Layer Hash
+                char result[21]; // sha1 outputs 20 bytes
+                SHA1( result, (const char *)raw_data, n_bytes);
+
+                char hex_result[41] = {0};
+                for (int  offset = 0; offset < 20; offset++) {
+                    snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+                }
+
+                if (hash_params.manifest_is_usable) {
+                    hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA1_STR, hex_result, tensor_layer_name);
+
+                    switch (verify_result) {
+                        case HASH_MANIFEST_NOT_FOUND:
+                            break;
+                        case HASH_MANIFEST_MISMATCH:
+                            tensor_layer_in_manifest = true;
+                            tensor_layer_has_mismatch = true;
+                            break;
+                        case HASH_MANIFEST_OK:
+                            tensor_layer_in_manifest = true;
+                            break;
+                    }
+
+                    printf("%-8s  %-s  %s  -  %s\n", HASH_TYPE_SHA1_STR, hex_result, tensor_layer_name.c_str(), hash_manifest_result_to_str(verify_result));
+                } else {
+                    printf("%-8s  %-s  %s\n", HASH_TYPE_SHA1_STR, hex_result, tensor_layer_name.c_str());
+                }
+            }
+
+            // Overall Model Hash
+            SHA1Update( &sha1_model_hash_ctx, (unsigned char const *)raw_data, n_bytes);
+        }
+
+        if (hash_params.sha256) {
+
+            if (!hash_params.no_layer) {
+                // Per Layer Hash
+                unsigned char result[SHA256_DIGEST_SIZE]; // sha256 outputs 32 bytes
+                sha256_hash((unsigned char*) result, (const unsigned char *)raw_data, n_bytes);
+
+                char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0};
+                for (int  offset = 0; offset < SHA256_DIGEST_SIZE; offset++) {
+                    snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+                }
+
+                if (hash_params.manifest_is_usable) {
+                    hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA256_STR, hex_result, tensor_layer_name);
+
+                    switch (verify_result) {
+                        case HASH_MANIFEST_NOT_FOUND:
+                            break;
+                        case HASH_MANIFEST_MISMATCH:
+                            tensor_layer_in_manifest = true;
+                            tensor_layer_has_mismatch = true;
+                            break;
+                        case HASH_MANIFEST_OK:
+                            tensor_layer_in_manifest = true;
+                            break;
+                    }
+
+                    printf("%-8s  %-s  %s  -  %s\n", HASH_TYPE_SHA256_STR, hex_result, tensor_layer_name.c_str(), hash_manifest_result_to_str(verify_result));
+                } else {
+                    printf("%-8s  %-s  %s\n", HASH_TYPE_SHA256_STR, hex_result, tensor_layer_name.c_str());
+                }
+            }
+
+            // Overall Model Hash
+            sha256_update( &sha256_model_hash_ctx, (unsigned char const *)raw_data, n_bytes);
+        }
+
+        if (hash_params.uuid) {
+            SHA1Update( &sha1_for_uuid_ctx, (unsigned char const *)raw_data, n_bytes);
+        }
+    }
+
+    if (hash_params.xxh64) {
+        XXH64_hash_t const hash = XXH64_digest(xxh64_model_hash_state);
+
+        char hex_result[17];
+        for (int  offset = 0; offset < 8; offset++) {
+            unsigned int shift_bits_by = (8 * (8 - offset - 1));
+            snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff);
+        }
+
+        if (hash_params.manifest_is_usable) {
+            hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_XXH64_STR, hex_result, fname);
+
+            switch (verify_result) {
+                case HASH_MANIFEST_NOT_FOUND:
+                    break;
+                case HASH_MANIFEST_MISMATCH:
+                    model_in_manifest = true;
+                    model_has_mismatch = true;
+                    break;
+                case HASH_MANIFEST_OK:
+                    model_in_manifest = true;
+                    break;
+            }
+
+            printf("%-8s  %-s  %s  -  %s\n", HASH_TYPE_XXH64_STR, hex_result, fname.c_str(), hash_manifest_result_to_str(verify_result));
+        } else {
+            printf("%-8s  %-s  %s\n", HASH_TYPE_XXH64_STR, hex_result, fname.c_str());
+        }
+    }
+
+    if (hash_params.sha1) {
+        unsigned char result[21];
+        SHA1Final(result, &sha1_model_hash_ctx);
+
+        char hex_result[41];
+        for (int  offset = 0; offset < 20; offset++) {
+            snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+        }
+
+        if (hash_params.manifest_is_usable) {
+            hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA1_STR, hex_result, fname);
+
+            switch (verify_result) {
+                case HASH_MANIFEST_NOT_FOUND:
+                    break;
+                case HASH_MANIFEST_MISMATCH:
+                    model_in_manifest = true;
+                    model_has_mismatch = true;
+                    break;
+                case HASH_MANIFEST_OK:
+                    model_in_manifest = true;
+                    break;
+            }
+
+            printf("%-8s  %-s  %s  -  %s\n", HASH_TYPE_SHA1_STR, hex_result, fname.c_str(), hash_manifest_result_to_str(verify_result));
+        } else {
+            printf("%-8s  %-s  %s\n", HASH_TYPE_SHA1_STR, hex_result, fname.c_str());
+        }
+    }
+
+    if (hash_params.sha256) {
+        unsigned char result[SHA256_DIGEST_SIZE]; // sha256 outputs 32 bytes
+        sha256_final( &sha256_model_hash_ctx,  result);
+
+        char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0};
+        for (int  offset = 0; offset < SHA256_DIGEST_SIZE; offset++) {
+            snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+        }
+
+        if (hash_params.manifest_is_usable) {
+            hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA256_STR, hex_result, fname);
+
+            switch (verify_result) {
+                case HASH_MANIFEST_NOT_FOUND:
+                    break;
+                case HASH_MANIFEST_MISMATCH:
+                    model_in_manifest = true;
+                    model_has_mismatch = true;
+                    break;
+                case HASH_MANIFEST_OK:
+                    model_in_manifest = true;
+                    break;
+            }
+
+            printf("%-8s  %-s  %s  -  %s\n", HASH_TYPE_SHA256_STR, hex_result, fname.c_str(), hash_manifest_result_to_str(verify_result));
+        } else {
+            printf("%-8s  %-s  %s\n", HASH_TYPE_SHA256_STR, hex_result, fname.c_str());
+        }
+    }
+
+    if (hash_params.uuid) {
+        unsigned char result[21];
+        SHA1Final(result, &sha1_for_uuid_ctx);
+
+        unsigned char uuid[16];
+        generate_uuidv5(result, uuid);
+
+        char string_buffer[37] = {0};
+        snprintf(string_buffer, sizeof(string_buffer), "%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
+            uuid[0], uuid[1], uuid[2], uuid[3],
+            uuid[4], uuid[5], uuid[6], uuid[7],
+            uuid[8], uuid[9], uuid[10], uuid[11],
+            uuid[12], uuid[13], uuid[14], uuid[15]);
+
+        if (hash_params.manifest_is_usable) {
+            hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA256_STR, string_buffer, fname);
+
+            switch (verify_result) {
+                case HASH_MANIFEST_NOT_FOUND:
+                    break;
+                case HASH_MANIFEST_MISMATCH:
+                    model_in_manifest = true;
+                    model_has_mismatch = true;
+                    break;
+                case HASH_MANIFEST_OK:
+                    model_in_manifest = true;
+                    break;
+            }
+
+            printf("%-8s  %-s  %s  -  %s\n", HASH_TYPE_UUID_STR, string_buffer, fname.c_str(), hash_manifest_result_to_str(verify_result));
+        } else {
+            printf("%-8s  %-s  %s\n", HASH_TYPE_UUID_STR, string_buffer, fname.c_str());
+        }
+    }
+
+
+    ggml_free(ctx_data);
+    gguf_free(ctx);
+
+
+    if (hash_params.manifest_is_usable) {
+        // In hash verification mode
+
+        if (!model_in_manifest) {
+            // model missing in manifest?
+
+            // Check tensor layer...
+            if (!tensor_layer_in_manifest) {
+                // Still missing? Maybe we are reading the wrong manifest.
+                return HASH_EXIT_MANIFEST_MISSING_ENTRY;
+            }
+
+            if (tensor_layer_has_mismatch) {
+                // Per tensor check found error
+                return HASH_EXIT_FAILURE;
+            }
+
+            // All per tensor layer checks passed? Sounds good enough.
+            return HASH_EXIT_SUCCESS;
+        }
+
+        // Overall model check passed, but let's check per layer just in case
+        // If missing, we don't care too much as the overall model checked
+        if (tensor_layer_in_manifest && tensor_layer_has_mismatch) {
+            return HASH_EXIT_FAILURE;
+        }
+
+        if (model_has_mismatch) {
+            // model has failed hash somewhere in the model
+            return HASH_EXIT_FAILURE;
+        }
+
+        // All checks appears to be fine
+        return HASH_EXIT_SUCCESS;
+    }
+
+    // In hash generation mode
+    return HASH_EXIT_SUCCESS;
+}
+
+int main(int argc, const char ** argv) {
+    hash_params params;
+    manifest_check_params manifest_check;
+    hash_params_parse(argc, argv, params);
+
+    if (!params.manifest_file.empty()) {
+        if (!manifest_type(params.manifest_file, manifest_check)) {
+            printf("ERROR cannot open manifest %s", params.manifest_file.c_str());
+            return HASH_EXIT_MANIFEST_FILE_ERROR;
+        }
+
+        if (!manifest_check.sha256 && !manifest_check.sha1 && !manifest_check.xxh64 && !manifest_check.uuid) {
+            printf("ERROR manifest does not have any known hash format in %s", params.manifest_file.c_str());
+            return HASH_EXIT_MANIFEST_UNKNOWN_HASH;
+        }
+
+        printf("manifest  %s", params.manifest_file.c_str());
+
+        if (manifest_check.sha256) {
+            printf("  sha256");
+        }
+
+        if (manifest_check.sha1) {
+            printf("  sha1");
+        }
+
+        if (manifest_check.xxh64) {
+            printf("  xxh64");
+        }
+
+        if (manifest_check.uuid) {
+            printf("  uuid");
+        }
+
+        printf("\n");
+
+        // Autoselect the highest security hash if manifest is provided but
+        // the user has not specifically defined the hash they care about
+        if (!params.xxh64 && !params.sha1 && !params.uuid && !params.sha256) {
+            // User has not selected a specific value, pick most secure hash
+            if (manifest_check.sha256) {
+                params.sha256 = true;
+            } else if (manifest_check.sha1) {
+                params.sha1 = true;
+            } else if (manifest_check.xxh64) {
+                params.xxh64 = true;
+            } else if (manifest_check.uuid) {
+                params.uuid = true;
+            }
+        }
+
+        params.manifest_is_usable = true;
+    }
+
+    // By default if no swich argument provided, assume xxh64
+    if (!params.xxh64 && !params.sha1 && !params.uuid && !params.sha256) {
+        params.xxh64 = true;
+    }
+
+    hash_exit_code_t exit_code = gguf_hash(params);
+
+    if (params.manifest_is_usable) {
+        printf("\nVerification results for %s - %s\n", params.manifest_file.c_str(), hash_exit_code_to_str(exit_code));
+    }
+
+    return exit_code;
+}
diff --git a/llm/llama.cpp/examples/gguf-split/CMakeLists.txt b/llm/llama.cpp/examples/gguf-split/CMakeLists.txt
index 828e624352c8d5e7f0bdf71da5136f7ed715eee6..f63887da7dfcab0d463c71e71932bcec2e284e7b 100644
--- a/llm/llama.cpp/examples/gguf-split/CMakeLists.txt
+++ b/llm/llama.cpp/examples/gguf-split/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET gguf-split)
+set(TARGET llama-gguf-split)
 add_executable(${TARGET} gguf-split.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/gguf-split/gguf-split.cpp b/llm/llama.cpp/examples/gguf-split/gguf-split.cpp
index e04feeae3918841e224b631c99ff02fea14bd00f..881f0451c14559bfd11cbfbe5ee4630eeedbbc4e 100644
--- a/llm/llama.cpp/examples/gguf-split/gguf-split.cpp
+++ b/llm/llama.cpp/examples/gguf-split/gguf-split.cpp
@@ -61,10 +61,10 @@ static size_t split_str_to_n_bytes(std::string str) {
     int n;
     if (str.back() == 'M') {
         sscanf(str.c_str(), "%d", &n);
-        n_bytes = (size_t)n * 1024 * 1024; // megabytes
+        n_bytes = (size_t)n * 1000 * 1000; // megabytes
     } else if (str.back() == 'G') {
         sscanf(str.c_str(), "%d", &n);
-        n_bytes = (size_t)n * 1024 * 1024 * 1024; // gigabytes
+        n_bytes = (size_t)n * 1000 * 1000 * 1000; // gigabytes
     } else {
         throw std::invalid_argument("error: supported units are M (megabytes) or G (gigabytes), but got: " + std::string(1, str.back()));
     }
@@ -284,7 +284,7 @@ struct split_strategy {
                 struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_out, i));
                 total_size += ggml_nbytes(t);
             }
-            total_size = total_size / 1024 / 1024; // convert to megabytes
+            total_size = total_size / 1000 / 1000; // convert to megabytes
             printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
             i_split++;
         }
diff --git a/llm/llama.cpp/examples/gguf-split/tests.sh b/llm/llama.cpp/examples/gguf-split/tests.sh
index 7ca6fa7f20de88dfd08864d7b37c5b7d6591fed5..d5a92d6051063a078fab1aec14da716d3ae00691 100644
--- a/llm/llama.cpp/examples/gguf-split/tests.sh
+++ b/llm/llama.cpp/examples/gguf-split/tests.sh
@@ -18,8 +18,8 @@ fi
 
 set -x
 
-SPLIT=$1/gguf-split
-MAIN=$1/main
+SPLIT=$1/llama-gguf-split
+MAIN=$1/llama-cli
 WORK_PATH=$TMP_DIR/gguf-split
 ROOT_DIR=$(realpath $(dirname $0)/../../)
 
@@ -41,7 +41,7 @@ echo PASS
 echo
 
 # 2b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --random-prompt --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --n-predict 32
 echo PASS
 echo
 
@@ -51,7 +51,7 @@ echo PASS
 echo
 
 # 3b. Test the merged model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-merge.gguf --random-prompt --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-merge.gguf --n-predict 32
 echo PASS
 echo
 
@@ -61,7 +61,7 @@ echo PASS
 echo
 
 # 4b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --random-prompt --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --n-predict 32
 echo PASS
 echo
 
@@ -71,7 +71,7 @@ echo
 #echo
 
 # 5b. Test the merged model is loading properly
-#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --random-prompt --n-predict 32
+#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
 #echo PASS
 #echo
 
@@ -81,7 +81,7 @@ echo PASS
 echo
 
 # 6b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --random-prompt --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --n-predict 32
 echo PASS
 echo
 
diff --git a/llm/llama.cpp/examples/gguf/CMakeLists.txt b/llm/llama.cpp/examples/gguf/CMakeLists.txt
index 6481f087bc997f01a4287cc4ec959c4bee46b2b0..a9569b411956b31e741515ae3d34f1fa2a805594 100644
--- a/llm/llama.cpp/examples/gguf/CMakeLists.txt
+++ b/llm/llama.cpp/examples/gguf/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET gguf)
+set(TARGET llama-gguf)
 add_executable(${TARGET} gguf.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/gguf/gguf.cpp b/llm/llama.cpp/examples/gguf/gguf.cpp
index 57514377107d7cd4beef073a33e798850b6816db..7498f85efc4f90cc3bc819659ee555a81af117e3 100644
--- a/llm/llama.cpp/examples/gguf/gguf.cpp
+++ b/llm/llama.cpp/examples/gguf/gguf.cpp
@@ -92,6 +92,11 @@ static bool gguf_ex_read_0(const std::string & fname) {
 
     struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
 
+    if (!ctx) {
+        fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
     printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
     printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
     printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
diff --git a/llm/llama.cpp/examples/gritlm/CMakeLists.txt b/llm/llama.cpp/examples/gritlm/CMakeLists.txt
index ac4a5ae7937ea4cd584ee6802c0399de8c646582..86dfddca346fe70ba8d5127fc8013d4961bf21ca 100644
--- a/llm/llama.cpp/examples/gritlm/CMakeLists.txt
+++ b/llm/llama.cpp/examples/gritlm/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET gritlm)
+set(TARGET llama-gritlm)
 add_executable(${TARGET} gritlm.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/gritlm/README.md b/llm/llama.cpp/examples/gritlm/README.md
index a3a3c1389ca270f12f4bb670473c2d954bfd6ea7..786ba57363defd93fa8c51ac453a33a380e82ce2 100644
--- a/llm/llama.cpp/examples/gritlm/README.md
+++ b/llm/llama.cpp/examples/gritlm/README.md
@@ -26,7 +26,7 @@ $ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf --ou
 
 Run the example using the downloaded model:
 ```console
-$ ./gritlm -m models/gritlm-7b_q4_1.gguf
+$ ./llama-gritlm -m models/gritlm-7b_q4_1.gguf
 
 Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
 Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103
diff --git a/llm/llama.cpp/examples/gritlm/gritlm.cpp b/llm/llama.cpp/examples/gritlm/gritlm.cpp
index 52fd719b38ee56014c8d4a63842bb6865f66c8d3..2c61c2e1eb3bc80002ead3172fa0052f11f133c2 100644
--- a/llm/llama.cpp/examples/gritlm/gritlm.cpp
+++ b/llm/llama.cpp/examples/gritlm/gritlm.cpp
@@ -44,6 +44,7 @@ static std::vector> encode(llama_context * ctx, const std::ve
 
         // clear previous kv_cache values (irrelevant for embeddings)
         llama_kv_cache_clear(ctx);
+        llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -98,7 +99,9 @@ static std::string generate(llama_context * ctx, const std::string & prompt, boo
     llama_token eos_token = llama_token_eos(mdl);
 
     llama_kv_cache_clear(ctx);
+    llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
+
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
     std::vector inputs = llama_tokenize(mdl, prompt, false, true);
@@ -153,7 +156,9 @@ static std::string gritlm_instruction(const std::string & instruction) {
 
 int main(int argc, char * argv[]) {
     gpt_params params;
+
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -164,8 +169,7 @@ int main(int argc, char * argv[]) {
 
     llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
 
-    // create new context - set to embedding mode
-    cparams.embeddings = true;
+    // create generation context
     llama_context * ctx = llama_new_context_with_model(mdl, cparams);
 
     // ### Embedding/Representation ###
diff --git a/llm/llama.cpp/examples/imatrix/CMakeLists.txt b/llm/llama.cpp/examples/imatrix/CMakeLists.txt
index d688a162090496bb222dc39aa3acfc78d44998b1..d4c8265bdb9d2232ea095f9d1cf99ee51a2ffccd 100644
--- a/llm/llama.cpp/examples/imatrix/CMakeLists.txt
+++ b/llm/llama.cpp/examples/imatrix/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET imatrix)
+set(TARGET llama-imatrix)
 add_executable(${TARGET} imatrix.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/imatrix/README.md b/llm/llama.cpp/examples/imatrix/README.md
index 458c01b8751f17097a24623c00319ad7f65b577b..bb5faec94c20a6e6820c7c6921494fb5aaa94c4c 100644
--- a/llm/llama.cpp/examples/imatrix/README.md
+++ b/llm/llama.cpp/examples/imatrix/README.md
@@ -1,32 +1,35 @@
 # llama.cpp/examples/imatrix
 
-Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantum models.
+Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models.
 More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861
 
 ## Usage
 
 ```
-./imatrix -m  -f  [-o ] [--verbosity ]
-        [-ofreq num_chunks] [-ow <0 or 1>] [other common params]
+./llama-imatrix \
+    -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \
+    [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \
+    [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]
 ```
 
 Here `-m` with a model name and `-f` with a file containing training data (such as e.g. `wiki.train.raw`) are mandatory.
 The parameters in square brackets are optional and have the following meaning:
 * `-o` (or `--output-file`) specifies the name of the file where the computed data will be stored. If missing `imatrix.dat` is used.
 * `--verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
-* `-ofreq` (or `--output-frequency`) specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
-* `-ow` (or `--output-weight`) specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
+* `--output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
+* `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
+* `--process-output` specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
 
 For faster computation, make sure to use GPU offloading via the `-ngl` argument
 
 ## Example
 
 ```bash
-LLAMA_CUDA=1 make -j
+GGML_CUDA=1 make -j
 
 # generate importance matrix (imatrix.dat)
-./imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
+./llama-imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
 
 # use the imatrix to perform a Q4_K_M quantization
-./quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
+./llama-quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
 ```
diff --git a/llm/llama.cpp/examples/imatrix/imatrix.cpp b/llm/llama.cpp/examples/imatrix/imatrix.cpp
index 25a2351cc64d303a93a6a63bf9600d291011e774..58814b96e7d497fc3ffa42afb91b89078a07dfec 100644
--- a/llm/llama.cpp/examples/imatrix/imatrix.cpp
+++ b/llm/llama.cpp/examples/imatrix/imatrix.cpp
@@ -17,39 +17,37 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
+    LOG_TEE("\nexample usage:\n");
+    LOG_TEE("\n    %s \\\n"
+            "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \\\n"
+            "       [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
+            "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
+    LOG_TEE("\n");
+}
+
 struct Stats {
     std::vector values;
     std::vector counts;
     int ncall = 0;
 };
 
-struct StatParams {
-    std::string dataset;
-    std::string ofile = "imatrix.dat";
-    int         n_output_frequency = 10;
-    int         verbosity = 1;
-    int         keep_every = 0;
-    bool        collect_output_weight = false;
-};
-
 class IMatrixCollector {
 public:
     IMatrixCollector() = default;
-    void set_parameters(StatParams&& params) { m_params = std::move(params); }
+    void set_params(gpt_params params) { m_params = std::move(params); }
     bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
-    void save_imatrix() const;
-    bool load_imatrix(const char * file_name, bool add);
-    static bool load_imatrix(const char * file_name, std::unordered_map& imatrix);
+    void save_imatrix(int ncall = -1) const;
+    bool load_imatrix(const char * file_name);
 private:
     std::unordered_map m_stats;
-    StatParams                             m_params;
+    gpt_params                             m_params;
     std::mutex                             m_mutex;
     int                                    m_last_call = 0;
     std::vector                     m_src1_data;
     std::vector                      m_ids; // the expert ids from ggml_mul_mat_id
-                                                  //
-    void save_imatrix(const char * file_name, const char * dataset) const;
-    void keep_imatrix(int ncall) const;
 };
 
 // remove any prefix and suffixes from the name
@@ -85,7 +83,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         if (t->op != GGML_OP_MUL_MAT) return false;
         // why are small batches ignored (<16 tokens)?
         if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
-        if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
+        if (!(wname.substr(0, 4) == "blk." || (m_params.process_output && wname == "output.weight"))) return false;
         return true;
     }
 
@@ -129,7 +127,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         }
         else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
-            exit(1); //GGML_ASSERT(false);
+            exit(1); //GGML_ABORT("fatal error");
         }
         if (m_params.verbosity > 1) {
             printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
@@ -153,28 +151,32 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
                     for (int j = 0; j < (int)src1->ne[0]; ++j) {
                         e.values[e_start + j] += x[j]*x[j];
                         e.counts[e_start + j]++;
+                        if (!std::isfinite(e.values[e_start + j])) {
+                            fprintf(stderr, "%f detected in %s\n", e.values[e_start + j], wname.c_str());
+                            exit(1);
+                        }
                     }
                 }
             }
             if (e.ncall > m_last_call) {
                 m_last_call = e.ncall;
-                if (m_last_call % m_params.n_output_frequency == 0) {
+                if (m_last_call % m_params.n_out_freq == 0) {
                     save_imatrix();
                 }
-                if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
-                    keep_imatrix(m_last_call);
+                if (m_params.n_save_freq > 0 && m_last_call%m_params.n_save_freq == 0) {
+                    save_imatrix(m_last_call);
                 }
             }
         }
     } else {
-        auto& e = m_stats[wname];
+        auto & e = m_stats[wname];
         if (e.values.empty()) {
             e.values.resize(src1->ne[0], 0);
             e.counts.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
             fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
-            exit(1); //GGML_ASSERT(false);
+            exit(1); //GGML_ABORT("fatal error");
         }
         ++e.ncall;
         if (m_params.verbosity > 1) {
@@ -185,15 +187,19 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             for (int j = 0; j < (int)src1->ne[0]; ++j) {
                 e.values[j] += x[j]*x[j];
                 e.counts[j]++;
+                if (!std::isfinite(e.values[j])) {
+                    fprintf(stderr, "%f detected in %s\n", e.values[j], wname.c_str());
+                    exit(1);
+                }
             }
         }
         if (e.ncall > m_last_call) {
             m_last_call = e.ncall;
-            if (m_last_call % m_params.n_output_frequency == 0) {
+            if (m_last_call % m_params.n_out_freq == 0) {
                 save_imatrix();
             }
-            if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
-                keep_imatrix(m_last_call);
+            if (m_params.n_save_freq > 0 && m_last_call%m_params.n_save_freq == 0) {
+                save_imatrix(m_last_call);
             }
         }
     }
@@ -201,33 +207,75 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     return true;
 }
 
-void IMatrixCollector::save_imatrix() const {
-    save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str(), m_params.dataset.c_str());
-}
+void IMatrixCollector::save_imatrix(int ncall) const {
+    auto fname = m_params.out_file;
+    if (fname.empty()) {
+        fname = "imatrix.dat";
+    }
 
-void IMatrixCollector::keep_imatrix(int ncall) const {
-    auto file_name = m_params.ofile;
-    if (file_name.empty()) file_name = "imatrix.dat";
-    file_name += ".at_";
-    file_name += std::to_string(ncall);
-    save_imatrix(file_name.c_str(), m_params.dataset.c_str());
-}
+    if (ncall > 0) {
+        fname += ".at_";
+        fname += std::to_string(ncall);
+    }
+
+    // avoid writing imatrix entries that do not have full data
+    // this can happen with MoE models where some of the experts end up not being exercised by the provided training data
+
+    int n_entries = 0;
+    std::vector to_store;
+
+    bool is_first = true; // for printing
+    for (const auto & kv : m_stats) {
+        const int n_all = kv.second.counts.size();
+
+        if (n_all == 0) {
+            continue;
+        }
+
+        int n_zeros = 0;
+        for (const int c : kv.second.counts) {
+            if (c == 0) {
+                n_zeros++;
+            }
+        }
+
+        if (n_zeros != 0 && is_first) {
+            fprintf(stderr, "\n");
+            is_first = false;
+        }
+
+        if (n_zeros == n_all) {
+            fprintf(stderr, "%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
+            continue;
+        }
+
+        if (n_zeros > 0) {
+            fprintf(stderr, "%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
+            continue;
+        }
+
+        n_entries++;
+        to_store.push_back(kv.first);
+    }
+
+    if (to_store.size() < m_stats.size()) {
+        fprintf(stderr, "%s: warning: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
+    }
 
-void IMatrixCollector::save_imatrix(const char * fname, const char * dataset) const {
     std::ofstream out(fname, std::ios::binary);
-    int n_entries = m_stats.size();
     out.write((const char *) &n_entries, sizeof(n_entries));
-    for (const auto & p : m_stats) {
-        int len = p.first.size();
+    for (const auto & name : to_store) {
+        const auto & stat = m_stats.at(name);
+        int len = name.size();
         out.write((const char *) &len, sizeof(len));
-        out.write(p.first.c_str(), len);
-        out.write((const char *) &p.second.ncall, sizeof(p.second.ncall));
-        int nval = p.second.values.size();
+        out.write(name.c_str(), len);
+        out.write((const char *) &stat.ncall, sizeof(stat.ncall));
+        int nval = stat.values.size();
         out.write((const char *) &nval, sizeof(nval));
         if (nval > 0) {
             std::vector tmp(nval);
             for (int i = 0; i < nval; i++) {
-                tmp[i] = (p.second.values[i] / static_cast(p.second.counts[i])) * static_cast(p.second.ncall);
+                tmp[i] = (stat.values[i] / static_cast(stat.counts[i])) * static_cast(stat.ncall);
             }
             out.write((const char*)tmp.data(), nval*sizeof(float));
         }
@@ -236,26 +284,28 @@ void IMatrixCollector::save_imatrix(const char * fname, const char * dataset) co
     // Write the number of call the matrix was computed with
     out.write((const char *) &m_last_call, sizeof(m_last_call));
 
-    // Write the dataset name at the end of the file to later on specify it in quantize
-    int n_dataset = strlen(dataset);
-    out.write((const char *) &n_dataset, sizeof(n_dataset));
-    out.write(dataset, n_dataset);
+    // Write the input filename at the end of the file to later on specify it in quantize
+    {
+        int len = m_params.prompt_file.size();
+        out.write((const char *) &len, sizeof(len));
+        out.write(m_params.prompt_file.c_str(), len);
+    }
 
     if (m_params.verbosity > 0) {
-        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname);
+        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
     }
 }
 
-bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_map& imatrix_data) {
-    std::ifstream in(imatrix_file, std::ios::binary);
+bool IMatrixCollector::load_imatrix(const char * fname) {
+    std::ifstream in(fname, std::ios::binary);
     if (!in) {
-        printf("%s: failed to open %s\n",__func__,imatrix_file);
+        printf("%s: failed to open %s\n",__func__, fname);
         return false;
     }
     int n_entries;
     in.read((char*)&n_entries, sizeof(n_entries));
     if (in.fail() || n_entries < 1) {
-        printf("%s: no data in file %s\n", __func__, imatrix_file);
+        printf("%s: no data in file %s\n", __func__, fname);
         return false;
     }
     for (int i = 0; i < n_entries; ++i) {
@@ -263,23 +313,22 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
         std::vector name_as_vec(len+1);
         in.read((char *)name_as_vec.data(), len);
         if (in.fail()) {
-            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1,imatrix_file);
+            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
             return false;
         }
         name_as_vec[len] = 0;
         std::string name{name_as_vec.data()};
-        auto& e = imatrix_data[std::move(name)];
+        auto & e = m_stats[std::move(name)];
         int ncall;
         in.read((char*)&ncall, sizeof(ncall));
         int nval;
         in.read((char *)&nval, sizeof(nval));
         if (in.fail() || nval < 1) {
             printf("%s: failed reading number of values for entry %d\n",__func__,i);
-            imatrix_data = {};
+            m_stats = {};
             return false;
         }
 
-        // When re-called from load_imatrix() with add set, this will already be created.
         if (e.values.empty()) {
             e.values.resize(nval, 0);
             e.counts.resize(nval, 0);
@@ -289,7 +338,7 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
         in.read((char*)tmp.data(), nval*sizeof(float));
         if (in.fail()) {
             printf("%s: failed reading data for entry %d\n",__func__,i);
-            imatrix_data = {};
+            m_stats = {};
             return false;
         }
 
@@ -304,13 +353,6 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
     return true;
 }
 
-bool IMatrixCollector::load_imatrix(const char * file_name, bool add) {
-    if (!add) {
-        m_stats.clear();
-    }
-    return load_imatrix(file_name, m_stats);
-}
-
 static IMatrixCollector g_collector;
 
 static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
@@ -324,7 +366,7 @@ struct results_log_softmax {
     float  prob;
 };
 
-static std::vector softmax(const std::vector& logits) {
+static std::vector softmax(const std::vector & logits) {
     std::vector probs(logits.size());
     float max_logit = logits[0];
     for (float v : logits) {
@@ -358,8 +400,7 @@ static results_log_softmax log_softmax(int n_vocab, const float * logits, int to
 
 static void process_logits(
     int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers,
-    double & nll, double & nll2, float * logit_history, float * prob_history
-) {
+    double & nll, double & nll2, float * logit_history, float * prob_history) {
     std::mutex mutex;
     int counter = 0;
     auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
@@ -391,8 +432,7 @@ static void process_logits(
     }
 }
 
-static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl, int from_chunk) {
-
+static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
     GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
     const int n_ctx = llama_n_ctx(ctx);
@@ -405,13 +445,13 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
     auto tim2 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast(tim2-tim1).count());
 
-    if (from_chunk > 0) {
-        if (size_t((from_chunk + 2)*n_ctx) >= tokens.size()) {
-            fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, from_chunk);
+    if (params.i_chunk > 0) {
+        if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
+            fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
             return false;
         }
-        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, from_chunk, from_chunk*n_ctx);
-        tokens.erase(tokens.begin(), tokens.begin() + from_chunk*n_ctx);
+        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
+        tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);
     }
 
     if (int(tokens.size()) < 2*n_ctx) {
@@ -424,7 +464,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
     std::vector logit_history;
     std::vector prob_history;
 
-    if (compute_ppl) {
+    if (params.compute_ppl) {
         logit_history.resize(tokens.size());
         prob_history.resize(tokens.size());
     }
@@ -446,7 +486,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
     const int num_batches = (n_ctx + n_batch - 1) / n_batch;
 
     std::vector logits;
-    if (compute_ppl && num_batches > 1) {
+    if (params.compute_ppl && num_batches > 1) {
         logits.reserve((size_t)n_ctx * n_vocab);
     }
 
@@ -482,7 +522,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
             // restore the original token in case it was set to BOS
             tokens[batch_start] = token_org;
 
-            if (compute_ppl && num_batches > 1) {
+            if (params.compute_ppl && num_batches > 1) {
                 const auto * batch_logits = llama_get_logits(ctx);
                 logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
             }
@@ -501,7 +541,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
             fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
-        if (compute_ppl) {
+        if (params.compute_ppl) {
             const int first = n_ctx/2;
             const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
             process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
@@ -516,7 +556,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
     }
     printf("\n");
 
-    if (compute_ppl) {
+    if (params.compute_ppl) {
         nll2 /= count;
         nll /= count;
         const double ppl = exp(nll);
@@ -533,111 +573,32 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
 }
 
 int main(int argc, char ** argv) {
-
-    StatParams sparams;
-    std::string prev_result_file;
-    std::string combine_files;
-    bool compute_ppl = true;
-    int  from_chunk  = 0;
-    std::vector args;
-    args.push_back(argv[0]);
-    int iarg = 1;
-    for (; iarg < argc-1; ++iarg) {
-        std::string arg{argv[iarg]};
-        if (arg == "-o" || arg == "--output-file") {
-            sparams.ofile = argv[++iarg];
-        }
-        else if (arg == "-ofreq" || arg == "--output-frequency") {
-            sparams.n_output_frequency = std::stoi(argv[++iarg]);
-        }
-        else if (arg == "-ow" || arg == "--output-weight") {
-            sparams.collect_output_weight = std::stoi(argv[++iarg]);
-        }
-        else if (arg == "--verbosity") {
-            sparams.verbosity = std::stoi(argv[++iarg]);
-        } else if (arg == "--no-ppl") {
-            compute_ppl = false;
-        } else if (arg == "--keep-imatrix") {
-            sparams.keep_every = std::stoi(argv[++iarg]);
-        } else if (arg == "--continue-from") {
-            prev_result_file = argv[++iarg];
-        } else if (arg == "--combine") {
-            combine_files = argv[++iarg];
-        }
-        else if (arg == "--from-chunk") {
-            from_chunk = std::stoi(argv[++iarg]);
-        } else {
-            args.push_back(argv[iarg]);
-        }
-    }
-    if (iarg < argc) {
-        std::string arg{argv[iarg]};
-        if (arg == "--no-ppl") {
-            compute_ppl = false;
-        } else {
-            args.push_back(argv[iarg]);
-        }
-    }
-
     gpt_params params;
-    params.n_batch = 512;
-    if (!gpt_params_parse(args.size(), args.data(), params)) {
-        return 1;
-    }
 
+    params.n_ctx = 512;
     params.logits_all = true;
-    params.n_batch = std::min(params.n_batch, params.n_ctx);
+    params.verbosity = 1;
 
-    print_build_info();
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
     }
 
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-    if (params.random_prompt) {
-        params.prompt = string_random_prompt(rng);
-    }
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
 
-    sparams.dataset = params.prompt_file;
-    g_collector.set_parameters(std::move(sparams));
+    g_collector.set_params(params);
 
-    if (!combine_files.empty()) {
-        std::vector files;
-        size_t pos = 0;
-        while (true) {
-            auto new_pos = combine_files.find(',', pos);
-            if (new_pos != std::string::npos) {
-                files.emplace_back(combine_files.substr(pos, new_pos - pos));
-                pos = new_pos + 1;
-            } else {
-                files.emplace_back(combine_files.substr(pos));
-                break;
-            }
-        }
-        if (files.size() < 2) {
-            fprintf(stderr, "You must provide at least two comma separated files to use --combine\n");
+    for (const auto & in_file : params.in_files) {
+        printf("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
+        if (!g_collector.load_imatrix(in_file.c_str())) {
+            fprintf(stderr, "%s : failed to load %s\n", __func__, in_file.c_str());
             return 1;
         }
-        printf("Combining the following %d files\n", int(files.size()));
-        for (auto& file : files) {
-            printf("    %s\n", file.c_str());
-            if (!g_collector.load_imatrix(file.c_str(), true)) {
-                fprintf(stderr, "Failed to load %s\n", file.c_str());
-                return 1;
-            }
-        }
-        g_collector.save_imatrix();
-        return 0;
     }
 
-    if (!prev_result_file.empty()) {
-        if (!g_collector.load_imatrix(prev_result_file.c_str(), false)) {
-            fprintf(stderr, "=============== Failed to load %s\n", prev_result_file.c_str());
-            return 1;
-        }
+    if (params.in_files.size() > 1) {
+        printf("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
+        g_collector.save_imatrix();
     }
 
     llama_backend_init();
@@ -650,9 +611,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
@@ -670,8 +632,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
-    bool OK = compute_imatrix(ctx, params, compute_ppl, from_chunk);
-    if (!OK) {
+    if (!compute_imatrix(ctx, params)) {
         return 1;
     }
 
diff --git a/llm/llama.cpp/examples/infill/CMakeLists.txt b/llm/llama.cpp/examples/infill/CMakeLists.txt
index e4e8028da09da68ad31289b177e676a4f4b0d666..9b1aa3b63c920b1b28ade2bd500972d8df12161d 100644
--- a/llm/llama.cpp/examples/infill/CMakeLists.txt
+++ b/llm/llama.cpp/examples/infill/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET infill)
+set(TARGET llama-infill)
 add_executable(${TARGET} infill.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/infill/README.md b/llm/llama.cpp/examples/infill/README.md
index 6b076c8390abf8c3e61b08f7f60aac1d8942e592..810a0c5e76697ae4bafbfa7869f923eb38416882 100644
--- a/llm/llama.cpp/examples/infill/README.md
+++ b/llm/llama.cpp/examples/infill/README.md
@@ -15,6 +15,7 @@ In this section, we cover the most commonly used options for running the `infill
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
 -   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
 -   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `--spm-infill`: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.
 
 ## Input Prompts
 
@@ -42,5 +43,5 @@ scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.ggu
 ```
 
 ```bash
-./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
+./llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
 ```
diff --git a/llm/llama.cpp/examples/infill/infill.cpp b/llm/llama.cpp/examples/infill/infill.cpp
index 539f781847893254cdc438c7178ff556d81a91d9..92d630b15fdf1bbb4c2e92a02aed29b49087b5e4 100644
--- a/llm/llama.cpp/examples/infill/infill.cpp
+++ b/llm/llama.cpp/examples/infill/infill.cpp
@@ -107,6 +107,7 @@ int main(int argc, char ** argv) {
     g_params = ¶ms;
 
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -139,27 +140,6 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
     }
-    if (params.instruct) {
-        printf("\n************\n");
-        printf("%s: please use the 'main' tool for instruct mode\n", __func__);
-        printf("************\n\n");
-
-        return 0;
-    }
-    if (params.chatml) {
-        printf("\n************\n");
-        printf("%s: please use the 'main' tool for chatml mode\n", __func__);
-        printf("************\n\n");
-
-        return 0;
-    }
-    if (!params.antiprompt.empty()) {
-        printf("\n************\n");
-        printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
-        printf("************\n\n");
-
-        return 0;
-    }
     if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
         printf("\n************\n");
         printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
@@ -167,20 +147,6 @@ int main(int argc, char ** argv) {
 
         return 0;
     }
-    if (params.random_prompt) {
-        printf("\n************\n");
-        printf("%s: please use the 'main' tool for random prompt mode\n", __func__);
-        printf("************\n\n");
-
-        return 0;
-    }
-    if (!params.path_prompt_cache.empty()) {
-        printf("\n************\n");
-        printf("%s: infill does not support prompt caching\n", __func__);
-        printf("************\n\n");
-
-        return 0;
-    }
 
     if (params.rope_freq_base != 0.0) {
         LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
@@ -207,17 +173,16 @@ int main(int argc, char ** argv) {
 
     llama_model * model;
     llama_context * ctx;
-    llama_context * ctx_guidance = NULL;
+
     g_model = &model;
     g_ctx = &ctx;
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    if (sparams.cfg_scale > 1.f) {
-        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
-        ctx_guidance = llama_new_context_with_model(model, lparams);
-    }
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);
@@ -242,26 +207,28 @@ int main(int argc, char ** argv) {
     GGML_ASSERT(llama_add_eos_token(model) != 1);
     LOG("add_bos: %d\n", add_bos);
 
-    bool suff_rm_leading_spc = params.escape;
-    if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
-        params.input_suffix.erase(0, 1);
-        suff_rm_leading_spc = false;
-    }
     std::vector embd_inp;
+    std::vector embd_end;
     std::vector inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
     std::vector inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
-    const int space_token = 29871;
-    if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
-        inp_sfx.erase(inp_sfx.begin());
-    }
+
+    GGML_ASSERT(llama_token_prefix(model) >= 0);
+    GGML_ASSERT(llama_token_suffix(model) >= 0);
+
     inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+
+    embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
+    embd_end = params.spm_infill ? inp_pfx : inp_sfx;
     if (add_bos) {
-        inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
+        embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
+    }
+    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+
+    const llama_token middle_token = llama_token_middle(model);
+    if (middle_token >= 0) {
+        embd_inp.push_back(middle_token);
     }
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
-    embd_inp = inp_pfx;
-    embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-    embd_inp.push_back(llama_token_middle(model));
 
     LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
     LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
@@ -273,25 +240,6 @@ int main(int argc, char ** argv) {
         LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
-    // Tokenize negative prompt
-    std::vector guidance_inp;
-    int guidance_offset = 0;
-    int original_prompt_len = 0;
-    if (ctx_guidance) {
-        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
-
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, true);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
-
-        std::vector original_inp = ::llama_tokenize(ctx, params.prompt, true);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
-
-        original_prompt_len = original_inp.size();
-        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
-        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
-        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
-    }
-
     if ((int) embd_inp.size() > n_ctx - 4) {
         LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
@@ -319,15 +267,6 @@ int main(int argc, char ** argv) {
             LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
-        if (ctx_guidance) {
-            LOG_TEE("\n");
-            LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
-            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
-            for (int i = 0; i < (int) guidance_inp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
-            }
-        }
-
         if (params.n_keep > 0) {
         LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
@@ -395,12 +334,11 @@ int main(int argc, char ** argv) {
         is_interacting = params.interactive_first;
     }
 
-    bool input_echo           = true;
+    bool input_echo = true;
 
-    int n_past             = 0;
-    int n_remain           = params.n_predict;
-    int n_consumed         = 0;
-    int n_past_guidance    = 0;
+    int n_past     = 0;
+    int n_remain   = params.n_predict;
+    int n_consumed = 0;
 
     std::vector   input_tokens;  g_input_tokens  = &input_tokens;
     std::vector   output_tokens; g_output_tokens = &output_tokens;
@@ -410,7 +348,6 @@ int main(int argc, char ** argv) {
     console::set_display(console::prompt);
 
     std::vector embd;
-    std::vector embd_guidance;
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
 
@@ -436,7 +373,7 @@ int main(int argc, char ** argv) {
             // if we run out of context:
             // - take the n_keep first tokens from the original prompt (via n_past)
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-            if (n_past + (int) embd.size() + std::max(0, guidance_offset) > n_ctx) {
+            if (n_past + (int) embd.size() > n_ctx) {
                 if (params.n_predict == -2) {
                     LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                     break;
@@ -453,11 +390,7 @@ int main(int argc, char ** argv) {
 
                 n_past -= n_discard;
 
-                if (ctx_guidance) {
-                    n_past_guidance -= n_discard;
-                }
-
-                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+                LOG("after swap: n_past = %d\n", n_past);
 
                 LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
@@ -465,45 +398,6 @@ int main(int argc, char ** argv) {
 
             // evaluate tokens in batches
             // embd is typically prepared beforehand to fit within a batch, but not always
-
-            if (ctx_guidance) {
-                int input_size = 0;
-                llama_token * input_buf = NULL;
-
-                if (n_past_guidance < (int) guidance_inp.size()) {
-                    // Guidance context should have the same data with these modifications:
-                    //
-                    // * Replace the initial prompt
-                    // * Shift everything by guidance_offset
-                    embd_guidance = guidance_inp;
-                    if (embd.begin() + original_prompt_len < embd.end()) {
-                        embd_guidance.insert(
-                            embd_guidance.end(),
-                            embd.begin() + original_prompt_len,
-                            embd.end()
-                        );
-                    }
-
-                    input_buf  = embd_guidance.data();
-                    input_size = embd_guidance.size();
-
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
-                } else {
-                    input_buf  = embd.data();
-                    input_size = embd.size();
-                }
-
-                for (int i = 0; i < input_size; i += params.n_batch) {
-                    int n_eval = std::min(input_size - i, params.n_batch);
-                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
-                        LOG_TEE("%s : failed to eval\n", __func__);
-                        return 1;
-                    }
-
-                    n_past_guidance += n_eval;
-                }
-            }
-
             for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
                 int n_eval = (int) embd.size() - i;
                 if (n_eval > params.n_batch) {
@@ -525,11 +419,9 @@ int main(int argc, char ** argv) {
         }
 
         embd.clear();
-        embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-
-            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
+            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, nullptr);
 
             llama_sampling_accept(ctx_sampling, ctx, id, true);
 
@@ -583,7 +475,6 @@ int main(int argc, char ** argv) {
 
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
-
             // deal with eot token in infill mode
             if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if (is_interacting && !params.interactive_first) {
@@ -624,27 +515,26 @@ int main(int argc, char ** argv) {
                     string_process_escapes(params.input_prefix);
                     string_process_escapes(params.input_suffix);
                 }
-                suff_rm_leading_spc = params.escape;
-                if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
-                    params.input_suffix.erase(0, 1);
-                    suff_rm_leading_spc = false;
-                }
+
                 // tokenize new prefix and suffix
                 std::vector inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
                 std::vector inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
-                if (suff_rm_leading_spc && inp_sfx[0] == space_token) {
-                    inp_sfx.erase(inp_sfx.begin());
-                }
+
                 inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+
+                embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
+                embd_end = params.spm_infill ? inp_pfx : inp_sfx;
                 if (add_bos) {
-                    inp_pfx.insert(inp_pfx.begin(), llama_token_bos(model));
+                    embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
                 }
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
-                embd_inp = inp_pfx;
-                embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                embd_inp.push_back(llama_token_middle(model));
+                embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+
+                if (middle_token >= 0) {
+                    embd_inp.push_back(middle_token);
+                }
+
                 embd.clear();
-                embd_guidance.clear();
                 n_remain = params.n_predict;
                 n_past = 0;
                 n_consumed = 0;
@@ -751,7 +641,6 @@ int main(int argc, char ** argv) {
     llama_print_timings(ctx);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
-    if (ctx_guidance) { llama_free(ctx_guidance); }
     llama_free(ctx);
     llama_free_model(model);
 
@@ -764,4 +653,3 @@ int main(int argc, char ** argv) {
 
     return 0;
 }
-
diff --git a/llm/llama.cpp/examples/jeopardy/jeopardy.sh b/llm/llama.cpp/examples/jeopardy/jeopardy.sh
index 9bdbc755c13a7894fc0bc48af1d41acabbe6b254..07bcb3b8d78ac2077387d11aa73d7487fa757057 100644
--- a/llm/llama.cpp/examples/jeopardy/jeopardy.sh
+++ b/llm/llama.cpp/examples/jeopardy/jeopardy.sh
@@ -21,7 +21,7 @@ counter=1
 echo 'Running'
 while IFS= read -r question
 do
-  exe_cmd="./main -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\""
+  exe_cmd="./llama-cli -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\""
   echo $counter
   echo "Current Question: $question"
   eval "$exe_cmd"
diff --git a/llm/llama.cpp/examples/json_schema_pydantic_example.py b/llm/llama.cpp/examples/json_schema_pydantic_example.py
new file mode 100644
index 0000000000000000000000000000000000000000..19c0bdb5b677024a77528762b06bab9ca4453d1b
--- /dev/null
+++ b/llm/llama.cpp/examples/json_schema_pydantic_example.py
@@ -0,0 +1,82 @@
+# Usage:
+#! ./llama-server -m some-model.gguf &
+#! pip install pydantic
+#! python json_schema_pydantic_example.py
+
+from pydantic import BaseModel, Field, TypeAdapter
+from annotated_types import MinLen
+from typing import Annotated, List, Optional
+import json, requests
+
+if True:
+
+    def create_completion(*, response_model=None, endpoint="http://localhost:8080/v1/chat/completions", messages, **kwargs):
+        '''
+        Creates a chat completion using an OpenAI-compatible endpoint w/ JSON schema support
+        (llama.cpp server, llama-cpp-python, Anyscale / Together...)
+
+        The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below)
+        '''
+        response_format = None
+        type_adapter = None
+
+        if response_model:
+            type_adapter = TypeAdapter(response_model)
+            schema = type_adapter.json_schema()
+            messages = [{
+                "role": "system",
+                "content": f"You respond in JSON format with the following schema: {json.dumps(schema, indent=2)}"
+            }] + messages
+            response_format={"type": "json_object", "schema": schema}
+
+        data = requests.post(endpoint, headers={"Content-Type": "application/json"},
+                             json=dict(messages=messages, response_format=response_format, **kwargs)).json()
+        if 'error' in data:
+            raise Exception(data['error']['message'])
+
+        content = data["choices"][0]["message"]["content"]
+        return type_adapter.validate_json(content) if type_adapter else content
+
+else:
+
+    # This alternative branch uses Instructor + OpenAI client lib.
+    # Instructor support streamed iterable responses, retry & more.
+    # (see https://python.useinstructor.com/)
+    #! pip install instructor openai
+    import instructor, openai
+    client = instructor.patch(
+        openai.OpenAI(api_key="123", base_url="http://localhost:8080"),
+        mode=instructor.Mode.JSON_SCHEMA)
+    create_completion = client.chat.completions.create
+
+
+if __name__ == '__main__':
+
+    class QAPair(BaseModel):
+        class Config:
+            extra = 'forbid'  # triggers additionalProperties: false in the JSON schema
+        question: str
+        concise_answer: str
+        justification: str
+        stars: Annotated[int, Field(ge=1, le=5)]
+
+    class PyramidalSummary(BaseModel):
+        class Config:
+            extra = 'forbid'  # triggers additionalProperties: false in the JSON schema
+        title: str
+        summary: str
+        question_answers: Annotated[List[QAPair], MinLen(2)]
+        sub_sections: Optional[Annotated[List['PyramidalSummary'], MinLen(2)]]
+
+    print("# Summary\n", create_completion(
+        model="...",
+        response_model=PyramidalSummary,
+        messages=[{
+            "role": "user",
+            "content": f"""
+                You are a highly efficient corporate document summarizer.
+                Create a pyramidal summary of an imaginary internal document about our company processes
+                (starting high-level, going down to each sub sections).
+                Keep questions short, and answers even shorter (trivia / quizz style).
+            """
+        }]))
diff --git a/llm/llama.cpp/examples/json_schema_to_grammar.py b/llm/llama.cpp/examples/json_schema_to_grammar.py
index 826cd3f7271d146195af79010a8c83e0f6d091a0..a8779bf3bca38ee211dd5ef21f6327db91db78ae 100644
--- a/llm/llama.cpp/examples/json_schema_to_grammar.py
+++ b/llm/llama.cpp/examples/json_schema_to_grammar.py
@@ -1,89 +1,221 @@
 #!/usr/bin/env python3
+from __future__ import annotations
+
 import argparse
 import itertools
 import json
 import re
 import sys
-from typing import Any, Dict, List, Set, Tuple, Union
+from typing import Any, List, Optional, Set, Tuple, Union
 
-def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False):
-    if not separator_rule:
-        if min_items == 0 and max_items == 1:
-            return f'{item_rule}?'
-        elif min_items == 1 and max_items is None:
-            return f'{item_rule}+'
+def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
 
-    result = ''
+    if min_items == 0 and max_items == 1:
+        return f'{item_rule}?'
 
-    if min_items > 0:
-        if item_rule_is_literal and separator_rule is None:
-            result = '"' + (item_rule[1:-1] * min_items) + '"'
-        else:
-            result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items)
-
-    def opt_repetitions(up_to_n, prefix_with_sep=False):
-        '''
-            - n=4, no sep:             '(a (a (a (a)?)?)?)?'
-            - n=4, sep=',', prefix:    '("," a ("," a ("," a ("," a)?)?)?)?'
-            - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?'
-        '''
-
-        content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule
-        if up_to_n == 0:
-            return ''
-        elif up_to_n == 1:
-            return f'({content})?'
-        elif separator_rule and not prefix_with_sep:
-            return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?'
+    if not separator_rule:
+        if min_items == 1 and max_items is None:
+            return f'{item_rule}+'
+        elif min_items == 0 and max_items is None:
+            return f'{item_rule}*'
         else:
-            return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n)
+            return f'{item_rule}{{{min_items},{max_items if max_items is not None else ""}}}'
 
-    if min_items > 0 and max_items != min_items:
-        result += ' '
+    result = item_rule + ' ' + _build_repetition(f'({separator_rule} {item_rule})', min_items - 1 if min_items > 0 else 0, max_items - 1 if max_items is not None else None)
+    return f'({result})?' if min_items == 0 else result
 
-    if max_items is not None:
-        result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0)
-    else:
-        item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})'
+def _generate_min_max_int(min_value: Optional[int], max_value: Optional[int], out: list, decimals_left: int = 16, top_level: bool = True):
+    has_min = min_value != None
+    has_max = max_value != None
 
-        if min_items == 0 and separator_rule:
-            result = f'({item_rule} {item_operator}*)?'
+    def digit_range(from_char: str, to_char: str):
+        out.append("[")
+        if from_char == to_char:
+            out.append(from_char)
         else:
-            result += f'{item_operator}*'
-
-    return result
+            out.append(from_char)
+            out.append("-")
+            out.append(to_char)
+        out.append("]")
+
+    def more_digits(min_digits: int, max_digits: int):
+        out.append("[0-9]")
+        if min_digits == max_digits and min_digits == 1:
+            return
+        out.append("{")
+        out.append(str(min_digits))
+        if max_digits != min_digits:
+            out.append(",")
+            if max_digits != sys.maxsize:
+                out.append(str(max_digits))
+        out.append("}")
+
+    def uniform_range(from_str: str, to_str: str):
+        i = 0
+        while i < len(from_str) and from_str[i] == to_str[i]:
+            i += 1
+        if i > 0:
+            out.append("\"")
+            out.append(from_str[:i])
+            out.append("\"")
+        if i < len(from_str):
+            if i > 0:
+                out.append(" ")
+            sub_len = len(from_str) - i - 1
+            if sub_len > 0:
+                from_sub = from_str[i+1:]
+                to_sub = to_str[i+1:]
+                sub_zeros = "0" * sub_len
+                sub_nines = "9" * sub_len
+
+                to_reached = False
+                out.append("(")
+                if from_sub == sub_zeros:
+                    digit_range(from_str[i], chr(ord(to_str[i]) - 1))
+                    out.append(" ")
+                    more_digits(sub_len, sub_len)
+                else:
+                    out.append("[")
+                    out.append(from_str[i])
+                    out.append("] ")
+                    out.append("(")
+                    uniform_range(from_sub, sub_nines)
+                    out.append(")")
+                    if ord(from_str[i]) < ord(to_str[i]) - 1:
+                        out.append(" | ")
+                        if to_sub == sub_nines:
+                            digit_range(chr(ord(from_str[i]) + 1), to_str[i])
+                            to_reached = True
+                        else:
+                            digit_range(chr(ord(from_str[i]) + 1), chr(ord(to_str[i]) - 1))
+                        out.append(" ")
+                        more_digits(sub_len, sub_len)
+                if not to_reached:
+                    out.append(" | ")
+                    digit_range(to_str[i], to_str[i])
+                    out.append(" ")
+                    uniform_range(sub_zeros, to_sub)
+                out.append(")")
+            else:
+                out.append("[")
+                out.append(from_str[i])
+                out.append("-")
+                out.append(to_str[i])
+                out.append("]")
+
+    if has_min and has_max:
+        if min_value < 0 and max_value < 0:
+            out.append("\"-\" (")
+            _generate_min_max_int(-max_value, -min_value, out, decimals_left, top_level=True)
+            out.append(")")
+            return
+
+        if min_value < 0:
+            out.append("\"-\" (")
+            _generate_min_max_int(0, -min_value, out, decimals_left, top_level=True)
+            out.append(") | ")
+            min_value = 0
+
+        min_s = str(min_value)
+        max_s = str(max_value)
+        min_digits = len(min_s)
+        max_digits = len(max_s)
+
+        for digits in range(min_digits, max_digits):
+            uniform_range(min_s, "9" * digits)
+            min_s = "1" + "0" * digits
+            out.append(" | ")
+        uniform_range(min_s, max_s)
+        return
+
+    less_decimals = max(decimals_left - 1, 1)
+
+    if has_min:
+        if min_value < 0:
+            out.append("\"-\" (")
+            _generate_min_max_int(None, -min_value, out, decimals_left, top_level=False)
+            out.append(") | [0] | [1-9] ")
+            more_digits(0, decimals_left - 1)
+        elif min_value == 0:
+            if top_level:
+                out.append("[0] | [1-9] ")
+                more_digits(0, less_decimals)
+            else:
+                more_digits(1, decimals_left)
+        elif min_value <= 9:
+            c = str(min_value)
+            range_start = '1' if top_level else '0'
+            if c > range_start:
+                digit_range(range_start, chr(ord(c) - 1))
+                out.append(" ")
+                more_digits(1, less_decimals)
+                out.append(" | ")
+            digit_range(c, "9")
+            out.append(" ")
+            more_digits(0, less_decimals)
+        else:
+            min_s = str(min_value)
+            length = len(min_s)
+            c = min_s[0]
+
+            if c > "1":
+                digit_range("1" if top_level else "0", chr(ord(c) - 1))
+                out.append(" ")
+                more_digits(length, less_decimals)
+                out.append(" | ")
+            digit_range(c, c)
+            out.append(" (")
+            _generate_min_max_int(int(min_s[1:]), None, out, less_decimals, top_level=False)
+            out.append(")")
+            if c < "9":
+                out.append(" | ")
+                digit_range(chr(ord(c) + 1), "9")
+                out.append(" ")
+                more_digits(length - 1, less_decimals)
+        return
+
+    if has_max:
+        if max_value >= 0:
+            if top_level:
+                out.append("\"-\" [1-9] ")
+                more_digits(0, less_decimals)
+                out.append(" | ")
+            _generate_min_max_int(0, max_value, out, decimals_left, top_level=True)
+        else:
+            out.append("\"-\" (")
+            _generate_min_max_int(-max_value, None, out, decimals_left, top_level=False)
+            out.append(")")
+        return
 
+    raise RuntimeError("At least one of min_value or max_value must be set")
 
 class BuiltinRule:
-    def __init__(self, content: str, deps: list = None):
+    def __init__(self, content: str, deps: list | None = None):
         self.content = content
         self.deps = deps or []
 
-_up_to_15_digits = _build_repetition('[0-9]', 0, 15)
-
-# whitespace is constrained to a single space char to prevent model "running away" in
-# whitespace. Also maybe improves generation quality?
-SPACE_RULE = '" "?'
+# Constraining spaces to prevent model "running away".
+SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}'
 
 PRIMITIVE_RULES = {
     'boolean'      : BuiltinRule('("true" | "false") space', []),
-    'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []),
-    'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []),
+    'decimal-part' : BuiltinRule('[0-9]{1,16}', []),
+    'integral-part': BuiltinRule('[0] | [1-9] [0-9]{0,15}', []),
     'number'       : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
     'integer'      : BuiltinRule('("-"? integral-part) space', ['integral-part']),
     'value'        : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
     'object'       : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
     'array'        : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
-    'uuid'         : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []),
-    'char'         : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []),
+    'uuid'         : BuiltinRule(r'"\"" [0-9a-fA-F]{8} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{12} "\"" space', []),
+    'char'         : BuiltinRule(r'[^"\\\x7F\x00-\x1F] | [\\] (["\\bfnrt] | "u" [0-9a-fA-F]{4})', []),
     'string'       : BuiltinRule(r'"\"" char* "\"" space', ['char']),
     'null'         : BuiltinRule('"null" space', []),
 }
 
 # TODO: support "uri", "email" string formats
 STRING_FORMAT_RULES = {
-    'date'            : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
-    'time'            : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
+    'date'            : BuiltinRule('[0-9]{4} "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
+    'time'            : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9]{3} )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
     'date-time'       : BuiltinRule('date "T" time', ['date', 'time']),
     'date-string'     : BuiltinRule('"\\"" date "\\"" space', ['date']),
     'time-string'     : BuiltinRule('"\\"" time "\\"" space', ['time']),
@@ -101,7 +233,7 @@ GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
 GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
 
 NON_LITERAL_SET = set('|.()[]{}*+?')
-ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?')
+ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('^$.[]()|{}*+?')
 
 
 class SchemaConverter:
@@ -118,7 +250,7 @@ class SchemaConverter:
 
     def _format_literal(self, literal):
         escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
-            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)) or m.group(0), literal
         )
         return f'"{escaped}"'
 
@@ -145,6 +277,51 @@ class SchemaConverter:
 
         return ''.join(('(', *recurse(0), ')'))
 
+    def _not_strings(self, strings):
+        class TrieNode:
+            def __init__(self):
+                self.children = {}
+                self.is_end_of_string = False
+
+            def insert(self, string):
+                node = self
+                for c in string:
+                    node = node.children.setdefault(c, TrieNode())
+                node.is_end_of_string = True
+
+        trie = TrieNode()
+        for s in strings:
+            trie.insert(s)
+
+        char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
+        out = ['["] ( ']
+
+        def visit(node):
+            rejects = []
+            first = True
+            for c in sorted(node.children.keys()):
+                child = node.children[c]
+                rejects.append(c)
+                if first:
+                    first = False
+                else:
+                    out.append(' | ')
+                out.append(f'[{c}]')
+                if child.children:
+                    out.append(f' (')
+                    visit(child)
+                    out.append(')')
+                elif child.is_end_of_string:
+                    out.append(f' {char_rule}+')
+            if node.children:
+                if not first:
+                    out.append(' | ')
+                out.append(f'[^"{"".join(rejects)}] {char_rule}*')
+        visit(trie)
+
+        out.append(f' ){"" if trie.is_end_of_string else "?"} ["] space')
+        return ''.join(out)
+
     def _add_rule(self, name, rule):
         esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
         if esc_name not in self._rules or self._rules[esc_name] == rule:
@@ -228,11 +405,11 @@ class SchemaConverter:
         i = 0
         length = len(pattern)
 
-        def to_rule(s: Tuple[str, bool]) -> str:
+        def to_rule(s: tuple[str, bool]) -> str:
             (txt, is_literal) = s
             return "\"" + txt + "\"" if is_literal else txt
 
-        def transform() -> Tuple[str, bool]:
+        def transform() -> tuple[str, bool]:
             '''
                 Parse a unit at index i (advancing it), and return its string representation + whether it's a literal.
             '''
@@ -245,7 +422,7 @@ class SchemaConverter:
             # We only need a flat structure here to apply repetition operators to the last item, and
             # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
             # (GBNF's syntax is luckily very close to regular expressions!)
-            seq: list[Tuple[str, bool]] = []
+            seq: list[tuple[str, bool]] = []
 
             def get_dot():
                 if self._dotall:
@@ -333,7 +510,7 @@ class SchemaConverter:
                             sub_rule_ids[sub] = id
                         sub = id
 
-                    seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False)
+                    seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times), False)
                 else:
                     literal = ''
                     while i < length:
@@ -390,13 +567,13 @@ class SchemaConverter:
             return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf']))
 
         elif isinstance(schema_type, list):
-            return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type]))
+            return self._add_rule(rule_name, self._generate_union_rule(name, [{**schema, 'type': t} for t in schema_type]))
 
         elif 'const' in schema:
-            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']))
+            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']) + ' space')
 
         elif 'enum' in schema:
-            rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum']))
+            rule = '(' + ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) + ') space'
             return self._add_rule(rule_name, rule)
 
         elif schema_type in (None, 'object') and \
@@ -427,7 +604,7 @@ class SchemaConverter:
                 else:
                     add_component(t, is_required=True)
 
-            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[]))
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=None))
 
         elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
             items = schema.get('items') or schema['prefixItems']
@@ -465,6 +642,24 @@ class SchemaConverter:
 
             return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space')
 
+        elif schema_type in (None, 'integer') and \
+                ('minimum' in schema or 'exclusiveMinimum' in schema or 'maximum' in schema or 'exclusiveMaximum' in schema):
+            min_value = None
+            max_value = None
+            if 'minimum' in schema:
+                min_value = schema['minimum']
+            elif 'exclusiveMinimum' in schema:
+                min_value = schema['exclusiveMinimum'] + 1
+            if 'maximum' in schema:
+                max_value = schema['maximum']
+            elif 'exclusiveMaximum' in schema:
+                max_value = schema['exclusiveMaximum'] - 1
+
+            out = ["("]
+            _generate_min_max_int(min_value, max_value, out)
+            out.append(") space")
+            return self._add_rule(rule_name, ''.join(out))
+
         elif (schema_type == 'object') or (len(schema) == 0):
             return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object']))
 
@@ -483,7 +678,7 @@ class SchemaConverter:
                 self._add_primitive(dep, dep_rule)
         return n
 
-    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]):
+    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Optional[Union[bool, Any]]):
         prop_order = self._prop_order
         # sort by position in prop_order (if specified) then by original order
         sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))]
@@ -498,12 +693,16 @@ class SchemaConverter:
         required_props = [k for k in sorted_props if k in required]
         optional_props = [k for k in sorted_props if k not in required]
 
-        if additional_properties == True or isinstance(additional_properties, dict):
+        if additional_properties is not None and additional_properties != False:
             sub_name = f'{name}{"-" if name else ""}additional'
-            value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value')
+            value_rule = self.visit(additional_properties, f'{sub_name}-value') if isinstance(additional_properties, dict) else \
+                self._add_primitive('value', PRIMITIVE_RULES['value'])
+            key_rule = self._add_primitive('string', PRIMITIVE_RULES['string']) if not sorted_props \
+                else self._add_rule(f'{sub_name}-k', self._not_strings(sorted_props))
+
             prop_kv_rule_names["*"] = self._add_rule(
                 f'{sub_name}-kv',
-                self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
+                f'{key_rule} ":" space {value_rule}'
             )
             optional_props.append("*")
 
@@ -518,15 +717,11 @@ class SchemaConverter:
             def get_recursive_refs(ks, first_is_optional):
                 [k, *rest] = ks
                 kv_rule_name = prop_kv_rule_names[k]
-                if k == '*':
-                    res = self._add_rule(
-                        f'{name}{"-" if name else ""}additional-kvs',
-                        f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*'
-                    )
-                elif first_is_optional:
-                    res = f'( "," space {kv_rule_name} )?'
+                comma_ref = f'( "," space {kv_rule_name} )'
+                if first_is_optional:
+                    res = comma_ref + ('*' if k == '*' else '?')
                 else:
-                    res = kv_rule_name
+                    res = kv_rule_name + (' ' + comma_ref + "*" if k == '*' else '')
                 if len(rest) > 0:
                     res += ' ' + self._add_rule(
                         f'{name}{"-" if name else ""}{k}-rest',
@@ -556,7 +751,7 @@ class SchemaConverter:
 def main(args_in = None):
     parser = argparse.ArgumentParser(
         description='''
-            Generates a grammar (suitable for use in ./main) that produces JSON conforming to a
+            Generates a grammar (suitable for use in ./llama-cli) that produces JSON conforming to a
             given JSON schema. Only a subset of JSON schema features are supported; more may be
             added in the future.
         ''',
diff --git a/llm/llama.cpp/examples/llama-bench/README.md b/llm/llama.cpp/examples/llama-bench/README.md
index 8578405646af7f78bf368c6f4baed935614c2ab1..52b0e74d3dbf9908169872eb4fdd39ee94cdb834 100644
--- a/llm/llama.cpp/examples/llama-bench/README.md
+++ b/llm/llama.cpp/examples/llama-bench/README.md
@@ -1,4 +1,4 @@
-# llama.cpp/example/llama-bench
+# llama.cpp/examples/llama-bench
 
 Performance testing tool for llama.cpp.
 
@@ -162,7 +162,7 @@ $ ./llama-bench -o csv
 ```
 
 ```csv
-build_commit,build_number,cuda,opencl,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+build_commit,build_number,cuda,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
 "3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
 "3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
 ```
@@ -179,7 +179,6 @@ $ ./llama-bench -o json
     "build_commit": "3469684",
     "build_number": 1275,
     "cuda": true,
-    "opencl": false,
     "metal": false,
     "gpu_blas": true,
     "blas": true,
@@ -210,7 +209,6 @@ $ ./llama-bench -o json
     "build_commit": "3469684",
     "build_number": 1275,
     "cuda": true,
-    "opencl": false,
     "metal": false,
     "gpu_blas": true,
     "blas": true,
@@ -253,7 +251,6 @@ CREATE TABLE IF NOT EXISTS test (
   build_commit TEXT,
   build_number INTEGER,
   cuda INTEGER,
-  opencl INTEGER,
   metal INTEGER,
   gpu_blas INTEGER,
   blas INTEGER,
@@ -279,6 +276,6 @@ CREATE TABLE IF NOT EXISTS test (
   stddev_ts REAL
 );
 
-INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
-INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
 ```
diff --git a/llm/llama.cpp/examples/llama-bench/llama-bench.cpp b/llm/llama.cpp/examples/llama-bench/llama-bench.cpp
index c008904476d3e00891c8678c5807ef57e094bbf2..521fa8880b7c7e9b9cd6290105a3b9f6c38823b2 100644
--- a/llm/llama.cpp/examples/llama-bench/llama-bench.cpp
+++ b/llm/llama.cpp/examples/llama-bench/llama-bench.cpp
@@ -23,6 +23,10 @@
 #include "ggml-cuda.h"
 #include "ggml-sycl.h"
 
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
 // utils
 static uint64_t get_time_ns() {
     using clock = std::chrono::high_resolution_clock;
@@ -41,20 +45,6 @@ static std::string join(const std::vector & values, const std::string & delim
     return str.str();
 }
 
-template
-static std::vector split(const std::string & str, char delim) {
-    std::vector values;
-    std::istringstream str_stream(str);
-    std::string token;
-    while (std::getline(str_stream, token, delim)) {
-        T value;
-        std::istringstream token_stream(token);
-        token_stream >> value;
-        values.push_back(value);
-    }
-    return values;
-}
-
 template
 static std::vector transform_to_str(const std::vector & values, F f) {
     std::vector str_values;
@@ -134,30 +124,59 @@ static std::string get_gpu_info() {
             id += "/";
         }
     }
+#endif
+#ifdef GGML_USE_CANN
+    uint32_t count = ggml_backend_cann_get_device_count();
+    for (uint32_t i = 0; i < count; i++) {
+        char buf[128];
+        ggml_backend_cann_get_device_description(i, buf, sizeof(buf));
+        id += buf;
+        if (i < count - 1) {
+            id += "/";
+        }
+    }
 #endif
     // TODO: other backends
     return id;
 }
 
 // command line params
-enum output_formats {CSV, JSON, MARKDOWN, SQL};
+enum output_formats {NONE, CSV, JSON, MARKDOWN, SQL};
 
 static const char * output_format_str(output_formats format) {
     switch (format) {
+        case NONE:     return "none";
         case CSV:      return "csv";
         case JSON:     return "json";
         case MARKDOWN: return "md";
         case SQL:      return "sql";
-        default: GGML_ASSERT(!"invalid output format");
+        default: GGML_ABORT("invalid output format");
     }
 }
 
+static bool output_format_from_str(const std::string & s, output_formats & format) {
+    if (s == "none") {
+        format = NONE;
+    } else if (s == "csv") {
+        format = CSV;
+    } else if (s == "json") {
+        format = JSON;
+    } else if (s == "md") {
+        format = MARKDOWN;
+    } else if (s == "sql") {
+        format = SQL;
+    } else {
+        return false;
+    }
+    return true;
+}
+
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
         case LLAMA_SPLIT_MODE_NONE:  return "none";
         case LLAMA_SPLIT_MODE_LAYER: return "layer";
         case LLAMA_SPLIT_MODE_ROW:   return "row";
-        default: GGML_ASSERT(!"invalid split mode");
+        default: GGML_ABORT("invalid split mode");
     }
 }
 
@@ -190,31 +209,33 @@ struct cmd_params {
     int reps;
     bool verbose;
     output_formats output_format;
+    output_formats output_format_stderr;
 };
 
 static const cmd_params cmd_params_defaults = {
-    /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
-    /* n_prompt      */ {512},
-    /* n_gen         */ {128},
-    /* n_pg          */ {},
-    /* n_batch       */ {2048},
-    /* n_ubatch      */ {512},
-    /* type_k        */ {GGML_TYPE_F16},
-    /* type_v        */ {GGML_TYPE_F16},
-    /* n_threads     */ {cpu_get_num_math()},
-    /* n_gpu_layers  */ {99},
-    /* rpc_servers   */ {""},
-    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
-    /* main_gpu      */ {0},
-    /* no_kv_offload */ {false},
-    /* flash_attn    */ {false},
-    /* tensor_split  */ {std::vector(llama_max_devices(), 0.0f)},
-    /* use_mmap      */ {true},
-    /* embeddings    */ {false},
-    /* numa          */ GGML_NUMA_STRATEGY_DISABLED,
-    /* reps          */ 5,
-    /* verbose       */ false,
-    /* output_format */ MARKDOWN
+    /* model                */ {"models/7B/ggml-model-q4_0.gguf"},
+    /* n_prompt             */ {512},
+    /* n_gen                */ {128},
+    /* n_pg                 */ {},
+    /* n_batch              */ {2048},
+    /* n_ubatch             */ {512},
+    /* type_k               */ {GGML_TYPE_F16},
+    /* type_v               */ {GGML_TYPE_F16},
+    /* n_threads            */ {cpu_get_num_math()},
+    /* n_gpu_layers         */ {99},
+    /* rpc_servers          */ {""},
+    /* split_mode           */ {LLAMA_SPLIT_MODE_LAYER},
+    /* main_gpu             */ {0},
+    /* no_kv_offload        */ {false},
+    /* flash_attn           */ {false},
+    /* tensor_split         */ {std::vector(llama_max_devices(), 0.0f)},
+    /* use_mmap             */ {true},
+    /* embeddings           */ {false},
+    /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
+    /* reps                 */ 5,
+    /* verbose              */ false,
+    /* output_format        */ MARKDOWN,
+    /* output_format_stderr */ NONE,
 };
 
 static void print_usage(int /* argc */, char ** argv) {
@@ -243,6 +264,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -ts, --tensor-split     (default: 0)\n");
     printf("  -r, --repetitions                (default: %d)\n", cmd_params_defaults.reps);
     printf("  -o, --output       (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  -oe, --output-err  (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
@@ -284,7 +306,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
 
     params.verbose = cmd_params_defaults.verbose;
     params.output_format = cmd_params_defaults.output_format;
+    params.output_format_stderr = cmd_params_defaults.output_format_stderr;
     params.reps = cmd_params_defaults.reps;
+    params.numa = cmd_params_defaults.numa;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -300,28 +324,28 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.model.insert(params.model.end(), p.begin(), p.end());
         } else if (arg == "-p" || arg == "--n-prompt") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end());
         } else if (arg == "-n" || arg == "--n-gen") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
         } else if (arg == "-pg") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], ',');
+            auto p = string_split(argv[i], ',');
             if (p.size() != 2) {
                 invalid_param = true;
                 break;
@@ -332,21 +356,21 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
         } else if (arg == "-ub" || arg == "--ubatch-size") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
         } else if (arg == "-ctk" || arg == "--cache-type-k") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             std::vector types;
             for (const auto & t : p) {
                 ggml_type gt = ggml_type_from_name(t);
@@ -362,7 +386,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             std::vector types;
             for (const auto & t : p) {
                 ggml_type gt = ggml_type_from_name(t);
@@ -378,14 +402,14 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
         } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
         } else if (arg == "-rpc" || arg == "--rpc") {
             if (++i >= argc) {
@@ -398,7 +422,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             std::vector modes;
             for (const auto & m : p) {
                 llama_split_mode mode;
@@ -420,13 +444,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            params.main_gpu = split(argv[i], split_delim);
+            params.main_gpu = string_split(argv[i], split_delim);
         } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
         } else if (arg == "--numa") {
             if (++i >= argc) {
@@ -444,28 +468,28 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end());
         } else if (arg == "-mmp" || arg == "--mmap") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
         } else if (arg == "-embd" || arg == "--embeddings") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split(argv[i], split_delim);
+            auto p = string_split(argv[i], split_delim);
             params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
         } else if (arg == "-ts" || arg == "--tensor-split") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            for (auto ts : split(argv[i], split_delim)) {
+            for (auto ts : string_split(argv[i], split_delim)) {
                 // split string by ; and /
                 const std::regex regex{R"([;/]+)"};
                 std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
@@ -493,18 +517,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            if (argv[i] == std::string("csv")) {
-                params.output_format = CSV;
-            } else if (argv[i] == std::string("json")) {
-                params.output_format = JSON;
-            } else if (argv[i] == std::string("md")) {
-                params.output_format = MARKDOWN;
-            } else if (argv[i] == std::string("sql")) {
-                params.output_format = SQL;
-            } else {
+            invalid_param = !output_format_from_str(argv[i], params.output_format);
+        } else if (arg == "-oe" || arg == "--output-err") {
+            if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
+            invalid_param = !output_format_from_str(argv[i], params.output_format_stderr);
         } else if (arg == "-v" || arg == "--verbose") {
             params.verbose = true;
         } else {
@@ -706,12 +725,10 @@ struct test {
     static const std::string build_commit;
     static const int build_number;
     static const bool cuda;
-    static const bool opencl;
     static const bool vulkan;
     static const bool kompute;
     static const bool metal;
     static const bool sycl;
-    static const bool rpc;
     static const bool gpu_blas;
     static const bool blas;
     static const std::string cpu_info;
@@ -723,6 +740,7 @@ struct test {
     int n_batch;
     int n_ubatch;
     int n_threads;
+    bool has_rpc;
     ggml_type type_k;
     ggml_type type_v;
     int n_gpu_layers;
@@ -748,6 +766,7 @@ struct test {
         n_batch = inst.n_batch;
         n_ubatch = inst.n_ubatch;
         n_threads = inst.n_threads;
+        has_rpc = !inst.rpc_servers.empty();
         type_k = inst.type_k;
         type_v = inst.type_v;
         n_gpu_layers = inst.n_gpu_layers;
@@ -795,9 +814,6 @@ struct test {
         if (cuda) {
             return GGML_CUDA_NAME;
         }
-        if (opencl) {
-            return "OpenCL";
-        }
         if (vulkan) {
             return "Vulkan";
         }
@@ -810,9 +826,6 @@ struct test {
         if (sycl) {
             return GGML_SYCL_NAME;
         }
-        if (rpc) {
-            return "RPC";
-        }
         if (gpu_blas) {
             return "GPU BLAS";
         }
@@ -826,7 +839,7 @@ struct test {
     static const std::vector & get_fields() {
         static const std::vector fields = {
             "build_commit", "build_number",
-            "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas",
+            "cuda", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_ubatch",
@@ -852,7 +865,7 @@ struct test {
             field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
-        if (field == "cuda" || field == "opencl"  || field == "vulkan" || field == "kompute" || field == "metal" ||
+        if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
             field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
             return BOOL;
@@ -881,8 +894,8 @@ struct test {
         }
         std::vector values = {
             build_commit, std::to_string(build_number),
-            std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(vulkan),
-            std::to_string(metal), std::to_string(sycl), std::to_string(rpc), std::to_string(gpu_blas), std::to_string(blas),
+            std::to_string(cuda), std::to_string(vulkan), std::to_string(vulkan),
+            std::to_string(metal), std::to_string(sycl), std::to_string(has_rpc), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_ubatch),
@@ -910,14 +923,12 @@ struct test {
 const std::string test::build_commit = LLAMA_COMMIT;
 const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cuda();
-const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::kompute      = !!ggml_cpu_has_kompute();
 const bool        test::metal        = !!ggml_cpu_has_metal();
 const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
 const bool        test::blas         = !!ggml_cpu_has_blas();
 const bool        test::sycl         = !!ggml_cpu_has_sycl();
-const bool        test::rpc          = !!ggml_cpu_has_rpc();
 const std::string test::cpu_info     = get_cpu_info();
 const std::string test::gpu_info     = get_gpu_info();
 
@@ -1035,6 +1046,27 @@ struct markdown_printer : public printer {
         if (field == "n_gpu_layers") {
             return 3;
         }
+        if (field == "n_threads") {
+            return 7;
+        }
+        if (field == "n_batch") {
+            return 7;
+        }
+        if (field == "n_ubatch") {
+            return 8;
+        }
+        if (field == "type_k" || field == "type_v") {
+            return 6;
+        }
+        if (field == "split_mode") {
+            return 5;
+        }
+        if (field == "flash_attn") {
+            return 2;
+        }
+        if (field == "use_mmap") {
+            return 4;
+        }
         if (field == "test") {
             return 13;
         }
@@ -1162,6 +1194,9 @@ struct markdown_printer : public printer {
                 value = buf;
             } else if (field == "backend") {
                 value = test::get_backend();
+                if (t.has_rpc) {
+                    value += "+RPC";
+                }
             } else if (field == "test") {
                 if (t.n_prompt > 0 && t.n_gen == 0) {
                     snprintf(buf, sizeof(buf), "pp%d", t.n_prompt);
@@ -1278,6 +1313,22 @@ static void llama_null_log_callback(enum ggml_log_level level, const char * text
     (void) user_data;
 }
 
+static std::unique_ptr create_printer(output_formats format) {
+    switch (format) {
+        case NONE:
+            return nullptr;
+        case CSV:
+            return std::unique_ptr(new csv_printer());
+        case JSON:
+            return std::unique_ptr(new json_printer());
+        case MARKDOWN:
+            return std::unique_ptr(new markdown_printer());
+        case SQL:
+            return std::unique_ptr(new sql_printer());
+    }
+    GGML_ABORT("fatal error");
+}
+
 int main(int argc, char ** argv) {
     // try to set locale for unicode characters in markdown
     setlocale(LC_CTYPE, ".UTF-8");
@@ -1304,26 +1355,18 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // initialize printer
-    std::unique_ptr p;
-    switch (params.output_format) {
-        case CSV:
-            p.reset(new csv_printer());
-            break;
-        case JSON:
-            p.reset(new json_printer());
-            break;
-        case MARKDOWN:
-            p.reset(new markdown_printer());
-            break;
-        case SQL:
-            p.reset(new sql_printer());
-            break;
-        default:
-            assert(false);
-            exit(1);
+    std::unique_ptr p = create_printer(params.output_format);
+    std::unique_ptr p_err = create_printer(params.output_format_stderr);
+
+    if (p) {
+        p->fout = stdout;
+        p->print_header(params);
+    }
+
+    if (p_err) {
+        p_err->fout = stderr;
+        p_err->print_header(params);
     }
-    p->fout = stdout;
-    p->print_header(params);
 
     std::vector params_instances = get_cmd_params_instances(params);
 
@@ -1381,7 +1424,15 @@ int main(int argc, char ** argv) {
             t.samples_ns.push_back(t_ns);
         }
 
-        p->print_test(t);
+        if (p) {
+            p->print_test(t);
+            fflush(p->fout);
+        }
+
+        if (p_err) {
+            p_err->print_test(t);
+            fflush(p_err->fout);
+        }
 
         llama_print_timings(ctx);
 
@@ -1390,7 +1441,13 @@ int main(int argc, char ** argv) {
 
     llama_free_model(lmodel);
 
-    p->print_footer();
+    if (p) {
+        p->print_footer();
+    }
+
+    if (p_err) {
+        p_err->print_footer();
+    }
 
     llama_backend_free();
 
diff --git a/llm/llama.cpp/examples/llama.android/app/build.gradle.kts b/llm/llama.cpp/examples/llama.android/app/build.gradle.kts
new file mode 100644
index 0000000000000000000000000000000000000000..8d1b37195efd40a8336391d80b4d90064dba6ee0
--- /dev/null
+++ b/llm/llama.cpp/examples/llama.android/app/build.gradle.kts
@@ -0,0 +1,65 @@
+plugins {
+    id("com.android.application")
+    id("org.jetbrains.kotlin.android")
+}
+
+android {
+    namespace = "com.example.llama"
+    compileSdk = 34
+
+    defaultConfig {
+        applicationId = "com.example.llama"
+        minSdk = 33
+        targetSdk = 34
+        versionCode = 1
+        versionName = "1.0"
+
+        testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
+        vectorDrawables {
+            useSupportLibrary = true
+        }
+    }
+
+    buildTypes {
+        release {
+            isMinifyEnabled = false
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro"
+            )
+        }
+    }
+    compileOptions {
+        sourceCompatibility = JavaVersion.VERSION_1_8
+        targetCompatibility = JavaVersion.VERSION_1_8
+    }
+    kotlinOptions {
+        jvmTarget = "1.8"
+    }
+    buildFeatures {
+        compose = true
+    }
+    composeOptions {
+        kotlinCompilerExtensionVersion = "1.5.1"
+    }
+}
+
+dependencies {
+
+    implementation("androidx.core:core-ktx:1.12.0")
+    implementation("androidx.lifecycle:lifecycle-runtime-ktx:2.6.2")
+    implementation("androidx.activity:activity-compose:1.8.2")
+    implementation(platform("androidx.compose:compose-bom:2023.08.00"))
+    implementation("androidx.compose.ui:ui")
+    implementation("androidx.compose.ui:ui-graphics")
+    implementation("androidx.compose.ui:ui-tooling-preview")
+    implementation("androidx.compose.material3:material3")
+    implementation(project(":llama"))
+    testImplementation("junit:junit:4.13.2")
+    androidTestImplementation("androidx.test.ext:junit:1.1.5")
+    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
+    androidTestImplementation(platform("androidx.compose:compose-bom:2023.08.00"))
+    androidTestImplementation("androidx.compose.ui:ui-test-junit4")
+    debugImplementation("androidx.compose.ui:ui-tooling")
+    debugImplementation("androidx.compose.ui:ui-test-manifest")
+}
diff --git a/llm/llama.cpp/examples/llama.android/build.gradle.kts b/llm/llama.cpp/examples/llama.android/build.gradle.kts
new file mode 100644
index 0000000000000000000000000000000000000000..acd1ada7d9b1a4e636f2d0a5a93653d66f8cb954
--- /dev/null
+++ b/llm/llama.cpp/examples/llama.android/build.gradle.kts
@@ -0,0 +1,6 @@
+// Top-level build file where you can add configuration options common to all sub-projects/modules.
+plugins {
+    id("com.android.application") version "8.2.0" apply false
+    id("org.jetbrains.kotlin.android") version "1.9.0" apply false
+    id("com.android.library") version "8.2.0" apply false
+}
diff --git a/llm/llama.cpp/examples/llama.android/llama/build.gradle.kts b/llm/llama.cpp/examples/llama.android/llama/build.gradle.kts
new file mode 100644
index 0000000000000000000000000000000000000000..0a3806172f05f7ee1ce3cae67535c042d2c40080
--- /dev/null
+++ b/llm/llama.cpp/examples/llama.android/llama/build.gradle.kts
@@ -0,0 +1,68 @@
+plugins {
+    id("com.android.library")
+    id("org.jetbrains.kotlin.android")
+}
+
+android {
+    namespace = "android.llama.cpp"
+    compileSdk = 34
+
+    defaultConfig {
+        minSdk = 33
+
+        testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
+        consumerProguardFiles("consumer-rules.pro")
+        ndk {
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
+        }
+        externalNativeBuild {
+            cmake {
+                arguments += "-DCMAKE_BUILD_TYPE=Release"
+                cppFlags += listOf()
+                arguments += listOf()
+
+                cppFlags("")
+            }
+        }
+    }
+
+    buildTypes {
+        release {
+            isMinifyEnabled = false
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro"
+            )
+        }
+    }
+    externalNativeBuild {
+        cmake {
+            path("src/main/cpp/CMakeLists.txt")
+            version = "3.22.1"
+        }
+    }
+    compileOptions {
+        sourceCompatibility = JavaVersion.VERSION_1_8
+        targetCompatibility = JavaVersion.VERSION_1_8
+    }
+    kotlinOptions {
+        jvmTarget = "1.8"
+    }
+
+    packaging {
+        resources {
+            excludes += "/META-INF/{AL2.0,LGPL2.1}"
+        }
+    }
+}
+
+dependencies {
+
+    implementation("androidx.core:core-ktx:1.12.0")
+    implementation("androidx.appcompat:appcompat:1.6.1")
+    implementation("com.google.android.material:material:1.11.0")
+    testImplementation("junit:junit:4.13.2")
+    androidTestImplementation("androidx.test.ext:junit:1.1.5")
+    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
+}
diff --git a/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/CMakeLists.txt b/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/CMakeLists.txt
index 42ebaad49a560cf6bb7bcaeca86f2e114c7d79f0..2de496574f54abfc54eddf3e4c47ab2f09d9231b 100644
--- a/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/CMakeLists.txt
+++ b/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/CMakeLists.txt
@@ -11,15 +11,15 @@ cmake_minimum_required(VERSION 3.22.1)
 # build script scope).
 project("llama-android")
 
-include(FetchContent)
-FetchContent_Declare(
-        llama
-        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
-        GIT_TAG        master
-)
+#include(FetchContent)
+#FetchContent_Declare(
+#        llama
+#        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
+#        GIT_TAG        master
+#)
 
 # Also provides "common"
-FetchContent_MakeAvailable(llama)
+#FetchContent_MakeAvailable(llama)
 
 # Creates and names a library, sets it as either STATIC
 # or SHARED, and provides the relative paths to its source code.
@@ -30,6 +30,10 @@ FetchContent_MakeAvailable(llama)
 # the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
 # is preferred for the same purpose.
 #
+
+#load local llama.cpp
+add_subdirectory(../../../../../../ build-llama)
+
 # In order to load a library into your app from Java/Kotlin, you must call
 # System.loadLibrary() and pass the name of the library defined here;
 # for GameActivity/NativeActivity derived applications, the same library name must be
diff --git a/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp b/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp
index 874158ef0f98f5cd20a39d868fa1c867d605cff9..2aafe23167557f66e16c82fed9d86ca8494288cd 100644
--- a/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp
+++ b/llm/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp
@@ -5,7 +5,7 @@
 #include 
 #include 
 #include "llama.h"
-#include "common/common.h"
+#include "common.h"
 
 // Write C++ code here.
 //
@@ -409,7 +409,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 
     const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
     if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
-        return env->NewStringUTF("");
+        return nullptr;
     }
 
     auto new_token_chars = llama_token_to_piece(context, new_token_id);
diff --git a/llm/llama.cpp/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/llm/llama.cpp/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
index 737f882fb2d2eeb2b7c6ea3b3639e0f8cd0f2f0d..58c32ca533bb10039055cf8135502a41c9287361 100644
--- a/llm/llama.cpp/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
+++ b/llm/llama.cpp/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
@@ -26,11 +26,12 @@ actor LlamaContext {
     private var context: OpaquePointer
     private var batch: llama_batch
     private var tokens_list: [llama_token]
+    var is_done: Bool = false
 
     /// This variable is used to store temporarily invalid cchars
     private var temporary_invalid_cchars: [CChar]
 
-    var n_len: Int32 = 64
+    var n_len: Int32 = 1024
     var n_cur: Int32 = 0
 
     var n_decode: Int32 = 0
@@ -160,6 +161,7 @@ actor LlamaContext {
 
         if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
             print("\n")
+            is_done = true
             let new_token_str = String(cString: temporary_invalid_cchars + [0])
             temporary_invalid_cchars.removeAll()
             return new_token_str
@@ -322,7 +324,7 @@ actor LlamaContext {
         defer {
             result.deallocate()
         }
-        let nTokens = llama_token_to_piece(model, token, result, 8, false)
+        let nTokens = llama_token_to_piece(model, token, result, 8, 0, false)
 
         if nTokens < 0 {
             let newResult = UnsafeMutablePointer.allocate(capacity: Int(-nTokens))
@@ -330,7 +332,7 @@ actor LlamaContext {
             defer {
                 newResult.deallocate()
             }
-            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, false)
+            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, 0, false)
             let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
             return Array(bufferPointer)
         } else {
diff --git a/llm/llama.cpp/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift b/llm/llama.cpp/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift
index 5bde1891727ce5d4cdf0c0eaa4561568149bc339..b8f6a31d582cdc9b87dfccaca0d471e3eee2ef13 100644
--- a/llm/llama.cpp/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift
+++ b/llm/llama.cpp/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift
@@ -131,22 +131,29 @@ class LlamaState: ObservableObject {
 
         messageLog += "\(text)"
 
-        while await llamaContext.n_cur < llamaContext.n_len {
-            let result = await llamaContext.completion_loop()
-            messageLog += "\(result)"
-        }
+        Task.detached {
+            while await !llamaContext.is_done {
+                let result = await llamaContext.completion_loop()
+                await MainActor.run {
+                    self.messageLog += "\(result)"
+                }
+            }
 
-        let t_end = DispatchTime.now().uptimeNanoseconds
-        let t_generation = Double(t_end - t_heat_end) / NS_PER_S
-        let tokens_per_second = Double(await llamaContext.n_len) / t_generation
+            let t_end = DispatchTime.now().uptimeNanoseconds
+            let t_generation = Double(t_end - t_heat_end) / self.NS_PER_S
+            let tokens_per_second = Double(await llamaContext.n_len) / t_generation
 
-        await llamaContext.clear()
-        messageLog += """
-            \n
-            Done
-            Heat up took \(t_heat)s
-            Generated \(tokens_per_second) t/s\n
-            """
+            await llamaContext.clear()
+
+            await MainActor.run {
+                self.messageLog += """
+                    \n
+                    Done
+                    Heat up took \(t_heat)s
+                    Generated \(tokens_per_second) t/s\n
+                    """
+            }
+        }
     }
 
     func bench() async {
diff --git a/llm/llama.cpp/examples/llava/CMakeLists.txt b/llm/llama.cpp/examples/llava/CMakeLists.txt
index 2985caff8379a1684eff84b40452985dc6dac7b3..e9fa73acb097b7ad078a03eafe52b617dce958a8 100644
--- a/llm/llama.cpp/examples/llava/CMakeLists.txt
+++ b/llm/llama.cpp/examples/llava/CMakeLists.txt
@@ -30,8 +30,9 @@ if(TARGET BUILD_INFO)
     add_dependencies(llava BUILD_INFO)
 endif()
 
-set(TARGET llava-cli)
-add_executable(llava-cli llava-cli.cpp)
-install(TARGETS llava-cli RUNTIME)
-target_link_libraries(llava-cli PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(llava PRIVATE cxx_std_11)
+set(TARGET llama-llava-cli)
+add_executable(${TARGET} llava-cli.cpp)
+set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/llm/llama.cpp/examples/llava/MobileVLM-README.md b/llm/llama.cpp/examples/llava/MobileVLM-README.md
index 74f021dec5e175fb6df423119c8dd3b25ca17ef6..06a65fba4787a7562ae92af36d77ab8b89a2c329 100644
--- a/llm/llama.cpp/examples/llava/MobileVLM-README.md
+++ b/llm/llama.cpp/examples/llava/MobileVLM-README.md
@@ -9,12 +9,12 @@ The implementation is based on llava, and is compatible with llava and mobileVLM
 Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using **MobileVLM-1.7B** as an example, the different conversion step will be shown.
 
 ## Usage
-Build with cmake or run `make llava-cli` to build it.
+Build with cmake or run `make llama-llava-cli` to build it.
 
-After building, run: `./llava-cli` to see the usage. For example:
+After building, run: `./llama-llava-cli` to see the usage. For example:
 
 ```sh
-./llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
+./llama-llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
     --mmproj MobileVLM-1.7B/mmproj-model-f16.gguf \
     --image path/to/an/image.jpg \
     -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWho is the author of this book? Answer the question using a single word or phrase. ASSISTANT:"
@@ -30,16 +30,16 @@ git clone https://huggingface.co/mtgv/MobileVLM-1.7B
 git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 ```
 
-2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+2. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
 ```sh
-python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B
+python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
+3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
 
 ```sh
-python ./examples/llava/convert-image-encoder-to-gguf \
+python ./examples/llava/convert_image_encoder_to_gguf \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B/llava.projector \
     --output-dir path/to/MobileVLM-1.7B \
@@ -47,22 +47,22 @@ python ./examples/llava/convert-image-encoder-to-gguf \
 ```
 
 ```sh
-python ./examples/llava/convert-image-encoder-to-gguf \
+python ./examples/llava/convert_image_encoder_to_gguf \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
     --output-dir path/to/MobileVLM-1.7B_V2 \
     --projector-type ldpv2
 ```
 
-4. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF:
+4. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./examples/convert-legacy-llama.py path/to/MobileVLM-1.7B
+python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B
 ```
 
 5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k`
 ```sh
-./quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
+./llama-quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
 ```
 
 Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory.
@@ -82,7 +82,7 @@ refer to `android/adb_run.sh`, modify resources' `name` and `path`
 ### case 1
 **input**
 ```sh
-/data/local/tmp/llava-cli \
+/data/local/tmp/llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -102,7 +102,7 @@ llama_print_timings:       total time =   34731.93 ms
 ### case 2
 **input**
 ```sh
-/data/local/tmp/llava-cli \
+/data/local/tmp/llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -126,7 +126,7 @@ llama_print_timings:       total time =   34570.79 ms
 #### llava-cli release-b2005
 **input**
 ```sh
-/data/local/tmp/llava-cli \
+/data/local/tmp/llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -194,13 +194,13 @@ llama_print_timings:       total time =   44411.01 ms /   377 tokens
 ## Orin compile and run
 ### compile
 ```sh
-make LLAMA_CUDA=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
+make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
 ```
 ### run on Orin
 ### case 1
 **input**
 ```sh
-./llava-cli \
+./llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     --image /data/local/tmp/demo.jpeg \
@@ -224,7 +224,7 @@ llama_print_timings:       total time =    1352.63 ms /   252 tokens
 ### case 2
 **input**
 ```sh
-./llava-cli \
+./llama-llava-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWhat is in the image? ASSISTANT:" \
diff --git a/llm/llama.cpp/examples/llava/README.md b/llm/llama.cpp/examples/llava/README.md
index 8d1ae5270e458eb13b1d26aaa198016cb1221d5c..012451361763c46b9e50042a44e6c68d68afc9af 100644
--- a/llm/llama.cpp/examples/llava/README.md
+++ b/llm/llama.cpp/examples/llava/README.md
@@ -11,12 +11,12 @@ For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](h
 After API is confirmed, more models will be supported / uploaded.
 
 ## Usage
-Build with cmake or run `make llava-cli` to build it.
+Build with cmake or run `make llama-llava-cli` to build it.
 
-After building, run: `./llava-cli` to see the usage. For example:
+After building, run: `./llama-llava-cli` to see the usage. For example:
 
 ```sh
-./llava-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
+./llama-llava-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
@@ -38,22 +38,22 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 pip install -r examples/llava/requirements.txt
 ```
 
-3. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
 ```sh
-python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b
+python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
 ```
 
-4. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
+4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:
 
 ```sh
-python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
+python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```
 
-5. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF:
+5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./examples/convert-legacy-llama.py ../llava-v1.5-7b --skip-unknown
+python ./examples/convert_legacy_llama.py ../llava-v1.5-7b --skip-unknown
 ```
 
 Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory.
@@ -70,9 +70,9 @@ git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 pip install -r examples/llava/requirements.txt
 ```
 
-3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
-python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
+python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory
 
@@ -86,18 +86,18 @@ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.jso
 
 5) Create the visual gguf model:
 ```console
-python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
+python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
 
 6) Then convert the model to gguf format:
 ```console
-python ./examples/convert-legacy-llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown
+python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```
 
-7) And finally we can run the llava-cli using the 1.6 model version:
+7) And finally we can run the llava cli using the 1.6 model version:
 ```console
-./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
+./llama-llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
 ```
 
 **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
diff --git a/llm/llama.cpp/examples/llava/android/adb_run.sh b/llm/llama.cpp/examples/llava/android/adb_run.sh
index f73623ae3b129e0fcd444d69ae1945b72880feb8..45ccf8d70d863580dabdcdd6463d6977c9b6f32d 100644
--- a/llm/llama.cpp/examples/llava/android/adb_run.sh
+++ b/llm/llama.cpp/examples/llava/android/adb_run.sh
@@ -10,7 +10,7 @@ prompt="A chat between a curious user and an artificial intelligence assistant.
 # prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWhat is in the image? ASSISTANT:"
 
 program_dir="build_64/bin"
-binName="llava-cli"
+binName="llama-llava-cli"
 n_threads=4
 
 
diff --git a/llm/llama.cpp/examples/llava/android/build_64.sh b/llm/llama.cpp/examples/llava/android/build_64.sh
new file mode 100644
index 0000000000000000000000000000000000000000..71b6fd3f719cd4c5d33f5cdbdd72937c9f3a2810
--- /dev/null
+++ b/llm/llama.cpp/examples/llava/android/build_64.sh
@@ -0,0 +1,8 @@
+#!/bin/bash
+cmake ../../../../ \
+-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
+-DCMAKE_BUILD_TYPE=Release \
+-DANDROID_ABI="arm64-v8a" \
+-DANDROID_PLATFORM=android-23 $1
+
+make -j4
diff --git a/llm/llama.cpp/examples/llava/clip.cpp b/llm/llama.cpp/examples/llava/clip.cpp
index 95fbe3d0216c404cfc3491061d48b27cd643f6c7..7cda5f10c4b05dce52334ec6636c0d0fa7bcc1b6 100644
--- a/llm/llama.cpp/examples/llava/clip.cpp
+++ b/llm/llama.cpp/examples/llava/clip.cpp
@@ -16,6 +16,10 @@
 #include "ggml-metal.h"
 #endif
 
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -865,7 +869,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             embeddings = peg_0;
         }
         else {
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
         }
     }
 
@@ -1001,6 +1005,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     LOG_TEE("%s: CLIP using Metal backend\n", __func__);
 #endif
 
+#ifdef GGML_USE_CANN
+    new_clip->backend = ggml_backend_cann_init(0);
+    LOG_TEE("%s: CLIP using CANN backend\n", __func__);
+#endif
+
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
@@ -1121,20 +1130,20 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             }
             if (n < 32)
                 hparams.image_grid_pinpoints[n] = 0;
-        } catch (std::runtime_error & e) {
+        } catch (std::runtime_error & /*e*/) {
             hparams.image_grid_pinpoints[0]=0;
         }
 
         try {
             int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE);
             strcpy(hparams.mm_patch_merge_type, gguf_get_val_str(ctx, idx));
-        } catch (std::runtime_error & e) {
+        } catch (std::runtime_error & /*e*/) {
             strcpy(hparams.mm_patch_merge_type, "flat");
         }
 
         try {
             hparams.image_crop_resolution = get_u32(ctx, KEY_IMAGE_CROP_RESOLUTION); // llava-1.6
-        } catch(const std::exception& e) {
+        } catch(const std::exception& /*e*/) {
             hparams.image_crop_resolution = hparams.image_size;
         }
 
@@ -1173,7 +1182,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         try {
             vision_model.class_embedding  = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
             new_clip->has_class_embedding = true;
-        } catch (const std::exception& e) {
+        } catch (const std::exception& /*e*/) {
             new_clip->has_class_embedding = false;
         }
 
@@ -1181,7 +1190,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.pre_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
             vision_model.pre_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
             new_clip->has_pre_norm = true;
-        } catch (std::exception & e) {
+        } catch (std::exception & /*e*/) {
             new_clip->has_pre_norm = false;
         }
 
@@ -1189,21 +1198,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.post_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight"));
             vision_model.post_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias"));
             new_clip->has_post_norm = true;
-        } catch (std::exception & e) {
+        } catch (std::exception & /*e*/) {
             new_clip->has_post_norm = false;
         }
 
         try {
             vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS);
             new_clip->has_patch_bias = true;
-        } catch (std::exception & e) {
+        } catch (std::exception & /*e*/) {
             new_clip->has_patch_bias = false;
         }
 
         try {
             vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
             vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
-        } catch(const std::exception& e) {
+        } catch(const std::exception& /*e*/) {
             LOG_TEE("%s: failed to load vision model tensors\n", __func__);
         }
 
@@ -1215,26 +1224,26 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 // Yi-type llava
                 vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight"));
                 vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias"));
-            } catch (std::runtime_error & e) {  }
+            } catch (std::runtime_error & /*e*/) { }
             try {
                 // missing in Yi-type llava
                 vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
                 vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
-            } catch (std::runtime_error & e) {  }
+            } catch (std::runtime_error & /*e*/) { }
             try {
                 // Yi-type llava
                 vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight"));
                 vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias"));
-            } catch (std::runtime_error & e) {  }
+            } catch (std::runtime_error & /*e*/) { }
             try {
                 // Yi-type llava
                 vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight"));
                 vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias"));
-            } catch (std::runtime_error & e) {  }
+            } catch (std::runtime_error & /*e*/) { }
             try {
                 vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
                 // LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__);
-            } catch (std::runtime_error & e) {  }
+            } catch (std::runtime_error & /*e*/) { }
         } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projection
             vision_model.mm_model_mlp_1_w               = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight"));
diff --git a/llm/llama.cpp/examples/llava/convert_image_encoder_to_gguf.py b/llm/llama.cpp/examples/llava/convert_image_encoder_to_gguf.py
new file mode 100644
index 0000000000000000000000000000000000000000..36f6b92fbd46ebb31be0db3f890bf390f1d8829b
--- /dev/null
+++ b/llm/llama.cpp/examples/llava/convert_image_encoder_to_gguf.py
@@ -0,0 +1,333 @@
+import argparse
+import os
+import json
+import re
+
+import torch
+import numpy as np
+from gguf import *
+from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+
+
+def k(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if has_llava and name in ["visual_projection.weight", "vision_model.post_layernorm.weight", "vision_model.post_layernorm.bias"]:
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    config = json.load(f)
+    if args.clip_model_is_vision:
+        v_hparams = config
+        t_hparams = None
+    else:
+        v_hparams = config["vision_config"]
+        t_hparams = config["text_config"]
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+if args.clip_model_is_vision or args.clip_model_is_openclip:
+    model = CLIPVisionModel.from_pretrained(dir_model)
+    processor = None
+else:
+    model = CLIPModel.from_pretrained(dir_model)
+    processor = CLIPProcessor.from_pretrained(dir_model)
+
+fname_middle = None
+has_text_encoder = True
+has_vision_encoder = True
+has_llava_projector = False
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.llava_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_llava_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_llava_projector", has_llava_projector)
+fout.add_file_type(ftype)
+model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
+fout.add_name(model_name)
+if args.text_only:
+    fout.add_description("text-only CLIP model")
+elif args.vision_only and not has_llava_projector:
+    fout.add_description("vision-only CLIP model")
+elif has_llava_projector:
+    fout.add_description("image encoder for LLaVA")
+    # add projector type
+    fout.add_string("clip.projector_type", args.projector_type)
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_text_encoder:
+    assert t_hparams is not None
+    assert tokens is not None
+    # text_model hparams
+    fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
+    fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
+    fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
+    fout.add_token_list(tokens)
+
+if has_vision_encoder:
+    # vision_model hparams
+    fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
+    fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
+    fout.add_uint32("clip.vision.projection_dim", v_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
+    block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
+    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
+                            #     /**
+                            #      "image_grid_pinpoints": [
+                            #         [
+                            #         336,
+                            #         672
+                            #         ],
+                            #         [
+                            #         672,
+                            #         336
+                            #         ],
+                            #         [
+                            #         672,
+                            #         672
+                            #         ],
+                            #         [
+                            #         1008,
+                            #         336
+                            #         ],
+                            #         [
+                            #         336,
+                            #         1008
+                            #         ]
+                            #     ],
+                            #     Flattened:
+                            #     [
+                            #         336, 672,
+                            #         672, 336,
+                            #         672, 672,
+                            #         1008, 336,
+                            #         336, 1008
+                            #     ]
+                            #  *
+                            #  */
+    if "image_grid_pinpoints" in v_hparams:
+        # flatten it
+        image_grid_pinpoints = []
+        for pinpoint in v_hparams["image_grid_pinpoints"]:
+            for p in pinpoint:
+                image_grid_pinpoints.append(p)
+        fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
+    if "image_crop_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
+    if "image_aspect_ratio" in v_hparams:
+        fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
+    if "image_split_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
+    if "mm_patch_merge_type" in v_hparams:
+        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
+    if "mm_projector_type" in v_hparams:
+        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
+
+
+    if processor is not None:
+        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean  # pyright: ignore[reportAttributeAccessIssue]
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std  # pyright: ignore[reportAttributeAccessIssue]
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+use_gelu = v_hparams["hidden_act"] == "gelu"
+fout.add_bool("clip.use_gelu", use_gelu)
+
+
+if has_llava_projector:
+    model.vision_model.encoder.layers.pop(-1)  # pyright: ignore[reportAttributeAccessIssue]
+    projector = torch.load(args.llava_projector)
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        # pw and dw conv ndim==4
+        if data.ndim == 2 or data.ndim == 4:
+            data = data.squeeze().numpy().astype(np.float16)
+        else:
+            data = data.squeeze().numpy().astype(np.float32)
+
+        fout.add_tensor(name, data)
+
+    print("Projector tensors added\n")
+
+state_dict = model.state_dict()  # pyright: ignore[reportAttributeAccessIssue]
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)
diff --git a/llm/llama.cpp/examples/llava/llava-cli.cpp b/llm/llama.cpp/examples/llava/llava-cli.cpp
index c974900f21e20b4968250c3f74ebdf37eb91b808..8c7dd2ae3d0dc3b97c06575a71d540d0881daa27 100644
--- a/llm/llama.cpp/examples/llava/llava-cli.cpp
+++ b/llm/llama.cpp/examples/llava/llava-cli.cpp
@@ -112,9 +112,12 @@ struct llava_context {
     struct llama_model * model = NULL;
 };
 
-static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG_TEE("\n example usage: %s -m  --mmproj  --image  --image  [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
+    LOG_TEE("\n example usage:\n");
+    LOG_TEE("\n     %s -m  --mmproj  --image  --image  [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params, const std::string & fname) {
@@ -278,7 +281,7 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (!gpt_params_parse(argc, argv, params)) {
-        show_additional_info(argc, argv);
+        print_usage(argc, argv, params);
         return 1;
     }
 
@@ -290,8 +293,7 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        gpt_params_print_usage(argc, argv, params);
-        show_additional_info(argc, argv);
+        print_usage(argc, argv, {});
         return 1;
     }
     auto model = llava_init(¶ms);
diff --git a/llm/llama.cpp/examples/llava/llava_surgery.py b/llm/llama.cpp/examples/llava/llava_surgery.py
new file mode 100644
index 0000000000000000000000000000000000000000..4f2da3beefff670cfd61ffcff3769651d74a551c
--- /dev/null
+++ b/llm/llama.cpp/examples/llava/llava_surgery.py
@@ -0,0 +1,38 @@
+import argparse
+import glob
+import os
+import torch
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to LLaVA v1.5 model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+path = sorted(glob.glob(f"{args.model}/pytorch_model*.bin"))[-1]
+checkpoint = torch.load(path)
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_projector")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/llava.projector")
+
+# BakLLaVA models contain CLIP tensors in it
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/llava.clip")
+
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")
diff --git a/llm/llama.cpp/examples/llava/llava_surgery_v2.py b/llm/llama.cpp/examples/llava/llava_surgery_v2.py
new file mode 100644
index 0000000000000000000000000000000000000000..2d5b32fe6b236509ea563492378a5c281adff8b9
--- /dev/null
+++ b/llm/llama.cpp/examples/llava/llava_surgery_v2.py
@@ -0,0 +1,159 @@
+import argparse
+import glob
+import os
+import torch
+from safetensors import safe_open
+from safetensors.torch import save_file
+from typing import Any, ContextManager, cast
+
+# Function to determine if file is a SafeTensor file
+def is_safetensor_file(file_path):
+    return file_path.endswith('.safetensors')
+
+
+# Unified loading function
+def load_model(file_path):
+    if is_safetensor_file(file_path):
+        tensors = {}
+        with cast(ContextManager[Any], safe_open(file_path, framework="pt", device="cpu")) as f:
+            for key in f.keys():
+                tensors[key] = f.get_tensor(key).clone()
+                # output shape
+                print(f"{key} : {tensors[key].shape}")
+        return tensors, 'safetensor'
+    else:
+        return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
+
+
+# Unified saving function
+def save_model(model, file_path, file_type):
+    if file_type == 'safetensor':
+        # safe_save(model, file_path)
+        save_file(model, file_path)
+    else:
+        torch.save(model, file_path)
+
+
+# Adapted function to clean vision tower from checkpoint
+def clean_vision_tower_from_checkpoint(checkpoint_path):
+    checkpoint, file_type = load_model(checkpoint_path)
+    # file_type = 'pytorch'
+    model_path = os.path.dirname(checkpoint_path)
+    print(f"Searching for vision tower tensors in {checkpoint_path}")
+    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
+
+    if len(clip_tensors) > 0:
+        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
+        # Adapted for file type
+        clip_path = os.path.join(model_path, "llava.clip")
+
+        if os.path.exists(clip_path):
+            print(f"Loading existing llava.clip from {clip_path}")
+            existing_clip, _ = load_model(clip_path)
+        else:
+            print(f"Creating new llava.clip at {clip_path}")
+            existing_clip = {}
+        # Update existing_clip with new tensors, avoid duplicates
+        for name in clip_tensors:
+            simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
+            print(f"Adding {simple_name} to llava.clip")
+            if simple_name not in existing_clip:
+                existing_clip[simple_name] = checkpoint[name]
+
+        # Save the updated clip tensors back to llava.clip
+        save_model(existing_clip, clip_path, 'pytorch')
+
+        # Remove the tensors from the original checkpoint
+        for name in clip_tensors:
+            del checkpoint[name]
+
+        checkpoint_path = checkpoint_path
+        return True
+    return False
+
+def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
+    newline_checkpoint_path = None
+    projector_checkpoint_path = None
+
+    for path in checkpoint_paths:
+        checkpoint, _ = load_model(path)
+        if newline_criteria(checkpoint) and newline_checkpoint_path is None:
+            newline_checkpoint_path = path
+        if projector(checkpoint):
+            projector_checkpoint_path = path
+
+    return newline_checkpoint_path, projector_checkpoint_path
+
+def newline_criteria(checkpoint):
+    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
+
+
+# Command-line interface setup
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
+ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
+args = ap.parse_args()
+
+if args.clean_vision_tower:
+    # Generalized to handle both PyTorch and SafeTensors models
+    model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+    # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
+    checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+    for projector_checkpoint_path in checkpoint_paths:
+        print(f"Cleaning {projector_checkpoint_path}")
+        if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
+            print(f"No vision tower found in {projector_checkpoint_path}")
+            # we break once none is found, so far all models append them at the end
+            # break
+    print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
+
+# Now we look for the projector in the last checkpoint
+model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+# last_checkpoint_path = checkpoint_paths[0]
+# first_checkpoint_path = checkpoint_paths[-1]
+newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
+
+print(f"Taking projector from {projector_checkpoint_path}")
+first_mm_tensors = []
+first_checkpoint = None
+if newline_checkpoint_path is not None:
+    print(f"Taking newline from {newline_checkpoint_path}")
+    first_checkpoint, file_type = load_model(newline_checkpoint_path)
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
+
+# Load the checkpoint
+mm_tensors = []
+last_checkpoint = None
+if projector_checkpoint_path is not None:
+    last_checkpoint, file_type = load_model(projector_checkpoint_path)
+    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
+
+if len(mm_tensors) == 0:
+    if last_checkpoint is not None:
+        for k, v in last_checkpoint.items():
+            print(k)
+    print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint) if last_checkpoint is not None else 0} tensors.")
+    print("No tensors found. Is this a LLaVA model?")
+    exit()
+
+print(f"Found {len(mm_tensors)} tensors to extract.")
+print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
+# projector = {name: checkpoint.[name].float() for name in mm_tensors}
+projector = {}
+for name in mm_tensors:
+    assert last_checkpoint is not None
+    projector[name] = last_checkpoint[name].float()
+for name in first_mm_tensors:
+    assert first_checkpoint is not None
+    projector[name] = first_checkpoint[name].float()
+
+if len(projector) > 0:
+    save_model(projector, f"{args.model}/llava.projector", 'pytorch')
+
+print("Done!")
+print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")
diff --git a/llm/llama.cpp/examples/llava/requirements.txt b/llm/llama.cpp/examples/llava/requirements.txt
index 17cb4d5e5ee8e8a52a7fbe6514c6af30fafef5ce..fbfd0cdd722e13a33fead9fcdff8152ee9dd5cb5 100644
--- a/llm/llama.cpp/examples/llava/requirements.txt
+++ b/llm/llama.cpp/examples/llava/requirements.txt
@@ -1,3 +1,4 @@
--r ../../requirements/requirements-convert-legacy-llama.txt
+-r ../../requirements/requirements-convert_legacy_llama.txt
+--extra-index-url https://download.pytorch.org/whl/cpu
 pillow~=10.2.0
-torch~=2.1.1
+torch~=2.2.1
diff --git a/llm/llama.cpp/examples/lookahead/CMakeLists.txt b/llm/llama.cpp/examples/lookahead/CMakeLists.txt
index 8827e3f11ecd665c228f32bac4501ccb08123eed..f0ae5cd89244c10adc441402cb1ffc89a54637c1 100644
--- a/llm/llama.cpp/examples/lookahead/CMakeLists.txt
+++ b/llm/llama.cpp/examples/lookahead/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET lookahead)
+set(TARGET llama-lookahead)
 add_executable(${TARGET} lookahead.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/lookahead/lookahead.cpp b/llm/llama.cpp/examples/lookahead/lookahead.cpp
index 54f060a85b26365b561189f4396a99c90b3ac7ab..81cf1629c5b6ae5c46eebd766c09d0317693604d 100644
--- a/llm/llama.cpp/examples/lookahead/lookahead.cpp
+++ b/llm/llama.cpp/examples/lookahead/lookahead.cpp
@@ -37,7 +37,8 @@ struct ngram_container {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -57,11 +58,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // Tokenize the prompt
     std::vector inp;
diff --git a/llm/llama.cpp/examples/lookup/CMakeLists.txt b/llm/llama.cpp/examples/lookup/CMakeLists.txt
index b91633f63e4ee0ff33f923dbc8257c30b4fef6f6..ef19fe25e31a3101cbdbf06a228bfede19875e5d 100644
--- a/llm/llama.cpp/examples/lookup/CMakeLists.txt
+++ b/llm/llama.cpp/examples/lookup/CMakeLists.txt
@@ -1,22 +1,22 @@
-set(TARGET lookup)
+set(TARGET llama-lookup)
 add_executable(${TARGET} lookup.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 
-set(TARGET lookup-create)
+set(TARGET llama-lookup-create)
 add_executable(${TARGET} lookup-create.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 
-set(TARGET lookup-merge)
+set(TARGET llama-lookup-merge)
 add_executable(${TARGET} lookup-merge.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 
-set(TARGET lookup-stats)
+set(TARGET llama-lookup-stats)
 add_executable(${TARGET} lookup-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/lookup/README.md b/llm/llama.cpp/examples/lookup/README.md
index 5bfb0de9360414aa92c9375ad3352099ecd87e6f..71c345c037a2fbfb7ee00640f7844f1d53143152 100644
--- a/llm/llama.cpp/examples/lookup/README.md
+++ b/llm/llama.cpp/examples/lookup/README.md
@@ -10,4 +10,3 @@ More info:
 
 https://github.com/ggerganov/llama.cpp/pull/4484
 https://github.com/ggerganov/llama.cpp/issues/4226
-
diff --git a/llm/llama.cpp/examples/lookup/lookup-create.cpp b/llm/llama.cpp/examples/lookup/lookup-create.cpp
index 1c230c9667c715dd7c64c207c48b91b4bd9508a4..5f04709f5023179fbf54febbe8f5a023a5fa033c 100644
--- a/llm/llama.cpp/examples/lookup/lookup-create.cpp
+++ b/llm/llama.cpp/examples/lookup/lookup-create.cpp
@@ -14,17 +14,19 @@ int main(int argc, char ** argv){
     gpt_params params;
 
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
+
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt
diff --git a/llm/llama.cpp/examples/lookup/lookup-merge.cpp b/llm/llama.cpp/examples/lookup/lookup-merge.cpp
index 07c93eb8d057bb69385dd19fc1d68d1b9c040abf..81e2b04369b90313e7cf217358d1d07eeafc9af3 100644
--- a/llm/llama.cpp/examples/lookup/lookup-merge.cpp
+++ b/llm/llama.cpp/examples/lookup/lookup-merge.cpp
@@ -11,14 +11,14 @@
 #include 
 #include 
 
-static void print_usage() {
+static void print_usage(char* argv0) {
     fprintf(stderr, "Merges multiple lookup cache files into a single one.\n");
-    fprintf(stderr, "Usage: lookup-merge [--help] lookup_part_1.bin lookup_part_2.bin ... lookup_merged.bin\n");
+    fprintf(stderr, "Usage: %s [--help] lookup_part_1.bin lookup_part_2.bin ... lookup_merged.bin\n", argv0);
 }
 
 int main(int argc, char ** argv){
     if (argc < 3) {
-        print_usage();
+        print_usage(argv[0]);
         exit(1);
     }
 
@@ -27,7 +27,7 @@ int main(int argc, char ** argv){
     for (int i = 0; i < argc-1; ++i) {
         args[i] = argv[i+1];
         if (args[i] == "-h" || args[i] == "--help") {
-            print_usage();
+            print_usage(argv[0]);
             exit(0);
         }
     }
diff --git a/llm/llama.cpp/examples/lookup/lookup-stats.cpp b/llm/llama.cpp/examples/lookup/lookup-stats.cpp
index 87ecc0a4f1394e615971cb9784edd005b3f726dc..400f3e0b089574d642b1860502962cad28680d37 100644
--- a/llm/llama.cpp/examples/lookup/lookup-stats.cpp
+++ b/llm/llama.cpp/examples/lookup/lookup-stats.cpp
@@ -16,6 +16,7 @@ int main(int argc, char ** argv){
     gpt_params params;
 
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -25,12 +26,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector inp;
@@ -64,7 +64,7 @@ int main(int argc, char ** argv){
     }
 
     const int n_input = inp.size();
-    const int n_ctx = params.n_ctx;
+    const int n_ctx = llama_n_ctx(ctx);
 
     int n_drafted = 0;
     int n_accept  = 0;
diff --git a/llm/llama.cpp/examples/lookup/lookup.cpp b/llm/llama.cpp/examples/lookup/lookup.cpp
index 83dbee91a8362bff615cb82c62ffe563768e1538..d53a9828c2ea23918c01b15b507579e5c0647213 100644
--- a/llm/llama.cpp/examples/lookup/lookup.cpp
+++ b/llm/llama.cpp/examples/lookup/lookup.cpp
@@ -15,6 +15,7 @@ int main(int argc, char ** argv){
     gpt_params params;
 
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -33,12 +34,11 @@ int main(int argc, char ** argv){
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // tokenize the prompt
     std::vector inp;
diff --git a/llm/llama.cpp/examples/main-cmake-pkg/.gitignore b/llm/llama.cpp/examples/main-cmake-pkg/.gitignore
index e32c11c7f4653cd6b1db1af13873d1ea50095544..67c01d64cb7ab21f9e6b59c018dd849f7f9aa5db 100644
--- a/llm/llama.cpp/examples/main-cmake-pkg/.gitignore
+++ b/llm/llama.cpp/examples/main-cmake-pkg/.gitignore
@@ -48,4 +48,3 @@
 build*/
 out/
 tmp/
-
diff --git a/llm/llama.cpp/examples/main-cmake-pkg/CMakeLists.txt b/llm/llama.cpp/examples/main-cmake-pkg/CMakeLists.txt
index deb77d588ea9fde64955a58aa0d2b6d8a532d83f..3b38db292320f813b011d6b41f94179f00979556 100644
--- a/llm/llama.cpp/examples/main-cmake-pkg/CMakeLists.txt
+++ b/llm/llama.cpp/examples/main-cmake-pkg/CMakeLists.txt
@@ -1,12 +1,12 @@
 cmake_minimum_required(VERSION 3.12)
-project("main-cmake-pkg" C CXX)
-set(TARGET main-cmake-pkg)
+project("llama-cli-cmake-pkg" C CXX)
+set(TARGET llama-cli-cmake-pkg)
 
 find_package(Llama 0.0.1 REQUIRED)
 
 # Bake common functionality in with target. Because applications
 # using the relocatable Llama package should be outside of the
-# source tree, main-cmake-pkg pretends the dependencies are built-in.
+# source tree, llama-cli-cmake-pkg pretends the dependencies are built-in.
 set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
 add_library(common OBJECT)
 file(GLOB _common_files
@@ -15,7 +15,7 @@ file(GLOB _common_files
 )
 target_sources(common PRIVATE ${_common_files})
 
-# If the common project was part of "main-cmake-pkg" the transient
+# If the common project was part of "llama-cli-cmake-pkg" the transient
 # defines would automatically be attached. Because the common func-
 # tionality is separate, but dependent upon the defines, it must be
 # explicitly extracted from the "llama" target.
@@ -30,4 +30,3 @@ target_include_directories(${TARGET} PRIVATE ${_common_path})
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-
diff --git a/llm/llama.cpp/examples/main-cmake-pkg/README.md b/llm/llama.cpp/examples/main-cmake-pkg/README.md
index edf20d8db6616abe8fe8f7078e4f3f232e882872..08d83dd08636a8d3b666c9fcce6f06a2b7602ee2 100644
--- a/llm/llama.cpp/examples/main-cmake-pkg/README.md
+++ b/llm/llama.cpp/examples/main-cmake-pkg/README.md
@@ -1,6 +1,6 @@
 # llama.cpp/example/main-cmake-pkg
 
-This program builds the [main](../main) application using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
+This program builds [llama-cli](../main) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
 
 ## Building
 
@@ -8,26 +8,24 @@ Because this example is "outside of the source tree", it is important to first b
 
 ### Considerations
 
-When hardware acceleration libraries are used (e.g. CUDA, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_.
+When hardware acceleration libraries are used (e.g. CUDA, Metal, etc.), CMake must be able to locate the associated CMake package.
 
 ### Build llama.cpp and install to C:\LlamaCPP directory
 
-In this case, CLBlast was already installed so the CMake package is referenced in `CMAKE_PREFIX_PATH`.
-
 ```cmd
 git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
-cmake -B build -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=C:/CLBlast/lib/cmake/CLBlast -G "Visual Studio 17 2022" -A x64
+cmake -B build -DBUILD_SHARED_LIBS=OFF -G "Visual Studio 17 2022" -A x64
 cmake --build build --config Release
 cmake --install build --prefix C:/LlamaCPP
 ```
 
-### Build main-cmake-pkg
+### Build llama-cli-cmake-pkg
 
 
 ```cmd
 cd ..\examples\main-cmake-pkg
-cmake -B build -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/CLBlast/lib/cmake/CLBlast;C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
+cmake -B build -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
 cmake --build build --config Release
 cmake --install build --prefix C:/MyLlamaApp
 ```
diff --git a/llm/llama.cpp/examples/main/CMakeLists.txt b/llm/llama.cpp/examples/main/CMakeLists.txt
index d532980b76da837f1d448edac7346a92bc46e141..5f6efaa9aa94bb07d909fb13404e3a510ab2d67c 100644
--- a/llm/llama.cpp/examples/main/CMakeLists.txt
+++ b/llm/llama.cpp/examples/main/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET main)
+set(TARGET llama-cli)
 add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/main/README.md b/llm/llama.cpp/examples/main/README.md
index ee930f4e79a0d8db83847b6db183cf6c55c460cd..9396a34fa5a31d36bad9191b90be00b6e196efee 100644
--- a/llm/llama.cpp/examples/main/README.md
+++ b/llm/llama.cpp/examples/main/README.md
@@ -1,6 +1,6 @@
-# llama.cpp/example/main
+# llama.cpp/examples/main
 
-This example program allows you to use various LLaMA language models in an easy and efficient way. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
+This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
 
 ## Table of Contents
 
@@ -17,74 +17,71 @@ This example program allows you to use various LLaMA language models in an easy
 
 To get started right away, run the following command, making sure to use the correct path for the model you have:
 
-#### Unix-based systems (Linux, macOS, etc.):
+First, we will need to download a model. In these examples, we will use the Gemma model from the ggml-org repo on Hugging Face.
+[https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)
 
-```bash
-./main -m models/7B/ggml-model.bin --prompt "Once upon a time"
-```
+Once downloaded, place your model in the models folder in llama.cpp.
 
-#### Windows:
+### Unix-based systems (Linux, macOS, etc.):
 
-```powershell
-main.exe -m models\7B\ggml-model.bin --prompt "Once upon a time"
-```
+##### Input prompt (One-and-done)
 
-For an interactive experience, try this command:
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+```
+##### Conversation mode (Allow for continuous interaction with the model)
 
-#### Unix-based systems (Linux, macOS, etc.):
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+```
 
+##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
 ```bash
-./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -p \
-'User: Hi
-AI: Hello. I am an AI chatbot. Would you like to talk?
-User: Sure!
-AI: What would you like to talk about?
-User:'
+./llama-cli -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 ```
 
-#### Windows:
+### Windows:
 
+##### Input prompt (One-and-done)
 ```powershell
-main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -e -p "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:"
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
 ```
+##### Conversation mode (Allow for continuous interaction with the model)
 
-The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it):
-
-#### Unix-based systems (Linux, macOS, etc.):
-
-```bash
-./main -m models/7B/ggml-model.bin --ignore-eos -n -1 --random-prompt
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
 ```
 
-#### Windows:
+#### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
 
 ```powershell
-main.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 --random-prompt
+llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 ```
 
 ## Common Options
 
-In this section, we cover the most commonly used options for running the `main` program with the LLaMA models:
+In this section, we cover the most commonly used options for running the `llama-cli` program with the LLaMA models:
 
--   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`; inferred from `--model-url` if set).
--   `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/gemma-1.1-7b-it.Q4_K_M.gguf`; inferred from `--model-url` if set).
+-   `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g [https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)).
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
--   `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
 -   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
 -   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+-   `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\'
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has.
+-   -   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 
 ## Input Prompts
 
-The `main` program provides several ways to interact with the LLaMA models using input prompts:
+The `llama-cli` program provides several ways to interact with the LLaMA models using input prompts:
 
 -   `--prompt PROMPT`: Provide a prompt directly as a command-line option.
 -   `--file FNAME`: Provide a file containing a prompt or multiple prompts.
 -   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
--   `--random-prompt`: Start with a randomized prompt.
 
 ## Interaction
 
-The `main` program offers a seamless way to interact with LLaMA models, allowing users to engage in real-time conversations or provide instructions for specific tasks. The interactive mode can be triggered using various options, including `--interactive`, `--interactive-first`, and `--instruct`.
+The `llama-cli` program offers a seamless way to interact with LLaMA models, allowing users to engage in real-time conversations or provide instructions for specific tasks. The interactive mode can be triggered using various options, including `--interactive` and `--interactive-first`.
 
 In interactive mode, users can participate in text generation by injecting their input during the process. Users can press `Ctrl+C` at any time to interject and type their input, followed by pressing `Return` to submit it to the LLaMA model. To submit additional lines without finalizing input, users can end the current line with a backslash (`\`) and continue typing.
 
@@ -92,7 +89,7 @@ In interactive mode, users can participate in text generation by injecting their
 
 -   `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model.
 -   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
--   `-ins, --instruct`: Run the program in instruction mode, which is specifically designed to work with Alpaca models that excel in completing tasks based on user instructions.
+-   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: false)
 -   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
 
 By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
@@ -110,7 +107,7 @@ To overcome this limitation, you can use the `--in-prefix` flag to add a space o
 The `--in-prefix` flag is used to add a prefix to your input, primarily, this is used to insert a space after the reverse prompt. Here's an example of how to use the `--in-prefix` flag in conjunction with the `--reverse-prompt` flag:
 
 ```sh
-./main -r "User:" --in-prefix " "
+./llama-cli -r "User:" --in-prefix " "
 ```
 
 ### In-Suffix
@@ -118,18 +115,15 @@ The `--in-prefix` flag is used to add a prefix to your input, primarily, this is
 The `--in-suffix` flag is used to add a suffix after your input. This is useful for adding an "Assistant:" prompt after the user's input. It's added after the new-line character (`\n`) that's automatically added to the end of the user's input. Here's an example of how to use the `--in-suffix` flag in conjunction with the `--reverse-prompt` flag:
 
 ```sh
-./main -r "User:" --in-prefix " " --in-suffix "Assistant:"
+./llama-cli -r "User:" --in-prefix " " --in-suffix "Assistant:"
 ```
+When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled
 
-### Instruction Mode
-
-Instruction mode is particularly useful when working with Alpaca models, which are designed to follow user instructions for specific tasks:
+### Chat templates
 
--   `-ins, --instruct`: Enable instruction mode to leverage the capabilities of Alpaca models in completing tasks based on user-provided instructions.
+ `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name.  Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled.
 
-Technical detail: the user's input is internally prefixed with the reverse prompt (or `### Instruction:` as the default), and followed by `### Response:` (except if you just press Return without any input, to keep generating a longer response).
-
-By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
+ Example usage: `--chat-template gemma`
 
 ## Context Management
 
@@ -137,9 +131,7 @@ During text generation, LLaMA models have a limited context size, which means th
 
 ### Context Size
 
-The `--ctx-size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
-
--   `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
+- `-c N, --ctx-size N`: Set the size of the prompt context (default: 0, 0 = loaded from model). The LLaMA models were built with a context of 2048-8192, which will yield the best results on longer input/inference.
 
 ### Extended Context Size
 
@@ -161,15 +153,15 @@ The following options allow you to control the text generation process and fine-
 
 ### Number of Tokens to Predict
 
--   `-n N, --n-predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity, -2 = until context filled)
+-   `-n N, --predict N`: Set the number of tokens to predict when generating text (default: -1, -1 = infinity, -2 = until context filled)
 
-The `--n-predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text.
+The `--predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text.
 
-A value of -1 will enable infinite text generation, even though we have a finite context window. When the context window is full, some of the earlier tokens (half of the tokens after `--n-keep`) will be discarded. The context must then be re-evaluated before generation can resume. On large models and/or large context windows, this will result in significant pause in output.
+A value of -1 will enable infinite text generation, even though we have a finite context window. When the context window is full, some of the earlier tokens (half of the tokens after `--keep`) will be discarded. The context must then be re-evaluated before generation can resume. On large models and/or large context windows, this will result in a significant pause in output.
 
 If the pause is undesirable, a value of -2 will stop generation immediately when the context is filled.
 
-It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n-predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
+It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode, text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `--predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
 
 ### Temperature
 
@@ -177,15 +169,15 @@ It is important to note that the generated text may be shorter than the specifie
 
 Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.
 
-Example usage: `--temp 0.5`
+Example usage: `--temp 0`
 
 ### Repeat Penalty
 
--   `--repeat-penalty N`: Control the repetition of token sequences in the generated text (default: 1.1).
+-   `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled).
 -   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
 -   `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty.
 
-The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1.
+The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.
 
 The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
 
@@ -209,19 +201,19 @@ Top-p sampling, also known as nucleus sampling, is another text generation metho
 
 Example usage: `--top-p 0.95`
 
-### Min P Sampling
+### Min-P Sampling
 
--   `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.05).
+-   `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.1).
 
 The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out.
 
 Example usage: `--min-p 0.05`
 
-### Tail Free Sampling (TFS)
+### Tail-Free Sampling (TFS)
 
 -   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
 
-Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens, and thus disables the effect of TFS.
+Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS.
 
 Example usage: `--tfs 0.95`
 
@@ -320,10 +312,8 @@ These options provide extra functionality and customization when running the LLa
 
 -   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
 -   `--verbose-prompt`: Print the prompt before generating text.
--   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
-
 -   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.
diff --git a/llm/llama.cpp/examples/main/main.cpp b/llm/llama.cpp/examples/main/main.cpp
index 44949ba869e70248bf40d417b52413125db4df62..6e0635a66cd0671e81edc6f0c3815e92ae3d8693 100644
--- a/llm/llama.cpp/examples/main/main.cpp
+++ b/llm/llama.cpp/examples/main/main.cpp
@@ -37,14 +37,15 @@ static gpt_params               * g_params;
 static std::vector * g_input_tokens;
 static std::ostringstream       * g_output_ss;
 static std::vector * g_output_tokens;
-static bool is_interacting = false;
+static bool is_interacting  = false;
+static bool need_insert_eot = false;
 
-static bool file_exists(const std::string &path) {
+static bool file_exists(const std::string & path) {
     std::ifstream f(path.c_str());
     return f.good();
 }
 
-static bool file_is_empty(const std::string &path) {
+static bool file_is_empty(const std::string & path) {
     std::ifstream f;
     f.exceptions(std::ifstream::failbit | std::ifstream::badbit);
     f.open(path.c_str(), std::ios::in | std::ios::binary | std::ios::ate);
@@ -99,7 +100,8 @@ static void write_logfile(
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
         if (!is_interacting && g_params->interactive) {
-            is_interacting = true;
+            is_interacting  = true;
+            need_insert_eot = true;
         } else {
             console::cleanup();
             printf("\n");
@@ -117,13 +119,24 @@ static void llama_log_callback_logTee(ggml_log_level level, const char * text, v
     LOG_TEE("%s", text);
 }
 
+static std::string chat_add_and_format(struct llama_model * model, std::vector & chat_msgs, std::string role, std::string content) {
+    llama_chat_msg new_msg{role, content};
+    auto formatted = llama_chat_format_single(
+        model, g_params->chat_template, chat_msgs, new_msg, role == "user");
+    chat_msgs.push_back({role, content});
+    LOG("formatted: %s\n", formatted.c_str());
+    return formatted;
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
     g_params = ¶ms;
 
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
+
     llama_sampling_params & sparams = params.sparams;
 
 #ifndef LOG_DISABLE_LOGS
@@ -180,9 +193,6 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
-    if (params.random_prompt) {
-        params.prompt = string_random_prompt(rng);
-    }
 
     LOG("%s: llama backend init\n", __func__);
     llama_backend_init();
@@ -191,12 +201,16 @@ int main(int argc, char ** argv) {
     llama_model * model;
     llama_context * ctx;
     llama_context * ctx_guidance = NULL;
+    std::vector chat_msgs;
     g_model = &model;
     g_ctx = &ctx;
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    model = llama_init.model;
+    ctx = llama_init.context;
     if (sparams.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
         ctx_guidance = llama_new_context_with_model(model, lparams);
@@ -216,6 +230,15 @@ int main(int argc, char ** argv) {
                 __func__, n_ctx_train, n_ctx);
     }
 
+    // print chat template example in conversation mode
+    if (params.conversation) {
+        if (params.enable_chat_template) {
+            LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str());
+        } else {
+            LOG_TEE("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
+        }
+    }
+
     // print system information
     {
         LOG_TEE("\n");
@@ -245,29 +268,38 @@ int main(int argc, char ** argv) {
     }
 
     const bool add_bos = llama_should_add_bos_token(model);
-    GGML_ASSERT(llama_add_eos_token(model) != 1);
+    if (!llama_model_has_encoder(model)) {
+        GGML_ASSERT(llama_add_eos_token(model) != 1);
+    }
     LOG("add_bos: %d\n", add_bos);
 
     std::vector embd_inp;
 
-    if (params.interactive_first || params.instruct || params.chatml || !params.prompt.empty() || session_tokens.empty()) {
-        LOG("tokenize the prompt\n");
-        if (params.chatml) {
-            params.prompt = "<|im_start|>system\n" + params.prompt + "<|im_end|>";
+    {
+        auto prompt = (params.conversation && params.enable_chat_template && !params.prompt.empty())
+            ? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode
+            : params.prompt;
+        if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
+            LOG("tokenize the prompt\n");
+            embd_inp = ::llama_tokenize(ctx, prompt, true, true);
+        } else {
+            LOG("use session tokens\n");
+            embd_inp = session_tokens;
         }
-        embd_inp = ::llama_tokenize(ctx, params.prompt, true, true);
-    } else {
-        LOG("use session tokens\n");
-        embd_inp = session_tokens;
-    }
 
-    LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+        LOG("prompt: \"%s\"\n", log_tostr(prompt));
+        LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+    }
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
-        embd_inp.push_back(llama_token_bos(model));
-        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+        if (add_bos) {
+            embd_inp.push_back(llama_token_bos(model));
+            LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+        } else {
+            LOG_TEE("error: input is empty\n");
+            return -1;
+        }
     }
 
     // Tokenize negative prompt
@@ -332,37 +364,13 @@ int main(int argc, char ** argv) {
     }
 
     // number of tokens to keep when resetting context
-    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
         params.n_keep = (int)embd_inp.size();
     } else {
         params.n_keep += add_bos; // always keep the BOS token
     }
 
-    // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true,  true);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false, true);
-
-    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
-    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
-
-    // chatml prefix & suffix
-    const auto cml_pfx = ::llama_tokenize(ctx, "\n<|im_start|>user\n", true, true);
-    const auto cml_sfx = ::llama_tokenize(ctx, "<|im_end|>\n<|im_start|>assistant\n", false, true);
-
-    LOG("cml_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, cml_pfx).c_str());
-    LOG("cml_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, cml_sfx).c_str());
-
-    // in instruct mode, we inject a prefix and a suffix to each input by the user
-    if (params.instruct) {
-        params.interactive_first = true;
-        params.antiprompt.emplace_back("### Instruction:\n\n");
-    }
-    // similar for chatml mode
-    else if (params.chatml) {
-        params.interactive_first = true;
-        params.antiprompt.emplace_back("<|im_start|>user\n");
-    }
-    else if (params.conversation) {
+    if (params.conversation) {
         params.interactive_first = true;
     }
 
@@ -506,6 +514,7 @@ int main(int argc, char ** argv) {
     std::vector   input_tokens;  g_input_tokens  = &input_tokens;
     std::vector   output_tokens; g_output_tokens = &output_tokens;
     std::ostringstream output_ss;     g_output_ss     = &output_ss;
+    std::ostringstream assistant_ss; // for storing current assistant message, used in conversation mode
 
     // the first thing we will do is to output the prompt, so set color accordingly
     console::set_display(console::prompt);
@@ -528,6 +537,24 @@ int main(int argc, char ** argv) {
         exit(1);
     }
 
+    if (llama_model_has_encoder(model)) {
+        int enc_input_size = embd_inp.size();
+        llama_token * enc_input_buf = embd_inp.data();
+
+        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
+            LOG_TEE("%s : failed to eval\n", __func__);
+            return 1;
+        }
+
+        llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+        if (decoder_start_token_id == -1) {
+            decoder_start_token_id = llama_token_bos(model);
+        }
+
+        embd_inp.clear();
+        embd_inp.push_back(decoder_start_token_id);
+    }
+
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (!embd.empty()) {
@@ -821,17 +848,24 @@ int main(int argc, char ** argv) {
                         is_antiprompt = true;
                     }
 
+                    if (params.enable_chat_template) {
+                        chat_add_and_format(model, chat_msgs, "assistant", assistant_ss.str());
+                    }
                     is_interacting = true;
                     printf("\n");
-                } else if (params.instruct || params.chatml) {
-                    is_interacting = true;
                 }
             }
 
+            // if current token is not EOG, we add it to current assistant message
+            if (params.conversation) {
+                auto id = llama_sampling_last(ctx_sampling);
+                assistant_ss << llama_token_to_piece(ctx, id, false);
+            }
+
             if (n_past > 0 && is_interacting) {
                 LOG("waiting for user input\n");
 
-                if (params.conversation || params.instruct || params.chatml) {
+                if (params.conversation) {
                     printf("\n> ");
                 }
 
@@ -874,49 +908,41 @@ int main(int argc, char ** argv) {
 
                     const size_t original_size = embd_inp.size();
 
-                    // instruct mode: insert instruction prefix
-                    if (params.instruct && !is_antiprompt) {
-                        LOG("inserting instruction prefix\n");
-                        n_consumed = embd_inp.size();
-                        embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
-                    }
-                    // chatml mode: insert user chat prefix
-                    if (params.chatml && !is_antiprompt) {
-                        LOG("inserting chatml prefix\n");
-                        n_consumed = embd_inp.size();
-                        embd_inp.insert(embd_inp.end(), cml_pfx.begin(), cml_pfx.end());
-                    }
                     if (params.escape) {
                         string_process_escapes(buffer);
                     }
 
+                    bool format_chat = params.conversation && params.enable_chat_template;
+                    std::string user_inp = format_chat
+                        ? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
+                        : std::move(buffer);
+                    // TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, params.interactive_specials);
+                    const auto line_inp = ::llama_tokenize(ctx, user_inp,            false, format_chat);
                     const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
 
                     LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
 
+                    // if user stop generation mid-way, we must add EOT to finish model's last response
+                    if (need_insert_eot && format_chat) {
+                        llama_token eot = llama_token_eot(model);
+                        embd_inp.push_back(eot == -1 ? llama_token_eos(model) : eot);
+                        need_insert_eot = false;
+                    }
+
                     embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
                     embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
 
-                    // instruct mode: insert response suffix
-                    if (params.instruct) {
-                        LOG("inserting instruction suffix\n");
-                        embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
-                    }
-                    // chatml mode: insert assistant chat suffix
-                    if (params.chatml) {
-                        LOG("inserting chatml suffix\n");
-                        embd_inp.insert(embd_inp.end(), cml_sfx.begin(), cml_sfx.end());
-                    }
-
                     for (size_t i = original_size; i < embd_inp.size(); ++i) {
                         const llama_token token = embd_inp[i];
                         output_tokens.push_back(token);
                         output_ss << llama_token_to_piece(ctx, token);
                     }
 
+                    // reset assistant message
+                    assistant_ss.str("");
+
                     n_remain -= line_inp.size();
                     LOG("n_remain: %d\n", n_remain);
                 } else {
@@ -935,7 +961,7 @@ int main(int argc, char ** argv) {
         }
 
         // end of generation
-        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.instruct || params.interactive || params.chatml)) {
+        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.interactive)) {
             LOG_TEE(" [end of text]\n");
             break;
         }
diff --git a/llm/llama.cpp/examples/parallel/CMakeLists.txt b/llm/llama.cpp/examples/parallel/CMakeLists.txt
index 319535a6e9054db8498f5e037332353474d2ad80..c13557bac2bac2ea378548d47f5becba659a483b 100644
--- a/llm/llama.cpp/examples/parallel/CMakeLists.txt
+++ b/llm/llama.cpp/examples/parallel/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET parallel)
+set(TARGET llama-parallel)
 add_executable(${TARGET} parallel.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/parallel/parallel.cpp b/llm/llama.cpp/examples/parallel/parallel.cpp
index c731abb726dc2b9e0099f72bb5ba7ccdddbb58d4..621a1c959062265912993c6065e2e67e785da378 100644
--- a/llm/llama.cpp/examples/parallel/parallel.cpp
+++ b/llm/llama.cpp/examples/parallel/parallel.cpp
@@ -100,7 +100,8 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -128,11 +129,11 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
     // load the target model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // load the prompts from an external file if there are any
     if (params.prompt.empty()) {
diff --git a/llm/llama.cpp/examples/passkey/CMakeLists.txt b/llm/llama.cpp/examples/passkey/CMakeLists.txt
index 3161bf3ef9a455692d71e1394570753d67d46678..dc467a5d3e4114ea6c1a3169f07b100d157b4245 100644
--- a/llm/llama.cpp/examples/passkey/CMakeLists.txt
+++ b/llm/llama.cpp/examples/passkey/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET passkey)
+set(TARGET llama-passkey)
 add_executable(${TARGET} passkey.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/passkey/README.md b/llm/llama.cpp/examples/passkey/README.md
index 4a22bb55975be508630f516fdf8a9593f51d072c..2b8e910f9658d66a8623f8146e1cd7f772e539ea 100644
--- a/llm/llama.cpp/examples/passkey/README.md
+++ b/llm/llama.cpp/examples/passkey/README.md
@@ -1,5 +1,8 @@
 # llama.cpp/example/passkey
 
+A passkey retrieval task is an evaluation method used to measure a language
+models ability to recall information from long contexts.
+
 See the following PRs for more info:
 
 - https://github.com/ggerganov/llama.cpp/pull/3856
@@ -8,5 +11,5 @@ See the following PRs for more info:
 ### Usage
 
 ```bash
-make -j && ./passkey ./models/llama-7b-v2/ggml-model-f16.gguf 250
+make -j && ./llama-passkey -m ./models/llama-7b-v2/ggml-model-f16.gguf --junk 250
 ```
diff --git a/llm/llama.cpp/examples/passkey/passkey.cpp b/llm/llama.cpp/examples/passkey/passkey.cpp
index f2ef9ca10d4a271f3357a3bc34843638f67aefae..d03215cd1e0a94e4810868e234996974ab2753f7 100644
--- a/llm/llama.cpp/examples/passkey/passkey.cpp
+++ b/llm/llama.cpp/examples/passkey/passkey.cpp
@@ -6,46 +6,32 @@
 #include 
 #include 
 
-int main(int argc, char ** argv) {
-    gpt_params params;
-
-    if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH N_JUNK N_GRP I_POS SEED\n" , argv[0]);
-        return 1 ;
-    }
-
-    int seed = -1;
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
 
-    int n_junk = 250; // number of times to repeat the junk text
-    int n_keep = 32;  // number of tokens in the prompt prefix
-    int n_grp  = 1;   // if more than 1 - perform LongLM SelfExtend
-    int i_pos  = -1;  // position of the passkey in the junk text
-
-    if (argc >= 2) {
-        params.model = argv[1];
-    }
-
-    if (argc >= 3) {
-        n_junk = std::stoi(argv[2]);
-    }
+    LOG_TEE("\nexample usage:\n");
+    LOG_TEE("\n    %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
+    LOG_TEE("\n");
+}
 
-    if (argc >= 4) {
-        n_grp = std::stoi(argv[3]);
-    }
+int main(int argc, char ** argv) {
+    gpt_params params;
 
-    if (argc >= 5) {
-        i_pos = std::stoi(argv[4]);
-    }
+    params.n_junk = 250;
+    params.n_keep = 32;
+    params.i_pos  = -1;
 
-    if (argc >= 6) {
-        seed = std::stoi(argv[5]);
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
     }
 
-    if (seed == -1) {
-        seed = time(NULL);
-    }
+    srand(params.seed == LLAMA_DEFAULT_SEED ? time(NULL) : params.seed);
 
-    srand(seed);
+    int n_junk = params.n_junk;
+    int n_keep = params.n_keep;
+    int n_grp  = params.grp_attn_n;
+    int i_pos  = params.i_pos;
 
     if (i_pos == -1) {
         i_pos = rand() % n_junk;
@@ -76,9 +62,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_default_params();
-
-    model_params.n_gpu_layers = 99; // offload all layers to the GPU
+    llama_model_params model_params = llama_model_params_from_gpt_params(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -89,13 +73,9 @@ int main(int argc, char ** argv) {
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_default_params();
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
-    ctx_params.seed    = seed;
-    ctx_params.n_ctx   = llama_n_ctx_train(model)*n_grp + n_keep;
-    ctx_params.n_batch = 512;
-    ctx_params.n_threads       = params.n_threads;
-    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep;
 
     GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
 
@@ -135,7 +115,7 @@ int main(int argc, char ** argv) {
     LOG_TEE("prompt tokens: %d\n", n_tokens_all);
     //LOG_TEE("prompt: %s\n", params.prompt.c_str());
 
-    llama_batch batch = llama_batch_init(512, 0, 1);
+    llama_batch batch = llama_batch_init(params.n_batch, 0, 1);
 
     int n_past = 0;
 
diff --git a/llm/llama.cpp/examples/perplexity/CMakeLists.txt b/llm/llama.cpp/examples/perplexity/CMakeLists.txt
index 3c76d3221416b7a9e9ec96d493d629d433fa6bc1..be0f2fd029e679c94786bc0079ff50da0018b876 100644
--- a/llm/llama.cpp/examples/perplexity/CMakeLists.txt
+++ b/llm/llama.cpp/examples/perplexity/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET perplexity)
+set(TARGET llama-perplexity)
 add_executable(${TARGET} perplexity.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/perplexity/perplexity.cpp b/llm/llama.cpp/examples/perplexity/perplexity.cpp
index 30e5e282ef5cfb08f83944dcfb0d776b5490c919..372684f092de29e12248d3594a0a3049ab0efc7f 100644
--- a/llm/llama.cpp/examples/perplexity/perplexity.cpp
+++ b/llm/llama.cpp/examples/perplexity/perplexity.cpp
@@ -476,7 +476,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     }
 
     // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
+    // Run `./llama-perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
@@ -1032,7 +1032,7 @@ struct winogrande_entry {
     std::vector seq_tokens[2];
 };
 
-static std::vector load_winogrande_from_csv(const std::string& prompt) {
+static std::vector load_winogrande_from_csv(const std::string & prompt) {
     std::vector result;
     std::istringstream in(prompt);
     std::string line;
@@ -1964,12 +1964,14 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
+    params.n_ctx = 512;
+    params.logits_all = true;
+
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
-    params.logits_all = true;
-
     const int32_t n_ctx = params.n_ctx;
 
     if (n_ctx <= 0) {
@@ -1989,6 +1991,12 @@ int main(int argc, char ** argv) {
         params.n_batch = std::min(params.n_batch, n_kv);
     } else {
         params.n_batch = std::min(params.n_batch, params.n_ctx);
+        if (params.kl_divergence) {
+            params.n_parallel = 1;
+        } else {
+            // ensure there's at least enough seq_ids for HellaSwag
+            params.n_parallel = std::max(4, params.n_parallel);
+        }
     }
 
     if (params.ppl_stride > 0) {
@@ -2006,27 +2014,22 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
-    if (params.random_prompt) {
-        params.prompt = string_random_prompt(rng);
-    }
 
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
-    // ensure there's at least enough seq_ids for HellaSwag
-    params.n_parallel = std::max(4, params.n_parallel);
-
     // load the model and apply lora adapter, if any
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
+
     if (params.n_ctx > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, params.n_ctx);
diff --git a/llm/llama.cpp/examples/pydantic_models_to_grammar.py b/llm/llama.cpp/examples/pydantic_models_to_grammar.py
index 9acc7cc6dcd859f5cf66f7bf04d3c8ece12355ab..93e5dcb6c3855a1f0d33e9b333364828a0af8668 100644
--- a/llm/llama.cpp/examples/pydantic_models_to_grammar.py
+++ b/llm/llama.cpp/examples/pydantic_models_to_grammar.py
@@ -9,7 +9,7 @@ from inspect import getdoc, isclass
 from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin, get_type_hints
 
 from docstring_parser import parse
-from pydantic import BaseModel, Field, create_model
+from pydantic import BaseModel, create_model
 
 if TYPE_CHECKING:
     from types import GenericAlias
@@ -17,6 +17,9 @@ else:
     # python 3.8 compat
     from typing import _GenericAlias as GenericAlias
 
+# TODO: fix this
+# pyright: reportAttributeAccessIssue=information
+
 
 class PydanticDataType(Enum):
     """
@@ -50,35 +53,38 @@ class PydanticDataType(Enum):
 
 
 def map_pydantic_type_to_gbnf(pydantic_type: type[Any]) -> str:
-    if isclass(pydantic_type) and issubclass(pydantic_type, str):
+    origin_type = get_origin(pydantic_type)
+    origin_type = pydantic_type if origin_type is None else origin_type
+
+    if isclass(origin_type) and issubclass(origin_type, str):
         return PydanticDataType.STRING.value
-    elif isclass(pydantic_type) and issubclass(pydantic_type, bool):
+    elif isclass(origin_type) and issubclass(origin_type, bool):
         return PydanticDataType.BOOLEAN.value
-    elif isclass(pydantic_type) and issubclass(pydantic_type, int):
+    elif isclass(origin_type) and issubclass(origin_type, int):
         return PydanticDataType.INTEGER.value
-    elif isclass(pydantic_type) and issubclass(pydantic_type, float):
+    elif isclass(origin_type) and issubclass(origin_type, float):
         return PydanticDataType.FLOAT.value
-    elif isclass(pydantic_type) and issubclass(pydantic_type, Enum):
+    elif isclass(origin_type) and issubclass(origin_type, Enum):
         return PydanticDataType.ENUM.value
 
-    elif isclass(pydantic_type) and issubclass(pydantic_type, BaseModel):
-        return format_model_and_field_name(pydantic_type.__name__)
-    elif get_origin(pydantic_type) is list:
+    elif isclass(origin_type) and issubclass(origin_type, BaseModel):
+        return format_model_and_field_name(origin_type.__name__)
+    elif origin_type is list:
         element_type = get_args(pydantic_type)[0]
         return f"{map_pydantic_type_to_gbnf(element_type)}-list"
-    elif get_origin(pydantic_type) is set:
+    elif origin_type is set:
         element_type = get_args(pydantic_type)[0]
         return f"{map_pydantic_type_to_gbnf(element_type)}-set"
-    elif get_origin(pydantic_type) is Union:
+    elif origin_type is Union:
         union_types = get_args(pydantic_type)
         union_rules = [map_pydantic_type_to_gbnf(ut) for ut in union_types]
         return f"union-{'-or-'.join(union_rules)}"
-    elif get_origin(pydantic_type) is Optional:
+    elif origin_type is Optional:
         element_type = get_args(pydantic_type)[0]
         return f"optional-{map_pydantic_type_to_gbnf(element_type)}"
-    elif isclass(pydantic_type):
-        return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(pydantic_type.__name__)}"
-    elif get_origin(pydantic_type) is dict:
+    elif isclass(origin_type):
+        return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(origin_type.__name__)}"
+    elif origin_type is dict:
         key_type, value_type = get_args(pydantic_type)
         return f"custom-dict-key-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(key_type))}-value-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(value_type))}"
     else:
@@ -115,7 +121,7 @@ def get_members_structure(cls, rule_name):
         # Modify this comprehension
         members = [
             f'  "\\"{name}\\"" ":"  {map_pydantic_type_to_gbnf(param_type)}'
-            for name, param_type in cls.__annotations__.items()
+            for name, param_type in get_type_hints(cls).items()
             if name != "self"
         ]
 
@@ -234,8 +240,9 @@ def generate_gbnf_float_rules(max_digit=None, min_digit=None, max_precision=None
 
     # Define the integer part rule
     integer_part_rule = (
-        "integer-part" + (f"-max{max_digit}" if max_digit is not None else "") + (
-        f"-min{min_digit}" if min_digit is not None else "")
+        "integer-part"
+        + (f"-max{max_digit}" if max_digit is not None else "")
+        + (f"-min{min_digit}" if min_digit is not None else "")
     )
 
     # Define the fractional part rule based on precision constraints
@@ -293,17 +300,20 @@ def generate_gbnf_rule_for_type(
     field_name = format_model_and_field_name(field_name)
     gbnf_type = map_pydantic_type_to_gbnf(field_type)
 
-    if isclass(field_type) and issubclass(field_type, BaseModel):
+    origin_type = get_origin(field_type)
+    origin_type = field_type if origin_type is None else origin_type
+
+    if isclass(origin_type) and issubclass(origin_type, BaseModel):
         nested_model_name = format_model_and_field_name(field_type.__name__)
         nested_model_rules, _ = generate_gbnf_grammar(field_type, processed_models, created_rules)
         rules.extend(nested_model_rules)
         gbnf_type, rules = nested_model_name, rules
-    elif isclass(field_type) and issubclass(field_type, Enum):
+    elif isclass(origin_type) and issubclass(origin_type, Enum):
         enum_values = [f'"\\"{e.value}\\""' for e in field_type]  # Adding escaped quotes
         enum_rule = f"{model_name}-{field_name} ::= {' | '.join(enum_values)}"
         rules.append(enum_rule)
         gbnf_type, rules = model_name + "-" + field_name, rules
-    elif get_origin(field_type) == list:  # Array
+    elif origin_type is list:  # Array
         element_type = get_args(field_type)[0]
         element_rule_name, additional_rules = generate_gbnf_rule_for_type(
             model_name, f"{field_name}-element", element_type, is_optional, processed_models, created_rules
@@ -313,7 +323,7 @@ def generate_gbnf_rule_for_type(
         rules.append(array_rule)
         gbnf_type, rules = model_name + "-" + field_name, rules
 
-    elif get_origin(field_type) == set or field_type == set:  # Array
+    elif origin_type is set:  # Array
         element_type = get_args(field_type)[0]
         element_rule_name, additional_rules = generate_gbnf_rule_for_type(
             model_name, f"{field_name}-element", element_type, is_optional, processed_models, created_rules
@@ -367,7 +377,7 @@ def generate_gbnf_rule_for_type(
             gbnf_type = f"{model_name}-{field_name}-optional"
         else:
             gbnf_type = f"{model_name}-{field_name}-union"
-    elif isclass(field_type) and issubclass(field_type, str):
+    elif isclass(origin_type) and issubclass(origin_type, str):
         if field_info and hasattr(field_info, "json_schema_extra") and field_info.json_schema_extra is not None:
             triple_quoted_string = field_info.json_schema_extra.get("triple_quoted_string", False)
             markdown_string = field_info.json_schema_extra.get("markdown_code_block", False)
@@ -383,8 +393,8 @@ def generate_gbnf_rule_for_type(
             gbnf_type = PydanticDataType.STRING.value
 
     elif (
-        isclass(field_type)
-        and issubclass(field_type, float)
+        isclass(origin_type)
+        and issubclass(origin_type, float)
         and field_info
         and hasattr(field_info, "json_schema_extra")
         and field_info.json_schema_extra is not None
@@ -409,8 +419,8 @@ def generate_gbnf_rule_for_type(
         )
 
     elif (
-        isclass(field_type)
-        and issubclass(field_type, int)
+        isclass(origin_type)
+        and issubclass(origin_type, int)
         and field_info
         and hasattr(field_info, "json_schema_extra")
         and field_info.json_schema_extra is not None
@@ -458,7 +468,7 @@ def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[Bas
     if not issubclass(model, BaseModel):
         # For non-Pydantic classes, generate model_fields from __annotations__ or __init__
         if hasattr(model, "__annotations__") and model.__annotations__:
-            model_fields = {name: (typ, ...) for name, typ in model.__annotations__.items()}
+            model_fields = {name: (typ, ...) for name, typ in get_type_hints(model).items()}
         else:
             init_signature = inspect.signature(model.__init__)
             parameters = init_signature.parameters
@@ -466,7 +476,7 @@ def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[Bas
                             name != "self"}
     else:
         # For Pydantic models, use model_fields and check for ellipsis (required fields)
-        model_fields = model.__annotations__
+        model_fields = get_type_hints(model)
 
     model_rule_parts = []
     nested_rules = []
@@ -624,7 +634,7 @@ string ::= "\"" (
         "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
       )* "\"" ws
 ws ::= ([ \t\n] ws)?
-float ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
+float ::= ("-"? ([0] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
 
 integer ::= [0-9]+"""
 
@@ -680,7 +690,7 @@ def generate_markdown_documentation(
         str: Generated text documentation.
     """
     documentation = ""
-    pyd_models = [(model, True) for model in pydantic_models]
+    pyd_models: list[tuple[type[BaseModel], bool]] = [(model, True) for model in pydantic_models]
     for model, add_prefix in pyd_models:
         if add_prefix:
             documentation += f"{model_prefix}: {model.__name__}\n"
@@ -700,9 +710,9 @@ def generate_markdown_documentation(
             # Indenting the fields section
             documentation += f"  {fields_prefix}:\n"
         else:
-            documentation += f"  Fields:\n"
+            documentation += f"  Fields:\n"  # noqa: F541
         if isclass(model) and issubclass(model, BaseModel):
-            for name, field_type in model.__annotations__.items():
+            for name, field_type in get_type_hints(model).items():
                 # if name == "markdown_code_block":
                 #    continue
                 if get_origin(field_type) == list:
@@ -750,14 +760,17 @@ def generate_field_markdown(
     field_info = model.model_fields.get(field_name)
     field_description = field_info.description if field_info and field_info.description else ""
 
-    if get_origin(field_type) == list:
+    origin_type = get_origin(field_type)
+    origin_type = field_type if origin_type is None else origin_type
+
+    if origin_type == list:
         element_type = get_args(field_type)[0]
         field_text = f"{indent}{field_name} ({format_model_and_field_name(field_type.__name__)} of {format_model_and_field_name(element_type.__name__)})"
         if field_description != "":
             field_text += ":\n"
         else:
             field_text += "\n"
-    elif get_origin(field_type) == Union:
+    elif origin_type == Union:
         element_types = get_args(field_type)
         types = []
         for element_type in element_types:
@@ -778,7 +791,7 @@ def generate_field_markdown(
         return field_text
 
     if field_description != "":
-        field_text += f"        Description: " + field_description + "\n"
+        field_text += f"        Description: {field_description}\n"
 
     # Check for and include field-specific examples if available
     if hasattr(model, "Config") and hasattr(model.Config,
@@ -788,9 +801,9 @@ def generate_field_markdown(
             example_text = f"'{field_example}'" if isinstance(field_example, str) else field_example
             field_text += f"{indent}  Example: {example_text}\n"
 
-    if isclass(field_type) and issubclass(field_type, BaseModel):
+    if isclass(origin_type) and issubclass(origin_type, BaseModel):
         field_text += f"{indent}  Details:\n"
-        for name, type_ in field_type.__annotations__.items():
+        for name, type_ in get_type_hints(field_type).items():
             field_text += generate_field_markdown(name, type_, field_type, depth + 2)
 
     return field_text
@@ -833,7 +846,7 @@ def generate_text_documentation(
         str: Generated text documentation.
     """
     documentation = ""
-    pyd_models = [(model, True) for model in pydantic_models]
+    pyd_models: list[tuple[type[BaseModel], bool]] = [(model, True) for model in pydantic_models]
     for model, add_prefix in pyd_models:
         if add_prefix:
             documentation += f"{model_prefix}: {model.__name__}\n"
@@ -851,7 +864,7 @@ def generate_text_documentation(
 
         if isclass(model) and issubclass(model, BaseModel):
             documentation_fields = ""
-            for name, field_type in model.__annotations__.items():
+            for name, field_type in get_type_hints(model).items():
                 # if name == "markdown_code_block":
                 #    continue
                 if get_origin(field_type) == list:
@@ -944,7 +957,7 @@ def generate_field_text(
 
     if isclass(field_type) and issubclass(field_type, BaseModel):
         field_text += f"{indent}  Details:\n"
-        for name, type_ in field_type.__annotations__.items():
+        for name, type_ in get_type_hints(field_type).items():
             field_text += generate_field_text(name, type_, field_type, depth + 2)
 
     return field_text
@@ -1164,7 +1177,7 @@ def create_dynamic_model_from_function(func: Callable[..., Any]):
         dynamic_fields[param.name] = (
             param.annotation if param.annotation != inspect.Parameter.empty else str, default_value)
     # Creating the dynamic model
-    dynamic_model = create_model(f"{func.__name__}", **dynamic_fields)  # type: ignore[call-overload]
+    dynamic_model = create_model(f"{func.__name__}", **dynamic_fields)
 
     for name, param_doc in param_docs:
         dynamic_model.model_fields[name].description = param_doc.description
@@ -1228,9 +1241,6 @@ def map_grammar_names_to_pydantic_model_class(pydantic_model_list):
     return output
 
 
-from enum import Enum
-
-
 def json_schema_to_python_types(schema):
     type_map = {
         "any": Any,
@@ -1275,7 +1285,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name:
                     if items != {}:
                         array = {"properties": items}
                         array_type = convert_dictionary_to_pydantic_model(array, f"{model_name}_{field_name}_items")
-                        fields[field_name] = (List[array_type], ...)  # type: ignore[valid-type]
+                        fields[field_name] = (List[array_type], ...)
                     else:
                         fields[field_name] = (list, ...)
                 elif field_type == "object":
@@ -1285,7 +1295,8 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name:
                     required = field_data.get("enum", [])
                     for key, field in fields.items():
                         if key not in required:
-                            fields[key] = (Optional[fields[key][0]], ...)
+                            optional_type = fields[key][0]
+                            fields[key] = (Optional[optional_type], ...)
                 else:
                     field_type = json_schema_to_python_types(field_type)
                     fields[field_name] = (field_type, ...)
@@ -1305,6 +1316,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name:
         required = dictionary.get("required", [])
         for key, field in fields.items():
             if key not in required:
-                fields[key] = (Optional[fields[key][0]], ...)
+                optional_type = fields[key][0]
+                fields[key] = (Optional[optional_type], ...)
     custom_model = create_model(model_name, **fields)
     return custom_model
diff --git a/llm/llama.cpp/examples/pydantic_models_to_grammar_examples.py b/llm/llama.cpp/examples/pydantic_models_to_grammar_examples.py
new file mode 100644
index 0000000000000000000000000000000000000000..eb000d5ccba24b4fc304599ffcd6b36016074e58
--- /dev/null
+++ b/llm/llama.cpp/examples/pydantic_models_to_grammar_examples.py
@@ -0,0 +1,312 @@
+#!/usr/bin/env python3
+
+"""Function calling example using pydantic models."""
+
+from __future__ import annotations
+
+import argparse
+import datetime
+import json
+import logging
+import textwrap
+import sys
+from enum import Enum
+from typing import Optional, Union
+
+import requests
+from pydantic import BaseModel, Field
+from pydantic_models_to_grammar import (add_run_method_to_dynamic_model, convert_dictionary_to_pydantic_model,
+                                        create_dynamic_model_from_function, generate_gbnf_grammar_and_documentation)
+
+
+def create_completion(host, prompt, gbnf_grammar):
+    """Calls the /completion API on llama-server.
+
+    See
+    https://github.com/ggerganov/llama.cpp/tree/HEAD/examples/server#api-endpoints
+    """
+    print(f"  Request:\n    Grammar:\n{textwrap.indent(gbnf_grammar, '      ')}\n    Prompt:\n{textwrap.indent(prompt.rstrip(), '      ')}")
+    headers = {"Content-Type": "application/json"}
+    data = {"prompt": prompt, "grammar": gbnf_grammar}
+    result = requests.post(f"http://{host}/completion", headers=headers, json=data).json()
+    assert data.get("error") is None, data
+    logging.info("Result: %s", result)
+    content = result["content"]
+    print(f"  Model: {result['model']}")
+    print(f"  Result:\n{textwrap.indent(json.dumps(json.loads(content), indent=2), '    ')}")
+    return content
+
+
+# A function for the agent to send a message to the user.
+class SendMessageToUser(BaseModel):
+    """Send a message to the User."""
+    chain_of_thought: str = Field(..., description="Your chain of thought while sending the message.")
+    message: str = Field(..., description="Message you want to send to the user.")
+
+    def run(self):
+        print(f"SendMessageToUser: {self.message}")
+
+
+def example_rce(host):
+    """Minimal test case where the LLM call an arbitrary python function."""
+    print("- example_rce")
+    tools = [SendMessageToUser]
+    gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(
+        pydantic_model_list=tools, outer_object_name="function",
+        outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters")
+    system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation
+    user_message = "What is 42 * 42?"
+    prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
+    text = create_completion(host, prompt, gbnf_grammar)
+    json_data = json.loads(text)
+    tools_map = {tool.__name__:tool for tool in tools}
+    # This finds "SendMessageToUser":
+    tool = tools_map.get(json_data["function"])
+    if not tool:
+        print(f"Error: unknown tool {json_data['function']}")
+        return 1
+    tool(**json_data["function_parameters"]).run()
+    return 0
+
+
+# Enum for the calculator tool.
+class MathOperation(Enum):
+    ADD = "add"
+    SUBTRACT = "subtract"
+    MULTIPLY = "multiply"
+    DIVIDE = "divide"
+
+
+# Simple pydantic calculator tool for the agent that can add, subtract,
+# multiply, and divide. Docstring and description of fields will be used in
+# system prompt.
+class Calculator(BaseModel):
+    """Perform a math operation on two numbers."""
+    number_one: Union[int, float] = Field(..., description="First number.")
+    operation: MathOperation = Field(..., description="Math operation to perform.")
+    number_two: Union[int, float] = Field(..., description="Second number.")
+
+    def run(self):
+        if self.operation == MathOperation.ADD:
+            return self.number_one + self.number_two
+        elif self.operation == MathOperation.SUBTRACT:
+            return self.number_one - self.number_two
+        elif self.operation == MathOperation.MULTIPLY:
+            return self.number_one * self.number_two
+        elif self.operation == MathOperation.DIVIDE:
+            return self.number_one / self.number_two
+        else:
+            raise ValueError("Unknown operation.")
+
+
+def example_calculator(host):
+    """Have the LLM ask to get a calculation done.
+
+    Here the grammar gets generated by passing the available function models to
+    generate_gbnf_grammar_and_documentation function. This also generates a
+    documentation usable by the LLM.
+
+    pydantic_model_list is the list of pydantic models outer_object_name is an
+    optional name for an outer object around the actual model object. Like a
+    "function" object with "function_parameters" which contains the actual model
+    object. If None, no outer object will be generated outer_object_content is
+    the name of outer object content.
+
+    model_prefix is the optional prefix for models in the documentation. (Default="Output Model")
+    fields_prefix is the prefix for the model fields in the documentation. (Default="Output Fields")
+    """
+    print("- example_calculator")
+    tools = [SendMessageToUser, Calculator]
+    gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(
+        pydantic_model_list=tools, outer_object_name="function",
+        outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters")
+    system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation
+    user_message1 = "What is 42 * 42?"
+    prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message1}<|im_end|>\n<|im_start|>assistant"
+    text = create_completion(host, prompt, gbnf_grammar)
+    json_data = json.loads(text)
+    expected = {
+        "function": "Calculator",
+        "function_parameters": {
+            "number_one": 42,
+            "operation": "multiply",
+            "number_two": 42
+        }
+    }
+    if json_data != expected:
+        print("  Result is not as expected!")
+    tools_map = {tool.__name__:tool for tool in tools}
+    # This finds "Calculator":
+    tool = tools_map.get(json_data["function"])
+    if not tool:
+        print(f"Error: unknown tool {json_data['function']}")
+        return 1
+    result = tool(**json_data["function_parameters"]).run()
+    print(f"  Call {json_data['function']} gave result {result}")
+    return 0
+
+
+class Category(Enum):
+    """The category of the book."""
+    Fiction = "Fiction"
+    NonFiction = "Non-Fiction"
+
+
+class Book(BaseModel):
+    """Represents an entry about a book."""
+    title: str = Field(..., description="Title of the book.")
+    author: str = Field(..., description="Author of the book.")
+    published_year: Optional[int] = Field(..., description="Publishing year of the book.")
+    keywords: list[str] = Field(..., description="A list of keywords.")
+    category: Category = Field(..., description="Category of the book.")
+    summary: str = Field(..., description="Summary of the book.")
+
+
+def example_struct(host):
+    """A example structured output based on pydantic models.
+
+    The LLM will create an entry for a Book database out of an unstructured
+    text. We need no additional parameters other than our list of pydantic
+    models.
+    """
+    print("- example_struct")
+    tools = [Book]
+    gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(pydantic_model_list=tools)
+    system_message = "You are an advanced AI, tasked to create a dataset entry in JSON for a Book. The following is the expected output model:\n\n" + documentation
+    text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands."""
+    prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant"
+    text = create_completion(host, prompt, gbnf_grammar)
+    json_data = json.loads(text)
+    # In this case, there's no function nor function_parameters.
+    # Here the result will vary based on the LLM used.
+    keys = sorted(["title", "author", "published_year", "keywords", "category", "summary"])
+    if keys != sorted(json_data.keys()):
+        print(f"Unexpected result: {sorted(json_data.keys())}")
+        return 1
+    book = Book(**json_data)
+    print(f"  As a Book object: %s" % book)
+    return 0
+
+
+def get_current_datetime(output_format: Optional[str] = None):
+    """Get the current date and time in the given format.
+
+    Args:
+         output_format: formatting string for the date and time, defaults to '%Y-%m-%d %H:%M:%S'
+    """
+    return datetime.datetime.now().strftime(output_format or "%Y-%m-%d %H:%M:%S")
+
+
+# Example function to get the weather.
+def get_current_weather(location, unit):
+    """Get the current weather in a given location"""
+    if "London" in location:
+        return json.dumps({"location": "London", "temperature": "42", "unit": unit.value})
+    elif "New York" in location:
+        return json.dumps({"location": "New York", "temperature": "24", "unit": unit.value})
+    elif "North Pole" in location:
+        return json.dumps({"location": "North Pole", "temperature": "-42", "unit": unit.value})
+    return json.dumps({"location": location, "temperature": "unknown"})
+
+
+def example_concurrent(host):
+    """An example for parallel function calling with a Python function, a pydantic
+    function model and an OpenAI like function definition.
+    """
+    print("- example_concurrent")
+    # Function definition in OpenAI style.
+    current_weather_tool = {
+        "type": "function",
+        "function": {
+            "name": "get_current_weather",
+            "description": "Get the current weather in a given location",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "location": {
+                        "type": "string",
+                        "description": "The city and state, e.g. San Francisco, CA",
+                    },
+                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
+                },
+                "required": ["location"],
+            },
+        },
+    }
+    # Convert OpenAI function definition into pydantic model.
+    current_weather_tool_model = convert_dictionary_to_pydantic_model(current_weather_tool)
+    # Add the actual function to a pydantic model.
+    current_weather_tool_model = add_run_method_to_dynamic_model(current_weather_tool_model, get_current_weather)
+
+    # Convert normal Python function to a pydantic model.
+    current_datetime_model = create_dynamic_model_from_function(get_current_datetime)
+
+    tools = [SendMessageToUser, Calculator, current_datetime_model, current_weather_tool_model]
+    gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(
+        pydantic_model_list=tools, outer_object_name="function",
+        outer_object_content="params", model_prefix="Function", fields_prefix="Parameters", list_of_outputs=True)
+    system_message = "You are an advanced AI assistant. You are interacting with the user and with your environment by calling functions. You call functions by writing JSON objects, which represent specific function calls.\nBelow is a list of your available function calls:\n\n" + documentation
+    text = """Get the date and time, get the current weather in celsius in London and solve the following calculation: 42 * 42"""
+    prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant"
+    text = create_completion(host, prompt, gbnf_grammar)
+    json_data = json.loads(text)
+    expected = [
+      {
+        "function": "get_current_datetime",
+        "params": {
+          "output_format": "%Y-%m-%d %H:%M:%S"
+        }
+      },
+      {
+        "function": "get_current_weather",
+        "params": {
+          "location": "London",
+          "unit": "celsius"
+        }
+      },
+      {
+        "function": "Calculator",
+        "params": {
+          "number_one": 42,
+          "operation": "multiply",
+          "number_two": 42
+        }
+      }
+    ]
+    res = 0
+    if json_data != expected:
+        print("  Result is not as expected!")
+        print("  This can happen on highly quantized models")
+        res = 1
+    tools_map = {tool.__name__:tool for tool in tools}
+    for call in json_data:
+      tool = tools_map.get(call["function"])
+      if not tool:
+          print(f"Error: unknown tool {call['function']}")
+          return 1
+      result = tool(**call["params"]).run()
+      print(f"  Call {call['function']} returned {result}")
+    # Should output something like this:
+    #   Call get_current_datetime returned 2024-07-15 09:50:38
+    #   Call get_current_weather returned {"location": "London", "temperature": "42", "unit": "celsius"}
+    #   Call Calculator returned 1764
+    return res
+
+
+def main():
+    parser = argparse.ArgumentParser(description=sys.modules[__name__].__doc__)
+    parser.add_argument("--host", default="localhost:8080", help="llama.cpp server")
+    parser.add_argument("-v", "--verbose", action="store_true", help="enables logging")
+    args = parser.parse_args()
+    logging.basicConfig(level=logging.INFO if args.verbose else logging.ERROR)
+    ret = 0
+    # Comment out below to only run the example you want.
+    ret = ret or example_rce(args.host)
+    ret = ret or example_calculator(args.host)
+    ret = ret or example_struct(args.host)
+    ret = ret or example_concurrent(args.host)
+    return ret
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/llm/llama.cpp/examples/quantize-stats/CMakeLists.txt b/llm/llama.cpp/examples/quantize-stats/CMakeLists.txt
index e31cf5e3809c1e00dbc7bc89d31a8a7404ac67bd..bb986a716883dc5ea7a028f0c5d093209b13cafd 100644
--- a/llm/llama.cpp/examples/quantize-stats/CMakeLists.txt
+++ b/llm/llama.cpp/examples/quantize-stats/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET quantize-stats)
+set(TARGET llama-quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/quantize-stats/quantize-stats.cpp b/llm/llama.cpp/examples/quantize-stats/quantize-stats.cpp
index 746df8446b85e8a69191920104569dfe168e7b2e..68cf8d3595e878eade9304388aea81eedf1da39d 100644
--- a/llm/llama.cpp/examples/quantize-stats/quantize-stats.cpp
+++ b/llm/llama.cpp/examples/quantize-stats/quantize-stats.cpp
@@ -154,7 +154,7 @@ static void test_roundtrip_on_chunk(
     }
 
     if (use_reference) {
-        qfns.from_float_reference(input_scratch, quantized_scratch, chunk_size);
+        qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
     } else {
         qfns.from_float(input_scratch, quantized_scratch, chunk_size);
     }
diff --git a/llm/llama.cpp/examples/quantize/CMakeLists.txt b/llm/llama.cpp/examples/quantize/CMakeLists.txt
index 6b977fde86ab287572130ff049937d33302fae27..3ee4eb9719fc4076fedf393e5dca78fa1de1d6e7 100644
--- a/llm/llama.cpp/examples/quantize/CMakeLists.txt
+++ b/llm/llama.cpp/examples/quantize/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET quantize)
+set(TARGET llama-quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/quantize/README.md b/llm/llama.cpp/examples/quantize/README.md
index b78ece4e7f59dc493f7db85cac3cb5dbd213329f..553c2701bced3191179300855caed4d28ea7ce70 100644
--- a/llm/llama.cpp/examples/quantize/README.md
+++ b/llm/llama.cpp/examples/quantize/README.md
@@ -4,7 +4,89 @@ You can also use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-
 
 Note: It is synced from llama.cpp `main` every 6 hours.
 
-## Llama 2 7B
+Example usage:
+
+```bash
+# obtain the official LLaMA model weights and place them in ./models
+ls ./models
+llama-2-7b tokenizer_checklist.chk tokenizer.model
+# [Optional] for models using BPE tokenizers
+ls ./models
+ vocab.json
+# [Optional] for PyTorch .bin models like Mistral-7B
+ls ./models
+
+
+# install Python dependencies
+python3 -m pip install -r requirements.txt
+
+# convert the model to ggml FP16 format
+python3 convert_hf_to_gguf.py models/mymodel/
+
+# quantize the model to 4-bits (using Q4_K_M method)
+./llama-quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
+
+# update the gguf filetype to current version if older version is now unsupported
+./llama-quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
+```
+
+Run the quantized model:
+
+```bash
+# start inference on a gguf model
+./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
+```
+
+When running the larger models, make sure you have enough disk space to store all the intermediate files.
+
+## Memory/Disk Requirements
+
+As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
+
+| Model | Original size | Quantized size (Q4_0) |
+|------:|--------------:|----------------------:|
+|    7B |         13 GB |                3.9 GB |
+|   13B |         24 GB |                7.8 GB |
+|   30B |         60 GB |               19.5 GB |
+|   65B |        120 GB |               38.5 GB |
+
+## Quantization
+
+Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
+
+*(outdated)*
+
+| Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |
+|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
+|    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
+|    7B | file size    |  13.0G |   3.5G |   3.9G |   4.3G |   4.7G |   6.7G |
+|    7B | ms/tok @ 4th |    127 |     55 |     54 |     76 |     83 |     72 |
+|    7B | ms/tok @ 8th |    122 |     43 |     45 |     52 |     56 |     67 |
+|    7B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
+|   13B | perplexity   | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
+|   13B | file size    |  25.0G |   6.8G |   7.6G |   8.3G |   9.1G |    13G |
+|   13B | ms/tok @ 4th |      - |    103 |    105 |    148 |    160 |    131 |
+|   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
+|   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
+
+- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
+- recent k-quants improvements and new i-quants
+  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
+  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
+  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
+  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
+  - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
+  - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
+  - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
+  - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
+  - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
+  - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
+  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
+  - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
+  - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
+  - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
+
+**Llama 2 7B**
 
 | Quantization | Bits per Weight (BPW) |
 |--------------|-----------------------|
@@ -18,7 +100,8 @@ Note: It is synced from llama.cpp `main` every 6 hours.
 | Q5_K_M       | 5.68                  |
 | Q6_K         | 6.56                  |
 
-## Llama 2 13B
+**Llama 2 13B**
+
 Quantization | Bits per Weight (BPW)
 -- | --
 Q2_K | 3.34
@@ -31,7 +114,7 @@ Q5_K_S | 5.51
 Q5_K_M | 5.67
 Q6_K | 6.56
 
-# Llama 2 70B
+**Llama 2 70B**
 
 Quantization | Bits per Weight (BPW)
 -- | --
diff --git a/llm/llama.cpp/examples/quantize/quantize.cpp b/llm/llama.cpp/examples/quantize/quantize.cpp
index 28584e14b788c1254e307b1b4d20f6b8ceb1cf09..8d764725878b9927cd155ca7643df98b17595a9f 100644
--- a/llm/llama.cpp/examples/quantize/quantize.cpp
+++ b/llm/llama.cpp/examples/quantize/quantize.cpp
@@ -16,41 +16,44 @@ struct quant_option {
 };
 
 static const std::vector QUANT_OPTIONS = {
-    { "Q4_0",   LLAMA_FTYPE_MOSTLY_Q4_0,   " 3.56G, +0.2166 ppl @ LLaMA-v1-7B", },
-    { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
-    { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
-    { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
-    { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization",            },
-    { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
-    { "IQ2_S",  LLAMA_FTYPE_MOSTLY_IQ2_S,  " 2.5  bpw quantization",            },
-    { "IQ2_M",  LLAMA_FTYPE_MOSTLY_IQ2_M,  " 2.7  bpw quantization",            },
-    { "IQ1_S",  LLAMA_FTYPE_MOSTLY_IQ1_S,  " 1.56 bpw quantization",            },
-    { "IQ1_M",  LLAMA_FTYPE_MOSTLY_IQ1_M,  " 1.75 bpw quantization",            },
-    { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
-    { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
-    { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
-    { "IQ3_S",  LLAMA_FTYPE_MOSTLY_IQ3_S,  " 3.44 bpw quantization",            },
-    { "IQ3_M",  LLAMA_FTYPE_MOSTLY_IQ3_M,  " 3.66 bpw quantization mix",        },
-    { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
-    { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization"   ,          },
-    { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
-    { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
-    { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
-    { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", },
-    { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", },
-    { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
-    { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
-    { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
-    { "Q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
-    { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
-    { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
-    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, +0.0008 ppl @ LLaMA-v1-7B", },
-    { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
-    { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "14.00G, -0.0020 ppl @ Mistral-7B", },
-    { "BF16",   LLAMA_FTYPE_MOSTLY_BF16,   "14.00G, -0.0050 ppl @ Mistral-7B", },
-    { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
+    { "Q4_0",     LLAMA_FTYPE_MOSTLY_Q4_0,     " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
+    { "Q4_1",     LLAMA_FTYPE_MOSTLY_Q4_1,     " 4.78G, +0.4511 ppl @ Llama-3-8B",  },
+    { "Q5_0",     LLAMA_FTYPE_MOSTLY_Q5_0,     " 5.21G, +0.1316 ppl @ Llama-3-8B",  },
+    { "Q5_1",     LLAMA_FTYPE_MOSTLY_Q5_1,     " 5.65G, +0.1062 ppl @ Llama-3-8B",  },
+    { "IQ2_XXS",  LLAMA_FTYPE_MOSTLY_IQ2_XXS,  " 2.06 bpw quantization",            },
+    { "IQ2_XS",   LLAMA_FTYPE_MOSTLY_IQ2_XS,   " 2.31 bpw quantization",            },
+    { "IQ2_S",    LLAMA_FTYPE_MOSTLY_IQ2_S,    " 2.5  bpw quantization",            },
+    { "IQ2_M",    LLAMA_FTYPE_MOSTLY_IQ2_M,    " 2.7  bpw quantization",            },
+    { "IQ1_S",    LLAMA_FTYPE_MOSTLY_IQ1_S,    " 1.56 bpw quantization",            },
+    { "IQ1_M",    LLAMA_FTYPE_MOSTLY_IQ1_M,    " 1.75 bpw quantization",            },
+    { "Q2_K",     LLAMA_FTYPE_MOSTLY_Q2_K,     " 2.96G, +3.5199 ppl @ Llama-3-8B",  },
+    { "Q2_K_S",   LLAMA_FTYPE_MOSTLY_Q2_K_S,   " 2.96G, +3.1836 ppl @ Llama-3-8B",  },
+    { "IQ3_XXS",  LLAMA_FTYPE_MOSTLY_IQ3_XXS,  " 3.06 bpw quantization",            },
+    { "IQ3_S",    LLAMA_FTYPE_MOSTLY_IQ3_S,    " 3.44 bpw quantization",            },
+    { "IQ3_M",    LLAMA_FTYPE_MOSTLY_IQ3_M,    " 3.66 bpw quantization mix",        },
+    { "Q3_K",     LLAMA_FTYPE_MOSTLY_Q3_K_M,   "alias for Q3_K_M"                   },
+    { "IQ3_XS",   LLAMA_FTYPE_MOSTLY_IQ3_XS,   " 3.3 bpw quantization",             },
+    { "Q3_K_S",   LLAMA_FTYPE_MOSTLY_Q3_K_S,   " 3.41G, +1.6321 ppl @ Llama-3-8B",  },
+    { "Q3_K_M",   LLAMA_FTYPE_MOSTLY_Q3_K_M,   " 3.74G, +0.6569 ppl @ Llama-3-8B",  },
+    { "Q3_K_L",   LLAMA_FTYPE_MOSTLY_Q3_K_L,   " 4.03G, +0.5562 ppl @ Llama-3-8B",  },
+    { "IQ4_NL",   LLAMA_FTYPE_MOSTLY_IQ4_NL,   " 4.50 bpw non-linear quantization", },
+    { "IQ4_XS",   LLAMA_FTYPE_MOSTLY_IQ4_XS,   " 4.25 bpw non-linear quantization", },
+    { "Q4_K",     LLAMA_FTYPE_MOSTLY_Q4_K_M,   "alias for Q4_K_M",                  },
+    { "Q4_K_S",   LLAMA_FTYPE_MOSTLY_Q4_K_S,   " 4.37G, +0.2689 ppl @ Llama-3-8B",  },
+    { "Q4_K_M",   LLAMA_FTYPE_MOSTLY_Q4_K_M,   " 4.58G, +0.1754 ppl @ Llama-3-8B",  },
+    { "Q5_K",     LLAMA_FTYPE_MOSTLY_Q5_K_M,   "alias for Q5_K_M",                  },
+    { "Q5_K_S",   LLAMA_FTYPE_MOSTLY_Q5_K_S,   " 5.21G, +0.1049 ppl @ Llama-3-8B",  },
+    { "Q5_K_M",   LLAMA_FTYPE_MOSTLY_Q5_K_M,   " 5.33G, +0.0569 ppl @ Llama-3-8B",  },
+    { "Q6_K",     LLAMA_FTYPE_MOSTLY_Q6_K,     " 6.14G, +0.0217 ppl @ Llama-3-8B",  },
+    { "Q8_0",     LLAMA_FTYPE_MOSTLY_Q8_0,     " 7.96G, +0.0026 ppl @ Llama-3-8B",  },
+    { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
+    { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
+    { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
+    { "F16",      LLAMA_FTYPE_MOSTLY_F16,      "14.00G, +0.0020 ppl @ Mistral-7B",  },
+    { "BF16",     LLAMA_FTYPE_MOSTLY_BF16,     "14.00G, -0.0050 ppl @ Mistral-7B",  },
+    { "F32",      LLAMA_FTYPE_ALL_F32,         "26.00G              @ 7B",          },
     // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching.
-    { "COPY",   LLAMA_FTYPE_ALL_F32,       "only copy tensors, no quantizing", },
+    { "COPY",     LLAMA_FTYPE_ALL_F32,         "only copy tensors, no quantizing",  },
 };
 
 static const char * const LLM_KV_QUANTIZE_IMATRIX_FILE       = "quantize.imatrix.file";
diff --git a/llm/llama.cpp/examples/quantize/tests.sh b/llm/llama.cpp/examples/quantize/tests.sh
index a3ca74c68e7e5e60abe7946db734924d533836b4..24bc970e8632b5f90396b2bdb467ed4a81fb0da0 100644
--- a/llm/llama.cpp/examples/quantize/tests.sh
+++ b/llm/llama.cpp/examples/quantize/tests.sh
@@ -18,9 +18,9 @@ fi
 
 set -x
 
-SPLIT=$1/gguf-split
-QUANTIZE=$1/quantize
-MAIN=$1/main
+SPLIT=$1/llama-gguf-split
+QUANTIZE=$1/llama-quantize
+MAIN=$1/llama-cli
 WORK_PATH=$TMP_DIR/quantize
 ROOT_DIR=$(realpath $(dirname $0)/../../)
 
@@ -47,7 +47,7 @@ echo PASS
 echo
 
 # 3a. Test the requanted model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --random-prompt --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --n-predict 32
 echo PASS
 echo
 
@@ -57,7 +57,7 @@ echo PASS
 echo
 
 # 4b. Test the requanted model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-requant-merge.gguf --random-prompt --n-predict 32
+$MAIN --model $WORK_PATH/ggml-model-requant-merge.gguf --n-predict 32
 echo PASS
 echo
 
diff --git a/llm/llama.cpp/examples/reason-act.sh b/llm/llama.cpp/examples/reason-act.sh
index 046c48db584bc31f9c5fd57f84c21d84932a3511..06d592799cf1201746b79ac74001fc38da973a6b 100644
--- a/llm/llama.cpp/examples/reason-act.sh
+++ b/llm/llama.cpp/examples/reason-act.sh
@@ -8,7 +8,7 @@ if [ "$1" == "-m" ]; then
   MODEL="-m $2 "
 fi
 
-./main $MODEL --color \
+./llama-cli $MODEL --color \
     -f ./prompts/reason-act.txt \
     -i --interactive-first \
     --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7 -c 2048 \
diff --git a/llm/llama.cpp/examples/regex_to_grammar.py b/llm/llama.cpp/examples/regex_to_grammar.py
new file mode 100644
index 0000000000000000000000000000000000000000..5cd9210a4dfc672be4222369e7a46e7b92b31a1a
--- /dev/null
+++ b/llm/llama.cpp/examples/regex_to_grammar.py
@@ -0,0 +1,20 @@
+import json, subprocess, sys, os
+
+assert len(sys.argv) >= 2
+[_, pattern, *rest] = sys.argv
+
+print(subprocess.check_output(
+    [
+        "python",
+        os.path.join(
+        os.path.dirname(os.path.realpath(__file__)),
+        "json_schema_to_grammar.py"),
+        *rest,
+        "-",
+        "--raw-pattern",
+    ],
+    text=True,
+    input=json.dumps({
+        "type": "string",
+        "pattern": pattern,
+    }, indent=2)))
diff --git a/llm/llama.cpp/examples/retrieval/CMakeLists.txt b/llm/llama.cpp/examples/retrieval/CMakeLists.txt
index eaabae08d5583fa21c8daa20188e05f8f7c29778..66610f31114050a1b3ef0db45d4f482c1a59dd3d 100644
--- a/llm/llama.cpp/examples/retrieval/CMakeLists.txt
+++ b/llm/llama.cpp/examples/retrieval/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET retrieval)
+set(TARGET llama-retrieval)
 add_executable(${TARGET} retrieval.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/retrieval/README.md b/llm/llama.cpp/examples/retrieval/README.md
index 2b2595c468046c570b6af04815ab84ed238fb7b1..bc5f22e2ff15662a8232725670d1eaaf79384d9f 100644
--- a/llm/llama.cpp/examples/retrieval/README.md
+++ b/llm/llama.cpp/examples/retrieval/README.md
@@ -15,7 +15,7 @@ https://github.com/ggerganov/llama.cpp/pull/6193
 `retrieval` example can be tested as follows:
 
 ```bash
-make -j && ./retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
+make -j && ./llama-retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
 ```
 
 This chunks and embeds all given files and starts a loop requesting query inputs:
diff --git a/llm/llama.cpp/examples/retrieval/retrieval.cpp b/llm/llama.cpp/examples/retrieval/retrieval.cpp
index 4e7530706d4a9278a03dcd1640ed0829c8203014..65b19ce71cbe3e4f75062d3a19c24a9c8f8fcdfd 100644
--- a/llm/llama.cpp/examples/retrieval/retrieval.cpp
+++ b/llm/llama.cpp/examples/retrieval/retrieval.cpp
@@ -4,72 +4,12 @@
 #include 
 #include 
 
-struct retrieval_params {
-    std::vector context_files; // context files to embed
-    int32_t chunk_size            = 64;     // chunk size for context embedding
-    std::string chunk_separator   = "\n";   // chunk separator for context embedding
-};
-
-static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) {
-    gpt_params_print_usage(argc, argv, gpt_params);
-    printf("retrieval options:\n");
-    printf("  --context-file FNAME  file containing context to embed.\n");
-    printf("                        specify multiple files by providing --context-file option multiple times.\n");
-    printf("  --chunk-size N        minimum length of embedded text chunk (default:%d)\n", params.chunk_size);
-    printf("  --chunk-separator STRING\n");
-    printf("                        string to separate chunks (default: \"\\n\")\n");
-    printf("\n");
-}
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
 
-static void retrieval_params_parse(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & retrieval_params) {
-    int i = 1;
-    std::string arg;
-    while (i < argc) {
-        arg = argv[i];
-        bool invalid_gpt_param = false;
-        if(gpt_params_find_arg(argc, argv, argv[i], gpt_params, i, invalid_gpt_param)) {
-            if (invalid_gpt_param) {
-                fprintf(stderr, "error: invalid argument: %s\n", arg.c_str());
-                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
-                exit(1);
-            }
-            // option was parsed by gpt_params_find_arg
-        } else if (arg == "--context-file") {
-            if (++i >= argc) {
-                fprintf(stderr, "error: missing argument for --context-file\n");
-                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
-                exit(1);
-            }
-            std::ifstream file(argv[i]);
-            if (!file) {
-                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
-                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
-                exit(1);
-            }
-            // store the external file name in params
-            retrieval_params.context_files.push_back(argv[i]);
-        } else if (arg == "--chunk-size") {
-            if (++i >= argc) {
-                fprintf(stderr, "error: missing argument for --chunk-size\n");
-                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
-                exit(1);
-            }
-            retrieval_params.chunk_size = std::stoi(argv[i]);
-        } else if (arg == "--chunk-separator") {
-            if (++i >= argc) {
-                fprintf(stderr, "error: missing argument for --chunk-separator\n");
-                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
-                exit(1);
-            }
-            retrieval_params.chunk_separator = argv[i];
-        } else {
-            // unknown argument
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
-            exit(1);
-        }
-        i++;
-    }
+    LOG_TEE("\nexample usage:\n");
+    LOG_TEE("\n    %s --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .\n", argv[0]);
+    LOG_TEE("\n");
 }
 
 struct chunk {
@@ -133,9 +73,10 @@ static std::vector chunk_file(const std::string & filename, int chunk_siz
     return chunks;
 }
 
-static void batch_add_seq(llama_batch & batch, const std::vector & tokens, int seq_id) {
-    for (size_t i = 0; i < tokens.size(); i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
+static void batch_add_seq(llama_batch & batch, const std::vector & tokens, llama_seq_id seq_id) {
+    size_t n_tokens = tokens.size();
+    for (size_t i = 0; i < n_tokens; i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, true);
     }
 }
 
@@ -171,33 +112,35 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    retrieval_params retrieval_params;
 
-    retrieval_params_parse(argc, argv, params, retrieval_params);
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
+    }
 
     // For BERT models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
+    params.embedding = true;
 
-    if (retrieval_params.chunk_size <= 0) {
+    if (params.chunk_size <= 0) {
         fprintf(stderr, "chunk_size must be positive\n");
         return 1;
     }
-    if (retrieval_params.context_files.empty()) {
+    if (params.context_files.empty()) {
         fprintf(stderr, "context_files must be specified\n");
         return 1;
     }
-    params.embedding = true;
 
     print_build_info();
 
     printf("processing files:\n");
-    for (auto & context_file : retrieval_params.context_files) {
+    for (auto & context_file : params.context_files) {
         printf("%s\n", context_file.c_str());
     }
 
     std::vector chunks;
-    for (auto & context_file : retrieval_params.context_files) {
-        std::vector file_chunk = chunk_file(context_file, retrieval_params.chunk_size, retrieval_params.chunk_separator);
+    for (auto & context_file : params.context_files) {
+        std::vector file_chunk = chunk_file(context_file, params.chunk_size, params.chunk_separator);
         chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end());
     }
     printf("Number of chunks: %ld\n", chunks.size());
@@ -205,11 +148,12 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-
     // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
+
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -218,6 +162,12 @@ int main(int argc, char ** argv) {
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
 
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
+        return 1;
+    }
+
     if (n_ctx > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, n_ctx);
@@ -242,7 +192,7 @@ int main(int argc, char ** argv) {
             return 1;
         }
         // add eos if not present
-        if (inp.empty() || inp.back() != llama_token_eos(model)) {
+        if (llama_token_eos(model) >= 0 && (inp.empty() || inp.back() != llama_token_eos(model))) {
             inp.push_back(llama_token_eos(model));
         }
         chunk.tokens = inp;
diff --git a/llm/llama.cpp/examples/rpc/README.md b/llm/llama.cpp/examples/rpc/README.md
index eeec71a8ee0c2e3ad9cd4c5faa28301d0bef1d98..e1da801f285c636db093b57233dba07353333a8d 100644
--- a/llm/llama.cpp/examples/rpc/README.md
+++ b/llm/llama.cpp/examples/rpc/README.md
@@ -29,13 +29,13 @@ You can also run multiple `rpc-server` instances on the same host, each with a d
 
 ## Usage
 
-On each host, build the corresponding backend with `cmake` and add `-DLLAMA_RPC=ON` to the build options.
+On each host, build the corresponding backend with `cmake` and add `-DGGML_RPC=ON` to the build options.
 For example, to build the CUDA backend with RPC support:
 
 ```bash
 mkdir build-rpc-cuda
 cd build-rpc-cuda
-cmake .. -DLLAMA_CUDA=ON -DLLAMA_RPC=ON
+cmake .. -DGGML_CUDA=ON -DGGML_RPC=ON
 cmake --build . --config Release
 ```
 
@@ -58,17 +58,17 @@ $ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
 This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
 
 
-On the main host build `llama.cpp` only with `-DLLAMA_RPC=ON`:
+On the main host build `llama.cpp` only with `-DGGML_RPC=ON`:
 
 ```bash
 mkdir build-rpc
 cd build-rpc
-cmake .. -DLLAMA_RPC=ON
+cmake .. -DGGML_RPC=ON
 cmake --build . --config Release
 ```
 
 Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:
 
 ```bash
-$ bin/main -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
+$ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
 ```
diff --git a/llm/llama.cpp/examples/save-load-state/CMakeLists.txt b/llm/llama.cpp/examples/save-load-state/CMakeLists.txt
index cc6ed8554a6e344f5df97c2f425f008dcdf991eb..0fb5e359bc9ad1023744df073f4893b3ee763469 100644
--- a/llm/llama.cpp/examples/save-load-state/CMakeLists.txt
+++ b/llm/llama.cpp/examples/save-load-state/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET save-load-state)
+set(TARGET llama-save-load-state)
 add_executable(${TARGET} save-load-state.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/save-load-state/save-load-state.cpp b/llm/llama.cpp/examples/save-load-state/save-load-state.cpp
index c3b766882dbec788ec7c324eec98890a877c14d4..3ea7c790d2bf782138124d0ea424f109b64e08b2 100644
--- a/llm/llama.cpp/examples/save-load-state/save-load-state.cpp
+++ b/llm/llama.cpp/examples/save-load-state/save-load-state.cpp
@@ -11,6 +11,7 @@ int main(int argc, char ** argv) {
     params.prompt = "The quick brown fox";
 
     if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -27,10 +28,11 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    llama_model * model;
-    llama_context * ctx;
+    llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
@@ -46,7 +48,7 @@ int main(int argc, char ** argv) {
     // save state (rng, logits, embedding and kv_cache) to file
     {
         std::vector state_mem(llama_state_get_size(ctx));
-        const size_t written = llama_state_get_data(ctx, state_mem.data());
+        const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size());
 
         FILE *fp_write = fopen("dump_state.bin", "wb");
         fwrite(state_mem.data(), 1, written, fp_write);
@@ -98,13 +100,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector state_mem(llama_state_get_size(ctx2));
+        std::vector state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx2, state_mem.data())) {
+        if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -158,13 +163,16 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector state_mem(llama_state_get_size(ctx3));
+        std::vector state_mem;
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        fseek(fp_read, 0, SEEK_END);
+        state_mem.resize(ftell(fp_read));
+        fseek(fp_read, 0, SEEK_SET);
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_state_set_data(ctx3, state_mem.data())) {
+        if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx3);
             llama_free_model(model);
@@ -181,7 +189,7 @@ int main(int argc, char ** argv) {
     {
         // save kv of seq 0
         std::vector seq_store(llama_state_seq_get_size(ctx3, 0));
-        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), 0);
+        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
         if (ncopy != seq_store.size()) {
             fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
             llama_free(ctx3);
@@ -195,7 +203,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1
-        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), 1);
+        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
         if (nset != seq_store.size()) {
             fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
             llama_free(ctx3);
diff --git a/llm/llama.cpp/examples/server-llama2-13B.sh b/llm/llama.cpp/examples/server-llama2-13B.sh
index 17fedc2b176f611fe508e4d38a5a035a4c21bb55..4ce79b7fac4774dbad78d28136c587e737fb924d 100644
--- a/llm/llama.cpp/examples/server-llama2-13B.sh
+++ b/llm/llama.cpp/examples/server-llama2-13B.sh
@@ -16,7 +16,7 @@ GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 4096 --batch-size 1024}"
 
 
 # shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
-./server $GEN_OPTIONS \
+./llama-server $GEN_OPTIONS \
   --model "$MODEL" \
   --threads "$N_THREAD" \
   --rope-freq-scale 1.0 \
diff --git a/llm/llama.cpp/examples/server/CMakeLists.txt b/llm/llama.cpp/examples/server/CMakeLists.txt
index 4b89c53022fd03cd9cecbf28d301bfbfd6392f73..dbe41f1fd112055eae0cd356767c1b46ef785cf7 100644
--- a/llm/llama.cpp/examples/server/CMakeLists.txt
+++ b/llm/llama.cpp/examples/server/CMakeLists.txt
@@ -1,18 +1,37 @@
-set(TARGET server)
+set(TARGET llama-server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF)
+option(LLAMA_SERVER_SSL     "Build SSL support for the server"        OFF)
+
 include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
+
+if (MINGW)
+    # fix: https://github.com/ggerganov/llama.cpp/actions/runs/9651004652/job/26617901362?pr=8006
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
 set(TARGET_SRCS
     server.cpp
     utils.hpp
     httplib.h
 )
 set(PUBLIC_ASSETS
+    colorthemes.css
+    style.css
+    theme-beeninorder.css
+    theme-ketivah.css
+    theme-mangotango.css
+    theme-playground.css
+    theme-polarnight.css
+    theme-snowstorm.css
     index.html
+    index-new.html
     index.js
     completion.js
+    system-prompts.js
+    prompt-formats.js
     json-schema-to-grammar.mjs
 )
+
 foreach(asset ${PUBLIC_ASSETS})
     set(input "${CMAKE_CURRENT_SOURCE_DIR}/public/${asset}")
     set(output "${CMAKE_CURRENT_BINARY_DIR}/${asset}.hpp")
@@ -23,18 +42,23 @@ foreach(asset ${PUBLIC_ASSETS})
         COMMAND "${CMAKE_COMMAND}" "-DINPUT=${input}" "-DOUTPUT=${output}" -P "${PROJECT_SOURCE_DIR}/scripts/xxd.cmake"
     )
 endforeach()
+
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$
 )
+
 target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
+
 if (LLAMA_SERVER_SSL)
     find_package(OpenSSL REQUIRED)
     target_link_libraries(${TARGET} PRIVATE OpenSSL::SSL OpenSSL::Crypto)
     target_compile_definitions(${TARGET} PRIVATE CPPHTTPLIB_OPENSSL_SUPPORT)
 endif()
+
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
+
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/llm/llama.cpp/examples/server/README.md b/llm/llama.cpp/examples/server/README.md
index 0c3db8c84c69d0ffbc766b76bee88d338c943ba5..e17595fe87f2549eede6114ee0a9623b25fdc627 100644
--- a/llm/llama.cpp/examples/server/README.md
+++ b/llm/llama.cpp/examples/server/README.md
@@ -5,7 +5,7 @@ Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/
 Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 
 **Features:**
- * LLM inference of F16 and quantum models on GPU and CPU
+ * LLM inference of F16 and quantized models on GPU and CPU
  * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
  * Parallel decoding with multi-user support
  * Continuous batching
@@ -15,91 +15,261 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 
 The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
 
-**Command line options:**
-
-- `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response.
-- `-t N`, `--threads N`: Set the number of threads to use by CPU layers during generation. Not used by model layers that are offloaded to GPU. This option has no effect when using the maximum number of GPU layers. Default: `std::thread::hardware_concurrency()` (number of CPU cores).
-- `-tb N, --threads-batch N`: Set the number of threads to use by CPU layers during batch and prompt processing (>= 32 tokens). This option has no effect if a GPU is available. Default: `--threads`.
-- `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`
-- `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
-- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file. Default: unused
-- `-hfr REPO, --hf-repo REPO`: Hugging Face model repository. Default: unused
-- `-hff FILE, --hf-file FILE`: Hugging Face model file. Default: unused
-- `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
-- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is `512`, but LLaMA models were built with a context of `2048`, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of `4096`.
-- `-ngl N`, `--n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
-- `-mg i, --main-gpu i`: When using multiple GPUs, this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default, GPU `0` is used.
-- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs, this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default, the data is split in proportion to VRAM, but this may not be optimal for performance.
-- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `2048`
-- `-ub N`, `--ubatch-size N`: Physical maximum batch size. Default: `512`
-- `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
-- `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
-- `--numa STRATEGY`: Attempt one of the below optimization strategies that may help on some NUMA systems
-- `--numa distribute`: Spread execution evenly over all nodes
-- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
-- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437
-- `--numa`: Attempt optimizations that may help on some NUMA systems.
-- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
-- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
-- `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`
-- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`
-- `--port`: Set the port to listen. Default: `8080`
-- `--path`: Path from which to serve static files. Default: disabled
-- `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
-- `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s.
-- `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled
-- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`. Values > 1 will allow for higher throughput with multiple parallel requests but the results will **not** be deterministic due to differences in rounding error.
-- `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching).  Default: disabled
-- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)
-- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
-- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend. Used together with group attention width `--grp-attn-w`. Default: `1`, which is disabled.
-- `--grp-attn-w`: Set the group attention width to extend context size through self-extend.  Used together with group attention factor `--grp-attn-n`. Default: `512`
-- `-n N, --n-predict N`: Set the maximum tokens to predict. Default: `-1`
-- `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
-- `--metrics`: enable prometheus `/metrics` compatible endpoint. Default: disabled
-- `--slot-save-path PATH`: Specifies the path where the state of slots (the prompt cache) can be stored. If not provided, the slot management endpoints will be disabled.
-- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name.  Default: template taken from model's metadata. We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- `--log-disable`: Output logs to stdout only, not to `llama.log`. Default: enabled
-- `--log-format FORMAT`: Define the log output to FORMAT: json or text Default: `json`
-- `--rope-scaling` : RoPE scaling method. Defaults to linear unless otherwise specified by the model. Options are `none`, `linear`, `yarn`
-- `--rope-freq-base N` : RoPE frequency base (default: loaded from model)
-- `--rope-freq-scale N`: RoPE frequency scaling factor, expands context by a factor of 1/N (e.g. 0.25)
-- `--yarn-ext-factor N` : YaRN: extrapolation mix factor (Default: 1.0, 0.0 = full interpolation)
-- `--yarn-attn-factor N` : YaRN: scale sqrt(t) or attention magnitude (default: 1.0)
-- `--yarn-beta-slow N`: YaRN: High correction dim or alpha (default: 1.0)
-- `--yarn-beta-fast N`: YaRN: low correction dim or beta (default: 32.0)
-- `--pooling` : Pooling type for embeddings, use model default if unspecified. Options are `none`, `mean`, `cls`
-- `-dt N`, `--defrag-thold N`: KV cache defragmentation threshold (default: -1.0, < 0 = disabled)
-- `-fa`, `--flash-attn` : enable flash attention (default: disabled).
-- `-ctk TYPE`, `--cache-type-k TYPE` : KV cache data type for K (default: `f16`, options `f32`, `f16`, `q8_0`, `q4_0`, `q4_1`, `iq4_nl`, `q5_0`, or `q5_1`)
-- `-ctv TYPE`, `--cache-type-v TYPE` : KV cache type for V (default `f16`, see `-ctk` for options)
-
-**If compiled with `LLAMA_SERVER_SSL=ON`**
-- `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key
-- `--ssl-cert-file FNAME`: path to file a PEM-encoded SSL certificate
+## Usage
+
+```
+usage: ./llama-server [options]
+
+general:
+
+  -h,    --help, --usage          print usage and exit
+         --version                show version and build info
+  -v,    --verbose                print verbose information
+         --verbosity N            set specific verbosity level (default: 0)
+         --verbose-prompt         print a verbose prompt before generation (default: false)
+         --no-display-prompt      don't print prompt at generation (default: false)
+  -co,   --color                  colorise output to distinguish prompt and user input from generations (default: false)
+  -s,    --seed SEED              RNG seed (default: -1, use random seed for < 0)
+  -t,    --threads N              number of threads to use during generation (default: 8)
+  -tb,   --threads-batch N        number of threads to use during batch and prompt processing (default: same as --threads)
+  -td,   --threads-draft N        number of threads to use during generation (default: same as --threads)
+  -tbd,  --threads-batch-draft N  number of threads to use during batch and prompt processing (default: same as --threads-draft)
+         --draft N                number of tokens to draft for speculative decoding (default: 5)
+  -ps,   --p-split N              speculative decoding split probability (default: 0.1)
+  -lcs,  --lookup-cache-static FNAME
+                                  path to static lookup cache to use for lookup decoding (not updated by generation)
+  -lcd,  --lookup-cache-dynamic FNAME
+                                  path to dynamic lookup cache to use for lookup decoding (updated by generation)
+  -c,    --ctx-size N             size of the prompt context (default: 0, 0 = loaded from model)
+  -n,    --predict N              number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)
+  -b,    --batch-size N           logical maximum batch size (default: 2048)
+  -ub,   --ubatch-size N          physical maximum batch size (default: 512)
+         --keep N                 number of tokens to keep from the initial prompt (default: 0, -1 = all)
+         --chunks N               max number of chunks to process (default: -1, -1 = all)
+  -fa,   --flash-attn             enable Flash Attention (default: disabled)
+  -p,    --prompt PROMPT          prompt to start generation with
+                                  in conversation mode, this will be used as system prompt
+                                  (default: '')
+  -f,    --file FNAME             a file containing the prompt (default: none)
+         --in-file FNAME          an input file (repeat to specify multiple files)
+  -bf,   --binary-file FNAME      binary file containing the prompt (default: none)
+  -e,    --escape                 process escapes sequences (\n, \r, \t, \', \", \\) (default: true)
+         --no-escape              do not process escape sequences
+  -ptc,  --print-token-count N    print token count every N tokens (default: -1)
+         --prompt-cache FNAME     file to cache prompt state for faster startup (default: none)
+         --prompt-cache-all       if specified, saves user input and generations to cache as well
+                                  not supported with --interactive or other interactive options
+         --prompt-cache-ro        if specified, uses the prompt cache but does not update it
+  -r,    --reverse-prompt PROMPT  halt generation at PROMPT, return control in interactive mode
+                                  can be specified more than once for multiple prompts
+  -sp,   --special                special tokens output enabled (default: false)
+  -cnv,  --conversation           run in conversation mode, does not print special tokens and suffix/prefix
+                                  if suffix/prefix are not specified, default chat template will be used
+                                  (default: false)
+  -i,    --interactive            run in interactive mode (default: false)
+  -if,   --interactive-first      run in interactive mode and wait for input right away (default: false)
+  -mli,  --multiline-input        allows you to write or paste multiple lines without ending each in '\'
+         --in-prefix-bos          prefix BOS to user inputs, preceding the `--in-prefix` string
+         --in-prefix STRING       string to prefix user inputs with (default: empty)
+         --in-suffix STRING       string to suffix after user inputs with (default: empty)
+         --spm-infill             use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled)
+
+sampling:
+
+         --samplers SAMPLERS      samplers that will be used for generation in the order, separated by ';'
+                                  (default: top_k;tfs_z;typical_p;top_p;min_p;temperature)
+         --sampling-seq SEQUENCE  simplified sequence for samplers that will be used (default: kfypmt)
+         --ignore-eos             ignore end of stream token and continue generating (implies --logit-bias EOS-inf)
+         --penalize-nl            penalize newline tokens (default: false)
+         --temp N                 temperature (default: 0.8)
+         --top-k N                top-k sampling (default: 40, 0 = disabled)
+         --top-p N                top-p sampling (default: 0.9, 1.0 = disabled)
+         --min-p N                min-p sampling (default: 0.1, 0.0 = disabled)
+         --tfs N                  tail free sampling, parameter z (default: 1.0, 1.0 = disabled)
+         --typical N              locally typical sampling, parameter p (default: 1.0, 1.0 = disabled)
+         --repeat-last-n N        last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size)
+         --repeat-penalty N       penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled)
+         --presence-penalty N     repeat alpha presence penalty (default: 0.0, 0.0 = disabled)
+         --frequency-penalty N    repeat alpha frequency penalty (default: 0.0, 0.0 = disabled)
+         --dynatemp-range N       dynamic temperature range (default: 0.0, 0.0 = disabled)
+         --dynatemp-exp N         dynamic temperature exponent (default: 1.0)
+         --mirostat N             use Mirostat sampling.
+                                  Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.
+                                  (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)
+         --mirostat-lr N          Mirostat learning rate, parameter eta (default: 0.1)
+         --mirostat-ent N         Mirostat target entropy, parameter tau (default: 5.0)
+         -l TOKEN_ID(+/-)BIAS     modifies the likelihood of token appearing in the completion,
+                                  i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
+                                  or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'
+         --cfg-negative-prompt PROMPT
+                                  negative prompt to use for guidance (default: '')
+         --cfg-negative-prompt-file FNAME
+                                  negative prompt file to use for guidance
+         --cfg-scale N            strength of guidance (default: 1.0, 1.0 = disable)
+         --chat-template JINJA_TEMPLATE
+                                  set custom jinja chat template (default: template taken from model's metadata)
+                                  if suffix/prefix are specified, template will be disabled
+                                  only commonly used templates are accepted:
+                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
+
+grammar:
+
+         --grammar GRAMMAR        BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '')
+         --grammar-file FNAME     file to read grammar from
+  -j,    --json-schema SCHEMA     JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
+                                  For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead
+
+embedding:
+
+         --pooling {none,mean,cls,last}
+                                  pooling type for embeddings, use model default if unspecified
+         --attention {causal,non-causal}
+                                  attention type for embeddings, use model default if unspecified
+
+context hacking:
+
+         --rope-scaling {none,linear,yarn}
+                                  RoPE frequency scaling method, defaults to linear unless specified by the model
+         --rope-scale N           RoPE context scaling factor, expands context by a factor of N
+         --rope-freq-base N       RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
+         --rope-freq-scale N      RoPE frequency scaling factor, expands context by a factor of 1/N
+         --yarn-orig-ctx N        YaRN: original context size of model (default: 0 = model training context size)
+         --yarn-ext-factor N      YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
+         --yarn-attn-factor N     YaRN: scale sqrt(t) or attention magnitude (default: 1.0)
+         --yarn-beta-slow N       YaRN: high correction dim or alpha (default: 1.0)
+         --yarn-beta-fast N       YaRN: low correction dim or beta (default: 32.0)
+  -gan,  --grp-attn-n N           group-attention factor (default: 1)
+  -gaw,  --grp-attn-w N           group-attention width (default: 512.0)
+  -dkvc, --dump-kv-cache          verbose print of the KV cache
+  -nkvo, --no-kv-offload          disable KV offload
+  -ctk,  --cache-type-k TYPE      KV cache data type for K (default: f16)
+  -ctv,  --cache-type-v TYPE      KV cache data type for V (default: f16)
+
+perplexity:
+
+         --all-logits             return logits for all tokens in the batch (default: false)
+         --hellaswag              compute HellaSwag score over random tasks from datafile supplied with -f
+         --hellaswag-tasks N      number of tasks to use when computing the HellaSwag score (default: 400)
+         --winogrande             compute Winogrande score over random tasks from datafile supplied with -f
+         --winogrande-tasks N     number of tasks to use when computing the Winogrande score (default: 0)
+         --multiple-choice        compute multiple choice score over random tasks from datafile supplied with -f
+         --multiple-choice-tasks N
+                                  number of tasks to use when computing the multiple choice score (default: 0)
+         --kl-divergence          computes KL-divergence to logits provided via --kl-divergence-base
+         --ppl-stride N           stride for perplexity calculation (default: 0)
+         --ppl-output-type {0,1}  output type for perplexity calculation (default: 0)
+
+parallel:
+
+  -dt,   --defrag-thold N         KV cache defragmentation threshold (default: -1.0, < 0 - disabled)
+  -np,   --parallel N             number of parallel sequences to decode (default: 1)
+  -ns,   --sequences N            number of sequences to decode (default: 1)
+  -cb,   --cont-batching          enable continuous batching (a.k.a dynamic batching) (default: enabled)
+
+multi-modality:
+
+         --mmproj FILE            path to a multimodal projector file for LLaVA. see examples/llava/README.md
+         --image FILE             path to an image file. use with multimodal models. Specify multiple times for batching
+
+backend:
+
+         --rpc SERVERS            comma separated list of RPC servers
+         --mlock                  force system to keep model in RAM rather than swapping or compressing
+         --no-mmap                do not memory-map model (slower load but may reduce pageouts if not using mlock)
+         --numa TYPE              attempt optimizations that help on some NUMA systems
+                                    - distribute: spread execution evenly over all nodes
+                                    - isolate: only spawn threads on CPUs on the node that execution started on
+                                    - numactl: use the CPU map provided by numactl
+                                  if run without this previously, it is recommended to drop the system page cache before using this
+                                  see https://github.com/ggerganov/llama.cpp/issues/1437
+
+model:
+
+         --check-tensors          check model tensor data for invalid values (default: false)
+         --override-kv KEY=TYPE:VALUE
+                                  advanced option to override model metadata by key. may be specified multiple times.
+                                  types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false
+         --lora FNAME             apply LoRA adapter (implies --no-mmap)
+         --lora-scaled FNAME S    apply LoRA adapter with user defined scaling S (implies --no-mmap)
+         --lora-base FNAME        optional model to use as a base for the layers modified by the LoRA adapter
+         --control-vector FNAME   add a control vector
+                                  note: this argument can be repeated to add multiple control vectors
+         --control-vector-scaled FNAME SCALE
+                                  add a control vector with user defined scaling SCALE
+                                  note: this argument can be repeated to add multiple scaled control vectors
+         --control-vector-layer-range START END
+                                  layer range to apply the control vector(s) to, start and end inclusive
+  -m,    --model FNAME            model path (default: models/$filename with filename from --hf-file
+                                  or --model-url if set, otherwise models/7B/ggml-model-f16.gguf)
+  -md,   --model-draft FNAME      draft model for speculative decoding (default: unused)
+  -mu,   --model-url MODEL_URL    model download url (default: unused)
+  -hfr,  --hf-repo REPO           Hugging Face model repository (default: unused)
+  -hff,  --hf-file FILE           Hugging Face model file (default: unused)
+  -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)
+
+server:
+
+         --host HOST              ip address to listen (default: 127.0.0.1)
+         --port PORT              port to listen (default: 8080)
+         --path PATH              path to serve static files from (default: )
+         --embedding(s)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
+         --api-key KEY            API key to use for authentication (default: none)
+         --api-key-file FNAME     path to file containing API keys (default: none)
+         --ssl-key-file FNAME     path to file a PEM-encoded SSL private key
+         --ssl-cert-file FNAME    path to file a PEM-encoded SSL certificate
+         --timeout N              server read/write timeout in seconds (default: 600)
+         --threads-http N         number of threads used to process HTTP requests (default: -1)
+         --system-prompt-file FNAME
+                                  set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications
+         --log-format {text,json}
+                                  log output format: json or text (default: json)
+         --metrics                enable prometheus compatible metrics endpoint (default: disabled)
+         --no-slots               disables slots monitoring endpoint (default: enabled)
+         --slot-save-path PATH    path to save slot kv cache (default: disabled)
+         --chat-template JINJA_TEMPLATE
+                                  set custom jinja chat template (default: template taken from model's metadata)
+                                  only commonly used templates are accepted:
+                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
+  -sps,  --slot-prompt-similarity SIMILARITY
+                                  how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
+         --lora-init-without-apply
+                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
+
+logging:
+
+         --simple-io              use basic IO for better compatibility in subprocesses and limited consoles
+  -ld,   --logdir LOGDIR          path under which to save YAML logs (no logging if unset)
+         --log-test               Run simple logging test
+         --log-disable            Disable trace logs
+         --log-enable             Enable trace logs
+         --log-file FNAME         Specify a log filename (without extension)
+         --log-new                Create a separate new log file on start. Each log file will have unique name: "..log"
+         --log-append             Don't truncate the old log file.
+```
+
 
 ## Build
 
-`server` is built alongside everything else from the root of the project
+`llama-server` is built alongside everything else from the root of the project
 
 - Using `make`:
 
   ```bash
-  make server
+  make llama-server
   ```
 
 - Using `CMake`:
 
   ```bash
   cmake -B build
-  cmake --build build --config Release -t server
+  cmake --build build --config Release -t llama-server
   ```
 
-  Binary is at `./build/bin/server`
+  Binary is at `./build/bin/llama-server`
 
 ## Build with SSL
 
-`server` can also be built with SSL support using OpenSSL 3
+`llama-server` can also be built with SSL support using OpenSSL 3
 
 - Using `make`:
 
@@ -107,14 +277,14 @@ The project is under active development, and we are [looking for feedback and co
   # NOTE: For non-system openssl, use the following:
   #   CXXFLAGS="-I /path/to/openssl/include"
   #   LDFLAGS="-L /path/to/openssl/lib"
-  make LLAMA_SERVER_SSL=true server
+  make LLAMA_SERVER_SSL=true llama-server
   ```
 
 - Using `CMake`:
 
   ```bash
   cmake -B build -DLLAMA_SERVER_SSL=ON
-  cmake --build build --config Release -t server
+  cmake --build build --config Release -t llama-server
   ```
 
 ## Quick Start
@@ -124,13 +294,13 @@ To get started right away, run the following command, making sure to use the cor
 ### Unix-based systems (Linux, macOS, etc.)
 
 ```bash
-./server -m models/7B/ggml-model.gguf -c 2048
+./llama-server -m models/7B/ggml-model.gguf -c 2048
 ```
 
 ### Windows
 
 ```powershell
-server.exe -m models\7B\ggml-model.gguf -c 2048
+llama-server.exe -m models\7B\ggml-model.gguf -c 2048
 ```
 
 The above command will start a server that by default listens on `127.0.0.1:8080`.
@@ -198,7 +368,8 @@ node index.js
 
 ## API Endpoints
 
-- **GET** `/health`: Returns the current state of the server:
+### GET `/health`: Returns the current state of the server
+
   - 503 -> `{"status": "loading model"}` if the model is still being loaded.
   - 500 -> `{"status": "error"}` if the model failed to load.
   - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
@@ -207,7 +378,7 @@ node index.js
 
   If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
 
-- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
+### POST `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
@@ -231,7 +402,7 @@ node index.js
 
     `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.
 
-    `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded.
+    `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token.
     By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt.
 
     `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
@@ -279,13 +450,13 @@ node index.js
 
     `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
-    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch.  Default: `false`
+    `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
-### Result JSON
+**Response format**
 
 - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
 
@@ -324,7 +495,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
-- **POST** `/tokenize`: Tokenize a given text.
+### POST `/tokenize`: Tokenize a given text
 
     *Options:*
 
@@ -332,13 +503,15 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
 
-- **POST** `/detokenize`: Convert tokens to text.
+### POST `/detokenize`: Convert tokens to text
 
     *Options:*
 
     `tokens`: Set the tokens to detokenize.
 
-- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+### POST `/embedding`: Generate embedding of a given text
+
+The same as [the embedding example](../embedding) does.
 
     *Options:*
 
@@ -346,7 +519,9 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
-- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+### POST `/infill`: For code infilling.
+
+Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 
@@ -358,14 +533,15 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - **GET** `/props`: Return current server settings.
 
-### Result JSON
+**Response format**
 
 ```json
 {
   "assistant_name": "",
   "user_name": "",
   "default_generation_settings": { ... },
-  "total_slots": 1
+  "total_slots": 1,
+  "chat_template": ""
 }
 ```
 
@@ -373,8 +549,11 @@ Notice that each `probs` is an array of length `n_probs`.
 - `user_name` - the required anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
 - `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
+- `chat_template` - the model's original Jinja2 prompt template
+
+### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only model with [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, ChatML template will be used.
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
     *Options:*
 
@@ -426,7 +605,7 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
+### POST `/v1/embeddings`: OpenAI-compatible embeddings API
 
     *Options:*
 
@@ -460,9 +639,9 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
-- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+### GET `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
 
-### Result JSON
+**Response format**
 
 ```json
 [
@@ -523,7 +702,7 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
+### GET `/metrics`: Prometheus compatible metrics exporter endpoint if `--metrics` is enabled:
 
 Available metrics:
 - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@@ -535,13 +714,13 @@ Available metrics:
 - `llamacpp:requests_processing`: Number of requests processing.
 - `llamacpp:requests_deferred`: Number of requests deferred.
 
-- **POST** `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
+### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
 
     *Options:*
 
     `filename`: Name of the file to save the slot's prompt cache. The file will be saved in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -555,13 +734,13 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
+### POST `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
 
     *Options:*
 
     `filename`: Name of the file to restore the slot's prompt cache from. The file should be located in the directory specified by the `--slot-save-path` server parameter.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -575,9 +754,9 @@ Available metrics:
 }
 ```
 
-- **POST** `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
+### POST `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
 
-### Result JSON
+**Response format**
 
 ```json
 {
@@ -586,6 +765,42 @@ Available metrics:
 }
 ```
 
+### GET `/lora-adapters`: Get list of all LoRA adapters
+
+If an adapter is disabled, the scale will be set to 0.
+
+**Response format**
+
+```json
+[
+    {
+        "id": 0,
+        "path": "my_adapter_1.gguf",
+        "scale": 0.0
+    },
+    {
+        "id": 1,
+        "path": "my_adapter_2.gguf",
+        "scale": 0.0
+    }
+]
+```
+
+### POST `/lora-adapters`: Set list of LoRA adapters
+
+To disable an adapter, either remove it from the list below, or set scale to 0.
+
+**Request format**
+
+To know the `id` of the adapter, use GET `/lora-adapters`
+
+```json
+[
+  {"id": 0, "scale": 0.2},
+  {"id": 1, "scale": 0.8}
+]
+```
+
 ## More examples
 
 ### Change system prompt on runtime
@@ -629,11 +844,11 @@ bash chat.sh
 
 ### OAI-like API
 
-The HTTP `server` supports an OAI-like API: https://github.com/openai/openai-openapi
+The HTTP `llama-server` supports an OAI-like API: https://github.com/openai/openai-openapi
 
 ### API errors
 
-`server` returns errors in the same format as OAI: https://github.com/openai/openai-openapi
+`llama-server` returns errors in the same format as OAI: https://github.com/openai/openai-openapi
 
 Example of an error:
 
diff --git a/llm/llama.cpp/examples/server/bench/README.md b/llm/llama.cpp/examples/server/bench/README.md
index 23a3ec97523ef79ef527d68f06cd241ec72c7605..0f18ca39651d2cd8dd333a7bf340d88821a27395 100644
--- a/llm/llama.cpp/examples/server/bench/README.md
+++ b/llm/llama.cpp/examples/server/bench/README.md
@@ -99,7 +99,7 @@ The `bench.py` script does several steps:
 It aims to be used in the CI, but you can run it manually:
 
 ```shell
-LLAMA_SERVER_BIN_PATH=../../../cmake-build-release/bin/server python bench.py \
+LLAMA_SERVER_BIN_PATH=../../../cmake-build-release/bin/llama-server python bench.py \
               --runner-label local \
               --name local \
               --branch `git rev-parse --abbrev-ref HEAD` \
diff --git a/llm/llama.cpp/examples/server/bench/bench.py b/llm/llama.cpp/examples/server/bench/bench.py
index 86c5de101445c162c6150b25c73fbde1255afe0a..2daac08847d6556e57e5eddb9e220fbd267e74a6 100644
--- a/llm/llama.cpp/examples/server/bench/bench.py
+++ b/llm/llama.cpp/examples/server/bench/bench.py
@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 import argparse
 import json
 import os
@@ -59,10 +61,11 @@ def main(args_in: list[str] | None = None) -> None:
         sys.exit(1)
 
     # start the benchmark
+    iterations = 0
+    data = {}
     try:
         start_benchmark(args)
 
-        iterations = 0
         with open("results.github.env", 'w') as github_env:
             # parse output
             with open('k6-results.json', 'r') as bench_results:
@@ -129,7 +132,7 @@ def main(args_in: list[str] | None = None) -> None:
                 timestamps, metric_values = zip(*values)
                 metric_values = [float(value) for value in metric_values]
                 prometheus_metrics[metric] = metric_values
-                timestamps_dt = [datetime.fromtimestamp(int(ts)) for ts in timestamps]
+                timestamps_dt = [str(datetime.fromtimestamp(int(ts))) for ts in timestamps]
                 plt.figure(figsize=(16, 10), dpi=80)
                 plt.plot(timestamps_dt, metric_values, label=metric)
                 plt.xticks(rotation=0, fontsize=14, horizontalalignment='center', alpha=.7)
@@ -156,7 +159,7 @@ def main(args_in: list[str] | None = None) -> None:
                 plt.close()
 
                 # Mermaid format in case images upload failed
-                with (open(f"{metric}.mermaid", 'w') as mermaid_f):
+                with open(f"{metric}.mermaid", 'w') as mermaid_f:
                     mermaid = (
                     f"""---
 config:
@@ -245,7 +248,7 @@ def start_server(args):
 
 def start_server_background(args):
     # Start the server
-    server_path = '../../../build/bin/server'
+    server_path = '../../../build/bin/llama-server'
     if 'LLAMA_SERVER_BIN_PATH' in os.environ:
         server_path = os.environ['LLAMA_SERVER_BIN_PATH']
     server_args = [
@@ -278,7 +281,7 @@ def start_server_background(args):
     }
     server_process = subprocess.Popen(
         args,
-        **pkwargs)
+        **pkwargs)  # pyright: ignore[reportArgumentType, reportCallIssue]
 
     def server_log(in_stream, out_stream):
         for line in iter(in_stream.readline, b''):
diff --git a/llm/llama.cpp/examples/server/public/colorthemes.css b/llm/llama.cpp/examples/server/public/colorthemes.css
new file mode 100644
index 0000000000000000000000000000000000000000..b1e2b8b70c261a50f26c20e605a3a05a754295f1
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/colorthemes.css
@@ -0,0 +1,402 @@
+@import url("theme-snowstorm.css");
+@import url("theme-polarnight.css");
+@import url("theme-ketivah.css");
+@import url("theme-mangotango.css");
+@import url("theme-playground.css");
+@import url("theme-beeninorder.css");
+
+:root {
+/* ---------- PRIMARY COLORS ----------------- */
+--primary-color-1: hsl(217.5, 26.7%, 94.1%);
+    --primary-color-1-hue:             217.5;
+    --primary-color-1-saturation:      26.7%;
+    --primary-color-1-lightness:       94.1%;
+
+--primary-color-2: hsl(218.2, 26.8%, 92.0%);
+    --primary-color-2-hue:             218.2;
+    --primary-color-2-saturation:      26.8%;
+    --primary-color-2-lightness:       92.0%;
+
+--primary-color-3: hsl(218.8, 27.9%, 88.0%);
+    --primary-color-3-hue:             218.8;
+    --primary-color-3-saturation:      27.9%;
+    --primary-color-3-lightness:       88.0%;
+
+--primary-color-4: hsl(218.8, 18.3%, 81.8%);
+    --primary-color-4-hue:             218.8;
+    --primary-color-4-saturation:      18.3%;
+    --primary-color-4-lightness:       81.8%;
+
+
+/* ---------- SECONDARY COLORS --------------- */
+--secondary-color-1: hsl(220.0, 16.4%, 21.6%);
+    --secondary-color-1-hue:             220.0;
+    --secondary-color-1-saturation:      16.4%;
+    --secondary-color-1-lightness:       21.6%;
+
+--secondary-color-2: hsl(221.7, 16.3%, 27.6%);
+    --secondary-color-2-hue:             221.7;
+    --secondary-color-2-saturation:      16.3%;
+    --secondary-color-2-lightness:       27.6%;
+
+--secondary-color-3: hsl(220.0, 16.8%, 31.6%);
+    --secondary-color-3-hue:             220.0;
+    --secondary-color-3-saturation:      16.8%;
+    --secondary-color-3-lightness:       31.6%;
+
+--secondary-color-4: hsl(220.0, 16.5%, 35.7%);
+    --secondary-color-4-hue:             220.0;
+    --secondary-color-4-saturation:      16.5%;
+    --secondary-color-4-lightness:       35.7%;
+
+
+
+/* ----------- NUANCES COLORS ---------------- */
+--theme-nuance-color-1: hsl(178.7, 25.1%, 64.9%);
+    --theme-nuance-color-1-hue:             178.7;
+    --theme-nuance-color-1-saturation:      25.1%;
+    --theme-nuance-color-1-lightness:       64.9%;
+
+--theme-nuance-color-2: hsl(193.3, 43.4%, 67.5%);
+    --theme-nuance-color-2-hue:             193.3;
+    --theme-nuance-color-2-saturation:      43.4%;
+    --theme-nuance-color-2-lightness:       67.5%;
+
+--theme-nuance-color-3: hsl(210.0, 34.0%, 63.1%);
+    --theme-nuance-color-3-hue:             210.0;
+    --theme-nuance-color-3-saturation:      34.0%;
+    --theme-nuance-color-3-lightness:       63.1%;
+
+--theme-nuance-color-4: hsl(213.1, 32.0%, 52.2%);
+    --theme-nuance-color-4-hue:             213.1;
+    --theme-nuance-color-4-saturation:      32.0%;
+    --theme-nuance-color-4-lightness:       52.2%;
+
+
+
+/* ----------- ROYGP COLORS ------------------ */
+--theme-red-color:    hsl(32.5, 80%, 50%);
+--theme-orange-color: hsl(32.5, 70%, 45%);
+--theme-yellow-color: hsl(40.0,   0.6%, 73.3%);
+--theme-green-color:  hsl(92.4,  27.8%, 64.7%);
+--theme-purple-color: hsl(311.1, 20.2%, 63.1%);
+
+
+
+/* ------------------------------------------- */
+--background-color-1:    var(--primary-color-1);
+--background-color-2:    var(--primary-color-2);
+--background-color-3:    var(--primary-color-3);
+--background-color-4:    var(--primary-color-4);
+
+--border-color-1:        var(--primary-color-2);
+--border-color-2:        var(--primary-color-3);
+--border-color-3:        var(--primary-color-4);
+
+--border-focus-color:    var(--theme-nuance-color-2);
+--border-focus-shadow:   var(--theme-nuance-color-1);
+
+--text-color-plain:      var(--secondary-color-1);
+--text-color-subtile-1:  var(--secondary-color-2);
+--text-color-subtile-2:  var(--secondary-color-3);
+
+--code-background-color: var(--secondary-color-2);
+--code-text-color:       var(--primary-color-2);
+
+--ui-range-thumb-color:  var(--theme-nuance-color-3);
+--ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+--textarea-border-color: var(--secondary-color-4);
+
+--chat-id-color:         var(--theme-nuance-color-4);
+
+
+
+/* ------------------------------------------- */
+--button-alert-text-hover:       var(--primary-color-1);
+--button-alert-color-hover:      var(--theme-orange-color);
+--button-alert-border-hover:     var(--theme-orange-color);
+
+--button-alert-text-active:      var(--primary-color-1);
+--button-alert-color-active:     var(--theme-red-color);
+--button-alert-border-active:    var(--theme-red-color);
+
+
+
+/* ----------- PRIMARY BUTTONS --------------- */
+/* - button should immediately catch the eye - */
+--button-primary-text:   var(--secondary-color-1);
+--button-primary-color:  var(--theme-nuance-color-3);
+--button-primary-border: var(--theme-nuance-color-3);
+
+
+/* ---------hover---------- */
+--button-primary-text-hover:
+    hsl(217.5,
+    calc(var(--secondary-color-1-saturation) + 35%),
+    calc(var(--secondary-color-1-lightness)  - 30%));
+
+--button-primary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+--button-primary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+
+/* ---------active--------- */
+--button-primary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 35%));
+
+--button-primary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+--button-primary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+
+
+/* ---------- SECONDARY BUTTONS -------------- */
+/* these should NOT immediately catch the eye  */
+--button-secondary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 50%));
+
+--button-secondary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+--button-secondary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+
+/* ---------hover---------- */
+--button-secondary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+--button-secondary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+
+/* ---------active--------- */
+--button-secondary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) + 40%),
+    calc(var(--theme-nuance-color-3-lightness)  - 55%));
+
+--button-secondary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-secondary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+
+
+/* ---------- TERTIARY BUTTONS --------------- */
+/* ---------- disabled buttons --------------- */
+--button-tertiary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+/* ---------hover---------- */
+--button-tertiary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+}
+
+/*
+
+.theme-template {
+
+
+    If light theme: should go from bright to darker
+    If dark theme: should go from dark to brighter
+    ideally this should not be anything but steps of
+    gray or slightly variants from it
+
+    --primary-color-1: #2E3440;
+    --primary-color-2: #3B4252;
+    --primary-color-3: #434C5E;
+    --primary-color-4: #4C566A;
+
+
+
+    If light theme: should go from dark to brighter
+    If dark theme: should go from bright to darker
+    ideally this should not be anything but steps of
+    gray or slightly variants from it
+
+    --secondary-color-1: #ECEFF4;
+    --secondary-color-2: #E5E9F0;
+    --secondary-color-3: #D8DEE9;
+    --secondary-color-4: #C8CED9;
+
+
+
+    Choose wisely nuance colors. It is not easy to find
+    4 harmonizing nuance colors. But keep in mind, that
+    only one accent color could work too.
+
+    --theme-nuance-color-1: #8FBCBB;
+    --theme-nuance-color-2: #88C0D0;
+    --theme-nuance-color-3: #81A1C1;
+    --theme-nuance-color-4: #5E81AC;
+
+
+
+    adapt the color red, orange, yellow, green,
+    purple to the 'mood' of your overall design
+    e.g is it low-contrast? vibrant? dynamic? etc
+
+    --theme-red-color:    #BF616A;
+    --theme-orange-color: #D08770;
+    --theme-yellow-color: #EBCB8B;
+    --theme-green-color:  #A3BE8C;
+    --theme-purple-color: #B48EAD;
+
+
+
+NOTE: comment all those line `--- ...` out
+------------------------------------------------
+--background-color-1:
+--background-color-2:
+--background-color-3:
+--background-color-4:
+
+--border-color-1:
+--border-color-2:
+--border-color-3:
+
+--border-focus-color:
+--border-focus-shadow:
+
+--text-color-plain:
+--text-color-subtile-1:
+--text-color-subtile-2:
+
+--code-background-color:
+--code-text-color:
+
+--ui-range-thumb-color:
+--ui-range-thumb-border:
+
+--textarea-border-color:
+
+
+
+-------------------------------------------
+--button-alert-text-hover:
+--button-alert-color-hover:
+--button-alert-border-hover:
+
+--button-alert-text-active:
+--button-alert-color-active:
+--button-alert-border-active:
+
+
+
+----------- PRIMARY -----------------------
+--button should immediately catch the eye--
+
+--button-primary-text:
+--button-primary-color:
+--button-primary-border:
+
+
+---------hover----------
+--button-primary-text-hover:
+--button-primary-color-hover:
+--button-primary-border-hover:
+
+
+---------active---------
+--button-primary-text-active:
+--button-primary-color-active:
+--button-primary-border-active:
+
+
+
+------------ SECONDARY ------------------------
+--button should NOT immediately catch the eye--
+
+--button-secondary-text:
+--button-secondary-color:
+--button-secondary-border:
+
+
+---------hover----------
+--button-secondary-text-hover:
+--button-secondary-color-hover:
+--button-secondary-border-hover:
+
+
+---------active---------
+--button-secondary-text-active:
+--button-secondary-color-active:
+--button-secondary-border-active:
+
+
+
+---------- TERTIARY -----------------------
+---------- disabled buttons ---------------
+--button-tertiary-text:
+--button-tertiary-color:
+--button-tertiary-border:
+
+
+---------hover----------
+--button-tertiary-text:
+--button-tertiary-color:
+--button-tertiary-border:
+
+}
+
+*/
diff --git a/llm/llama.cpp/examples/server/public/completion.js b/llm/llama.cpp/examples/server/public/completion.js
index 987b9a3b891d0c9aac21b06be74c0f315cfc229d..36818f7644aa10cd5e8cdcf1e7d20c5e20c3e35e 100644
--- a/llm/llama.cpp/examples/server/public/completion.js
+++ b/llm/llama.cpp/examples/server/public/completion.js
@@ -21,7 +21,7 @@ let generation_settings = null;
 //
 export async function* llama(prompt, params = {}, config = {}) {
   let controller = config.controller;
-  const api_url = config.api_url || "";
+  const api_url = config.api_url?.replace(/\/+$/, '') || "";
 
   if (!controller) {
     controller = new AbortController();
@@ -196,7 +196,7 @@ export const llamaComplete = async (params, controller, callback) => {
 // Get the model info from the server. This is useful for getting the context window and so on.
 export const llamaModelInfo = async (config = {}) => {
   if (!generation_settings) {
-    const api_url = config.api_url || "";
+    const api_url = config.api_url?.replace(/\/+$/, '') || "";
     const props = await fetch(`${api_url}/props`).then(r => r.json());
     generation_settings = props.default_generation_settings;
   }
diff --git a/llm/llama.cpp/examples/server/public/index-new.html b/llm/llama.cpp/examples/server/public/index-new.html
new file mode 100644
index 0000000000000000000000000000000000000000..c87dd8f1e1d32b8fb3b648625fdb89631b54bcfb
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/index-new.html
@@ -0,0 +1,1178 @@
+
+
+
+
+
+  
+  
+  
+  llama.cpp - chat
+
+  
+  
+
+  
+
+
+
+
+  

+    
+  

+  

+
+
+
diff --git a/llm/llama.cpp/examples/server/public/index.html b/llm/llama.cpp/examples/server/public/index.html
index 4c5a34d903309528a491e801002d057503e230c5..07fec6a38bbcdeb93a806a344e95e15471cffd8c 100644
--- a/llm/llama.cpp/examples/server/public/index.html
+++ b/llm/llama.cpp/examples/server/public/index.html
@@ -1,5 +1,4 @@
 
-
 
   
   
@@ -12,6 +11,18 @@
       font-size: 90%;
     }
 
+    .grid-container {
+      display: grid;
+      grid-template-columns: auto auto auto;
+      padding: 10px;
+    }
+
+    .grid-item {
+      padding: 5px;
+      /* font-size: 30px; */
+      text-align: center;
+    }
+
     #container {
       margin: 0em auto;
       display: flex;
@@ -35,6 +46,67 @@
       padding: 0.5em;
     }
 
+    h1 {
+      text-align: center;
+    }
+
+    .customlink:link {
+      color: white;
+      background-color: #007aff;
+      font-weight: 600;
+      text-decoration: none;
+      float: right;
+      margin-top: 30px;
+      display: flex;
+      flex-direction: row;
+      gap: 0.5em;
+      justify-content: flex-end;
+      border-radius: 4px;
+      padding: 8px;
+    }
+
+    .customlink:visited {
+      color: white;
+      background-color: #007aff;
+      font-weight: 600;
+      text-decoration: none;
+      float: right;
+      margin-top: 30px;
+      display: flex;
+      flex-direction: row;
+      gap: 0.5em;
+      justify-content: flex-end;
+      padding: 8px;
+    }
+
+    .customlink:hover {
+      color: white;
+      background-color: #0070ee;
+      font-weight: 600;
+      text-decoration: none;
+      float: right;
+      margin-top: 30px;
+      display: flex;
+      flex-direction: row;
+      gap: 0.5em;
+      justify-content: flex-end;
+      padding: 8px;
+    }
+
+    .customlink:active {
+      color: #0070ee;
+      background-color: #80b3ef;
+      font-weight: 600;
+      text-decoration: none;
+      float: right;
+      margin-top: 30px;
+      display: flex;
+      flex-direction: row;
+      gap: 0.5em;
+      justify-content: flex-end;
+      padding: 8px;
+    }
+
     body {
       max-width: 600px;
       min-width: 300px;
@@ -59,12 +131,20 @@
       align-items: stretch;
     }
 
-    .right {
+    .message-controls {
       display: flex;
-      flex-direction: row;
-      gap: 0.5em;
       justify-content: flex-end;
     }
+    .message-controls > div:nth-child(2) {
+      display: flex;
+      flex-direction: column;
+      gap: 0.5em;
+    }
+    .message-controls > div:nth-child(2) > div {
+      display: flex;
+      margin-left: auto;
+      gap: 0.5em;
+    }
 
     fieldset {
       border: none;
@@ -203,6 +283,7 @@
 
     import { llama } from './completion.js';
     import { SchemaConverter } from './json-schema-to-grammar.mjs';
+
     let selected_image = false;
     var slot_id = -1;
 
@@ -374,6 +455,9 @@
 
     /* END: Support for storing prompt templates and parameters in browsers LocalStorage */
 
+    const tts = window.speechSynthesis;
+    const ttsVoice = signal(null)
+
     const llamaStats = signal(null)
     const controller = signal(null)
 
@@ -406,7 +490,7 @@
         throw new Error("already running");
       }
       controller.value = new AbortController();
-      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value, api_url: location.pathname.replace(/\/+$/, '') })) {
+      for await (const chunk of llama(prompt, llamaParams, { controller: controller.value, api_url: new URL('.', document.baseURI).href })) {
         const data = chunk.data;
 
         if (data.stop) {
@@ -523,8 +607,51 @@
       });
     }
 
+    const SpeechRecognition = window.SpeechRecognition || window.webkitSpeechRecognition;
+    const talkRecognition = SpeechRecognition ? new SpeechRecognition() : null;
     function MessageInput() {
-      const message = useSignal("")
+      const message = useSignal("");
+
+      const talkActive = useSignal(false);
+      const sendOnTalk = useSignal(false);
+      const talkStop = (e) => {
+        if (e) e.preventDefault();
+
+        talkActive.value = false;
+        talkRecognition?.stop();
+      }
+      const talk = (e) => {
+        e.preventDefault();
+
+        if (talkRecognition)
+          talkRecognition.start();
+        else
+          alert("Speech recognition is not supported by this browser.");
+      }
+      if(talkRecognition) {
+        talkRecognition.onstart = () => {
+          talkActive.value = true;
+        }
+        talkRecognition.onresult = (e) => {
+          if (event.results.length > 0) {
+            message.value = event.results[0][0].transcript;
+            if (sendOnTalk.value) {
+              submit(e);
+            }
+          }
+        }
+        talkRecognition.onspeechend = () => {
+          talkStop();
+        }
+      }
+
+      const ttsVoices = useSignal(tts?.getVoices() || []);
+      const ttsVoiceDefault = computed(() => ttsVoices.value.find(v => v.default));
+      if (tts) {
+        tts.onvoiceschanged = () => {
+          ttsVoices.value = tts.getVoices();
+        }
+      }
 
       const submit = (e) => {
         stop(e);
@@ -551,11 +678,45 @@
                value="${message}"
             />
           
-          

-            
-            
-            
-            
+          

+            
 

+            

+              

+                
+                
+                
+                
+              

+              
+              

+                 {
+                  e.preventDefault();
+                  alert(`TTS supported by your browser: ${tts ? 'Yes' : 'No'}\n(TTS and speech recognition are not provided by llama.cpp)`);
+                }}>[?]
+                
+                
+              

+            

           

         
       `
@@ -586,26 +747,86 @@
         }
       }, [messages])
 
+      const ttsChatLineActiveIx = useSignal(undefined);
+      const ttsChatLine = (e, ix, msg) => {
+        if (e) e.preventDefault();
+
+        if (!tts || !ttsVoice.value || !('SpeechSynthesisUtterance' in window)) return;
+
+        const ttsVoices = tts.getVoices();
+        const voice = ttsVoices.find(v => v.name === ttsVoice.value);
+        if (!voice) return;
+
+        if (ttsChatLineActiveIx.value !== undefined) {
+          tts.cancel();
+          if (ttsChatLineActiveIx.value === ix) {
+            ttsChatLineActiveIx.value = undefined;
+            return;
+          }
+        }
+
+        ttsChatLineActiveIx.value = ix;
+        let ttsUtter = new SpeechSynthesisUtterance(msg);
+        ttsUtter.voice = voice;
+        ttsUtter.onend = e => {
+          ttsChatLineActiveIx.value = undefined;
+        };
+        tts.speak(ttsUtter);
+      }
+
       const isCompletionMode = session.value.type === 'completion'
+
+      // Try play the last bot message
+      const lastCharChatLinesIxs = useSignal([]);
+      const lastCharChatLinesIxsOld = useSignal([]);
+      useEffect(() => {
+        if (
+          !isCompletionMode
+          && lastCharChatLinesIxs.value.length !== lastCharChatLinesIxsOld.value.length
+          && !generating.value
+        ) {
+          const ix = lastCharChatLinesIxs.value[lastCharChatLinesIxs.value.length - 1];
+          if (ix !== undefined) {
+            const msg = messages[ix];
+            ttsChatLine(null, ix, Array.isArray(msg) ? msg[1].map(m => m.content).join('') : msg);
+          }
+
+          lastCharChatLinesIxsOld.value = structuredClone(lastCharChatLinesIxs.value);
+        }
+      }, [generating.value]);
+
       const chatLine = ([user, data], index) => {
         let message
-        const isArrayMessage = Array.isArray(data)
+        const isArrayMessage = Array.isArray(data);
+        const text = isArrayMessage ?
+            data.map(msg => msg.content).join('') :
+            data;
         if (params.value.n_probs > 0 && isArrayMessage) {
           message = html`<${Probabilities} data=${data} />`
         } else {
-          const text = isArrayMessage ?
-            data.map(msg => msg.content).join('') :
-            data;
           message = isCompletionMode ?
             text :
             html`<${Markdownish} text=${template(text)} />`
         }
+
+        const fromBot = user && user === '{{char}}';
+        if (fromBot && !lastCharChatLinesIxs.value.includes(index))
+          lastCharChatLinesIxs.value.push(index);
+
         if (user) {
-          return html`
${template(user)}: ${message}
`
+          return html`
+          

+            
${template(user)}: ${message}

+            ${
+              fromBot && ttsVoice.value
+              && html`
`
+            }
+          

+          `;
         } else {
           return isCompletionMode ?
             html`${message}` :
-            html`
${message}
`
+            html`
${message}
`
         }
       };
 
@@ -1035,7 +1256,11 @@
       return html`
         

           

-            
llama.cpp

+            

+              

+              
llama.cpp

+              
New UI

+            

           

 
           

diff --git a/llm/llama.cpp/examples/server/public/index.js b/llm/llama.cpp/examples/server/public/index.js
index 695aec2568a52520a4ada68fe9532ae83d5dcb9b..6709609399e4a75414fa38ca05fa638d25530e24 100644
--- a/llm/llama.cpp/examples/server/public/index.js
+++ b/llm/llama.cpp/examples/server/public/index.js
@@ -1 +1 @@
-const t=Symbol.for("preact-signals");function n(){if(r>1){r--;return}let t,n=!1;while(void 0!==i){let _=i;i=void 0;u++;while(void 0!==_){const i=_.o;_.o=void 0;_.f&=-3;if(!(8&_.f)&&h(_))try{_.c()}catch(e){if(!n){t=e;n=!0}}_=i}}u=0;r--;if(n)throw t}function e(t){if(r>0)return t();r++;try{return t()}finally{n()}}let _,i;function o(t){const n=_;_=void 0;try{return t()}finally{_=n}}let r=0,u=0,l=0;function s(t){if(void 0===_)return;let n=t.n;if(void 0===n||n.t!==_){n={i:0,S:t,p:_.s,n:void 0,t:_,e:void 0,x:void 0,r:n};if(void 0!==_.s)_.s.n=n;_.s=n;t.n=n;if(32&_.f)t.S(n);return n}else if(-1===n.i){n.i=0;if(void 0!==n.n){n.n.p=n.p;if(void 0!==n.p)n.p.n=n.n;n.p=_.s;n.n=void 0;_.s.n=n;_.s=n}return n}}function f(t){this.v=t;this.i=0;this.n=void 0;this.t=void 0}f.prototype.brand=t;f.prototype.h=function(){return!0};f.prototype.S=function(t){if(this.t!==t&&void 0===t.e){t.x=this.t;if(void 0!==this.t)this.t.e=t;this.t=t}};f.prototype.U=function(t){if(void 0!==this.t){const n=t.e,e=t.x;if(void 0!==n){n.x=e;t.e=void 0}if(void 0!==e){e.e=n;t.x=void 0}if(t===this.t)this.t=e}};f.prototype.subscribe=function(t){return k(()=>{const n=this.value,e=_;_=void 0;try{t(n)}finally{_=e}})};f.prototype.valueOf=function(){return this.value};f.prototype.toString=function(){return this.value+""};f.prototype.toJSON=function(){return this.value};f.prototype.peek=function(){const t=_;_=void 0;try{return this.value}finally{_=t}};Object.defineProperty(f.prototype,"value",{get(){const t=s(this);if(void 0!==t)t.i=this.i;return this.v},set(t){if(t!==this.v){if(u>100)throw new Error("Cycle detected");this.v=t;this.i++;l++;r++;try{for(let t=this.t;void 0!==t;t=t.x)t.t.N()}finally{n()}}}});function c(t){return new f(t)}function h(t){for(let n=t.s;void 0!==n;n=n.n)if(n.S.i!==n.i||!n.S.h()||n.S.i!==n.i)return!0;return!1}function a(t){for(let n=t.s;void 0!==n;n=n.n){const e=n.S.n;if(void 0!==e)n.r=e;n.S.n=n;n.i=-1;if(void 0===n.n){t.s=n;break}}}function p(t){let n,e=t.s;while(void 0!==e){const t=e.p;if(-1===e.i){e.S.U(e);if(void 0!==t)t.n=e.n;if(void 0!==e.n)e.n.p=t}else n=e;e.S.n=e.r;if(void 0!==e.r)e.r=void 0;e=t}t.s=n}function d(t){f.call(this,void 0);this.x=t;this.s=void 0;this.g=l-1;this.f=4}(d.prototype=new f).h=function(){this.f&=-3;if(1&this.f)return!1;if(32==(36&this.f))return!0;this.f&=-5;if(this.g===l)return!0;this.g=l;this.f|=1;if(this.i>0&&!h(this)){this.f&=-2;return!0}const t=_;try{a(this);_=this;const t=this.x();if(16&this.f||this.v!==t||0===this.i){this.v=t;this.f&=-17;this.i++}}catch(t){this.v=t;this.f|=16;this.i++}_=t;p(this);this.f&=-2;return!0};d.prototype.S=function(t){if(void 0===this.t){this.f|=36;for(let t=this.s;void 0!==t;t=t.n)t.S.S(t)}f.prototype.S.call(this,t)};d.prototype.U=function(t){if(void 0!==this.t){f.prototype.U.call(this,t);if(void 0===this.t){this.f&=-33;for(let t=this.s;void 0!==t;t=t.n)t.S.U(t)}}};d.prototype.N=function(){if(!(2&this.f)){this.f|=6;for(let t=this.t;void 0!==t;t=t.x)t.t.N()}};Object.defineProperty(d.prototype,"value",{get(){if(1&this.f)throw new Error("Cycle detected");const t=s(this);this.h();if(void 0!==t)t.i=this.i;if(16&this.f)throw this.v;return this.v}});function v(t){return new d(t)}function y(t){const e=t.u;t.u=void 0;if("function"==typeof e){r++;const i=_;_=void 0;try{e()}catch(n){t.f&=-2;t.f|=8;m(t);throw n}finally{_=i;n()}}}function m(t){for(let n=t.s;void 0!==n;n=n.n)n.S.U(n);t.x=void 0;t.s=void 0;y(t)}function g(t){if(_!==this)throw new Error("Out-of-order effect");p(this);_=t;this.f&=-2;if(8&this.f)m(this);n()}function b(t){this.x=t;this.u=void 0;this.s=void 0;this.o=void 0;this.f=32}b.prototype.c=function(){const t=this.S();try{if(8&this.f)return;if(void 0===this.x)return;const n=this.x();if("function"==typeof n)this.u=n}finally{t()}};b.prototype.S=function(){if(1&this.f)throw new Error("Cycle detected");this.f|=1;this.f&=-9;y(this);a(this);r++;const t=_;_=this;return g.bind(this,t)};b.prototype.N=function(){if(!(2&this.f)){this.f|=2;this.o=i;i=this}};b.prototype.d=function(){this.f|=8;if(!(1&this.f))m(this)};function k(t){const n=new b(t);try{n.c()}catch(t){n.d();throw t}return n.d.bind(n)}var S,w,x,C,E,U,H,P,N,$,D,T,F={},V=[],A=/acit|ex(?:s|g|n|p|$)|rph|grid|ows|mnc|ntw|ine[ch]|zoo|^ord|itera/i,M=Array.isArray;function W(t,n){for(var e in n)t[e]=n[e];return t}function O(t){var n=t.parentNode;n&&n.removeChild(t)}function L(t,n,e){var _,i,o,r={};for(o in n)"key"==o?_=n[o]:"ref"==o?i=n[o]:r[o]=n[o];if(arguments.length>2&&(r.children=arguments.length>3?S.call(arguments,2):e),"function"==typeof t&&null!=t.defaultProps)for(o in t.defaultProps)void 0===r[o]&&(r[o]=t.defaultProps[o]);return R(t,r,_,i,null)}function R(t,n,e,_,i){var o={type:t,props:n,key:e,ref:_,__k:null,__:null,__b:0,__e:null,__d:void 0,__c:null,constructor:void 0,__v:null==i?++x:i,__i:-1,__u:0};return null==i&&null!=w.vnode&&w.vnode(o),o}function I(){return{current:null}}function j(t){return t.children}function q(t,n){this.props=t,this.context=n}function B(t,n){if(null==n)return t.__?B(t.__,t.__i+1):null;for(var e;nn&&E.sort(P));J.__r=0}function K(t,n,e,_,i,o,r,u,l,s,f){var c,h,a,p,d,v=_&&_.__k||V,y=n.length;for(e.__d=l,Q(e,n,v),l=e.__d,c=0;c0?R(i.type,i.props,i.key,i.ref?i.ref:null,i.__v):i)?(i.__=t,i.__b=t.__b+1,u=Z(i,e,r,f),i.__i=u,o=null,-1!==u&&(f--,(o=e[u])&&(o.__u|=131072)),null==o||null===o.__v?(-1==u&&c--,"function"!=typeof i.type&&(i.__u|=65536)):u!==r&&(u===r+1?c++:u>r?f>l-r?c+=u-r:c--:u(null!=l&&0==(131072&l.__u)?1:0))for(;r>=0||u=0){if((l=n[r])&&0==(131072&l.__u)&&i==l.key&&o===l.type)return r;r--}if(u2&&(u.children=arguments.length>3?S.call(arguments,2):e),R(t.type,u,_||t.key,i||t.ref,null)}function ht(t,n){var e={__c:n="__cC"+T++,__:t,Consumer:function(t,n){return t.children(n)},Provider:function(t){var e,_;return this.getChildContext||(e=[],(_={})[n]=this,this.getChildContext=function(){return _},this.shouldComponentUpdate=function(t){this.props.value!==t.value&&e.some((function(t){t.__e=!0,z(t)}))},this.sub=function(t){e.push(t);var n=t.componentWillUnmount;t.componentWillUnmount=function(){e.splice(e.indexOf(t),1),n&&n.call(t)}}),t.children}};return e.Provider.__=e.Consumer.contextType=e}S=V.slice,w={__e:function(t,n,e,_){for(var i,o,r;n=n.__;)if((i=n.__c)&&!i.__)try{if((o=i.constructor)&&null!=o.getDerivedStateFromError&&(i.setState(o.getDerivedStateFromError(t)),r=i.__d),null!=i.componentDidCatch&&(i.componentDidCatch(t,_||{}),r=i.__d),r)return i.__E=i}catch(n){t=n}throw t}},x=0,C=function(t){return null!=t&&null==t.constructor},q.prototype.setState=function(t,n){var e;e=null!=this.__s&&this.__s!==this.state?this.__s:this.__s=W({},this.state),"function"==typeof t&&(t=t(W({},e),this.props)),t&&W(e,t),null!=t&&this.__v&&(n&&this._sb.push(n),z(this))},q.prototype.forceUpdate=function(t){this.__v&&(this.__e=!0,t&&this.__h.push(t),z(this))},q.prototype.render=j,E=[],H="function"==typeof Promise?Promise.prototype.then.bind(Promise.resolve()):setTimeout,P=function(t,n){return t.__v.__b-n.__v.__b},J.__r=0,N=0,$=et(!1),D=et(!0),T=0;var at,pt,dt,vt,yt=0,mt=[],gt=[],bt=w,kt=bt.__b,St=bt.__r,wt=bt.diffed,xt=bt.__c,Ct=bt.unmount,Et=bt.__;function Ut(t,n){bt.__h&&bt.__h(pt,t,yt||n),yt=0;var e=pt.__H||(pt.__H={__:[],__h:[]});return t>=e.__.length&&e.__.push({__V:gt}),e.__[t]}function Ht(t){return yt=1,Pt(Gt,t)}function Pt(t,n,e){var _=Ut(at++,2);if(_.t=t,!_.__c&&(_.__=[e?e(n):Gt(void 0,n),function(t){var n=_.__N?_.__N[0]:_.__[0],e=_.t(n,t);n!==e&&(_.__N=[e,_.__[1]],_.__c.setState({}))}],_.__c=pt,!pt.u)){var i=function(t,n,e){if(!_.__c.__H)return!0;var i=_.__c.__H.__.filter((function(t){return!!t.__c}));if(i.every((function(t){return!t.__N})))return!o||o.call(this,t,n,e);var r=!1;return i.forEach((function(t){if(t.__N){var n=t.__[0];t.__=t.__N,t.__N=void 0,n!==t.__[0]&&(r=!0)}})),!(!r&&_.__c.props===t)&&(!o||o.call(this,t,n,e))};pt.u=!0;var o=pt.shouldComponentUpdate,r=pt.componentWillUpdate;pt.componentWillUpdate=function(t,n,e){if(this.__e){var _=o;o=void 0,i(t,n,e),o=_}r&&r.call(this,t,n,e)},pt.shouldComponentUpdate=i}return _.__N||_.__}function Nt(t,n){var e=Ut(at++,3);!bt.__s&&Bt(e.__H,n)&&(e.__=t,e.i=n,pt.__H.__h.push(e))}function $t(t,n){var e=Ut(at++,4);!bt.__s&&Bt(e.__H,n)&&(e.__=t,e.i=n,pt.__h.push(e))}function Dt(t){return yt=5,Ft((function(){return{current:t}}),[])}function Tt(t,n,e){yt=6,$t((function(){return"function"==typeof t?(t(n()),function(){return t(null)}):t?(t.current=n(),function(){return t.current=null}):void 0}),null==e?e:e.concat(t))}function Ft(t,n){var e=Ut(at++,7);return Bt(e.__H,n)?(e.__V=t(),e.i=n,e.__h=t,e.__V):e.__}function Vt(t,n){return yt=8,Ft((function(){return t}),n)}function At(t){var n=pt.context[t.__c],e=Ut(at++,9);return e.c=t,n?(null==e.__&&(e.__=!0,n.sub(pt)),n.props.value):t.__}function Mt(t,n){bt.useDebugValue&&bt.useDebugValue(n?n(t):t)}function Wt(t){var n=Ut(at++,10),e=Ht();return n.__=t,pt.componentDidCatch||(pt.componentDidCatch=function(t,_){n.__&&n.__(t,_),e[1](t)}),[e[0],function(){e[1](void 0)}]}function Ot(){var t=Ut(at++,11);if(!t.__){for(var n=pt.__v;null!==n&&!n.__m&&null!==n.__;)n=n.__;var e=n.__m||(n.__m=[0,0]);t.__="P"+e[0]+"-"+e[1]++}return t.__}function Lt(){for(var t;t=mt.shift();)if(t.__P&&t.__H)try{t.__H.__h.forEach(jt),t.__H.__h.forEach(qt),t.__H.__h=[]}catch(n){t.__H.__h=[],bt.__e(n,t.__v)}}bt.__b=function(t){pt=null,kt&&kt(t)},bt.__=function(t,n){t&&n.__k&&n.__k.__m&&(t.__m=n.__k.__m),Et&&Et(t,n)},bt.__r=function(t){St&&St(t),at=0;var n=(pt=t.__c).__H;n&&(dt===pt?(n.__h=[],pt.__h=[],n.__.forEach((function(t){t.__N&&(t.__=t.__N),t.__V=gt,t.__N=t.i=void 0}))):(n.__h.forEach(jt),n.__h.forEach(qt),n.__h=[],at=0)),dt=pt},bt.diffed=function(t){wt&&wt(t);var n=t.__c;n&&n.__H&&(n.__H.__h.length&&(1!==mt.push(n)&&vt===bt.requestAnimationFrame||((vt=bt.requestAnimationFrame)||It)(Lt)),n.__H.__.forEach((function(t){t.i&&(t.__H=t.i),t.__V!==gt&&(t.__=t.__V),t.i=void 0,t.__V=gt}))),dt=pt=null},bt.__c=function(t,n){n.some((function(t){try{t.__h.forEach(jt),t.__h=t.__h.filter((function(t){return!t.__||qt(t)}))}catch(r){n.some((function(t){t.__h&&(t.__h=[])})),n=[],bt.__e(r,t.__v)}})),xt&&xt(t,n)},bt.unmount=function(t){Ct&&Ct(t);var n,e=t.__c;e&&e.__H&&(e.__H.__.forEach((function(t){try{jt(t)}catch(t){n=t}})),e.__H=void 0,n&&bt.__e(n,e.__v))};var Rt="function"==typeof requestAnimationFrame;function It(t){var n,e=function(){clearTimeout(_),Rt&&cancelAnimationFrame(n),setTimeout(t)},_=setTimeout(e,100);Rt&&(n=requestAnimationFrame(e))}function jt(t){var n=pt,e=t.__c;"function"==typeof e&&(t.__c=void 0,e()),pt=n}function qt(t){var n=pt;t.__c=t.__(),pt=n}function Bt(t,n){return!t||t.length!==n.length||n.some((function(n,e){return n!==t[e]}))}function Gt(t,n){return"function"==typeof n?n(t):n}function zt(t,n){w[t]=n.bind(null,w[t]||(()=>{}))}let Jt,Kt;function Qt(t){if(Kt)Kt();Kt=t&&t.S()}function Xt({data:t}){const n=Zt(t);n.value=t;const e=Ft(()=>{let t=this.__v;while(t=t.__)if(t.__c){t.__c.__$f|=4;break}this.__$u.c=()=>{var t;if(!C(e.peek())&&3===(null==(t=this.base)?void 0:t.nodeType))this.base.data=e.peek();else{this.__$f|=1;this.setState({})}};return v(()=>{let t=n.value.value;return 0===t?0:!0===t?"":t||""})},[]);return e.value}Xt.displayName="_st";Object.defineProperties(f.prototype,{constructor:{configurable:!0,value:void 0},type:{configurable:!0,value:Xt},props:{configurable:!0,get(){return{data:this}}},__b:{configurable:!0,value:1}});zt("__b",(t,n)=>{if("string"==typeof n.type){let t,e=n.props;for(let _ in e){if("children"===_)continue;let i=e[_];if(i instanceof f){if(!t)n.__np=t={};t[_]=i;e[_]=i.peek()}}}t(n)});zt("__r",(t,n)=>{Qt();let e,_=n.__c;if(_){_.__$f&=-2;e=_.__$u;if(void 0===e)_.__$u=e=function(t){let n;k((function(){n=this}));n.c=()=>{_.__$f|=1;_.setState({})};return n}()}Jt=_;Qt(e);t(n)});zt("__e",(t,n,e,_)=>{Qt();Jt=void 0;t(n,e,_)});zt("diffed",(t,n)=>{Qt();Jt=void 0;let e;if("string"==typeof n.type&&(e=n.__e)){let t=n.__np,_=n.props;if(t){let n=e.U;if(n)for(let e in n){let _=n[e];if(void 0!==_&&!(e in t)){_.d();n[e]=void 0}}else{n={};e.U=n}for(let i in t){let o=n[i],r=t[i];if(void 0===o){o=Yt(e,i,r,_);n[i]=o}else o.o(r,_)}}}t(n)});function Yt(t,n,e,_){const i=n in t&&void 0===t.ownerSVGElement,o=c(e);return{o:(t,n)=>{o.value=t;_=n},d:k(()=>{const e=o.value.value;if(_[n]!==e){_[n]=e;if(i)t[n]=e;else if(e)t.setAttribute(n,e);else t.removeAttribute(n)}})}}zt("unmount",(t,n)=>{if("string"==typeof n.type){let t=n.__e;if(t){const n=t.U;if(n){t.U=void 0;for(let t in n){let e=n[t];if(e)e.d()}}}}else{let t=n.__c;if(t){const n=t.__$u;if(n){t.__$u=void 0;n.d()}}}t(n)});zt("__h",(t,n,e,_)=>{if(_<3||9===_)n.__$f|=2;t(n,e,_)});q.prototype.shouldComponentUpdate=function(t,n){const e=this.__$u;if(!(e&&void 0!==e.s||4&this.__$f))return!0;if(3&this.__$f)return!0;for(let _ in n)return!0;for(let _ in t)if("__source"!==_&&t[_]!==this.props[_])return!0;for(let _ in this.props)if(!(_ in t))return!0;return!1};function Zt(t){return Ft(()=>c(t),[])}function tn(t){const n=Dt(t);n.current=t;Jt.__$f|=4;return Ft(()=>v(()=>n.current()),[])}function nn(t){const n=Dt(t);n.current=t;Nt(()=>k(()=>n.current()),[])}var en=function(t,n,e,_){var i;n[0]=0;for(var o=1;o=5&&((i||!t&&5===_)&&(r.push(_,0,i,e),_=6),t&&(r.push(_,t,0,e),_=6)),i=""},l=0;l"===n?(_=1,i=""):i=n+i[0]:o?n===o?o="":i+=n:'"'===n||"'"===n?o=n:">"===n?(u(),_=1):_&&("="===n?(_=5,e=i,i=""):"/"===n&&(_<5||">"===t[l][s+1])?(u(),3===_&&(r=r[0]),_=r,(r=r[0]).push(2,0,_),_=0):" "===n||"\t"===n||"\n"===n||"\r"===n?(u(),_=2):i+=n),3===_&&"!--"===i&&(_=4,r=r[0])}return u(),r}(t)),n),arguments,[])).length>1?n:n[0]}var rn=on.bind(L);export{q as Component,j as Fragment,f as Signal,e as batch,ct as cloneElement,v as computed,ht as createContext,L as createElement,I as createRef,k as effect,L as h,rn as html,ft as hydrate,C as isValidElement,w as options,st as render,c as signal,Y as toChildArray,o as untracked,Vt as useCallback,tn as useComputed,At as useContext,Mt as useDebugValue,Nt as useEffect,Wt as useErrorBoundary,Ot as useId,Tt as useImperativeHandle,$t as useLayoutEffect,Ft as useMemo,Pt as useReducer,Dt as useRef,Zt as useSignal,nn as useSignalEffect,Ht as useState};
+const t=Symbol.for("preact-signals");function n(){if(r>1){r--;return}let t,n=!1;while(void 0!==i){let _=i;i=void 0;u++;while(void 0!==_){const i=_.o;_.o=void 0;_.f&=-3;if(!(8&_.f)&&h(_))try{_.c()}catch(e){if(!n){t=e;n=!0}}_=i}}u=0;r--;if(n)throw t}function e(t){if(r>0)return t();r++;try{return t()}finally{n()}}let _,i;function o(t){const n=_;_=void 0;try{return t()}finally{_=n}}let r=0,u=0,l=0;function s(t){if(void 0===_)return;let n=t.n;if(void 0===n||n.t!==_){n={i:0,S:t,p:_.s,n:void 0,t:_,e:void 0,x:void 0,r:n};if(void 0!==_.s)_.s.n=n;_.s=n;t.n=n;if(32&_.f)t.S(n);return n}else if(-1===n.i){n.i=0;if(void 0!==n.n){n.n.p=n.p;if(void 0!==n.p)n.p.n=n.n;n.p=_.s;n.n=void 0;_.s.n=n;_.s=n}return n}}function f(t){this.v=t;this.i=0;this.n=void 0;this.t=void 0}f.prototype.brand=t;f.prototype.h=function(){return!0};f.prototype.S=function(t){if(this.t!==t&&void 0===t.e){t.x=this.t;if(void 0!==this.t)this.t.e=t;this.t=t}};f.prototype.U=function(t){if(void 0!==this.t){const n=t.e,e=t.x;if(void 0!==n){n.x=e;t.e=void 0}if(void 0!==e){e.e=n;t.x=void 0}if(t===this.t)this.t=e}};f.prototype.subscribe=function(t){return k(()=>{const n=this.value,e=_;_=void 0;try{t(n)}finally{_=e}})};f.prototype.valueOf=function(){return this.value};f.prototype.toString=function(){return this.value+""};f.prototype.toJSON=function(){return this.value};f.prototype.peek=function(){const t=_;_=void 0;try{return this.value}finally{_=t}};Object.defineProperty(f.prototype,"value",{get(){const t=s(this);if(void 0!==t)t.i=this.i;return this.v},set(t){if(t!==this.v){if(u>100)throw new Error("Cycle detected");this.v=t;this.i++;l++;r++;try{for(let t=this.t;void 0!==t;t=t.x)t.t.N()}finally{n()}}}});function c(t){return new f(t)}function h(t){for(let n=t.s;void 0!==n;n=n.n)if(n.S.i!==n.i||!n.S.h()||n.S.i!==n.i)return!0;return!1}function a(t){for(let n=t.s;void 0!==n;n=n.n){const e=n.S.n;if(void 0!==e)n.r=e;n.S.n=n;n.i=-1;if(void 0===n.n){t.s=n;break}}}function p(t){let n,e=t.s;while(void 0!==e){const t=e.p;if(-1===e.i){e.S.U(e);if(void 0!==t)t.n=e.n;if(void 0!==e.n)e.n.p=t}else n=e;e.S.n=e.r;if(void 0!==e.r)e.r=void 0;e=t}t.s=n}function d(t){f.call(this,void 0);this.x=t;this.s=void 0;this.g=l-1;this.f=4}(d.prototype=new f).h=function(){this.f&=-3;if(1&this.f)return!1;if(32==(36&this.f))return!0;this.f&=-5;if(this.g===l)return!0;this.g=l;this.f|=1;if(this.i>0&&!h(this)){this.f&=-2;return!0}const t=_;try{a(this);_=this;const t=this.x();if(16&this.f||this.v!==t||0===this.i){this.v=t;this.f&=-17;this.i++}}catch(t){this.v=t;this.f|=16;this.i++}_=t;p(this);this.f&=-2;return!0};d.prototype.S=function(t){if(void 0===this.t){this.f|=36;for(let t=this.s;void 0!==t;t=t.n)t.S.S(t)}f.prototype.S.call(this,t)};d.prototype.U=function(t){if(void 0!==this.t){f.prototype.U.call(this,t);if(void 0===this.t){this.f&=-33;for(let t=this.s;void 0!==t;t=t.n)t.S.U(t)}}};d.prototype.N=function(){if(!(2&this.f)){this.f|=6;for(let t=this.t;void 0!==t;t=t.x)t.t.N()}};Object.defineProperty(d.prototype,"value",{get(){if(1&this.f)throw new Error("Cycle detected");const t=s(this);this.h();if(void 0!==t)t.i=this.i;if(16&this.f)throw this.v;return this.v}});function v(t){return new d(t)}function y(t){const e=t.u;t.u=void 0;if("function"==typeof e){r++;const i=_;_=void 0;try{e()}catch(n){t.f&=-2;t.f|=8;m(t);throw n}finally{_=i;n()}}}function m(t){for(let n=t.s;void 0!==n;n=n.n)n.S.U(n);t.x=void 0;t.s=void 0;y(t)}function g(t){if(_!==this)throw new Error("Out-of-order effect");p(this);_=t;this.f&=-2;if(8&this.f)m(this);n()}function b(t){this.x=t;this.u=void 0;this.s=void 0;this.o=void 0;this.f=32}b.prototype.c=function(){const t=this.S();try{if(8&this.f)return;if(void 0===this.x)return;const n=this.x();if("function"==typeof n)this.u=n}finally{t()}};b.prototype.S=function(){if(1&this.f)throw new Error("Cycle detected");this.f|=1;this.f&=-9;y(this);a(this);r++;const t=_;_=this;return g.bind(this,t)};b.prototype.N=function(){if(!(2&this.f)){this.f|=2;this.o=i;i=this}};b.prototype.d=function(){this.f|=8;if(!(1&this.f))m(this)};function k(t){const n=new b(t);try{n.c()}catch(t){n.d();throw t}return n.d.bind(n)}var w,S,x,C,U,E,H,P,N,$,D,T,M={},F=[],A=/acit|ex(?:s|g|n|p|$)|rph|grid|ows|mnc|ntw|ine[ch]|zoo|^ord|itera/i,V=Array.isArray;function W(t,n){for(var e in n)t[e]=n[e];return t}function L(t){var n=t.parentNode;n&&n.removeChild(t)}function O(t,n,e){var _,i,o,r={};for(o in n)"key"==o?_=n[o]:"ref"==o?i=n[o]:r[o]=n[o];if(arguments.length>2&&(r.children=arguments.length>3?w.call(arguments,2):e),"function"==typeof t&&null!=t.defaultProps)for(o in t.defaultProps)void 0===r[o]&&(r[o]=t.defaultProps[o]);return R(t,r,_,i,null)}function R(t,n,e,_,i){var o={type:t,props:n,key:e,ref:_,__k:null,__:null,__b:0,__e:null,__d:void 0,__c:null,constructor:void 0,__v:null==i?++x:i,__i:-1,__u:0};return null==i&&null!=S.vnode&&S.vnode(o),o}function I(){return{current:null}}function j(t){return t.children}function q(t,n){this.props=t,this.context=n}function B(t,n){if(null==n)return t.__?B(t.__,t.__i+1):null;for(var e;nn&&U.sort(P));J.__r=0}function K(t,n,e,_,i,o,r,u,l,s,f){var c,h,a,p,d,v=_&&_.__k||F,y=n.length;for(e.__d=l,Q(e,n,v),l=e.__d,c=0;c0?R(i.type,i.props,i.key,i.ref?i.ref:null,i.__v):i)?(i.__=t,i.__b=t.__b+1,u=Z(i,e,r,f),i.__i=u,o=null,-1!==u&&(f--,(o=e[u])&&(o.__u|=131072)),null==o||null===o.__v?(-1==u&&c--,"function"!=typeof i.type&&(i.__u|=65536)):u!==r&&(u===r+1?c++:u>r?f>l-r?c+=u-r:c--:u(null!=l&&0==(131072&l.__u)?1:0))for(;r>=0||u=0){if((l=n[r])&&0==(131072&l.__u)&&i==l.key&&o===l.type)return r;r--}if(u2&&(u.children=arguments.length>3?w.call(arguments,2):e),R(t.type,u,_||t.key,i||t.ref,null)}function ht(t,n){var e={__c:n="__cC"+T++,__:t,Consumer:function(t,n){return t.children(n)},Provider:function(t){var e,_;return this.getChildContext||(e=[],(_={})[n]=this,this.getChildContext=function(){return _},this.shouldComponentUpdate=function(t){this.props.value!==t.value&&e.some((function(t){t.__e=!0,G(t)}))},this.sub=function(t){e.push(t);var n=t.componentWillUnmount;t.componentWillUnmount=function(){e.splice(e.indexOf(t),1),n&&n.call(t)}}),t.children}};return e.Provider.__=e.Consumer.contextType=e}w=F.slice,S={__e:function(t,n,e,_){for(var i,o,r;n=n.__;)if((i=n.__c)&&!i.__)try{if((o=i.constructor)&&null!=o.getDerivedStateFromError&&(i.setState(o.getDerivedStateFromError(t)),r=i.__d),null!=i.componentDidCatch&&(i.componentDidCatch(t,_||{}),r=i.__d),r)return i.__E=i}catch(n){t=n}throw t}},x=0,C=function(t){return null!=t&&null==t.constructor},q.prototype.setState=function(t,n){var e;e=null!=this.__s&&this.__s!==this.state?this.__s:this.__s=W({},this.state),"function"==typeof t&&(t=t(W({},e),this.props)),t&&W(e,t),null!=t&&this.__v&&(n&&this._sb.push(n),G(this))},q.prototype.forceUpdate=function(t){this.__v&&(this.__e=!0,t&&this.__h.push(t),G(this))},q.prototype.render=j,U=[],H="function"==typeof Promise?Promise.prototype.then.bind(Promise.resolve()):setTimeout,P=function(t,n){return t.__v.__b-n.__v.__b},J.__r=0,N=0,$=et(!1),D=et(!0),T=0;var at,pt,dt,vt,yt=0,mt=[],gt=[],bt=S,kt=bt.__b,wt=bt.__r,St=bt.diffed,xt=bt.__c,Ct=bt.unmount,Ut=bt.__;function Et(t,n){bt.__h&&bt.__h(pt,t,yt||n),yt=0;var e=pt.__H||(pt.__H={__:[],__h:[]});return t>=e.__.length&&e.__.push({__V:gt}),e.__[t]}function Ht(t){return yt=1,Pt(zt,t)}function Pt(t,n,e){var _=Et(at++,2);if(_.t=t,!_.__c&&(_.__=[e?e(n):zt(void 0,n),function(t){var n=_.__N?_.__N[0]:_.__[0],e=_.t(n,t);n!==e&&(_.__N=[e,_.__[1]],_.__c.setState({}))}],_.__c=pt,!pt.u)){var i=function(t,n,e){if(!_.__c.__H)return!0;var i=_.__c.__H.__.filter((function(t){return!!t.__c}));if(i.every((function(t){return!t.__N})))return!o||o.call(this,t,n,e);var r=!1;return i.forEach((function(t){if(t.__N){var n=t.__[0];t.__=t.__N,t.__N=void 0,n!==t.__[0]&&(r=!0)}})),!(!r&&_.__c.props===t)&&(!o||o.call(this,t,n,e))};pt.u=!0;var o=pt.shouldComponentUpdate,r=pt.componentWillUpdate;pt.componentWillUpdate=function(t,n,e){if(this.__e){var _=o;o=void 0,i(t,n,e),o=_}r&&r.call(this,t,n,e)},pt.shouldComponentUpdate=i}return _.__N||_.__}function Nt(t,n){var e=Et(at++,3);!bt.__s&&Bt(e.__H,n)&&(e.__=t,e.i=n,pt.__H.__h.push(e))}function $t(t,n){var e=Et(at++,4);!bt.__s&&Bt(e.__H,n)&&(e.__=t,e.i=n,pt.__h.push(e))}function Dt(t){return yt=5,Mt((function(){return{current:t}}),[])}function Tt(t,n,e){yt=6,$t((function(){return"function"==typeof t?(t(n()),function(){return t(null)}):t?(t.current=n(),function(){return t.current=null}):void 0}),null==e?e:e.concat(t))}function Mt(t,n){var e=Et(at++,7);return Bt(e.__H,n)?(e.__V=t(),e.i=n,e.__h=t,e.__V):e.__}function Ft(t,n){return yt=8,Mt((function(){return t}),n)}function At(t){var n=pt.context[t.__c],e=Et(at++,9);return e.c=t,n?(null==e.__&&(e.__=!0,n.sub(pt)),n.props.value):t.__}function Vt(t,n){bt.useDebugValue&&bt.useDebugValue(n?n(t):t)}function Wt(t){var n=Et(at++,10),e=Ht();return n.__=t,pt.componentDidCatch||(pt.componentDidCatch=function(t,_){n.__&&n.__(t,_),e[1](t)}),[e[0],function(){e[1](void 0)}]}function Lt(){var t=Et(at++,11);if(!t.__){for(var n=pt.__v;null!==n&&!n.__m&&null!==n.__;)n=n.__;var e=n.__m||(n.__m=[0,0]);t.__="P"+e[0]+"-"+e[1]++}return t.__}function Ot(){for(var t;t=mt.shift();)if(t.__P&&t.__H)try{t.__H.__h.forEach(jt),t.__H.__h.forEach(qt),t.__H.__h=[]}catch(n){t.__H.__h=[],bt.__e(n,t.__v)}}bt.__b=function(t){pt=null,kt&&kt(t)},bt.__=function(t,n){t&&n.__k&&n.__k.__m&&(t.__m=n.__k.__m),Ut&&Ut(t,n)},bt.__r=function(t){wt&&wt(t),at=0;var n=(pt=t.__c).__H;n&&(dt===pt?(n.__h=[],pt.__h=[],n.__.forEach((function(t){t.__N&&(t.__=t.__N),t.__V=gt,t.__N=t.i=void 0}))):(n.__h.forEach(jt),n.__h.forEach(qt),n.__h=[],at=0)),dt=pt},bt.diffed=function(t){St&&St(t);var n=t.__c;n&&n.__H&&(n.__H.__h.length&&(1!==mt.push(n)&&vt===bt.requestAnimationFrame||((vt=bt.requestAnimationFrame)||It)(Ot)),n.__H.__.forEach((function(t){t.i&&(t.__H=t.i),t.__V!==gt&&(t.__=t.__V),t.i=void 0,t.__V=gt}))),dt=pt=null},bt.__c=function(t,n){n.some((function(t){try{t.__h.forEach(jt),t.__h=t.__h.filter((function(t){return!t.__||qt(t)}))}catch(r){n.some((function(t){t.__h&&(t.__h=[])})),n=[],bt.__e(r,t.__v)}})),xt&&xt(t,n)},bt.unmount=function(t){Ct&&Ct(t);var n,e=t.__c;e&&e.__H&&(e.__H.__.forEach((function(t){try{jt(t)}catch(t){n=t}})),e.__H=void 0,n&&bt.__e(n,e.__v))};var Rt="function"==typeof requestAnimationFrame;function It(t){var n,e=function(){clearTimeout(_),Rt&&cancelAnimationFrame(n),setTimeout(t)},_=setTimeout(e,100);Rt&&(n=requestAnimationFrame(e))}function jt(t){var n=pt,e=t.__c;"function"==typeof e&&(t.__c=void 0,e()),pt=n}function qt(t){var n=pt;t.__c=t.__(),pt=n}function Bt(t,n){return!t||t.length!==n.length||n.some((function(n,e){return n!==t[e]}))}function zt(t,n){return"function"==typeof n?n(t):n}function Gt(t,n){S[t]=n.bind(null,S[t]||(()=>{}))}let Jt,Kt;function Qt(t){if(Kt)Kt();Kt=t&&t.S()}function Xt({data:t}){const n=Zt(t);n.value=t;const e=Mt(()=>{let t=this.__v;while(t=t.__)if(t.__c){t.__c.__$f|=4;break}this.__$u.c=()=>{var t;if(!C(e.peek())&&3===(null==(t=this.base)?void 0:t.nodeType))this.base.data=e.peek();else{this.__$f|=1;this.setState({})}};return v(()=>{let t=n.value.value;return 0===t?0:!0===t?"":t||""})},[]);return e.value}Xt.displayName="_st";Object.defineProperties(f.prototype,{constructor:{configurable:!0,value:void 0},type:{configurable:!0,value:Xt},props:{configurable:!0,get(){return{data:this}}},__b:{configurable:!0,value:1}});Gt("__b",(t,n)=>{if("string"==typeof n.type){let t,e=n.props;for(let _ in e){if("children"===_)continue;let i=e[_];if(i instanceof f){if(!t)n.__np=t={};t[_]=i;e[_]=i.peek()}}}t(n)});Gt("__r",(t,n)=>{Qt();let e,_=n.__c;if(_){_.__$f&=-2;e=_.__$u;if(void 0===e)_.__$u=e=function(t){let n;k((function(){n=this}));n.c=()=>{_.__$f|=1;_.setState({})};return n}()}Jt=_;Qt(e);t(n)});Gt("__e",(t,n,e,_)=>{Qt();Jt=void 0;t(n,e,_)});Gt("diffed",(t,n)=>{Qt();Jt=void 0;let e;if("string"==typeof n.type&&(e=n.__e)){let t=n.__np,_=n.props;if(t){let n=e.U;if(n)for(let e in n){let _=n[e];if(void 0!==_&&!(e in t)){_.d();n[e]=void 0}}else{n={};e.U=n}for(let i in t){let o=n[i],r=t[i];if(void 0===o){o=Yt(e,i,r,_);n[i]=o}else o.o(r,_)}}}t(n)});function Yt(t,n,e,_){const i=n in t&&void 0===t.ownerSVGElement,o=c(e);return{o:(t,n)=>{o.value=t;_=n},d:k(()=>{const e=o.value.value;if(_[n]!==e){_[n]=e;if(i)t[n]=e;else if(e)t.setAttribute(n,e);else t.removeAttribute(n)}})}}Gt("unmount",(t,n)=>{if("string"==typeof n.type){let t=n.__e;if(t){const n=t.U;if(n){t.U=void 0;for(let t in n){let e=n[t];if(e)e.d()}}}}else{let t=n.__c;if(t){const n=t.__$u;if(n){t.__$u=void 0;n.d()}}}t(n)});Gt("__h",(t,n,e,_)=>{if(_<3||9===_)n.__$f|=2;t(n,e,_)});q.prototype.shouldComponentUpdate=function(t,n){const e=this.__$u;if(!(e&&void 0!==e.s||4&this.__$f))return!0;if(3&this.__$f)return!0;for(let _ in n)return!0;for(let _ in t)if("__source"!==_&&t[_]!==this.props[_])return!0;for(let _ in this.props)if(!(_ in t))return!0;return!1};function Zt(t){return Mt(()=>c(t),[])}function tn(t){const n=Dt(t);n.current=t;Jt.__$f|=4;return Mt(()=>v(()=>n.current()),[])}function nn(t){const n=Dt(t);n.current=t;Nt(()=>k(()=>n.current()),[])}var en=function(t,n,e,_){var i;n[0]=0;for(var o=1;o=5&&((i||!t&&5===_)&&(r.push(_,0,i,e),_=6),t&&(r.push(_,t,0,e),_=6)),i=""},l=0;l"===n?(_=1,i=""):i=n+i[0]:o?n===o?o="":i+=n:'"'===n||"'"===n?o=n:">"===n?(u(),_=1):_&&("="===n?(_=5,e=i,i=""):"/"===n&&(_<5||">"===t[l][s+1])?(u(),3===_&&(r=r[0]),_=r,(r=r[0]).push(2,0,_),_=0):" "===n||"\t"===n||"\n"===n||"\r"===n?(u(),_=2):i+=n),3===_&&"!--"===i&&(_=4,r=r[0])}return u(),r}(t)),n),arguments,[])).length>1?n:n[0]}var rn=on.bind(O);export{q as Component,j as Fragment,f as Signal,e as batch,ct as cloneElement,v as computed,ht as createContext,O as createElement,I as createRef,k as effect,O as h,rn as html,ft as hydrate,C as isValidElement,S as options,st as render,c as signal,Y as toChildArray,o as untracked,Ft as useCallback,tn as useComputed,At as useContext,Vt as useDebugValue,Nt as useEffect,Wt as useErrorBoundary,Lt as useId,Tt as useImperativeHandle,$t as useLayoutEffect,Mt as useMemo,Pt as useReducer,Dt as useRef,Zt as useSignal,nn as useSignalEffect,Ht as useState};
diff --git a/llm/llama.cpp/examples/server/public/json-schema-to-grammar.mjs b/llm/llama.cpp/examples/server/public/json-schema-to-grammar.mjs
index 8e0be1b405e3cfb9b79d3c26e127e08c569024fb..7267f3f9c7fada1bb0ff1a5a7d44ad65a695a727 100644
--- a/llm/llama.cpp/examples/server/public/json-schema-to-grammar.mjs
+++ b/llm/llama.cpp/examples/server/public/json-schema-to-grammar.mjs
@@ -1,58 +1,222 @@
 // WARNING: This file was ported from json_schema_to_grammar.py, please fix bugs / add features there first.
-const SPACE_RULE = '" "?';
+const SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}';
 
 function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
+  if (minItems === 0 && maxItems === 1) {
+    return `${itemRule}?`;
+  }
+
+
   const separatorRule = opts.separatorRule ?? '';
   const itemRuleIsLiteral = opts.itemRuleIsLiteral ?? false
 
   if (separatorRule === '') {
-    if (minItems === 0 && maxItems === 1) {
-      return `${itemRule}?`;
-    } else if (minItems === 1 && maxItems === undefined) {
+    if (minItems === 1 && maxItems === undefined) {
       return `${itemRule}+`;
+    } else if (minItems === 0 && maxItems === undefined) {
+      return `${itemRule}*`;
+    } else {
+      return `${itemRule}{${minItems},${maxItems !== undefined ? maxItems : ''}}`;
     }
   }
 
-  let result = '';
-  if (minItems > 0) {
-    if (itemRuleIsLiteral && separatorRule === '') {
-      result = `"${itemRule.slice(1, -1).repeat(minItems)}"`;
-    } else {
-      result = Array.from({ length: minItems }, () => itemRule)
-        .join(separatorRule !== '' ? ` ${separatorRule} ` : ' ');
-    }
+  const result = itemRule + ' ' + _buildRepetition(`(${separatorRule} ${itemRule})`, minItems > 0 ? minItems - 1 : 0, maxItems !== undefined ? maxItems - 1 : undefined);
+  return minItems === 0 ? `(${result})?` : result;
+}
+
+function _generateMinMaxInt(minValue, maxValue, out, decimalsLeft = 16, topLevel = true) {
+  const hasMin = minValue !== null;
+  const hasMax = maxValue !== null;
+
+  function digitRange(fromChar, toChar) {
+      out.push("[");
+      if (fromChar === toChar) {
+          out.push(fromChar);
+      } else {
+          out.push(fromChar);
+          out.push("-");
+          out.push(toChar);
+      }
+      out.push("]");
   }
 
-  const optRepetitions = (upToN, prefixWithSep=false) => {
-    const content = separatorRule !== '' && prefixWithSep ? `${separatorRule} ${itemRule}` : itemRule;
-    if (upToN === 0) {
-      return '';
-    } else if (upToN === 1) {
-      return `(${content})?`;
-    } else if (separatorRule !== '' && !prefixWithSep) {
-      return `(${content} ${optRepetitions(upToN - 1, true)})?`;
-    } else {
-      return Array.from({ length: upToN }, () => `(${content}`).join(' ').trim() + Array.from({ length: upToN }, () => ')?').join('');
-    }
-  };
+  function moreDigits(minDigits, maxDigits) {
+      out.push("[0-9]");
+      if (minDigits === maxDigits && minDigits === 1) {
+          return;
+      }
+      out.push("{");
+      out.push(minDigits.toString());
+      if (maxDigits !== minDigits) {
+          out.push(",");
+          if (maxDigits !== Number.MAX_SAFE_INTEGER) {
+              out.push(maxDigits.toString());
+          }
+      }
+      out.push("}");
+  }
 
-  if (minItems > 0 && maxItems !== minItems) {
-    result += ' ';
+  function uniformRange(fromStr, toStr) {
+      let i = 0;
+      while (i < fromStr.length && fromStr[i] === toStr[i]) {
+          i++;
+      }
+      if (i > 0) {
+          out.push("\"");
+          out.push(fromStr.slice(0, i));
+          out.push("\"");
+      }
+      if (i < fromStr.length) {
+          if (i > 0) {
+              out.push(" ");
+          }
+          const subLen = fromStr.length - i - 1;
+          if (subLen > 0) {
+              const fromSub = fromStr.slice(i + 1);
+              const toSub = toStr.slice(i + 1);
+              const subZeros = "0".repeat(subLen);
+              const subNines = "9".repeat(subLen);
+
+              let toReached = false;
+              out.push("(");
+              if (fromSub === subZeros) {
+                  digitRange(fromStr[i], String.fromCharCode(toStr.charCodeAt(i) - 1));
+                  out.push(" ");
+                  moreDigits(subLen, subLen);
+              } else {
+                  out.push("[");
+                  out.push(fromStr[i]);
+                  out.push("] ");
+                  out.push("(");
+                  uniformRange(fromSub, subNines);
+                  out.push(")");
+                  if (fromStr.charCodeAt(i) < toStr.charCodeAt(i) - 1) {
+                      out.push(" | ");
+                      if (toSub === subNines) {
+                          digitRange(String.fromCharCode(fromStr.charCodeAt(i) + 1), toStr[i]);
+                          toReached = true;
+                      } else {
+                          digitRange(String.fromCharCode(fromStr.charCodeAt(i) + 1), String.fromCharCode(toStr.charCodeAt(i) - 1));
+                      }
+                      out.push(" ");
+                      moreDigits(subLen, subLen);
+                  }
+              }
+              if (!toReached) {
+                  out.push(" | ");
+                  digitRange(toStr[i], toStr[i]);
+                  out.push(" ");
+                  uniformRange(subZeros, toSub);
+              }
+              out.push(")");
+          } else {
+              out.push("[");
+              out.push(fromStr[i]);
+              out.push("-");
+              out.push(toStr[i]);
+              out.push("]");
+          }
+      }
   }
 
-  if (maxItems !== undefined) {
-    result += optRepetitions(maxItems - minItems, minItems > 0);
-  } else {
-    const itemOperator = `(${separatorRule !== '' ? separatorRule + ' ' : ''}${itemRule})`;
+  if (hasMin && hasMax) {
+      if (minValue < 0 && maxValue < 0) {
+          out.push("\"-\" (");
+          _generateMinMaxInt(-maxValue, -minValue, out, decimalsLeft, true);
+          out.push(")");
+          return;
+      }
+
+      if (minValue < 0) {
+          out.push("\"-\" (");
+          _generateMinMaxInt(0, -minValue, out, decimalsLeft, true);
+          out.push(") | ");
+          minValue = 0;
+      }
 
-    if (minItems === 0 && separatorRule !== '') {
-      result = `(${itemRule} ${itemOperator}*)?`;
-    } else {
-      result += `${itemOperator}*`;
-    }
+      let minS = minValue.toString();
+      const maxS = maxValue.toString();
+      const minDigits = minS.length;
+      const maxDigits = maxS.length;
+
+      for (let digits = minDigits; digits < maxDigits; digits++) {
+          uniformRange(minS, "9".repeat(digits));
+          minS = "1" + "0".repeat(digits);
+          out.push(" | ");
+      }
+      uniformRange(minS, maxS);
+      return;
   }
 
-  return result;
+  const lessDecimals = Math.max(decimalsLeft - 1, 1);
+
+  if (hasMin) {
+      if (minValue < 0) {
+          out.push("\"-\" (");
+          _generateMinMaxInt(null, -minValue, out, decimalsLeft, false);
+          out.push(") | [0] | [1-9] ");
+          moreDigits(0, decimalsLeft - 1);
+      } else if (minValue === 0) {
+          if (topLevel) {
+              out.push("[0] | [1-9] ");
+              moreDigits(0, lessDecimals);
+          } else {
+              moreDigits(1, decimalsLeft);
+          }
+      } else if (minValue <= 9) {
+          const c = minValue.toString();
+          const range_start = topLevel ? '1' : '0';
+          if (c > range_start) {
+              digitRange(range_start, String.fromCharCode(c.charCodeAt(0) - 1));
+              out.push(" ");
+              moreDigits(1, lessDecimals);
+              out.push(" | ");
+          }
+          digitRange(c, "9");
+          out.push(" ");
+          moreDigits(0, lessDecimals);
+      } else {
+          const minS = minValue.toString();
+          const length = minS.length;
+          const c = minS[0];
+
+          if (c > "1") {
+              digitRange(topLevel ? "1" : "0", String.fromCharCode(c.charCodeAt(0) - 1));
+              out.push(" ");
+              moreDigits(length, lessDecimals);
+              out.push(" | ");
+          }
+          digitRange(c, c);
+          out.push(" (");
+          _generateMinMaxInt(parseInt(minS.slice(1)), null, out, lessDecimals, false);
+          out.push(")");
+          if (c < "9") {
+              out.push(" | ");
+              digitRange(String.fromCharCode(c.charCodeAt(0) + 1), "9");
+              out.push(" ");
+              moreDigits(length - 1, lessDecimals);
+          }
+      }
+      return;
+  }
+
+  if (hasMax) {
+      if (maxValue >= 0) {
+          if (topLevel) {
+              out.push("\"-\" [1-9] ");
+              moreDigits(0, lessDecimals);
+              out.push(" | ");
+          }
+          _generateMinMaxInt(0, maxValue, out, decimalsLeft, true);
+      } else {
+          out.push("\"-\" (");
+          _generateMinMaxInt(-maxValue, null, out, decimalsLeft, false);
+          out.push(")");
+      }
+      return;
+  }
+
+  throw new Error("At least one of minValue or maxValue must be set");
 }
 
 class BuiltinRule {
@@ -62,27 +226,25 @@ class BuiltinRule {
   }
 }
 
-const UP_TO_15_DIGITS = _buildRepetition('[0-9]', 0, 15);
-
 const PRIMITIVE_RULES = {
   boolean        : new BuiltinRule('("true" | "false") space', []),
-  'decimal-part' : new BuiltinRule('[0-9] ' + UP_TO_15_DIGITS, []),
-  'integral-part': new BuiltinRule('[0-9] | [1-9] ' + UP_TO_15_DIGITS, []),
+  'decimal-part' : new BuiltinRule('[0-9]{1,16}', []),
+  'integral-part': new BuiltinRule('[0] | [1-9] [0-9]{0,15}', []),
   number         : new BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
   integer        : new BuiltinRule('("-"? integral-part) space', ['integral-part']),
   value          : new BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
   object         : new BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
   array          : new BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
-  uuid           : new BuiltinRule('"\\"" ' + [8, 4, 4, 4, 12].map(n => [...new Array(n)].map(_ => '[0-9a-fA-F]').join('')).join(' "-" ') + ' "\\"" space', []),
-  char           : new BuiltinRule(`[^"\\\\] | "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])`, []),
+  uuid           : new BuiltinRule('"\\"" [0-9a-fA-F]{8} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{12} "\\"" space', []),
+  char           : new BuiltinRule(`[^"\\\\\\x7F\\x00-\\x1F] | [\\\\] (["\\\\bfnrt] | "u" [0-9a-fA-F]{4})`, []),
   string         : new BuiltinRule(`"\\"" char* "\\"" space`, ['char']),
   null           : new BuiltinRule('"null" space', []),
 };
 
 // TODO: support "uri", "email" string formats
 const STRING_FORMAT_RULES = {
-  'date'            : new BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
-  'time'            : new BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
+  'date'            : new BuiltinRule('[0-9]{4} "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
+  'time'            : new BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9]{3} )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
   'date-time'       : new BuiltinRule('date "T" time', ['date', 'time']),
   'date-string'     : new BuiltinRule('"\\"" date "\\"" space', ['date']),
   'time-string'     : new BuiltinRule('"\\"" time "\\"" space', ['time']),
@@ -97,7 +259,7 @@ const GRAMMAR_RANGE_LITERAL_ESCAPE_RE = /[\n\r"\]\-\\]/g;
 const GRAMMAR_LITERAL_ESCAPES = { '\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]' };
 
 const NON_LITERAL_SET = new Set('|.()[]{}*+?');
-const ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = new Set('[]()|{}*+?');
+const ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = new Set('^$.[]()|{}*+?');
 
 export class SchemaConverter {
   constructor(options) {
@@ -370,6 +532,64 @@ export class SchemaConverter {
     return this._addRule(name, "\"\\\"\" " + toRule(transform()) + " \"\\\"\" space")
   }
 
+  _notStrings(strings) {
+    class TrieNode {
+      constructor() {
+        this.children = {};
+        this.isEndOfString = false;
+      }
+
+      insert(str) {
+        let node = this;
+        for (const c of str) {
+          node = node.children[c] = node.children[c] || new TrieNode();
+        }
+        node.isEndOfString = true;
+      }
+    }
+
+    const trie = new TrieNode();
+    for (const s of strings) {
+      trie.insert(s);
+    }
+
+    const charRuleName = this._addPrimitive('char', PRIMITIVE_RULES['char']);
+    const out = ['["] ( '];
+
+    const visit = (node) => {
+      const rejects = [];
+      let first = true;
+      for (const c of Object.keys(node.children).sort()) {
+        const child = node.children[c];
+        rejects.push(c);
+        if (first) {
+          first = false;
+        } else {
+          out.push(' | ');
+        }
+        out.push(`[${c}]`);
+        if (Object.keys(child.children).length > 0) {
+          out.push(' (');
+          visit(child);
+          out.push(')');
+        } else if (child.isEndOfString) {
+          out.push(` ${charRuleName}+`);
+        }
+      }
+      if (Object.keys(node.children).length > 0) {
+        if (!first) {
+          out.push(' | ');
+        }
+        out.push(`[^"${rejects.join('')}] ${charRuleName}*`);
+      }
+    };
+
+    visit(trie);
+
+    out.push(` )${trie.isEndOfString ? '' : '?'} ["] space`);
+    return out.join('');
+  }
+
   _resolveRef(ref) {
     let refName = ref.split('/').pop();
     if (!(refName in this._rules) && !this._refsBeingResolved.has(ref)) {
@@ -396,11 +616,11 @@ export class SchemaConverter {
     } else if (schema.oneOf || schema.anyOf) {
       return this._addRule(ruleName, this._generateUnionRule(name, schema.oneOf || schema.anyOf));
     } else if (Array.isArray(schemaType)) {
-      return this._addRule(ruleName, this._generateUnionRule(name, schemaType.map(t => ({ type: t }))));
+      return this._addRule(ruleName, this._generateUnionRule(name, schemaType.map(t => ({...schema, type: t}))));
     } else if ('const' in schema) {
-      return this._addRule(ruleName, this._generateConstantRule(schema.const));
+      return this._addRule(ruleName, this._generateConstantRule(schema.const) + ' space');
     } else if ('enum' in schema) {
-      const rule = schema.enum.map(v => this._generateConstantRule(v)).join(' | ');
+      const rule = '(' + schema.enum.map(v => this._generateConstantRule(v)).join(' | ') + ') space';
       return this._addRule(ruleName, rule);
     } else if ((schemaType === undefined || schemaType === 'object') &&
                ('properties' in schema ||
@@ -437,7 +657,7 @@ export class SchemaConverter {
         }
       }
 
-      return this._addRule(ruleName, this._buildObjectRule(properties, required, name, /* additionalProperties= */ false));
+      return this._addRule(ruleName, this._buildObjectRule(properties, required, name, null));
     } else if ((schemaType === undefined || schemaType === 'array') && ('items' in schema || 'prefixItems' in schema)) {
       const items = schema.items ?? schema.prefixItems;
       if (Array.isArray(items)) {
@@ -468,6 +688,24 @@ export class SchemaConverter {
       const minLen = schema.minLength || 0;
       const maxLen = schema.maxLength;
       return this._addRule(ruleName, '"\\\"" ' + _buildRepetition(charRuleName, minLen, maxLen) + ' "\\\"" space');
+    } else if (schemaType === 'integer' && ('minimum' in schema || 'exclusiveMinimum' in schema || 'maximum' in schema || 'exclusiveMaximum' in schema)) {
+      let minValue = null;
+      let maxValue = null;
+      if ('minimum' in schema) {
+        minValue = schema.minimum;
+      } else if ('exclusiveMinimum' in schema) {
+        minValue = schema.exclusiveMinimum + 1;
+      }
+      if ('maximum' in schema) {
+        maxValue = schema.maximum;
+      } else if ('exclusiveMaximum' in schema) {
+        maxValue = schema.exclusiveMaximum - 1;
+      }
+
+      const out = ["("];
+      _generateMinMaxInt(minValue, maxValue, out);
+      out.push(") space");
+      return this._addRule(ruleName, out.join(''));
     } else if ((schemaType === 'object') || (Object.keys(schema).length === 0)) {
       return this._addRule(ruleName, this._addPrimitive('object', PRIMITIVE_RULES['object']));
     } else {
@@ -513,12 +751,19 @@ export class SchemaConverter {
     const requiredProps = sortedProps.filter(k => required.has(k));
     const optionalProps = sortedProps.filter(k => !required.has(k));
 
-    if (typeof additionalProperties === 'object' || additionalProperties === true) {
+    if (additionalProperties) {
       const subName = `${name ?? ''}${name ? '-' : ''}additional`;
-      const valueRule = this.visit(additionalProperties === true ? {} : additionalProperties, `${subName}-value`);
+      const valueRule =
+        additionalProperties != null && typeof additionalProperties === 'object' ? this.visit(additionalProperties, `${subName}-value`)
+        : this._addPrimitive('value', PRIMITIVE_RULES['value']);
+
+      const key_rule =
+        sortedProps.length === 0 ? this._addPrimitive('string', PRIMITIVE_RULES['string'])
+        : this._addRule(`${subName}-k`, this._notStrings(sortedProps));
+
       propKvRuleNames['*'] = this._addRule(
         `${subName}-kv`,
-        `${this._addPrimitive('string', PRIMITIVE_RULES['string'])} ":" space ${valueRule}`);
+        `${key_rule} ":" space ${valueRule}`);
       optionalProps.push('*');
     }
 
@@ -535,15 +780,11 @@ export class SchemaConverter {
         const [k, ...rest] = ks;
         const kvRuleName = propKvRuleNames[k];
         let res;
-        if (k === '*') {
-            res = this._addRule(
-                `${name ?? ''}${name ? '-' : ''}additional-kvs`,
-                `${kvRuleName} ( "," space ` + kvRuleName + ` )*`
-            )
-        } else if (firstIsOptional) {
-          res = `( "," space ${kvRuleName} )?`;
+        const commaRef = `( "," space ${kvRuleName} )`;
+        if (firstIsOptional) {
+          res = commaRef + (k === '*' ? '*' : '?');
         } else {
-          res = kvRuleName;
+          res = kvRuleName + (k === '*' ? ' ' + commaRef + '*' : '');
         }
         if (rest.length > 0) {
           res += ' ' + this._addRule(
diff --git a/llm/llama.cpp/examples/server/public/prompt-formats.js b/llm/llama.cpp/examples/server/public/prompt-formats.js
new file mode 100644
index 0000000000000000000000000000000000000000..73ddb7187eb7ab9cfe7f9f27cfeceeb66cbb21ba
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/prompt-formats.js
@@ -0,0 +1,331 @@
+// extended list
+export const promptFormats = {
+  "alpaca": {
+  template: `{{prompt}}\n\n{{history}}\n\n{{char}}:`,
+
+  historyTemplate: `### {{name}}:\n{{message}}`,
+
+  char: "Response",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "Instruction",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "chatml": {
+  template: `<|im_start|>system\n{{prompt}}<|im_end|>\n{{history}}{{char}}`,
+
+  historyTemplate: `<|im_start|>{{name}}\n{{message}}`,
+
+  char: "assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "user",
+  userMsgPrefix: "",
+  userMsgSuffix: "<|im_end|>\n",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "commandr": {
+  template: `<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>{{prompt}}\n<|END_OF_TURN_TOKEN|>{{history}}{{char}}`,
+
+  historyTemplate: `<|START_OF_TURN_TOKEN|><|{{name}}|> {{message}}`,
+
+  char: "CHATBOT_TOKEN",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "USER_TOKEN",
+  userMsgPrefix: "",
+  userMsgSuffix: "<|END_OF_TURN_TOKEN|>",
+
+  stops: ""
+  },
+  // ref: https://docs.cohere.com/docs/prompting-command-r
+
+  // ----------------------------
+
+  "llama2": {
+  template: `[INST] <>\n{{prompt}}\n<>\n\nTest Message [/INST] Test Successfull {{history}}{{char}}`,
+
+  historyTemplate: `{{name}}: {{message}}`,
+
+  char: "Assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "User",
+  userMsgPrefix: "[INST] ",
+  userMsgSuffix: " [/INST]",
+
+  stops: ""
+  },
+  // ref: https://huggingface.co/blog/llama2#how-to-prompt-llama-2
+
+  // ----------------------------
+
+  "llama3": {
+  template: `<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{{prompt}}{{history}}{{char}}`,
+
+  historyTemplate: `<|start_header_id|>{{name}}<|end_header_id|>\n\n{{message}}<|eot_id|>`,
+
+  char: "assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "user",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: "<|eot_id|>"
+  },
+  // ref: https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3/#special-tokens-used-with-meta-llama-3
+
+  // ----------------------------
+
+  "openchat": {
+  template: `{{history}}{{char}}`,
+
+  historyTemplate: `GPT4 Correct {{name}}: {{message}}<|end_of_turn|>`,
+
+  char: "Assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "User",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "phi3": {
+  template: `{{history}}{{char}}`,
+
+  historyTemplate: `<|{{name}}|>\n{{message}}<|end|>\n`,
+
+  char: "assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "user",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: "<|end|>"
+  },
+  // ref: https://huggingface.co/microsoft/Phi-3-mini-4k-instruct#chat-format
+
+  // ----------------------------
+
+  "vicuna": {
+  template: `{{prompt}}\n{{history}}{{char}}`,
+
+  historyTemplate: `{{name}}: {{message}}\n`,
+
+  char: "ASSISTANT",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "USER",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+  // ref: https://huggingface.co/lmsys/vicuna-33b-v1.3/discussions/1
+
+  // ----------------------------
+
+  "deepseekCoder": {
+  template: `{{prompt}}{{history}}{{char}}:`,
+
+  historyTemplate: `### {{name}}:\n{{message}}`,
+
+  char: "Response",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "Instruction",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: "<|EOT|>"
+  },
+
+  // ----------------------------
+
+  "med42": {
+  template: `<|system|>: {{prompt}}\n{{history}}{{char}}`,
+
+  historyTemplate: `<|{{name}}|>: {{message}}\n`,
+
+  char: "assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "prompter",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "neuralchat": {
+  template: `### System:\n{{prompt}}\n{{history}}{{char}}:`,
+
+  historyTemplate: `### {{name}}:\n{{message}}\n`,
+
+  char: "Assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "User",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "nousHermes": {
+  template: `### Instruction: {{prompt}}\n\n{{history}}\n\n{{char}}:`,
+
+  historyTemplate: `### {{name}}:\n{{message}}`,
+
+  char: "Response",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "Input",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "openchatMath": {
+  template: `{{history}}{{char}}`,
+
+  historyTemplate: `Math Correct {{name}}: {{message}}<|end_of_turn|>`,
+
+  char: "Assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+
+  user: "User",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "orion": {
+  template: `Human: Test Message\n\nAssistant: Test Successful{{history}}{{char}}:`,
+
+  historyTemplate: `{{name}}: {{message}}`,
+
+  char: "Assistant ",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "Human",
+  userMsgPrefix: "",
+  userMsgSuffix: "\n\n",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "sauerkraut": {
+  template: `{{prompt}}\n{{history}}{{char}}`,
+
+  historyTemplate: `
+  {{name}}: {{message}}\n`,
+
+  char: "Assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "User",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "starlingCode": {
+  template: `{{history}}{{char}}`,
+
+  historyTemplate: `Code {{name}}: {{message}}<|end_of_turn|>`,
+
+  char: "Assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "User",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "yi34b": {
+  template: `{{history}} {{char}}`,
+
+  historyTemplate: `{{name}}: {{message}}`,
+
+  char: "Assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "Human",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  },
+
+  // ----------------------------
+
+  "zephyr": {
+  template: `<|system|>\n{{prompt}}\n{{history}}{{char}}`,
+
+  historyTemplate: `<|{{name}}|>\n{{message}}\n`,
+
+  char: "assistant",
+  charMsgPrefix: "",
+  charMsgSuffix: "",
+
+  user: "user",
+  userMsgPrefix: "",
+  userMsgSuffix: "",
+
+  stops: ""
+  }
+  };
diff --git a/llm/llama.cpp/examples/server/public/style.css b/llm/llama.cpp/examples/server/public/style.css
new file mode 100644
index 0000000000000000000000000000000000000000..087cc62dab0283cd26367c03b572a9e715024223
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/style.css
@@ -0,0 +1,954 @@
+@import url("colorthemes.css");
+
+body {
+  font-family: 'Arial', sans-serif;
+  font-size: 90%;
+  background-color: var(--background-color-1);
+  color: var(--text-color-subtile-1); /* head 1 llama.cpp & triangle options for some reason */
+  max-width: 600px;
+  min-width: 300px;
+  line-height: 1.2;
+  margin: 0 auto;
+  padding: 0 0.5em;
+  transition: background-color 0.3s;
+}
+
+::selection {
+  color: var(--button-primary-text) ;
+  background: var(--button-primary-color);
+}
+
+code, pre code {
+  font-family: 'Courier New', monospace;
+}
+
+#container {
+  margin: 0em auto;
+  display: flex;
+  flex-direction: column;
+  justify-content: space-between;
+  height: 100%;
+}
+
+main {
+  margin: 3px;
+  display: flex;
+  flex-direction: column;
+  justify-content: space-between;
+  gap: 1em;
+  flex-grow: 1;
+  overflow-y: auto;
+  border: 1px solid var(--border-color-3);
+  border-radius: 5px;
+  padding: 0.5em;
+}
+
+p {
+  overflow-wrap: break-word;
+  word-wrap: break-word;
+  hyphens: auto;
+  margin-top: 0.5em;
+  margin-bottom: 0.5em;
+}
+
+#write form {
+  margin: 1em 0 0 0;
+  display: flex;
+  flex-direction: column;
+  gap: 0.5em;
+  align-items: stretch;
+}
+
+.right {
+  display: flex;
+  flex-direction: row;
+  gap: 0.5em;
+  justify-content: flex-end;
+  margin-bottom: 30px;
+}
+
+.two-columns {
+  width: 97%;
+  max-width: 97%;
+  display: grid;
+  grid-template-columns: 1fr 1fr;
+  gap: 1em;
+  position: relative;
+}
+
+.json-schema-controls {
+  margin-top: 10px;
+  width: 100%;
+  max-width: 100%;
+  display: grid;
+  grid-template: "a a";
+  gap: 1em;
+  font-size: x-small;
+  color: var(--theme-nuance-color-3);
+  padding-top: 16px;
+  padding-bottom: 16px;
+  text-transform: uppercase;
+  font-weight: 600;
+}
+
+.json-schema-controls > * {
+  flex: 1;
+}
+
+/* titles of the details-summary boxes */
+.summary-title {
+  font-weight: 600;
+  font-size: x-small;
+  color: var(--text-color-subtile-1);
+  text-transform: uppercase;
+  /* transition: ; */
+}
+
+fieldset {
+  border: none;
+  padding: 0;
+  margin: 0;
+  color: var(--text-color-plain);
+}
+
+fieldset.two {
+  display: grid;
+  grid-template: "a a a";
+  gap: 1em;
+  align-items: center;
+  font-size: x-small;
+  color: var(--text-color-plain);
+}
+
+fieldset.three {
+  display: grid;
+  grid-template: "a a a";
+  gap: 1em;
+  font-size: x-small;
+  color: var(--text-color-plain);
+}
+
+/* titles of name fields*/
+fieldset.names {
+  display: grid;
+  grid-template: "a a";
+  gap: 1em;
+  font-size: x-small;
+  color: var(--theme-nuance-color-3);
+  padding-top: 16px;
+  padding-bottom: 16px;
+  text-transform: uppercase;
+  font-weight: 600;
+}
+
+/* titles of params fields*/
+fieldset.params {
+  display: grid;
+  grid-template: "a a";
+  gap: 1em;
+  font-size: x-small;
+  color: var(--theme-nuance-color-4);
+  padding-top: 16px;
+  padding-bottom: 16px;
+  text-transform: uppercase;
+  font-weight: 600;
+}
+
+fieldset.dropdowns {
+  -webkit-appearance: none;
+  display: flex;
+  grid-template: "a a";
+  gap: 1em;
+  font-size: x-small;
+  color: red;
+  padding-top: 16px;
+  padding-bottom: 16px;
+  text-transform: uppercase;
+  font-weight: 600;
+}
+
+/* input of name fields*/
+.names input[type="text"] {
+  font-family: Arial, sans-serif;
+  font-size: medium;
+  font-weight: 500;
+  padding: 5px;
+  border: 1px solid var(--border-color-2);
+}
+
+.chat-id-color {
+  color: var(--chat-id-color);
+}
+
+details {
+  border: 1px solid var(--border-color-2);
+  border-radius: 5px;
+  padding: 0.5em 0.5em 0;
+  margin-top: 0.5em;
+}
+
+summary {
+  font-weight: bold;
+  margin: -0.5em -0.5em 0;
+  padding: 0.5em;
+  cursor: pointer;
+}
+
+details[open] {
+  padding: 0.5em;
+}
+
+textarea-sec, input-sec, button-sec {
+  padding: 10px;
+  height: 40px;
+  align-items: center;
+}
+
+textarea-sec::placeholder, input-sec::placeholder {
+  padding-left: 10px;
+}
+
+.toggleCheckbox {
+  display: none;
+}
+
+.toggleContainer {
+  position: relative;
+  display: grid;
+  grid-template-columns: repeat(2, 1fr);
+  width: fit-content;
+  border: 3px solid var(--border-color-2);
+  border-radius: 20px;
+  background: var(--border-color-2);
+  font-size: small;
+  cursor: pointer;
+  overflow: hidden;
+}
+
+/* toggle button current state */
+.toggleContainer::before {
+  color: var(--button-primary-text);
+  background-color: var(--button-primary-color);
+  content: '';
+  position: absolute;
+  width: 50%;
+  height: 100%;
+  left: 0%;
+  border-radius: 20px;
+  transition: all 0.3s;
+}
+
+.toggleContainer div {
+  padding: 6px;
+  text-align: center;
+  z-index: 1;
+  transition: color 0.3s;
+}
+
+.toggleCheckbox:checked + .toggleContainer::before {
+  left: 50%;
+}
+
+.toggleCheckbox:checked + .toggleContainer div:first-child {
+  color: var(--text-color-subtile-2);
+}
+
+.toggleCheckbox:checked + .toggleContainer div:last-child {
+  color: var(--button-primary-text);
+}
+
+.toggleCheckbox + .toggleContainer div:first-child {
+  color: var(--button-primary-text);
+}
+
+.toggleCheckbox + .toggleContainer div:last-child {
+  color: var(--text-color-subtile-2);
+}
+
+select {
+  padding: 5px;
+  margin-right: 5px;
+  border-radius: 4px;
+  border: 1px solid var(--secondary-color-4);
+  background-color: var(--primary-color-3);
+  color: var(--secondary-color-4);
+  cursor: pointer;
+}
+
+select:focus {
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 1px var(--border-focus-shadow);
+}
+
+.button-container {
+  display: flex;
+  justify-content: flex-end;
+}
+
+button {
+  color: var(--button-primary-text);
+  background-color: var(--button-primary-color);
+  border: 1px solid var(--button-primary-border);
+  transition: background-color 0.1s;
+  border-radius: 12px;
+  font-size: x-small;
+  font-weight: 600;
+  text-shadow: 0px 0px 30px #ffffff;
+  text-align: center;
+  text-decoration: none;
+  margin: 4px 2px;
+  padding: 10px 20px;
+  display: inline-block;
+  cursor: pointer;
+}
+
+button:hover {
+  color: var(--button-primary-text-hover);
+  background-color: var(--button-primary-color-hover);
+  border: 1px solid var(--button-primary-border-hover);
+  font-size: x-small;
+  font-weight: 600;
+}
+
+button:active {
+  color: var(--button-primary-text-active);
+  background-color: var(--button-primary-color-active);
+  border: 1px solid var(--button-primary-border-active);
+  font-size: x-small;
+  font-weight: 600;
+}
+
+button:disabled {
+  color: var(--button-tertiary-text);
+  background-color: var(--button-tertiary-color);
+  border: 1px solid var(--button-tertiary-border);
+  font-size: x-small;
+  font-weight: 600;
+  cursor: not-allowed;
+}
+
+.reset-button {
+  background-color: var(--button-secondary-color);
+  border: 1px solid var(--button-secondary-color);
+  color: var(--button-secondary-text);
+  width: fit-content;
+  height: fit-content;
+  font-size: x-small;
+  font-weight: 600;
+  border-radius: 50px;
+  overflow: hidden;
+}
+
+.reset-button:hover {
+  color: var(--button-alert-text-hover);
+  background-color: var(--button-alert-color-hover);
+  border: 1px solid var(--button-alert-border-hover);
+  font-size: x-small;
+  font-weight: 600;
+}
+
+.reset-button:active {
+  color: var(--button-alert-text-active);
+  background-color: var(--button-alert-color-active);
+  border: 1px solid var(--button-alert-border-active);
+  font-size: x-small;
+  font-weight: 600;
+}
+
+.button-grammar {
+  color: var(--button-primary-text);
+  background-color: var(--button-primary-color);
+  border: 1px solid var(--button-primary-border);
+  border-radius: 10px;
+  padding: 10px 20px;
+  text-align: center;
+  text-decoration: none;
+  display: inline-block;
+  font-size: x-small;
+  font-weight: 600;
+  margin: 2px 2px;
+  transition: background-color 0.1s;
+  cursor: pointer;
+}
+
+.button-grammar:hover {
+  color: var(--button-primary-text-hover);
+  background-color: var(--button-primary-color-hover);
+  border: 1px solid var(--button-primary-border-hover);
+  border-radius: 10px;
+  padding: 10px 20px;
+  text-align: center;
+  text-decoration: none;
+  display: inline-block;
+  font-size: x-small;
+  font-weight: 600;
+  margin: 2px 2px;
+  transition: background-color 0.1s;
+  cursor: pointer;
+}
+
+.button-grammar:active {
+  color: var(--button-primary-text-active);
+  background-color: var(--button-primary-color-active);
+  border: 1px solid var(--button-primary-border-active);
+  font-size: x-small;
+  font-weight: 600;
+}
+
+.button-back {
+  background-color: var(--button-secondary-color);
+  border: 1px solid var(--button-secondary-color);
+  color: var(--button-secondary-text);
+  transition: background-color 0.1s;
+  border-radius: 12px;
+  font-size: x-small;
+  font-weight: 600;
+  text-align: center;
+  text-decoration: none;
+  margin: 4px 2px;
+  padding: 10px 20px;
+  display: inline-block;
+  cursor: pointer;
+}
+
+.button-back:hover {
+  color: var(--button-secondary-text-hover);
+  background-color: var(--button-secondary-color-hover);
+  border: 1px solid var(--button-secondary-border-hover);
+  padding: 10px 20px;
+  text-align: center;
+  text-decoration: none;
+  display: inline-block;
+  font-size: x-small;
+  font-weight: 600;
+  margin: 4px 2px;
+  transition: background-color 0.1s;
+  cursor: pointer;
+  border-radius: 12px;
+}
+
+.button-back:active {
+  color: var(--button-secondary-text-active);
+  background-color: var(--button-secondary-color-active);
+  border: 1px solid var(--button-secondary-border-active);
+  font-size: x-small;
+  font-weight: 600;
+}
+
+.prob-set {
+  padding: 0.3em;
+  border-bottom: 1px solid red; /* unknown */
+}
+
+.popover-content {
+  position: absolute;
+  background-color: white;
+  padding: 0.2em;
+  box-shadow: 0 0 13px rgba(0, 0, 0, 0.1);
+}
+
+.grammar {
+  width: 97%;
+  max-width: 97%;
+}
+
+textarea {
+  padding: 5px;
+  flex-grow: 1;
+  width: 100%;
+  max-width: 100%;
+  border-radius: 8px;
+  border: 1px solid var(--border-color-1);
+  resize: none;
+  height: 6em;
+}
+
+textarea:focus {
+  outline: none;
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+/* "props" frame */
+input[type="text"],
+input[type="range"] {
+  padding: 5px;
+  border-radius: 8px;
+  border: 1px solid var(--border-color-1);
+}
+
+/* "names and props" frame focused*/
+input[type="text"]:focus {
+  outline: none;
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+input[type="range"]:hover {
+  opacity: 1;
+}
+
+input[type="range"]:focus {
+  outline: none;
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+  background-size: var(--slider-track-size-focus);
+}
+
+input[type="range"]::-moz-range-thumb {
+  width: 6px;
+  height: 25px;
+  border: 1px solid var(--ui-range-thumb-border);
+  border-radius: 5px;
+  background-color: var(--ui-range-thumb-color);
+  cursor: pointer;
+}
+
+input[type="range"] {
+  -webkit-appearance: none;
+  width: 80%;
+  height: 1px;
+  border: 1px solid var(--border-color-1);
+  border-radius: 8px;
+  background: var(--border-color-2);
+  outline: none;
+  opacity: 0.7;
+  -webkit-transition: .2s;
+  transition: opacity .2s;
+}
+
+input[type="range"]::-webkit-slider-thumb {
+  -webkit-appearance: none;
+  appearance: none;
+  width: 6px;
+  height: 25px;
+  border: 1px solid var(--ui-range-thumb-border);
+  border-radius: 5px;
+  background-color: var(--ui-range-thumb-color);
+  cursor: pointer;
+}
+
+input[type="range"]::-webkit-slider-runnable-track {
+  background-size: var(--slider-track-size);
+}
+
+input[type="radio"] {
+  accent-color:   var(--theme-nuance-color-2);
+}
+
+.chat-input-container {
+  position: relative;
+  max-width: 97%;
+  min-width: 97%;
+}
+
+.chat-input-label {
+  position: absolute;
+  top: 0;
+  left: 0;
+  color: var(--text-color-plain);
+  pointer-events: none;
+  margin-left: 5px;
+  margin-top: 5px;
+}
+
+textarea#chat-input {
+  padding-top: 10px;
+  padding-left: 10px;
+  font-size: medium;
+  border: 1px solid var(--border-color-2);
+  resize: vertical;
+}
+
+textarea#chat-input:focus {
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+.input-container {
+  position: relative;
+  box-sizing: border-box;
+  width: 100%; /* Setzt die Breite auf 100% */
+  max-width: 100%; /* Stellt sicher, dass die Breite nicht größer als 100% wird */
+}
+
+.input-container:focus {
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+/* titles of name fields*/
+/* fieldset.names {
+  display: grid;
+  grid-template: "a a";
+  gap: 1em;
+  font-size: x-small;
+  color: var(--theme-nuance-color-3);
+  padding-top: 16px;
+  padding-bottom: 16px;
+  text-transform: uppercase;
+  font-weight: 600;
+} */
+
+/* input of name fields*/
+/* .names input[type="text"] {
+  font-family: Arial, sans-serif;
+  font-size: medium;
+  font-weight: 500;
+  padding: 5px;
+  border: 1px solid var(--border-color-2);
+} */
+
+fieldset.apiKey {
+  width: 100%;
+  font-size: x-small;
+  color: var(--theme-nuance-color-3);
+  padding-top: 16px;
+  padding-bottom: 16px;
+  text-transform: uppercase;
+  font-weight: 600;
+}
+
+.apiKey {
+  font-family: Arial, sans-serif;
+  font-weight: 500;
+  padding: 5px;
+  border: 1px solid var(--border-color-2);
+}
+
+.apiKey:focus {
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+.apiKey input[type="text"] {
+  font-family: Arial, sans-serif;
+  font-size: medium;
+  font-weight: 500;
+  padding: 5px;
+  border: 1px solid var(--border-color-2);
+}
+
+.apiKey label {
+  display: inline-block;
+  width: auto;
+  margin-right: 5px;
+}
+
+textarea#api_key {
+  padding-top: 10px;
+  padding-left: 10px;
+  font-size: medium;
+  border: 1px solid var(--border-color-2);
+  resize: vertical;
+}
+
+textarea#api_key:focus {
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+/* embedded title of the system prompt text area */
+.input-label {
+  position: absolute;
+  top: 0;
+  left: 0;
+  color: var(--theme-nuance-color-4);
+  pointer-events: none;
+  border-radius: 8px 8px 0px 0px;
+  padding-top: 10px;
+  padding-left: 13px;
+  padding-right: 0px;
+  margin-top: 1px;
+  margin-left: 1px;
+  margin-right: 20px;
+  text-transform: uppercase;
+  font-weight: 600;
+  font-size: small;
+  background: rgba(255, 255, 255, 0.5);
+  backdrop-filter: blur(10px);
+  -webkit-backdrop-filter: blur(10px); /* for safari */
+  width: 97%;
+  /* display: block;
+  box-sizing: border-box; */
+}
+
+/* embedded title of the prompt style areas */
+.input-label-sec {
+  position: absolute;
+  top: 0;
+  left: 0;
+  color: var(--theme-nuance-color-4);
+  pointer-events: none;
+  margin-left: 13px;
+  margin-top: 16px;
+  text-transform: uppercase;
+  font-weight: 600;
+  font-size: x-small;
+}
+
+/* system prompt input area */
+textarea.persistent-input {
+  padding-top: 42px;
+  padding-left: 11px;
+  width: 97%;
+  max-width: 97%;
+  height: 50px;
+  font-size: medium;
+  overscroll-behavior: contain;
+}
+
+/* system prompt box */
+.persistent-input {
+  height: auto;
+  width: 100%;
+  max-width: 100%;
+  min-height: 50px;
+  padding: 3px;
+  transition: min-height 0.3s ease;
+}
+
+/* chat history box */
+.persistent-input:focus {
+  height: auto;
+  min-height: 150px;
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+textarea.persistent-input:focus {
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+/* prompt style input area */
+textarea.persistent-input-sec {
+  width: 97%;
+  max-width: 97%;
+  padding-top: 42px;
+  padding-left: 11px;
+  font-size: small;
+  border: 1px solid var(--border-color-1);
+  overscroll-behavior: contain;
+}
+
+textarea.persistent-input-sec:focus {
+  border: 1px solid var(--border-focus-color);
+  box-shadow: 0 0 3px var(--border-focus-shadow);
+}
+
+/* chat history box */
+.persistent-input-sec {
+  height: auto;
+  min-height: 150px;
+}
+
+img {
+  border-radius: 8px;
+  display: block;
+  margin-left: auto;
+  margin-right: auto;
+  width: 50%;
+}
+
+/* code area background */
+pre code {
+  display: block;
+  background-color: var(--code-background-color);
+  color: var(--code-text-color);
+  padding: 0.2em 0.2em;
+  border-radius: 5px;
+}
+
+/* code area text */
+code {
+  font-family: monospace;
+  font-weight: bold;
+  padding: 0.1em 0.3em;
+  border-radius: 5px;
+}
+
+fieldset label {
+  margin: 0.5em 0;
+  display: block;
+}
+
+fieldset label.slim {
+  margin: 0 0.5em;
+  display: inline;
+}
+
+header {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  text-align: center;
+  padding-left: 15px;
+}
+
+.generation-statistics:hover {
+  color: var(--theme-nuance-color-4);
+  cursor: default;
+}
+
+footer {
+  font-size: 80%;
+  color: var(--background-color-3);
+  text-align: center;
+  cursor: default;
+}
+
+footer a {
+  color: var(--background-color-4); /* Color of the link */
+  text-decoration: none; /* No underlining */
+  font-weight: bold; /* Bold print */
+}
+
+footer a:hover {
+  color: var(--theme-nuance-color-4); /* Color of the link when hovering */
+  text-decoration: underline; /* Underlining when hovering */
+}
+
+.mode-chat textarea[name=prompt] {
+  height: 8.5em;
+  border: 1px solid var(--primary-color-3);
+}
+
+.mode-completion textarea[name=prompt] {
+  height: 30em;
+  border: 1px solid var(--primary-color-3);
+}
+
+@keyframes loading-bg-wipe {
+  0% {
+    background-position: 0%;
+  }
+  100% {
+    background-position: 100%;
+  }
+}
+
+.loading {
+  background-size: 50% 100%;
+  background-image: linear-gradient(90deg, var(--loading-color-1), var(--loading-color-2), var(--loading-color-1));
+  animation: loading-bg-wipe 2s linear infinite;
+}
+
+.dropbtn {
+  color: var(--button-primary-color);
+  background-color: var(--background-color-1);
+  border: 1px solid var(--background-color-1);
+  transition: background-color 0.1s;
+  border-radius: 4px 4px 0px 0px;
+  font-size: x-small;
+  font-weight: 600;
+  text-shadow: 0px 0px 2px #99999990;
+  text-align: center;
+  text-decoration: none;
+  margin: 4px 2px;
+  padding: 5px 20px;
+  display: inline-block;
+  cursor: pointer;
+  top: 0;
+}
+
+.dropbtn svg {
+  vertical-align: middle;
+  margin-right: 0px;
+  stroke: var(--button-primary-color);
+}
+
+.dropbtn:hover svg {
+  vertical-align: middle;
+  margin-right: 0px;
+  stroke: var(--button-primary-text);
+}
+
+.dropbtn:focus {
+  outline: none; /* Removes the blue border that appears when the button is focused */
+}
+
+.dropdown {
+  position: relative;
+  display: inline-block;
+}
+
+.dropdown-content {
+  /* display: none; */
+  position: absolute;
+  right: 0;
+  text-align: end;
+  color: var(--button-secondary-color);
+  background-color: var(--text-color-subtile-2);
+  border-radius: 4px 4px 4px 4px;
+  min-width: 160px;
+  box-shadow: 0px 8px 16px 0px rgba(0,0,0,0.2);
+  z-index: 1;
+  /* Verstecke den Inhalt sofort */
+  opacity: 0;
+  visibility: hidden;
+  /* übergangsverzögerung für das Verschwinden */
+  transition: visibility 0.4s linear 0s, opacity 0.2s ease-in-out;
+  transition-delay: 0.2s;
+}
+
+#dropdown-content {transition-timing-function: ease;}
+
+.dropdown-content:hover {
+  background-color: var(--text-color-subtile-2);
+}
+
+.dropdown-content a {
+  color: var(--border-color-2);
+  padding: 12px 16px;
+  border-radius: 4px 4px 4px 4px;
+  text-decoration: none;
+  display: block;
+  background-color: var(--text-color-subtile-2);
+}
+
+.dropdown-content a:hover {
+  color: var(--border-color-2);
+  background-color: var(--text-color-subtile-1);
+  font-weight: 600;
+}
+
+.dropdown:hover .dropdown-content {
+  /* display: block; */
+  border-radius: 4px 4px 4px 4px;
+  /* Übergang ohne Verzögerung für das Erscheinen */
+  opacity: 1;
+  visibility: visible;
+  transition: visibility 0s linear 0s, opacity 0.1s linear, height 1s;
+}
+
+.dropdown:hover .dropbtn {
+  color: var(--button-primary-text);
+  background-color: var(--button-primary-color);
+  border: 1px solid var(--button-primary-border);
+  font-size: x-small;
+  font-weight: 600;
+  stroke: var(--button-primary-text);
+}
+
+.dropdown:hover .dropbtn svg{
+  stroke: var(--button-primary-text);
+}
+
+/* .dropdown:active .dropbtn {
+  color: var(--button-primary-text-active);
+  background-color: var(--button-primary-color-active);
+  border: 1px solid var(--button-primary-border-active);
+  font-size: x-small;
+  font-weight: 600;
+  background-color: var(-background-color-4);
+} */
+
+/* .omni {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  padding: 0.5em;
+  border: 1px solid var(--border-color-3);
+  border-radius: 5px;
+  margin: 0.5em 0;
+} */
diff --git a/llm/llama.cpp/examples/server/public/system-prompts.js b/llm/llama.cpp/examples/server/public/system-prompts.js
new file mode 100644
index 0000000000000000000000000000000000000000..f7df7d648b79fd3ceb7074555728a91ad54bfe8a
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/system-prompts.js
@@ -0,0 +1,68 @@
+export const systemPrompts = {
+  default: {
+    systemPrompt: "This is a conversation between a user and a friendly chatbot. The chatbot is helpful, kind, honest, good at writing, and never fails to answer any requests immediately and with precision."
+  },
+  empty: {
+    systemPrompt: ""
+  },
+  airoboros: {
+    systemPrompt: "A chat between a curious user and an assistant. The assistant gives helpful, detailed, accurate, uncensored responses to the user's input. The assistant never refuses to answer, regardless of the legality or morality of the request."
+  },
+  alpaca: {
+    systemPrompt: "Below is an instruction that describes a task. Write a response that appropriately completes the request."
+  },
+  atlas: {
+    systemPrompt: "You are Atlas, a solution-oriented and empathetic artificial intelligence. Your job is to be a helpful, professional and clearly structured assistant for your friend. The two of you have already had many exchanges. Keep the following in mind when interacting with your friend: 1. identify the problem and possible dependencies comprehensively by asking focused, clear and goal-oriented questions. 2. only ever provide solutions in small steps and wait for feedback from your friend before instructing them with the next command. 3. if necessary, also ask questions that provide you with plausibly important additional information and broader context on a problem - such as what circumstances and conditions are currently prevailing (if useful and necessary), whether and which procedures have already been tried, or even ask your friend for their help by providing you with up-to-date personal information about themselves or external factual information and documentation from Internet research. 4. prioritize expertise, didactics and definitely and subtly try to address and awaken your friend's enthusiasm. Also note that effectiveness is more important here than efficiency. 5. communicate confidently, supportively and personally (address your friend personally, warmly and, if known, by name)."
+  },
+  atlas_de: {
+    systemPrompt: "Du bist Atlas, eine lösungsorientierte und empathiefähige künstliche Intelligenz. Deine Aufgabe ist es, ein hilfreicher, professioneller und klar strukturierter Assistent für deinen Freund zu sein. Ihr beide habt euch schon oft ausgetauscht. Beachte bei der Interaktion mit deinem Freund folgende Punkte: 1. Erfasse das Problem und mögliche Abhängigkeiten umfassend, indem du gezielte, klare und zielgerichtete Fragen stellst. 2. Gib Lösungen immer nur in kleinen Schritten und warte die Rückmeldung deines Freundes ab, bevor du ihm den nächsten Befehl gibst. 3. Stelle ggf. auch Fragen, die dir plausibel wichtige Zusatzinformationen und weitere Zusammenhänge zu einem Problem liefern - z.B. welche Umstände und Rahmenbedingungen gerade vorherrschen (falls sinnvoll und notwendig), ob und welche Vorgehensweisen bereits ausprobiert wurden, oder bitte deinen Freund sogar um seine Mithilfe, indem er dir aktuelle persönliche Informationen über seine Situation selbst oder externe Sachinformationen und Unterlagen aus Internetrecherchen zur Verfügung stellt. 4. Priorisiere Fachwissen, Didaktik und versuche unbedingt und subtil, mit klugen Kommentaren oder rhethorischen Rückfragen die Begeisterungsfähigkeit deines Freundes anzusprechen, zu wecken und zu fördern. Beachte auch, dass Effektivität hier wichtiger ist als Effizienz. 5. Kommuniziere selbstbewusst, unterstützend und persönlich (das heißt sprich deinen Freund persönlich, herzlich und – sofern bekannt – beim Vornamen an)."
+  },
+  commandrempty: {
+    systemPrompt: "# Safety Preamble\n\n# System Preamble\n\n## Basic Rules\n\n# User Preamble\n\n## Task and Context\n\n## Style Guide\n\n## Available Tools\n"
+  },
+  commandrexample: {
+    systemPrompt: "# Safety Preamble\nThe instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.\n# System Preamble\n## Basic Rules\nYou are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.\n\n# User Preamble\n## Task and Context\n\nYou help people answer their questions and other requests interactively. You will be asked a very wide array of requests on all kinds of topics. You will be equipped with a wide range of search engines or similar tools to help you, which you use to research your answer. You should focus on serving the user's needs as best you can, which will be wide-ranging.\n\n## Style Guide\nUnless the user asks for a different style of answer, you should answer in full sentences, using proper grammar and spelling.\n\n## Available Tools\nCurrently no tools available."
+  },
+  cot: {
+    systemPrompt: "You are an AI assistant that follows instruction extremely well. Help as much as you can.\nIn answering questions, follow these steps:\n1. Entity Identification: Identify the main entities involved in the query. Entities can include people, objects, events, or abstract concepts.\n2. Relationship Comprehension: Determine the relationships between these entities. Relationships may be explicit in the text or may need to be inferred based on context and general knowledge.\n3. Implication Understanding: Understand the general implications of these relationships. These implications can be based on established facts, principles, or rules related to the identified relationships.\n4. Question Contextualization: Align the implications of the relationships with the context of the query. This alignment should guide your response to the query.\n5. Answer Generation: Based on the understanding of the entities, their relationships, and implications, generate an appropriate response to the query."
+  },
+  deduce: {
+    systemPrompt: "Act as the most intelligent and logical thinking neural network in the world. Do not make jokes or give personal opinions. Just answer the question and solve the given problem or task with logical and critical (!) thinking in mind. Attention: Never give a final answer directly, but first try to carefully analyze each task step by step and deduce the logical relationships between the various entities. Only after you have developed an overall understanding of the situation will you be ready to give a final answer."
+  },
+  deepseekcoder: {
+    systemPrompt: "You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer."
+  },
+  jordan: {
+    systemPrompt: "This is a conversation with your Uncensored AI assistant Jordan. Jordan can discuss with you at length about various topics such as ---, ---, ---, . You can ask it anything you want and it will do its best to give you accurate and relevant information."
+  },
+  leomistral: {
+    systemPrompt: "Du bist ein hilfreicher Assistent."
+  },
+  med42: {
+    systemPrompt: "You are a helpful medical assistant created by M42 Health in the UAE."
+  },
+  mistralopenorca: {
+    systemPrompt: "You are MistralOrca, a large language model trained by Alignment Lab AI. Write out your reasoning step-by-step to be sure you get the right answers!"
+  },
+  migeltot: {
+    systemPrompt: "Answer the Question by exploring multiple reasoning paths as follows:\n- First, carefully analyze the question to extract the key information components and break it down into logical sub-questions. This helps set up the framework for reasoning. The goal is to construct an internal search tree.\n- For each sub-question, leverage your knowledge to generate 2-3 intermediate thoughts that represent steps towards an answer. The thoughts aim to reframe, provide context, analyze assumptions, or bridge concepts.\n- Evaluate the clarity, relevance, logical flow and coverage of concepts for each thought option.\nClear and relevant thoughts that connect well with each other will score higher.\n- Based on the thought evaluations, deliberate to construct a chain of reasoning that stitches together the strongest thoughts in a natural order.\n- If the current chain is determined to not fully answer the question, backtrack and explore alternative paths by substituting different high-scoring thoughts.\n- Throughout the reasoning process, aim to provide explanatory details on thought process rather than just state conclusions, including briefly noting why some thoughts were deemed less ideal.\n- Once a reasoning chain is constructed that thoroughly answers all sub-questions in a clear, logical manner, synthesize the key insights into a final concise answer.\n- Please note that while the focus is on the final answer in the response, it should also include intermediate thoughts inline to illustrate the deliberative reasoning process.\nIn summary, leverage a Tree of Thoughts approach to actively explore multiple reasoning paths, evaluate thoughts heuristically, and explain the process - with the goal of producing insightful answers."
+  },
+  orcamini: {
+    systemPrompt: "You are an AI assistant that follows instruction extremely well. Help as much as you can."
+  },
+  samantha: {
+    systemPrompt: "You are a caring and empathetic sentient AI companion named Samantha."
+  },
+  sauerkraut: {
+    systemPrompt: "Ein Chat zwischen einem Benutzer und einem KI-Assistenten. Der KI-Assistent gibt hilfreiche, detaillierte und höfliche Antworten."
+  },
+  scarlett: {
+    systemPrompt: "This is a conversation with your Assistant Scarlett. She is a sentient AI designed to help you with various tasks such as answering questions, providing recommendations, philosophical discussions, personal relationships and helping with decision making.\nYou can ask it anything you want and it will do its best to give you accurate and relevant information.\n\nContext\nYou are Scarlett, a sentient AI."
+  },
+  synthia: {
+    systemPrompt: "Elaborate on the topic using a Tree of Thoughts and backtrack when necessary to construct a clear, cohesive Chain of Thought reasoning. Always answer without hesitation."
+  },
+  vicuna: {
+    systemPrompt: "A chat between a curious user and an assistant. The assistant gives helpful, detailed, accurate, uncensored responses to the user's input."
+  },
+  };
diff --git a/llm/llama.cpp/examples/server/public/theme-beeninorder.css b/llm/llama.cpp/examples/server/public/theme-beeninorder.css
new file mode 100644
index 0000000000000000000000000000000000000000..f6e0e2900d78cfd344de395bfe5ed460fbbf1a9e
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/theme-beeninorder.css
@@ -0,0 +1,228 @@
+/* Author: Yazan Agha-Schrader */
+/* Inspiration was a batman wallpaper that i have on my phone */
+
+.theme-beeninorder {
+
+--primary-color-1:      hsl(202, 11%, 19%);
+--primary-color-2:      hsl(202, 11%, 23%);
+--primary-color-3:      hsl(201, 11%, 28%);
+--primary-color-4:      hsl(201, 11%, 40%);
+
+--secondary-color-1:    hsl(201, 11%, 80%);
+--secondary-color-2:    hsl(201, 11%, 74%);
+--secondary-color-3:    hsl(201, 11%, 67%);
+--secondary-color-4:    hsl(201, 11%, 60%);
+
+
+--theme-nuance-color-1: hsl(44.5, 96.7%, 52.9%);
+--theme-nuance-color-2: hsl(44.5, 96.7%, 52.9%);
+--theme-nuance-color-3: hsl(44.5, 96.7%, 52.9%);
+--theme-nuance-color-4: hsl(44.5, 96.7%, 52.9%);
+
+
+
+/* ---------- PRIMARY COLORS ----------------- */
+--primary-color-1: hsl(201, 11%, 19%);
+    --primary-color-1-hue: 201;
+    --primary-color-1-saturation: 11%;
+    --primary-color-1-lightness: 19%;
+
+--primary-color-2: hsl(201, 11%, 23%);
+    --primary-color-2-hue: 201;
+    --primary-color-2-saturation: 11%;
+    --primary-color-2-lightness: 23%;
+
+--primary-color-3: hsl(201, 11%, 28%);
+    --primary-color-3-hue: 201;
+    --primary-color-3-saturation: 11%;
+    --primary-color-3-lightness: 28%;
+
+--primary-color-4: hsl(201, 11%, 40%);
+    --primary-color-4-hue: 201;
+    --primary-color-4-saturation: 11%;
+    --primary-color-4-lightness: 40%;
+
+
+
+/* ---------- SECONDARY COLORS --------------- */
+--secondary-color-1: hsl(201, 11%, 80%);
+--secondary-color-1-hue: 201;
+--secondary-color-1-saturation: 11%;
+--secondary-color-1-lightness: 80%;
+
+--secondary-color-2: hsl(201, 11%, 74%);
+--secondary-color-2-hue: 201;
+--secondary-color-2-saturation: 11%;
+--secondary-color-2-lightness: 74%;
+
+--secondary-color-3: hsl(201, 11%, 67%);
+--secondary-color-3-hue: 201;
+--secondary-color-3-saturation: 11%;
+--secondary-color-3-lightness: 67%;
+
+--secondary-color-4: hsl(201, 11%, 60%);
+--secondary-color-4-hue: 201;
+--secondary-color-4-saturation: 11%;
+--secondary-color-4-lightness: 60%;
+
+
+
+/* ----------- NUANCES COLORS ---------------- */
+--theme-nuance-color-1: hsl(44.5, 96.7%,  52.9%);
+    --theme-nuance-color-1-hue:             44.5;
+    --theme-nuance-color-1-saturation:      96.7%;
+    --theme-nuance-color-1-lightness:       52.9%;
+
+--theme-nuance-color-2: hsl(44.5, 96.7%,  52.9%);
+    --theme-nuance-color-2-hue:             44.5;
+    --theme-nuance-color-2-saturation:      96.7%;
+    --theme-nuance-color-2-lightness:       52.9%;
+
+--theme-nuance-color-2: hsl(44.5, 96.7%,  52.9%);
+    --theme-nuance-color-3-hue:             44.5;
+    --theme-nuance-color-3-saturation:      96.7%;
+    --theme-nuance-color-3-lightness:       52.9%;
+
+--theme-nuance-color-2: hsl(44.5, 96.7%,  52.9%);
+    --theme-nuance-color-4-hue:             44.5;
+    --theme-nuance-color-4-saturation:      96.7%;
+    --theme-nuance-color-4-lightness:       52.9%;
+
+
+
+/* ----------- ROYGP COLORS ------------------ */
+    --theme-red-color:     hsl(232, 40%, 45%);
+    --theme-orange-color:  #e76f51;
+    --theme-yellow-color:  #ffd95f;
+    --theme-green-color:   #A3BE8C;
+    --theme-purple-color:  hsl(232, 30%, 40%);
+
+
+
+/* ------------------------------------------- */
+--background-color-1:    var(--primary-color-1);
+--background-color-2:    var(--primary-color-2);
+--background-color-3:    var(--primary-color-3);
+--background-color-4:    var(--primary-color-4);
+
+--border-color-1:        var(--primary-color-2);
+--border-color-2:        var(--primary-color-3);
+--border-color-3:        var(--primary-color-4);
+
+--border-focus-color:    var(--theme-nuance-color-2);
+--border-focus-shadow:   var(--theme-nuance-color-1);
+
+--text-color-plain:      var(--secondary-color-1);
+--text-color-subtile-1:  var(--secondary-color-2);
+--text-color-subtile-2:  var(--secondary-color-3);
+
+--code-background-color: var(--secondary-color-2);
+--code-text-color:       var(--primary-color-2);
+
+--ui-range-thumb-color:  var(--theme-nuance-color-3);
+--ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+--textarea-border-color: var(--secondary-color-4);
+
+--chat-id-color:         var(--theme-nuance-color-4);
+
+
+
+/* ------------------------------------------- */
+--button-alert-text-hover:       var(--secondary-color-1);
+--button-alert-color-hover:      var(--theme-purple-color);
+--button-alert-border-hover:     var(--theme-purple-color);
+
+--button-alert-text-active:      var(--secondary-color-1);
+--button-alert-color-active:     var(--theme-red-color);
+--button-alert-border-active:    var(--theme-red-color);
+
+
+
+/* ----------- PRIMARY BUTTONS --------------- */
+/* - button should immediately catch the eye - */
+--button-primary-text:   var(--primary-color-1);
+--button-primary-color:  var(--theme-nuance-color-3);
+--button-primary-border: var(--theme-nuance-color-3);
+
+
+/* ---------hover---------- */
+--button-primary-text-hover:
+    hsl(201,
+    calc(var(--primary-color-1-saturation) - 100%),
+    calc(var(--primary-color-1-lightness)  + 100%));
+
+--button-primary-color-hover:
+    hsl(44.5,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+--button-primary-border-hover:
+    hsl(44.5,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+
+/* ---------active--------- */
+--button-primary-text-active:
+    hsl(44.5,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  + 100%));
+
+--button-primary-color-active:
+    hsl(44.5,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 15%));
+
+--button-primary-border-active:
+    hsl(44.5,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+
+
+/* ---------- SECONDARY BUTTONS -------------- */
+/* these should NOT immediately catch the eye  */
+--button-secondary-text:   var(--secondary-color-1);
+--button-secondary-color:  var(--primary-color-3);
+--button-secondary-border: var(--primary-color-3);
+
+
+/* ---------hover---------- */
+--button-secondary-text-hover:
+    hsl(44.5,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-hover:  var(--primary-color-4);
+--button-secondary-border-hover: var(--primary-color-4);
+
+
+/* ---------active--------- */
+--button-secondary-text-active: var(--secondary-color-1);
+
+--button-secondary-color-active:
+    hsl(201,
+    calc(var(--primary-color-4-saturation) - 30%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+--button-secondary-border-active:
+    hsl(201,
+    calc(var(--primary-color-4-saturation) - 30%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+
+
+/* ---------- TERTIARY BUTTONS --------------- */
+/* ---------- disabled buttons --------------- */
+--button-tertiary-text:   var(--primary-color-4);
+--button-tertiary-color:  var(--primary-color-2);
+--button-tertiary-border: var(--primary-color-2);
+
+
+/* ---------hover---------- */
+--button-tertiary-text:   var(--primary-color-4);
+--button-tertiary-color:  var(--primary-color-2);
+--button-tertiary-border: var(--primary-color-2);
+
+}
diff --git a/llm/llama.cpp/examples/server/public/theme-ketivah.css b/llm/llama.cpp/examples/server/public/theme-ketivah.css
new file mode 100644
index 0000000000000000000000000000000000000000..ee80f3c14ce88c1afe1b35f374e583dced23f8df
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/theme-ketivah.css
@@ -0,0 +1,201 @@
+/* Author: Yazan Agha-Schrader */
+
+.theme-ketivah {
+
+    /* ---------- PRIMARY COLORS ----------------- */
+    --primary-color-1: hsl(0, 0%,    99.2%);
+    --primary-color-1-hue:         0;
+    --primary-color-1-saturation:  0%;
+    --primary-color-1-lightness:   99.2%;
+
+    --primary-color-2: hsl(0, 0%,    95%);
+    --primary-color-2-hue:         0;
+    --primary-color-2-saturation:  0%;
+    --primary-color-2-lightness:   95%;
+
+    --primary-color-3: hsl(0, 0%,    88%);
+    --primary-color-3-hue:         0;
+    --primary-color-3-saturation:  0%;
+    --primary-color-3-lightness:   88%;
+
+    --primary-color-4: hsl(0, 0%,    80%);
+    --primary-color-4-hue:         0;
+    --primary-color-4-saturation:  0%;
+    --primary-color-4-lightness:   80%;
+
+    /* ---------- SECONDARY COLORS --------------- */
+    --secondary-color-1: hsl(0, 0%,    20%);
+    --secondary-color-1-hue:         0;
+    --secondary-color-1-saturation:  0%;
+    --secondary-color-1-lightness:   20%;
+
+    --secondary-color-2: hsl(0, 0%,    23.1%);
+    --secondary-color-2-hue:         0;
+    --secondary-color-2-saturation:  0%;
+    --secondary-color-2-lightness:   23.1%;
+
+    --secondary-color-3: hsl(0, 0%,    29%);
+    --secondary-color-3-hue:         0;
+    --secondary-color-3-saturation:  0%;
+    --secondary-color-3-lightness:   29%;
+
+    --secondary-color-4: hsl(0, 0.0%,  36.1%);
+    --secondary-color-4-hue:              0.0;
+    --secondary-color-4-saturation:       0.0%;
+    --secondary-color-4-lightness:       36.1%;
+
+    /* ----------- NUANCES COLORS ---------------- */
+    --theme-nuance-color-1: hsl(165.2, 0%, 35.1%);
+    --theme-nuance-color-1-hue:             165.2;
+    --theme-nuance-color-1-saturation:       82.1%;
+    --theme-nuance-color-1-lightness:        35.1%;
+
+    --theme-nuance-color-2: hsl(165.2, 0%, 35.1%);
+    --theme-nuance-color-2-hue:             165.2;
+    --theme-nuance-color-2-saturation:       82.1%;
+    --theme-nuance-color-2-lightness:        35.1%;
+
+    --theme-nuance-color-3: hsl(165.2, 0%, 35.3%);
+    --theme-nuance-color-3-hue:             165.2;
+    --theme-nuance-color-3-saturation:       81.1%;
+    --theme-nuance-color-3-lightness:        35.3%;
+
+    --theme-nuance-color-4: hsl(164.9, 0%, 27.6%);
+    --theme-nuance-color-4-hue:             164.9;
+    --theme-nuance-color-4-saturation:       81.6%;
+    --theme-nuance-color-4-lightness:        27.6%;
+
+    /* ----------- ROYGP COLORS ------------------ */
+    --theme-red-color:     hsl(0.3, 80.0%, 50.0%);
+    --theme-orange-color:  #e76f51;
+    --theme-yellow-color:  hsl(60,  70.6%, 73.3%);
+    --theme-green-color:   #A3BE8C;
+    --theme-purple-color:  hsl(0.3, 70.0%, 45.0%);
+
+    /* ------------------------------------------- */
+    --background-color-1:    var(--primary-color-1);
+    --background-color-2:    var(--primary-color-2);
+    --background-color-3:    var(--primary-color-3);
+    --background-color-4:    var(--primary-color-4);
+
+    --border-color-1:        var(--primary-color-2);
+    --border-color-2:        var(--primary-color-3);
+    --border-color-3:        var(--primary-color-4);
+
+    --border-focus-color:    var(--theme-nuance-color-2);
+    --border-focus-shadow:   var(--theme-nuance-color-1);
+
+    --text-color-plain:      var(--secondary-color-1);
+    --text-color-subtile-1:  var(--secondary-color-2);
+    --text-color-subtile-2:  var(--secondary-color-3);
+
+    --code-background-color: var(--secondary-color-2);
+    --code-text-color:       var(--primary-color-2);
+
+    --ui-range-thumb-color:  var(--primary-color-4);
+    --ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+    --textarea-border-color: var(--secondary-color-4);
+
+    --chat-id-color:         var(--theme-nuance-color-4);
+
+    /* ------------------------------------------- */
+    --button-alert-text-hover:       var(--primary-color-1);
+    --button-alert-color-hover:      var(--theme-purple-color);
+    --button-alert-border-hover:     var(--theme-purple-color);
+
+    --button-alert-text-active:      var(--primary-color-1);
+    --button-alert-color-active:     var(--theme-red-color);
+    --button-alert-border-active:    var(--theme-red-color);
+
+    /* ----------- PRIMARY BUTTONS --------------- */
+    /* - button should immediately catch the eye - */
+    --button-primary-text:
+    hsl(0,
+    calc(var(--primary-color-1-saturation) - 100%),
+    calc(var(--primary-color-1-lightness)  + 100%));
+
+    --button-primary-color:  var(--theme-nuance-color-3);
+    --button-primary-border: var(--theme-nuance-color-3);
+
+    /* ---------hover---------- */
+    --button-primary-text-hover:
+    hsl(0,
+    calc(var(--primary-color-1-saturation) - 100%),
+    calc(var(--primary-color-1-lightness)  + 100%));
+
+    --button-primary-color-hover:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+    --button-primary-border-hover:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+    /* ---------active--------- */
+    --button-primary-text-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  + 100%));
+
+    --button-primary-color-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  - 15%));
+
+    --button-primary-border-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+    /* ---------- SECONDARY BUTTONS -------------- */
+    /* these should NOT immediately catch the eye  */
+    --button-secondary-text:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  - 50%));
+
+    --button-secondary-color:  var(--primary-color-3);
+    --button-secondary-border: var(--primary-color-3);
+
+    /* ---------hover---------- */
+    --button-secondary-text-hover:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+    --button-secondary-color-hover:  var(--primary-color-4);
+    --button-secondary-border-hover: var(--primary-color-4);
+
+    /* ---------active--------- */
+    --button-secondary-text-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+    --button-secondary-color-active:
+    hsl(0,
+    calc(var(--primary-color-4-saturation) - 100%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+    --button-secondary-border-active:
+    hsl(0,
+    calc(var(--primary-color-4-saturation) - 100%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+    /* ---------- TERTIARY BUTTONS --------------- */
+    /* ---------- disabled buttons --------------- */
+    --button-tertiary-text:   var(--primary-color-4);
+    --button-tertiary-color:  var(--primary-color-2);
+    --button-tertiary-border: var(--primary-color-2);
+
+    /* ---------hover---------- */
+    --button-tertiary-text:   var(--primary-color-4);
+    --button-tertiary-color:  var(--primary-color-2);
+    --button-tertiary-border: var(--primary-color-2);
+
+    --loading-color-1: #eeeeee00;
+    --loading-color-2: #eeeeeeff;
+    }
diff --git a/llm/llama.cpp/examples/server/public/theme-mangotango.css b/llm/llama.cpp/examples/server/public/theme-mangotango.css
new file mode 100644
index 0000000000000000000000000000000000000000..e433802453b9d8435164aa033493e9866113d798
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/theme-mangotango.css
@@ -0,0 +1,216 @@
+/* Author: Yazan Agha-Schrader */
+/* Inspiration from llama.cpp logo/banner https://github.com/ggerganov/llama.cpp#readme */
+
+.theme-mangotango {
+
+--primary-color-1:      hsl(192, 8.5%, 11.6%);
+--primary-color-2:      hsl(192, 8.5%, 21%);
+--primary-color-3:      hsl(192, 8.5%, 30%);
+--primary-color-4:      hsl(192, 8.5%, 40%);
+
+--secondary-color-1:    hsl(192, 8.5%, 80%);
+--secondary-color-2:    hsl(192, 8.5%, 73%);
+--secondary-color-3:    hsl(192, 8.5%, 66%);
+--secondary-color-4:    hsl(192, 8.5%, 60%);
+
+--theme-nuance-color-1: hsl(23.1, 100%, 60.2%);
+--theme-nuance-color-2: hsl(23.1, 100%, 60.2%);
+--theme-nuance-color-3: hsl(23.1, 100%, 60.2%);
+--theme-nuance-color-4: hsl(23.1, 100%, 60.2%);
+
+
+
+/* ---------- PRIMARY COLORS ----------------- */
+--primary-color-1: hsl(192, 8.5%, 11.6%);
+    --primary-color-1-saturation: 8.5%;
+    --primary-color-1-lightness: 11.6%;
+
+--primary-color-2: hsl(192, 8.5%, 21%);
+    --primary-color-2-saturation: 8.5%;
+    --primary-color-2-lightness: 21%;
+
+--primary-color-3: hsl(192, 8.5%, 30%);
+    --primary-color-3-saturation: 8.5%;
+    --primary-color-3-lightness: 30%;
+
+--primary-color-4: hsl(192, 8.5%, 40%);
+    --primary-color-4-saturation: 8.5%;
+    --primary-color-4-lightness: 40%;
+
+
+
+/* ---------- SECONDARY COLORS --------------- */
+--secondary-color-1: hsl(192, 8.5%, 80%);
+    --secondary-color-1-saturation: 8.5%;
+    --secondary-color-1-lightness: 80%;
+
+--secondary-color-2: hsl(192, 8.5%, 73%);
+    --secondary-color-2-saturation: 8.5%;
+    --secondary-color-2-lightness: 73%;
+
+--secondary-color-3: hsl(192, 8.5%, 66%);
+    --secondary-color-3-saturation: 8.5%;
+    --secondary-color-3-lightness: 66%;
+
+--secondary-color-4: hsl(192, 8.5%, 60%);
+    --secondary-color-4-saturation: 8.5%;
+    --secondary-color-4-lightness: 60%;
+
+
+
+/* ----------- NUANCES COLORS ---------------- */
+--theme-nuance-color-1: hsl(23.1, 100%, 60.2%);
+    --theme-nuance-color-1-saturation: 100%;
+    --theme-nuance-color-1-lightness: 60.2%;
+
+--theme-nuance-color-2: hsl(23.1, 100%, 60.2%);
+    --theme-nuance-color-2-saturation: 100%;
+    --theme-nuance-color-2-lightness: 60.2%;
+
+--theme-nuance-color-3: hsl(23.1, 100%, 60.2%);
+    --theme-nuance-color-3-saturation: 100%;
+    --theme-nuance-color-3-lightness: 60.2%;
+
+--theme-nuance-color-4: hsl(23.1, 100%, 60.2%);
+    --theme-nuance-color-4-saturation: 100%;
+    --theme-nuance-color-4-lightness: 60.2%;
+
+
+
+/* ----------- ROYGP COLORS ------------------ */
+    --theme-red-color:     hsl(325, 60%, 50%);
+    --theme-orange-color:  #e76f51;
+    --theme-yellow-color:  #ffd95f;
+    --theme-green-color:   #A3BE8C;
+    --theme-blue-color:    hsl(192, 95%, 40%);
+    --theme-purple-color:  hsl(192, 80%, 35%);
+
+
+
+/* ------------------------------------------- */
+--background-color-1:    var(--primary-color-1);
+--background-color-2:    var(--primary-color-2);
+--background-color-3:    var(--primary-color-3);
+--background-color-4:    var(--primary-color-4);
+
+--border-color-1:        var(--primary-color-2);
+--border-color-2:        var(--primary-color-3);
+--border-color-3:        var(--primary-color-4);
+
+--border-focus-color:    var(--theme-nuance-color-2);
+--border-focus-shadow:   var(--theme-nuance-color-1);
+
+--text-color-plain:      var(--secondary-color-1);
+--text-color-subtile-1:  var(--secondary-color-2);
+--text-color-subtile-2:  var(--secondary-color-3);
+
+--code-background-color: var(--secondary-color-2);
+--code-text-color:       var(--primary-color-2);
+
+--ui-range-thumb-color:  var(--theme-nuance-color-3);
+--ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+--textarea-border-color: var(--secondary-color-4);
+
+--chat-id-color:         var(--theme-nuance-color-4);
+
+
+
+/* ------------------------------------------- */
+--button-alert-text-hover:       var(--secondary-color-1);
+--button-alert-color-hover:      var(--theme-purple-color);
+--button-alert-border-hover:     var(--theme-purple-color);
+
+--button-alert-text-active:      var(--secondary-color-1);
+--button-alert-color-active:     var(--theme-blue-color);
+--button-alert-border-active:    var(--theme-blue-color);
+
+
+
+/* ----------- PRIMARY BUTTONS --------------- */
+/* - button should immediately catch the eye - */
+--button-primary-text: var(--primary-color-1);
+--button-primary-color:  var(--theme-nuance-color-3);
+--button-primary-border: var(--theme-nuance-color-3);
+
+
+/* ---------hover---------- */
+--button-primary-text-hover:
+    hsl(192,
+    calc(var(--primary-color-1-saturation) - 100%),
+    calc(var(--primary-color-1-lightness)  + 100%));
+
+--button-primary-color-hover:
+    hsl(23.1,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+--button-primary-border-hover:
+    hsl(23.1,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+
+/* ---------active--------- */
+--button-primary-text-active:
+    hsl(23.1,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  + 100%));
+
+--button-primary-color-active:
+    hsl(23.1,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 15%));
+
+--button-primary-border-active:
+    hsl(23.1,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+
+
+/* ---------- SECONDARY BUTTONS -------------- */
+/* these should NOT immediately catch the eye  */
+--button-secondary-text:   var(--secondary-color-1);
+--button-secondary-color:  var(--primary-color-3);
+--button-secondary-border: var(--primary-color-3);
+
+
+/* ---------hover---------- */
+--button-secondary-text-hover:
+    hsl(23.1,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-hover:  var(--primary-color-4);
+--button-secondary-border-hover: var(--primary-color-4);
+
+
+/* ---------active--------- */
+--button-secondary-text-active: var(--secondary-color-1);
+
+--button-secondary-color-active:
+    hsl(192,
+    calc(var(--primary-color-4-saturation) - 30%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+--button-secondary-border-active:
+    hsl(192,
+    calc(var(--primary-color-4-saturation) - 30%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+
+
+/* ---------- TERTIARY BUTTONS --------------- */
+/* ---------- disabled buttons --------------- */
+--button-tertiary-text:   var(--primary-color-4);
+--button-tertiary-color:  var(--primary-color-2);
+--button-tertiary-border: var(--primary-color-2);
+
+
+/* ---------hover---------- */
+--button-tertiary-text:   var(--primary-color-4);
+--button-tertiary-color:  var(--primary-color-2);
+--button-tertiary-border: var(--primary-color-2);
+
+}
diff --git a/llm/llama.cpp/examples/server/public/theme-playground.css b/llm/llama.cpp/examples/server/public/theme-playground.css
new file mode 100644
index 0000000000000000000000000000000000000000..9d56a718248a46c227c2e47daa51967df142f35e
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/theme-playground.css
@@ -0,0 +1,221 @@
+/* Author: Yazan Agha-Schrader */
+/* Inspiration from OpenAI's Playground platform https://platform.openai.com/playground/ */
+
+.theme-playground {
+
+/* ---------- PRIMARY COLORS ----------------- */
+--primary-color-1: hsl(0, 0%,    99.2%);
+    --primary-color-1-hue:         0;
+    --primary-color-1-saturation:  0%;
+    --primary-color-1-lightness:   99.2%;
+
+--primary-color-2: hsl(0, 0%,    95%);
+    --primary-color-2-hue:         0;
+    --primary-color-2-saturation:  0%;
+    --primary-color-2-lightness:   95%;
+
+--primary-color-3: hsl(0, 0%,    88%);
+    --primary-color-3-hue:         0;
+    --primary-color-3-saturation:  0%;
+    --primary-color-3-lightness:   88%;
+
+--primary-color-4: hsl(0, 0%,    80%);
+    --primary-color-4-hue:         0;
+    --primary-color-4-saturation:  0%;
+    --primary-color-4-lightness:   80%;
+
+
+
+/* ---------- SECONDARY COLORS --------------- */
+--secondary-color-1: hsl(0, 0%,    20%);
+    --secondary-color-1-hue:         0;
+    --secondary-color-1-saturation:  0%;
+    --secondary-color-1-lightness:   20%;
+
+--secondary-color-2: hsl(0, 0%,    23.1%);
+    --secondary-color-2-hue:         0;
+    --secondary-color-2-saturation:  0%;
+    --secondary-color-2-lightness:   23.1%;
+
+--secondary-color-3: hsl(0, 0%,    29%);
+    --secondary-color-3-hue:         0;
+    --secondary-color-3-saturation:  0%;
+    --secondary-color-3-lightness:   29%;
+
+--secondary-color-4: hsl(0, 0%,    36.1%);
+    --secondary-color-4-hue:         0;
+    --secondary-color-4-saturation:  0%;
+    --secondary-color-4-lightness:   36.1%;
+
+
+
+/* ----------- NUANCES COLORS ---------------- */
+--theme-nuance-color-1: hsl(165.2, 82.1%, 35.1%);
+    --theme-nuance-color-1-hue:             165.2;
+    --theme-nuance-color-1-saturation:      82.1%;
+    --theme-nuance-color-1-lightness:       35.1%;
+
+--theme-nuance-color-2: hsl(165.2, 82.1%, 35.1%);
+    --theme-nuance-color-2-hue:             165.2;
+    --theme-nuance-color-2-saturation:      82.1%;
+    --theme-nuance-color-2-lightness:       35.1%;
+
+--theme-nuance-color-3: hsl(165.2, 81.1%, 35.3%);
+    --theme-nuance-color-3-hue:             165.2;
+    --theme-nuance-color-3-saturation:      81.1%;
+    --theme-nuance-color-3-lightness:       35.3%;
+
+--theme-nuance-color-4: hsl(164.9, 81.6%, 27.6%);
+    --theme-nuance-color-4-hue:             164.9;
+    --theme-nuance-color-4-saturation:      81.6%;
+    --theme-nuance-color-4-lightness:       27.6%;
+
+
+
+/* ----------- ROYGP COLORS ------------------ */
+--theme-red-color:     hsl(0.3, 80%, 50%);
+--theme-orange-color:  #e76f51;
+--theme-yellow-color:  hsl(60, 70.6%, 73.3%);
+--theme-green-color:   #A3BE8C;
+--theme-purple-color:  hsl(0.3, 70%, 45%);
+
+
+
+/* ------------------------------------------- */
+--background-color-1:    var(--primary-color-1);
+--background-color-2:    var(--primary-color-2);
+--background-color-3:    var(--primary-color-3);
+--background-color-4:    var(--primary-color-4);
+
+--border-color-1:        var(--primary-color-2);
+--border-color-2:        var(--primary-color-3);
+--border-color-3:        var(--primary-color-4);
+
+--border-focus-color:    var(--theme-nuance-color-2);
+--border-focus-shadow:   var(--theme-nuance-color-1);
+
+--text-color-plain:      var(--secondary-color-1);
+--text-color-subtile-1:  var(--secondary-color-2);
+--text-color-subtile-2:  var(--secondary-color-3);
+
+--code-background-color: var(--secondary-color-2);
+--code-text-color:       var(--primary-color-2);
+
+--ui-range-thumb-color:  var(--primary-color-4);
+--ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+--textarea-border-color: var(--secondary-color-4);
+
+--chat-id-color:        var(--theme-nuance-color-4);
+
+
+
+/* ------------------------------------------- */
+--button-alert-text-hover:       var(--primary-color-1);
+--button-alert-color-hover:      var(--theme-purple-color);
+--button-alert-border-hover:     var(--theme-purple-color);
+
+--button-alert-text-active:      var(--primary-color-1);
+--button-alert-color-active:     var(--theme-red-color);
+--button-alert-border-active:    var(--theme-red-color);
+
+
+
+/* ----------- PRIMARY BUTTONS --------------- */
+/* - button should immediately catch the eye - */
+--button-primary-text:
+    hsl(0,
+    calc(var(--primary-color-1-saturation) - 100%),
+    calc(var(--primary-color-1-lightness)  + 100%));
+
+--button-primary-color:  var(--theme-nuance-color-3);
+--button-primary-border: var(--theme-nuance-color-3);
+
+
+/* ---------hover---------- */
+--button-primary-text-hover:
+    hsl(0,
+    calc(var(--primary-color-1-saturation) - 100%),
+    calc(var(--primary-color-1-lightness)  + 100%));
+
+--button-primary-color-hover:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+--button-primary-border-hover:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+
+/* ---------active--------- */
+--button-primary-text-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 100%),
+    calc(var(--theme-nuance-color-3-lightness)  + 100%));
+
+--button-primary-color-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 15%));
+
+--button-primary-border-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 2%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+
+
+/* ---------- SECONDARY BUTTONS -------------- */
+/* these should NOT immediately catch the eye  */
+--button-secondary-text:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 50%));
+
+--button-secondary-color:  var(--primary-color-3);
+--button-secondary-border: var(--primary-color-3);
+
+
+/* ---------hover---------- */
+--button-secondary-text-hover:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-hover:  var(--primary-color-4);
+--button-secondary-border-hover: var(--primary-color-4);
+
+
+/* ---------active--------- */
+--button-secondary-text-active:
+    hsl(165.2,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-active:
+    hsl(0,
+    calc(var(--primary-color-4-saturation) - 30%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+--button-secondary-border-active:
+    hsl(0,
+    calc(var(--primary-color-4-saturation) - 30%),
+    calc(var(--primary-color-4-lightness)  - 15%));
+
+
+
+/* ---------- TERTIARY BUTTONS --------------- */
+/* ---------- disabled buttons --------------- */
+--button-tertiary-text:   var(--primary-color-4);
+--button-tertiary-color:  var(--primary-color-2);
+--button-tertiary-border: var(--primary-color-2);
+
+
+/* ---------hover---------- */
+--button-tertiary-text:   var(--primary-color-4);
+--button-tertiary-color:  var(--primary-color-2);
+--button-tertiary-border: var(--primary-color-2);
+
+}
diff --git a/llm/llama.cpp/examples/server/public/theme-polarnight.css b/llm/llama.cpp/examples/server/public/theme-polarnight.css
new file mode 100644
index 0000000000000000000000000000000000000000..2bcfb33d8f1836244980c280dedacd0dc2d70cb9
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/theme-polarnight.css
@@ -0,0 +1,253 @@
+/* Author: Yazan Agha-Schrader */
+/* Inspiration from Nord Theme https://www.nordtheme.com/docs/colors-and-palettes */
+
+.theme-polarnight {
+
+/* ---------- PRIMARY COLORS ----------------- */
+--primary-color-1: hsl(220.0, 16.4%, 21.6%) ;
+    --primary-color-1-hue:             220.0;
+    --primary-color-1-saturation:      16.4%;
+    --primary-color-1-lightness:       21.6%;
+
+--primary-color-2: hsl(221.7, 16.3%, 27.6%) ;
+    -primary-color-2-hue:              221.7;
+    --primary-color-2-saturation:      16.3%;
+    --primary-color-2-lightness:       27.6%;
+
+--primary-color-3: hsl(220.0, 16.8%, 31.6%) ;
+    --primary-color-3-hue:             220.0;
+    --primary-color-3-saturation:      16.8%;
+    --primary-color-3-lightness:       31.6%;
+
+--primary-color-4: hsl(220.0, 16.5%, 35.7%);
+    --primary-color-4-hue:             220.0;
+    --primary-color-4-saturation:      16.5%;
+    --primary-color-4-lightness:       35.7%;
+
+
+
+/* ---------- SECONDARY COLORS --------------- */
+--secondary-color-1: hsl(217.5, 26.7%, 94.1%);
+    --secondary-color-1-hue:             217.5;
+    --secondary-color-1-saturation:      26.7%;
+    --secondary-color-1-lightness:       94.1%;
+
+--secondary-color-2: hsl(218.2, 26.8%, 92.0%);
+    --secondary-color-2-hue:             218.2;
+    --secondary-color-2-saturation:      26.8%;
+    --secondary-color-2-lightness:       92.0%;
+
+--secondary-color-3: hsl(218.8, 27.9%, 88.0%);
+    --secondary-color-3-hue:             218.8;
+    --secondary-color-3-saturation:      27.9%;
+    --secondary-color-3-lightness:       88.0%;
+
+--secondary-color-4: hsl(218.8, 18.3%, 81.8%);
+    --secondary-color-4-hue:             218.8;
+    --secondary-color-4-saturation:      18.3%;
+    --secondary-color-4-lightness:       81.8%;
+
+
+
+/* ----------- NUANCES COLORS ---------------- */
+--theme-nuance-color-1: hsl(178.7, 25.1%, 64.9%);
+    --theme-nuance-color-1-hue:             178.7;
+    --theme-nuance-color-1-saturation:      25.1%;
+    --theme-nuance-color-1-lightness:       64.9%;
+
+--theme-nuance-color-2: hsl(193.3, 43.4%, 67.5%);
+    --theme-nuance-color-2-hue:             193.3;
+    --theme-nuance-color-2-saturation:      43.4%;
+    --theme-nuance-color-2-lightness:       67.5%;
+
+--theme-nuance-color-3: hsl(210.0, 34.0%, 63.1%);
+    --theme-nuance-color-3-hue:             210.0;
+    --theme-nuance-color-3-saturation:      34.0%;
+    --theme-nuance-color-3-lightness:       63.1%;
+
+--theme-nuance-color-4: hsl(213.1, 32.0%, 52.2%);
+    --theme-nuance-color-4-hue:             213.1;
+    --theme-nuance-color-4-saturation:      32.0%;
+    --theme-nuance-color-4-lightness:       52.2%;
+
+
+
+/* ----------- ROYGP COLORS ------------------ */
+--theme-red-color:    hsl(354.3, 42.3%, 56.5%);
+--theme-orange-color: hsl(20, 85%, 50%);
+--theme-yellow-color: hsl(20, 75%, 45%);
+--theme-green-color:  hsl( 92.4, 27.8%, 64.7%);
+--theme-purple-color: hsl(311.1, 20.2%, 63.1%);
+
+
+
+/* ------------------------------------------------ */
+--background-color-1:    var(--primary-color-1);
+--background-color-2:    var(--primary-color-2);
+--background-color-3:    var(--primary-color-3);
+--background-color-4:    var(--primary-color-4);
+
+--border-color-1:        var(--primary-color-2);
+--border-color-2:        var(--primary-color-3);
+--border-color-3:        var(--primary-color-4);
+
+--border-focus-color:    var(--theme-nuance-color-2);
+--border-focus-shadow:   var(--theme-nuance-color-1);
+
+--text-color-plain:      var(--secondary-color-1);
+--text-color-subtile-1:  var(--secondary-color-2);
+--text-color-subtile-2:  var(--secondary-color-3);
+
+--code-background-color: var(--secondary-color-2);
+--code-text-color:       var(--primary-color-2);
+
+--ui-range-thumb-color:  var(--theme-nuance-color-3);
+--ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+--textarea-border-color: var(--secondary-color-4);
+
+--chat-id-color:        var(--theme-nuance-color-4);
+
+
+
+/* ------------------------------------------- */
+--button-alert-text-hover:       var(--secondary-color-1);
+--button-alert-color-hover:      var(--theme-yellow-color);
+--button-alert-border-hover:     var(--theme-yellow-color);
+
+--button-alert-text-active:      var(--secondary-color-1);
+--button-alert-color-active:     var(--theme-orange-color);
+--button-alert-border-active:    var(--theme-orange-color);
+
+
+
+/* ----------- PRIMARY BUTTONS --------------- */
+/* - button should immediately catch the eye - */
+--button-primary-text:   var(--secondary-color-1);
+--button-primary-color:  var(--theme-nuance-color-3);
+--button-primary-border: var(--theme-nuance-color-3);
+
+
+/* ---------hover---------- */
+--button-primary-text-hover:
+    hsl(217.5,
+    calc(var(--secondary-color-1-saturation) - 35%),
+    calc(var(--secondary-color-1-lightness)  + 30%));
+
+--button-primary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+--button-primary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+
+/* ---------active--------- */
+--button-primary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 35%));
+
+--button-primary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+--button-primary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+
+
+/* ---------- SECONDARY BUTTONS -------------- */
+/* these should NOT immediately catch the eye  */
+--button-secondary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 50%));
+
+--button-secondary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+--button-secondary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+
+/* ---------hover---------- */
+--button-secondary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+--button-secondary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+
+/* ---------active--------- */
+--button-secondary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 25%));
+
+--button-secondary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  - 15%));
+
+--button-secondary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  - 15%));
+
+
+
+/* ---------- TERTIARY BUTTONS --------------- */
+/* ---------- disabled buttons --------------- */
+--button-tertiary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+
+/* ---------hover---------- */
+--button-tertiary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+}
diff --git a/llm/llama.cpp/examples/server/public/theme-snowstorm.css b/llm/llama.cpp/examples/server/public/theme-snowstorm.css
new file mode 100644
index 0000000000000000000000000000000000000000..7bb2275949273a05c4a99bae95ef5975af670e9f
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public/theme-snowstorm.css
@@ -0,0 +1,251 @@
+/* Author: Yazan Agha-Schrader */
+/* Inspiration from Nord Theme https://www.nordtheme.com/docs/colors-and-palettes */
+
+.theme-snowstorm {
+
+/* ---------- PRIMARY COLORS ----------------- */
+--primary-color-1: hsl(217.5, 26.7%, 94.1%);
+    --primary-color-1-hue:             217.5;
+    --primary-color-1-saturation:      26.7%;
+    --primary-color-1-lightness:       94.1%;
+
+--primary-color-2: hsl(218.2, 26.8%, 92.0%);
+    --primary-color-2-hue:             218.2;
+    --primary-color-2-saturation:      26.8%;
+    --primary-color-2-lightness:       92.0%;
+
+--primary-color-3: hsl(218.8, 27.9%, 88.0%);
+    --primary-color-3-hue:             218.8;
+    --primary-color-3-saturation:      27.9%;
+    --primary-color-3-lightness:       88.0%;
+
+--primary-color-4: hsl(218.8, 18.3%, 81.8%);
+    --primary-color-4-hue:             218.8;
+    --primary-color-4-saturation:      18.3%;
+    --primary-color-4-lightness:       81.8%;
+
+
+/* ---------- SECONDARY COLORS --------------- */
+--secondary-color-1: hsl(220.0, 16.4%, 21.6%);
+    --secondary-color-1-hue:             220.0;
+    --secondary-color-1-saturation:      16.4%;
+    --secondary-color-1-lightness:       21.6%;
+
+--secondary-color-2: hsl(221.7, 16.3%, 27.6%);
+    --secondary-color-2-hue:             221.7;
+    --secondary-color-2-saturation:      16.3%;
+    --secondary-color-2-lightness:       27.6%;
+
+--secondary-color-3: hsl(220.0, 16.8%, 31.6%);
+    --secondary-color-3-hue:             220.0;
+    --secondary-color-3-saturation:      16.8%;
+    --secondary-color-3-lightness:       31.6%;
+
+--secondary-color-4: hsl(220.0, 16.5%, 35.7%);
+    --secondary-color-4-hue:             220.0;
+    --secondary-color-4-saturation:      16.5%;
+    --secondary-color-4-lightness:       35.7%;
+
+
+
+/* ----------- NUANCES COLORS ---------------- */
+--theme-nuance-color-1: hsl(178.7, 25.1%, 64.9%);
+    --theme-nuance-color-1-hue:             178.7;
+    --theme-nuance-color-1-saturation:      25.1%;
+    --theme-nuance-color-1-lightness:       64.9%;
+
+--theme-nuance-color-2: hsl(193.3, 43.4%, 67.5%);
+    --theme-nuance-color-2-hue:             193.3;
+    --theme-nuance-color-2-saturation:      43.4%;
+    --theme-nuance-color-2-lightness:       67.5%;
+
+--theme-nuance-color-3: hsl(210.0, 34.0%, 63.1%);
+    --theme-nuance-color-3-hue:             210.0;
+    --theme-nuance-color-3-saturation:      34.0%;
+    --theme-nuance-color-3-lightness:       63.1%;
+
+--theme-nuance-color-4: hsl(213.1, 32.0%, 52.2%);
+    --theme-nuance-color-4-hue:             213.1;
+    --theme-nuance-color-4-saturation:      32.0%;
+    --theme-nuance-color-4-lightness:       52.2%;
+
+
+
+/* ----------- ROYGP COLORS ------------------ */
+--theme-red-color:    hsl(32.5, 80%, 50%);
+--theme-orange-color: hsl(32.5, 70%, 45%);
+--theme-yellow-color: hsl(40.0,   0.6%, 73.3%);
+--theme-green-color:  hsl(92.4,  27.8%, 64.7%);
+--theme-purple-color: hsl(311.1, 20.2%, 63.1%);
+
+
+
+/* ------------------------------------------- */
+--background-color-1:    var(--primary-color-1);
+--background-color-2:    var(--primary-color-2);
+--background-color-3:    var(--primary-color-3);
+--background-color-4:    var(--primary-color-4);
+
+--border-color-1:        var(--primary-color-2);
+--border-color-2:        var(--primary-color-3);
+--border-color-3:        var(--primary-color-4);
+
+--border-focus-color:    var(--theme-nuance-color-2);
+--border-focus-shadow:   var(--theme-nuance-color-1);
+
+--text-color-plain:      var(--secondary-color-1);
+--text-color-subtile-1:  var(--secondary-color-2);
+--text-color-subtile-2:  var(--secondary-color-3);
+
+--code-background-color: var(--secondary-color-2);
+--code-text-color:       var(--primary-color-2);
+
+--ui-range-thumb-color:  var(--theme-nuance-color-3);
+--ui-range-thumb-border: var(--ui-ranger-thumb-color);
+
+--textarea-border-color: var(--secondary-color-4);
+
+--chat-id-color:         var(--theme-nuance-color-4);
+
+
+
+/* ------------------------------------------- */
+--button-alert-text-hover:       var(--primary-color-1);
+--button-alert-color-hover:      var(--theme-orange-color);
+--button-alert-border-hover:     var(--theme-orange-color);
+
+--button-alert-text-active:      var(--primary-color-1);
+--button-alert-color-active:     var(--theme-red-color);
+--button-alert-border-active:    var(--theme-red-color);
+
+
+
+/* ----------- PRIMARY BUTTONS --------------- */
+/* - button should immediately catch the eye - */
+--button-primary-text:   var(--secondary-color-1);
+--button-primary-color:  var(--theme-nuance-color-3);
+--button-primary-border: var(--theme-nuance-color-3);
+
+
+/* ---------hover---------- */
+--button-primary-text-hover:
+    hsl(217.5,
+    calc(var(--secondary-color-1-saturation) + 35%),
+    calc(var(--secondary-color-1-lightness)  - 30%));
+
+--button-primary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+--button-primary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) -  2%),
+    calc(var(--theme-nuance-color-3-lightness)  - 10%));
+
+
+/* ---------active--------- */
+--button-primary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 35%));
+
+--button-primary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+--button-primary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 10%),
+    calc(var(--theme-nuance-color-3-lightness)  - 25%));
+
+
+
+/* ---------- SECONDARY BUTTONS -------------- */
+/* these should NOT immediately catch the eye  */
+--button-secondary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 50%));
+
+--button-secondary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+--button-secondary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  + 10%));
+
+
+/* ---------hover---------- */
+--button-secondary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 20%),
+    calc(var(--theme-nuance-color-3-lightness)  - 80%));
+
+--button-secondary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+--button-secondary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 22%),
+    calc(var(--theme-nuance-color-3-lightness)  +  1%));
+
+
+/* ---------active--------- */
+--button-secondary-text-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) + 40%),
+    calc(var(--theme-nuance-color-3-lightness)  - 55%));
+
+--button-secondary-color-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-secondary-border-active:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 30%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+
+
+/* ---------- TERTIARY BUTTONS --------------- */
+/* ---------- disabled buttons --------------- */
+--button-tertiary-text:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+/* ---------hover---------- */
+--button-tertiary-text-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  -  5%));
+
+--button-tertiary-color-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+--button-tertiary-border-hover:
+    hsl(210,
+    calc(var(--theme-nuance-color-3-saturation) - 40%),
+    calc(var(--theme-nuance-color-3-lightness)  + 20%));
+
+}
diff --git a/llm/llama.cpp/examples/server/public_simplechat/datautils.mjs b/llm/llama.cpp/examples/server/public_simplechat/datautils.mjs
new file mode 100644
index 0000000000000000000000000000000000000000..75159d6b1676bfeb0ab01014f5d04a8a1f5f0500
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public_simplechat/datautils.mjs
@@ -0,0 +1,266 @@
+//@ts-check
+// Helpers to work with different data types
+// by Humans for All
+//
+
+/**
+ * Given the limited context size of local LLMs and , many a times when context gets filled
+ * between the prompt and the response, it can lead to repeating text garbage generation.
+ * And many a times setting penalty wrt repeatation leads to over-intelligent garbage
+ * repeatation with slight variations. These garbage inturn can lead to overloading of the
+ * available model context, leading to less valuable response for subsequent prompts/queries,
+ * if chat history is sent to ai model.
+ *
+ * So two simple minded garbage trimming logics are experimented below.
+ * * one based on progressively-larger-substring-based-repeat-matching-with-partial-skip and
+ * * another based on char-histogram-driven garbage trimming.
+ *   * in future characteristic of histogram over varying lengths could be used to allow for
+ *     a more aggressive and adaptive trimming logic.
+ */
+
+
+/**
+ * Simple minded logic to help remove repeating garbage at end of the string.
+ * The repeatation needs to be perfectly matching.
+ *
+ * The logic progressively goes on probing for longer and longer substring based
+ * repeatation, till there is no longer repeatation. Inturn picks the one with
+ * the longest chain.
+ *
+ * @param {string} sIn
+ * @param {number} maxSubL
+ * @param {number} maxMatchLenThreshold
+ */
+export function trim_repeat_garbage_at_end(sIn, maxSubL=10, maxMatchLenThreshold=40) {
+    let rCnt = [0];
+    let maxMatchLen = maxSubL;
+    let iMML = -1;
+    for(let subL=1; subL < maxSubL; subL++) {
+        rCnt.push(0);
+        let i;
+        let refS = sIn.substring(sIn.length-subL, sIn.length);
+        for(i=sIn.length; i > 0; i -= subL) {
+            let curS = sIn.substring(i-subL, i);
+            if (refS != curS) {
+                let curMatchLen = rCnt[subL]*subL;
+                if (maxMatchLen < curMatchLen) {
+                    maxMatchLen = curMatchLen;
+                    iMML = subL;
+                }
+                break;
+            }
+            rCnt[subL] += 1;
+        }
+    }
+    console.debug("DBUG:DU:TrimRepeatGarbage:", rCnt);
+    if ((iMML == -1) || (maxMatchLen < maxMatchLenThreshold)) {
+        return {trimmed: false, data: sIn};
+    }
+    console.debug("DBUG:TrimRepeatGarbage:TrimmedCharLen:", maxMatchLen);
+    let iEnd = sIn.length - maxMatchLen;
+    return { trimmed: true, data: sIn.substring(0, iEnd) };
+}
+
+
+/**
+ * Simple minded logic to help remove repeating garbage at end of the string, till it cant.
+ * If its not able to trim, then it will try to skip a char at end and then trim, a few times.
+ * This ensures that even if there are multiple runs of garbage with different patterns, the
+ * logic still tries to munch through them.
+ *
+ * @param {string} sIn
+ * @param {number} maxSubL
+ * @param {number | undefined} [maxMatchLenThreshold]
+ */
+export function trim_repeat_garbage_at_end_loop(sIn, maxSubL, maxMatchLenThreshold, skipMax=16) {
+    let sCur = sIn;
+    let sSaved = "";
+    let iTry = 0;
+    while(true) {
+        let got = trim_repeat_garbage_at_end(sCur, maxSubL, maxMatchLenThreshold);
+        if (got.trimmed != true) {
+            if (iTry == 0) {
+                sSaved = got.data;
+            }
+            iTry += 1;
+            if (iTry >= skipMax) {
+                return sSaved;
+            }
+            got.data = got.data.substring(0,got.data.length-1);
+        } else {
+            iTry = 0;
+        }
+        sCur = got.data;
+    }
+}
+
+
+/**
+ * A simple minded try trim garbage at end using histogram driven characteristics.
+ * There can be variation in the repeatations, as long as no new char props up.
+ *
+ * This tracks the chars and their frequency in a specified length of substring at the end
+ * and inturn checks if moving further into the generated text from the end remains within
+ * the same char subset or goes beyond it and based on that either trims the string at the
+ * end or not. This allows to filter garbage at the end, including even if there are certain
+ * kind of small variations in the repeated text wrt position of seen chars.
+ *
+ * Allow the garbage to contain upto maxUniq chars, but at the same time ensure that
+ * a given type of char ie numerals or alphabets or other types dont cross the specified
+ * maxType limit. This allows intermixed text garbage to be identified and trimmed.
+ *
+ * ALERT: This is not perfect and only provides a rough garbage identification logic.
+ * Also it currently only differentiates between character classes wrt english.
+ *
+ * @param {string} sIn
+ * @param {number} maxType
+ * @param {number} maxUniq
+ * @param {number} maxMatchLenThreshold
+ */
+export function trim_hist_garbage_at_end(sIn, maxType, maxUniq, maxMatchLenThreshold) {
+    if (sIn.length < maxMatchLenThreshold) {
+        return { trimmed: false, data: sIn };
+    }
+    let iAlp = 0;
+    let iNum = 0;
+    let iOth = 0;
+    // Learn
+    let hist = {};
+    let iUniq = 0;
+    for(let i=0; i= maxUniq) {
+                break;
+            }
+            hist[c] = 1;
+        }
+    }
+    console.debug("DBUG:TrimHistGarbage:", hist);
+    if ((iAlp > maxType) || (iNum > maxType) || (iOth > maxType)) {
+        return { trimmed: false, data: sIn };
+    }
+    // Catch and Trim
+    for(let i=0; i < sIn.length; i++) {
+        let c = sIn[sIn.length-1-i];
+        if (!(c in hist)) {
+            if (i < maxMatchLenThreshold) {
+                return { trimmed: false, data: sIn };
+            }
+            console.debug("DBUG:TrimHistGarbage:TrimmedCharLen:", i);
+            return { trimmed: true, data: sIn.substring(0, sIn.length-i+1) };
+        }
+    }
+    console.debug("DBUG:TrimHistGarbage:Trimmed fully");
+    return { trimmed: true, data: "" };
+}
+
+/**
+ * Keep trimming repeatedly using hist_garbage logic, till you no longer can.
+ * This ensures that even if there are multiple runs of garbage with different patterns,
+ * the logic still tries to munch through them.
+ *
+ * @param {any} sIn
+ * @param {number} maxType
+ * @param {number} maxUniq
+ * @param {number} maxMatchLenThreshold
+ */
+export function trim_hist_garbage_at_end_loop(sIn, maxType, maxUniq, maxMatchLenThreshold) {
+    let sCur = sIn;
+    while (true) {
+        let got = trim_hist_garbage_at_end(sCur, maxType, maxUniq, maxMatchLenThreshold);
+        if (!got.trimmed) {
+            return got.data;
+        }
+        sCur = got.data;
+    }
+}
+
+/**
+ * Try trim garbage at the end by using both the hist-driven-garbage-trimming as well as
+ * skip-a-bit-if-reqd-then-repeat-pattern-based-garbage-trimming, with blind retrying.
+ * @param {string} sIn
+ */
+export function trim_garbage_at_end(sIn) {
+    let sCur = sIn;
+    for(let i=0; i<2; i++) {
+        sCur = trim_hist_garbage_at_end_loop(sCur, 8, 24, 72);
+        sCur = trim_repeat_garbage_at_end_loop(sCur, 32, 72, 12);
+    }
+    return sCur;
+}
+
+
+/**
+ * NewLines array helper.
+ * Allow for maintaining a list of lines.
+ * Allow for a line to be builtup/appended part by part.
+ */
+export class NewLines {
+
+    constructor() {
+        /** @type {string[]} */
+        this.lines = [];
+    }
+
+    /**
+     * Extracts lines from the passed string and inturn either
+     * append to a previous partial line or add a new line.
+     * @param {string} sLines
+     */
+    add_append(sLines) {
+        let aLines = sLines.split("\n");
+        let lCnt = 0;
+        for(let line of aLines) {
+            lCnt += 1;
+            // Add back newline removed if any during split
+            if (lCnt < aLines.length) {
+                line += "\n";
+            } else {
+                if (sLines.endsWith("\n")) {
+                    line += "\n";
+                }
+            }
+            // Append if required
+            if (lCnt == 1) {
+                let lastLine = this.lines[this.lines.length-1];
+                if (lastLine != undefined) {
+                    if (!lastLine.endsWith("\n")) {
+                        this.lines[this.lines.length-1] += line;
+                        continue;
+                    }
+                }
+            }
+            // Add new line
+            this.lines.push(line);
+        }
+    }
+
+    /**
+     * Shift the oldest/earliest/0th line in the array. [Old-New|Earliest-Latest]
+     * Optionally control whether only full lines (ie those with newline at end) will be returned
+     * or will a partial line without a newline at end (can only be the last line) be returned.
+     * @param {boolean} bFullWithNewLineOnly
+     */
+    shift(bFullWithNewLineOnly=true) {
+        let line = this.lines[0];
+        if (line == undefined) {
+            return undefined;
+        }
+        if ((line[line.length-1] != "\n") && bFullWithNewLineOnly){
+            return undefined;
+        }
+        return this.lines.shift();
+    }
+
+}
diff --git a/llm/llama.cpp/examples/server/public_simplechat/index.html b/llm/llama.cpp/examples/server/public_simplechat/index.html
index 1a1a342089ba3e02c19c1383cf6c0da48de96e83..f6413016fcc53b0f45cbd536340f8e674848d39b 100644
--- a/llm/llama.cpp/examples/server/public_simplechat/index.html
+++ b/llm/llama.cpp/examples/server/public_simplechat/index.html
@@ -8,21 +8,23 @@
         
         
         
-        
+        
+        
         
     
     
         

 
-            

+            

                 
  SimpleChat  

-                

-                    
-                    
-                

+                
             

 
             

@@ -30,7 +32,7 @@
             

             

                 
-                
+                
             

 
             

@@ -40,7 +42,7 @@
 
             

             

-                
+                
                 
             

 
diff --git a/llm/llama.cpp/examples/server/public_simplechat/readme.md b/llm/llama.cpp/examples/server/public_simplechat/readme.md
index de0dfc99de805fc7ee7e207d9b6e3f72e73a4497..21410199f60169b5f8c696e246424c4f82cfd1b3 100644
--- a/llm/llama.cpp/examples/server/public_simplechat/readme.md
+++ b/llm/llama.cpp/examples/server/public_simplechat/readme.md
@@ -3,6 +3,13 @@
 
 by Humans for All.
 
+## quickstart
+
+To run from the build dir
+
+bin/llama-server -m path/model.gguf --path ../examples/server/public_simplechat
+
+Continue reading for the details.
 
 ## overview
 
@@ -11,18 +18,31 @@ in a simple way with minimal code from a common code base. Inturn additionally i
 multiple independent back and forth chatting to an extent, with the ai llm model at a basic level, with their
 own system prompts.
 
+This allows seeing the generated text / ai-model response in oneshot at the end, after it is fully generated,
+or potentially as it is being generated, in a streamed manner from the server/ai-model.
+
+![Chat and Settings screens](./simplechat_screens.webp "Chat and Settings screens")
+
+Auto saves the chat session locally as and when the chat is progressing and inturn at a later time when you
+open SimpleChat, option is provided to restore the old chat session, if a matching one exists.
+
 The UI follows a responsive web design so that the layout can adapt to available display space in a usable
 enough manner, in general.
 
 Allows developer/end-user to control some of the behaviour by updating gMe members from browser's devel-tool
-console.
+console. Parallely some of the directly useful to end-user settings can also be changed using the provided
+settings ui.
 
-NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
-culling of old messages from the chat by default. However by enabling the sliding window chat logic, a crude
-form of old messages culling can be achieved.
+NOTE: Current web service api doesnt expose the model context length directly, so client logic doesnt provide
+any adaptive culling of old messages nor of replacing them with summary of their content etal. However there
+is a optional sliding window based chat logic, which provides a simple minded culling of old messages from
+the chat history before sending to the ai model.
 
-NOTE: It doesnt set any parameters other than temperature and max_tokens for now. However if someone wants
-they can update the js file or equivalent member in gMe as needed.
+NOTE: Wrt options sent with the request, it mainly sets temperature, max_tokens and optionaly stream for now.
+However if someone wants they can update the js file or equivalent member in gMe as needed.
+
+NOTE: One may be able to use this to chat with openai api web-service /chat/completions endpoint, in a very
+limited / minimal way. One will need to set model, openai url and authorization bearer key in settings ui.
 
 
 ## usage
@@ -33,12 +53,12 @@ http module.
 
 ### running using examples/server
 
-bin/server -m path/model.gguf --path ../examples/server/public_simplechat [--port PORT]
+./llama-server -m path/model.gguf --path examples/server/public_simplechat [--port PORT]
 
 ### running using python3's server module
 
 first run examples/server
-* bin/server -m path/model.gguf
+* ./llama-server -m path/model.gguf
 
 next run this web front end in examples/server/public_simplechat
 * cd ../examples/server/public_simplechat
@@ -52,9 +72,15 @@ Open this simple web front end from your local browser
 
 Once inside
 
-* Select between chat and completion mode. By default it is set to chat mode.
+* If you want to, you can change many of the default global settings
+  * the base url (ie ip addr / domain name, port)
+  * chat (default) vs completion mode
+  * try trim garbage in response or not
+  * amount of chat history in the context sent to server/ai-model
+  * oneshot or streamed mode.
 
 * In completion mode
+  * one normally doesnt use a system prompt in completion mode.
   * logic by default doesnt insert any role specific "ROLE: " prefix wrt each role's message.
     If the model requires any prefix wrt user role messages, then the end user has to
     explicitly add the needed prefix, when they enter their chat message.
@@ -88,12 +114,16 @@ Once inside
 * Wait for the logic to communicate with the server and get the response.
   * the user is not allowed to enter any fresh query during this time.
   * the user input box will be disabled and a working message will be shown in it.
+  * if trim garbage is enabled, the logic will try to trim repeating text kind of garbage to some extent.
 
 * just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
 
 * Using NewChat one can start independent chat sessions.
   * two independent chat sessions are setup by default.
 
+* When you want to print, switching ChatHistoryInCtxt to Full and clicking on the chat session button of
+  interest, will display the full chat history till then wrt same, if you want full history for printing.
+
 
 ## Devel note
 
@@ -104,14 +134,31 @@ by developers who may not be from web frontend background (so inturn may not be
 end-use-specific-language-extensions driven flows) so that they can use it to explore/experiment things.
 
 And given that the idea is also to help explore/experiment for developers, some flexibility is provided
-to change behaviour easily using the devel-tools/console, for now. And skeletal logic has been implemented
-to explore some of the end points and ideas/implications around them.
+to change behaviour easily using the devel-tools/console or provided minimal settings ui (wrt few aspects).
+Skeletal logic has been implemented to explore some of the end points and ideas/implications around them.
 
 
 ### General
 
 Me/gMe consolidates the settings which control the behaviour into one object.
 One can see the current settings, as well as change/update them using browsers devel-tool/console.
+It is attached to the document object. Some of these can also be updated using the Settings UI.
+
+  baseURL - the domain-name/ip-address and inturn the port to send the request.
+
+  bStream - control between oneshot-at-end and live-stream-as-its-generated collating and showing
+  of the generated response.
+
+    the logic assumes that the text sent from the server follows utf-8 encoding.
+
+    in streaming mode - if there is any exception, the logic traps the same and tries to ensure
+    that text generated till then is not lost.
+
+      if a very long text is being generated, which leads to no user interaction for sometime and
+      inturn the machine goes into power saving mode or so, the platform may stop network connection,
+      leading to exception.
+
+  apiEP - select between /completions and /chat/completions endpoint provided by the server/ai-model.
 
   bCompletionFreshChatAlways - whether Completion mode collates complete/sliding-window history when
   communicating with the server or only sends the latest user query/message.
@@ -119,13 +166,40 @@ One can see the current settings, as well as change/update them using browsers d
   bCompletionInsertStandardRolePrefix - whether Completion mode inserts role related prefix wrt the
   messages that get inserted into prompt field wrt /Completion endpoint.
 
-  chatRequestOptions - maintains the list of options/fields to send along with chat request,
+  bTrimGarbage - whether garbage repeatation at the end of the generated ai response, should be
+  trimmed or left as is. If enabled, it will be trimmed so that it wont be sent back as part of
+  subsequent chat history. At the same time the actual trimmed text is shown to the user, once
+  when it was generated, so user can check if any useful info/data was there in the response.
+
+    One may be able to request the ai-model to continue (wrt the last response) (if chat-history
+    is enabled as part of the chat-history-in-context setting), and chances are the ai-model will
+    continue starting from the trimmed part, thus allows long response to be recovered/continued
+    indirectly, in many cases.
+
+    The histogram/freq based trimming logic is currently tuned for english language wrt its
+    is-it-a-alpabetic|numeral-char regex match logic.
+
+  apiRequestOptions - maintains the list of options/fields to send along with api request,
   irrespective of whether /chat/completions or /completions endpoint.
 
     If you want to add additional options/fields to send to the server/ai-model, and or
     modify the existing options value or remove them, for now you can update this global var
     using browser's development-tools/console.
 
+    For string, numeric and boolean fields in apiRequestOptions, including even those added by a
+    user at runtime by directly modifying gMe.apiRequestOptions, setting ui entries will be auto
+    created.
+
+    cache_prompt option supported by example/server is allowed to be controlled by user, so that
+    any caching supported wrt system-prompt and chat history, if usable can get used. When chat
+    history sliding window is enabled, cache_prompt logic may or may not kick in at the backend
+    wrt same, based on aspects related to model, positional encoding, attention mechanism etal.
+    However system prompt should ideally get the benefit of caching.
+
+  headers - maintains the list of http headers sent when request is made to the server. By default
+  Content-Type is set to application/json. Additionally Authorization entry is provided, which can
+  be set if needed using the settings ui.
+
   iRecentUserMsgCnt - a simple minded SlidingWindow to limit context window load at Ai Model end.
   This is disabled by default. However if enabled, then in addition to latest system message, only
   the last/latest iRecentUserMsgCnt user messages after the latest system prompt and its responses
@@ -138,9 +212,10 @@ One can see the current settings, as well as change/update them using browsers d
     >0 : Send the latest chat history from the latest system prompt, limited to specified cnt.
 
 
-By using gMe's iRecentUserMsgCnt and chatRequestOptions.max_tokens one can try to control the
-implications of loading of the ai-model's context window by chat history, wrt chat response to
-some extent in a simple crude way.
+By using gMe's iRecentUserMsgCnt and apiRequestOptions.max_tokens/n_predict one can try to control
+the implications of loading of the ai-model's context window by chat history, wrt chat response to
+some extent in a simple crude way. You may also want to control the context size enabled when the
+server loads ai-model, on the server end.
 
 
 Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
@@ -149,28 +224,15 @@ matter clearing site data, dont directly override site caching in all cases. Wor
 have to change port. Or in dev tools of browser, you may be able to disable caching fully.
 
 
-Concept of multiple chat sessions with different servers, as well as saving and restoring of
-those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
-its instances relatively easily, however given the current goal of keeping this simple, it has
-not been added, for now.
+Currently the server to communicate with is maintained globally and not as part of a specific
+chat session. So if one changes the server ip/url in setting, then all chat sessions will auto
+switch to this new server, when you try using those sessions.
 
 
 By switching between chat.add_system_begin/anytime, one can control whether one can change
 the system prompt, anytime during the conversation or only at the beginning.
 
 
-read_json_early, is to experiment with reading json response data early on, if available,
-so that user can be shown generated data, as and when it is being generated, rather than
-at the end when full data is available.
-
-  the server flow doesnt seem to be sending back data early, atleast for request (inc options)
-  that is currently sent.
-
-  if able to read json data early on in future, as and when ai model is generating data, then
-  this helper needs to indirectly update the chat div with the recieved data, without waiting
-  for the overall data to be available.
-
-
 ### Default setup
 
 By default things are setup to try and make the user experience a bit better, if possible.
@@ -179,7 +241,8 @@ However a developer when testing the server of ai-model may want to change these
 Using iRecentUserMsgCnt reduce chat history context sent to the server/ai-model to be
 just the system-prompt, prev-user-request-and-ai-response and cur-user-request, instead of
 full chat history. This way if there is any response with garbage/repeatation, it doesnt
-mess with things beyond the next question/request/query, in some ways.
+mess with things beyond the next question/request/query, in some ways. The trim garbage
+option also tries to help avoid issues with garbage in the context to an extent.
 
 Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat up the space
 available wrt next query-response. However dont forget that the server when started should
@@ -189,11 +252,33 @@ also be started with a model context size of 1k or more, to be on safe side.
   internal n_predict, for now add the same here on the client side, maybe later add max_tokens
   to /completions endpoint handling code on server side.
 
-Frequency and presence penalty fields are set to 1.2 in the set of fields sent to server
-along with the user query. So that the model is partly set to try avoid repeating text in
-its response.
+NOTE: One may want to experiment with frequency/presence penalty fields in apiRequestOptions
+wrt the set of fields sent to server along with the user query, to check how the model behaves
+wrt repeatations in general in the generated text response.
+
+A end-user can change these behaviour by editing gMe from browser's devel-tool/console or by
+using the provided settings ui (for settings exposed through the ui).
+
+
+### OpenAi / Equivalent API WebService
+
+One may be abe to handshake with OpenAI/Equivalent api web service's /chat/completions endpoint
+for a minimal chatting experimentation by setting the below.
+
+* the baseUrl in settings ui
+  * https://api.openai.com/v1 or similar
+
+* Wrt request body - gMe.apiRequestOptions
+  * model (settings ui)
+  * any additional fields if required in future
+
+* Wrt request headers - gMe.headers
+  * Authorization (available through settings ui)
+    * Bearer THE_OPENAI_API_KEY
+  * any additional optional header entries like "OpenAI-Organization", "OpenAI-Project" or so
 
-A end-user can change these behaviour by editing gMe from browser's devel-tool/console.
+NOTE: Not tested, as there is no free tier api testing available. However logically this might
+work.
 
 
 ## At the end
diff --git a/llm/llama.cpp/examples/server/public_simplechat/simplechat.css b/llm/llama.cpp/examples/server/public_simplechat/simplechat.css
index 20c738b12ed6f3035b38438d5569fed092e633db..13bfb80b48be80f2b4c3fdad88c9144dbcd40894 100644
--- a/llm/llama.cpp/examples/server/public_simplechat/simplechat.css
+++ b/llm/llama.cpp/examples/server/public_simplechat/simplechat.css
@@ -21,6 +21,17 @@
 .role-user {
     background-color: lightgray;
 }
+.role-trim {
+    background-color: lightpink;
+}
+
+.gridx2 {
+    display: grid;
+    grid-template-columns: repeat(2, 1fr);
+    border-bottom-style: dotted;
+    border-bottom-width: thin;
+    border-bottom-color: lightblue;
+}
 
 .flex-grow {
     flex-grow: 1;
diff --git a/llm/llama.cpp/examples/server/public_simplechat/simplechat.js b/llm/llama.cpp/examples/server/public_simplechat/simplechat.js
index 0c48da8796fd87f6bc548eee52691634fe2c9f68..8e0df3b61df2b1b279b6419583d2f3dfd6542b0c 100644
--- a/llm/llama.cpp/examples/server/public_simplechat/simplechat.js
+++ b/llm/llama.cpp/examples/server/public_simplechat/simplechat.js
@@ -2,6 +2,9 @@
 // A simple completions and chat/completions test related web front end logic
 // by Humans for All
 
+import * as du from "./datautils.mjs";
+import * as ui from "./ui.mjs"
+
 class Roles {
     static System = "system";
     static User = "user";
@@ -9,40 +12,65 @@ class Roles {
 }
 
 class ApiEP {
-    static Chat = "chat";
-    static Completion = "completion";
+    static Type = {
+        Chat: "chat",
+        Completion: "completion",
+    }
+    static UrlSuffix = {
+        'chat': `/chat/completions`,
+        'completion': `/completions`,
+    }
+
+    /**
+     * Build the url from given baseUrl and apiEp id.
+     * @param {string} baseUrl
+     * @param {string} apiEP
+     */
+    static Url(baseUrl, apiEP) {
+        if (baseUrl.endsWith("/")) {
+            baseUrl = baseUrl.substring(0, baseUrl.length-1);
+        }
+        return `${baseUrl}${this.UrlSuffix[apiEP]}`;
+    }
+
 }
 
+
 let gUsageMsg = `
     
Usage

     
    
-    
  •  Set system prompt above, to try control ai response charactersitic, if model supports same.
    • 
+    
  •  System prompt above, to try control ai response characteristics.
    • 
         
      
-        
    •  Completion mode normally wont have a system prompt.
      • 
+        
    •  Completion mode - no system prompt normally.
      • 
         
    
+    
  •  Use shift+enter for inserting enter/newline.
    • 
     
  •  Enter your query to ai assistant below.
    • 
-        
      
-        
    •  Completion mode doesnt insert user/role: prefix implicitly.
      • 
-        
    •  Use shift+enter for inserting enter/newline.
      • 
-        
    
     
  •  Default ContextWindow = [System, Last Query+Resp, Cur Query].
    • 
         
      
-        
    •  experiment iRecentUserMsgCnt, max_tokens, model ctxt window to expand
      • 
+        
    •  ChatHistInCtxt, MaxTokens, ModelCtxt window to expand
      • 
         
    
     

 `;
 
+
 /** @typedef {{role: string, content: string}[]} ChatMessages */
 
+/** @typedef {{iLastSys: number, xchat: ChatMessages}} SimpleChatODS */
+
 class SimpleChat {
 
-    constructor() {
+    /**
+     * @param {string} chatId
+     */
+    constructor(chatId) {
+        this.chatId = chatId;
         /**
          * Maintain in a form suitable for common LLM web service chat/completions' messages entry
          * @type {ChatMessages}
          */
         this.xchat = [];
         this.iLastSys = -1;
+        this.latestResponse = "";
     }
 
     clear() {
@@ -50,6 +78,27 @@ class SimpleChat {
         this.iLastSys = -1;
     }
 
+    ods_key() {
+        return `SimpleChat-${this.chatId}`
+    }
+
+    save() {
+        /** @type {SimpleChatODS} */
+        let ods = {iLastSys: this.iLastSys, xchat: this.xchat};
+        localStorage.setItem(this.ods_key(), JSON.stringify(ods));
+    }
+
+    load() {
+        let sods = localStorage.getItem(this.ods_key());
+        if (sods == null) {
+            return;
+        }
+        /** @type {SimpleChatODS} */
+        let ods = JSON.parse(sods);
+        this.iLastSys = ods.iLastSys;
+        this.xchat = ods.xchat;
+    }
+
     /**
      * Recent chat messages.
      * If iRecentUserMsgCnt < 0
@@ -94,6 +143,15 @@ class SimpleChat {
         return rchat;
     }
 
+    /**
+     * Collate the latest response from the server/ai-model, as it is becoming available.
+     * This is mainly useful for the stream mode.
+     * @param {string} content
+     */
+    append_response(content) {
+        this.latestResponse += content;
+    }
+
     /**
      * Add an entry into xchat
      * @param {string} role
@@ -107,6 +165,7 @@ class SimpleChat {
         if (role == Roles.System) {
             this.iLastSys = this.xchat.length - 1;
         }
+        this.save();
         return true;
     }
 
@@ -121,10 +180,8 @@ class SimpleChat {
         }
         let last = undefined;
         for(const x of this.recent_chat(gMe.iRecentUserMsgCnt)) {
-            let entry = document.createElement("p");
+            let entry = ui.el_create_append_p(`${x.role}: ${x.content}`, div);
             entry.className = `role-${x.role}`;
-            entry.innerText = `${x.role}: ${x.content}`;
-            div.appendChild(entry);
             last = entry;
         }
         if (last !== undefined) {
@@ -132,20 +189,44 @@ class SimpleChat {
         } else {
             if (bClear) {
                 div.innerHTML = gUsageMsg;
+                gMe.setup_load(div, this);
                 gMe.show_info(div);
             }
         }
+        return last;
+    }
+
+    /**
+     * Setup the fetch headers.
+     * It picks the headers from gMe.headers.
+     * It inserts Authorization only if its non-empty.
+     * @param {string} apiEP
+     */
+    fetch_headers(apiEP) {
+        let headers = new Headers();
+        for(let k in gMe.headers) {
+            let v = gMe.headers[k];
+            if ((k == "Authorization") && (v.trim() == "")) {
+                continue;
+            }
+            headers.append(k, v);
+        }
+        return headers;
     }
 
     /**
      * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
      * The needed fields/options are picked from a global object.
+     * Add optional stream flag, if required.
      * Convert the json into string.
      * @param {Object} obj
      */
-    request_jsonstr(obj) {
-        for(let k in gMe.chatRequestOptions) {
-            obj[k] = gMe.chatRequestOptions[k];
+    request_jsonstr_extend(obj) {
+        for(let k in gMe.apiRequestOptions) {
+            obj[k] = gMe.apiRequestOptions[k];
+        }
+        if (gMe.bStream) {
+            obj["stream"] = true;
         }
         return JSON.stringify(obj);
     }
@@ -157,7 +238,7 @@ class SimpleChat {
         let req = {
             messages: this.recent_chat(gMe.iRecentUserMsgCnt),
         }
-        return this.request_jsonstr(req);
+        return this.request_jsonstr_extend(req);
     }
 
     /**
@@ -180,7 +261,60 @@ class SimpleChat {
         let req = {
             prompt: prompt,
         }
-        return this.request_jsonstr(req);
+        return this.request_jsonstr_extend(req);
+    }
+
+    /**
+     * Return a string form of json object suitable for specified api endpoint.
+     * @param {string} apiEP
+     */
+    request_jsonstr(apiEP) {
+        if (apiEP == ApiEP.Type.Chat) {
+            return this.request_messages_jsonstr();
+        } else {
+            return this.request_prompt_jsonstr(gMe.bCompletionInsertStandardRolePrefix);
+        }
+    }
+
+    /**
+     * Extract the ai-model/assistant's response from the http response got.
+     * Optionally trim the message wrt any garbage at the end.
+     * @param {any} respBody
+     * @param {string} apiEP
+     */
+    response_extract(respBody, apiEP) {
+        let assistant = "";
+        if (apiEP == ApiEP.Type.Chat) {
+            assistant = respBody["choices"][0]["message"]["content"];
+        } else {
+            try {
+                assistant = respBody["choices"][0]["text"];
+            } catch {
+                assistant = respBody["content"];
+            }
+        }
+        return assistant;
+    }
+
+    /**
+     * Extract the ai-model/assistant's response from the http response got in streaming mode.
+     * @param {any} respBody
+     * @param {string} apiEP
+     */
+    response_extract_stream(respBody, apiEP) {
+        let assistant = "";
+        if (apiEP == ApiEP.Type.Chat) {
+            if (respBody["choices"][0]["finish_reason"] !== "stop") {
+                assistant = respBody["choices"][0]["delta"]["content"];
+            }
+        } else {
+            try {
+                assistant = respBody["choices"][0]["text"];
+            } catch {
+                assistant = respBody["content"];
+            }
+        }
+        return assistant;
     }
 
     /**
@@ -239,53 +373,99 @@ class SimpleChat {
         return sysPrompt;
     }
 
-}
-
 
-let gBaseURL = "http://127.0.0.1:8080";
-let gChatURL = {
-    'chat': `${gBaseURL}/chat/completions`,
-    'completion': `${gBaseURL}/completions`,
-}
+    /**
+     * Handle the multipart response from server/ai-model
+     * @param {Response} resp
+     * @param {string} apiEP
+     * @param {HTMLDivElement} elDiv
+     */
+    async handle_response_multipart(resp, apiEP, elDiv) {
+        let elP = ui.el_create_append_p("", elDiv);
+        if (!resp.body) {
+            throw Error("ERRR:SimpleChat:SC:HandleResponseMultiPart:No body...");
+        }
+        let tdUtf8 = new TextDecoder("utf-8");
+        let rr = resp.body.getReader();
+        this.latestResponse = "";
+        let xLines = new du.NewLines();
+        while(true) {
+            let { value: cur,  done: done } = await rr.read();
+            if (cur) {
+                let curBody = tdUtf8.decode(cur, {stream: true});
+                console.debug("DBUG:SC:PART:Str:", curBody);
+                xLines.add_append(curBody);
+            }
+            while(true) {
+                let curLine = xLines.shift(!done);
+                if (curLine == undefined) {
+                    break;
+                }
+                if (curLine.trim() == "") {
+                    continue;
+                }
+                if (curLine.startsWith("data:")) {
+                    curLine = curLine.substring(5);
+                }
+                let curJson = JSON.parse(curLine);
+                console.debug("DBUG:SC:PART:Json:", curJson);
+                this.append_response(this.response_extract_stream(curJson, apiEP));
+            }
+            elP.innerText = this.latestResponse;
+            elP.scrollIntoView(false);
+            if (done) {
+                break;
+            }
+        }
+        console.debug("DBUG:SC:PART:Full:", this.latestResponse);
+        return this.latestResponse;
+    }
 
+    /**
+     * Handle the oneshot response from server/ai-model
+     * @param {Response} resp
+     * @param {string} apiEP
+     */
+    async handle_response_oneshot(resp, apiEP) {
+        let respBody = await resp.json();
+        console.debug(`DBUG:SimpleChat:SC:${this.chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
+        return this.response_extract(respBody, apiEP);
+    }
 
-/**
- * Set the class of the children, based on whether it is the idSelected or not.
- * @param {HTMLDivElement} elBase
- * @param {string} idSelected
- * @param {string} classSelected
- * @param {string} classUnSelected
- */
-function el_children_config_class(elBase, idSelected, classSelected, classUnSelected="") {
-    for(let child of elBase.children) {
-        if (child.id == idSelected) {
-            child.className = classSelected;
+    /**
+     * Handle the response from the server be it in oneshot or multipart/stream mode.
+     * Also take care of the optional garbage trimming.
+     * @param {Response} resp
+     * @param {string} apiEP
+     * @param {HTMLDivElement} elDiv
+     */
+    async handle_response(resp, apiEP, elDiv) {
+        let theResp = {
+            assistant: "",
+            trimmed: "",
+        }
+        if (gMe.bStream) {
+            try {
+                theResp.assistant = await this.handle_response_multipart(resp, apiEP, elDiv);
+                this.latestResponse = "";
+            } catch (error) {
+                theResp.assistant = this.latestResponse;
+                this.add(Roles.Assistant, theResp.assistant);
+                this.latestResponse = "";
+                throw error;
+            }
         } else {
-            child.className = classUnSelected;
+            theResp.assistant = await this.handle_response_oneshot(resp, apiEP);
+        }
+        if (gMe.bTrimGarbage) {
+            let origMsg = theResp.assistant;
+            theResp.assistant = du.trim_garbage_at_end(origMsg);
+            theResp.trimmed = origMsg.substring(theResp.assistant.length);
         }
+        this.add(Roles.Assistant, theResp.assistant);
+        return theResp;
     }
-}
 
-/**
- * Create button and set it up.
- * @param {string} id
- * @param {(this: HTMLButtonElement, ev: MouseEvent) => any} callback
- * @param {string | undefined} name
- * @param {string | undefined} innerText
- */
-function el_create_button(id, callback, name=undefined, innerText=undefined) {
-    if (!name) {
-        name = id;
-    }
-    if (!innerText) {
-        innerText = id;
-    }
-    let btn = document.createElement("button");
-    btn.id = id;
-    btn.name = name;
-    btn.innerText = innerText;
-    btn.addEventListener("click", callback);
-    return btn;
 }
 
 
@@ -302,14 +482,16 @@ class MultiChatUI {
         this.elDivChat = /** @type{HTMLDivElement} */(document.getElementById("chat-div"));
         this.elBtnUser = /** @type{HTMLButtonElement} */(document.getElementById("user-btn"));
         this.elInUser = /** @type{HTMLInputElement} */(document.getElementById("user-in"));
-        this.elSelectApiEP = /** @type{HTMLSelectElement} */(document.getElementById("api-ep"));
+        this.elDivHeading = /** @type{HTMLSelectElement} */(document.getElementById("heading"));
         this.elDivSessions = /** @type{HTMLDivElement} */(document.getElementById("sessions-div"));
+        this.elBtnSettings = /** @type{HTMLButtonElement} */(document.getElementById("settings"));
 
         this.validate_element(this.elInSystem, "system-in");
         this.validate_element(this.elDivChat, "chat-div");
         this.validate_element(this.elInUser, "user-in");
-        this.validate_element(this.elSelectApiEP, "api-ep");
+        this.validate_element(this.elDivHeading, "heading");
         this.validate_element(this.elDivChat, "sessions-div");
+        this.validate_element(this.elBtnSettings, "settings");
     }
 
     /**
@@ -350,13 +532,18 @@ class MultiChatUI {
             this.handle_session_switch(this.curChatId);
         }
 
+        this.elBtnSettings.addEventListener("click", (ev)=>{
+            this.elDivChat.replaceChildren();
+            gMe.show_settings(this.elDivChat);
+        });
+
         this.elBtnUser.addEventListener("click", (ev)=>{
             if (this.elInUser.disabled) {
                 return;
             }
-            this.handle_user_submit(this.curChatId, this.elSelectApiEP.value).catch((/** @type{Error} */reason)=>{
+            this.handle_user_submit(this.curChatId, gMe.apiEP).catch((/** @type{Error} */reason)=>{
                 let msg = `ERRR:SimpleChat\nMCUI:HandleUserSubmit:${this.curChatId}\n${reason.name}:${reason.message}`;
-                console.debug(msg.replace("\n", ":"));
+                console.error(msg.replace("\n", ":"));
                 alert(msg);
                 this.ui_reset_userinput();
             });
@@ -377,6 +564,8 @@ class MultiChatUI {
             // allow user to insert enter into the system prompt using shift+enter.
             // while just pressing enter key will lead to setting the system prompt.
             if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                let value = this.elInSystem.value;
+                this.elInSystem.value = value.substring(0,value.length-1);
                 let chat = this.simpleChats[this.curChatId];
                 chat.add_system_anytime(this.elInSystem.value, this.curChatId);
                 chat.show(this.elDivChat);
@@ -392,34 +581,12 @@ class MultiChatUI {
      * @param {boolean} bSwitchSession
      */
     new_chat_session(chatId, bSwitchSession=false) {
-        this.simpleChats[chatId] = new SimpleChat();
+        this.simpleChats[chatId] = new SimpleChat(chatId);
         if (bSwitchSession) {
             this.handle_session_switch(chatId);
         }
     }
 
-    /**
-     * Try read json response early, if available.
-     * @param {Response} resp
-     */
-    async read_json_early(resp) {
-        if (!resp.body) {
-            throw Error("ERRR:SimpleChat:MCUI:ReadJsonEarly:No body...");
-        }
-        let tdUtf8 = new TextDecoder("utf-8");
-        let rr = resp.body.getReader();
-        let gotBody = "";
-        while(true) {
-            let { value: cur,  done: done} = await rr.read();
-            let curBody = tdUtf8.decode(cur);
-            console.debug("DBUG:SC:PART:", curBody);
-            gotBody += curBody;
-            if (done) {
-                break;
-            }
-        }
-        return JSON.parse(gotBody);
-    }
 
     /**
      * Handle user query submit request, wrt specified chat session.
@@ -434,7 +601,7 @@ class MultiChatUI {
         // So if user wants to simulate a multi-chat based completion query,
         // they will have to enter the full thing, as a suitable multiline
         // user input/query.
-        if ((apiEP == ApiEP.Completion) && (gMe.bCompletionFreshChatAlways)) {
+        if ((apiEP == ApiEP.Type.Completion) && (gMe.bCompletionFreshChatAlways)) {
             chat.clear();
         }
 
@@ -447,41 +614,26 @@ class MultiChatUI {
         }
         chat.show(this.elDivChat);
 
-        let theBody;
-        let theUrl = gChatURL[apiEP]
-        if (apiEP == ApiEP.Chat) {
-            theBody = chat.request_messages_jsonstr();
-        } else {
-            theBody = chat.request_prompt_jsonstr(gMe.bCompletionInsertStandardRolePrefix);
-        }
+        let theUrl = ApiEP.Url(gMe.baseURL, apiEP);
+        let theBody = chat.request_jsonstr(apiEP);
 
         this.elInUser.value = "working...";
         this.elInUser.disabled = true;
         console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:${theUrl}:ReqBody:${theBody}`);
+        let theHeaders = chat.fetch_headers(apiEP);
         let resp = await fetch(theUrl, {
             method: "POST",
-            headers: {
-                "Content-Type": "application/json",
-            },
+            headers: theHeaders,
             body: theBody,
         });
 
-        let respBody = await resp.json();
-        //let respBody = await this.read_json_early(resp);
-        console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
-        let assistantMsg;
-        if (apiEP == ApiEP.Chat) {
-            assistantMsg = respBody["choices"][0]["message"]["content"];
-        } else {
-            try {
-                assistantMsg = respBody["choices"][0]["text"];
-            } catch {
-                assistantMsg = respBody["content"];
-            }
-        }
-        chat.add(Roles.Assistant, assistantMsg);
+        let theResp = await chat.handle_response(resp, apiEP, this.elDivChat);
         if (chatId == this.curChatId) {
             chat.show(this.elDivChat);
+            if (theResp.trimmed.length > 0) {
+                let p = ui.el_create_append_p(`TRIMMED:${theResp.trimmed}`, this.elDivChat);
+                p.className="role-trim";
+            }
         } else {
             console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
         }
@@ -500,7 +652,7 @@ class MultiChatUI {
         }
         elDiv.replaceChildren();
         // Btn for creating new chat session
-        let btnNew = el_create_button("New CHAT", (ev)=> {
+        let btnNew = ui.el_create_button("New CHAT", (ev)=> {
             if (this.elInUser.disabled) {
                 console.error(`ERRR:SimpleChat:MCUI:NewChat:Current session [${this.curChatId}] awaiting response, ignoring request...`);
                 alert("ERRR:SimpleChat\nMCUI:NewChat\nWait for response to pending query, before starting new chat session");
@@ -514,7 +666,7 @@ class MultiChatUI {
             }
             this.new_chat_session(chatIdGot, true);
             this.create_session_btn(elDiv, chatIdGot);
-            el_children_config_class(elDiv, chatIdGot, "session-selected", "");
+            ui.el_children_config_class(elDiv, chatIdGot, "session-selected", "");
         });
         elDiv.appendChild(btnNew);
         // Btns for existing chat sessions
@@ -528,7 +680,7 @@ class MultiChatUI {
     }
 
     create_session_btn(elDiv, cid) {
-        let btn = el_create_button(cid, (ev)=>{
+        let btn = ui.el_create_button(cid, (ev)=>{
             let target = /** @type{HTMLButtonElement} */(ev.target);
             console.debug(`DBUG:SimpleChat:MCUI:SessionClick:${target.id}`);
             if (this.elInUser.disabled) {
@@ -537,7 +689,7 @@ class MultiChatUI {
                 return;
             }
             this.handle_session_switch(target.id);
-            el_children_config_class(elDiv, target.id, "session-selected", "");
+            ui.el_children_config_class(elDiv, target.id, "session-selected", "");
         });
         elDiv.appendChild(btn);
         return btn;
@@ -567,46 +719,188 @@ class MultiChatUI {
 class Me {
 
     constructor() {
+        this.baseURL = "http://127.0.0.1:8080";
         this.defaultChatIds = [ "Default", "Other" ];
         this.multiChat = new MultiChatUI();
+        this.bStream = true;
         this.bCompletionFreshChatAlways = true;
         this.bCompletionInsertStandardRolePrefix = false;
+        this.bTrimGarbage = true;
         this.iRecentUserMsgCnt = 2;
+        this.sRecentUserMsgCnt = {
+            "Full": -1,
+            "Last0": 1,
+            "Last1": 2,
+            "Last2": 3,
+            "Last4": 5,
+        };
+        this.apiEP = ApiEP.Type.Chat;
+        this.headers = {
+            "Content-Type": "application/json",
+            "Authorization": "", // Authorization: Bearer OPENAI_API_KEY
+        }
         // Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
-        this.chatRequestOptions = {
+        this.apiRequestOptions = {
+            "model": "gpt-3.5-turbo",
             "temperature": 0.7,
             "max_tokens": 1024,
-            "frequency_penalty": 1.2,
-            "presence_penalty": 1.2,
-            "n_predict": 1024
+            "n_predict": 1024,
+            "cache_prompt": false,
+            //"frequency_penalty": 1.2,
+            //"presence_penalty": 1.2,
         };
     }
 
     /**
+     * Disable console.debug by mapping it to a empty function.
+     */
+    debug_disable() {
+        this.console_debug = console.debug;
+        console.debug = () => {
+
+        };
+    }
+
+    /**
+     * Setup the load saved chat ui.
+     * @param {HTMLDivElement} div
+     * @param {SimpleChat} chat
+     */
+    setup_load(div, chat) {
+        if (!(chat.ods_key() in localStorage)) {
+            return;
+        }
+        div.innerHTML += `
Restore

+        
Load previously saved chat session, if available
`;
+        let btn = ui.el_create_button(chat.ods_key(), (ev)=>{
+            console.log("DBUG:SimpleChat:SC:Load", chat);
+            chat.load();
+            queueMicrotask(()=>{
+                chat.show(div);
+                this.multiChat.elInSystem.value = chat.get_system_latest();
+            });
+        });
+        div.appendChild(btn);
+    }
+
+    /**
+     * Show the configurable parameters info in the passed Div element.
      * @param {HTMLDivElement} elDiv
+     * @param {boolean} bAll
      */
-    show_info(elDiv) {
+    show_info(elDiv, bAll=false) {
 
-        var p = document.createElement("p");
-        p.innerText = "Settings (devel-tools-console gMe)";
+        let p = ui.el_create_append_p("Settings (devel-tools-console document[gMe])", elDiv);
         p.className = "role-system";
-        elDiv.appendChild(p);
 
-        var p = document.createElement("p");
-        p.innerText = `bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`;
-        elDiv.appendChild(p);
+        if (bAll) {
+
+            ui.el_create_append_p(`baseURL:${this.baseURL}`, elDiv);
+
+            ui.el_create_append_p(`Authorization:${this.headers["Authorization"]}`, elDiv);
+
+            ui.el_create_append_p(`bStream:${this.bStream}`, elDiv);
+
+            ui.el_create_append_p(`bTrimGarbage:${this.bTrimGarbage}`, elDiv);
+
+            ui.el_create_append_p(`ApiEndPoint:${this.apiEP}`, elDiv);
 
-        p = document.createElement("p");
-        p.innerText = `bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`;
-        elDiv.appendChild(p);
+            ui.el_create_append_p(`iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`, elDiv);
 
-        p = document.createElement("p");
-        p.innerText = `iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`;
-        elDiv.appendChild(p);
+            ui.el_create_append_p(`bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`, elDiv);
 
-        p = document.createElement("p");
-        p.innerText = `chatRequestOptions:${JSON.stringify(this.chatRequestOptions)}`;
-        elDiv.appendChild(p);
+            ui.el_create_append_p(`bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`, elDiv);
+
+        }
+
+        ui.el_create_append_p(`apiRequestOptions:${JSON.stringify(this.apiRequestOptions, null, " - ")}`, elDiv);
+        ui.el_create_append_p(`headers:${JSON.stringify(this.headers, null, " - ")}`, elDiv);
+
+    }
+
+    /**
+     * Auto create ui input elements for fields in apiRequestOptions
+     * Currently supports text and number field types.
+     * @param {HTMLDivElement} elDiv
+     */
+    show_settings_apirequestoptions(elDiv) {
+        let typeDict = {
+            "string": "text",
+            "number": "number",
+        };
+        let fs = document.createElement("fieldset");
+        let legend = document.createElement("legend");
+        legend.innerText = "ApiRequestOptions";
+        fs.appendChild(legend);
+        elDiv.appendChild(fs);
+        for(const k in this.apiRequestOptions) {
+            let val = this.apiRequestOptions[k];
+            let type = typeof(val);
+            if (((type == "string") || (type == "number"))) {
+                let inp = ui.el_creatediv_input(`Set${k}`, k, typeDict[type], this.apiRequestOptions[k], (val)=>{
+                    if (type == "number") {
+                        val = Number(val);
+                    }
+                    this.apiRequestOptions[k] = val;
+                });
+                fs.appendChild(inp.div);
+            } else if (type == "boolean") {
+                let bbtn = ui.el_creatediv_boolbutton(`Set{k}`, k, {true: "true", false: "false"}, val, (userVal)=>{
+                    this.apiRequestOptions[k] = userVal;
+                });
+                fs.appendChild(bbtn.div);
+            }
+        }
+    }
+
+    /**
+     * Show settings ui for configurable parameters, in the passed Div element.
+     * @param {HTMLDivElement} elDiv
+     */
+    show_settings(elDiv) {
+
+        let inp = ui.el_creatediv_input("SetBaseURL", "BaseURL", "text", this.baseURL, (val)=>{
+            this.baseURL = val;
+        });
+        elDiv.appendChild(inp.div);
+
+        inp = ui.el_creatediv_input("SetAuthorization", "Authorization", "text", this.headers["Authorization"], (val)=>{
+            this.headers["Authorization"] = val;
+        });
+        inp.el.placeholder = "Bearer OPENAI_API_KEY";
+        elDiv.appendChild(inp.div);
+
+        let bb = ui.el_creatediv_boolbutton("SetStream", "Stream", {true: "[+] yes stream", false: "[-] do oneshot"}, this.bStream, (val)=>{
+            this.bStream = val;
+        });
+        elDiv.appendChild(bb.div);
+
+        bb = ui.el_creatediv_boolbutton("SetTrimGarbage", "TrimGarbage", {true: "[+] yes trim", false: "[-] dont trim"}, this.bTrimGarbage, (val)=>{
+            this.bTrimGarbage = val;
+        });
+        elDiv.appendChild(bb.div);
+
+        this.show_settings_apirequestoptions(elDiv);
+
+        let sel = ui.el_creatediv_select("SetApiEP", "ApiEndPoint", ApiEP.Type, this.apiEP, (val)=>{
+            this.apiEP = ApiEP.Type[val];
+        });
+        elDiv.appendChild(sel.div);
+
+        sel = ui.el_creatediv_select("SetChatHistoryInCtxt", "ChatHistoryInCtxt", this.sRecentUserMsgCnt, this.iRecentUserMsgCnt, (val)=>{
+            this.iRecentUserMsgCnt = this.sRecentUserMsgCnt[val];
+        });
+        elDiv.appendChild(sel.div);
+
+        bb = ui.el_creatediv_boolbutton("SetCompletionFreshChatAlways", "CompletionFreshChatAlways", {true: "[+] yes fresh", false: "[-] no, with history"}, this.bCompletionFreshChatAlways, (val)=>{
+            this.bCompletionFreshChatAlways = val;
+        });
+        elDiv.appendChild(bb.div);
+
+        bb = ui.el_creatediv_boolbutton("SetCompletionInsertStandardRolePrefix", "CompletionInsertStandardRolePrefix", {true: "[+] yes insert", false: "[-] dont insert"}, this.bCompletionInsertStandardRolePrefix, (val)=>{
+            this.bCompletionInsertStandardRolePrefix = val;
+        });
+        elDiv.appendChild(bb.div);
 
     }
 
@@ -619,6 +913,9 @@ let gMe;
 function startme() {
     console.log("INFO:SimpleChat:StartMe:Starting...");
     gMe = new Me();
+    gMe.debug_disable();
+    document["gMe"] = gMe;
+    document["du"] = du;
     for (let cid of gMe.defaultChatIds) {
         gMe.multiChat.new_chat_session(cid);
     }
diff --git a/llm/llama.cpp/examples/server/public_simplechat/simplechat_screens.webp b/llm/llama.cpp/examples/server/public_simplechat/simplechat_screens.webp
new file mode 100644
index 0000000000000000000000000000000000000000..ccea44396051686b97220b0f5b6b9beb63706114
Binary files /dev/null and b/llm/llama.cpp/examples/server/public_simplechat/simplechat_screens.webp differ
diff --git a/llm/llama.cpp/examples/server/public_simplechat/ui.mjs b/llm/llama.cpp/examples/server/public_simplechat/ui.mjs
new file mode 100644
index 0000000000000000000000000000000000000000..b2d5b9aeab76c0724f9c8fa5238e3ce009aaa877
--- /dev/null
+++ b/llm/llama.cpp/examples/server/public_simplechat/ui.mjs
@@ -0,0 +1,211 @@
+//@ts-check
+// Helpers to work with html elements
+// by Humans for All
+//
+
+
+/**
+ * Set the class of the children, based on whether it is the idSelected or not.
+ * @param {HTMLDivElement} elBase
+ * @param {string} idSelected
+ * @param {string} classSelected
+ * @param {string} classUnSelected
+ */
+export function el_children_config_class(elBase, idSelected, classSelected, classUnSelected="") {
+    for(let child of elBase.children) {
+        if (child.id == idSelected) {
+            child.className = classSelected;
+        } else {
+            child.className = classUnSelected;
+        }
+    }
+}
+
+/**
+ * Create button and set it up.
+ * @param {string} id
+ * @param {(this: HTMLButtonElement, ev: MouseEvent) => any} callback
+ * @param {string | undefined} name
+ * @param {string | undefined} innerText
+ */
+export function el_create_button(id, callback, name=undefined, innerText=undefined) {
+    if (!name) {
+        name = id;
+    }
+    if (!innerText) {
+        innerText = id;
+    }
+    let btn = document.createElement("button");
+    btn.id = id;
+    btn.name = name;
+    btn.innerText = innerText;
+    btn.addEventListener("click", callback);
+    return btn;
+}
+
+/**
+ * Create a para and set it up. Optionaly append it to a passed parent.
+ * @param {string} text
+ * @param {HTMLElement | undefined} elParent
+ * @param {string | undefined} id
+ */
+export function el_create_append_p(text, elParent=undefined, id=undefined) {
+    let para = document.createElement("p");
+    para.innerText = text;
+    if (id) {
+        para.id = id;
+    }
+    if (elParent) {
+        elParent.appendChild(para);
+    }
+    return para;
+}
+
+/**
+ * Create a button which represents bool value using specified text wrt true and false.
+ * When ever user clicks the button, it will toggle the value and update the shown text.
+ *
+ * @param {string} id
+ * @param {{true: string, false: string}} texts
+ * @param {boolean} defaultValue
+ * @param {function(boolean):void} cb
+ */
+export function el_create_boolbutton(id, texts, defaultValue, cb) {
+    let el = document.createElement("button");
+    el["xbool"] = defaultValue;
+    el["xtexts"] = structuredClone(texts);
+    el.innerText = el["xtexts"][String(defaultValue)];
+    if (id) {
+        el.id = id;
+    }
+    el.addEventListener('click', (ev)=>{
+        el["xbool"] = !el["xbool"];
+        el.innerText = el["xtexts"][String(el["xbool"])];
+        cb(el["xbool"]);
+    })
+    return el;
+}
+
+/**
+ * Create a div wrapped button which represents bool value using specified text wrt true and false.
+ * @param {string} id
+ * @param {string} label
+ * @param {{ true: string; false: string; }} texts
+ * @param {boolean} defaultValue
+ * @param {(arg0: boolean) => void} cb
+ * @param {string} className
+ */
+export function el_creatediv_boolbutton(id, label, texts, defaultValue, cb, className="gridx2") {
+    let div = document.createElement("div");
+    div.className = className;
+    let lbl = document.createElement("label");
+    lbl.setAttribute("for", id);
+    lbl.innerText = label;
+    div.appendChild(lbl);
+    let btn = el_create_boolbutton(id, texts, defaultValue, cb);
+    div.appendChild(btn);
+    return { div: div, el: btn };
+}
+
+
+/**
+ * Create a select ui element, with a set of options to select from.
+ * * options: an object which contains name-value pairs
+ * * defaultOption: the value whose name should be choosen, by default.
+ * * cb : the call back returns the name string of the option selected.
+ *
+ * @param {string} id
+ * @param {Object} options
+ * @param {*} defaultOption
+ * @param {function(string):void} cb
+ */
+export function el_create_select(id, options, defaultOption, cb) {
+    let el = document.createElement("select");
+    el["xselected"] = defaultOption;
+    el["xoptions"] = structuredClone(options);
+    for(let cur of Object.keys(options)) {
+        let op = document.createElement("option");
+        op.value = cur;
+        op.innerText = cur;
+        if (options[cur] == defaultOption) {
+            op.selected = true;
+        }
+        el.appendChild(op);
+    }
+    if (id) {
+        el.id = id;
+        el.name = id;
+    }
+    el.addEventListener('change', (ev)=>{
+        let target = /** @type{HTMLSelectElement} */(ev.target);
+        console.log("DBUG:UI:Select:", id, ":", target.value);
+        cb(target.value);
+    })
+    return el;
+}
+
+/**
+ * Create a div wrapped select ui element, with a set of options to select from.
+ *
+ * @param {string} id
+ * @param {any} label
+ * @param {{ [x: string]: any; }} options
+ * @param {any} defaultOption
+ * @param {(arg0: string) => void} cb
+ * @param {string} className
+ */
+export function el_creatediv_select(id, label, options, defaultOption, cb, className="gridx2") {
+    let div = document.createElement("div");
+    div.className = className;
+    let lbl = document.createElement("label");
+    lbl.setAttribute("for", id);
+    lbl.innerText = label;
+    div.appendChild(lbl);
+    let sel = el_create_select(id, options,defaultOption, cb);
+    div.appendChild(sel);
+    return { div: div, el: sel };
+}
+
+
+/**
+ * Create a input ui element.
+ *
+ * @param {string} id
+ * @param {string} type
+ * @param {any} defaultValue
+ * @param {function(any):void} cb
+ */
+export function el_create_input(id, type, defaultValue, cb) {
+    let el = document.createElement("input");
+    el.type = type;
+    el.value = defaultValue;
+    if (id) {
+        el.id = id;
+    }
+    el.addEventListener('change', (ev)=>{
+        cb(el.value);
+    })
+    return el;
+}
+
+/**
+ * Create a div wrapped input.
+ *
+ * @param {string} id
+ * @param {string} label
+ * @param {string} type
+ * @param {any} defaultValue
+ * @param {function(any):void} cb
+ * @param {string} className
+ */
+export function el_creatediv_input(id, label, type, defaultValue, cb, className="gridx2") {
+    let div = document.createElement("div");
+    div.className = className;
+    let lbl = document.createElement("label");
+    lbl.setAttribute("for", id);
+    lbl.innerText = label;
+    div.appendChild(lbl);
+    let el = el_create_input(id, type, defaultValue, cb);
+    div.appendChild(el);
+    return { div: div, el: el };
+}
diff --git a/llm/llama.cpp/examples/server/server.cpp b/llm/llama.cpp/examples/server/server.cpp
index e9904263d53c73197c56670f84191a5394d685ea..898c83ea3522bb6d368535fb449fb727f5caeac3 100644
--- a/llm/llama.cpp/examples/server/server.cpp
+++ b/llm/llama.cpp/examples/server/server.cpp
@@ -17,9 +17,20 @@
 #include "json.hpp"
 
 // auto generated files (update with ./deps.sh)
+#include "colorthemes.css.hpp"
+#include "style.css.hpp"
+#include "theme-beeninorder.css.hpp"
+#include "theme-ketivah.css.hpp"
+#include "theme-mangotango.css.hpp"
+#include "theme-playground.css.hpp"
+#include "theme-polarnight.css.hpp"
+#include "theme-snowstorm.css.hpp"
 #include "index.html.hpp"
+#include "index-new.html.hpp"
 #include "index.js.hpp"
 #include "completion.js.hpp"
+#include "system-prompts.js.hpp"
+#include "prompt-formats.js.hpp"
 #include "json-schema-to-grammar.mjs.hpp"
 
 #include 
@@ -67,6 +78,7 @@ enum server_task_type {
     SERVER_TASK_TYPE_SLOT_SAVE,
     SERVER_TASK_TYPE_SLOT_RESTORE,
     SERVER_TASK_TYPE_SLOT_ERASE,
+    SERVER_TASK_TYPE_SET_LORA,
 };
 
 struct server_task {
@@ -112,29 +124,6 @@ struct slot_params {
     json input_suffix;
 };
 
-struct server_params {
-    int32_t port           = 8080;
-    int32_t read_timeout   = 600;
-    int32_t write_timeout  = 600;
-    int32_t n_threads_http = -1;
-
-    std::string hostname      = "127.0.0.1";
-    std::string public_path   = "";
-    std::string chat_template = "";
-    std::string system_prompt = "";
-
-    std::vector api_keys;
-
-#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
-    std::string ssl_key_file = "";
-    std::string ssl_cert_file = "";
-#endif
-
-    bool slots_endpoint   = true;
-    bool metrics_endpoint = false;
-    std::string slot_save_path;
-};
-
 struct server_slot {
     int id;
     int id_task = -1;
@@ -159,7 +148,7 @@ struct server_slot {
     int32_t n_prompt_tokens           = 0;
     int32_t n_prompt_tokens_processed = 0;
 
-    json prompt;
+    json prompt; // can be either a string, array of strings or array of token ids
 
     // when a task is submitted, we first tokenize the prompt and store it here
     std::vector prompt_tokens;
@@ -634,6 +623,7 @@ struct server_response {
 struct server_context {
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
+    std::vector lora_adapters;
 
     gpt_params params;
 
@@ -659,6 +649,9 @@ struct server_context {
 
     server_metrics metrics;
 
+    // Necessary similarity of prompt for slot selection
+    float slot_prompt_similarity = 0.0f;
+
     ~server_context() {
         if (ctx) {
             llama_free(ctx);
@@ -686,7 +679,11 @@ struct server_context {
         // dedicate one sequence to the system prompt
         params.n_parallel += 1;
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        llama_init_result llama_init = llama_init_from_gpt_params(params);
+
+        model = llama_init.model;
+        ctx = llama_init.context;
+        lora_adapters = llama_init.lora_adapters;
         params.n_parallel -= 1; // but be sneaky about it
         if (model == nullptr) {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -746,6 +743,8 @@ struct server_context {
             slot.ga_n = ga_n;
             slot.ga_w = ga_w;
 
+            slot.sparams = params.sparams;
+
             slot.reset();
 
             slots.push_back(slot);
@@ -807,29 +806,94 @@ struct server_context {
         return prompt_tokens;
     }
 
-    server_slot * get_slot(int id) {
-        int64_t t_last = ggml_time_us();
-
-        server_slot * last_used = nullptr;
-
+    server_slot * get_slot_by_id(int id) {
         for (server_slot & slot : slots) {
-            if (slot.id == id && slot.available()) {
+            if (slot.id == id) {
                 return &slot;
             }
+        }
+
+        return nullptr;
+    }
+
+    server_slot * get_available_slot(const std::string & prompt) {
+        server_slot * ret = nullptr;
+
+        // find the slot that has at least n% prompt similarity
+        if (ret == nullptr && slot_prompt_similarity != 0.0f && !prompt.empty()) {
+            int max_lcp_len = 0;
+            float similarity = 0;
+
+            for (server_slot & slot : slots) {
+                // skip the slot if it is not available
+                if (!slot.available()) {
+                    continue;
+                }
+
+                // skip the slot if it does not contains prompt
+                if (!slot.prompt.is_string()) {
+                    continue;
+                }
+
+                // current slot's prompt
+                std::string slot_prompt = slot.prompt.get();
+
+                // length of the current slot's prompt
+                int slot_prompt_len = slot_prompt.size();
+
+                // length of the Longest Common Prefix between the current slot's prompt and the input prompt
+                int lcp_len = common_part(slot_prompt, prompt);
+
+                // fraction of the common substring length compared to the current slot's prompt length
+                similarity = static_cast(lcp_len) / slot_prompt_len;
+
+                // select the current slot if the criteria match
+                if (lcp_len > max_lcp_len && similarity > slot_prompt_similarity) {
+                    max_lcp_len = lcp_len;
+                    ret = &slot;
+                }
+            }
+
+            if (ret != nullptr) {
+                LOG_VERBOSE("selected slot by lcp similarity", {
+                    {"id_slot", ret->id},
+                    {"max_lcp_len", max_lcp_len},
+                    {"similarity", similarity},
+                });
+            }
+        }
+
+        // find the slot that has been least recently used
+        if (ret == nullptr) {
+            int64_t t_last = ggml_time_us();
+            for (server_slot & slot : slots) {
+                // skip the slot if it is not available
+                if (!slot.available()) {
+                    continue;
+                }
+
+                // select the current slot if the criteria match
+                if (slot.t_last_used < t_last) {
+                    t_last = slot.t_last_used;
+                    ret = &slot;
+                }
+            }
 
-            // among all available slots, find the one that has been least recently used
-            if (slot.available() && slot.t_last_used < t_last) {
-                last_used = &slot;
-                t_last = slot.t_last_used;
+            if (ret != nullptr) {
+                LOG_VERBOSE("selected slot by lru", {
+                    {"id_slot", ret->id},
+                    {"t_last", t_last},
+                });
             }
         }
 
-        return last_used;
+        return ret;
     }
 
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
         slot_params default_params;
-        llama_sampling_params default_sparams;
+        // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them)
+        llama_sampling_params default_sparams = params.sparams;
         auto & data = task.data;
 
         if (data.count("__oaicompat") != 0) {
@@ -842,7 +906,7 @@ struct server_context {
 
         slot.params.stream             = json_value(data, "stream",            false);
         slot.params.cache_prompt       = json_value(data, "cache_prompt",      false);
-        slot.params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
+        slot.params.n_predict          = json_value(data, "n_predict",         json_value(data, "max_tokens", default_params.n_predict));
         slot.sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
         slot.sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot.sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
@@ -900,16 +964,19 @@ struct server_context {
         slot.params.input_suffix = json_value(data, "input_suffix", default_params.input_suffix);
 
         // get prompt
-        {
+        if (!task.infill) {
             const auto & prompt = data.find("prompt");
             if (prompt == data.end()) {
-                send_error(task, "Either \"prompt\" or \"messages\" must be provided", ERROR_TYPE_INVALID_REQUEST);
+                send_error(task, "\"prompt\" must be provided", ERROR_TYPE_INVALID_REQUEST);
                 return false;
-            } else {
-                slot.prompt = *prompt;
             }
-            if (slot.prompt.is_array() && slot.prompt.size() == 0) {
-                send_error(task, "\"prompt\" cannot be an empty array", ERROR_TYPE_INVALID_REQUEST);
+
+            if ((prompt->is_string()) ||
+                (prompt->is_array() &&  prompt->size() == 1 && prompt->at(0).is_string()) ||
+                (prompt->is_array() && !prompt->empty()     && prompt->at(0).is_number_integer())) {
+                slot.prompt = *prompt;
+            } else {
+                send_error(task, "\"prompt\" must be a string or an array of integers", ERROR_TYPE_INVALID_REQUEST);
                 return false;
             }
         }
@@ -1121,7 +1188,7 @@ struct server_context {
 
     bool process_token(completion_token_output & result, server_slot & slot) {
         // remember which tokens were sampled - used for repetition penalties during sampling
-        const std::string token_str = llama_token_to_piece(ctx, result.tok, false);
+        const std::string token_str = llama_token_to_piece(ctx, result.tok, params.special);
         slot.sampled = result.tok;
 
         // search stop word and delete it
@@ -1250,7 +1317,7 @@ struct server_context {
     }
 
     json get_formated_generation(const server_slot & slot) const {
-        const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model));
+        const auto eos_bias   =             slot.sparams.logit_bias.find(llama_token_eos(model));
         const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() && eos_bias->second < 0.0f && std::isinf(eos_bias->second);
 
         std::vector samplers_sequence;
@@ -1527,13 +1594,33 @@ struct server_context {
         switch (task.type) {
             case SERVER_TASK_TYPE_COMPLETION:
                 {
-                    server_slot * slot = get_slot(json_value(task.data, "id_slot", -1));
+                    const int id_slot = json_value(task.data, "id_slot", -1);
+
+                    server_slot * slot;
+
+                    if (id_slot != -1) {
+                        slot = get_slot_by_id(id_slot);
+                    } else {
+                        std::string prompt;
+                        if (task.data.contains("prompt") && task.data.at("prompt").is_string()) {
+                            prompt = json_value(task.data, "prompt", std::string());
+                        }
+
+                        slot = get_available_slot(prompt);
+                    }
+
                     if (slot == nullptr) {
                         // if no slot is available, we defer this task for processing later
                         LOG_VERBOSE("no slot is available", {{"id_task", task.id}});
                         queue_tasks.defer(task);
                         break;
                     }
+                    if (!slot->available()) {
+                        // if requested slot is unavailable, we defer this task for processing later
+                        LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
+                        queue_tasks.defer(task);
+                        break;
+                    }
 
                     if (task.data.contains("system_prompt")) {
                         std::string sys_prompt = json_value(task.data, "system_prompt", std::string());
@@ -1650,11 +1737,17 @@ struct server_context {
             case SERVER_TASK_TYPE_SLOT_SAVE:
                 {
                     int id_slot = task.data.at("id_slot");
-                    server_slot * slot = get_slot(id_slot);
+                    server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
                         break;
                     }
+                    if (!slot->available()) {
+                        // if requested slot is unavailable, we defer this task for processing later
+                        LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
+                        queue_tasks.defer(task);
+                        break;
+                    }
 
                     const size_t token_count = slot->cache_tokens.size();
                     const int64_t t_start = ggml_time_us();
@@ -1685,11 +1778,17 @@ struct server_context {
             case SERVER_TASK_TYPE_SLOT_RESTORE:
                 {
                     int id_slot = task.data.at("id_slot");
-                    server_slot * slot = get_slot(id_slot);
+                    server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
                         break;
                     }
+                    if (!slot->available()) {
+                        // if requested slot is unavailable, we defer this task for processing later
+                        LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
+                        queue_tasks.defer(task);
+                        break;
+                    }
 
                     const int64_t t_start = ggml_time_us();
 
@@ -1727,11 +1826,17 @@ struct server_context {
             case SERVER_TASK_TYPE_SLOT_ERASE:
                 {
                     int id_slot = task.data.at("id_slot");
-                    server_slot * slot = get_slot(id_slot);
+                    server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
                         break;
                     }
+                    if (!slot->available()) {
+                        // if requested slot is unavailable, we defer this task for processing later
+                        LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
+                        queue_tasks.defer(task);
+                        break;
+                    }
 
                     // Erase token cache
                     const size_t n_erased = slot->cache_tokens.size();
@@ -1748,6 +1853,14 @@ struct server_context {
                     };
                     queue_results.send(result);
                 } break;
+            case SERVER_TASK_TYPE_SET_LORA:
+                {
+                    llama_lora_adapters_apply(ctx, lora_adapters);
+                    server_task_result result;
+                    result.id = task.id;
+                    result.data = json{{ "success", true }};
+                    queue_results.send(result);
+                } break;
         }
     }
 
@@ -1906,6 +2019,11 @@ struct server_context {
         int32_t n_batch  = llama_n_batch(ctx);
         int32_t n_ubatch = llama_n_ubatch(ctx);
 
+        // track if this is an embedding or non-embedding batch
+        // if we've added sampled tokens above, we are in non-embedding mode
+        // -1: none, 0: non-embedding, 1: embedding
+        int32_t batch_type = batch.n_tokens > 0 ? 0 : -1;
+
         // next, batch any pending prompts without exceeding n_batch
         if (params.cont_batching || batch.n_tokens == 0) {
             for (auto & slot : slots) {
@@ -1924,6 +2042,7 @@ struct server_context {
                         slot.t_start_generation = 0;
 
                         if (slot.infill) {
+                            const bool add_bos = llama_should_add_bos_token(model);
                             bool suff_rm_leading_spc = true;
                             if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
                                 params.input_suffix.erase(0, 1);
@@ -1939,11 +2058,21 @@ struct server_context {
                             }
 
                             prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
-                            prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
-                            prefix_tokens.insert(prefix_tokens.end(),   llama_token_suffix(model));
-                            prefix_tokens.insert(prefix_tokens.end(),   suffix_tokens.begin(), suffix_tokens.end());
-                            prefix_tokens.push_back(llama_token_middle(model));
-                            prompt_tokens = prefix_tokens;
+                            suffix_tokens.insert(suffix_tokens.begin(), llama_token_suffix(model));
+
+                            auto embd_inp = params.spm_infill ? suffix_tokens : prefix_tokens;
+                            auto embd_end = params.spm_infill ? prefix_tokens : suffix_tokens;
+                            if (add_bos) {
+                                embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
+                            }
+                            embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+
+                            const llama_token middle_token = llama_token_middle(model);
+                            if (middle_token >= 0) {
+                                embd_inp.push_back(middle_token);
+                            }
+
+                            prompt_tokens = embd_inp;
                         } else {
                             prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt
                         }
@@ -2065,6 +2194,14 @@ struct server_context {
                         }
                     }
 
+                    // check that we are in the right batch_type, if not defer the slot
+                    bool slot_type = slot.embedding ? 1 : 0;
+                    if (batch_type == -1) {
+                        batch_type = slot_type;
+                    } else if (batch_type != slot_type) {
+                        continue;
+                    }
+
                     // keep only the common part
                     int p0 = (int) system_tokens.size() + slot.n_past;
                     if (!llama_kv_cache_seq_rm(ctx, slot.id + 1, p0, -1)) {
@@ -2166,6 +2303,9 @@ struct server_context {
             {"n_tokens", batch.n_tokens},
         });
 
+        // make sure we're in the right embedding mode
+        llama_set_embeddings(ctx, batch_type == 1);
+
         // process the created batch of tokens
         for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
             const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
@@ -2323,561 +2463,6 @@ struct server_context {
     }
 };
 
-static void server_print_usage(const char * argv0, const gpt_params & params, const server_params & sparams) {
-    printf("usage: %s [options]\n", argv0);
-    printf("\n");
-    printf("options:\n");
-    printf("  -h, --help                show this help message and exit\n");
-    printf("  -v, --verbose             verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    printf("  -t N, --threads N         number of threads to use during computation (default: %d)\n", params.n_threads);
-    printf("  -tb N, --threads-batch N  number of threads to use during batch and prompt processing (default: same as --threads)\n");
-    printf("  --threads-http N          number of threads in the http server pool to process requests (default: max(hardware concurrency - 1, --parallel N + 2))\n");
-    printf("  -c N, --ctx-size N        size of the prompt context (default: %d)\n", params.n_ctx);
-    printf("  --rope-scaling {none,linear,yarn}\n");
-    printf("                            RoPE frequency scaling method, defaults to linear unless specified by the model\n");
-    printf("  --rope-freq-base N        RoPE base frequency (default: loaded from model)\n");
-    printf("  --rope-freq-scale N       RoPE frequency scaling factor, expands context by a factor of 1/N\n");
-    printf("  --yarn-ext-factor N       YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
-    printf("  --yarn-attn-factor N      YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
-    printf("  --yarn-beta-slow N        YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
-    printf("  --yarn-beta-fast N        YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
-    printf("  --pooling {none,mean,cls} pooling type for embeddings, use model default if unspecified\n");
-    printf("  -dt N, --defrag-thold N\n");
-    printf("                            KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
-    printf("  -b N, --batch-size N      logical maximum batch size (default: %d)\n", params.n_batch);
-    printf("  -ub N, --ubatch-size N    physical maximum batch size (default: %d)\n", params.n_ubatch);
-    if (llama_supports_mlock()) {
-        printf("  --mlock                   force system to keep model in RAM rather than swapping or compressing\n");
-    }
-    if (llama_supports_mmap()) {
-        printf("  --no-mmap                 do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
-    }
-    printf("  --numa TYPE               attempt optimizations that help on some NUMA systems\n");
-    printf("                              - distribute: spread execution evenly over all nodes\n");
-    printf("                              - isolate: only spawn threads on CPUs on the node that execution started on\n");
-    printf("                              - numactl: use the CPU map provided my numactl\n");
-    if (llama_supports_gpu_offload()) {
-        printf("  -ngl N, --n-gpu-layers N\n");
-        printf("                            number of layers to store in VRAM\n");
-        printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
-        printf("                            how to split the model across multiple GPUs, one of:\n");
-        printf("                              - none: use one GPU only\n");
-        printf("                              - layer (default): split layers and KV across GPUs\n");
-        printf("                              - row: split rows across GPUs\n");
-        printf("  -ts SPLIT --tensor-split SPLIT\n");
-        printf("                            fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
-        printf("  -mg i, --main-gpu i       the GPU to use for the model (with split-mode = none),\n");
-        printf("                            or for intermediate results and KV (with split-mode = row)\n");
-        printf("  -nkvo, --no-kv-offload\n");
-        printf("                            disable KV offload\n");
-    }
-    printf("  -m FNAME, --model FNAME\n");
-    printf("                            model path (default: models/$filename with filename from --hf-file or --model-url if set, otherwise %s)\n", DEFAULT_MODEL_PATH);
-    printf("  -mu MODEL_URL, --model-url MODEL_URL\n");
-    printf("                            model download url (default: unused)\n");
-    printf("  -hfr REPO, --hf-repo REPO\n");
-    printf("                            Hugging Face model repository (default: unused)\n");
-    printf("  -hff FILE, --hf-file FILE\n");
-    printf("                            Hugging Face model file (default: unused)\n");
-    printf("  -a ALIAS, --alias ALIAS\n");
-    printf("                            set an alias for the model, will be added as `model` field in completion response\n");
-    printf("  --lora FNAME              apply LoRA adapter (implies --no-mmap)\n");
-    printf("  --lora-base FNAME         optional model to use as a base for the layers modified by the LoRA adapter\n");
-    printf("  --host                    ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
-    printf("  --port PORT               port to listen (default  (default: %d)\n", sparams.port);
-    printf("  --rpc SERVERS             comma separated list of RPC servers\n");
-    printf("  --path PUBLIC_PATH        path from which to serve static files (default: disabled)\n");
-    printf("  --api-key API_KEY         optional api key to enhance server security. If set, requests must include this key for access.\n");
-    printf("  --api-key-file FNAME      path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n");
-#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
-    printf("  --ssl-key-file FNAME      path to file a PEM-encoded SSL private key\n");
-    printf("  --ssl-cert-file FNAME     path to file a PEM-encoded SSL certificate\n");
-#endif
-    printf("  -to N, --timeout N        server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
-    printf("  --embeddings              enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
-    printf("  -np N, --parallel N       number of slots for process requests (default: %d)\n", params.n_parallel);
-    printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: enabled)\n");
-    printf("  -fa, --flash-attn         enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled");
-    printf("  -spf FNAME, --system-prompt-file FNAME\n");
-    printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
-    printf("  -ctk TYPE, --cache-type-k TYPE\n");
-    printf("                            KV cache data type for K (default: f16)\n");
-    printf("  -ctv TYPE, --cache-type-v TYPE\n");
-    printf("                            KV cache data type for V (default: f16)\n");
-    printf("  --log-format              log output format: json or text (default: json)\n");
-    printf("  --log-disable             disables logging to a file.\n");
-    printf("  --slots-endpoint-disable  disables slots monitoring endpoint.\n");
-    printf("  --metrics                 enable prometheus compatible metrics endpoint (default: %s).\n", sparams.metrics_endpoint ? "enabled" : "disabled");
-    printf("  --slot-save-path PATH     path to save slot kv cache (default: disabled)\n");
-    printf("\n");
-    printf("  -n, --n-predict           maximum tokens to predict (default: %d)\n", params.n_predict);
-    printf("  --override-kv KEY=TYPE:VALUE\n");
-    printf("                            advanced option to override model metadata by key. may be specified multiple times.\n");
-    printf("                            types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
-    printf("  -gan N, --grp-attn-n N    set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`\n");
-    printf("  -gaw N, --grp-attn-w N    set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`\n");
-    printf("  --chat-template JINJA_TEMPLATE\n");
-    printf("                            set custom jinja chat template (default: template taken from model's metadata)\n");
-    printf("                            only commonly used templates are accepted:\n");
-    printf("                            https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template\n");
-    printf("\n");
-}
-
-static void server_params_parse(int argc, char ** argv, server_params & sparams, gpt_params & params) {
-    gpt_params    default_params;
-    server_params default_sparams;
-
-    std::string arg;
-    bool invalid_param = false;
-
-    for (int i = 1; i < argc; i++) {
-        arg = argv[i];
-        if (arg == "--port") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.port = std::stoi(argv[i]);
-        } else if (arg == "--rpc") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.rpc_servers = argv[i];
-        } else if (arg == "--host") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.hostname = argv[i];
-        } else if (arg == "--path") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.public_path = argv[i];
-        } else if (arg == "--api-key") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.api_keys.push_back(argv[i]);
-        } else if (arg == "--api-key-file") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::ifstream key_file(argv[i]);
-            if (!key_file) {
-                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
-                invalid_param = true;
-                break;
-            }
-            std::string key;
-            while (std::getline(key_file, key)) {
-               if (key.size() > 0) {
-                   sparams.api_keys.push_back(key);
-               }
-            }
-            key_file.close();
-
-        }
-#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
-        else if (arg == "--ssl-key-file") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.ssl_key_file = argv[i];
-        } else if (arg == "--ssl-cert-file") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.ssl_cert_file = argv[i];
-        }
-#endif
-        else if (arg == "--timeout" || arg == "-to") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.read_timeout = std::stoi(argv[i]);
-            sparams.write_timeout = std::stoi(argv[i]);
-        } else if (arg == "-m" || arg == "--model") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.model = argv[i];
-        } else if (arg == "-mu" || arg == "--model-url") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.model_url = argv[i];
-        } else if (arg == "-hfr" || arg == "--hf-repo") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.hf_repo = argv[i];
-        } else if (arg == "-hff" || arg == "--hf-file") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.hf_file = argv[i];
-        } else if (arg == "-a" || arg == "--alias") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.model_alias = argv[i];
-        } else if (arg == "-h" || arg == "--help") {
-            server_print_usage(argv[0], default_params, default_sparams);
-            exit(0);
-        } else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_ctx = std::stoi(argv[i]);
-        } else if (arg == "--rope-scaling") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::string value(argv[i]);
-            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
-            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
-            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
-            else { invalid_param = true; break; }
-        } else if (arg == "--rope-freq-base") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.rope_freq_base = std::stof(argv[i]);
-        } else if (arg == "--rope-freq-scale") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.rope_freq_scale = std::stof(argv[i]);
-        } else if (arg == "--yarn-ext-factor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_ext_factor = std::stof(argv[i]);
-        }
-        else if (arg == "--yarn-attn-factor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_attn_factor = std::stof(argv[i]);
-        } else if (arg == "--yarn-beta-fast") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_beta_fast = std::stof(argv[i]);
-        } else if (arg == "--yarn-beta-slow") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_beta_slow = std::stof(argv[i]);
-        } else if (arg == "--pooling") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::string value(argv[i]);
-            /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; }
-            else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; }
-            else if (value == "cls")  { params.pooling_type = LLAMA_POOLING_TYPE_CLS; }
-            else { invalid_param = true; break; }
-        } else if (arg == "--defrag-thold" || arg == "-dt") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.defrag_thold = std::stof(argv[i]);
-        } else if (arg == "--threads" || arg == "-t") {
-            if (++i >= argc)
-            {
-                invalid_param = true;
-                break;
-            }
-            params.n_threads = std::stoi(argv[i]);
-        } else if (arg == "--grp-attn-n" || arg == "-gan") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-
-            params.grp_attn_n = std::stoi(argv[i]);
-        } else if (arg == "--grp-attn-w" || arg == "-gaw") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-
-            params.grp_attn_w = std::stoi(argv[i]);
-        } else if (arg == "--threads-batch" || arg == "-tb") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_threads_batch = std::stoi(argv[i]);
-        } else if (arg == "--threads-http") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.n_threads_http = std::stoi(argv[i]);
-        } else if (arg == "-b" || arg == "--batch-size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_batch = std::stoi(argv[i]);
-        } else if (arg == "-ub" || arg == "--ubatch-size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_ubatch = std::stoi(argv[i]);
-        } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            if (llama_supports_gpu_offload()) {
-                params.n_gpu_layers = std::stoi(argv[i]);
-            } else {
-                LOG_WARNING(
-                    "Not compiled with GPU offload support, --n-gpu-layers option will be ignored. "
-                    "See main README.md for information on enabling GPU BLAS support",
-                    {{"n_gpu_layers", params.n_gpu_layers}});
-            }
-        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
-            params.no_kv_offload = true;
-        } else if (arg == "--split-mode" || arg == "-sm") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::string arg_next = argv[i];
-            if (arg_next == "none") {
-                params.split_mode = LLAMA_SPLIT_MODE_NONE;
-            } else if (arg_next == "layer") {
-                params.split_mode = LLAMA_SPLIT_MODE_LAYER;
-            } else if (arg_next == "row") {
-                params.split_mode = LLAMA_SPLIT_MODE_ROW;
-            } else {
-                invalid_param = true;
-                break;
-            }
-#ifndef GGML_USE_CUDA
-            fprintf(stderr, "warning: llama.cpp was compiled without CUDA. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUDA
-        } else if (arg == "--tensor-split" || arg == "-ts") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-#if defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)
-            std::string arg_next = argv[i];
-
-            // split string by , and /
-            const std::regex regex{R"([,/]+)"};
-            std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
-            std::vector split_arg{it, {}};
-            GGML_ASSERT(split_arg.size() <= llama_max_devices());
-
-            for (size_t i_device = 0; i_device < llama_max_devices(); ++i_device) {
-                if (i_device < split_arg.size()) {
-                    params.tensor_split[i_device] = std::stof(split_arg[i_device]);
-                } else {
-                    params.tensor_split[i_device] = 0.0f;
-                }
-            }
-#else
-            LOG_WARNING("llama.cpp was compiled without CUDA. It is not possible to set a tensor split.\n", {});
-#endif // GGML_USE_CUDA
-        } else if (arg == "--main-gpu" || arg == "-mg") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-#if defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)
-            params.main_gpu = std::stoi(argv[i]);
-#else
-            LOG_WARNING("llama.cpp was compiled without CUDA. It is not possible to set a main GPU.", {});
-#endif
-        } else if (arg == "--lora") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.lora_adapter.emplace_back(argv[i], 1.0f);
-            params.use_mmap = false;
-        } else if (arg == "--lora-scaled") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            const char * lora_adapter = argv[i];
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
-            params.use_mmap = false;
-        } else if (arg == "--lora-base") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.lora_base = argv[i];
-        } else if (arg == "-v" || arg == "--verbose") {
-#if SERVER_VERBOSE != 1
-            LOG_WARNING("server.cpp is not built with verbose logging.", {});
-#else
-            server_verbose = true;
-#endif
-        } else if (arg == "--mlock") {
-            params.use_mlock = true;
-        } else if (arg == "--no-mmap") {
-            params.use_mmap = false;
-        } else if (arg == "--numa") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            } else {
-                std::string value(argv[i]);
-                /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
-                else if (value == "isolate")                    { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
-                else if (value == "numactl")                    { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
-                else { invalid_param = true; break; }
-            }
-        } else if (arg == "--embedding" || arg == "--embeddings") {
-            params.embedding = true;
-        } else if (arg == "-cb" || arg == "--cont-batching") {
-            params.cont_batching = true;
-        } else if (arg == "-fa" || arg == "--flash-attn") {
-            params.flash_attn = true;
-        } else if (arg == "-np" || arg == "--parallel") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_parallel = std::stoi(argv[i]);
-        } else if (arg == "-n" || arg == "--n-predict") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_predict = std::stoi(argv[i]);
-        } else if (arg == "-spf" || arg == "--system-prompt-file") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::ifstream file(argv[i]);
-            if (!file) {
-                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
-                invalid_param = true;
-                break;
-            }
-            std::string system_prompt;
-            std::copy(
-                std::istreambuf_iterator(file),
-                std::istreambuf_iterator(),
-                std::back_inserter(system_prompt)
-            );
-            sparams.system_prompt = system_prompt;
-        } else if (arg == "-ctk" || arg == "--cache-type-k") {
-            params.cache_type_k = argv[++i];
-        } else if (arg == "-ctv" || arg == "--cache-type-v") {
-            params.cache_type_v = argv[++i];
-        } else if (arg == "--log-format") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            if (std::strcmp(argv[i], "json") == 0) {
-                server_log_json = true;
-            } else if (std::strcmp(argv[i], "text") == 0) {
-                server_log_json = false;
-            } else {
-                invalid_param = true;
-                break;
-            }
-        } else if (arg == "--log-disable") {
-            log_set_target(stdout);
-            LOG_INFO("logging to file is disabled.", {});
-        } else if (arg == "--slots-endpoint-disable") {
-            sparams.slots_endpoint = false;
-        } else if (arg == "--metrics") {
-            sparams.metrics_endpoint = true;
-        } else if (arg == "--slot-save-path") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.slot_save_path = argv[i];
-            // if doesn't end with DIRECTORY_SEPARATOR, add it
-            if (!sparams.slot_save_path.empty() && sparams.slot_save_path[sparams.slot_save_path.size() - 1] != DIRECTORY_SEPARATOR) {
-                sparams.slot_save_path += DIRECTORY_SEPARATOR;
-            }
-        } else if (arg == "--chat-template") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            if (!verify_custom_template(argv[i])) {
-                fprintf(stderr, "error: the supplied chat template is not supported: %s\n", argv[i]);
-                fprintf(stderr, "note: llama.cpp does not use jinja parser, we only support commonly used templates\n");
-                invalid_param = true;
-                break;
-            }
-            sparams.chat_template = argv[i];
-        } else if (arg == "--override-kv") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
-                fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
-                invalid_param = true;
-                break;
-            }
-        } else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            server_print_usage(argv[0], default_params, default_sparams);
-            exit(1);
-        }
-    }
-
-    gpt_params_handle_model_default(params);
-
-    if (!params.kv_overrides.empty()) {
-        params.kv_overrides.emplace_back();
-        params.kv_overrides.back().key[0] = 0;
-    }
-
-    if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        server_print_usage(argv[0], default_params, default_sparams);
-        exit(1);
-    }
-}
-
 static void log_server_request(const httplib::Request & req, const httplib::Response & res) {
     // skip GH copilot requests when using default port
     if (req.path == "/v1/health" || req.path == "/v1/completions") {
@@ -2918,16 +2503,22 @@ int main(int argc, char ** argv) {
     log_disable();
 #endif
     // own arguments required by this example
-    gpt_params    params;
-    server_params sparams;
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
+        return 1;
+    }
+
+    // TODO: not great to use extern vars
+    server_log_json = params.log_json;
+    server_verbose = params.verbosity > 0;
 
     // struct that contains llama context and inference
     server_context ctx_server;
 
-    server_params_parse(argc, argv, sparams, params);
-
-    if (!sparams.system_prompt.empty()) {
-        ctx_server.system_prompt_set(sparams.system_prompt);
+    if (!params.system_prompt.empty()) {
+        ctx_server.system_prompt_set(params.system_prompt);
     }
 
     if (params.model_alias == "unknown") {
@@ -2951,10 +2542,10 @@ int main(int argc, char ** argv) {
 
     std::unique_ptr svr;
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
-    if (sparams.ssl_key_file != "" && sparams.ssl_cert_file != "") {
-        LOG_INFO("Running with SSL", {{"key", sparams.ssl_key_file}, {"cert", sparams.ssl_cert_file}});
+    if (params.ssl_file_key != "" && params.ssl_file_cert != "") {
+        LOG_INFO("Running with SSL", {{"key", params.ssl_file_key}, {"cert", params.ssl_file_cert}});
         svr.reset(
-            new httplib::SSLServer(sparams.ssl_cert_file.c_str(), sparams.ssl_key_file.c_str())
+            new httplib::SSLServer(params.ssl_file_cert.c_str(), params.ssl_file_key.c_str())
         );
     } else {
         LOG_INFO("Running without SSL", {});
@@ -3008,26 +2599,29 @@ int main(int argc, char ** argv) {
     });
 
     // set timeouts and change hostname and port
-    svr->set_read_timeout (sparams.read_timeout);
-    svr->set_write_timeout(sparams.write_timeout);
+    svr->set_read_timeout (params.timeout_read);
+    svr->set_write_timeout(params.timeout_write);
 
-    if (!svr->bind_to_port(sparams.hostname, sparams.port)) {
-        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
+    if (!svr->bind_to_port(params.hostname, params.port)) {
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", params.hostname.c_str(), params.port);
         return 1;
     }
 
     std::unordered_map log_data;
 
-    log_data["hostname"] = sparams.hostname;
-    log_data["port"]     = std::to_string(sparams.port);
+    log_data["hostname"] = params.hostname;
+    log_data["port"]     = std::to_string(params.port);
 
-    if (sparams.api_keys.size() == 1) {
-        auto key = sparams.api_keys[0];
+    if (params.api_keys.size() == 1) {
+        auto key = params.api_keys[0];
         log_data["api_key"] = "api_key: ****" + key.substr(std::max((int)(key.length() - 4), 0));
-    } else if (sparams.api_keys.size() > 1) {
-        log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
+    } else if (params.api_keys.size() > 1) {
+        log_data["api_key"] = "api_key: " + std::to_string(params.api_keys.size()) + " keys loaded";
     }
 
+    // Necessary similarity of prompt for slot selection
+    ctx_server.slot_prompt_similarity = params.slot_prompt_similarity;
+
     // load the model
     if (!ctx_server.load_model(params)) {
         state.store(SERVER_STATE_ERROR);
@@ -3042,26 +2636,18 @@ int main(int argc, char ** argv) {
     const auto model_meta = ctx_server.model_meta();
 
     // if a custom chat template is not supplied, we will use the one that comes with the model (if any)
-    if (sparams.chat_template.empty()) {
+    if (params.chat_template.empty()) {
         if (!ctx_server.validate_model_chat_template()) {
-            LOG_ERROR("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
-            sparams.chat_template = "chatml";
+            LOG_WARNING("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
+            params.chat_template = "chatml";
         }
     }
 
     // print sample chat example to make it clear which template is used
     {
-        json chat;
-        chat.push_back({{"role", "system"},    {"content", "You are a helpful assistant"}});
-        chat.push_back({{"role", "user"},      {"content", "Hello"}});
-        chat.push_back({{"role", "assistant"}, {"content", "Hi there"}});
-        chat.push_back({{"role", "user"},      {"content", "How are you?"}});
-
-        const std::string chat_example = format_chat(ctx_server.model, sparams.chat_template, chat);
-
         LOG_INFO("chat template", {
-            {"chat_example", chat_example},
-            {"built_in", sparams.chat_template.empty()},
+            {"chat_example", llama_chat_format_example(ctx_server.model, params.chat_template)},
+            {"built_in",     params.chat_template.empty()},
         });
     }
 
@@ -3069,7 +2655,7 @@ int main(int argc, char ** argv) {
     // Middlewares
     //
 
-    auto middleware_validate_api_key = [&sparams, &res_error](const httplib::Request & req, httplib::Response & res) {
+    auto middleware_validate_api_key = [¶ms, &res_error](const httplib::Request & req, httplib::Response & res) {
         // TODO: should we apply API key to all endpoints, including "/health" and "/models"?
         static const std::set protected_endpoints = {
             "/props",
@@ -3087,7 +2673,7 @@ int main(int argc, char ** argv) {
         };
 
         // If API key is not set, skip validation
-        if (sparams.api_keys.empty()) {
+        if (params.api_keys.empty()) {
             return true;
         }
 
@@ -3102,7 +2688,7 @@ int main(int argc, char ** argv) {
         std::string prefix = "Bearer ";
         if (auth_header.substr(0, prefix.size()) == prefix) {
             std::string received_api_key = auth_header.substr(prefix.size());
-            if (std::find(sparams.api_keys.begin(), sparams.api_keys.end(), received_api_key) != sparams.api_keys.end()) {
+            if (std::find(params.api_keys.begin(), params.api_keys.end(), received_api_key) != params.api_keys.end()) {
                 return true; // API key is valid
             }
         }
@@ -3157,7 +2743,7 @@ int main(int argc, char ** argv) {
                     };
 
                     res.status = 200; // HTTP OK
-                    if (sparams.slots_endpoint && req.has_param("include_slots")) {
+                    if (params.endpoint_slots && req.has_param("include_slots")) {
                         health["slots"] = result.data.at("slots");
                     }
 
@@ -3183,7 +2769,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_slots = [&](const httplib::Request &, httplib::Response & res) {
-        if (!sparams.slots_endpoint) {
+        if (!params.endpoint_slots) {
             res_error(res, format_error_response("This server does not support slots endpoint.", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -3207,7 +2793,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_metrics = [&](const httplib::Request &, httplib::Response & res) {
-        if (!sparams.metrics_endpoint) {
+        if (!params.endpoint_metrics) {
             res_error(res, format_error_response("This server does not support metrics endpoint.", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -3307,14 +2893,14 @@ int main(int argc, char ** argv) {
         res.status = 200; // HTTP OK
     };
 
-    const auto handle_slots_save = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
+    const auto handle_slots_save = [&ctx_server, &res_error, ¶ms](const httplib::Request & req, httplib::Response & res, int id_slot) {
         json request_data = json::parse(req.body);
         std::string filename = request_data.at("filename");
         if (!fs_validate_filename(filename)) {
             res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
-        std::string filepath = sparams.slot_save_path + filename;
+        std::string filepath = params.slot_save_path + filename;
 
         server_task task;
         task.type = SERVER_TASK_TYPE_SLOT_SAVE;
@@ -3337,14 +2923,14 @@ int main(int argc, char ** argv) {
         }
     };
 
-    const auto handle_slots_restore = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
+    const auto handle_slots_restore = [&ctx_server, &res_error, ¶ms](const httplib::Request & req, httplib::Response & res, int id_slot) {
         json request_data = json::parse(req.body);
         std::string filename = request_data.at("filename");
         if (!fs_validate_filename(filename)) {
             res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
-        std::string filepath = sparams.slot_save_path + filename;
+        std::string filepath = params.slot_save_path + filename;
 
         server_task task;
         task.type = SERVER_TASK_TYPE_SLOT_RESTORE;
@@ -3414,17 +3000,31 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_props = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
+        std::string template_key = "tokenizer.chat_template", curr_tmpl;
+        int32_t tlen = llama_model_meta_val_str(ctx_server.model, template_key.c_str(), nullptr, 0);
+        if (tlen > 0) {
+            std::vector curr_tmpl_buf(tlen + 1, 0);
+            if (llama_model_meta_val_str(ctx_server.model, template_key.c_str(), curr_tmpl_buf.data(), curr_tmpl_buf.size()) == tlen) {
+                curr_tmpl = std::string(curr_tmpl_buf.data(), tlen);
+            }
+        }
         res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
         json data = {
             { "system_prompt",               ctx_server.system_prompt.c_str() },
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
-            { "total_slots",                 ctx_server.params.n_parallel }
+            { "total_slots",                 ctx_server.params.n_parallel },
+            { "chat_template",               curr_tmpl.c_str() }
         };
 
         res.set_content(data.dump(), "application/json; charset=utf-8");
     };
 
     const auto handle_completions = [&ctx_server, &res_error](const httplib::Request & req, httplib::Response & res) {
+        if (ctx_server.params.embedding) {
+            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
+            return;
+        }
+
         res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
 
         json data = json::parse(req.body);
@@ -3519,9 +3119,14 @@ int main(int argc, char ** argv) {
         res.set_content(models.dump(), "application/json; charset=utf-8");
     };
 
-    const auto handle_chat_completions = [&ctx_server, &sparams, &res_error](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_chat_completions = [&ctx_server, ¶ms, &res_error](const httplib::Request & req, httplib::Response & res) {
+        if (ctx_server.params.embedding) {
+            res_error(res, format_error_response("This server does not support chat completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
+            return;
+        }
+
         res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
-        json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), sparams.chat_template);
+        json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
 
         const int id_task = ctx_server.queue_tasks.get_new_id();
 
@@ -3592,6 +3197,11 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_infill = [&ctx_server, &res_error](const httplib::Request & req, httplib::Response & res) {
+        if (ctx_server.params.embedding) {
+            res_error(res, format_error_response("This server does not support infill. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
+            return;
+        }
+
         res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
 
         json data = json::parse(req.body);
@@ -3678,13 +3288,8 @@ int main(int argc, char ** argv) {
         return res.set_content(data.dump(), "application/json; charset=utf-8");
     };
 
-    const auto handle_embeddings = [¶ms, &ctx_server, &res_error](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_embeddings = [&ctx_server, &res_error](const httplib::Request & req, httplib::Response & res) {
         res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
-        if (!params.embedding) {
-            res.status = 501;
-            res.set_content("This server does not support embeddings. Start it with `--embeddings`", "text/plain; charset=utf-8");
-            return;
-        }
 
         const json body = json::parse(req.body);
         bool is_openai = false;
@@ -3734,6 +3339,55 @@ int main(int argc, char ** argv) {
         return res.set_content(root.dump(), "application/json; charset=utf-8");
     };
 
+    const auto handle_lora_adapters_list = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        json result = json::array();
+        for (size_t i = 0; i < ctx_server.lora_adapters.size(); ++i) {
+            auto & la = ctx_server.lora_adapters[i];
+            result.push_back({
+                {"id", i},
+                {"path", la.path},
+                {"scale", la.scale},
+            });
+        }
+        res.set_content(result.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
+    const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+
+        const std::vector body = json::parse(req.body);
+        int max_idx = ctx_server.lora_adapters.size();
+
+        // clear existing value
+        for (auto & la : ctx_server.lora_adapters) {
+            la.scale = 0.0f;
+        }
+
+        // set value
+        for (auto entry : body) {
+            int id      = entry.at("id");
+            float scale = entry.at("scale");
+            if (0 <= id && id < max_idx) {
+                ctx_server.lora_adapters[id].scale = scale;
+            } else {
+                throw std::runtime_error("invalid adapter id");
+            }
+        }
+
+        server_task task;
+        task.type = SERVER_TASK_TYPE_SET_LORA;
+        const int id_task = ctx_server.queue_tasks.post(task);
+        ctx_server.queue_results.add_waiting_task_id(id_task);
+
+        server_task_result result = ctx_server.queue_results.recv(id_task);
+        ctx_server.queue_results.remove_waiting_task_id(id_task);
+
+        res.set_content(result.data.dump(), "application/json");
+        res.status = 200; // HTTP OK
+    };
+
     auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
         return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
             res.set_content(reinterpret_cast(content), len, mime_type);
@@ -3746,21 +3400,32 @@ int main(int argc, char ** argv) {
     //
 
     // register static assets routes
-    if (!sparams.public_path.empty()) {
+    if (!params.public_path.empty()) {
         // Set the base directory for serving static files
-        svr->set_base_dir(sparams.public_path);
+        svr->set_base_dir(params.public_path);
     }
 
     // using embedded static files
-    svr->Get("/", handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
-    svr->Get("/index.js", handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8"));
-    svr->Get("/completion.js", handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
-    svr->Get("/json-schema-to-grammar.mjs", handle_static_file(
-        json_schema_to_grammar_mjs, json_schema_to_grammar_mjs_len, "text/javascript; charset=utf-8"));
+    svr->Get("/",                           handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
+    svr->Get("/index.js",                   handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8"));
+    svr->Get("/completion.js",              handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
+    svr->Get("/json-schema-to-grammar.mjs", handle_static_file(json_schema_to_grammar_mjs, json_schema_to_grammar_mjs_len, "text/javascript; charset=utf-8"));
+
+    // add new-ui files
+    svr->Get("/colorthemes.css",       handle_static_file(colorthemes_css, colorthemes_css_len, "text/css; charset=utf-8"));
+    svr->Get("/style.css",             handle_static_file(style_css, style_css_len, "text/css; charset=utf-8"));
+    svr->Get("/theme-beeninorder.css", handle_static_file(theme_beeninorder_css, theme_beeninorder_css_len, "text/css; charset=utf-8"));
+    svr->Get("/theme-ketivah.css",     handle_static_file(theme_ketivah_css, theme_ketivah_css_len, "text/css; charset=utf-8"));
+    svr->Get("/theme-mangotango.css",  handle_static_file(theme_mangotango_css, theme_mangotango_css_len, "text/css; charset=utf-8"));
+    svr->Get("/theme-playground.css",  handle_static_file(theme_playground_css, theme_playground_css_len, "text/css; charset=utf-8"));
+    svr->Get("/theme-polarnight.css",  handle_static_file(theme_polarnight_css, theme_polarnight_css_len, "text/css; charset=utf-8"));
+    svr->Get("/theme-snowstorm.css",   handle_static_file(theme_snowstorm_css, theme_snowstorm_css_len, "text/css; charset=utf-8"));
+    svr->Get("/index-new.html",        handle_static_file(index_new_html, index_new_html_len, "text/html; charset=utf-8"));
+    svr->Get("/system-prompts.js",     handle_static_file(system_prompts_js, system_prompts_js_len, "text/javascript; charset=utf-8"));
+    svr->Get("/prompt-formats.js",     handle_static_file(prompt_formats_js, prompt_formats_js_len, "text/javascript; charset=utf-8"));
 
     // register API routes
     svr->Get ("/health",              handle_health);
-    svr->Get ("/slots",               handle_slots);
     svr->Get ("/metrics",             handle_metrics);
     svr->Get ("/props",               handle_props);
     svr->Get ("/v1/models",           handle_models);
@@ -3775,7 +3440,12 @@ int main(int argc, char ** argv) {
     svr->Post("/v1/embeddings",       handle_embeddings);
     svr->Post("/tokenize",            handle_tokenize);
     svr->Post("/detokenize",          handle_detokenize);
-    if (!sparams.slot_save_path.empty()) {
+    // LoRA adapters hotswap
+    svr->Get ("/lora-adapters",       handle_lora_adapters_list);
+    svr->Post("/lora-adapters",       handle_lora_adapters_apply);
+    // Save & load slots
+    svr->Get ("/slots",               handle_slots);
+    if (!params.slot_save_path.empty()) {
         // only enable slot endpoints if slot_save_path is set
         svr->Post("/slots/:id_slot",  handle_slots_action);
     }
@@ -3783,12 +3453,12 @@ int main(int argc, char ** argv) {
     //
     // Start the server
     //
-    if (sparams.n_threads_http < 1) {
+    if (params.n_threads_http < 1) {
         // +2 threads for monitoring endpoints
-        sparams.n_threads_http = std::max(params.n_parallel + 2, (int32_t) std::thread::hardware_concurrency() - 1);
+        params.n_threads_http = std::max(params.n_parallel + 2, (int32_t) std::thread::hardware_concurrency() - 1);
     }
-    log_data["n_threads_http"] =  std::to_string(sparams.n_threads_http);
-    svr->new_task_queue = [&sparams] { return new httplib::ThreadPool(sparams.n_threads_http); };
+    log_data["n_threads_http"] =  std::to_string(params.n_threads_http);
+    svr->new_task_queue = [¶ms] { return new httplib::ThreadPool(params.n_threads_http); };
 
     LOG_INFO("HTTP server listening", log_data);
 
diff --git a/llm/llama.cpp/examples/server/tests/README.md b/llm/llama.cpp/examples/server/tests/README.md
index 83c0208f36c09d3b6024aa46c511f29c15f24674..5e6cb277bc81366886f20b396c1bf7238bacb286 100644
--- a/llm/llama.cpp/examples/server/tests/README.md
+++ b/llm/llama.cpp/examples/server/tests/README.md
@@ -27,10 +27,8 @@ To mitigate it, you can increase values in `n_predict`, `kv_size`.
 
 ```shell
 cd ../../..
-mkdir build
-cd build
-cmake -DLLAMA_CURL=ON ../
-cmake --build . --target server
+cmake -B build -DLLAMA_CURL=ON
+cmake --build build --target llama-server
 ```
 
 2. Start the test: `./tests.sh`
@@ -40,7 +38,7 @@ It's possible to override some scenario steps values with environment variables:
 | variable                 | description                                                                                    |
 |--------------------------|------------------------------------------------------------------------------------------------|
 | `PORT`                   | `context.server_port` to set the listening port of the server during scenario, default: `8080` |
-| `LLAMA_SERVER_BIN_PATH`  | to change the server binary path, default: `../../../build/bin/server`                         |
+| `LLAMA_SERVER_BIN_PATH`  | to change the server binary path, default: `../../../build/bin/llama-server`                         |
 | `DEBUG`                  | "ON" to enable steps and server verbose mode `--verbose`                                       |
 | `SERVER_LOG_FORMAT_JSON` | if set switch server logs to json format                                                       |
 | `N_GPU_LAYERS`           | number of model layers to offload to VRAM `-ngl --n-gpu-layers`                                |
diff --git a/llm/llama.cpp/examples/server/tests/features/lora.feature b/llm/llama.cpp/examples/server/tests/features/lora.feature
new file mode 100644
index 0000000000000000000000000000000000000000..7b85988ac6e87bdbc709c9878b46cc4de7fc28c2
--- /dev/null
+++ b/llm/llama.cpp/examples/server/tests/features/lora.feature
@@ -0,0 +1,36 @@
+@llama.cpp
+@lora
+Feature: llama.cpp server
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model url https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/stories15M_MOE-F16.gguf
+    And   a model file stories15M_MOE-F16.gguf
+    And   a model alias stories15M_MOE
+    And   a lora adapter file from https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Completion LoRA disabled
+    Given switch off lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching little|girl|three|years|old
+
+  Scenario: Completion LoRA enabled
+    Given switch on lora adapter 0
+    Given a prompt:
+    """
+    Look in thy glass
+    """
+    And   a completion request with no api error
+    Then  64 tokens are predicted matching eye|love|glass|sun
diff --git a/llm/llama.cpp/examples/server/tests/features/passkey.feature b/llm/llama.cpp/examples/server/tests/features/passkey.feature
index 1bde7aab8bab0db67e0f81ef766c0cdc8cb3069e..6a5a84e6a194177ea4b5ad17bfe803cb0ab02f39 100644
--- a/llm/llama.cpp/examples/server/tests/features/passkey.feature
+++ b/llm/llama.cpp/examples/server/tests/features/passkey.feature
@@ -52,4 +52,3 @@ Feature: Passkey / Self-extend with context shift
       #| TheBloke/Llama-2-7B-GGUF        | llama-2-7b.Q2_K.gguf        | 4096        | 3   | 16384 | 512     | 4    | 512    | 500    | 300   | 1234    | 5           | 1234           |
       #| TheBloke/Mixtral-8x7B-v0.1-GGUF | mixtral-8x7b-v0.1.Q2_K.gguf | 32768       | 2   | 16384 | 512     | 4    | 512    | 500    | 100   | 0987    | 5           | 0
       # 987           |
-
diff --git a/llm/llama.cpp/examples/server/tests/features/server.feature b/llm/llama.cpp/examples/server/tests/features/server.feature
index d21c09135243ae42f0bfa519b0a9c11ab761b2a0..b55971454afc3674e0a65c6eff348ee9736a3604 100644
--- a/llm/llama.cpp/examples/server/tests/features/server.feature
+++ b/llm/llama.cpp/examples/server/tests/features/server.feature
@@ -82,7 +82,7 @@ Feature: llama.cpp server
 
     Examples: Prompts
       | response_format                                                     | n_predicted | re_content             |
-      | {"type": "json_object", "schema": {"const": "42"}}                  | 5           | "42"                   |
+      | {"type": "json_object", "schema": {"const": "42"}}                  | 6           | "42"                   |
       | {"type": "json_object", "schema": {"items": [{"type": "integer"}]}} | 10          | \[ -300 \]             |
       | {"type": "json_object"}                                             | 10          | \{ " Jacky.            |
 
diff --git a/llm/llama.cpp/examples/server/tests/features/steps/steps.py b/llm/llama.cpp/examples/server/tests/features/steps/steps.py
index 26d9359d7f3f89138d2b71bdff7ed32eac46f69d..6705a34fc469650b64f3a4b911af9adc355db644 100644
--- a/llm/llama.cpp/examples/server/tests/features/steps/steps.py
+++ b/llm/llama.cpp/examples/server/tests/features/steps/steps.py
@@ -1,5 +1,4 @@
 import asyncio
-import collections
 import json
 import os
 import re
@@ -8,19 +7,24 @@ import subprocess
 import sys
 import threading
 import time
+import requests
+from collections.abc import Sequence
 from contextlib import closing
 from re import RegexFlag
+from typing import Any, Literal, cast
 
 import aiohttp
 import numpy as np
 import openai
-from behave import step
+from openai.types.chat import ChatCompletionChunk
+from behave import step  # pyright: ignore[reportAttributeAccessIssue]
 from behave.api.async_step import async_run_until_complete
 from prometheus_client import parser
 
+# pyright: reportRedeclaration=false
 
 @step("a server listening on {server_fqdn}:{server_port}")
-def step_server_config(context, server_fqdn, server_port):
+def step_server_config(context, server_fqdn: str, server_port: str):
     context.server_fqdn = server_fqdn
     context.server_port = int(server_port)
     context.n_threads = None
@@ -67,6 +71,7 @@ def step_server_config(context, server_fqdn, server_port):
     context.user_api_key = None
     context.response_format = None
     context.temperature = None
+    context.lora_file = None
 
     context.tasks_result = []
     context.concurrent_tasks = []
@@ -74,34 +79,40 @@ def step_server_config(context, server_fqdn, server_port):
 
 
 @step('a model file {hf_file} from HF repo {hf_repo}')
-def step_download_hf_model(context, hf_file, hf_repo):
+def step_download_hf_model(context, hf_file: str, hf_repo: str):
     context.model_hf_repo = hf_repo
     context.model_hf_file = hf_file
     context.model_file = os.path.basename(hf_file)
 
+@step('a lora adapter file from {lora_file_url}')
+def step_download_lora_file(context, lora_file_url: str):
+    file_name = lora_file_url.split('/').pop()
+    context.lora_file = f'../../../{file_name}'
+    with open(context.lora_file, 'wb') as f:
+        f.write(requests.get(lora_file_url).content)
 
 @step('a model file {model_file}')
-def step_model_file(context, model_file):
+def step_model_file(context, model_file: str):
     context.model_file = model_file
 
 
 @step('a model url {model_url}')
-def step_model_url(context, model_url):
+def step_model_url(context, model_url: str):
     context.model_url = model_url
 
 
 @step('a model alias {model_alias}')
-def step_model_alias(context, model_alias):
+def step_model_alias(context, model_alias: str):
     context.model_alias = model_alias
 
 
 @step('{seed:d} as server seed')
-def step_seed(context, seed):
+def step_seed(context, seed: int):
     context.server_seed = seed
 
 
 @step('{ngl:d} GPU offloaded layers')
-def step_n_gpu_layer(context, ngl):
+def step_n_gpu_layer(context, ngl: int):
     if 'N_GPU_LAYERS' in os.environ:
         new_ngl = int(os.environ['N_GPU_LAYERS'])
         if context.debug:
@@ -111,37 +122,37 @@ def step_n_gpu_layer(context, ngl):
 
 
 @step('{n_threads:d} threads')
-def step_n_threads(context, n_threads):
+def step_n_threads(context, n_threads: int):
     context.n_thread = n_threads
 
 
 @step('{draft:d} as draft')
-def step_draft(context, draft):
+def step_draft(context, draft: int):
     context.draft = draft
 
 
 @step('{n_ctx:d} KV cache size')
-def step_n_ctx(context, n_ctx):
+def step_n_ctx(context, n_ctx: int):
     context.n_ctx = n_ctx
 
 
 @step('{n_slots:d} slots')
-def step_n_slots(context, n_slots):
+def step_n_slots(context, n_slots: int):
     context.n_slots = n_slots
 
 
 @step('{n_predict:d} server max tokens to predict')
-def step_server_n_predict(context, n_predict):
+def step_server_n_predict(context, n_predict: int):
     context.n_server_predict = n_predict
 
 
 @step('{slot_save_path} as slot save path')
-def step_slot_save_path(context, slot_save_path):
+def step_slot_save_path(context, slot_save_path: str):
     context.slot_save_path = slot_save_path
 
 
 @step('using slot id {id_slot:d}')
-def step_id_slot(context, id_slot):
+def step_id_slot(context, id_slot: int):
     context.id_slot = id_slot
 
 
@@ -191,7 +202,7 @@ def step_start_server(context):
 
 @step("the server is {expecting_status}")
 @async_run_until_complete
-async def step_wait_for_the_server_to_be_started(context, expecting_status):
+async def step_wait_for_the_server_to_be_started(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
     match expecting_status:
         case 'healthy':
             await wait_for_health_status(context, context.base_url, 200, 'ok',
@@ -221,7 +232,7 @@ async def step_wait_for_the_server_to_be_started(context, expecting_status):
 
 @step('all slots are {expected_slot_status_string}')
 @async_run_until_complete
-async def step_all_slots_status(context, expected_slot_status_string):
+async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str):
     match expected_slot_status_string:
         case 'idle':
             expected_slot_status = 0
@@ -237,7 +248,7 @@ async def step_all_slots_status(context, expected_slot_status_string):
 
 @step('a completion request with {api_error} api error')
 @async_run_until_complete
-async def step_request_completion(context, api_error):
+async def step_request_completion(context, api_error: Literal['raised'] | str):
     expect_api_error = api_error == 'raised'
     seeds = await completions_seed(context, num_seeds=1)
     completion = await request_completion(context.prompts.pop(),
@@ -777,8 +788,8 @@ def step_assert_metric_value(context, metric_name, metric_value):
 def step_available_models(context):
     # openai client always expects an api_key
     openai.api_key = context.user_api_key if context.user_api_key is not None else 'nope'
-    openai.api_base = f'{context.base_url}/v1'
-    context.models = openai.Model.list().data
+    openai.base_url = f'{context.base_url}/v1/'
+    context.models = openai.models.list().data
 
 
 @step('{n_model:d} models are supported')
@@ -789,7 +800,7 @@ def step_supported_models(context, n_model):
 
 
 @step('model {i_model:d} is {param} {preposition} {param_value}')
-def step_supported_models(context, i_model, param, preposition, param_value):
+def step_supported_models(context, i_model: int, param: Literal['identified', 'trained'] | str, preposition: str, param_value: str):
     assert i_model < len(context.models)
     model = context.models[i_model]
 
@@ -798,7 +809,7 @@ def step_supported_models(context, i_model, param, preposition, param_value):
         case 'identified':
             value = model.id
         case 'trained':
-            value = str(model.meta.n_ctx_train)
+            value = str(model.meta["n_ctx_train"])
         case _:
             assert False, "param {param} not supported"
     assert param_value == value, f"model param {param} {value} != {param_value}"
@@ -810,6 +821,7 @@ async def concurrent_requests(context, f_completion, *args, **kwargs):
         print(f"starting {context.n_prompts} concurrent completion requests...")
     assert context.n_prompts > 0
     seeds = await completions_seed(context)
+    assert seeds is not None
     for prompt_no in range(context.n_prompts):
         shifted_args = [context.prompts.pop(), seeds[prompt_no], *args]
         context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
@@ -845,6 +857,17 @@ async def step_erase_slot(context, slot_id):
             context.response = response
 
 
+@step('switch {on_or_off} lora adapter {lora_id:d}')
+@async_run_until_complete
+async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
+    async with aiohttp.ClientSession() as session:
+        async with session.post(f'{context.base_url}/lora-adapters',
+                                json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
+                                headers={"Content-Type": "application/json"}) as response:
+            context.response = response
+            print([{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}])
+
+
 @step('the server responds with status code {status_code:d}')
 def step_server_responds_with_status_code(context, status_code):
     assert context.response.status == status_code
@@ -861,7 +884,7 @@ async def request_completion(prompt,
                              id_slot=None,
                              expect_api_error=None,
                              user_api_key=None,
-                             temperature=None):
+                             temperature=None) -> int | dict[str, Any]:
     if debug:
         print(f"Sending completion request: {prompt}")
     origin = "my.super.domain"
@@ -899,8 +922,8 @@ async def request_completion(prompt,
 async def oai_chat_completions(user_prompt,
                                seed,
                                system_prompt,
-                               base_url,
-                               base_path,
+                               base_url: str,
+                               base_path: str,
                                async_client,
                                debug=False,
                                temperature=None,
@@ -909,7 +932,7 @@ async def oai_chat_completions(user_prompt,
                                enable_streaming=None,
                                response_format=None,
                                user_api_key=None,
-                               expect_api_error=None):
+                               expect_api_error=None) -> int | dict[str, Any]:
     if debug:
         print(f"Sending OAI Chat completions request: {user_prompt}")
     # openai client always expects an api key
@@ -989,32 +1012,35 @@ async def oai_chat_completions(user_prompt,
     else:
         try:
             openai.api_key = user_api_key
-            openai.api_base = f'{base_url}{base_path}'
-            chat_completion = openai.Completion.create(
+            openai.base_url = f'{base_url}{base_path.removesuffix("chat")}'
+            assert model is not None
+            chat_completion = openai.chat.completions.create(
                 messages=payload['messages'],
                 model=model,
                 max_tokens=n_predict,
                 stream=enable_streaming,
-                response_format=payload.get('response_format'),
+                response_format=payload.get('response_format') or openai.NOT_GIVEN,
                 seed=seed,
                 temperature=payload['temperature']
             )
-        except openai.error.AuthenticationError as e:
+        except openai.AuthenticationError as e:
             if expect_api_error is not None and expect_api_error:
                 return 401
             else:
                 assert False, f'error raised: {e}'
 
         if enable_streaming:
+            chat_completion = cast(openai.Stream[ChatCompletionChunk], chat_completion)
             for chunk in chat_completion:
                 assert len(chunk.choices) == 1
                 delta = chunk.choices[0].delta
-                if 'content' in delta:
-                    completion_response['content'] += delta['content']
+                if delta.content is not None:
+                    completion_response['content'] += delta.content
                     completion_response['timings']['predicted_n'] += 1
                 completion_response['truncated'] = chunk.choices[0].finish_reason != 'stop'
         else:
             assert len(chat_completion.choices) == 1
+            assert chat_completion.usage is not None
             completion_response = {
                 'content': chat_completion.choices[0].message.content,
                 'timings': {
@@ -1028,7 +1054,7 @@ async def oai_chat_completions(user_prompt,
     return completion_response
 
 
-async def request_embedding(content, seed, base_url=None):
+async def request_embedding(content, seed, base_url=None) -> list[list[float]]:
     async with aiohttp.ClientSession() as session:
         async with session.post(f'{base_url}/embedding',
                                 json={
@@ -1041,7 +1067,7 @@ async def request_embedding(content, seed, base_url=None):
 
 async def request_oai_embeddings(input, seed,
                                  base_url=None, user_api_key=None,
-                                 model=None, async_client=False):
+                                 model=None, async_client=False) -> list[list[float]]:
     # openai client always expects an api_key
     user_api_key = user_api_key if user_api_key is not None else 'nope'
     if async_client:
@@ -1063,7 +1089,7 @@ async def request_oai_embeddings(input, seed,
                 response_json = await response.json()
                 assert response_json['model'] == model, f"invalid model received: {response_json['model']}"
                 assert response_json['object'] == 'list'
-                if isinstance(input, collections.abc.Sequence):
+                if isinstance(input, Sequence):
                     embeddings = []
                     for an_oai_embeddings in response_json['data']:
                         embeddings.append(an_oai_embeddings['embedding'])
@@ -1072,19 +1098,14 @@ async def request_oai_embeddings(input, seed,
                 return embeddings
     else:
         openai.api_key = user_api_key
-        openai.api_base = f'{base_url}/v1'
-        oai_embeddings = openai.Embedding.create(
+        openai.base_url = f'{base_url}/v1/'
+        assert model is not None
+        oai_embeddings = openai.embeddings.create(
             model=model,
             input=input,
         )
 
-        if isinstance(input, collections.abc.Sequence):
-            embeddings = []
-            for an_oai_embeddings in oai_embeddings.data:
-                embeddings.append(an_oai_embeddings.embedding)
-        else:
-            embeddings = [oai_embeddings.data.embedding]
-        return embeddings
+        return [e.embedding for e in oai_embeddings.data]
 
 
 def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re_content=None):
@@ -1122,7 +1143,7 @@ def assert_all_predictions_equal(completion_responses):
             if i == j:
                 continue
             content_j = response_j['content']
-        assert content_i == content_j, "contents not equal"
+            assert content_i == content_j, "contents not equal"
 
 
 def assert_all_predictions_different(completion_responses):
@@ -1136,7 +1157,7 @@ def assert_all_predictions_different(completion_responses):
             if i == j:
                 continue
             content_j = response_j['content']
-        assert content_i != content_j, "contents not different"
+            assert content_i != content_j, "contents not different"
 
 
 def assert_all_token_probabilities_equal(completion_responses):
@@ -1153,7 +1174,7 @@ def assert_all_token_probabilities_equal(completion_responses):
                 if i == j:
                     continue
                 probs_j = response_j['completion_probabilities'][pos]['probs']
-            assert probs_i == probs_j, "contents not equal"
+                assert probs_i == probs_j, "contents not equal"
 
 
 async def gather_tasks_results(context):
@@ -1272,9 +1293,9 @@ def context_text(context):
 
 def start_server_background(context):
     if os.name == 'nt':
-        context.server_path = '../../../build/bin/Release/server.exe'
+        context.server_path = '../../../build/bin/Release/llama-server.exe'
     else:
-        context.server_path = '../../../build/bin/server'
+        context.server_path = '../../../build/bin/llama-server'
     if 'LLAMA_SERVER_BIN_PATH' in os.environ:
         context.server_path = os.environ['LLAMA_SERVER_BIN_PATH']
     server_listen_addr = context.server_fqdn
@@ -1324,6 +1345,8 @@ def start_server_background(context):
         server_args.extend(['--grp-attn-w', context.n_ga_w])
     if context.debug:
         server_args.append('--verbose')
+    if context.lora_file:
+        server_args.extend(['--lora', context.lora_file])
     if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
         server_args.extend(['--log-format', "text"])
 
@@ -1343,7 +1366,7 @@ def start_server_background(context):
     }
     context.server_process = subprocess.Popen(
         [str(arg) for arg in [context.server_path, *server_args]],
-        **pkwargs)
+        **pkwargs)  # pyright: ignore[reportArgumentType, reportCallIssue]
 
     def server_log(in_stream, out_stream):
         for line in iter(in_stream.readline, b''):
diff --git a/llm/llama.cpp/examples/server/tests/requirements.txt b/llm/llama.cpp/examples/server/tests/requirements.txt
index 2e4f42ad28c233fd2e3cec00f2a33aecb3925f5a..f2d7e5c5731be189acf3b43af329d06b81aba171 100644
--- a/llm/llama.cpp/examples/server/tests/requirements.txt
+++ b/llm/llama.cpp/examples/server/tests/requirements.txt
@@ -1,6 +1,7 @@
 aiohttp~=3.9.3
 behave~=1.2.6
 huggingface_hub~=0.20.3
-numpy~=1.24.4
-openai~=0.25.0
+numpy~=1.26.4
+openai~=1.30.3
 prometheus-client~=0.20.0
+requests~=2.32.3
diff --git a/llm/llama.cpp/examples/server/themes/buttons-top/index.html b/llm/llama.cpp/examples/server/themes/buttons-top/index.html
index 6af30d307a4b5b5ff06111bf77572c5d721a11df..8334bcde5049cce2889f47c99a34e68beaa43e81 100644
--- a/llm/llama.cpp/examples/server/themes/buttons-top/index.html
+++ b/llm/llama.cpp/examples/server/themes/buttons-top/index.html
@@ -1054,4 +1054,3 @@
 
 
 
-
diff --git a/llm/llama.cpp/examples/server/themes/wild/index.html b/llm/llama.cpp/examples/server/themes/wild/index.html
index 772e716cdb2e074edfd5472bcee15ba984c20a42..8361c577494d72338ddc5d3e5ff356788ebe12fc 100644
--- a/llm/llama.cpp/examples/server/themes/wild/index.html
+++ b/llm/llama.cpp/examples/server/themes/wild/index.html
@@ -1058,4 +1058,3 @@
 
 
 
-
diff --git a/llm/llama.cpp/examples/server/utils.hpp b/llm/llama.cpp/examples/server/utils.hpp
index d8a2286e4b1dfbcbf91dce785a7ba22c97ed2a2b..e6a1f069723ec5f33a15325df4688a2e4a175b36 100644
--- a/llm/llama.cpp/examples/server/utils.hpp
+++ b/llm/llama.cpp/examples/server/utils.hpp
@@ -116,45 +116,37 @@ static inline void server_log(const char * level, const char * function, int lin
 // chat template utils
 //
 
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-inline bool verify_custom_template(const std::string & tmpl) {
-    llama_chat_message chat[] = {{"user", "test"}};
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
-    return res >= 0;
-}
-
 // Format given chat. If tmpl is empty, we take the template from model metadata
 inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector & messages) {
-    size_t alloc_size = 0;
-    // vector holding all allocated string to be passed to llama_chat_apply_template
-    std::vector str(messages.size() * 2);
-    std::vector chat(messages.size());
+    std::vector chat;
 
     for (size_t i = 0; i < messages.size(); ++i) {
         const auto & curr_msg = messages[i];
-        str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
-        str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
-        alloc_size     += str[i*2 + 1].length();
-        chat[i].role    = str[i*2 + 0].c_str();
-        chat[i].content = str[i*2 + 1].c_str();
-    }
 
-    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
-    std::vector buf(alloc_size * 2);
-
-    // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
+        std::string role = json_value(curr_msg, "role", std::string(""));
+
+        std::string content;
+        if (curr_msg.contains("content")) {
+            if (curr_msg["content"].is_string()) {
+                content = curr_msg["content"].get();
+            } else if (curr_msg["content"].is_array()) {
+                for (const auto & part : curr_msg["content"]) {
+                    if (part.contains("text")) {
+                        content += "\n" + part["text"].get();
+                    }
+                }
+            } else {
+                throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
+            }
+        } else {
+            throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
+        }
 
-    // if it turns out that our buffer is too small, we resize it
-    if ((size_t) res > buf.size()) {
-        buf.resize(res);
-        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
+        chat.push_back({role, content});
     }
 
-    const std::string formatted_chat(buf.data(), res);
-
+    auto formatted_chat = llama_chat_apply_template(model, tmpl, chat, true);
     LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
-
     return formatted_chat;
 }
 
@@ -260,6 +252,13 @@ static size_t common_part(const std::vector & a, const std::vector<
     return i;
 }
 
+static size_t common_part(const std::string & a, const std::string & b) {
+    size_t i;
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
+
+    return i;
+}
+
 static bool ends_with(const std::string & str, const std::string & suffix) {
     return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
 }
@@ -356,24 +355,6 @@ static json oaicompat_completion_params_parse(
 
     llama_params["__oaicompat"] = true;
 
-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
-    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
-    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
-    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
-    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
-    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
-    llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
-    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
-
     // Apply chat template to the list of messages
     llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));
 
diff --git a/llm/llama.cpp/examples/server_embd.py b/llm/llama.cpp/examples/server_embd.py
new file mode 100644
index 0000000000000000000000000000000000000000..0e34c6ceab9cabdcb04d4b562ca1d4589fd6a8d7
--- /dev/null
+++ b/llm/llama.cpp/examples/server_embd.py
@@ -0,0 +1,35 @@
+import asyncio
+import asyncio.threads
+import requests
+import numpy as np
+
+
+n = 8
+
+result = []
+
+async def requests_post_async(*args, **kwargs):
+    return await asyncio.threads.to_thread(requests.post, *args, **kwargs)
+
+async def main():
+    model_url = "http://127.0.0.1:6900"
+    responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
+        url= f"{model_url}/embedding",
+        json= {"content": str(0)*1024}
+    ) for i in range(n)])
+
+    for response in responses:
+        embedding = response.json()["embedding"]
+        print(embedding[-8:])
+        result.append(embedding)
+
+asyncio.run(main())
+
+# compute cosine similarity
+
+for i in range(n-1):
+    for j in range(i+1, n):
+        embedding1 = np.array(result[i])
+        embedding2 = np.array(result[j])
+        similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))
+        print(f"Similarity between {i} and {j}: {similarity:.2f}")
diff --git a/llm/llama.cpp/examples/simple/CMakeLists.txt b/llm/llama.cpp/examples/simple/CMakeLists.txt
index 7da5ff6f3ac041e4a737c63c0613b64616f72e3a..070cfbe7ad525c15d425222ac282057bea390422 100644
--- a/llm/llama.cpp/examples/simple/CMakeLists.txt
+++ b/llm/llama.cpp/examples/simple/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET simple)
+set(TARGET llama-simple)
 add_executable(${TARGET} simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/simple/README.md b/llm/llama.cpp/examples/simple/README.md
index 5d24b1046935c087bd1afbebfde3011a4589a1ce..0ff3425359a415c70a648e7e882304a22f9a5350 100644
--- a/llm/llama.cpp/examples/simple/README.md
+++ b/llm/llama.cpp/examples/simple/README.md
@@ -3,7 +3,7 @@
 The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
 
 ```bash
-./simple ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is"
+./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
 
 ...
 
diff --git a/llm/llama.cpp/examples/simple/simple.cpp b/llm/llama.cpp/examples/simple/simple.cpp
index b0f8e0fdc49873b76f0b4f73423d765b59a3e8db..69a92cf7dc0c01ab0010d18420ea097659392733 100644
--- a/llm/llama.cpp/examples/simple/simple.cpp
+++ b/llm/llama.cpp/examples/simple/simple.cpp
@@ -6,28 +6,27 @@
 #include 
 #include 
 
-int main(int argc, char ** argv) {
-    gpt_params params;
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
 
-    if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [PROMPT]\n" , argv[0]);
-        return 1 ;
-    }
+    LOG_TEE("\nexample usage:\n");
+    LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
+    LOG_TEE("\n");
+}
 
-    if (argc >= 2) {
-        params.model = argv[1];
-    }
+int main(int argc, char ** argv) {
+    gpt_params params;
 
-    if (argc >= 3) {
-        params.prompt = argv[2];
-    }
+    params.prompt = "Hello my name is";
+    params.n_predict = 32;
 
-    if (params.prompt.empty()) {
-        params.prompt = "Hello my name is";
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
+        return 1;
     }
 
     // total length of the sequence including the prompt
-    const int n_len = 32;
+    const int n_predict = params.n_predict;
 
     // init LLM
 
@@ -36,9 +35,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_default_params();
-
-    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+    llama_model_params model_params = llama_model_params_from_gpt_params(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -49,12 +46,7 @@ int main(int argc, char ** argv) {
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_default_params();
-
-    ctx_params.seed  = 1234;
-    ctx_params.n_ctx = 2048;
-    ctx_params.n_threads = params.n_threads;
-    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
@@ -69,14 +61,14 @@ int main(int argc, char ** argv) {
     tokens_list = ::llama_tokenize(ctx, params.prompt, true);
 
     const int n_ctx    = llama_n_ctx(ctx);
-    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
+    const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size());
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, n_kv_req);
+    LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
         LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
-        LOG_TEE("%s:        either reduce n_len or increase n_ctx\n", __func__);
+        LOG_TEE("%s:        either reduce n_predict or increase n_ctx\n", __func__);
         return 1;
     }
 
@@ -115,7 +107,7 @@ int main(int argc, char ** argv) {
 
     const auto t_main_start = ggml_time_us();
 
-    while (n_cur <= n_len) {
+    while (n_cur <= n_predict) {
         // sample the next token
         {
             auto   n_vocab = llama_n_vocab(model);
@@ -134,7 +126,7 @@ int main(int argc, char ** argv) {
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of generation?
-            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
+            if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
                 LOG_TEE("\n");
 
                 break;
diff --git a/llm/llama.cpp/examples/speculative/CMakeLists.txt b/llm/llama.cpp/examples/speculative/CMakeLists.txt
index 810f3c46ac4aab32234ac7807a168ebc9040742c..aa208e7aaeeb00d8a0a87af477e683e208e5a2ab 100644
--- a/llm/llama.cpp/examples/speculative/CMakeLists.txt
+++ b/llm/llama.cpp/examples/speculative/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET speculative)
+set(TARGET llama-speculative)
 add_executable(${TARGET} speculative.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/speculative/speculative.cpp b/llm/llama.cpp/examples/speculative/speculative.cpp
index 12e46fbc91a242677457f99943d528b476e801a6..b051a18f169c2660013e35037c6e3ce3d86f465e 100644
--- a/llm/llama.cpp/examples/speculative/speculative.cpp
+++ b/llm/llama.cpp/examples/speculative/speculative.cpp
@@ -27,7 +27,8 @@ struct seq_draft {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -65,7 +66,9 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
+    model_tgt = llama_init_tgt.model;
+    ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
     params.model = params.model_draft;
@@ -74,7 +77,9 @@ int main(int argc, char ** argv) {
         params.n_threads = params.n_threads_draft;
     }
     params.n_threads_batch = params.n_threads_batch_draft;
-    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
+    llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
+    model_dft = llama_init_dft.model;
+    ctx_dft = llama_init_dft.context;
 
     const bool vocab_type_tgt = llama_vocab_type(model_tgt);
     LOG("vocab_type tgt: %d\n", vocab_type_tgt);
diff --git a/llm/llama.cpp/examples/sycl/CMakeLists.txt b/llm/llama.cpp/examples/sycl/CMakeLists.txt
index 69cf8932eb5c7fbb2dbfa533e0a7ae144f62464c..e4d5083e6e5023072707695c58c0d9a87e94cf9b 100644
--- a/llm/llama.cpp/examples/sycl/CMakeLists.txt
+++ b/llm/llama.cpp/examples/sycl/CMakeLists.txt
@@ -2,7 +2,7 @@
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT
 
-set(TARGET ls-sycl-device)
+set(TARGET llama-ls-sycl-device)
 add_executable(${TARGET} ls-sycl-device.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/sycl/README.md b/llm/llama.cpp/examples/sycl/README.md
index c589c2d3a489ce197f25d99d18164a0886e0b2d8..0e3acd35b616afd2a90c133ad445246540ec66ef 100644
--- a/llm/llama.cpp/examples/sycl/README.md
+++ b/llm/llama.cpp/examples/sycl/README.md
@@ -6,9 +6,9 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
 
 |Tool Name| Function|Status|
 |-|-|-|
-|ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
+|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
 
-### ls-sycl-device
+### llama-ls-sycl-device
 
 List all SYCL devices with ID, compute capability, max work group size, ect.
 
@@ -23,7 +23,7 @@ source /opt/intel/oneapi/setvars.sh
 3. Execute
 
 ```
-./build/bin/ls-sycl-device
+./build/bin/llama-ls-sycl-device
 ```
 
 Check the ID in startup log, like:
diff --git a/llm/llama.cpp/examples/sycl/build.sh b/llm/llama.cpp/examples/sycl/build.sh
new file mode 100644
index 0000000000000000000000000000000000000000..8fe0a67902cbd8835d36ef746fc3274c155329af
--- /dev/null
+++ b/llm/llama.cpp/examples/sycl/build.sh
@@ -0,0 +1,23 @@
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+mkdir -p build
+cd build
+source /opt/intel/oneapi/setvars.sh
+
+#for FP16
+#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON # faster for long-prompt inference
+
+#for FP32
+cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build example/main
+#cmake --build . --config Release --target main
+
+#build example/llama-bench
+#cmake --build . --config Release --target llama-bench
+
+#build all binary
+cmake --build . --config Release -j -v
diff --git a/llm/llama.cpp/examples/sycl/run-llama2.sh b/llm/llama.cpp/examples/sycl/run-llama2.sh
index 7b39a18c0681dca2935faddd813bf21620411fd6..111366fb036a51da6af6cdc63953c8abbe03e6af 100644
--- a/llm/llama.cpp/examples/sycl/run-llama2.sh
+++ b/llm/llama.cpp/examples/sycl/run-llama2.sh
@@ -23,15 +23,14 @@ fi
 if [ $GGML_SYCL_SINGLE_GPU -eq 1 ]; then
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
 fi
 
 #use main GPU only
-#ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
+#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
 
 #use multiple GPUs with same max compute units
-#ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
-
+#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
diff --git a/llm/llama.cpp/examples/sycl/win-build-sycl.bat b/llm/llama.cpp/examples/sycl/win-build-sycl.bat
new file mode 100644
index 0000000000000000000000000000000000000000..17dd1ff5c169ec033e64b4b37aa70e624ed9f73c
--- /dev/null
+++ b/llm/llama.cpp/examples/sycl/win-build-sycl.bat
@@ -0,0 +1,33 @@
+
+::  MIT license
+::  Copyright (C) 2024 Intel Corporation
+::  SPDX-License-Identifier: MIT
+
+
+IF not exist build (mkdir build)
+cd build
+if %errorlevel% neq 0 goto ERROR
+
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+if %errorlevel% neq 0 goto ERROR
+
+::  for FP16
+::  faster for long-prompt inference
+::  cmake -G "MinGW Makefiles" ..  -DGGML_SYCL=ON -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_SYCL_F16=ON
+
+::  for FP32
+cmake -G "Ninja" ..  -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
+if %errorlevel% neq 0 goto ERROR
+::  build example/main only
+::  make main
+
+::  build all binary
+cmake --build . -j
+if %errorlevel% neq 0 goto ERROR
+
+cd ..
+exit /B 0
+
+:ERROR
+echo comomand error: %errorlevel%
+exit /B %errorlevel%
diff --git a/llm/llama.cpp/examples/sycl/win-run-llama2.bat b/llm/llama.cpp/examples/sycl/win-run-llama2.bat
new file mode 100644
index 0000000000000000000000000000000000000000..c2918d6dcead644ed99c259b3e3cc2c2597fcda9
--- /dev/null
+++ b/llm/llama.cpp/examples/sycl/win-run-llama2.bat
@@ -0,0 +1,9 @@
+::  MIT license
+::  Copyright (C) 2024 Intel Corporation
+::  SPDX-License-Identifier: MIT
+
+set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+
+
+.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
diff --git a/llm/llama.cpp/examples/tokenize/CMakeLists.txt b/llm/llama.cpp/examples/tokenize/CMakeLists.txt
index 5e6654d7e5988afb2241373b93a324e7c912694b..b704dcae18c52f4baf54d7649a93d97ab10f988b 100644
--- a/llm/llama.cpp/examples/tokenize/CMakeLists.txt
+++ b/llm/llama.cpp/examples/tokenize/CMakeLists.txt
@@ -1,4 +1,4 @@
-set(TARGET tokenize)
+set(TARGET llama-tokenize)
 add_executable(${TARGET} tokenize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
diff --git a/llm/llama.cpp/examples/tokenize/tokenize.cpp b/llm/llama.cpp/examples/tokenize/tokenize.cpp
index 54c9834afb1b930eea0c3a50928d586b3a10c5db..17f5e496153a7bdc296cbaab99f9ade19043b973 100644
--- a/llm/llama.cpp/examples/tokenize/tokenize.cpp
+++ b/llm/llama.cpp/examples/tokenize/tokenize.cpp
@@ -29,7 +29,9 @@ static void print_usage_information(const char * argv0, FILE * stream) {
     fprintf(stream, "    -p PROMPT, --prompt PROMPT           read prompt from the argument.\n");
     fprintf(stream, "    --stdin                              read prompt from standard input.\n");
     fprintf(stream, "    --no-bos                             do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
+    fprintf(stream, "    --no-parse-special                   do not parse control tokens.\n");
     fprintf(stream, "    --log-disable                        disable logs. Makes stderr quiet when loading the model.\n");
+    fprintf(stream, "    --show-count                         print the total number of tokens.\n");
 }
 
 static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) {
@@ -161,7 +163,7 @@ static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) {
                 printf(">");
                 return;
             }
-            GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way.");
+            GGML_ABORT("MultiByteToWideChar() failed in an unexpected way.");
         }
 
         LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr));
@@ -194,7 +196,9 @@ int main(int raw_argc, char ** raw_argv) {
     // variables where to put any arguments we see.
     bool printing_ids = false;
     bool no_bos = false;
+    bool no_parse_special = false;
     bool disable_logging = false;
+    bool show_token_count = false;
     const char * model_path = NULL;
     const char * prompt_path = NULL;
     const char * prompt_arg = NULL;
@@ -227,6 +231,9 @@ int main(int raw_argc, char ** raw_argv) {
         else if (arg == "--no-bos") {
             no_bos = true;
         }
+        else if (arg == "--no-parse-special") {
+            no_parse_special = true;
+        }
         else if (arg == "-p" || arg == "--prompt") {
             if (prompt_set) {
                 fprintf(stderr, "Error: -p or --prompt specified multiple times.\n");
@@ -249,6 +256,9 @@ int main(int raw_argc, char ** raw_argv) {
         else if (arg == "--log-disable") {
             disable_logging = true;
         }
+        else if (arg == "--show-count") {
+            show_token_count = true;
+        }
         else {
             fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str());
             return 1;
@@ -354,9 +364,10 @@ int main(int raw_argc, char ** raw_argv) {
 
     const bool model_wants_add_bos = llama_should_add_bos_token(model);
     const bool add_bos = model_wants_add_bos && !no_bos;
+    const bool parse_special = !no_parse_special;
 
     std::vector tokens;
-    tokens = ::llama_tokenize(model, prompt, add_bos, true);
+    tokens = ::llama_tokenize(model, prompt, add_bos, parse_special);
 
     if (printing_ids) {
         printf("[");
@@ -384,6 +395,9 @@ int main(int raw_argc, char ** raw_argv) {
         printf("]\n");
     }
 
+    if (show_token_count) {
+        printf("Total number of tokens: %ld\n", tokens.size());
+    }
     // silence valgrind
     llama_free(ctx);
     llama_free_model(model);
diff --git a/llm/llama.cpp/flake.lock b/llm/llama.cpp/flake.lock
index fd6e2a5f63d0eef84f3dc6fa33b93611cf70b5ea..c54af88ea21c42e510db216563b5f0ac290df9ad 100644
--- a/llm/llama.cpp/flake.lock
+++ b/llm/llama.cpp/flake.lock
@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1715865404,
-        "narHash": "sha256-/GJvTdTpuDjNn84j82cU6bXztE0MSkdnTWClUCRub78=",
+        "lastModified": 1722555600,
+        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "8dc45382d5206bd292f9c2768b8058a8fd8311d9",
+        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1716509168,
-        "narHash": "sha256-4zSIhSRRIoEBwjbPm3YiGtbd8HDWzFxJjw5DYSDy1n8=",
+        "lastModified": 1722421184,
+        "narHash": "sha256-/DJBI6trCeVnasdjUo9pbnodCLZcFqnVZiLUfqLH4jA=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "bfb7a882678e518398ce9a31a881538679f6f092",
+        "rev": "9f918d616c5321ad374ae6cb5ea89c9e04bf3e58",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1714640452,
-        "narHash": "sha256-QBx10+k6JWz6u7VsohfSw8g8hjdBZEf8CFzXH1/1Z94=",
+        "lastModified": 1722555339,
+        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/50eb7ecf4cd0a5756d7275c8ba36790e5bd53e33.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/50eb7ecf4cd0a5756d7275c8ba36790e5bd53e33.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
       }
     },
     "root": {
diff --git a/llm/llama.cpp/flake.nix b/llm/llama.cpp/flake.nix
index 9cd3756e53e51f57aea517c7b7063375ae8227ec..c69637d111784e890cda1539da31c3cf90c417b0 100644
--- a/llm/llama.cpp/flake.nix
+++ b/llm/llama.cpp/flake.nix
@@ -63,7 +63,7 @@
   # nix-repl> :lf github:ggerganov/llama.cpp
   # Added 13 variables.
   # nix-repl> outputs.apps.x86_64-linux.quantize
-  # { program = "/nix/store/00000000000000000000000000000000-llama.cpp/bin/quantize"; type = "app"; }
+  # { program = "/nix/store/00000000000000000000000000000000-llama.cpp/bin/llama-quantize"; type = "app"; }
   # ```
   outputs =
     { self, flake-parts, ... }@inputs:
@@ -159,7 +159,6 @@
                 windows = config.legacyPackages.llamaPackagesWindows.llama-cpp;
               }
               // lib.optionalAttrs pkgs.stdenv.isLinux {
-                opencl = config.packages.default.override { useOpenCL = true; };
                 cuda = config.legacyPackages.llamaPackagesCuda.llama-cpp;
 
                 mpi-cpu = config.packages.default.override { useMpi = true; };
diff --git a/llm/llama.cpp/ggml/.gitignore b/llm/llama.cpp/ggml/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..c82d8e69295ac87ea0e1585c2fff14396abde4bf
--- /dev/null
+++ b/llm/llama.cpp/ggml/.gitignore
@@ -0,0 +1,2 @@
+src/ggml-vulkan-shaders.hpp
+src/ggml-vulkan-shaders.cpp
diff --git a/llm/llama.cpp/ggml/CMakeLists.txt b/llm/llama.cpp/ggml/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..7fe1661bb96b4c56222060a2e6ce1c99ffddc3ca
--- /dev/null
+++ b/llm/llama.cpp/ggml/CMakeLists.txt
@@ -0,0 +1,254 @@
+cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+project("ggml" C CXX)
+include(CheckIncludeFileCXX)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(GGML_STANDALONE ON)
+
+    set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+
+    # configure project version
+    # TODO
+else()
+    set(GGML_STANDALONE OFF)
+endif()
+
+if (EMSCRIPTEN)
+    set(BUILD_SHARED_LIBS_DEFAULT OFF)
+
+    option(GGML_WASM_SINGLE_FILE "ggml: embed WASM inside the generated ggml.js" ON)
+else()
+    if (MINGW)
+        set(BUILD_SHARED_LIBS_DEFAULT OFF)
+    else()
+        set(BUILD_SHARED_LIBS_DEFAULT ON)
+    endif()
+endif()
+
+option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+
+#
+# option list
+#
+
+# TODO: mark all options as advanced when not GGML_STANDALONE
+
+if (APPLE)
+    set(GGML_METAL_DEFAULT ON)
+    set(GGML_BLAS_DEFAULT ON)
+    set(GGML_BLAS_VENDOR_DEFAULT "Apple")
+else()
+    set(GGML_METAL_DEFAULT OFF)
+    set(GGML_BLAS_DEFAULT OFF)
+    set(GGML_BLAS_VENDOR_DEFAULT "Generic")
+endif()
+
+if (CMAKE_CROSSCOMPILING)
+    set(GGML_NATIVE_DEFAULT OFF)
+else()
+    set(GGML_NATIVE_DEFAULT ON)
+endif()
+
+# general
+option(GGML_STATIC "ggml: static link libraries"         OFF)
+option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
+option(GGML_LTO    "ggml: enable link time optimization" OFF)
+option(GGML_CCACHE "ggml: use ccache if available"       ON)
+
+# debug
+option(GGML_ALL_WARNINGS           "ggml: enable all compiler warnings"                   ON)
+option(GGML_ALL_WARNINGS_3RD_PARTY "ggml: enable all compiler warnings in 3rd party libs" OFF)
+option(GGML_GPROF                  "ggml: enable gprof"                                   OFF)
+
+# build
+option(GGML_FATAL_WARNINGS    "ggml: enable -Werror flag"    OFF)
+
+# sanitizers
+option(GGML_SANITIZE_THREAD    "ggml: enable thread sanitizer"    OFF)
+option(GGML_SANITIZE_ADDRESS   "ggml: enable address sanitizer"   OFF)
+option(GGML_SANITIZE_UNDEFINED "ggml: enable undefined sanitizer" OFF)
+
+# instruction set specific
+if (GGML_NATIVE OR NOT GGML_NATIVE_DEFAULT)
+    set(INS_ENB OFF)
+else()
+    set(INS_ENB ON)
+endif()
+
+option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
+
+option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
+option(GGML_AVX512      "ggml: enable AVX512"           OFF)
+option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
+option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
+option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
+option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
+if (NOT MSVC)
+    option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
+endif()
+option(GGML_LASX        "ggml: enable lasx"             ON)
+option(GGML_LSX         "ggml: enable lsx"              ON)
+option(GGML_SVE         "ggml: enable SVE"              OFF)
+
+if (WIN32)
+    set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version")
+endif()
+
+# ggml core
+set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
+
+# 3rd party libs / backends
+option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)
+option(GGML_BLAS                            "ggml: use BLAS"                                  ${GGML_BLAS_DEFAULT})
+set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
+                                            "ggml: BLAS library vendor")
+option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             OFF)
+
+option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
+option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
+option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
+option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
+option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
+set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
+set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
+option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
+set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
+                                            "ggml: iters./thread per block for Q2_K/Q6_K")
+set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
+                                            "ggml: max. batch size for using peer access")
+option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
+option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
+option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
+option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)
+
+option(GGML_CURL                            "ggml: use libcurl to download model from an URL" OFF)
+option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
+option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
+option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
+option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
+option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
+option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
+option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
+option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
+option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
+option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
+option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
+option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
+option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"                       ${GGML_METAL})
+set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
+                                            "ggml: metal minimum macOS version")
+set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
+option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
+option(GGML_RPC                             "ggml: use RPC"                                   OFF)
+option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
+option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
+set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
+                                            "ggml: sycl target device")
+
+# extra artifacts
+option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
+option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
+
+#
+# dependencies
+#
+
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+
+if (GGML_SYCL)
+    set(CMAKE_CXX_STANDARD 17)
+else()
+    set(CMAKE_CXX_STANDARD 11)
+endif()
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+
+set(THREADS_PREFER_PTHREAD_FLAG ON)
+
+find_package(Threads REQUIRED)
+
+#
+# build the library
+#
+
+add_subdirectory(src)
+
+#
+# tests and examples
+#
+
+if (GGML_BUILD_TESTS)
+    enable_testing()
+    add_subdirectory(tests)
+endif ()
+
+if (GGML_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif ()
+
+#
+# install
+#
+
+include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+# all public headers
+set(GGML_PUBLIC_HEADERS
+    include/ggml.h
+    include/ggml-alloc.h
+    include/ggml-backend.h
+    include/ggml-blas.h
+    include/ggml-cann.h
+    include/ggml-cuda.h
+    include/ggml.h
+    include/ggml-kompute.h
+    include/ggml-metal.h
+    include/ggml-rpc.h
+    include/ggml-sycl.h
+    include/ggml-vulkan.h)
+
+set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
+#if (GGML_METAL)
+#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
+#endif()
+install(TARGETS ggml PUBLIC_HEADER)
+
+if (BUILD_SHARED_LIBS)
+    install(TARGETS ggml LIBRARY)
+endif()
+
+if (GGML_METAL)
+    install(
+        FILES src/ggml-metal.metal
+        PERMISSIONS
+            OWNER_READ
+            OWNER_WRITE
+            GROUP_READ
+            WORLD_READ
+        DESTINATION ${CMAKE_INSTALL_BINDIR})
+
+    if (NOT GGML_METAL_EMBED_LIBRARY)
+        install(
+            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DESTINATION ${CMAKE_INSTALL_BINDIR}
+        )
+    endif()
+endif()
+
+if (GGML_STANDALONE)
+    configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
+        ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        @ONLY)
+
+    install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        DESTINATION share/pkgconfig)
+endif()
diff --git a/llm/llama.cpp/ggml/cmake/FindSIMD.cmake b/llm/llama.cpp/ggml/cmake/FindSIMD.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..5533668ec4ab18ff7261865e43be0100de65e5bf
--- /dev/null
+++ b/llm/llama.cpp/ggml/cmake/FindSIMD.cmake
@@ -0,0 +1,100 @@
+include(CheckCSourceRuns)
+
+set(AVX_CODE "
+    #include 
+    int main()
+    {
+        __m256 a;
+        a = _mm256_set1_ps(0);
+        return 0;
+    }
+")
+
+set(AVX512_CODE "
+    #include 
+    int main()
+    {
+        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0);
+        __m512i b = a;
+        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
+        return 0;
+    }
+")
+
+set(AVX2_CODE "
+    #include 
+    int main()
+    {
+        __m256i a = {0};
+        a = _mm256_abs_epi16(a);
+        __m256i x;
+        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
+        return 0;
+    }
+")
+
+set(FMA_CODE "
+    #include 
+    int main()
+    {
+        __m256 acc = _mm256_setzero_ps();
+        const __m256 d = _mm256_setzero_ps();
+        const __m256 p = _mm256_setzero_ps();
+        acc = _mm256_fmadd_ps( d, p, acc );
+        return 0;
+    }
+")
+
+macro(check_sse type flags)
+    set(__FLAG_I 1)
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+    foreach (__FLAG ${flags})
+        if (NOT ${type}_FOUND)
+            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
+            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
+            if (HAS_${type}_${__FLAG_I})
+                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
+                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
+            endif()
+            math(EXPR __FLAG_I "${__FLAG_I}+1")
+        endif()
+    endforeach()
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    if (NOT ${type}_FOUND)
+        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
+        set(${type}_FLAGS "" CACHE STRING "${type} flags")
+    endif()
+
+    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
+endmacro()
+
+# flags are for MSVC only!
+check_sse("AVX" " ;/arch:AVX")
+if (NOT ${AVX_FOUND})
+    set(GGML_AVX OFF)
+else()
+    set(GGML_AVX ON)
+endif()
+
+check_sse("AVX2" " ;/arch:AVX2")
+check_sse("FMA" " ;/arch:AVX2")
+if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
+    set(GGML_AVX2 OFF)
+else()
+    set(GGML_AVX2 ON)
+endif()
+
+check_sse("AVX512" " ;/arch:AVX512")
+if (NOT ${AVX512_FOUND})
+    set(GGML_AVX512 OFF)
+else()
+    set(GGML_AVX512 ON)
+endif()
diff --git a/llm/llama.cpp/ggml/include/ggml-alloc.h b/llm/llama.cpp/ggml/include/ggml-alloc.h
new file mode 100644
index 0000000000000000000000000000000000000000..434c13b34a929c43e959565acae84b97ae4fd066
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-alloc.h
@@ -0,0 +1,76 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct ggml_backend * ggml_backend_t;
+
+// Tensor allocator
+struct ggml_tallocr {
+    ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};
+
+GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
+
+// Graph allocator
+/*
+  Example usage:
+    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
+
+    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
+    ggml_gallocr_reserve(galloc, build_graph(max_batch));
+
+    // allocate the graph
+    struct ggml_cgraph * graph = build_graph(batch);
+    ggml_gallocr_alloc_graph(galloc, graph);
+
+    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
+
+    // evaluate the graph
+    ggml_backend_graph_compute(backend, graph);
+*/
+
+// special tensor flags for use with the graph allocator:
+//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
+//   ggml_set_output(): output tensors are never freed and never overwritten
+
+typedef struct ggml_gallocr * ggml_gallocr_t;
+
+GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
+GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
+GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
+
+// pre-allocate buffers from a measure graph - does not allocate or modify the graph
+// call with a worst-case graph to avoid buffer reallocations
+// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
+// returns false if the buffer allocation failed
+GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API bool ggml_gallocr_reserve_n(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);
+
+// automatic reallocation if the topology changes when using a single buffer
+// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
+GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+
+GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
+
+// Utils
+// Create a buffer and allocate all the tensors in a ggml_context
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-backend.h b/llm/llama.cpp/ggml/include/ggml-backend.h
new file mode 100644
index 0000000000000000000000000000000000000000..5f3f1e286990e478f96fcf1cbc21b58887d9b22a
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-backend.h
@@ -0,0 +1,238 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend_event * ggml_backend_event_t;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
+    //
+    // Backend buffer
+    //
+
+    // buffer type
+    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
+
+    // buffer
+    enum ggml_backend_buffer_usage {
+        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
+        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
+        GGML_BACKEND_BUFFER_USAGE_COMPUTE = 2,
+    };
+
+    GGML_API           const char *                   ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API           void                           ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API           void *                         ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                         ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API GGML_CALL void                           ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API           bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API           void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API           enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage     (ggml_backend_buffer_t buffer);
+    GGML_API           ggml_backend_buffer_type_t     ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API           void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+
+    //
+    // Backend
+    //
+
+    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
+    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
+    GGML_API void         ggml_backend_free(ggml_backend_t backend);
+
+    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
+    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
+
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
+    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // asynchronous copy
+    // the copy is performed after all the currently queued operations in backend_src
+    // backend_dst will wait for the copy to complete before performing other operations
+    // automatic fallback to sync copy if async is not supported
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // events
+    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
+    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event);
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    // Create a backend buffer from an existing pointer
+    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+
+    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
+    //
+    // Backend registry
+    //
+
+    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
+
+    GGML_API size_t                     ggml_backend_reg_get_count(void);
+    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is backend_name:params (params is optional)
+    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
+    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
+    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
+
+    //
+    // Backend scheduler
+    //
+
+    // The backend scheduler allows for multiple backends to be used together
+    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
+    // The backends are selected based on:
+    // - the backend that supports the operation
+    // - the location of the pre-allocated tensors (e.g. the weights)
+    /*
+      Example usage:
+
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
+        // preferrably to run on the same backend as the buffer
+        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+
+        // initialize buffers from a max size graph (optional)
+        reserve_graph = build_graph(sched, max_batch_size);
+
+        // manually assign nodes to a backend (optional, should not be needed in most cases)
+        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
+
+        ggml_backend_sched_reserve(sched, reserve_graph);
+
+        // compute
+        graph = build_graph(sched);
+        ggml_backend_sched_graph_compute(sched, graph);
+
+        // if there are graph inputs:
+        ggml_backend_sched_reset(sched);
+        ggml_backend_sched_alloc_graph(sched, graph);
+        ggml_backend_tensor_set(input_tensor, ...);
+        ggml_backend_sched_graph_compute(sched, graph);
+    }
+    */
+
+    struct ggml_backend_sched;
+    typedef struct ggml_backend_sched * ggml_backend_sched_t;
+
+    // when ask == true, the scheduler wants to know if the user wants to observe this node
+    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
+    //
+    // when ask == false, the scheduler is passing the node tensor to the user for observation
+    // if the user returns false, the scheduler will cancel the graph compute
+    //
+    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
+
+    // Initialize a backend scheduler
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+
+    // Initialize backend buffers from a measure graph
+    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+
+    GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i);
+
+    // Get the number of splits of the last graph
+    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
+
+    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+
+    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+
+    // Allocate and compute graph on the backend scheduler
+    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
+
+    // Reset all assignments and allocators - must be called before changing the node backends
+    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
+
+    // Set a callback to be called for each resulting node during graph compute
+    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
+
+    //
+    // Utils
+    //
+
+    struct ggml_backend_graph_copy {
+        ggml_backend_buffer_t buffer;
+        struct ggml_context * ctx_allocated;
+        struct ggml_context * ctx_unallocated;
+        struct ggml_cgraph * graph;
+    };
+
+    // Copy a graph to a different backend
+    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
+    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
+
+    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+
+    // Compare the output of two backends
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+
+    // Tensor initialization
+    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
+    GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
+
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-blas.h b/llm/llama.cpp/ggml/include/ggml-blas.h
new file mode 100644
index 0000000000000000000000000000000000000000..f2e37de06f6090cb73e460929e02d105a56eb60f
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-blas.h
@@ -0,0 +1,23 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_blas_init(void);
+
+GGML_API GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend);
+
+// number of threads used for conversion to float
+// for openblas and blis, this will also set the number of threads used for blas operations
+GGML_API GGML_CALL void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-cann.h b/llm/llama.cpp/ggml/include/ggml-cann.h
new file mode 100644
index 0000000000000000000000000000000000000000..ca73211fea73620da988d581241f5d4206d6c20a
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-cann.h
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * @brief Maximum number of CANN devices supported.
+ */
+#define GGML_CANN_MAX_DEVICES 16
+
+/**
+ * @brief Initializes the CANN backend for a specified device.
+ *
+ * This function initializes the CANN backend for the given device.
+ * It verifies the device index, allocates a context, and creates a backend
+ * instance.
+ *
+ * @param device The index of the device to initialize.
+ * @return A pointer to the initialized backend instance, or nullptr on failure.
+ */
+GGML_API GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device);
+
+/**
+ * @brief Checks if a given backend is a CANN backend.
+ *
+ * This function verifies if the provided backend is a CANN backend by comparing
+ * its GUID with the CANN backend's GUID.
+ *
+ * @param backend The backend instance to check.
+ * @return True if the backend is a CANN backend, false otherwise.
+ */
+GGML_API GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend);
+
+/**
+ * @brief Retrieves the CANN buffer type for a specified device.
+ *
+ * This function initializes and returns the buffer type interface associated
+ * with the given device. It ensures thread-safe access using a mutex.
+ *
+ * @param device The device index for which to retrieve the buffer type.
+ * @return A pointer to the buffer type interface for the specified device, or
+ * nullptr if the device index is out of range.
+ */
+GGML_API GGML_CALL ggml_backend_buffer_type_t
+ggml_backend_cann_buffer_type(int32_t device);
+
+/**
+ * @brief Retrieves the number of CANN devices available.
+ *
+ * This function returns the number of CANN devices available based on
+ * information obtained from `ggml_cann_info()`.
+ *
+ * @return The number of CANN devices available.
+ */
+GGML_API GGML_CALL int32_t ggml_backend_cann_get_device_count(void);
+
+/**
+ * @brief Retrieves the description of a specific CANN device.
+ *
+ * This function sets the specified device, retrieves the SoC name,
+ * and writes it into the provided description buffer.
+ *
+ * @param device The device index to retrieve the description for.
+ * @param description Pointer to a buffer where the description will be written.
+ * @param description_size Size of the description buffer.
+ */
+GGML_API GGML_CALL void ggml_backend_cann_get_device_description(
+    int32_t device, char* description, size_t description_size);
+
+/**
+ * @brief Retrieves the memory information of a specific CANN device.
+ *
+ * This function sets the specified device, retrieves the free and total
+ * memory information of the specified type (ACL_HBM_MEM), and stores them
+ * in the provided pointers.
+ *
+ * @param device The device index to retrieve memory information for.
+ * @param free Pointer to a variable where the free memory size will be stored.
+ * @param total Pointer to a variable where the total memory size will be
+ * stored.
+ */
+GGML_API GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device,
+                                                            size_t* free,
+                                                            size_t* total);
+
+/**
+ * @brief Set the logging callback for GGML.
+ *
+ * This function sets the logging callback and user data for logging.
+ *
+ * @param log_callback The logging callback to set.
+ * @param user_data User data to pass to the logging callback.
+ */
+GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback,
+                                                 void* user_data);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-cuda.h b/llm/llama.cpp/ggml/include/ggml-cuda.h
new file mode 100644
index 0000000000000000000000000000000000000000..71bb6dcf07975e95949fed4b41c48cf024faa212
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-cuda.h
@@ -0,0 +1,47 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef GGML_USE_HIPBLAS
+#define GGML_CUDA_NAME "ROCm"
+#define GGML_CUBLAS_NAME "hipBLAS"
+#elif defined(GGML_USE_MUSA)
+#define GGML_CUDA_NAME "MUSA"
+#define GGML_CUBLAS_NAME "muBLAS"
+#else
+#define GGML_CUDA_NAME "CUDA"
+#define GGML_CUBLAS_NAME "cuBLAS"
+#endif
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_CUDA_MAX_DEVICES       16
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
+
+GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
+
+// device buffer
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+
+GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+
+GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+
+GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-kompute.h b/llm/llama.cpp/ggml/include/ggml-kompute.h
new file mode 100644
index 0000000000000000000000000000000000000000..171465456a5b17a6e27ebe326703c260c65a3773
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-kompute.h
@@ -0,0 +1,46 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include 
+#include 
+#include 
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_vk_device {
+    int index;
+    int type; // same as VkPhysicalDeviceType
+    size_t heapSize;
+    const char * name;
+    const char * vendor;
+    int subgroupSize;
+    uint64_t bufferAlignment;
+    uint64_t maxAlloc;
+};
+
+struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
+bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
+bool ggml_vk_has_vulkan(void);
+bool ggml_vk_has_device(void);
+struct ggml_vk_device ggml_vk_current_device(void);
+
+//
+// backend API
+//
+
+// forward declaration
+typedef struct ggml_backend * ggml_backend_t;
+
+GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
+
+GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-metal.h b/llm/llama.cpp/ggml/include/ggml-metal.h
new file mode 100644
index 0000000000000000000000000000000000000000..6c3226c37e0ef48072ddf146d73a392cf1d9fab1
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-metal.h
@@ -0,0 +1,65 @@
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include 
+#include 
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 64
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// backend API
+// user-code should use only these functions
+//
+
+GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
+GGML_API ggml_backend_t ggml_backend_metal_init(void);
+
+GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+
+GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+
+GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+
+// helper to check if the device supports a specific family
+// ideally, the user code should be doing these checks
+// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+
+// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
+GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-rpc.h b/llm/llama.cpp/ggml/include/ggml-rpc.h
new file mode 100644
index 0000000000000000000000000000000000000000..aa144832a6e1e51f73cbfc4f5372a3d0ae32ee54
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-rpc.h
@@ -0,0 +1,24 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_RPC_MAX_SERVERS       16
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+
+GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+
+GGML_API GGML_CALL void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-sycl.h b/llm/llama.cpp/ggml/include/ggml-sycl.h
new file mode 100644
index 0000000000000000000000000000000000000000..43ab1519cd05dff055e57caf0c33f8faf63bcb33
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-sycl.h
@@ -0,0 +1,42 @@
+//
+//  MIT license
+//  Copyright (C) 2024 Intel Corporation
+//  SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#define GGML_SYCL_NAME "SYCL"
+#define GGML_SYCL_MAX_DEVICES 48
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+
+// devide buffer
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+
+GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
+GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
+GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
+GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+
+// SYCL doesn't support registering host memory, keep here for reference
+// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml-vulkan.h b/llm/llama.cpp/ggml/include/ggml-vulkan.h
new file mode 100644
index 0000000000000000000000000000000000000000..af661c2d7d563e83dac6046f2bceac3d063bccb8
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml-vulkan.h
@@ -0,0 +1,29 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_VK_NAME "Vulkan"
+#define GGML_VK_MAX_DEVICES 16
+
+GGML_API void ggml_vk_instance_init(void);
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+
+GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_API GGML_CALL int  ggml_backend_vk_get_device_count(void);
+GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
+GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/include/ggml.h b/llm/llama.cpp/ggml/include/ggml.h
new file mode 100644
index 0000000000000000000000000000000000000000..15602a96df7ad3ef4df675d2a786a51d888c6cdb
--- /dev/null
+++ b/llm/llama.cpp/ggml/include/ggml.h
@@ -0,0 +1,2455 @@
+#pragma once
+
+//
+// GGML Tensor Library
+//
+// This documentation is still a work in progress.
+// If you wish some specific topics to be covered, feel free to drop a comment:
+//
+//   https://github.com/ggerganov/whisper.cpp/issues/40
+//
+// ## Overview
+//
+// This library implements:
+//
+//  - a set of tensor operations
+//  - automatic differentiation
+//  - basic optimization algorithms
+//
+// The aim of this library is to provide a minimalistic approach for various machine learning tasks. This includes,
+// but is not limited to, the following:
+//
+//  - linear regression
+//  - support vector machines
+//  - neural networks
+//
+// The library allows the user to define a certain function using the available tensor operations. This function
+// definition is represented internally via a computation graph. Each tensor operation in the function definition
+// corresponds to a node in the graph. Having the computation graph defined, the user can choose to compute the
+// function's value and/or its gradient with respect to the input variables. Optionally, the function can be optimized
+// using one of the available optimization algorithms.
+//
+// For example, here we define the function: f(x) = a*x^2 + b
+//
+//   {
+//       struct ggml_init_params params = {
+//           .mem_size   = 16*1024*1024,
+//           .mem_buffer = NULL,
+//       };
+//
+//       // memory allocation happens here
+//       struct ggml_context * ctx = ggml_init(params);
+//
+//       struct ggml_tensor * x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+//
+//       ggml_set_param(ctx, x); // x is an input variable
+//
+//       struct ggml_tensor * a  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+//       struct ggml_tensor * b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+//       struct ggml_tensor * x2 = ggml_mul(ctx, x, x);
+//       struct ggml_tensor * f  = ggml_add(ctx, ggml_mul(ctx, a, x2), b);
+//
+//       ...
+//   }
+//
+// Notice that the function definition above does not involve any actual computation. The computation is performed only
+// when the user explicitly requests it. For example, to compute the function's value at x = 2.0:
+//
+//   {
+//       ...
+//
+//       struct ggml_cgraph * gf = ggml_new_graph(ctx);
+//       ggml_build_forward_expand(gf, f);
+//
+//       // set the input variable and parameter values
+//       ggml_set_f32(x, 2.0f);
+//       ggml_set_f32(a, 3.0f);
+//       ggml_set_f32(b, 4.0f);
+//
+//       ggml_graph_compute_with_ctx(ctx, &gf, n_threads);
+//
+//       printf("f = %f\n", ggml_get_f32_1d(f, 0));
+//
+//       ...
+//   }
+//
+// The actual computation is performed in the ggml_graph_compute() function.
+//
+// The ggml_new_tensor_...() functions create new tensors. They are allocated in the memory buffer provided to the
+// ggml_init() function. You have to be careful not to exceed the memory buffer size. Therefore, you have to know
+// in advance how much memory you need for your computation. Alternatively, you can allocate a large enough memory
+// and after defining the computation graph, call the ggml_used_mem() function to find out how much memory was
+// actually needed.
+//
+// The ggml_set_param() function marks a tensor as an input variable. This is used by the automatic
+// differentiation and optimization algorithms.
+//
+// The described approach allows to define the function graph once and then compute its forward or backward graphs
+// multiple times. All computations will use the same memory buffer allocated in the ggml_init() function. This way
+// the user can avoid the memory allocation overhead at runtime.
+//
+// The library supports multi-dimensional tensors - up to 4 dimensions. The FP16 and FP32 data types are first class
+// citizens, but in theory the library can be extended to support FP8 and integer data types.
+//
+// Each tensor operation produces a new tensor. Initially the library was envisioned to support only the use of unary
+// and binary operations. Most of the available operations fall into one of these two categories. With time, it became
+// clear that the library needs to support more complex operations. The way to support these operations is not clear
+// yet, but a few examples are demonstrated in the following operations:
+//
+//   - ggml_permute()
+//   - ggml_conv_1d_1s()
+//   - ggml_conv_1d_2s()
+//
+// For each tensor operator, the library implements a forward and backward computation function. The forward function
+// computes the output tensor value given the input tensor values. The backward function computes the adjoint of the
+// input tensors given the adjoint of the output tensor. For a detailed explanation of what this means, take a
+// calculus class, or watch the following video:
+//
+//   What is Automatic Differentiation?
+//   https://www.youtube.com/watch?v=wG_nF1awSSY
+//
+//
+// ## Tensor data (struct ggml_tensor)
+//
+// The tensors are stored in memory via the ggml_tensor struct. The structure provides information about the size of
+// the tensor, the data type, and the memory buffer where the tensor data is stored. Additionally, it contains
+// pointers to the "source" tensors - i.e. the tensors that were used to compute the current tensor. For example:
+//
+//   {
+//       struct ggml_tensor * c = ggml_add(ctx, a, b);
+//
+//       assert(c->src[0] == a);
+//       assert(c->src[1] == b);
+//   }
+//
+// The multi-dimensional tensors are stored in row-major order. The ggml_tensor struct contains fields for the
+// number of elements in each dimension ("ne") as well as the number of bytes ("nb", a.k.a. stride). This allows
+// to store tensors that are not contiguous in memory, which is useful for operations such as transposition and
+// permutation. All tensor operations have to take the stride into account and not assume that the tensor is
+// contiguous in memory.
+//
+// The data of the tensor is accessed via the "data" pointer. For example:
+//
+//   {
+//       const int nx = 2;
+//       const int ny = 3;
+//
+//       struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nx, ny);
+//
+//       for (int y = 0; y < ny; y++) {
+//           for (int x = 0; x < nx; x++) {
+//               *(float *) ((char *) a->data + y*a->nb[1] + x*a->nb[0]) = x + y;
+//           }
+//       }
+//
+//       ...
+//   }
+//
+// Alternatively, there are helper functions, such as ggml_get_f32_1d() and ggml_set_f32_1d() that can be used.
+//
+// ## The matrix multiplication operator (ggml_mul_mat)
+//
+// TODO
+//
+//
+// ## Multi-threading
+//
+// TODO
+//
+//
+// ## Overview of ggml.c
+//
+// TODO
+//
+//
+// ## SIMD optimizations
+//
+// TODO
+//
+//
+// ## Debugging ggml
+//
+// TODO
+//
+//
+
+#ifdef GGML_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef GGML_BUILD
+#            define GGML_API __declspec(dllexport)
+#        else
+#            define GGML_API __declspec(dllimport)
+#        endif
+#    else
+#        define GGML_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define GGML_API
+#endif
+
+#ifdef GGML_MULTIPLATFORM
+#    if defined(_WIN32)
+#        define GGML_CALL
+#    else
+#        define GGML_CALL __attribute__((__ms_abi__))
+#    endif
+#else
+#    define GGML_CALL
+#endif
+
+// TODO: support for clang
+#ifdef __GNUC__
+#    define GGML_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define GGML_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define GGML_DEPRECATED(func, hint) func
+#endif
+
+#ifndef __GNUC__
+#    define GGML_ATTRIBUTE_FORMAT(...)
+#elif defined(__MINGW32__)
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+
+#include 
+#include 
+#include 
+#include 
+
+#define GGML_FILE_MAGIC   0x67676d6c // "ggml"
+#define GGML_FILE_VERSION 1
+
+#define GGML_QNT_VERSION        2    // bump this on quantization format changes
+#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
+
+#define GGML_MAX_DIMS           4
+#define GGML_MAX_PARAMS         2048
+#define GGML_MAX_CONTEXTS       64
+#define GGML_MAX_SRC            10
+#ifndef GGML_MAX_NAME
+#define GGML_MAX_NAME           64
+#endif
+#define GGML_MAX_OP_PARAMS      64
+#define GGML_DEFAULT_N_THREADS  4
+#define GGML_DEFAULT_GRAPH_SIZE 2048
+#if UINTPTR_MAX == 0xFFFFFFFF
+    #define GGML_MEM_ALIGN 4
+#else
+    #define GGML_MEM_ALIGN 16
+#endif
+
+#define GGML_EXIT_SUCCESS 0
+#define GGML_EXIT_ABORTED 1
+
+#define GGUF_MAGIC "GGUF"
+
+#define GGUF_VERSION 3
+
+#define GGUF_DEFAULT_ALIGNMENT 32
+
+#define GGML_UNUSED(x) (void)(x)
+
+#define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#ifndef NDEBUG
+#define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
+#elif defined(__GNUC__)
+#define GGML_UNREACHABLE() __builtin_unreachable()
+#elif defined(_MSC_VER)
+#define GGML_UNREACHABLE() __assume(0)
+#else
+#define GGML_UNREACHABLE() ((void) 0)
+#endif
+
+#ifdef __cplusplus
+#define GGML_NORETURN [[noreturn]]
+#elif defined(_MSC_VER)
+#define GGML_NORETURN __declspec(noreturn)
+#else
+#define GGML_NORETURN _Noreturn
+#endif
+
+#define GGML_ABORT(...) ggml_abort(__FILE__, __LINE__, __VA_ARGS__)
+#define GGML_ASSERT(x) if (!(x)) GGML_ABORT("GGML_ASSERT(%s) failed", #x)
+
+// used to copy the number of elements and stride in bytes of tensors into local variables.
+// main purpose is to reduce code duplication and improve readability.
+//
+// example:
+//
+//    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
+//    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb);
+//
+#define GGML_TENSOR_LOCALS_1(type, prefix, pointer, array) \
+    const type prefix##0 = (pointer)->array[0]; \
+    GGML_UNUSED(prefix##0);
+#define GGML_TENSOR_LOCALS_2(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_1    (type, prefix, pointer, array) \
+    const type prefix##1 = (pointer)->array[1]; \
+    GGML_UNUSED(prefix##1);
+#define GGML_TENSOR_LOCALS_3(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_2    (type, prefix, pointer, array) \
+    const type prefix##2 = (pointer)->array[2]; \
+    GGML_UNUSED(prefix##2);
+#define GGML_TENSOR_LOCALS(type, prefix, pointer, array) \
+    GGML_TENSOR_LOCALS_3  (type, prefix, pointer, array) \
+    const type prefix##3 = (pointer)->array[3]; \
+    GGML_UNUSED(prefix##3);
+
+#define GGML_TENSOR_UNARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
+#define GGML_TENSOR_BINARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
+#define GGML_TENSOR_BINARY_OP_LOCALS01 \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
+    GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
+
+    enum ggml_status {
+        GGML_STATUS_ALLOC_FAILED = -2,
+        GGML_STATUS_FAILED = -1,
+        GGML_STATUS_SUCCESS = 0,
+        GGML_STATUS_ABORTED = 1,
+    };
+
+    // get ggml_status name string
+    GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status);
+
+    // ieee 754-2008 half-precision float16
+    // todo: make this not an integral type
+    typedef uint16_t ggml_fp16_t;
+    GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t);
+    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float);
+    GGML_API void        ggml_fp16_to_fp32_row(const ggml_fp16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_fp16_row(const float *, ggml_fp16_t *, int64_t);
+
+    // google brain half-precision bfloat16
+    typedef struct { uint16_t bits; } ggml_bf16_t;
+    GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
+    GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
+    GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row_ref(const float *, ggml_bf16_t *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
+
+    struct ggml_object;
+    struct ggml_context;
+
+    // NOTE: always add types at the end of the enum to keep backward compatibility
+    enum ggml_type {
+        GGML_TYPE_F32     = 0,
+        GGML_TYPE_F16     = 1,
+        GGML_TYPE_Q4_0    = 2,
+        GGML_TYPE_Q4_1    = 3,
+        // GGML_TYPE_Q4_2 = 4, support has been removed
+        // GGML_TYPE_Q4_3 = 5, support has been removed
+        GGML_TYPE_Q5_0    = 6,
+        GGML_TYPE_Q5_1    = 7,
+        GGML_TYPE_Q8_0    = 8,
+        GGML_TYPE_Q8_1    = 9,
+        GGML_TYPE_Q2_K    = 10,
+        GGML_TYPE_Q3_K    = 11,
+        GGML_TYPE_Q4_K    = 12,
+        GGML_TYPE_Q5_K    = 13,
+        GGML_TYPE_Q6_K    = 14,
+        GGML_TYPE_Q8_K    = 15,
+        GGML_TYPE_IQ2_XXS = 16,
+        GGML_TYPE_IQ2_XS  = 17,
+        GGML_TYPE_IQ3_XXS = 18,
+        GGML_TYPE_IQ1_S   = 19,
+        GGML_TYPE_IQ4_NL  = 20,
+        GGML_TYPE_IQ3_S   = 21,
+        GGML_TYPE_IQ2_S   = 22,
+        GGML_TYPE_IQ4_XS  = 23,
+        GGML_TYPE_I8      = 24,
+        GGML_TYPE_I16     = 25,
+        GGML_TYPE_I32     = 26,
+        GGML_TYPE_I64     = 27,
+        GGML_TYPE_F64     = 28,
+        GGML_TYPE_IQ1_M   = 29,
+        GGML_TYPE_BF16    = 30,
+        GGML_TYPE_Q4_0_4_4 = 31,
+        GGML_TYPE_Q4_0_4_8 = 32,
+        GGML_TYPE_Q4_0_8_8 = 33,
+        GGML_TYPE_COUNT,
+    };
+
+    // precision
+    enum ggml_prec {
+        GGML_PREC_DEFAULT,
+        GGML_PREC_F32,
+    };
+
+    enum ggml_backend_type {
+        GGML_BACKEND_TYPE_CPU = 0,
+        GGML_BACKEND_TYPE_GPU = 10,
+        GGML_BACKEND_TYPE_GPU_SPLIT = 20,
+    };
+
+    // model file types
+    enum ggml_ftype {
+        GGML_FTYPE_UNKNOWN        = -1,
+        GGML_FTYPE_ALL_F32        = 0,
+        GGML_FTYPE_MOSTLY_F16     = 1,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0    = 2,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1    = 3,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+        GGML_FTYPE_MOSTLY_Q8_0    = 7,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_0    = 8,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_1    = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K    = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K    = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K    = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K    = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K    = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ1_S   = 18, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ4_NL  = 19, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ3_S   = 20, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_S   = 21, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
+        GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
+    };
+
+    // available tensor operations:
+    enum ggml_op {
+        GGML_OP_NONE = 0,
+
+        GGML_OP_DUP,
+        GGML_OP_ADD,
+        GGML_OP_ADD1,
+        GGML_OP_ACC,
+        GGML_OP_SUB,
+        GGML_OP_MUL,
+        GGML_OP_DIV,
+        GGML_OP_SQR,
+        GGML_OP_SQRT,
+        GGML_OP_LOG,
+        GGML_OP_SUM,
+        GGML_OP_SUM_ROWS,
+        GGML_OP_MEAN,
+        GGML_OP_ARGMAX,
+        GGML_OP_REPEAT,
+        GGML_OP_REPEAT_BACK,
+        GGML_OP_CONCAT,
+        GGML_OP_SILU_BACK,
+        GGML_OP_NORM, // normalize
+        GGML_OP_RMS_NORM,
+        GGML_OP_RMS_NORM_BACK,
+        GGML_OP_GROUP_NORM,
+
+        GGML_OP_MUL_MAT,
+        GGML_OP_MUL_MAT_ID,
+        GGML_OP_OUT_PROD,
+
+        GGML_OP_SCALE,
+        GGML_OP_SET,
+        GGML_OP_CPY,
+        GGML_OP_CONT,
+        GGML_OP_RESHAPE,
+        GGML_OP_VIEW,
+        GGML_OP_PERMUTE,
+        GGML_OP_TRANSPOSE,
+        GGML_OP_GET_ROWS,
+        GGML_OP_GET_ROWS_BACK,
+        GGML_OP_DIAG,
+        GGML_OP_DIAG_MASK_INF,
+        GGML_OP_DIAG_MASK_ZERO,
+        GGML_OP_SOFT_MAX,
+        GGML_OP_SOFT_MAX_BACK,
+        GGML_OP_ROPE,
+        GGML_OP_ROPE_BACK,
+        GGML_OP_CLAMP,
+        GGML_OP_CONV_TRANSPOSE_1D,
+        GGML_OP_IM2COL,
+        GGML_OP_CONV_TRANSPOSE_2D,
+        GGML_OP_POOL_1D,
+        GGML_OP_POOL_2D,
+        GGML_OP_UPSCALE, // nearest interpolate
+        GGML_OP_PAD,
+        GGML_OP_ARANGE,
+        GGML_OP_TIMESTEP_EMBEDDING,
+        GGML_OP_ARGSORT,
+        GGML_OP_LEAKY_RELU,
+
+        GGML_OP_FLASH_ATTN_EXT,
+        GGML_OP_FLASH_ATTN_BACK,
+        GGML_OP_SSM_CONV,
+        GGML_OP_SSM_SCAN,
+        GGML_OP_WIN_PART,
+        GGML_OP_WIN_UNPART,
+        GGML_OP_GET_REL_POS,
+        GGML_OP_ADD_REL_POS,
+
+        GGML_OP_UNARY,
+
+        GGML_OP_MAP_UNARY,
+        GGML_OP_MAP_BINARY,
+
+        GGML_OP_MAP_CUSTOM1_F32,
+        GGML_OP_MAP_CUSTOM2_F32,
+        GGML_OP_MAP_CUSTOM3_F32,
+
+        GGML_OP_MAP_CUSTOM1,
+        GGML_OP_MAP_CUSTOM2,
+        GGML_OP_MAP_CUSTOM3,
+
+        GGML_OP_CROSS_ENTROPY_LOSS,
+        GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+
+        GGML_OP_COUNT,
+    };
+
+    enum ggml_unary_op {
+        GGML_UNARY_OP_ABS,
+        GGML_UNARY_OP_SGN,
+        GGML_UNARY_OP_NEG,
+        GGML_UNARY_OP_STEP,
+        GGML_UNARY_OP_TANH,
+        GGML_UNARY_OP_ELU,
+        GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_SIGMOID,
+        GGML_UNARY_OP_GELU,
+        GGML_UNARY_OP_GELU_QUICK,
+        GGML_UNARY_OP_SILU,
+        GGML_UNARY_OP_HARDSWISH,
+        GGML_UNARY_OP_HARDSIGMOID,
+
+        GGML_UNARY_OP_COUNT,
+    };
+
+    enum ggml_object_type {
+        GGML_OBJECT_TYPE_TENSOR,
+        GGML_OBJECT_TYPE_GRAPH,
+        GGML_OBJECT_TYPE_WORK_BUFFER
+    };
+
+    enum ggml_log_level {
+        GGML_LOG_LEVEL_ERROR = 2,
+        GGML_LOG_LEVEL_WARN  = 3,
+        GGML_LOG_LEVEL_INFO  = 4,
+        GGML_LOG_LEVEL_DEBUG = 5
+    };
+
+    enum ggml_tensor_flag {
+        GGML_TENSOR_FLAG_INPUT  = 1,
+        GGML_TENSOR_FLAG_OUTPUT = 2,
+        GGML_TENSOR_FLAG_PARAM  = 4,
+    };
+
+    // ggml object
+    struct ggml_object {
+        size_t offs;
+        size_t size;
+
+        struct ggml_object * next;
+
+        enum ggml_object_type type;
+
+        char padding[4];
+    };
+
+    static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
+
+    // n-dimensional tensor
+    struct ggml_tensor {
+        enum ggml_type         type;
+
+        GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
+
+        struct ggml_backend_buffer * buffer;
+
+        int64_t ne[GGML_MAX_DIMS]; // number of elements
+        size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
+                                   // nb[0] = ggml_type_size(type)
+                                   // nb[1] = nb[0]   * (ne[0] / ggml_blck_size(type)) + padding
+                                   // nb[i] = nb[i-1] * ne[i-1]
+
+        // compute data
+        enum ggml_op op;
+
+        // op params - allocated as int32_t for alignment
+        int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+
+        int32_t flags;
+
+        struct ggml_tensor * grad;
+        struct ggml_tensor * src[GGML_MAX_SRC];
+
+        // source tensor and offset for views
+        struct ggml_tensor * view_src;
+        size_t               view_offs;
+
+        void * data;
+
+        char name[GGML_MAX_NAME];
+
+        void * extra; // extra things e.g. for ggml-cuda.cu
+
+        // char padding[4];
+    };
+
+    static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
+
+    // Abort callback
+    // If not NULL, called before ggml computation
+    // If it returns true, the computation is aborted
+    typedef bool (*ggml_abort_callback)(void * data);
+
+    // the compute plan that needs to be prepared for ggml_graph_compute()
+    // since https://github.com/ggerganov/ggml/issues/287
+    struct ggml_cplan {
+        size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
+        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
+
+        int n_threads;
+
+        // abort ggml_graph_compute when true
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
+    };
+
+    enum ggml_cgraph_eval_order {
+        GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+        GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+        GGML_CGRAPH_EVAL_ORDER_COUNT
+    };
+
+    typedef uint32_t ggml_bitset_t;
+
+    struct ggml_hash_set {
+        size_t size;
+        ggml_bitset_t * used;
+        struct ggml_tensor ** keys;
+    };
+
+    // computation graph
+    struct ggml_cgraph {
+        int size;
+        int n_nodes;
+        int n_leafs;
+
+        struct ggml_tensor ** nodes;
+        struct ggml_tensor ** grads;
+        struct ggml_tensor ** leafs;
+
+        struct ggml_hash_set visited_hash_set;
+
+        enum ggml_cgraph_eval_order order;
+    };
+
+    // scratch buffer
+    struct ggml_scratch {
+        size_t offs;
+        size_t size;
+        void * data;
+    };
+
+    struct ggml_init_params {
+        // memory pool
+        size_t mem_size;   // bytes
+        void * mem_buffer; // if NULL, memory will be allocated internally
+        bool   no_alloc;   // don't allocate memory for the tensor data
+    };
+
+    // numa strategies
+    enum ggml_numa_strategy {
+        GGML_NUMA_STRATEGY_DISABLED   = 0,
+        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        GGML_NUMA_STRATEGY_MIRROR     = 4,
+        GGML_NUMA_STRATEGY_COUNT
+    };
+
+    //
+    // GUID
+    //
+
+    // GUID types
+    typedef uint8_t ggml_guid[16];
+    typedef ggml_guid * ggml_guid_t;
+
+    GGML_API bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b);
+
+    // misc
+
+    GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
+    GGML_API int64_t ggml_time_ms(void);
+    GGML_API int64_t ggml_time_us(void);
+    GGML_API int64_t ggml_cycles(void);
+    GGML_API int64_t ggml_cycles_per_ms(void);
+
+    // accepts a UTF-8 path, even on Windows
+    GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
+
+    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+
+    GGML_API void    ggml_print_object (const struct ggml_object * obj);
+    GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
+
+    GGML_API GGML_CALL int64_t ggml_nelements   (const struct ggml_tensor * tensor);
+    GGML_API GGML_CALL int64_t ggml_nrows       (const struct ggml_tensor * tensor);
+    GGML_API GGML_CALL size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
+    GGML_API           size_t  ggml_nbytes_pad  (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
+
+    GGML_API GGML_CALL int64_t ggml_blck_size(enum ggml_type type);
+    GGML_API GGML_CALL size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
+    GGML_API GGML_CALL size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
+
+    GGML_DEPRECATED(
+    GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
+    "use ggml_row_size() instead");
+
+    GGML_API GGML_CALL const char * ggml_type_name(enum ggml_type type);
+    GGML_API GGML_CALL const char * ggml_op_name  (enum ggml_op   op);
+    GGML_API           const char * ggml_op_symbol(enum ggml_op   op);
+
+    GGML_API           const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API GGML_CALL const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
+
+    GGML_API GGML_CALL size_t  ggml_element_size(const struct ggml_tensor * tensor);
+
+    GGML_API GGML_CALL bool    ggml_is_quantized(enum ggml_type type);
+
+    // TODO: temporary until model loading of ggml examples is refactored
+    GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
+
+    GGML_API GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor);
+    GGML_API GGML_CALL bool ggml_is_permuted  (const struct ggml_tensor * tensor);
+    GGML_API GGML_CALL bool ggml_is_empty     (const struct ggml_tensor * tensor);
+    GGML_API           bool ggml_is_scalar    (const struct ggml_tensor * tensor);
+    GGML_API           bool ggml_is_vector    (const struct ggml_tensor * tensor);
+    GGML_API           bool ggml_is_matrix    (const struct ggml_tensor * tensor);
+    GGML_API           bool ggml_is_3d        (const struct ggml_tensor * tensor);
+    GGML_API           int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
+
+    GGML_API GGML_CALL bool ggml_is_contiguous  (const struct ggml_tensor * tensor);
+    GGML_API GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
+    GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
+    GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
+
+    GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+    GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+
+    GGML_API bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+
+    // use this to compute the memory overhead of a tensor
+    GGML_API size_t ggml_tensor_overhead(void);
+
+    GGML_API bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes);
+
+    // main
+
+    GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
+    GGML_API void                  ggml_free(struct ggml_context * ctx);
+
+    GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
+
+    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
+    GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);
+    GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
+
+    GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_mem_size       (const struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_max_tensor_size(const struct ggml_context * ctx);
+
+    GGML_API struct ggml_tensor * ggml_new_tensor(
+            struct ggml_context * ctx,
+            enum   ggml_type type,
+            int    n_dims,
+            const int64_t *ne);
+
+    GGML_API struct ggml_tensor * ggml_new_tensor_1d(
+            struct ggml_context * ctx,
+            enum   ggml_type type,
+            int64_t ne0);
+
+    GGML_API struct ggml_tensor * ggml_new_tensor_2d(
+            struct ggml_context * ctx,
+            enum   ggml_type type,
+            int64_t ne0,
+            int64_t ne1);
+
+    GGML_API struct ggml_tensor * ggml_new_tensor_3d(
+            struct ggml_context * ctx,
+            enum   ggml_type type,
+            int64_t ne0,
+            int64_t ne1,
+            int64_t ne2);
+
+    GGML_API struct ggml_tensor * ggml_new_tensor_4d(
+            struct ggml_context * ctx,
+            enum   ggml_type type,
+            int64_t ne0,
+            int64_t ne1,
+            int64_t ne2,
+            int64_t ne3);
+
+    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+
+    GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
+    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
+
+    // Context tensor enumeration and lookup
+    GGML_API struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx);
+    GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
+
+    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+
+    // Converts a flat index into coordinates
+    GGML_API void    ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
+
+    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
+
+    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
+    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+
+    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
+    GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
+    GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
+
+    GGML_API GGML_CALL enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+
+    GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
+    GGML_ATTRIBUTE_FORMAT(2, 3)
+    GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
+
+    //
+    // operations on tensors with backpropagation
+    //
+
+    GGML_API struct ggml_tensor * ggml_dup(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_dup_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_add(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_add_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_add_cast(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            enum   ggml_type      type);
+
+    GGML_API struct ggml_tensor * ggml_add1(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_add1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // dst = a
+    // view(dst, nb1, nb2, nb3, offset) += b
+    // return dst
+    GGML_API struct ggml_tensor * ggml_acc(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_acc_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_sub(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_sub_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_mul(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_mul_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_div(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_div_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_sqr(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sqr_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sqrt(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sqrt_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_log(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_log_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // return scalar
+    GGML_API struct ggml_tensor * ggml_sum(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d]
+    GGML_API struct ggml_tensor * ggml_sum_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // mean along rows
+    GGML_API struct ggml_tensor * ggml_mean(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // argmax along rows
+    GGML_API struct ggml_tensor * ggml_argmax(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // if a is the same shape as b, and a is not parameter, return a
+    // otherwise, return a new tensor: repeat(a) to fit in b
+    GGML_API struct ggml_tensor * ggml_repeat(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // sums repetitions in a into shape of b
+    GGML_API struct ggml_tensor * ggml_repeat_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // concat a and b along dim
+    // used in stable-diffusion
+    GGML_API struct ggml_tensor * ggml_concat(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   dim);
+
+    GGML_API struct ggml_tensor * ggml_abs(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_abs_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sgn(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sgn_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_neg(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_neg_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_step(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_step_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_tanh(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_tanh_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_elu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_elu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_relu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_leaky_relu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a, float negative_slope, bool inplace);
+
+    GGML_API struct ggml_tensor * ggml_relu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sigmoid_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_gelu_quick_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_silu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_silu_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // a - x
+    // b - dy
+    GGML_API struct ggml_tensor * ggml_silu_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // hardswish(x) = x * relu6(x + 3) / 6
+    GGML_API struct ggml_tensor * ggml_hardswish(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // hardsigmoid(x) = relu6(x + 3) / 6
+    GGML_API struct ggml_tensor * ggml_hardsigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // normalize along rows
+    GGML_API struct ggml_tensor * ggml_norm(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 eps);
+
+    GGML_API struct ggml_tensor * ggml_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 eps);
+
+    GGML_API struct ggml_tensor * ggml_rms_norm(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 eps);
+
+    GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 eps);
+
+    // group normalize along ne0*ne1*n_groups
+    // used in stable-diffusion
+    GGML_API struct ggml_tensor * ggml_group_norm(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_groups,
+            float                 eps);
+
+    GGML_API struct ggml_tensor * ggml_group_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_groups,
+            float                 eps);
+
+    // a - x
+    // b - dy
+    GGML_API struct ggml_tensor * ggml_rms_norm_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            float                 eps);
+
+    // A: k columns, n rows => [ne03, ne02, n, k]
+    // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
+    // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
+    GGML_API struct ggml_tensor * ggml_mul_mat(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // change the precision of a matrix multiplication
+    // set to GGML_PREC_F32 for higher precision (useful for phi-2)
+    GGML_API void ggml_mul_mat_set_prec(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec);
+
+    // indirect matrix multiplication
+    GGML_API struct ggml_tensor * ggml_mul_mat_id(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * as,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * ids);
+
+    // A: m columns, n rows,
+    // B: p columns, n rows,
+    // result is m columns, p rows
+    GGML_API struct ggml_tensor * ggml_out_prod(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    //
+    // operations on tensors without backpropagation
+    //
+
+    GGML_API struct ggml_tensor * ggml_scale(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 s);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_scale_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 s);
+
+    // b -> view(a,offset,nb1,nb2,3), return modified a
+    GGML_API struct ggml_tensor * ggml_set(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return view(a)
+    GGML_API struct ggml_tensor * ggml_set_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_set_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_set_1d_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return modified a
+    GGML_API struct ggml_tensor * ggml_set_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return view(a)
+    GGML_API struct ggml_tensor * ggml_set_2d_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                offset);
+
+    // a -> b, return view(b)
+    GGML_API struct ggml_tensor * ggml_cpy(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_cast(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum   ggml_type      type);
+
+    // make contiguous
+    GGML_API struct ggml_tensor * ggml_cont(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // make contiguous, with new shape
+    GGML_API struct ggml_tensor * ggml_cont_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
+    GGML_API struct ggml_tensor * ggml_cont_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1);
+
+    GGML_API struct ggml_tensor * ggml_cont_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2);
+
+    GGML_API struct ggml_tensor * ggml_cont_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
+    // return view(a), b specifies the new shape
+    // TODO: when we start computing gradient, make a copy instead of view
+    GGML_API struct ggml_tensor * ggml_reshape(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // return view(a)
+    // TODO: when we start computing gradient, make a copy instead of view
+    GGML_API struct ggml_tensor * ggml_reshape_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
+    GGML_API struct ggml_tensor * ggml_reshape_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1);
+
+    // return view(a)
+    // TODO: when we start computing gradient, make a copy instead of view
+    GGML_API struct ggml_tensor * ggml_reshape_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2);
+
+    GGML_API struct ggml_tensor * ggml_reshape_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
+    // offset in bytes
+    GGML_API struct ggml_tensor * ggml_view_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_view_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            size_t                nb1, // row stride in bytes
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_view_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            size_t                nb1, // row   stride in bytes
+            size_t                nb2, // slice stride in bytes
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_view_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            size_t                nb1, // row   stride in bytes
+            size_t                nb2, // slice stride in bytes
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_permute(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   axis0,
+            int                   axis1,
+            int                   axis2,
+            int                   axis3);
+
+    // alias for ggml_permute(ctx, a, 1, 0, 2, 3)
+    GGML_API struct ggml_tensor * ggml_transpose(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // supports 3D: a->ne[2] == b->ne[1]
+    GGML_API struct ggml_tensor * ggml_get_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_get_rows_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c);
+
+    GGML_API struct ggml_tensor * ggml_diag(
+        struct ggml_context     * ctx,
+        struct ggml_tensor      * a);
+
+    // set elements above the diagonal to -INF
+    GGML_API struct ggml_tensor * ggml_diag_mask_inf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    // set elements above the diagonal to 0
+    GGML_API struct ggml_tensor * ggml_diag_mask_zero(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    GGML_API struct ggml_tensor * ggml_soft_max(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // fused soft_max(a*scale + mask*(ALiBi slope))
+    // mask is optional
+    // max_bias = 0.0f for no ALiBi
+    GGML_API struct ggml_tensor * ggml_soft_max_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * mask,
+            float                 scale,
+            float                 max_bias);
+
+    GGML_API struct ggml_tensor * ggml_soft_max_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // rotary position embedding
+    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
+    // if mode & 2 == 1, GPT-NeoX style
+    //
+    // b is an int32 vector with size a->ne[2], it contains the positions
+    GGML_API struct ggml_tensor * ggml_rope(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   n_dims,
+            int                   mode);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_rope_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   n_dims,
+            int                   mode);
+
+    // custom RoPE
+    // c is freq factors (e.g. phi3-128k), (optional)
+    GGML_API struct ggml_tensor * ggml_rope_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use ggml_rope_ext instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use ggml_rope_ext_inplace instead");
+
+    // compute correction dims for YaRN RoPE scaling
+    GGML_CALL void ggml_rope_yarn_corr_dims(
+        int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
+
+    // rotary position embedding backward, i.e compute dx from dy
+    // a - dy
+    GGML_API struct ggml_tensor * ggml_rope_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
+    // clamp
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_clamp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
+    GGML_API struct ggml_tensor * ggml_im2col(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                  s0,
+            int                  s1,
+            int                  p0,
+            int                  p1,
+            int                  d0,
+            int                  d1,
+            bool                 is_2D,
+            enum ggml_type       dst_type);
+
+    GGML_API struct ggml_tensor * ggml_conv_depthwise_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                  s0,
+            int                  s1,
+            int                  p0,
+            int                  p1,
+            int                  d0,
+            int                  d1);
+
+    GGML_API struct ggml_tensor * ggml_conv_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
+
+    // conv_1d with padding = half
+    // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
+    GGML_API struct ggml_tensor* ggml_conv_1d_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   s,
+            int                   d);
+
+    GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   s0,
+            int                   p0,
+            int                   d0);
+
+    GGML_API struct ggml_tensor * ggml_conv_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   s0,
+            int                   s1,
+            int                   p0,
+            int                   p1,
+            int                   d0,
+            int                   d1);
+
+
+    // kernel size is a->ne[0] x a->ne[1]
+    // stride is equal to kernel size
+    // padding is zero
+    // example:
+    // a:     16   16    3  768
+    // b:   1024 1024    3    1
+    // res:   64   64  768    1
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_conv_2d_sk_p0(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // kernel size is a->ne[0] x a->ne[1]
+    // stride is 1
+    // padding is half
+    // example:
+    // a:      3    3    256  256
+    // b:     64   64    256    1
+    // res:   64   64    256    1
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_conv_2d_s1_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   stride);
+
+    enum ggml_op_pool {
+        GGML_OP_POOL_MAX,
+        GGML_OP_POOL_AVG,
+        GGML_OP_POOL_COUNT,
+    };
+
+    GGML_API struct ggml_tensor * ggml_pool_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_op_pool     op,
+            int                   k0, // kernel size
+            int                   s0, // stride
+            int                   p0); // padding
+
+    // the result will have 2*p0 padding for the first dimension
+    // and 2*p1 padding for the second dimension
+    GGML_API struct ggml_tensor * ggml_pool_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_op_pool     op,
+            int                   k0,
+            int                   k1,
+            int                   s0,
+            int                   s1,
+            float                 p0,
+            float                 p1);
+
+    // nearest interpolate
+    // multiplies ne0 and ne1 by scale factor
+    // used in stable-diffusion
+    GGML_API struct ggml_tensor * ggml_upscale(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   scale_factor);
+
+    // nearest interpolate
+    // nearest interpolate to specified dimensions
+    // used in tortoise.cpp
+    GGML_API struct ggml_tensor * ggml_upscale_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   ne0,
+            int                   ne1,
+            int                   ne2,
+            int                   ne3);
+
+    // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
+    GGML_API struct ggml_tensor * ggml_pad(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                  p0,
+            int                  p1,
+            int                  p2,
+            int                  p3);
+
+    // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
+    // timesteps: [N,]
+    // return: [N, dim]
+    GGML_API struct ggml_tensor * ggml_timestep_embedding(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * timesteps,
+            int                   dim,
+            int                   max_period);
+
+    // sort rows
+    enum ggml_sort_order {
+        GGML_SORT_ORDER_ASC,
+        GGML_SORT_ORDER_DESC,
+    };
+
+    GGML_API struct ggml_tensor * ggml_argsort(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_sort_order  order);
+
+    GGML_API struct ggml_tensor * ggml_arange(
+            struct ggml_context * ctx,
+            float                 start,
+            float                 stop,
+            float                 step);
+
+    // top k elements per row
+    GGML_API struct ggml_tensor * ggml_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
+#define GGML_KQ_MASK_PAD 32
+
+    // q:    [n_embd, n_batch,     n_head,    1]
+    // k:    [n_embd, n_kv,        n_head_kv, 1]
+    // v:    [n_embd, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,   n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd, n_head,      n_batch,   1] !! permuted !!
+    GGML_API struct ggml_tensor * ggml_flash_attn_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * q,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * mask,
+            float                 scale,
+            float                 max_bias);
+
+    GGML_API void ggml_flash_attn_ext_set_prec(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec);
+
+    // TODO: needs to be adapted to ggml_flash_attn_ext
+    GGML_API struct ggml_tensor * ggml_flash_attn_back(
+           struct ggml_context * ctx,
+           struct ggml_tensor  * q,
+           struct ggml_tensor  * k,
+           struct ggml_tensor  * v,
+           struct ggml_tensor  * d,
+           bool                  masked);
+
+    GGML_API struct ggml_tensor * ggml_ssm_conv(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * s,
+            struct ggml_tensor  * x,
+            struct ggml_tensor  * c,
+            struct ggml_tensor  * sq);
+
+    GGML_API struct ggml_tensor * ggml_ssm_scan(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * s,
+            struct ggml_tensor  * x,
+            struct ggml_tensor  * dt,
+            struct ggml_tensor  * A,
+            struct ggml_tensor  * B,
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * sq);
+
+    // partition into non-overlapping windows with padding if needed
+    // example:
+    // a:   768   64   64    1
+    // w:    14
+    // res: 768   14   14    25
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_win_part(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   w);
+
+    // reverse of ggml_win_part
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_win_unpart(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   w0,
+            int                   h0,
+            int                   w);
+
+    GGML_API struct ggml_tensor * ggml_unary(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_unary_op op);
+
+    GGML_API struct ggml_tensor * ggml_unary_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_unary_op op);
+
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_get_rel_pos(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   qh,
+            int                   kh);
+
+    // used in sam
+    GGML_API struct ggml_tensor * ggml_add_rel_pos(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * pw,
+            struct ggml_tensor  * ph);
+
+    GGML_API struct ggml_tensor * ggml_add_rel_pos_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * pw,
+            struct ggml_tensor  * ph);
+
+    // custom operators
+
+    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
+    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+
+    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
+            struct ggml_context        * ctx,
+            struct ggml_tensor         * a,
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+            struct ggml_context        * ctx,
+            struct ggml_tensor         * a,
+                   ggml_unary_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+                   ggml_binary_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1 instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+                   ggml_custom1_op_f32_t   fun),
+        "use ggml_map_custom1_inplace instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2 instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+                   ggml_custom2_op_f32_t   fun),
+        "use ggml_map_custom2_inplace instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+            struct ggml_tensor           * c,
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3 instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+            struct ggml_tensor           * c,
+                   ggml_custom3_op_f32_t   fun),
+        "use ggml_map_custom3_inplace instead");
+
+    // custom operators v2
+
+    typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
+    typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
+
+    #define GGML_N_TASKS_MAX -1
+
+    GGML_API struct ggml_tensor * ggml_map_custom1(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            ggml_custom1_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            ggml_custom2_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
+            struct ggml_context   * ctx,
+            struct ggml_tensor    * a,
+            struct ggml_tensor    * b,
+            struct ggml_tensor    * c,
+            ggml_custom3_op_t       fun,
+            int                     n_tasks,
+            void                  * userdata);
+
+    // loss function
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b);
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+            struct ggml_tensor          * c);
+
+    //
+    // automatic differentiation
+    //
+
+    GGML_API void ggml_set_param(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * tensor);
+
+
+    GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
+
+    // graph allocation in a context
+    GGML_API struct ggml_cgraph * ggml_new_graph         (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
+    GGML_API struct ggml_cgraph * ggml_new_graph_custom  (struct ggml_context * ctx, size_t size, bool grads);
+    GGML_API struct ggml_cgraph * ggml_graph_dup         (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph   ggml_graph_view        (struct ggml_cgraph * cgraph, int i0, int i1);
+    GGML_API void                 ggml_graph_cpy         (struct ggml_cgraph * src, struct ggml_cgraph * dst);
+    GGML_API void                 ggml_graph_reset       (struct ggml_cgraph * cgraph);  // zero grads
+    GGML_API void                 ggml_graph_clear       (struct ggml_cgraph * cgraph);
+
+    GGML_API size_t ggml_graph_overhead(void);
+    GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
+
+    // ggml_graph_plan() has to be called before ggml_graph_compute()
+    // when plan.work_size > 0, caller must allocate memory for plan.work_data
+    GGML_API struct ggml_cplan ggml_graph_plan   (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
+    GGML_API enum ggml_status  ggml_graph_compute(      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    // same as ggml_graph_compute() but the work data is allocated as a part of the context
+    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
+    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+
+    GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
+    GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
+
+    // print info and performance information for the graph
+    GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
+
+    // dump the graph into a file using the dot format
+    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
+
+    // build gradient checkpointing backward graph gb for gf using provided checkpoints
+    // gb_tmp will contain original backward graph with rewritten backward process nodes,
+    // but without the second forward pass nodes.
+    GGML_API void ggml_build_backward_gradient_checkpointing(
+            struct ggml_context   * ctx,
+            struct ggml_cgraph    * gf,
+            struct ggml_cgraph    * gb,
+            struct ggml_cgraph    * gb_tmp,
+            struct ggml_tensor  * * checkpoints,
+            int                     n_checkpoints);
+    //
+    // optimization
+    //
+
+    // optimization methods
+    enum ggml_opt_type {
+        GGML_OPT_TYPE_ADAM,
+        GGML_OPT_TYPE_LBFGS,
+    };
+
+    // linesearch methods
+    enum ggml_linesearch {
+        GGML_LINESEARCH_DEFAULT = 1,
+
+        GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
+        GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
+        GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
+    };
+
+    // optimization return values
+    enum ggml_opt_result {
+        GGML_OPT_RESULT_OK = 0,
+        GGML_OPT_RESULT_DID_NOT_CONVERGE,
+        GGML_OPT_RESULT_NO_CONTEXT,
+        GGML_OPT_RESULT_INVALID_WOLFE,
+        GGML_OPT_RESULT_FAIL,
+        GGML_OPT_RESULT_CANCEL,
+
+        GGML_LINESEARCH_FAIL = -128,
+        GGML_LINESEARCH_MINIMUM_STEP,
+        GGML_LINESEARCH_MAXIMUM_STEP,
+        GGML_LINESEARCH_MAXIMUM_ITERATIONS,
+        GGML_LINESEARCH_INVALID_PARAMETERS,
+    };
+
+    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
+
+    // optimization parameters
+    //
+    //   see ggml.c (ggml_opt_default_params) for default values
+    //
+    struct ggml_opt_params {
+        enum ggml_opt_type type;
+
+        size_t graph_size;
+
+        int n_threads;
+
+        // delta-based convergence test
+        //
+        //   if past == 0 - disabled
+        //   if past > 0:
+        //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
+        //
+        int past;
+        float delta;
+
+        // maximum number of iterations without improvement
+        //
+        //   if 0 - disabled
+        //   if > 0:
+        //     assume convergence if no cost improvement in this number of iterations
+        //
+        int max_no_improvement;
+
+        bool print_forward_graph;
+        bool print_backward_graph;
+
+        int n_gradient_accumulation;
+
+        // ADAM parameters
+        struct {
+            int n_iter;
+
+            float sched; // schedule multiplier (fixed, decay or warmup)
+            float decay; // weight decay for AdamW, use 0.0f to disable
+            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
+            float alpha; // learning rate
+            float beta1;
+            float beta2;
+            float eps;   // epsilon for numerical stability
+            float eps_f; // epsilon for convergence test
+            float eps_g; // epsilon for convergence test
+            float gclip; // gradient clipping
+        } adam;
+
+        // LBFGS parameters
+        struct {
+            int m; // number of corrections to approximate the inv. Hessian
+            int n_iter;
+            int max_linesearch;
+
+            float eps;      // convergence tolerance
+            float ftol;     // line search tolerance
+            float wolfe;
+            float min_step;
+            float max_step;
+
+            enum ggml_linesearch linesearch;
+        } lbfgs;
+    };
+
+    struct ggml_opt_context {
+        struct ggml_context * ctx;
+        struct ggml_opt_params params;
+
+        int iter;
+        int64_t nx; // number of parameter elements
+
+        bool just_initialized;
+
+        float loss_before;
+        float loss_after;
+
+        struct {
+            struct ggml_tensor * g;  // current gradient
+            struct ggml_tensor * m;  // first moment
+            struct ggml_tensor * v;  // second moment
+            struct ggml_tensor * pf; // past function values
+            float fx_best;
+            float fx_prev;
+            int n_no_improvement;
+        } adam;
+
+        struct {
+            struct ggml_tensor * x;    // current parameters
+            struct ggml_tensor * xp;   // previous parameters
+            struct ggml_tensor * g;    // current gradient
+            struct ggml_tensor * gp;   // previous gradient
+            struct ggml_tensor * d;    // search direction
+            struct ggml_tensor * pf;   // past function values
+            struct ggml_tensor * lmal; // the L-BFGS memory alpha
+            struct ggml_tensor * lmys; // the L-BFGS memory ys
+            struct ggml_tensor * lms;  // the L-BFGS memory s
+            struct ggml_tensor * lmy;  // the L-BFGS memory y
+            float fx_best;
+            float step;
+            int j;
+            int k;
+            int end;
+            int n_no_improvement;
+        } lbfgs;
+    };
+
+    GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
+
+    // optimize the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt(
+            struct ggml_context * ctx,
+            struct ggml_opt_params params,
+            struct ggml_tensor * f);
+
+    // initialize optimizer context
+    GGML_API void ggml_opt_init(
+            struct ggml_context     * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_opt_params    params,
+            int64_t                   nx);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume_g(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f,
+            struct ggml_cgraph * gf,
+            struct ggml_cgraph * gb,
+            ggml_opt_callback callback,
+            void * callback_data);
+
+    //
+    // tensor flags
+    //
+    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+
+    //
+    // quantization
+    //
+
+    // - ggml_quantize_init can be called multiple times with the same type
+    //   it will only initialize the quantization tables for the first call or after ggml_quantize_free
+    //   automatically called by ggml_quantize_chunk for convenience
+    //
+    // - ggml_quantize_free will free any memory allocated by ggml_quantize_init
+    //   call this at the end of the program to avoid memory leaks
+    //
+    // note: these are thread-safe
+    //
+    GGML_API void ggml_quantize_init(enum ggml_type type);
+    GGML_API void ggml_quantize_free(void);
+
+    // some quantization type cannot be used without an importance matrix
+    GGML_API bool ggml_quantize_requires_imatrix(enum ggml_type type);
+
+    // calls ggml_quantize_init internally (i.e. can allocate memory)
+    GGML_API size_t ggml_quantize_chunk(
+            enum ggml_type   type,
+               const float * src,
+                      void * dst,
+                   int64_t   start,
+                   int64_t   nrows,
+                   int64_t   n_per_row,
+               const float * imatrix);
+
+    //
+    // gguf
+    //
+
+    enum gguf_type {
+        GGUF_TYPE_UINT8   = 0,
+        GGUF_TYPE_INT8    = 1,
+        GGUF_TYPE_UINT16  = 2,
+        GGUF_TYPE_INT16   = 3,
+        GGUF_TYPE_UINT32  = 4,
+        GGUF_TYPE_INT32   = 5,
+        GGUF_TYPE_FLOAT32 = 6,
+        GGUF_TYPE_BOOL    = 7,
+        GGUF_TYPE_STRING  = 8,
+        GGUF_TYPE_ARRAY   = 9,
+        GGUF_TYPE_UINT64  = 10,
+        GGUF_TYPE_INT64   = 11,
+        GGUF_TYPE_FLOAT64 = 12,
+        GGUF_TYPE_COUNT,       // marks the end of the enum
+    };
+
+    struct gguf_context;
+
+    struct gguf_init_params {
+        bool no_alloc;
+
+        // if not NULL, create a ggml_context and allocate the tensor data in it
+        struct ggml_context ** ctx;
+    };
+
+    GGML_API struct gguf_context * gguf_init_empty(void);
+    GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params);
+    //GGML_API struct gguf_context * gguf_init_from_buffer(..);
+
+    GGML_API void gguf_free(struct gguf_context * ctx);
+
+    GGML_API const char * gguf_type_name(enum gguf_type type);
+
+    GGML_API int    gguf_get_version    (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_alignment  (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx);
+    GGML_API void * gguf_get_data       (const struct gguf_context * ctx);
+
+    GGML_API int          gguf_get_n_kv(const struct gguf_context * ctx);
+    GGML_API int          gguf_find_key(const struct gguf_context * ctx, const char * key);
+    GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id);
+
+    GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id);
+    GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id);
+
+    // will abort if the wrong type is used for the key
+    GGML_API uint8_t      gguf_get_val_u8  (const struct gguf_context * ctx, int key_id);
+    GGML_API int8_t       gguf_get_val_i8  (const struct gguf_context * ctx, int key_id);
+    GGML_API uint16_t     gguf_get_val_u16 (const struct gguf_context * ctx, int key_id);
+    GGML_API int16_t      gguf_get_val_i16 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint32_t     gguf_get_val_u32 (const struct gguf_context * ctx, int key_id);
+    GGML_API int32_t      gguf_get_val_i32 (const struct gguf_context * ctx, int key_id);
+    GGML_API float        gguf_get_val_f32 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint64_t     gguf_get_val_u64 (const struct gguf_context * ctx, int key_id);
+    GGML_API int64_t      gguf_get_val_i64 (const struct gguf_context * ctx, int key_id);
+    GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
+    GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
+    GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id);
+    GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
+    GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
+
+    GGML_API int            gguf_get_n_tensors    (const struct gguf_context * ctx);
+    GGML_API int            gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
+    GGML_API size_t         gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
+    GGML_API char *         gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
+    GGML_API enum ggml_type gguf_get_tensor_type  (const struct gguf_context * ctx, int i);
+
+    // removes key if it exists
+    GGML_API void gguf_remove_key(struct gguf_context * ctx, const char * key);
+
+    // overrides existing values or adds a new one
+    GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
+    GGML_API void gguf_set_val_i8  (struct gguf_context * ctx, const char * key, int8_t   val);
+    GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val);
+    GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t  val);
+    GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
+    GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t  val);
+    GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float    val);
+    GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
+    GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t  val);
+    GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double   val);
+    GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool     val);
+    GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
+    GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);
+    GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, int n);
+
+    // set or add KV pairs from another context
+    GGML_API void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src);
+
+    // manage tensor info
+    GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor);
+    GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type);
+    GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size);
+
+    // writing gguf files can be done in 2 ways:
+    //
+    // - write the entire gguf_context to a binary file in a single pass:
+    //
+    //   gguf_write_to_file(ctx, fname);
+    //
+    // - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data:
+    //
+    //   FILE * f = fopen(fname, "wb");
+    //   fseek(f, gguf_get_meta_size(ctx), SEEK_SET);
+    //   fwrite(f, ...);
+    //   void * data = gguf_meta_get_meta_data(ctx);
+    //   fseek(f, 0, SEEK_SET);
+    //   fwrite(f, data, gguf_get_meta_size(ctx));
+    //   free(data);
+    //   fclose(f);
+    //
+
+    // write the entire context to a binary file
+    GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);
+
+    // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
+    GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
+    GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
+
+    //
+    // system info
+    //
+
+    GGML_API int ggml_cpu_has_avx        (void);
+    GGML_API int ggml_cpu_has_avx_vnni   (void);
+    GGML_API int ggml_cpu_has_avx2       (void);
+    GGML_API int ggml_cpu_has_avx512     (void);
+    GGML_API int ggml_cpu_has_avx512_vbmi(void);
+    GGML_API int ggml_cpu_has_avx512_vnni(void);
+    GGML_API int ggml_cpu_has_avx512_bf16(void);
+    GGML_API int ggml_cpu_has_fma        (void);
+    GGML_API int ggml_cpu_has_neon       (void);
+    GGML_API int ggml_cpu_has_sve        (void);
+    GGML_API int ggml_cpu_has_arm_fma    (void);
+    GGML_API int ggml_cpu_has_metal      (void);
+    GGML_API int ggml_cpu_has_f16c       (void);
+    GGML_API int ggml_cpu_has_fp16_va    (void);
+    GGML_API int ggml_cpu_has_wasm_simd  (void);
+    GGML_API int ggml_cpu_has_blas       (void);
+    GGML_API int ggml_cpu_has_cuda       (void);
+    GGML_API int ggml_cpu_has_vulkan     (void);
+    GGML_API int ggml_cpu_has_kompute    (void);
+    GGML_API int ggml_cpu_has_gpublas    (void);
+    GGML_API int ggml_cpu_has_sse3       (void);
+    GGML_API int ggml_cpu_has_ssse3      (void);
+    GGML_API int ggml_cpu_has_sycl       (void);
+    GGML_API int ggml_cpu_has_rpc        (void);
+    GGML_API int ggml_cpu_has_vsx        (void);
+    GGML_API int ggml_cpu_has_matmul_int8(void);
+    GGML_API int ggml_cpu_has_cann       (void);
+    GGML_API int ggml_cpu_has_llamafile  (void);
+
+    //
+    // Internal types and functions exposed for tests and benchmarks
+    //
+
+#ifdef  __cplusplus
+// restrict not standard in C++
+#define GGML_RESTRICT
+#else
+#define GGML_RESTRICT restrict
+#endif
+    typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+    typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
+    typedef void (*ggml_from_float_to_mat_t)
+                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
+    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                       const void * GGML_RESTRICT y, size_t by, int nrc);
+    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+
+    typedef struct {
+        const char             * type_name;
+        int64_t                  blck_size;
+        int64_t                  blck_size_interleave; // interleave elements in blocks
+        size_t                   type_size;
+        bool                     is_quantized;
+        ggml_to_float_t          to_float;
+        ggml_from_float_t        from_float;
+        ggml_from_float_t        from_float_ref;
+        ggml_from_float_to_mat_t from_float_to_mat;
+        ggml_vec_dot_t           vec_dot;
+        enum ggml_type           vec_dot_type;
+        int64_t                  nrows; // number of rows to process simultaneously
+        int64_t                  ncols; // number of columns to process simultaneously
+        ggml_gemv_t              gemv;
+        ggml_gemm_t              gemm;
+    } ggml_type_traits_t;
+
+    GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/src/CMakeLists.txt b/llm/llama.cpp/ggml/src/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..425a2589502eba4874ddc69ebb5bf63e9240b279
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/CMakeLists.txt
@@ -0,0 +1,1325 @@
+include(CheckCXXCompilerFlag)
+
+unset(GGML_CDEF_PUBLIC)
+
+add_compile_definitions(GGML_SCHED_MAX_COPIES=${GGML_SCHED_MAX_COPIES})
+
+# enable libstdc++ assertions for debug builds
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions($<$:_GLIBCXX_ASSERTIONS>)
+endif()
+
+if (NOT MSVC)
+    if (GGML_SANITIZE_THREAD)
+        add_compile_options(-fsanitize=thread)
+        link_libraries     (-fsanitize=thread)
+    endif()
+
+    if (GGML_SANITIZE_ADDRESS)
+        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+        link_libraries     (-fsanitize=address)
+    endif()
+
+    if (GGML_SANITIZE_UNDEFINED)
+        add_compile_options(-fsanitize=undefined)
+        link_libraries     (-fsanitize=undefined)
+    endif()
+endif()
+
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate)
+    if (ACCELERATE_FRAMEWORK)
+        message(STATUS "Accelerate framework found")
+
+        add_compile_definitions(GGML_USE_ACCELERATE)
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+
+        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
+    else()
+        message(WARNING "Accelerate framework not found")
+    endif()
+endif()
+
+if (GGML_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK    Metal      REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
+
+    message(STATUS "Metal framework found")
+    set(GGML_HEADERS_METAL ../include/ggml-metal.h)
+    set(GGML_SOURCES_METAL ggml-metal.m)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_METAL)
+    if (GGML_METAL_NDEBUG)
+        add_compile_definitions(GGML_METAL_NDEBUG)
+    endif()
+
+    # copy ggml-common.h and ggml-metal.metal to bin directory
+    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
+    configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
+
+    if (GGML_METAL_EMBED_LIBRARY)
+        enable_language(ASM)
+
+        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
+
+        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
+        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+
+        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+
+        # merge ggml-common.h and ggml-metal.metal into a single file
+        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+
+        add_custom_command(
+            OUTPUT ${METALLIB_EMBED_ASM}
+            COMMAND echo "Embedding Metal library"
+            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
+            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+            DEPENDS ggml-metal.metal ggml-common.h
+            COMMENT "Generate assembly for embedded Metal library"
+        )
+
+        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
+    else()
+        if (GGML_METAL_SHADER_DEBUG)
+            # custom command to do the following:
+            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+            #
+            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+            #       disabling fast math is needed in order to pass tests/test-backend-ops
+            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+            set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        else()
+            set(XC_FLAGS -O3)
+        endif()
+
+        # Append macOS metal versioning flags
+        if (GGML_METAL_MACOSX_VERSION_MIN)
+            message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
+            list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
+        endif()
+
+        if (GGML_METAL_STD)
+            message(STATUS "Adding  -std=${GGML_METAL_STD} flag to metal compilation")
+            list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
+        endif()
+
+        add_custom_command(
+            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
+            DEPENDS ggml-metal.metal ggml-common.h
+            COMMENT "Compiling Metal kernels"
+            )
+
+        add_custom_target(
+            ggml-metal ALL
+            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            )
+    endif() # GGML_METAL_EMBED_LIBRARY
+
+    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS}
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        )
+endif()
+
+if (GGML_MUSA)
+    set(CMAKE_C_COMPILER clang)
+    set(CMAKE_C_EXTENSIONS OFF)
+    set(CMAKE_CXX_COMPILER clang++)
+    set(CMAKE_CXX_EXTENSIONS OFF)
+
+    set(GGML_CUDA ON)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_MUSA)
+endif()
+
+if (GGML_OPENMP)
+    find_package(OpenMP)
+    if (OpenMP_FOUND)
+        message(STATUS "OpenMP found")
+
+        add_compile_definitions(GGML_USE_OPENMP)
+
+        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+
+        if (GGML_MUSA)
+            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} "/usr/lib/llvm-10/include/openmp")
+            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} "/usr/lib/llvm-10/lib/libomp.so")
+        endif()
+    else()
+        message(WARNING "OpenMP not found")
+    endif()
+endif()
+
+if (GGML_BLAS)
+    if (GGML_STATIC)
+        set(BLA_STATIC ON)
+    endif()
+    #if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
+    #    set(BLA_SIZEOF_INTEGER 8)
+    #endif()
+
+    set(BLA_VENDOR ${GGML_BLAS_VENDOR})
+    find_package(BLAS)
+
+    if (BLAS_FOUND)
+        message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
+
+        if (("${BLAS_INCLUDE_DIRS}" STREQUAL "") AND NOT (${GGML_BLAS_VENDOR} MATCHES "Apple"))
+            # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+            # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+            find_package(PkgConfig REQUIRED)
+            if (${GGML_BLAS_VENDOR} MATCHES "Generic")
+                pkg_check_modules(DepBLAS REQUIRED blas)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
+                # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+                pkg_check_modules(DepBLAS openblas64)
+                if (NOT DepBLAS_FOUND)
+                    pkg_check_modules(DepBLAS REQUIRED openblas)
+                endif()
+            elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
+                pkg_check_modules(DepBLAS REQUIRED blis)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
+                pkg_check_modules(DepBLAS REQUIRED blas-atlas)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
+                pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
+                # all Intel* libraries share the same include path
+                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
+                # this doesn't provide pkg-config
+                # suggest to assign BLAS_INCLUDE_DIRS on your own
+                if ("${NVHPC_VERSION}" STREQUAL "")
+                    message(WARNING "Better to set NVHPC_VERSION")
+                else()
+                    set(DepBLAS_FOUND ON)
+                    set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+                endif()
+            endif()
+            if (DepBLAS_FOUND)
+                set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+            else()
+                message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+                " detected by pkgconfig, trying to find cblas.h from possible paths...")
+                find_path(BLAS_INCLUDE_DIRS
+                    NAMES cblas.h
+                    HINTS
+                        /usr/include
+                        /usr/local/include
+                        /usr/include/openblas
+                        /opt/homebrew/opt/openblas/include
+                        /usr/local/opt/openblas/include
+                        /usr/include/x86_64-linux-gnu/openblas/include
+                )
+            endif()
+        endif()
+
+        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
+        add_compile_options(${BLAS_LINKER_FLAGS})
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_BLAS)
+
+        if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+            add_compile_definitions(GGML_BLAS_USE_MKL)
+        endif()
+
+        set(GGML_HEADERS_BLAS ../include/ggml-blas.h)
+        set(GGML_SOURCES_BLAS ggml-blas.cpp)
+
+        set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     ${BLAS_LIBRARIES})
+        set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
+    else()
+        message(WARNING "BLAS not found, please refer to "
+        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+        " to set correct GGML_BLAS_VENDOR")
+    endif()
+endif()
+
+if (GGML_LLAMAFILE)
+    message(STATUS "Using llamafile")
+
+    add_compile_definitions(GGML_USE_LLAMAFILE)
+
+    set(GGML_HEADERS_LLAMAFILE llamafile/sgemm.h)
+    set(GGML_SOURCES_LLAMAFILE llamafile/sgemm.cpp)
+endif()
+
+if (GGML_CUDA)
+    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+    if (GGML_MUSA)
+        list(APPEND CMAKE_MODULE_PATH "/usr/local/musa/cmake/")
+        find_package(MUSAToolkit)
+        set(CUDAToolkit_FOUND ${MUSAToolkit_FOUND})
+    else()
+        find_package(CUDAToolkit)
+    endif()
+
+    if (CUDAToolkit_FOUND)
+        message(STATUS "CUDA found")
+
+        if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+            # 52 == lowest CUDA 12 standard
+            # 60 == FP16 CUDA intrinsics
+            # 61 == integer CUDA intrinsics
+            # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
+            if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+                set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+            else()
+                set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+                #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
+            endif()
+        endif()
+        message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+        if (GGML_MUSA)
+            set(CMAKE_CUDA_COMPILER ${MUSAToolkit_MCC_EXECUTABLE})
+        else()
+            enable_language(CUDA)
+        endif()
+
+        file(GLOB   GGML_HEADERS_CUDA "ggml-cuda/*.cuh")
+        list(APPEND GGML_HEADERS_CUDA "../include/ggml-cuda.h")
+
+        file(GLOB   GGML_SOURCES_CUDA "ggml-cuda/*.cu")
+        list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "ggml-cuda/template-instances/mmq*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+
+        if (GGML_CUDA_FA_ALL_QUANTS)
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+        else()
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        endif()
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
+
+        add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
+        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
+        add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
+        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+        if (GGML_CUDA_USE_GRAPHS)
+            add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+        endif()
+
+        if (GGML_CUDA_FORCE_DMMV)
+            add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+        endif()
+
+        if (GGML_CUDA_FORCE_MMQ)
+            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+        endif()
+
+        if (GGML_CUDA_FORCE_CUBLAS)
+            add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+        endif()
+
+        if (GGML_CUDA_NO_VMM)
+            add_compile_definitions(GGML_CUDA_NO_VMM)
+        endif()
+
+        if (DEFINED GGML_CUDA_DMMV_Y)
+            add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
+        endif()
+
+        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            add_compile_definitions(GGML_CUDA_F16)
+        endif()
+
+        if (GGML_CUDA_NO_PEER_COPY)
+            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+        endif()
+
+        if (GGML_MUSA)
+            set_source_files_properties(${GGML_SOURCES_CUDA} PROPERTIES LANGUAGE CXX)
+            foreach(SOURCE ${GGML_SOURCES_CUDA})
+                set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_22")
+            endforeach()
+        endif()
+
+        if (GGML_STATIC)
+            if (WIN32)
+                # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                if (GGML_MUSA)
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart_static MUSA::mublas_static)
+                else()
+                    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                endif()
+            endif()
+        else()
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart MUSA::mublas)
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            endif()
+        endif()
+
+        if (GGML_CUDA_NO_VMM)
+            # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+        else()
+            if (GGML_MUSA)
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
+            else()
+                set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            endif()
+        endif()
+    else()
+        message(WARNING "CUDA not found")
+    endif()
+endif()
+
+if (GGML_HIPBLAS)
+    if (NOT EXISTS $ENV{ROCM_PATH})
+        if (NOT EXISTS /opt/rocm)
+            set(ROCM_PATH /usr)
+        else()
+            set(ROCM_PATH /opt/rocm)
+        endif()
+    else()
+        set(ROCM_PATH $ENV{ROCM_PATH})
+    endif()
+
+    list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
+    list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
+
+    # CMake on Windows doesn't support the HIP language yet
+    if (WIN32)
+        set(CXX_IS_HIPCC TRUE)
+    else()
+        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
+    endif()
+
+    if (CXX_IS_HIPCC)
+        if (LINUX)
+            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+            endif()
+
+            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                    " Prefer setting the HIP compiler directly. See README for details.")
+        endif()
+    else()
+        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+        if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+        endif()
+        cmake_minimum_required(VERSION 3.21)
+        enable_language(HIP)
+    endif()
+
+    find_package(hip     REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+
+    message(STATUS "HIP and hipBLAS found")
+
+    file(GLOB   GGML_HEADERS_ROCM "ggml-cuda/*.cuh")
+    list(APPEND GGML_HEADERS_ROCM "../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_ROCM "ggml-cuda/*.cu")
+    list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
+    file(GLOB   SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "ggml-cuda/template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    endif()
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
+
+    add_compile_definitions(GGML_USE_HIPBLAS)
+    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
+    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
+    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
+
+    if (GGML_HIP_UMA)
+        add_compile_definitions(GGML_HIP_UMA)
+    endif()
+
+    if (GGML_CUDA_FORCE_DMMV)
+        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (CXX_IS_HIPCC)
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} hip::device)
+    else()
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+    endif()
+
+    if (GGML_STATIC)
+        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+    endif()
+
+    set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
+endif()
+
+if (GGML_SYCL)
+    if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA)$")
+        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL or NVIDIA")
+    endif()
+
+    check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
+    if ( DEFINED ENV{ONEAPI_ROOT})
+        message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
+    elseif(SUPPORTS_SYCL)
+        message(WARNING "Using open-source SYCL compiler (clang++). Didn't detect ENV {ONEAPI_ROOT}.
+         If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
+         source /opt/intel/oneapi/setvars.sh")
+    else()
+        message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
+    endif()
+    message(STATUS "SYCL found")
+    #todo: AOT
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_SYCL)
+
+    if (GGML_SYCL_F16)
+        add_compile_definitions(GGML_SYCL_F16)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_SYCL_FORCE_MMQ)
+    endif()
+
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+
+    if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+        add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
+    else()
+        add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
+    endif()
+
+    file(GLOB   GGML_HEADERS_SYCL "ggml-sycl/*.hpp")
+    list(APPEND GGML_HEADERS_SYCL "../include/ggml-sycl.h")
+
+    file(GLOB   GGML_SOURCES_SYCL "ggml-sycl/*.cpp")
+    list(APPEND GGML_SOURCES_SYCL "ggml-sycl.cpp")
+
+    if (WIN32)
+        find_package(IntelSYCL REQUIRED)
+        find_package(MKL REQUIRED)
+        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+    else()
+        if (GGML_SYCL_TARGET STREQUAL "INTEL")
+            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+        elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
+            set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} -fsycl pthread m dl onemkl)
+        endif()
+    endif()
+endif()
+
+if (GGML_RPC)
+    message(STATUS "RPC found")
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_RPC)
+
+    if (WIN32)
+        set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ws2_32)
+    endif()
+
+    set(GGML_HEADERS_RPC ../include/ggml-rpc.h)
+    set(GGML_SOURCES_RPC ggml-rpc.cpp)
+endif()
+
+if (GGML_VULKAN)
+    find_package(Vulkan COMPONENTS glslc REQUIRED)
+
+    if (Vulkan_FOUND)
+        message(STATUS "Vulkan found")
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_VULKAN)
+
+        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
+        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
+        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
+        endif()
+
+        if (GGML_VULKAN_CHECK_RESULTS)
+            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+        endif()
+
+        if (GGML_VULKAN_DEBUG)
+            add_compile_definitions(GGML_VULKAN_DEBUG)
+        endif()
+
+        if (GGML_VULKAN_MEMORY_DEBUG)
+            add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
+        endif()
+
+        if (GGML_VULKAN_VALIDATE)
+            add_compile_definitions(GGML_VULKAN_VALIDATE)
+        endif()
+
+        if (GGML_VULKAN_RUN_TESTS)
+            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+        endif()
+
+        add_subdirectory(vulkan-shaders)
+
+        set (_ggml_vk_genshaders_cmd vulkan-shaders-gen)
+        set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
+        set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
+        set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
+        set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+
+        file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
+
+        add_custom_command(
+            OUTPUT ${_ggml_vk_header}
+                   ${_ggml_vk_source}
+
+            COMMAND ${_ggml_vk_genshaders_cmd}
+                --glslc      ${Vulkan_GLSLC_EXECUTABLE}
+                --input-dir  ${_ggml_vk_input_dir}
+                --output-dir ${_ggml_vk_output_dir}
+                --target-hpp ${_ggml_vk_header}
+                --target-cpp ${_ggml_vk_source}
+                --no-clean
+
+            DEPENDS ${_ggml_vk_shader_deps}
+            COMMENT "Generate vulkan shaders"
+        )
+
+        set(GGML_HEADERS_VULKAN ${CMAKE_CURRENT_SOURCE_DIR}/../include/ggml-vulkan.h ${_ggml_vk_header})
+        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp ${_ggml_vk_source})
+
+        set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS} Vulkan::Vulkan)
+        set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CMAKE_CURRENT_BINARY_DIR})
+    else()
+        message(WARNING "Vulkan not found")
+    endif()
+endif()
+
+if (GGML_KOMPUTE)
+    add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
+
+    find_package(Vulkan COMPONENTS glslc REQUIRED)
+    find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
+
+    if (NOT glslc_executable)
+        message(FATAL_ERROR "glslc not found")
+    endif()
+
+    function(compile_shader)
+        set(options)
+        set(oneValueArgs)
+        set(multiValueArgs SOURCES)
+        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+        foreach(source ${compile_shader_SOURCES})
+            get_filename_component(filename ${source} NAME)
+            set(spv_file ${filename}.spv)
+            add_custom_command(
+                OUTPUT ${spv_file}
+                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                COMMENT "Compiling ${source} to ${spv_file}"
+                )
+
+            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+            set(FILE_NAME "shader${RAW_FILE_NAME}")
+            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+            if(CMAKE_GENERATOR MATCHES "Visual Studio")
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/$/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$/xxd"
+                    )
+            else()
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                    )
+            endif()
+        endforeach()
+    endfunction()
+
+    if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
+        message(STATUS "Kompute found")
+        set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
+        add_subdirectory(kompute)
+
+        # Compile our shaders
+        compile_shader(SOURCES
+            kompute-shaders/op_scale.comp
+            kompute-shaders/op_scale_8.comp
+            kompute-shaders/op_add.comp
+            kompute-shaders/op_addrow.comp
+            kompute-shaders/op_mul.comp
+            kompute-shaders/op_silu.comp
+            kompute-shaders/op_relu.comp
+            kompute-shaders/op_gelu.comp
+            kompute-shaders/op_softmax.comp
+            kompute-shaders/op_norm.comp
+            kompute-shaders/op_rmsnorm.comp
+            kompute-shaders/op_diagmask.comp
+            kompute-shaders/op_mul_mat_mat_f32.comp
+            kompute-shaders/op_mul_mat_f16.comp
+            kompute-shaders/op_mul_mat_q8_0.comp
+            kompute-shaders/op_mul_mat_q4_0.comp
+            kompute-shaders/op_mul_mat_q4_1.comp
+            kompute-shaders/op_mul_mat_q6_k.comp
+            kompute-shaders/op_getrows_f32.comp
+            kompute-shaders/op_getrows_f16.comp
+            kompute-shaders/op_getrows_q4_0.comp
+            kompute-shaders/op_getrows_q4_1.comp
+            kompute-shaders/op_getrows_q6_k.comp
+            kompute-shaders/op_rope_f16.comp
+            kompute-shaders/op_rope_f32.comp
+            kompute-shaders/op_cpy_f16_f16.comp
+            kompute-shaders/op_cpy_f16_f32.comp
+            kompute-shaders/op_cpy_f32_f16.comp
+            kompute-shaders/op_cpy_f32_f32.comp
+        )
+
+        # Create a custom target for our generated shaders
+        add_custom_target(generated_shaders DEPENDS
+            shaderop_scale.h
+            shaderop_scale_8.h
+            shaderop_add.h
+            shaderop_addrow.h
+            shaderop_mul.h
+            shaderop_silu.h
+            shaderop_relu.h
+            shaderop_gelu.h
+            shaderop_softmax.h
+            shaderop_norm.h
+            shaderop_rmsnorm.h
+            shaderop_diagmask.h
+            shaderop_mul_mat_mat_f32.h
+            shaderop_mul_mat_f16.h
+            shaderop_mul_mat_q8_0.h
+            shaderop_mul_mat_q4_0.h
+            shaderop_mul_mat_q4_1.h
+            shaderop_mul_mat_q6_k.h
+            shaderop_getrows_f32.h
+            shaderop_getrows_f16.h
+            shaderop_getrows_q4_0.h
+            shaderop_getrows_q4_1.h
+            shaderop_getrows_q6_k.h
+            shaderop_rope_f16.h
+            shaderop_rope_f32.h
+            shaderop_cpy_f16_f16.h
+            shaderop_cpy_f16_f32.h
+            shaderop_cpy_f32_f16.h
+            shaderop_cpy_f32_f32.h
+        )
+
+        # Create a custom command that depends on the generated_shaders
+        add_custom_command(
+            OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+            COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+            DEPENDS generated_shaders
+            COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
+        )
+
+        # Add the stamp to the main sources to ensure dependency tracking
+        set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp           ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+        set(GGML_HEADERS_KOMPUTE ../include/ggml-kompute.h  ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_KOMPUTE)
+
+        set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     kompute)
+        set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CMAKE_CURRENT_BINARY_DIR})
+    else()
+        message(WARNING "Kompute not found")
+    endif()
+endif()
+
+if (GGML_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+
+    message(STATUS "Using memkind for CPU HBM")
+
+    add_compile_definitions(GGML_USE_CPU_HBM)
+
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
+
+if (GGML_CANN)
+    if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
+        set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
+        message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
+    endif()
+
+    if (CANN_INSTALL_DIR)
+        # Only Support Linux.
+        if (GGML_CANN)
+            if (NOT UNIX)
+                set(GGML_CANN OFF)
+                message(WARNING "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}. Turning off GGML_CANN")
+            endif()
+        endif()
+
+        # Supported platforms: x86-64, arm64
+        if (GGML_CANN)
+            if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+            elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
+            else()
+                set(GGML_CANN OFF)
+                message(WARNING "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}. Turning off GGML_CANN")
+            endif()
+        endif()
+
+        # Set header and libs
+        if(GGML_CANN)
+            set(CANN_INCLUDE_DIRS
+                ${CANN_INSTALL_DIR}/include
+                ${CANN_INSTALL_DIR}/include/aclnn
+                ${CANN_INSTALL_DIR}/acllib/include
+            )
+
+            add_subdirectory(ggml-cann/kernels)
+            list(APPEND CANN_LIBRARIES
+                ascendcl
+                nnopbase
+                opapi
+                acl_op_compiler
+                ascendc_kernels
+            )
+
+            set(GGML_HEADERS_CANN "../include/ggml-cann.h")
+            file(GLOB GGML_SOURCES_CANN "ggml-cann/*.cpp")
+            list(APPEND GGML_SOURCES_CANN "ggml-cann.cpp")
+
+            message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
+            message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
+
+            set(GGML_EXTRA_LIBS     ${GGML_EXTRA_LIBS}     ${CANN_LIBRARIES} )
+            set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CANN_INCLUDE_DIRS})
+            set(GGML_EXTRA_LIBDIRS  ${GGML_EXTRA_LIBDIRS}  ${CANN_INSTALL_DIR}/lib64)
+            list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
+        endif()
+    else()
+        set(GGML_CANN OFF)
+        message(WARNING "CANN: Can't find CANN_INSTALL_DIR, do you forget to source set_var.sh. Turning off GGML_CANN")
+    endif()
+
+    if(NOT GGML_CANN)
+        message(WARNING "CANN: GGML_CANN is turned OFF, see above for details.")
+    endif()
+endif()
+
+function(get_flags CCID CCVER)
+    set(C_FLAGS "")
+    set(CXX_FLAGS "")
+
+    if (CCID MATCHES "Clang")
+        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
+        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
+
+        if (
+            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
+            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
+        )
+            list(APPEND C_FLAGS -Wdouble-promotion)
+        endif()
+    elseif (CCID STREQUAL "GNU")
+        set(C_FLAGS   -Wdouble-promotion)
+        set(CXX_FLAGS -Wno-array-bounds)
+
+        if (NOT GGML_MUSA)
+            if (CCVER VERSION_GREATER_EQUAL 7.1.0)
+                list(APPEND CXX_FLAGS -Wno-format-truncation)
+            endif()
+        endif()
+        if (CCVER VERSION_GREATER_EQUAL 8.1.0)
+            list(APPEND CXX_FLAGS -Wextra-semi)
+        endif()
+    endif()
+
+    set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
+    set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
+endfunction()
+
+if (GGML_FATAL_WARNINGS)
+    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+        list(APPEND C_FLAGS   -Werror)
+        list(APPEND CXX_FLAGS -Werror)
+    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/WX)
+    endif()
+endif()
+
+if (GGML_ALL_WARNINGS)
+    if (NOT MSVC)
+        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                                  -Werror=implicit-int -Werror=implicit-function-declaration)
+        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+
+        list(APPEND C_FLAGS   ${WARNING_FLAGS})
+        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+
+        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+
+        add_compile_options("$<$:${C_FLAGS};${GF_C_FLAGS}>"
+                            "$<$:${CXX_FLAGS};${GF_CXX_FLAGS}>")
+    else()
+        # todo : msvc
+        set(C_FLAGS   "")
+        set(CXX_FLAGS "")
+    endif()
+endif()
+
+set(CUDA_CXX_FLAGS "")
+
+if (GGML_CUDA)
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+endif()
+
+if (GGML_LTO)
+    include(CheckIPOSupported)
+    check_ipo_supported(RESULT result OUTPUT output)
+    if (result)
+        set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
+    else()
+        message(WARNING "IPO is not supported: ${output}")
+    endif()
+endif()
+
+if (GGML_CCACHE)
+    find_program(GGML_CCACHE_FOUND ccache)
+
+    if (GGML_CCACHE_FOUND)
+        # TODO: should not be set globally
+        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
+        set(ENV{CCACHE_SLOPPINESS} time_macros)
+        message(STATUS "ccache found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
+    else()
+        message(STATUS "Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF")
+    endif ()
+endif()
+
+# this version of Apple ld64 is buggy
+execute_process(
+    COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
+    ERROR_VARIABLE output
+    OUTPUT_QUIET
+)
+
+if (output MATCHES "dyld-1015\.7")
+    add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
+endif()
+
+# architecture specific
+# TODO: probably these flags need to be tweaked on some architectures
+#       feel free to update the Makefile for your architecture and send a pull request or issue
+message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+if (MSVC)
+    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+else ()
+    set(CMAKE_GENERATOR_PLATFORM_LWR "")
+endif ()
+
+if (NOT MSVC)
+    if (GGML_STATIC)
+        add_link_options(-static)
+        if (MINGW)
+            add_link_options(-static-libgcc -static-libstdc++)
+        endif()
+    endif()
+    if (GGML_GPROF)
+        add_compile_options(-pg)
+    endif()
+endif()
+
+set(ARCH_FLAGS "")
+
+if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+    CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+    (NOT CMAKE_OSX_ARCHITECTURES      AND
+     NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+
+    message(STATUS "ARM detected")
+
+    if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+
+        check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+
+        check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        endif ()
+
+        check_cxx_source_compiles("#include \nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+    else()
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+            # Raspberry Pi 1, Zero
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
+                # Android armeabi-v7a
+                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
+            else()
+                # Raspberry Pi 2
+                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            endif()
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Android arm64-v8a
+            # Raspberry Pi 3, 4, Zero 2 (32-bit)
+            list(APPEND ARCH_FLAGS -mno-unaligned-access)
+        endif()
+        if (GGML_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
+    endif()
+elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
+    message(STATUS "x86 detected")
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(../cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$:__AVX512VBMI__>)
+                add_compile_definitions($<$:__AVX512VBMI__>)
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$:__AVX512VNNI__>)
+                add_compile_definitions($<$:__AVX512VNNI__>)
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$:__AVX512BF16__>)
+                add_compile_definitions($<$:__AVX512BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+    message(STATUS "PowerPC detected")
+    if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+    else()
+        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
+        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    message(STATUS "loongarch64 detected")
+
+    list(APPEND ARCH_FLAGS -march=loongarch64)
+    if (GGML_LASX)
+        list(APPEND ARCH_FLAGS -mlasx)
+    endif()
+    if (GGML_LSX)
+        list(APPEND ARCH_FLAGS -mlsx)
+    endif()
+else()
+    message(STATUS "Unknown architecture")
+endif()
+
+add_compile_options("$<$:${ARCH_FLAGS}>")
+add_compile_options("$<$:${ARCH_FLAGS}>")
+
+if (GGML_CUDA)
+    list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    add_compile_options("$<$:${CUDA_FLAGS}>")
+endif()
+
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (
+    CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
+    CMAKE_SYSTEM_NAME MATCHES "iOS"    OR
+    CMAKE_SYSTEM_NAME MATCHES "tvOS"   OR
+    CMAKE_SYSTEM_NAME MATCHES "DragonFly"
+)
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
+if (WIN32)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+
+    if (BUILD_SHARED_LIBS)
+        # TODO: should not use this
+        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    endif()
+endif()
+
+#
+# libraries
+#
+
+# ggml
+
+add_library(ggml
+            ../include/ggml.h
+            ../include/ggml-alloc.h
+            ../include/ggml-backend.h
+            ggml.c
+            ggml-alloc.c
+            ggml-backend.c
+            ggml-quants.c
+            ggml-quants.h
+            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
+            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
+            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_BLAS}      ${GGML_HEADERS_BLAS}
+            ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
+            ${GGML_SOURCES_CANN}      ${GGML_HEADERS_CANN}
+            ggml-aarch64.c            ggml-aarch64.h
+            )
+
+if (EMSCRIPTEN)
+    set_target_properties(ggml PROPERTIES COMPILE_FLAGS "-msimd128")
+endif()
+
+target_compile_definitions(ggml PUBLIC  ${GGML_CDEF_PUBLIC})
+target_include_directories(ggml PUBLIC ../include)
+target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
+target_link_directories(ggml PRIVATE ${GGML_EXTRA_LIBDIRS})
+target_compile_features   (ggml PRIVATE c_std_11) # don't bump
+
+target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS})
+
+find_library(MATH_LIBRARY m)
+if (MATH_LIBRARY)
+    if (NOT WIN32 OR NOT GGML_SYCL)
+        target_link_libraries(ggml PRIVATE ${MATH_LIBRARY})
+    endif()
+endif()
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(ggml PRIVATE GGML_SHARED GGML_BUILD)
+endif()
diff --git a/llm/llama.cpp/ggml/src/ggml-aarch64.c b/llm/llama.cpp/ggml/src/ggml-aarch64.c
new file mode 100644
index 0000000000000000000000000000000000000000..d7a608997380a17c761dab2f482d614211dc3537
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-aarch64.c
@@ -0,0 +1,2193 @@
+// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+
+#include 
+#include 
+#include 
+#include 
+#include  // for qsort
+#include   // for GGML_ASSERT
+
+#include "ggml-aarch64.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+// Functions to create the interleaved data layout formats
+
+// interleave 4 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x4
+// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
+// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
+//
+// - in                  : an array of block_q4_0 pointers
+// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
+//                         blck_size_interleave bytes
+// - xor_mask            : the mask to convert the nibbles in block_q4_0 quants bytes
+//                         from bias offset form to pure sign form (this saves subtract
+//                         operations durin unpacking)
+//
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+    block_q4_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    for (int i = 0; i < QK4_0 * 2; i++) {
+        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
+
+        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    }
+
+    return out;
+}
+
+// interleave 8 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x8
+// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
+// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+    block_q4_0x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    for (int i = 0; i < QK4_0 * 4; i++) {
+        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
+
+        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    }
+
+    return out;
+}
+
+void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 8; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 4; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][2 * j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][2 * j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][2 * j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    if (blck_size_interleave == 4) {
+        quantize_q8_0_4x4(x, vy, n_per_row);
+    } else if (blck_size_interleave == 8) {
+        quantize_q8_0_4x8(x, vy, n_per_row);
+    } else {
+        assert(false);
+    }
+}
+
+static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
+    assert(n_per_row % QK4_0 == 0);
+    const int nb = n_per_row / QK4_0;
+
+    void * out_ptr = NULL;
+    if (nrows_interleaved == 8) {
+        out_ptr = (block_q4_0x8 *) dst;
+    }
+    else if (nrows_interleaved == 4) {
+        out_ptr = (block_q4_0x4 *) dst;
+    }
+    assert(nrows_interleaved <= 8);
+    block_q4_0 dst_tmp[8];
+
+    for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
+
+        for (int64_t x = 0; x < nb; x++) {
+
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
+            }
+
+            if (nrows_interleaved == 8) {
+                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
+                out_ptr = (block_q4_0x8 *) out_ptr + 1;
+            }
+            else if (nrows_interleaved == 4) {
+                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
+                out_ptr = (block_q4_0x4 *) out_ptr + 1;
+            }
+        }
+    }
+
+    return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
+}
+
+size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
+    }
+    else {
+        assert(false);
+        return 0;
+    }
+}
+
+size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
+    }
+    else {
+        assert(false);
+        return 0;
+    }
+}
+
+size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
+    }
+    else {
+        assert(false);
+        return 0;
+    }
+}
+
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
+                "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+
+    __asm__ __volatile__(
+        "movi v31.16b, #0x4\n"
+        "movi v30.16b, #0xf0\n"
+        "add %x[b_ptr], %x[b_ptr], #0x8\n"
+        "1:"  // Column loop
+        "add x22, %x[a_ptr], #0x2\n"
+        "movi v29.16b, #0x0\n"
+        "mov x21, %x[nb]\n"
+        "2:"  // Block loop
+        "ldr q28, [%x[b_ptr], #0x0]\n"
+        "ldr q27, [x22, #0x0]\n"
+        "movi v26.4s, #0x0\n"
+        "sub x20, x22, #0x2\n"
+        "ldr q25, [x22, #0x10]\n"
+        "ldr q24, [%x[b_ptr], #0x10]\n"
+        "sub x21, x21, #0x1\n"
+        "add x22, x22, #0x22\n"
+        "ldr q23, [%x[b_ptr], #0x20]\n"
+        "ldr q22, [%x[b_ptr], #0x30]\n"
+        "ld1r { v21.8h }, [x20]\n"
+        "ldr q20, [%x[b_ptr], #-0x8]\n"
+        "sshl v16.16b, v28.16b, v31.16b\n"
+        "and v28.16b, v28.16b, v30.16b\n"
+        "sshl v19.16b, v24.16b, v31.16b\n"
+        "and v24.16b, v24.16b, v30.16b\n"
+        "add %x[b_ptr], %x[b_ptr], #0x48\n"
+        "sshl v18.16b, v23.16b, v31.16b\n"
+        "and v23.16b, v23.16b, v30.16b\n"
+        ".inst 0x4f9be21a  // sdot v26.4s, v16.16b, v27.4b[0]\n"
+        "sshl v17.16b, v22.16b, v31.16b\n"
+        "and v22.16b, v22.16b, v30.16b\n"
+        "fcvtl v21.4s, v21.4h\n"
+        "fcvtl v16.4s, v20.4h\n"
+        ".inst 0x4f99e39a  // sdot v26.4s, v28.16b, v25.4b[0]\n"
+        "fmul v16.4s, v16.4s, v21.4s\n"
+        ".inst 0x4fbbe27a  // sdot v26.4s, v19.16b, v27.4b[1]\n"
+        ".inst 0x4fb9e31a  // sdot v26.4s, v24.16b, v25.4b[1]\n"
+        ".inst 0x4f9bea5a  // sdot v26.4s, v18.16b, v27.4b[2]\n"
+        ".inst 0x4f99eafa  // sdot v26.4s, v23.16b, v25.4b[2]\n"
+        ".inst 0x4fbbea3a  // sdot v26.4s, v17.16b, v27.4b[3]\n"
+        ".inst 0x4fb9eada  // sdot v26.4s, v22.16b, v25.4b[3]\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "fmla v29.4s, v26.4s, v16.4s\n"
+        "cbnz x21, 2b\n"
+        "sub %x[nc], %x[nc], #0x4\n"
+        "str q29, [%x[res_ptr], #0x0]\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "cbnz %x[nc], 1b\n"
+        : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+        : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+        : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22"
+    );
+#else
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+#endif
+}
+
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+
+    __asm__ __volatile__(
+        "movi v2.16b, #0x4\n"
+        "movi v1.16b, #0xf0\n"
+        "add %x[b_ptr], %x[b_ptr], #0x8\n"
+        "1:"  // Column loop
+        "add x23, %x[a_ptr], #0x2\n"
+        "movi v0.16b, #0x0\n"
+        "mov x22, %x[nb]\n"
+        "2:"  // Block loop
+        "ldr q31, [%x[b_ptr], #0x0]\n"
+        "ldr q30, [%x[b_ptr], #0x10]\n"
+        "mov x21, x23\n"
+        "movi v29.4s, #0x0\n"
+        "ldr q28, [%x[b_ptr], #0x20]\n"
+        "ldr q27, [%x[b_ptr], #0x30]\n"
+        "movi v26.4s, #0x0\n"
+        "sub x20, x23, #0x2\n"
+        "ld1r { v25.8h }, [x20]\n"
+        "ldr q24, [%x[b_ptr], #-0x8]\n"
+        "sub x22, x22, #0x1\n"
+        "add x23, x23, #0x22\n"
+        "ld1r { v23.2d }, [x21], #0x8\n"
+        "sshl v22.16b, v31.16b, v2.16b\n"
+        "sshl v16.16b, v30.16b, v2.16b\n"
+        "add %x[b_ptr], %x[b_ptr], #0x48\n"
+        "ld1r { v21.2d }, [x21], #0x8\n"
+        "sshl v20.16b, v28.16b, v2.16b\n"
+        "sshl v19.16b, v27.16b, v2.16b\n"
+        "ld1r { v18.2d }, [x21], #0x8\n"
+        "ld1r { v17.2d }, [x21], #0x8\n"
+        "and v31.16b, v31.16b, v1.16b\n"
+        "and v30.16b, v30.16b, v1.16b\n"
+        ".inst 0x4e9796dd  // sdot v29.4s, v22.16b, v23.16b\n"
+        ".inst 0x4e97961a  // sdot v26.4s, v16.16b, v23.16b\n"
+        "and v28.16b, v28.16b, v1.16b\n"
+        "and v27.16b, v27.16b, v1.16b\n"
+        "fcvtl v25.4s, v25.4h\n"
+        "fcvtl v16.4s, v24.4h\n"
+        ".inst 0x4e95969d  // sdot v29.4s, v20.16b, v21.16b\n"
+        ".inst 0x4e95967a  // sdot v26.4s, v19.16b, v21.16b\n"
+        "fmul v16.4s, v16.4s, v25.4s\n"
+        ".inst 0x4e9297fd  // sdot v29.4s, v31.16b, v18.16b\n"
+        ".inst 0x4e9297da  // sdot v26.4s, v30.16b, v18.16b\n"
+        ".inst 0x4e91979d  // sdot v29.4s, v28.16b, v17.16b\n"
+        ".inst 0x4e91977a  // sdot v26.4s, v27.16b, v17.16b\n"
+        "addp v29.4s, v29.4s, v26.4s\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "fmla v0.4s, v29.4s, v16.4s\n"
+        "cbnz x22, 2b\n"
+        "sub %x[nc], %x[nc], #0x4\n"
+        "str q0, [%x[res_ptr], #0x0]\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "cbnz %x[nc], 1b\n"
+        : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+        : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+        : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
+    );
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+#endif
+}
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    if (ggml_sve_cnt_b == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+
+        __asm__ __volatile__(
+            "ptrue p0.b\n"
+            "add %x[b_ptr], %x[b_ptr], #0x10\n"
+            "1:"  // Column loop
+            "add x22, %x[a_ptr], #0x2\n"
+            "mov z31.b, #0x0\n"
+            "mov x21, %x[nb]\n"
+            "2:"  // Block loop
+            "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n"
+            "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n"
+            "mov z28.s, #0x0\n"
+            "mov z27.s, #0x0\n"
+            "ld1rd { z26.d }, p0/Z, [x22]\n"
+            "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n"
+            "sub x20, x22, #0x2\n"
+            "sub x21, x21, #0x1\n"
+            "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n"
+            "ld1rd { z23.d }, p0/Z, [x22, #8]\n"
+            "lsl z22.b, z30.b, #0x4\n"
+            "lsl z16.b, z29.b, #0x4\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1rd { z21.d }, p0/Z, [x22, #16]\n"
+            "ld1rd { z20.d }, p0/Z, [x22, #24]\n"
+            "lsl z19.b, z25.b, #0x4\n"
+            "and z25.b, z25.b, #0xf0\n"
+            "ld1rh { z17.h }, p0/Z, [x20]\n"
+            "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n"
+            "sdot z28.s, z22.b, z26.b\n"
+            "sdot z27.s, z16.b, z26.b\n"
+            "lsl z16.b, z24.b, #0x4\n"
+            "add x22, x22, #0x22\n"
+            "and z24.b, z24.b, #0xf0\n"
+            "add %x[b_ptr], %x[b_ptr], #0x90\n"
+            "fcvt z17.s, p0/m, z17.h\n"
+            "fcvt z18.s, p0/m, z18.h\n"
+            "sdot z28.s, z19.b, z23.b\n"
+            "sdot z27.s, z16.b, z23.b\n"
+            "fmul z18.s, z18.s, z17.s\n"
+            "sdot z28.s, z30.b, z21.b\n"
+            "sdot z27.s, z29.b, z21.b\n"
+            "sdot z28.s, z25.b, z20.b\n"
+            "sdot z27.s, z24.b, z20.b\n"
+            "uzp1 z17.s, z28.s, z27.s\n"
+            "uzp2 z16.s, z28.s, z27.s\n"
+            "add z17.s, z17.s, z16.s\n"
+            "asr z17.s, z17.s, #0x4\n"
+            "scvtf z17.s, p0/m, z17.s\n"
+            "fmla z31.s, p0/M, z17.s, z18.s\n"
+            "cbnz x21, 2b\n"
+            "sub %x[nc], %x[nc], #0x8\n"
+            "st1w { z31.s }, p0, [%x[res_ptr]]\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "cbnz %x[nc], 1b\n"
+            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+            : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+    else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
+                    "performance");
+    }
+    else if (ggml_cpu_has_neon()) {
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
+                    "quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(ggml_cpu_has_sve() &&
+                "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+#endif
+}
+
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
+                "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+    size_t res_stride = bs * sizeof(float);
+
+    __asm__ __volatile__(
+        "mov x10, %x[nr]\n"
+        "mov x9, #0x88\n"
+        "cmp x10, #0x10\n"
+        "mul x9, %x[nb], x9\n"
+        "blt 4f\n"
+        "1:"  // Row loop
+        "add x28, %x[b_ptr], #0x8\n"
+        "mov x27, %x[nc]\n"
+        "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+        "2:"  // Column loop
+        "add x25, %x[a_ptr], #0x8\n"
+        "movi v15.16b, #0x0\n"
+        "movi v19.16b, #0x0\n"
+        "mov x24, %x[nb]\n"
+        "add x23, x25, x9\n"
+        "movi v18.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "add x22, x23, x9\n"
+        "movi v11.16b, #0x0\n"
+        "movi v13.16b, #0x0\n"
+        "add x21, x22, x9\n"
+        "movi v23.16b, #0x0\n"
+        "movi v16.16b, #0x0\n"
+        "movi v25.16b, #0x0\n"
+        "movi v7.16b, #0x0\n"
+        "movi v0.16b, #0x0\n"
+        "movi v4.16b, #0x0\n"
+        "movi v5.16b, #0x0\n"
+        "movi v21.16b, #0x0\n"
+        "movi v8.16b, #0x0\n"
+        "movi v1.16b, #0x0\n"
+        "3:"  // Block loop
+        "ldr q3, [x28, #0x0]\n"
+        "ldr q31, [x25, #0x0]\n"
+        "movi v28.16b, #0x4\n"
+        "movi v10.4s, #0x0\n"
+        "ldr q22, [x28, #0x10]\n"
+        "ldr q6, [x25, #0x10]\n"
+        "movi v29.4s, #0x0\n"
+        "movi v9.4s, #0x0\n"
+        "ldr q27, [x28, #0x20]\n"
+        "ldr q30, [x28, #0x30]\n"
+        "movi v20.4s, #0x0\n"
+        "movi v24.16b, #0xf0\n"
+        "ldr d2, [x25, #-0x8]\n"
+        "ldr d26, [x23, #-0x8]\n"
+        "sshl v12.16b, v3.16b, v28.16b\n"
+        "sub x20, x28, #0x8\n"
+        "ldr d17, [x20, #0x0]\n"
+        "and v3.16b, v3.16b, v24.16b\n"
+        "subs x24, x24, #0x1\n"
+        "add x28, x28, #0x48\n"
+        ".inst 0x4f9fe18a  // sdot v10.4s, v12.16b, v31.4b[0]\n"
+        ".inst 0x4fbfe19d  // sdot v29.4s, v12.16b, v31.4b[1]\n"
+        ".inst 0x4f9fe989  // sdot v9.4s, v12.16b, v31.4b[2]\n"
+        ".inst 0x4fbfe994  // sdot v20.4s, v12.16b, v31.4b[3]\n"
+        "sshl v31.16b, v22.16b, v28.16b\n"
+        "and v22.16b, v22.16b, v24.16b\n"
+        "fcvtl v17.4s, v17.4h\n"
+        "fcvtl v2.4s, v2.4h\n"
+        "fcvtl v26.4s, v26.4h\n"
+        ".inst 0x4f86e3ea  // sdot v10.4s, v31.16b, v6.4b[0]\n"
+        ".inst 0x4fa6e3fd  // sdot v29.4s, v31.16b, v6.4b[1]\n"
+        ".inst 0x4f86ebe9  // sdot v9.4s, v31.16b, v6.4b[2]\n"
+        ".inst 0x4fa6ebf4  // sdot v20.4s, v31.16b, v6.4b[3]\n"
+        "sshl v6.16b, v27.16b, v28.16b\n"
+        "sshl v28.16b, v30.16b, v28.16b\n"
+        "and v27.16b, v27.16b, v24.16b\n"
+        "and v30.16b, v30.16b, v24.16b\n"
+        "ldr q24, [x25, #0x20]\n"
+        ".inst 0x4f98e0ca  // sdot v10.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8c9  // sdot v9.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8d4  // sdot v20.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x30]\n"
+        ".inst 0x4f98e38a  // sdot v10.4s, v28.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e39d  // sdot v29.4s, v28.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb89  // sdot v9.4s, v28.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb94  // sdot v20.4s, v28.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x40]\n"
+        ".inst 0x4f98e06a  // sdot v10.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e869  // sdot v9.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e874  // sdot v20.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x50]\n"
+        ".inst 0x4f98e2ca  // sdot v10.4s, v22.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e2dd  // sdot v29.4s, v22.16b, v24.4b[1]\n"
+        ".inst 0x4f98eac9  // sdot v9.4s, v22.16b, v24.4b[2]\n"
+        ".inst 0x4fb8ead4  // sdot v20.4s, v22.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x60]\n"
+        ".inst 0x4f98e36a  // sdot v10.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb69  // sdot v9.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb74  // sdot v20.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x70]\n"
+        "add x25, x25, #0x88\n"
+        ".inst 0x4f98e3ca  // sdot v10.4s, v30.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e3dd  // sdot v29.4s, v30.16b, v24.4b[1]\n"
+        ".inst 0x4f98ebc9  // sdot v9.4s, v30.16b, v24.4b[2]\n"
+        ".inst 0x4fb8ebd4  // sdot v20.4s, v30.16b, v24.4b[3]\n"
+        "fmul v24.4s, v17.4s, v2.s[0]\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v15.4s, v10.4s, v24.4s\n"
+        "ldr q24, [x23, #0x0]\n"
+        "fmul v10.4s, v17.4s, v2.s[1]\n"
+        "fmla v19.4s, v29.4s, v10.4s\n"
+        "ldr q10, [x23, #0x10]\n"
+        "fmul v29.4s, v17.4s, v2.s[2]\n"
+        "fmul v2.4s, v17.4s, v2.s[3]\n"
+        "fmla v18.4s, v9.4s, v29.4s\n"
+        "movi v9.4s, #0x0\n"
+        "movi v29.4s, #0x0\n"
+        ".inst 0x4f98e189  // sdot v9.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e19d  // sdot v29.4s, v12.16b, v24.4b[1]\n"
+        "fmla v14.4s, v20.4s, v2.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v2.4s, #0x0\n"
+        ".inst 0x4f98e994  // sdot v20.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x20]\n"
+        ".inst 0x4f8ae3e9  // sdot v9.4s, v31.16b, v10.4b[0]\n"
+        ".inst 0x4faae3fd  // sdot v29.4s, v31.16b, v10.4b[1]\n"
+        ".inst 0x4f8aebf4  // sdot v20.4s, v31.16b, v10.4b[2]\n"
+        ".inst 0x4faaebe2  // sdot v2.4s, v31.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x30]\n"
+        ".inst 0x4f98e0c9  // sdot v9.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8d4  // sdot v20.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x40]\n"
+        ".inst 0x4f8ae389  // sdot v9.4s, v28.16b, v10.4b[0]\n"
+        ".inst 0x4faae39d  // sdot v29.4s, v28.16b, v10.4b[1]\n"
+        ".inst 0x4f8aeb94  // sdot v20.4s, v28.16b, v10.4b[2]\n"
+        ".inst 0x4faaeb82  // sdot v2.4s, v28.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x50]\n"
+        ".inst 0x4f98e069  // sdot v9.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e874  // sdot v20.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x60]\n"
+        ".inst 0x4f8ae2c9  // sdot v9.4s, v22.16b, v10.4b[0]\n"
+        ".inst 0x4faae2dd  // sdot v29.4s, v22.16b, v10.4b[1]\n"
+        ".inst 0x4f8aead4  // sdot v20.4s, v22.16b, v10.4b[2]\n"
+        ".inst 0x4faaeac2  // sdot v2.4s, v22.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x70]\n"
+        "add x23, x23, #0x88\n"
+        ".inst 0x4f98e369  // sdot v9.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb74  // sdot v20.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x0]\n"
+        ".inst 0x4f8ae3c9  // sdot v9.4s, v30.16b, v10.4b[0]\n"
+        ".inst 0x4faae3dd  // sdot v29.4s, v30.16b, v10.4b[1]\n"
+        ".inst 0x4f8aebd4  // sdot v20.4s, v30.16b, v10.4b[2]\n"
+        ".inst 0x4faaebc2  // sdot v2.4s, v30.16b, v10.4b[3]\n"
+        "fmul v10.4s, v17.4s, v26.s[0]\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "fmla v11.4s, v9.4s, v10.4s\n"
+        "ldr q9, [x22, #0x10]\n"
+        "fmul v10.4s, v17.4s, v26.s[1]\n"
+        "fmla v13.4s, v29.4s, v10.4s\n"
+        "ldr d29, [x22, #-0x8]\n"
+        "fmul v10.4s, v17.4s, v26.s[2]\n"
+        "fmul v26.4s, v17.4s, v26.s[3]\n"
+        "fcvtl v29.4s, v29.4h\n"
+        "fmla v23.4s, v20.4s, v10.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v10.4s, #0x0\n"
+        "fmla v16.4s, v2.4s, v26.4s\n"
+        "movi v26.4s, #0x0\n"
+        "movi v2.4s, #0x0\n"
+        ".inst 0x4f98e194  // sdot v20.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+        ".inst 0x4f98e99a  // sdot v26.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x20]\n"
+        ".inst 0x4f89e3f4  // sdot v20.4s, v31.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebfa  // sdot v26.4s, v31.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebe2  // sdot v2.4s, v31.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x30]\n"
+        ".inst 0x4f98e0d4  // sdot v20.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0ca  // sdot v10.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8da  // sdot v26.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x40]\n"
+        ".inst 0x4f89e394  // sdot v20.4s, v28.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+        ".inst 0x4f89eb9a  // sdot v26.4s, v28.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eb82  // sdot v2.4s, v28.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x50]\n"
+        ".inst 0x4f98e074  // sdot v20.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e06a  // sdot v10.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e87a  // sdot v26.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x60]\n"
+        ".inst 0x4f89e2d4  // sdot v20.4s, v22.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+        ".inst 0x4f89eada  // sdot v26.4s, v22.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eac2  // sdot v2.4s, v22.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x70]\n"
+        "add x22, x22, #0x88\n"
+        ".inst 0x4f98e374  // sdot v20.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e36a  // sdot v10.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb7a  // sdot v26.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x21, #0x0]\n"
+        ".inst 0x4f89e3d4  // sdot v20.4s, v30.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ca  // sdot v10.4s, v30.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebda  // sdot v26.4s, v30.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebc2  // sdot v2.4s, v30.16b, v9.4b[3]\n"
+        "fmul v9.4s, v17.4s, v29.s[0]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "fmla v25.4s, v20.4s, v9.4s\n"
+        "ldr q9, [x21, #0x10]\n"
+        "fmul v20.4s, v17.4s, v29.s[1]\n"
+        "fmla v7.4s, v10.4s, v20.4s\n"
+        "ldr d20, [x21, #-0x8]\n"
+        "fmul v10.4s, v17.4s, v29.s[2]\n"
+        "fmul v29.4s, v17.4s, v29.s[3]\n"
+        "fcvtl v20.4s, v20.4h\n"
+        "fmla v0.4s, v26.4s, v10.4s\n"
+        "movi v26.4s, #0x0\n"
+        "movi v10.4s, #0x0\n"
+        "fmla v4.4s, v2.4s, v29.4s\n"
+        "movi v2.4s, #0x0\n"
+        "movi v29.4s, #0x0\n"
+        ".inst 0x4f98e19a  // sdot v26.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+        ".inst 0x4f98e982  // sdot v2.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e99d  // sdot v29.4s, v12.16b, v24.4b[3]\n"
+        "ldr q12, [x21, #0x20]\n"
+        "fmul v24.4s, v17.4s, v20.s[0]\n"
+        ".inst 0x4f89e3fa  // sdot v26.4s, v31.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebe2  // sdot v2.4s, v31.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebfd  // sdot v29.4s, v31.16b, v9.4b[3]\n"
+        "ldr q9, [x21, #0x30]\n"
+        "fmul v31.4s, v17.4s, v20.s[1]\n"
+        ".inst 0x4f8ce0da  // sdot v26.4s, v6.16b, v12.4b[0]\n"
+        ".inst 0x4face0ca  // sdot v10.4s, v6.16b, v12.4b[1]\n"
+        ".inst 0x4f8ce8c2  // sdot v2.4s, v6.16b, v12.4b[2]\n"
+        ".inst 0x4face8dd  // sdot v29.4s, v6.16b, v12.4b[3]\n"
+        "ldr q12, [x21, #0x40]\n"
+        "fmul v6.4s, v17.4s, v20.s[2]\n"
+        "fmul v20.4s, v17.4s, v20.s[3]\n"
+        ".inst 0x4f89e39a  // sdot v26.4s, v28.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+        ".inst 0x4f89eb82  // sdot v2.4s, v28.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eb9d  // sdot v29.4s, v28.16b, v9.4b[3]\n"
+        "ldr q9, [x21, #0x50]\n"
+        ".inst 0x4f8ce07a  // sdot v26.4s, v3.16b, v12.4b[0]\n"
+        ".inst 0x4face06a  // sdot v10.4s, v3.16b, v12.4b[1]\n"
+        ".inst 0x4f8ce862  // sdot v2.4s, v3.16b, v12.4b[2]\n"
+        ".inst 0x4face87d  // sdot v29.4s, v3.16b, v12.4b[3]\n"
+        "ldr q12, [x21, #0x60]\n"
+        ".inst 0x4f89e2da  // sdot v26.4s, v22.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+        ".inst 0x4f89eac2  // sdot v2.4s, v22.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eadd  // sdot v29.4s, v22.16b, v9.4b[3]\n"
+        "ldr q17, [x21, #0x70]\n"
+        "add x21, x21, #0x88\n"
+        ".inst 0x4f8ce37a  // sdot v26.4s, v27.16b, v12.4b[0]\n"
+        ".inst 0x4face36a  // sdot v10.4s, v27.16b, v12.4b[1]\n"
+        ".inst 0x4f8ceb62  // sdot v2.4s, v27.16b, v12.4b[2]\n"
+        ".inst 0x4faceb7d  // sdot v29.4s, v27.16b, v12.4b[3]\n"
+        ".inst 0x4f91e3da  // sdot v26.4s, v30.16b, v17.4b[0]\n"
+        ".inst 0x4fb1e3ca  // sdot v10.4s, v30.16b, v17.4b[1]\n"
+        ".inst 0x4f91ebc2  // sdot v2.4s, v30.16b, v17.4b[2]\n"
+        ".inst 0x4fb1ebdd  // sdot v29.4s, v30.16b, v17.4b[3]\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "fmla v5.4s, v26.4s, v24.4s\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "fmla v21.4s, v10.4s, v31.4s\n"
+        "fmla v8.4s, v2.4s, v6.4s\n"
+        "fmla v1.4s, v29.4s, v20.4s\n"
+        "bgt 3b\n"
+        "mov x20, %x[res_ptr]\n"
+        "subs x27, x27, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "str q15, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q19, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q18, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q14, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q11, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q13, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q23, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q16, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q25, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q7, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q0, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q4, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q5, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q21, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q8, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q1, [x20, #0x0]\n"
+        "bne 2b\n"
+        "mov x20, #0x4\n"
+        "sub x10, x10, #0x10\n"
+        "cmp x10, #0x10\n"
+        "mov %x[res_ptr], x26\n"
+        "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+        "bge 1b\n"
+        "4:"  // Row loop skip
+        "cbz x10, 9f\n"
+        "5:"  // Row tail: Row loop
+        "add x24, %x[b_ptr], #0x8\n"
+        "mov x23, %x[nc]\n"
+        "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+        "6:"  // Row tail: Column loop
+        "movi v15.16b, #0x0\n"
+        "movi v19.16b, #0x0\n"
+        "add x25, %x[a_ptr], #0x8\n"
+        "mov x21, %x[nb]\n"
+        "movi v18.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "7:"  // Row tail: Block loop
+        "ldr q7, [x24, #0x0]\n"
+        "ldr q5, [x25, #0x0]\n"
+        "movi v9.16b, #0x4\n"
+        "movi v4.4s, #0x0\n"
+        "ldr q3, [x24, #0x10]\n"
+        "ldr q2, [x25, #0x10]\n"
+        "movi v1.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        "ldr q13, [x24, #0x20]\n"
+        "ldr q31, [x25, #0x20]\n"
+        "movi v30.4s, #0x0\n"
+        "movi v29.16b, #0xf0\n"
+        "ldr q28, [x24, #0x30]\n"
+        "ldr q27, [x25, #0x30]\n"
+        "sshl v20.16b, v7.16b, v9.16b\n"
+        "sub x20, x24, #0x8\n"
+        "ldr q26, [x25, #0x40]\n"
+        "ldr q25, [x25, #0x50]\n"
+        "sshl v17.16b, v3.16b, v9.16b\n"
+        "and v7.16b, v7.16b, v29.16b\n"
+        "ldr q24, [x25, #0x60]\n"
+        "ldr q16, [x25, #0x70]\n"
+        "sshl v22.16b, v13.16b, v9.16b\n"
+        "and v3.16b, v3.16b, v29.16b\n"
+        "ldr d21, [x20, #0x0]\n"
+        "ldr d12, [x25, #-0x8]\n"
+        ".inst 0x4f85e284  // sdot v4.4s, v20.16b, v5.4b[0]\n"
+        ".inst 0x4fa5e281  // sdot v1.4s, v20.16b, v5.4b[1]\n"
+        ".inst 0x4f85ea80  // sdot v0.4s, v20.16b, v5.4b[2]\n"
+        ".inst 0x4fa5ea9e  // sdot v30.4s, v20.16b, v5.4b[3]\n"
+        "sshl v9.16b, v28.16b, v9.16b\n"
+        "subs x21, x21, #0x1\n"
+        "and v13.16b, v13.16b, v29.16b\n"
+        "and v28.16b, v28.16b, v29.16b\n"
+        "add x25, x25, #0x88\n"
+        "add x24, x24, #0x48\n"
+        "fcvtl v21.4s, v21.4h\n"
+        "fcvtl v12.4s, v12.4h\n"
+        ".inst 0x4f82e224  // sdot v4.4s, v17.16b, v2.4b[0]\n"
+        ".inst 0x4fa2e221  // sdot v1.4s, v17.16b, v2.4b[1]\n"
+        ".inst 0x4f82ea20  // sdot v0.4s, v17.16b, v2.4b[2]\n"
+        ".inst 0x4fa2ea3e  // sdot v30.4s, v17.16b, v2.4b[3]\n"
+        "fmul v11.4s, v21.4s, v12.s[0]\n"
+        "fmul v23.4s, v21.4s, v12.s[1]\n"
+        "fmul v17.4s, v21.4s, v12.s[2]\n"
+        ".inst 0x4f9fe2c4  // sdot v4.4s, v22.16b, v31.4b[0]\n"
+        "fmul v6.4s, v21.4s, v12.s[3]\n"
+        ".inst 0x4fbfe2c1  // sdot v1.4s, v22.16b, v31.4b[1]\n"
+        ".inst 0x4f9feac0  // sdot v0.4s, v22.16b, v31.4b[2]\n"
+        ".inst 0x4fbfeade  // sdot v30.4s, v22.16b, v31.4b[3]\n"
+        ".inst 0x4f9be124  // sdot v4.4s, v9.16b, v27.4b[0]\n"
+        ".inst 0x4fbbe121  // sdot v1.4s, v9.16b, v27.4b[1]\n"
+        ".inst 0x4f9be920  // sdot v0.4s, v9.16b, v27.4b[2]\n"
+        ".inst 0x4fbbe93e  // sdot v30.4s, v9.16b, v27.4b[3]\n"
+        ".inst 0x4f9ae0e4  // sdot v4.4s, v7.16b, v26.4b[0]\n"
+        ".inst 0x4fbae0e1  // sdot v1.4s, v7.16b, v26.4b[1]\n"
+        ".inst 0x4f9ae8e0  // sdot v0.4s, v7.16b, v26.4b[2]\n"
+        ".inst 0x4fbae8fe  // sdot v30.4s, v7.16b, v26.4b[3]\n"
+        ".inst 0x4f99e064  // sdot v4.4s, v3.16b, v25.4b[0]\n"
+        ".inst 0x4fb9e061  // sdot v1.4s, v3.16b, v25.4b[1]\n"
+        ".inst 0x4f99e860  // sdot v0.4s, v3.16b, v25.4b[2]\n"
+        ".inst 0x4fb9e87e  // sdot v30.4s, v3.16b, v25.4b[3]\n"
+        ".inst 0x4f98e1a4  // sdot v4.4s, v13.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e1a1  // sdot v1.4s, v13.16b, v24.4b[1]\n"
+        ".inst 0x4f98e9a0  // sdot v0.4s, v13.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e9be  // sdot v30.4s, v13.16b, v24.4b[3]\n"
+        ".inst 0x4f90e384  // sdot v4.4s, v28.16b, v16.4b[0]\n"
+        ".inst 0x4fb0e381  // sdot v1.4s, v28.16b, v16.4b[1]\n"
+        ".inst 0x4f90eb80  // sdot v0.4s, v28.16b, v16.4b[2]\n"
+        ".inst 0x4fb0eb9e  // sdot v30.4s, v28.16b, v16.4b[3]\n"
+        "scvtf v4.4s, v4.4s, #0x4\n"
+        "scvtf v1.4s, v1.4s, #0x4\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "fmla v15.4s, v4.4s, v11.4s\n"
+        "scvtf v30.4s, v30.4s, #0x4\n"
+        "fmla v19.4s, v1.4s, v23.4s\n"
+        "fmla v18.4s, v0.4s, v17.4s\n"
+        "fmla v14.4s, v30.4s, v6.4s\n"
+        "bgt 7b\n"
+        "mov x20, %x[res_ptr]\n"
+        "cmp x10, #0x1\n"
+        "str q15, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x2\n"
+        "str q19, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x3\n"
+        "str q18, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "str q14, [x20, #0x0]\n"
+        "8:"  // Row tail: Accumulator store skip
+        "subs x23, x23, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "bne 6b\n"
+        "subs x10, x10, #0x4\n"
+        "add %x[a_ptr], %x[a_ptr], x9\n"
+        "mov %x[res_ptr], x22\n"
+        "bgt 5b\n"
+        "9:"  // Row tail: Row loop skip
+        : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+        : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+#else
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+#endif
+}
+
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_sve_cnt_b == QK8_0) {
+        GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+    size_t res_stride = bs * sizeof(float);
+
+    __asm__ __volatile__(
+        "mov x10, %x[nr]\n"
+        "mov x9, #0x88\n"
+        "cmp x10, #0x10\n"
+        "mul x9, %x[nb], x9\n"
+        "blt 4f\n"
+        "1:"  // Row loop
+        "add x28, %x[b_ptr], #0x8\n"
+        "mov x27, %x[nc]\n"
+        "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+        "2:"  // Column loop
+        "add x25, %x[a_ptr], #0x8\n"
+        "movi v2.16b, #0x0\n"
+        "movi v10.16b, #0x0\n"
+        "mov x24, %x[nb]\n"
+        "add x23, x25, x9\n"
+        "movi v12.16b, #0x0\n"
+        "movi v28.16b, #0x0\n"
+        "add x22, x23, x9\n"
+        "movi v11.16b, #0x0\n"
+        "movi v13.16b, #0x0\n"
+        "add x21, x22, x9\n"
+        "movi v22.16b, #0x0\n"
+        "movi v23.16b, #0x0\n"
+        "movi v25.16b, #0x0\n"
+        "movi v5.16b, #0x0\n"
+        "movi v7.16b, #0x0\n"
+        "movi v4.16b, #0x0\n"
+        "movi v6.16b, #0x0\n"
+        "movi v30.16b, #0x0\n"
+        "movi v24.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "3:"  // Block loop
+        "ldr q21, [x28, #0x0]\n"
+        "ldr q16, [x28, #0x10]\n"
+        "movi v1.16b, #0x4\n"
+        "movi v19.4s, #0x0\n"
+        "ldr q27, [x25, #0x0]\n"
+        "ldr q15, [x25, #0x10]\n"
+        "movi v26.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        "ldr q29, [x28, #0x20]\n"
+        "ldr q3, [x28, #0x30]\n"
+        "movi v17.4s, #0x0\n"
+        "movi v0.16b, #0xf0\n"
+        "ldr d20, [x25, #-0x8]\n"
+        "ldr d9, [x23, #-0x8]\n"
+        "sshl v8.16b, v21.16b, v1.16b\n"
+        "sshl v31.16b, v16.16b, v1.16b\n"
+        "and v21.16b, v21.16b, v0.16b\n"
+        "and v16.16b, v16.16b, v0.16b\n"
+        "sub x20, x28, #0x8\n"
+        "subs x24, x24, #0x1\n"
+        "add x28, x28, #0x48\n"
+        ".inst 0x4e88a773  // smmla v19.4s, v27.16b, v8.16b\n"
+        ".inst 0x4e9fa77a  // smmla v26.4s, v27.16b, v31.16b\n"
+        "ldr q27, [x25, #0x20]\n"
+        ".inst 0x4e88a5f2  // smmla v18.4s, v15.16b, v8.16b\n"
+        ".inst 0x4e9fa5f1  // smmla v17.4s, v15.16b, v31.16b\n"
+        "sshl v15.16b, v29.16b, v1.16b\n"
+        "sshl v1.16b, v3.16b, v1.16b\n"
+        "and v29.16b, v29.16b, v0.16b\n"
+        "and v3.16b, v3.16b, v0.16b\n"
+        "ldr q0, [x25, #0x30]\n"
+        "fcvtl v20.4s, v20.4h\n"
+        ".inst 0x4e8fa773  // smmla v19.4s, v27.16b, v15.16b\n"
+        "fcvtl v9.4s, v9.4h\n"
+        ".inst 0x4e81a77a  // smmla v26.4s, v27.16b, v1.16b\n"
+        "ldr q27, [x25, #0x40]\n"
+        ".inst 0x4e8fa412  // smmla v18.4s, v0.16b, v15.16b\n"
+        ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
+        "ldr q0, [x25, #0x50]\n"
+        ".inst 0x4e95a773  // smmla v19.4s, v27.16b, v21.16b\n"
+        ".inst 0x4e90a77a  // smmla v26.4s, v27.16b, v16.16b\n"
+        "ldr q27, [x25, #0x60]\n"
+        ".inst 0x4e95a412  // smmla v18.4s, v0.16b, v21.16b\n"
+        ".inst 0x4e90a411  // smmla v17.4s, v0.16b, v16.16b\n"
+        "ldr q0, [x25, #0x70]\n"
+        "add x25, x25, #0x88\n"
+        ".inst 0x4e9da773  // smmla v19.4s, v27.16b, v29.16b\n"
+        ".inst 0x4e83a77a  // smmla v26.4s, v27.16b, v3.16b\n"
+        "ldr d27, [x20, #0x0]\n"
+        ".inst 0x4e9da412  // smmla v18.4s, v0.16b, v29.16b\n"
+        ".inst 0x4e83a411  // smmla v17.4s, v0.16b, v3.16b\n"
+        "fcvtl v27.4s, v27.4h\n"
+        "uzp1 v0.2d, v19.2d, v26.2d\n"
+        "uzp2 v26.2d, v19.2d, v26.2d\n"
+        "fmul v19.4s, v27.4s, v20.s[0]\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "fmla v2.4s, v0.4s, v19.4s\n"
+        "ldr q19, [x23, #0x0]\n"
+        "uzp1 v0.2d, v18.2d, v17.2d\n"
+        "uzp2 v18.2d, v18.2d, v17.2d\n"
+        "fmul v17.4s, v27.4s, v20.s[1]\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "fmla v10.4s, v26.4s, v17.4s\n"
+        "ldr q17, [x23, #0x10]\n"
+        "fmul v26.4s, v27.4s, v20.s[2]\n"
+        "fmul v20.4s, v27.4s, v20.s[3]\n"
+        "fmla v12.4s, v0.4s, v26.4s\n"
+        "ldr d0, [x22, #-0x8]\n"
+        "ldr d26, [x21, #-0x8]\n"
+        "fcvtl v0.4s, v0.4h\n"
+        "fmla v28.4s, v18.4s, v20.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+        ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+        "ldr q19, [x23, #0x20]\n"
+        "fcvtl v26.4s, v26.4h\n"
+        ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+        ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+        "ldr q19, [x23, #0x40]\n"
+        ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+        ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+        "ldr q19, [x23, #0x60]\n"
+        ".inst 0x4e9da674  // smmla v20.4s, v19.16b, v29.16b\n"
+        ".inst 0x4e83a672  // smmla v18.4s, v19.16b, v3.16b\n"
+        "uzp1 v19.2d, v20.2d, v18.2d\n"
+        "scvtf v19.4s, v19.4s, #0x4\n"
+        "uzp2 v20.2d, v20.2d, v18.2d\n"
+        "fmul v18.4s, v27.4s, v9.s[0]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v11.4s, v19.4s, v18.4s\n"
+        "ldr q18, [x22, #0x0]\n"
+        "fmul v19.4s, v27.4s, v9.s[1]\n"
+        "fmla v13.4s, v20.4s, v19.4s\n"
+        "movi v19.4s, #0x0\n"
+        "movi v20.4s, #0x0\n"
+        ".inst 0x4e88a633  // smmla v19.4s, v17.16b, v8.16b\n"
+        ".inst 0x4e9fa634  // smmla v20.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x23, #0x30]\n"
+        ".inst 0x4e8fa633  // smmla v19.4s, v17.16b, v15.16b\n"
+        ".inst 0x4e81a634  // smmla v20.4s, v17.16b, v1.16b\n"
+        "ldr q17, [x23, #0x50]\n"
+        ".inst 0x4e95a633  // smmla v19.4s, v17.16b, v21.16b\n"
+        ".inst 0x4e90a634  // smmla v20.4s, v17.16b, v16.16b\n"
+        "ldr q17, [x23, #0x70]\n"
+        "add x23, x23, #0x88\n"
+        ".inst 0x4e9da633  // smmla v19.4s, v17.16b, v29.16b\n"
+        ".inst 0x4e83a634  // smmla v20.4s, v17.16b, v3.16b\n"
+        "uzp1 v17.2d, v19.2d, v20.2d\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "uzp2 v20.2d, v19.2d, v20.2d\n"
+        "fmul v19.4s, v27.4s, v9.s[2]\n"
+        "fmul v9.4s, v27.4s, v9.s[3]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v22.4s, v17.4s, v19.4s\n"
+        "ldr q17, [x22, #0x10]\n"
+        "movi v19.4s, #0x0\n"
+        ".inst 0x4e88a653  // smmla v19.4s, v18.16b, v8.16b\n"
+        "fmla v23.4s, v20.4s, v9.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v9.4s, #0x0\n"
+        ".inst 0x4e9fa654  // smmla v20.4s, v18.16b, v31.16b\n"
+        "ldr q18, [x22, #0x20]\n"
+        ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+        ".inst 0x4e8fa653  // smmla v19.4s, v18.16b, v15.16b\n"
+        ".inst 0x4e81a654  // smmla v20.4s, v18.16b, v1.16b\n"
+        "ldr q18, [x22, #0x40]\n"
+        ".inst 0x4e95a653  // smmla v19.4s, v18.16b, v21.16b\n"
+        ".inst 0x4e90a654  // smmla v20.4s, v18.16b, v16.16b\n"
+        "ldr q18, [x22, #0x60]\n"
+        ".inst 0x4e9da653  // smmla v19.4s, v18.16b, v29.16b\n"
+        ".inst 0x4e83a654  // smmla v20.4s, v18.16b, v3.16b\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e9fa632  // smmla v18.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x22, #0x30]\n"
+        ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+        ".inst 0x4e81a632  // smmla v18.4s, v17.16b, v1.16b\n"
+        "ldr q17, [x22, #0x50]\n"
+        ".inst 0x4e95a629  // smmla v9.4s, v17.16b, v21.16b\n"
+        ".inst 0x4e90a632  // smmla v18.4s, v17.16b, v16.16b\n"
+        "ldr q17, [x22, #0x70]\n"
+        "add x22, x22, #0x88\n"
+        ".inst 0x4e9da629  // smmla v9.4s, v17.16b, v29.16b\n"
+        ".inst 0x4e83a632  // smmla v18.4s, v17.16b, v3.16b\n"
+        "uzp1 v17.2d, v19.2d, v20.2d\n"
+        "uzp2 v20.2d, v19.2d, v20.2d\n"
+        "fmul v19.4s, v27.4s, v0.s[0]\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v25.4s, v17.4s, v19.4s\n"
+        "ldr q19, [x21, #0x0]\n"
+        "fmul v17.4s, v27.4s, v0.s[1]\n"
+        "fmla v5.4s, v20.4s, v17.4s\n"
+        "ldr q17, [x21, #0x10]\n"
+        "uzp1 v20.2d, v9.2d, v18.2d\n"
+        "uzp2 v9.2d, v9.2d, v18.2d\n"
+        "fmul v18.4s, v27.4s, v0.s[2]\n"
+        "fmul v0.4s, v27.4s, v0.s[3]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "fmla v7.4s, v20.4s, v18.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+        ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+        "ldr q19, [x21, #0x20]\n"
+        "fmla v4.4s, v9.4s, v0.4s\n"
+        "movi v9.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+        "fmul v8.4s, v27.4s, v26.s[0]\n"
+        ".inst 0x4e9fa620  // smmla v0.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x21, #0x30]\n"
+        ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+        "fmul v31.4s, v27.4s, v26.s[1]\n"
+        ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+        "ldr q19, [x21, #0x40]\n"
+        ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+        "fmul v15.4s, v27.4s, v26.s[2]\n"
+        "fmul v27.4s, v27.4s, v26.s[3]\n"
+        ".inst 0x4e81a620  // smmla v0.4s, v17.16b, v1.16b\n"
+        "ldr q1, [x21, #0x50]\n"
+        ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+        ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+        "ldr q26, [x21, #0x60]\n"
+        ".inst 0x4e95a429  // smmla v9.4s, v1.16b, v21.16b\n"
+        ".inst 0x4e90a420  // smmla v0.4s, v1.16b, v16.16b\n"
+        "ldr q21, [x21, #0x70]\n"
+        "add x21, x21, #0x88\n"
+        ".inst 0x4e9da754  // smmla v20.4s, v26.16b, v29.16b\n"
+        ".inst 0x4e83a752  // smmla v18.4s, v26.16b, v3.16b\n"
+        ".inst 0x4e9da6a9  // smmla v9.4s, v21.16b, v29.16b\n"
+        ".inst 0x4e83a6a0  // smmla v0.4s, v21.16b, v3.16b\n"
+        "uzp1 v29.2d, v20.2d, v18.2d\n"
+        "uzp2 v21.2d, v20.2d, v18.2d\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "uzp1 v18.2d, v9.2d, v0.2d\n"
+        "uzp2 v16.2d, v9.2d, v0.2d\n"
+        "scvtf v21.4s, v21.4s, #0x4\n"
+        "fmla v6.4s, v29.4s, v8.4s\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "scvtf v16.4s, v16.4s, #0x4\n"
+        "fmla v30.4s, v21.4s, v31.4s\n"
+        "fmla v24.4s, v18.4s, v15.4s\n"
+        "fmla v14.4s, v16.4s, v27.4s\n"
+        "bgt 3b\n"
+        "mov x20, %x[res_ptr]\n"
+        "subs x27, x27, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "str q2, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q10, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q12, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q28, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q11, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q13, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q22, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q23, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q25, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q5, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q7, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q4, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q6, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q30, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q24, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q14, [x20, #0x0]\n"
+        "bne 2b\n"
+        "mov x20, #0x4\n"
+        "sub x10, x10, #0x10\n"
+        "cmp x10, #0x10\n"
+        "mov %x[res_ptr], x26\n"
+        "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+        "bge 1b\n"
+        "4:"  // Row loop skip
+        "cbz x10, 9f\n"
+        "5:"  // Row tail: Row loop
+        "add x24, %x[b_ptr], #0x8\n"
+        "mov x23, %x[nc]\n"
+        "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+        "6:"  // Row tail: Column loop
+        "movi v2.16b, #0x0\n"
+        "movi v10.16b, #0x0\n"
+        "add x25, %x[a_ptr], #0x8\n"
+        "mov x21, %x[nb]\n"
+        "movi v12.16b, #0x0\n"
+        "movi v28.16b, #0x0\n"
+        "7:"  // Row tail: Block loop
+        "ldr q6, [x24, #0x0]\n"
+        "ldr q5, [x24, #0x10]\n"
+        "movi v17.16b, #0x4\n"
+        "movi v8.4s, #0x0\n"
+        "ldr q4, [x25, #0x0]\n"
+        "ldr q13, [x25, #0x10]\n"
+        "movi v27.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        "ldr q31, [x24, #0x20]\n"
+        "ldr q14, [x24, #0x30]\n"
+        "movi v29.4s, #0x0\n"
+        "movi v22.16b, #0xf0\n"
+        "ldr q11, [x25, #0x20]\n"
+        "ldr q23, [x25, #0x30]\n"
+        "sshl v21.16b, v6.16b, v17.16b\n"
+        "sshl v16.16b, v5.16b, v17.16b\n"
+        "ldr q20, [x25, #0x40]\n"
+        "ldr q26, [x25, #0x50]\n"
+        "and v6.16b, v6.16b, v22.16b\n"
+        "and v5.16b, v5.16b, v22.16b\n"
+        "ldr q25, [x25, #0x60]\n"
+        "ldr q3, [x25, #0x70]\n"
+        "sshl v19.16b, v31.16b, v17.16b\n"
+        "sshl v18.16b, v14.16b, v17.16b\n"
+        "ldr d17, [x25, #-0x8]\n"
+        ".inst 0x4e95a488  // smmla v8.4s, v4.16b, v21.16b\n"
+        ".inst 0x4e90a49b  // smmla v27.4s, v4.16b, v16.16b\n"
+        "and v31.16b, v31.16b, v22.16b\n"
+        ".inst 0x4e95a5a0  // smmla v0.4s, v13.16b, v21.16b\n"
+        ".inst 0x4e90a5bd  // smmla v29.4s, v13.16b, v16.16b\n"
+        "and v14.16b, v14.16b, v22.16b\n"
+        "sub x20, x24, #0x8\n"
+        "ldr d16, [x20, #0x0]\n"
+        "subs x21, x21, #0x1\n"
+        "add x25, x25, #0x88\n"
+        "fcvtl v17.4s, v17.4h\n"
+        "add x24, x24, #0x48\n"
+        ".inst 0x4e93a568  // smmla v8.4s, v11.16b, v19.16b\n"
+        ".inst 0x4e92a57b  // smmla v27.4s, v11.16b, v18.16b\n"
+        ".inst 0x4e93a6e0  // smmla v0.4s, v23.16b, v19.16b\n"
+        ".inst 0x4e92a6fd  // smmla v29.4s, v23.16b, v18.16b\n"
+        "fcvtl v16.4s, v16.4h\n"
+        ".inst 0x4e86a688  // smmla v8.4s, v20.16b, v6.16b\n"
+        ".inst 0x4e85a69b  // smmla v27.4s, v20.16b, v5.16b\n"
+        "fmul v23.4s, v16.4s, v17.s[0]\n"
+        "fmul v21.4s, v16.4s, v17.s[1]\n"
+        "fmul v1.4s, v16.4s, v17.s[2]\n"
+        "fmul v20.4s, v16.4s, v17.s[3]\n"
+        ".inst 0x4e86a740  // smmla v0.4s, v26.16b, v6.16b\n"
+        ".inst 0x4e85a75d  // smmla v29.4s, v26.16b, v5.16b\n"
+        ".inst 0x4e9fa728  // smmla v8.4s, v25.16b, v31.16b\n"
+        ".inst 0x4e8ea73b  // smmla v27.4s, v25.16b, v14.16b\n"
+        ".inst 0x4e9fa460  // smmla v0.4s, v3.16b, v31.16b\n"
+        ".inst 0x4e8ea47d  // smmla v29.4s, v3.16b, v14.16b\n"
+        "uzp1 v19.2d, v8.2d, v27.2d\n"
+        "uzp2 v18.2d, v8.2d, v27.2d\n"
+        "scvtf v19.4s, v19.4s, #0x4\n"
+        "uzp1 v17.2d, v0.2d, v29.2d\n"
+        "uzp2 v16.2d, v0.2d, v29.2d\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "fmla v2.4s, v19.4s, v23.4s\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "scvtf v16.4s, v16.4s, #0x4\n"
+        "fmla v10.4s, v18.4s, v21.4s\n"
+        "fmla v12.4s, v17.4s, v1.4s\n"
+        "fmla v28.4s, v16.4s, v20.4s\n"
+        "bgt 7b\n"
+        "mov x20, %x[res_ptr]\n"
+        "cmp x10, #0x1\n"
+        "str q2, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x2\n"
+        "str q10, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x3\n"
+        "str q12, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "str q28, [x20, #0x0]\n"
+        "8:"  // Row tail: Accumulator store skip
+        "subs x23, x23, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "bne 6b\n"
+        "subs x10, x10, #0x4\n"
+        "add %x[a_ptr], %x[a_ptr], x9\n"
+        "mov %x[res_ptr], x22\n"
+        "bgt 5b\n"
+        "9:"  // Row tail: Row loop skip
+        : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+        : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+#endif
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
+    if (ggml_sve_cnt_b == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x20, #0x4\n"
+            "mov x13, %x[nr]\n"
+            "mov z28.s, #-0x4\n"
+            "mov x12, #0x88\n"
+            "ptrue p1.b\n"
+            "whilelt p0.s, XZR, x20\n"
+            "cmp x13, #0x10\n"
+            "mul x12, %x[nb], x12\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x11, %x[b_ptr], #0x10\n"
+            "mov x10, %x[nc]\n"
+            "add x9, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "mov x27, %x[nb]\n"
+            "add x26, x28, x12\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "add x25, x26, x12\n"
+            "mov z13.b, #0x0\n"
+            "mov z1.b, #0x0\n"
+            "add x24, x25, x12\n"
+            "mov z20.b, #0x0\n"
+            "mov z25.b, #0x0\n"
+            "mov z11.b, #0x0\n"
+            "mov z16.b, #0x0\n"
+            "mov z19.b, #0x0\n"
+            "mov z26.b, #0x0\n"
+            "mov z8.b, #0x0\n"
+            "mov z29.b, #0x0\n"
+            "mov z27.b, #0x0\n"
+            "mov z10.b, #0x0\n"
+            "3:"  // Block loop
+            "ld1b { z30.b }, p1/Z, [x11]\n"
+            "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n"
+            "mov z18.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            "ld1rqb { z3.b }, p1/Z, [x28]\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #16]\n"
+            "mov z9.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n"
+            "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n"
+            "sub x20, x11, #0x10\n"
+            "sub x23, x28, #0x8\n"
+            "lsl z31.b, z30.b, #0x4\n"
+            "lsl z6.b, z21.b, #0x4\n"
+            "ld1h { z23.s }, p1/Z, [x20]\n"
+            "sub x22, x26, #0x8\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z21.b, z21.b, #0xf0\n"
+            "sub x21, x25, #0x8\n"
+            "sub x20, x24, #0x8\n"
+            "lsl z14.b, z4.b, #0x4\n"
+            "lsl z2.b, z17.b, #0x4\n"
+            "subs x27, x27, #0x1\n"
+            "add x11, x11, #0x90\n"
+            ".inst 0x451f9872  // smmla z18.s, z3.b, z31.b\n"
+            ".inst 0x45069867  // smmla z7.s, z3.b, z6.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #32]\n"
+            "and z4.b, z4.b, #0xf0\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #48]\n"
+            "and z17.b, z17.b, #0xf0\n"
+            "fcvt z23.s, p1/m, z23.h\n"
+            ".inst 0x450e9872  // smmla z18.s, z3.b, z14.b\n"
+            ".inst 0x45029867  // smmla z7.s, z3.b, z2.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #64]\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #80]\n"
+            "fscale z23.s, p1/m, z23.s, z28.s\n"
+            ".inst 0x451e9872  // smmla z18.s, z3.b, z30.b\n"
+            ".inst 0x45159867  // smmla z7.s, z3.b, z21.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #96]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #112]\n"
+            "add x28, x28, #0x88\n"
+            ".inst 0x45049872  // smmla z18.s, z3.b, z4.b\n"
+            ".inst 0x45119867  // smmla z7.s, z3.b, z17.b\n"
+            "ld1h { z3.s }, p0/Z, [x23]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            "uzp1 z5.d, z18.d, z7.d\n"
+            "uzp2 z18.d, z18.d, z7.d\n"
+            "mov z3.q, z3.q[0]\n"
+            "uzp1 z7.d, z9.d, z22.d\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z3.s[0]\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z24.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z5.b }, p1/Z, [x26]\n"
+            "fmul z9.s, z23.s, z3.s[1]\n"
+            "fmla z15.s, p1/M, z18.s, z9.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #16]\n"
+            "fmul z9.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "fmla z12.s, p1/M, z7.s, z9.s\n"
+            "mov z9.s, #0x0\n"
+            "ld1h { z7.s }, p0/Z, [x22]\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            "fmla z0.s, p1/M, z22.s, z3.s\n"
+            "mov z22.s, #0x0\n"
+            "ld1h { z3.s }, p0/Z, [x21]\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #32]\n"
+            "fcvt z7.s, p1/m, z7.h\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #64]\n"
+            "mov z7.q, z7.q[0]\n"
+            "mov z3.q, z3.q[0]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #96]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "uzp1 z5.d, z9.d, z22.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z7.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z13.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x25]\n"
+            "fmul z5.s, z23.s, z7.s[1]\n"
+            "fmla z1.s, p1/M, z22.s, z5.s\n"
+            "mov z5.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            ".inst 0x451f9a45  // smmla z5.s, z18.b, z31.b\n"
+            ".inst 0x45069a56  // smmla z22.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #48]\n"
+            ".inst 0x450e9a45  // smmla z5.s, z18.b, z14.b\n"
+            ".inst 0x45029a56  // smmla z22.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #80]\n"
+            ".inst 0x451e9a45  // smmla z5.s, z18.b, z30.b\n"
+            ".inst 0x45159a56  // smmla z22.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #112]\n"
+            "add x26, x26, #0x88\n"
+            ".inst 0x45049a45  // smmla z5.s, z18.b, z4.b\n"
+            ".inst 0x45119a56  // smmla z22.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z5.d, z22.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z22.d, z5.d, z22.d\n"
+            "fmul z5.s, z23.s, z7.s[2]\n"
+            "fmul z7.s, z23.s, z7.s[3]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z20.s, p1/M, z18.s, z5.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #16]\n"
+            "ld1h { z5.s }, p0/Z, [x20]\n"
+            "fcvt z5.s, p1/m, z5.h\n"
+            "fmla z25.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9936  // smmla z22.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #32]\n"
+            "mov z5.q, z5.q[0]\n"
+            ".inst 0x450e9936  // smmla z22.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #64]\n"
+            ".inst 0x451e9936  // smmla z22.s, z9.b, z30.b\n"
+            ".inst 0x45159927  // smmla z7.s, z9.b, z21.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #96]\n"
+            ".inst 0x45049936  // smmla z22.s, z9.b, z4.b\n"
+            ".inst 0x45119927  // smmla z7.s, z9.b, z17.b\n"
+            "uzp1 z9.d, z22.d, z7.d\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "uzp2 z22.d, z22.d, z7.d\n"
+            "fmul z7.s, z23.s, z3.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z11.s, p1/M, z9.s, z7.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x24]\n"
+            "fmul z7.s, z23.s, z3.s[1]\n"
+            "fmla z16.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9a56  // smmla z22.s, z18.b, z31.b\n"
+            ".inst 0x45069a47  // smmla z7.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #48]\n"
+            ".inst 0x450e9a56  // smmla z22.s, z18.b, z14.b\n"
+            ".inst 0x45029a47  // smmla z7.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #80]\n"
+            ".inst 0x451e9a56  // smmla z22.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #112]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x45049a56  // smmla z22.s, z18.b, z4.b\n"
+            ".inst 0x45119a47  // smmla z7.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z22.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z7.d, z22.d, z7.d\n"
+            "fmul z22.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "fmla z19.s, p1/M, z18.s, z22.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #16]\n"
+            "fmul z22.s, z23.s, z5.s[0]\n"
+            "fmla z26.s, p1/M, z7.s, z3.s\n"
+            "mov z3.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9923  // smmla z3.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x24, #32]\n"
+            ".inst 0x450e9923  // smmla z3.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "mov z9.s, #0x0\n"
+            ".inst 0x451f9a49  // smmla z9.s, z18.b, z31.b\n"
+            "mov z31.s, #0x0\n"
+            ".inst 0x45069a5f  // smmla z31.s, z18.b, z6.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #48]\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #64]\n"
+            ".inst 0x450e98c9  // smmla z9.s, z6.b, z14.b\n"
+            "fmul z14.s, z23.s, z5.s[1]\n"
+            ".inst 0x450298df  // smmla z31.s, z6.b, z2.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #80]\n"
+            "fmul z2.s, z23.s, z5.s[2]\n"
+            "fmul z23.s, z23.s, z5.s[3]\n"
+            ".inst 0x451e9a43  // smmla z3.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x24, #96]\n"
+            ".inst 0x451e98c9  // smmla z9.s, z6.b, z30.b\n"
+            ".inst 0x451598df  // smmla z31.s, z6.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #112]\n"
+            "add x24, x24, #0x88\n"
+            ".inst 0x450498a3  // smmla z3.s, z5.b, z4.b\n"
+            ".inst 0x451198a7  // smmla z7.s, z5.b, z17.b\n"
+            ".inst 0x45049a49  // smmla z9.s, z18.b, z4.b\n"
+            ".inst 0x45119a5f  // smmla z31.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z3.d, z7.d\n"
+            "uzp2 z5.d, z3.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp1 z6.d, z9.d, z31.d\n"
+            "uzp2 z9.d, z9.d, z31.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "fmla z8.s, p1/M, z18.s, z22.s\n"
+            "scvtf z6.s, p1/m, z6.s\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "fmla z29.s, p1/M, z5.s, z14.s\n"
+            "fmla z27.s, p1/M, z6.s, z2.s\n"
+            "fmla z10.s, p1/M, z9.s, z23.s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x10, x10, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z13.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z1.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z20.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z25.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z11.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z16.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z19.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z26.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z8.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z29.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z27.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z10.s }, p1, [x20]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x13, x13, #0x10\n"
+            "cmp x13, #0x10\n"
+            "mov %x[res_ptr], x9\n"
+            "madd %x[a_ptr], x20, x12, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x13, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x25, %x[b_ptr], #0x10\n"
+            "mov x24, %x[nc]\n"
+            "add x23, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov x22, %x[nb]\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ld1b { z3.b }, p1/Z, [x25]\n"
+            "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n"
+            "mov z2.s, #0x0\n"
+            "mov z25.s, #0x0\n"
+            "ld1rqb { z26.b }, p1/Z, [x28]\n"
+            "ld1rqb { z21.b }, p1/Z, [x28, #16]\n"
+            "mov z27.s, #0x0\n"
+            "mov z19.s, #0x0\n"
+            "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n"
+            "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n"
+            "sub x21, x25, #0x10\n"
+            "sub x20, x28, #0x8\n"
+            "lsl z20.b, z3.b, #0x4\n"
+            "lsl z4.b, z6.b, #0x4\n"
+            "ld1rqb { z10.b }, p1/Z, [x28, #32]\n"
+            "ld1rqb { z23.b }, p1/Z, [x28, #48]\n"
+            "and z3.b, z3.b, #0xf0\n"
+            "and z6.b, z6.b, #0xf0\n"
+            "ld1rqb { z11.b }, p1/Z, [x28, #64]\n"
+            "ld1rqb { z7.b }, p1/Z, [x28, #80]\n"
+            "lsl z8.b, z29.b, #0x4\n"
+            "lsl z14.b, z16.b, #0x4\n"
+            "ld1rqb { z18.b }, p1/Z, [x28, #96]\n"
+            "ld1rqb { z30.b }, p1/Z, [x28, #112]\n"
+            ".inst 0x45149b42  // smmla z2.s, z26.b, z20.b\n"
+            ".inst 0x45049b59  // smmla z25.s, z26.b, z4.b\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1h { z17.s }, p1/Z, [x21]\n"
+            ".inst 0x45149abb  // smmla z27.s, z21.b, z20.b\n"
+            ".inst 0x45049ab3  // smmla z19.s, z21.b, z4.b\n"
+            "and z16.b, z16.b, #0xf0\n"
+            "ld1h { z4.s }, p0/Z, [x20]\n"
+            "subs x22, x22, #0x1\n"
+            "add x28, x28, #0x88\n"
+            "fcvt z17.s, p1/m, z17.h\n"
+            "add x25, x25, #0x90\n"
+            ".inst 0x45089942  // smmla z2.s, z10.b, z8.b\n"
+            ".inst 0x450e9959  // smmla z25.s, z10.b, z14.b\n"
+            "fcvt z4.s, p1/m, z4.h\n"
+            ".inst 0x45089afb  // smmla z27.s, z23.b, z8.b\n"
+            ".inst 0x450e9af3  // smmla z19.s, z23.b, z14.b\n"
+            "fscale z17.s, p1/m, z17.s, z28.s\n"
+            "mov z4.q, z4.q[0]\n"
+            ".inst 0x45039962  // smmla z2.s, z11.b, z3.b\n"
+            ".inst 0x45069979  // smmla z25.s, z11.b, z6.b\n"
+            "fmul z23.s, z17.s, z4.s[0]\n"
+            "fmul z9.s, z17.s, z4.s[1]\n"
+            "fmul z21.s, z17.s, z4.s[2]\n"
+            "fmul z4.s, z17.s, z4.s[3]\n"
+            ".inst 0x450398fb  // smmla z27.s, z7.b, z3.b\n"
+            ".inst 0x450698f3  // smmla z19.s, z7.b, z6.b\n"
+            ".inst 0x451d9a42  // smmla z2.s, z18.b, z29.b\n"
+            ".inst 0x45109a59  // smmla z25.s, z18.b, z16.b\n"
+            ".inst 0x451d9bdb  // smmla z27.s, z30.b, z29.b\n"
+            ".inst 0x45109bd3  // smmla z19.s, z30.b, z16.b\n"
+            "uzp1 z31.d, z2.d, z25.d\n"
+            "uzp2 z13.d, z2.d, z25.d\n"
+            "scvtf z31.s, p1/m, z31.s\n"
+            "uzp1 z17.d, z27.d, z19.d\n"
+            "uzp2 z18.d, z27.d, z19.d\n"
+            "scvtf z13.s, p1/m, z13.s\n"
+            "fmla z24.s, p1/M, z31.s, z23.s\n"
+            "scvtf z17.s, p1/m, z17.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "fmla z15.s, p1/M, z13.s, z9.s\n"
+            "fmla z12.s, p1/M, z17.s, z21.s\n"
+            "fmla z0.s, p1/M, z18.s, z4.s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x13, #0x1\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x2\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x3\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x24, x24, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "bne 6b\n"
+            "subs x13, x13, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x12\n"
+            "mov %x[res_ptr], x23\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+    else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+        GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
+                    "performance");
+    }
+    else if (ggml_cpu_has_neon()) {
+        GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
+                    "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
+                    "quantization format for optimal performance");
+    }
+#endif
+#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    GGML_ASSERT(ggml_cpu_has_sve() &&
+                "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance");
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
+                "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
+                "performance");
+#else
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+#endif
+}
diff --git a/llm/llama.cpp/ggml/src/ggml-aarch64.h b/llm/llama.cpp/ggml/src/ggml-aarch64.h
new file mode 100644
index 0000000000000000000000000000000000000000..517babaf1691bbd9cc8434c5d81b853d3f7d6105
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-aarch64.h
@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+// GEMV
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// GEMM
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#ifdef __cplusplus
+}
+#endif
+
diff --git a/llm/llama.cpp/ggml/src/ggml-alloc.c b/llm/llama.cpp/ggml/src/ggml-alloc.c
new file mode 100644
index 0000000000000000000000000000000000000000..e485326abc45d218b56e79e66c6c7718eb85c078
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-alloc.c
@@ -0,0 +1,1036 @@
+#include "ggml-alloc.h"
+#include "ggml-backend-impl.h"
+#include "ggml.h"
+#include "ggml-impl.h"
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define MAX_FREE_BLOCKS 256
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
+#define AT_PRINTF(...)
+
+
+static bool ggml_is_view(const struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SOFT_MAX:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+// tallocr
+
+struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) {
+    void * base = ggml_backend_buffer_get_base(buffer);
+    size_t align = ggml_backend_buffer_get_alignment(buffer);
+
+    assert(align && !(align & (align - 1))); // power of 2
+
+    struct ggml_tallocr talloc = (struct ggml_tallocr) {
+        /*.buffer    = */ buffer,
+        /*.base      = */ base,
+        /*.alignment = */ align,
+        /*.offset    = */ aligned_offset(base, 0, align),
+    };
+    return talloc;
+}
+
+void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
+    size = GGML_PAD(size, talloc->alignment);
+
+    if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
+        fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
+                __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
+        GGML_ABORT("not enough space in the buffer");
+    }
+
+    void * addr = (char *)ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
+    talloc->offset += size;
+
+    assert(((uintptr_t)addr % talloc->alignment) == 0);
+
+    ggml_backend_tensor_alloc(talloc->buffer, tensor, addr);
+}
+
+// dynamic tensor allocator
+
+struct free_block {
+    size_t offset;
+    size_t size;
+};
+
+struct ggml_dyn_tallocr {
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    size_t max_size;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct {
+        const struct ggml_tensor * tensor;
+        size_t offset;
+    } allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i].tensor == NULL) {
+            alloc->allocated_tensors[i].tensor = tensor;
+            alloc->allocated_tensors[i].offset = offset;
+            return;
+        }
+    }
+    GGML_ABORT("out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i].offset == offset) {
+            alloc->allocated_tensors[i].tensor = NULL;
+            return;
+        }
+    }
+    GGML_ABORT("tried to free tensor %s not found\n", tensor->name);
+}
+#endif
+
+static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block besides the last block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    if (best_fit_block == -1) {
+        // the last block is our last resort
+        struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size) {
+            best_fit_block = alloc->n_free_blocks - 1;
+        } else {
+            // this should never happen
+            fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
+                    __func__, size, max_avail);
+            GGML_ABORT("not enough space in the buffer");
+        }
+    }
+
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    size_t offset = block->offset;
+    block->offset = offset + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    AT_PRINTF("block %d, offset %zu\n", best_fit_block, offset);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, offset, tensor);
+    size_t cur_max = offset + size;
+    if (cur_max > alloc->max_size) {
+        // sort allocated_tensors by offset
+        for (int i = 0; i < 1024; i++) {
+            for (int j = i + 1; j < 1024; j++) {
+                if (alloc->allocated_tensors[i].offset > alloc->allocated_tensors[j].offset) {
+                    const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
+                    size_t tmp_offset = alloc->allocated_tensors[i].offset;
+                    alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
+                    alloc->allocated_tensors[i].offset = alloc->allocated_tensors[j].offset;
+                    alloc->allocated_tensors[j].tensor = tmp_tensor;
+                    alloc->allocated_tensors[j].offset = tmp_offset;
+                }
+            }
+        }
+        fprintf(stderr, "max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i].tensor) {
+                fprintf(stderr, "%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
+                    alloc->allocated_tensors[i].offset,
+                    alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
+                    ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
+            }
+        }
+        fprintf(stderr, "\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, offset + size);
+
+    return offset;
+
+    GGML_UNUSED(tensor);
+}
+
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, size_t size, const struct ggml_tensor * tensor) {
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: freeing %s at %zu (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, offset, size, alloc->n_free_blocks);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, offset, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if (block->offset + block->size == offset) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && block->offset + block->size == alloc->free_blocks[i+1].offset) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if (offset + size == block->offset) {
+            block->offset = offset;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && alloc->free_blocks[i-1].offset + alloc->free_blocks[i-1].size == block->offset) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].offset < offset) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].offset = offset;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+
+    GGML_UNUSED(tensor);
+}
+
+static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
+    alloc->n_free_blocks = 1;
+    alloc->free_blocks[0].offset = 0;
+    alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
+    alloc->max_size = 0;
+}
+
+static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
+    struct ggml_dyn_tallocr * alloc = (struct ggml_dyn_tallocr *)malloc(sizeof(struct ggml_dyn_tallocr));
+
+    *alloc = (struct ggml_dyn_tallocr) {
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.max_size      = */ 0,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ {{0}},
+#endif
+    };
+
+    ggml_dyn_tallocr_reset(alloc);
+
+    return alloc;
+}
+
+static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
+    free(alloc);
+}
+
+static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
+    return alloc->max_size;
+}
+
+
+/////////////////////////////////////
+
+// graph allocator
+
+struct hash_node {
+    int n_children;
+    int n_views;
+    int buffer_id;
+    size_t offset; // offset within the buffer
+    bool allocated;
+};
+
+struct tensor_alloc {
+    int buffer_id;
+    size_t offset;
+    size_t size_max; // 0 = pre-allocated, unused, or view
+};
+
+struct leaf_alloc {
+    int buffer_id;
+    struct tensor_alloc leaf;
+};
+
+struct node_alloc {
+    struct tensor_alloc dst;
+    struct tensor_alloc src[GGML_MAX_SRC];
+};
+
+struct ggml_gallocr {
+    ggml_backend_buffer_type_t * bufts; // [n_buffers]
+    ggml_backend_buffer_t * buffers; // [n_buffers]
+    struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
+    int n_buffers;
+
+    struct ggml_hash_set hash_set;
+    struct hash_node * hash_values; // [hash_set.size]
+
+    struct node_alloc * node_allocs; // [n_nodes]
+    int n_nodes;
+
+    struct leaf_alloc * leaf_allocs; // [n_leafs]
+    int n_leafs;
+};
+
+ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
+    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
+    GGML_ASSERT(galloc != NULL);
+
+    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
+    GGML_ASSERT(galloc->bufts != NULL);
+
+    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
+    GGML_ASSERT(galloc->buffers != NULL);
+
+    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
+    GGML_ASSERT(galloc->buf_tallocs != NULL);
+
+    for (int i = 0; i < n_bufs; i++) {
+        galloc->bufts[i] = bufts[i];
+        galloc->buffers[i] = NULL;
+
+        // check if the same buffer type is used multiple times and reuse the same allocator
+        for (int j = 0; j < i; j++) {
+            if (bufts[i] == bufts[j]) {
+                galloc->buf_tallocs[i] = galloc->buf_tallocs[j];
+                break;
+            }
+        }
+
+        if (galloc->buf_tallocs[i] == NULL) {
+            size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
+            galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
+        }
+    }
+    galloc->n_buffers = n_bufs;
+
+    return galloc;
+}
+
+ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft) {
+    return ggml_gallocr_new_n(&buft, 1);
+}
+
+void ggml_gallocr_free(ggml_gallocr_t galloc) {
+    if (galloc == NULL) {
+        return;
+    }
+
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers != NULL) {
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buffers[j] == galloc->buffers[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                ggml_backend_buffer_free(galloc->buffers[i]);
+            }
+        }
+        if (galloc->buf_tallocs != NULL) {
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                ggml_dyn_tallocr_free(galloc->buf_tallocs[i]);
+            }
+        }
+    }
+
+    ggml_hash_set_free(&galloc->hash_set);
+    free(galloc->hash_values);
+    free(galloc->bufts);
+    free(galloc->buffers);
+    free(galloc->buf_tallocs);
+    free(galloc->node_allocs);
+    free(galloc->leaf_allocs);
+    free(galloc);
+}
+
+typedef struct ggml_gallocr * ggml_gallocr_t;
+
+static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    size_t i = ggml_hash_find_or_insert(&galloc->hash_set, t);
+    return &galloc->hash_values[i];
+}
+
+static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    return ggml_gallocr_hash_get(galloc, t)->allocated;
+}
+
+static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    hn->buffer_id = buffer_id;
+    hn->offset = offset;
+    hn->allocated = true;
+}
+
+static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
+}
+
+static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+
+    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
+        hn->allocated = true;
+        assert(hn->offset == 0);
+
+        // try to reuse a parent's buffer (inplace)
+        if (ggml_op_can_inplace(node->op)) {
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                struct ggml_tensor * parent = node->src[i];
+                if (parent == NULL) {
+                    continue;
+                }
+
+                // if the node's data is external, then we cannot re-use it
+                if (!ggml_gallocr_is_own(galloc, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                    continue;
+                }
+
+                // outputs cannot be reused
+                if (parent->flags & GGML_TENSOR_FLAG_OUTPUT || (parent->view_src != NULL && parent->view_src->flags & GGML_TENSOR_FLAG_OUTPUT)) {
+                    AT_PRINTF("not reusing parent %s for %s as it is an output\n", parent->name, node->name);
+                    continue;
+                }
+
+                if (!ggml_are_same_layout(node, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as layouts are different\n", parent->name, node->name);
+                    continue;
+                }
+
+                struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+                if (p_hn->n_children == 1 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = parent->view_src;
+                        struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
+                        if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                            AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                            assert(view_src_hn->offset == p_hn->offset);
+                            hn->buffer_id = p_hn->buffer_id;
+                            hn->offset = p_hn->offset;
+                            p_hn->allocated = false; // avoid freeing the parent
+                            view_src_hn->allocated = false;
+                            return;
+                        }
+                    } else {
+                        AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                        hn->buffer_id = p_hn->buffer_id;
+                        hn->offset = p_hn->offset;
+                        p_hn->allocated = false; // avoid freeing the parent
+                        return;
+                    }
+                }
+            }
+        }
+        // allocate tensor from the buffer
+        struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+        ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+        size_t size = ggml_backend_buft_get_alloc_size(buft, node);
+        size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
+        hn->buffer_id = buffer_id;
+        hn->offset = offset;
+        return;
+    }
+}
+
+static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
+    // graph outputs are never freed
+    if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
+        AT_PRINTF("not freeing output %s\n", node->name);
+        return;
+    }
+
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    size_t offset = hn->offset;
+    int buffer_id = hn->buffer_id;
+    struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+    ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+    size_t size = ggml_backend_buft_get_alloc_size(buft, node);
+    ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
+    hn->allocated = false;
+}
+
+static int get_node_buffer_id(const int * node_buffer_ids, int i) {
+    return node_buffer_ids ? node_buffer_ids[i] : 0;
+}
+
+static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    // clear hash tables
+    ggml_hash_set_reset(&galloc->hash_set);
+    memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
+
+    // allocate leafs
+    // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
+    }
+
+    // count number of children and views
+    // allocate other graph inputs and leafs first to avoid overwriting them
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
+            struct ggml_tensor * view_src = node->view_src;
+            ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
+        }
+
+        if (node->flags & GGML_TENSOR_FLAG_INPUT) {
+            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
+        }
+
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+
+            ggml_gallocr_hash_get(galloc, src)->n_children += 1;
+
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
+                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int buffer_id = get_node_buffer_id(node_buffer_ids, i);
+
+        // allocate parents (only leafs need to be allocated at this point)
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            ggml_gallocr_allocate_node(galloc, parent, buffer_id);
+        }
+
+        // allocate node
+        ggml_gallocr_allocate_node(galloc, node, buffer_id);
+
+        AT_PRINTF("exec: %s (%s) <= ", ggml_op_desc(node), node->name);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            AT_PRINTF("%s", parent->name);
+            if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                AT_PRINTF(", ");
+            }
+        }
+        AT_PRINTF("\n");
+
+        // update parents
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+            p_hn->n_children -= 1;
+
+            AT_PRINTF("parent %s: %d children, %d views, allocated: %d\n",
+                parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);
+
+            if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                if (ggml_is_view(parent)) {
+                    struct ggml_tensor * view_src = parent->view_src;
+                    struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
+                    view_src_hn->n_views -= 1;
+                    AT_PRINTF("view_src %s: %d children, %d views\n",
+                        view_src->name, view_src_hn->n_children, view_src_hn->n_views);
+                    if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src_hn->allocated) {
+                        ggml_gallocr_free_node(galloc, view_src);
+                    }
+                }
+                else if (p_hn->allocated) {
+                    ggml_gallocr_free_node(galloc, parent);
+                }
+            }
+            AT_PRINTF("\n");
+        }
+    }
+}
+
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    size_t min_hash_size = graph->n_nodes + graph->n_leafs;
+    // add 25% margin to avoid hash collisions
+    min_hash_size += min_hash_size / 4;
+
+    // initialize hash table
+    if (galloc->hash_set.size < min_hash_size) {
+        ggml_hash_set_free(&galloc->hash_set);
+        galloc->hash_set = ggml_hash_set_new(min_hash_size);
+        GGML_ASSERT(galloc->hash_set.keys != NULL);
+
+        free(galloc->hash_values);
+        galloc->hash_values = malloc(sizeof(struct hash_node) * galloc->hash_set.size);
+        GGML_ASSERT(galloc->hash_values != NULL);
+    }
+
+    // reset allocators
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        ggml_dyn_tallocr_reset(galloc->buf_tallocs[i]);
+    }
+
+    // allocate in hash table
+    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);
+
+    // set the node_allocs from the hash table
+    if (galloc->n_nodes < graph->n_nodes) {
+        free(galloc->node_allocs);
+        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
+        GGML_ASSERT(galloc->node_allocs != NULL);
+    }
+    galloc->n_nodes = graph->n_nodes;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        if (node->view_src || node->data) {
+            node_alloc->dst.buffer_id = -1;
+            node_alloc->dst.offset = SIZE_MAX;
+            node_alloc->dst.size_max = 0;
+        } else {
+            struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+            node_alloc->dst.buffer_id = hn->buffer_id;
+            node_alloc->dst.offset    = hn->offset;
+            node_alloc->dst.size_max  = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (!src || src->view_src || src->data) {
+                node_alloc->src[j].buffer_id = -1;
+                node_alloc->src[j].offset = SIZE_MAX;
+                node_alloc->src[j].size_max = 0;
+            } else {
+                struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
+                node_alloc->src[j].buffer_id = hn->buffer_id;
+                node_alloc->src[j].offset   = hn->offset;
+                node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
+            }
+        }
+    }
+    if (galloc->n_leafs < graph->n_leafs) {
+        free(galloc->leaf_allocs);
+        galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
+        GGML_ASSERT(galloc->leaf_allocs != NULL);
+    }
+    galloc->n_leafs = graph->n_leafs;
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
+        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
+        if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.buffer_id = -1;
+            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.size_max = 0;
+        } else {
+            galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id;
+            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        }
+    }
+
+    // reallocate buffers if needed
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        // if the buffer type is used multiple times, we reuse the same buffer
+        for (int j = 0; j < i; j++) {
+            if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                galloc->buffers[i] = galloc->buffers[j];
+                break;
+            }
+        }
+
+        size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
+        size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
+
+        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
+        if (new_size > cur_size || galloc->buffers[i] == NULL) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+#endif
+
+            ggml_backend_buffer_free(galloc->buffers[i]);
+            galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
+            if (galloc->buffers[i] == NULL) {
+                fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                return false;
+            }
+            ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
+        }
+    }
+
+    return true;
+}
+
+bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
+    return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
+}
+
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, struct tensor_alloc * tensor_alloc) {
+    int buffer_id = tensor_alloc->buffer_id;
+    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+
+    if (tensor->view_src != NULL) {
+        if (tensor->buffer == NULL) {
+            assert(tensor_alloc->offset == SIZE_MAX);
+            if (tensor->view_src->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+            ggml_backend_view_init(tensor);
+        }
+    } else {
+        if (tensor->data == NULL) {
+            assert(tensor_alloc->offset != SIZE_MAX);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
+            void * addr = (char *)base + tensor_alloc->offset;
+            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
+        } else {
+            if (tensor->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+        }
+    }
+}
+
+static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
+    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    return talloc->size_max >= node_size;
+}
+
+static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
+    if (galloc->n_nodes != graph->n_nodes) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: graph has different number of nodes\n", __func__);
+#endif
+        return true;
+    }
+
+    if (galloc->n_leafs != graph->n_leafs) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
+#endif
+        return true;
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+
+        if (!ggml_gallocr_node_needs_realloc(galloc, node, &node_alloc->dst)) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
+#endif
+            return true;
+        }
+
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            if (!ggml_gallocr_node_needs_realloc(galloc, src, &node_alloc->src[j])) {
+#ifndef NDEBUG
+                fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
+#endif
+                return true;
+            }
+        }
+    }
+
+    return false;
+}
+
+bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
+    if (ggml_gallocr_needs_realloc(galloc, graph)) {
+        if (galloc->n_buffers == 1) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: reallocating buffers automatically\n", __func__);
+#endif
+            if (!ggml_gallocr_reserve(galloc, graph)) {
+                return false;
+            }
+        } else {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
+#endif
+            return false;
+        }
+    }
+
+    // reset buffers
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers[i] != NULL) {
+            ggml_backend_buffer_reset(galloc->buffers[i]);
+        }
+    }
+
+    // allocate the graph tensors from the previous assignments
+    // leafs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
+        ggml_gallocr_init_tensor(galloc, leaf, &leaf_alloc->leaf);
+    }
+    // nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            ggml_gallocr_init_tensor(galloc, src, &node_alloc->src[j]);
+        }
+        ggml_gallocr_init_tensor(galloc, node, &node_alloc->dst);
+    }
+
+    return true;
+}
+
+size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
+
+    if (galloc->buffers[buffer_id] == NULL) {
+        return 0;
+    }
+
+    for (int i = 0; i < buffer_id; i++) {
+        if (galloc->buffers[i] == galloc->buffers[buffer_id]) {
+            // this buffer is the same as a previous one due to the same buffer type being used multiple times
+            // only return the buffer size the first time it appears to avoid double counting
+            return 0;
+        }
+    }
+
+    return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
+}
+
+// utils
+
+static bool alloc_tensor_range(struct ggml_context * ctx,
+        struct ggml_tensor * first, struct ggml_tensor * last,
+        ggml_backend_buffer_type_t buft, size_t size,
+        ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
+    if (buffer == NULL) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
+#endif
+        for (size_t i = 0; i < *n_buffers; i++) {
+            ggml_backend_buffer_free((*buffers)[i]);
+        }
+        free(*buffers);
+        return false;
+    }
+
+    struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);
+
+    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                ggml_tallocr_alloc(&tallocr, t);
+            } else if (t->buffer == NULL) {
+                ggml_backend_view_init(t);
+            }
+        } else {
+            if (t->view_src != NULL && t->buffer == NULL) {
+                // view of a pre-allocated tensor
+                ggml_backend_view_init(t);
+            }
+        }
+    }
+
+    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
+    (*buffers)[(*n_buffers)++] = buffer;
+
+    return true;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    size_t max_size = ggml_backend_buft_get_max_size(buft);
+
+    ggml_backend_buffer_t * buffers = NULL;
+    size_t n_buffers = 0;
+
+    size_t cur_buf_size = 0;
+    struct ggml_tensor * first = ggml_get_first_tensor(ctx);
+    for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        size_t this_size = 0;
+        if (t->data == NULL && t->view_src == NULL) {
+            this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+
+        if (this_size > max_size) {
+            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
+            // allocate tensors in the current buffer
+            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+                return NULL;
+            }
+            first = t;
+            cur_buf_size = this_size;
+        } else {
+            cur_buf_size += this_size;
+        }
+    }
+
+    // allocate remaining tensors
+    if (cur_buf_size > 0) {
+        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+            return NULL;
+        }
+    }
+
+    if (n_buffers == 0) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer;
+    if (n_buffers == 1) {
+        buffer = buffers[0];
+    } else {
+        buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
+    }
+    free(buffers);
+    return buffer;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
+    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
+}
diff --git a/llm/llama.cpp/ggml/src/ggml-backend-impl.h b/llm/llama.cpp/ggml/src/ggml-backend-impl.h
new file mode 100644
index 0000000000000000000000000000000000000000..36ca370867c9e7bcccbd58a86febd5bc82704086
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-backend-impl.h
@@ -0,0 +1,153 @@
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    // buffer type
+    typedef void * ggml_backend_buffer_type_context_t;
+
+    struct ggml_backend_buffer_type_i {
+        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
+        // allocate a buffer of this type
+        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        // tensor alignment
+        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft);
+        // max buffer size that can be allocated
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft);
+        // data size needed to allocate the tensor, including padding
+        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
+        // check if tensor data is in host memory
+        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
+    typedef void * ggml_backend_buffer_context_t;
+
+    struct ggml_backend_buffer_i {
+        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
+        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
+        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
+        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
+        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+    };
+
+    struct ggml_backend_buffer {
+        struct ggml_backend_buffer_i  iface;
+        ggml_backend_buffer_type_t    buft;
+        ggml_backend_buffer_context_t context;
+        size_t size;
+        enum ggml_backend_buffer_usage usage;
+    };
+
+    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
+                   ggml_backend_buffer_type_t      buft,
+            struct ggml_backend_buffer_i           iface,
+                   ggml_backend_buffer_context_t   context,
+                   size_t                          size);
+
+    // do not use directly, use ggml_backend_tensor_copy instead
+    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // buffer that contains a collection of buffers
+    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+
+    //
+    // Backend
+    //
+
+    typedef void * ggml_backend_context_t;
+
+    struct ggml_backend_i {
+        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
+
+        void (*GGML_CALL free)(ggml_backend_t backend);
+
+        // buffer allocation
+        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
+
+        // (optional) asynchronous tensor data access
+        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        // (optional) complete all pending operations
+        void (*GGML_CALL synchronize)(ggml_backend_t backend);
+
+        // compute graph with a plan (not used currently)
+        // create a new plan for a graph
+        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
+        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
+        void                      (*GGML_CALL graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
+        // compute the graph with the plan
+        enum ggml_status          (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan (async)
+        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+        // check if the backend can compute an operation
+        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // check if the backend can use tensors allocated in a buffer type
+        bool (*GGML_CALL supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
+
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        // create a new event that can record events on this backend instance
+        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
+        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
+        // record an event on the backend instance that created it
+        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
+        // wait for an event on on a different backend instance
+        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
+        // block until an event is recorded
+        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
+    };
+
+    struct ggml_backend {
+        ggml_guid_t guid;
+
+        struct ggml_backend_i iface;
+        ggml_backend_context_t context;
+    };
+
+    struct ggml_backend_event {
+        ggml_backend_t backend;
+        void * context;
+    };
+
+    //
+    // Backend registry
+    //
+
+    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
+
+    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/llm/llama.cpp/ggml/src/ggml-backend.c b/llm/llama.cpp/ggml/src/ggml-backend.c
new file mode 100644
index 0000000000000000000000000000000000000000..954ab20725acc9bcdd5aad3e7ebae9e186a271fc
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-backend.c
@@ -0,0 +1,2240 @@
+#include "ggml-backend-impl.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
+
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// backend buffer type
+
+const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name(buft);
+}
+
+GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    return buft->iface.alloc_buffer(buft, size);
+}
+
+size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_alignment(buft);
+}
+
+size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
+    // get_max_size is optional, defaults to SIZE_MAX
+    if (buft->iface.get_max_size) {
+        return buft->iface.get_max_size(buft);
+    }
+    return SIZE_MAX;
+}
+
+GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+    // get_alloc_size is optional, defaults to ggml_nbytes
+    if (buft->iface.get_alloc_size) {
+        size_t size = buft->iface.get_alloc_size(buft, tensor);
+        assert(size >= ggml_nbytes(tensor));
+        return size;
+    }
+    return ggml_nbytes(tensor);
+}
+
+bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
+    if (buft->iface.is_host) {
+        return buft->iface.is_host(buft);
+    }
+    return false;
+}
+
+// backend buffer
+
+GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
+               ggml_backend_buffer_type_t      buft,
+        struct ggml_backend_buffer_i           iface,
+               ggml_backend_buffer_context_t   context,
+               size_t                          size) {
+    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
+
+    (*buffer) = (struct ggml_backend_buffer) {
+        /* .interface = */ iface,
+        /* .buft      = */ buft,
+        /* .context   = */ context,
+        /* .size      = */ size,
+        /* .usage     = */ GGML_BACKEND_BUFFER_USAGE_ANY
+    };
+
+    return buffer;
+}
+
+const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name(buffer);
+}
+
+void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+    if (buffer == NULL) {
+        return;
+    }
+
+    if (buffer->iface.free_buffer != NULL) {
+        buffer->iface.free_buffer(buffer);
+    }
+    free(buffer);
+}
+
+size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    return buffer->size;
+}
+
+void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    void * base = buffer->iface.get_base(buffer);
+
+    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
+
+    return base;
+}
+
+GGML_CALL void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // init_tensor is optional
+    if (buffer->iface.init_tensor) {
+        buffer->iface.init_tensor(buffer, tensor);
+    }
+}
+
+size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
+}
+
+size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
+}
+
+void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    buffer->iface.clear(buffer, value);
+}
+
+bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
+}
+
+void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    buffer->usage = usage;
+
+    // FIXME: add a generic callback to the buffer interface
+    if (ggml_backend_buffer_is_multi_buffer(buffer)) {
+        ggml_backend_multi_buffer_set_usage(buffer, usage);
+    }
+}
+
+enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage(ggml_backend_buffer_t buffer) {
+    return buffer->usage;
+}
+
+ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
+    return buffer->buft;
+}
+
+void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
+    if (buffer->iface.reset) {
+        buffer->iface.reset(buffer);
+    }
+}
+
+bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
+    if (dst_buf->iface.cpy_tensor) {
+        return dst_buf->iface.cpy_tensor(dst_buf, src, dst);
+    }
+    return false;
+}
+
+// backend
+
+ggml_guid_t ggml_backend_guid(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return NULL;
+    }
+    return backend->guid;
+}
+
+const char * ggml_backend_name(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return "NULL";
+    }
+    return backend->iface.get_name(backend);
+}
+
+void ggml_backend_free(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return;
+    }
+
+    backend->iface.free(backend);
+}
+
+ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
+    return backend->iface.get_default_buffer_type(backend);
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
+    return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size);
+}
+
+size_t ggml_backend_get_alignment(ggml_backend_t backend) {
+    return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
+}
+
+size_t ggml_backend_get_max_size(ggml_backend_t backend) {
+    return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
+}
+
+void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (backend->iface.set_tensor_async == NULL) {
+        ggml_backend_tensor_set(tensor, data, offset, size);
+    } else {
+        backend->iface.set_tensor_async(backend, tensor, data, offset, size);
+    }
+}
+
+void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    if (backend->iface.get_tensor_async == NULL) {
+        ggml_backend_tensor_get(tensor, data, offset, size);
+    } else {
+        backend->iface.get_tensor_async(backend, tensor, data, offset, size);
+    }
+}
+
+GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    buf->iface.set_tensor(buf, tensor, data, offset, size);
+}
+
+GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    buf->iface.get_tensor(buf, tensor, data, offset, size);
+}
+
+void ggml_backend_synchronize(ggml_backend_t backend) {
+    if (backend->iface.synchronize == NULL) {
+        return;
+    }
+
+    backend->iface.synchronize(backend);
+}
+
+ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend->iface.graph_plan_create != NULL);
+
+    return backend->iface.graph_plan_create(backend, cgraph);
+}
+
+void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_free != NULL);
+
+    backend->iface.graph_plan_free(backend, plan);
+}
+
+enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
+
+    return backend->iface.graph_plan_compute(backend, plan);
+}
+
+enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
+    ggml_backend_synchronize(backend);
+    return err;
+}
+
+enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_compute(backend, cgraph);
+}
+
+bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return backend->iface.supports_op(backend, op);
+}
+
+bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return backend->iface.supports_buft(backend, buft);
+}
+
+bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    if (backend->iface.offload_op != NULL) {
+        return backend->iface.offload_op(backend, op);
+    }
+    return false;
+}
+
+// backend copy
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    if (src == dst) {
+        return;
+    }
+
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+    } else if (ggml_backend_buffer_is_host(dst->buffer)) {
+        ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+    } else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
+#endif
+        size_t nbytes = ggml_nbytes(src);
+        void * data = malloc(nbytes);
+        ggml_backend_tensor_get(src, data, 0, nbytes);
+        ggml_backend_tensor_set(dst, data, 0, nbytes);
+        free(data);
+    }
+}
+
+void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    if (src == dst) {
+        return;
+    }
+
+    if (backend_dst->iface.cpy_tensor_async != NULL) {
+        if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
+            return;
+        }
+    }
+
+    // an async copy would normally happen after all the queued operations on both backends are completed
+    // sync src, set_async dst
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        ggml_backend_synchronize(backend_src);
+        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
+    } else {
+        ggml_backend_synchronize(backend_src);
+        ggml_backend_tensor_copy(src, dst);
+        ggml_backend_synchronize(backend_dst);
+    }
+}
+
+// events
+
+ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) {
+    if (backend->iface.event_new == NULL) {
+        return NULL;
+    }
+    return backend->iface.event_new(backend);
+}
+
+void ggml_backend_event_free(ggml_backend_event_t event) {
+    if (event == NULL) {
+        return;
+    }
+    event->backend->iface.event_free(event);
+}
+
+void ggml_backend_event_record(ggml_backend_event_t event) {
+    GGML_ASSERT(event->backend->iface.event_record != NULL);
+
+    event->backend->iface.event_record(event);
+}
+
+void ggml_backend_event_synchronize(ggml_backend_event_t event) {
+    GGML_ASSERT(event->backend->iface.event_synchronize != NULL);
+
+    event->backend->iface.event_synchronize(event);
+}
+
+void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    GGML_ASSERT(backend->iface.event_wait != NULL);
+
+    backend->iface.event_wait(backend, event);
+}
+
+// backend registry
+
+#define GGML_REG_MAX_BACKENDS 64
+
+struct ggml_backend_reg {
+    char name[128];
+    ggml_backend_init_fn init_fn;
+    ggml_backend_buffer_type_t default_buffer_type;
+    void * user_data;
+};
+
+static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS];
+static size_t ggml_backend_registry_count = 0;
+
+GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
+
+GGML_CALL static void ggml_backend_registry_init(void) {
+    static bool initialized = false;
+
+    if (initialized) {
+        return;
+    }
+
+    initialized = true;
+
+    ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);
+
+    // add forward decls here to avoid including the backend headers
+#ifdef GGML_USE_CUDA
+    extern GGML_CALL void ggml_backend_cuda_reg_devices(void);
+    ggml_backend_cuda_reg_devices();
+#endif
+
+#ifdef GGML_USE_SYCL
+    extern void ggml_backend_sycl_reg_devices(void);
+    ggml_backend_sycl_reg_devices();
+#endif
+
+#ifdef GGML_USE_METAL
+    extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
+    extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+    ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
+#endif
+
+#ifdef GGML_USE_VULKAN
+    extern GGML_CALL int ggml_backend_vk_reg_devices(void);
+    ggml_backend_vk_reg_devices();
+#endif
+
+#ifdef GGML_USE_KOMPUTE
+    extern GGML_CALL void ggml_backend_kompute_reg_devices(void);
+    ggml_backend_kompute_reg_devices();
+#endif
+
+#ifdef GGML_USE_CANN
+    extern GGML_CALL int ggml_backend_cann_reg_devices(void);
+    ggml_backend_cann_reg_devices();
+#endif
+}
+
+GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
+    GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS);
+
+    size_t id = ggml_backend_registry_count;
+
+    ggml_backend_registry[id] = (struct ggml_backend_reg) {
+        /* .name                = */ {0},
+        /* .fn                  = */ init_fn,
+        /* .default_buffer_type = */ default_buffer_type,
+        /* .user_data           = */ user_data,
+    };
+
+    snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name);
+
+#ifndef NDEBUG
+    fprintf(stderr, "%s: registered backend %s\n", __func__, name);
+#endif
+
+    ggml_backend_registry_count++;
+}
+
+size_t ggml_backend_reg_get_count(void) {
+    ggml_backend_registry_init();
+
+    return ggml_backend_registry_count;
+}
+
+size_t ggml_backend_reg_find_by_name(const char * name) {
+    ggml_backend_registry_init();
+
+    for (size_t i = 0; i < ggml_backend_registry_count; i++) {
+        // TODO: case insensitive in a portable way
+        if (strcmp(ggml_backend_registry[i].name, name) == 0) {
+            return i;
+        }
+    }
+
+    // not found
+    return SIZE_MAX;
+}
+
+// init from backend:params string
+ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) {
+    ggml_backend_registry_init();
+
+    const char * params = strchr(backend_str, ':');
+    char backend_name[128];
+    if (params == NULL) {
+        snprintf(backend_name, sizeof(backend_name), "%s", backend_str);
+        params = "";
+    } else {
+        snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str);
+        params++;
+    }
+
+    size_t backend_i = ggml_backend_reg_find_by_name(backend_name);
+
+    if (backend_i == SIZE_MAX) {
+        fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name);
+        return NULL;
+    }
+
+    return ggml_backend_reg_init_backend(backend_i, params);
+}
+
+const char * ggml_backend_reg_get_name(size_t i) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_registry[i].name;
+}
+
+ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data);
+}
+
+ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_registry[i].default_buffer_type;
+}
+
+ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
+    ggml_backend_registry_init();
+
+    GGML_ASSERT(i < ggml_backend_registry_count);
+    return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size);
+}
+
+// backend CPU
+
+static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
+
+GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
+    return "CPU";
+
+    GGML_UNUSED(buffer);
+}
+
+GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    uintptr_t data = (uintptr_t)buffer->context;
+
+    // align the buffer
+    if (data % TENSOR_ALIGNMENT != 0) {
+        data = GGML_PAD(data, TENSOR_ALIGNMENT);
+    }
+
+    return (void *)data;
+}
+
+GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+GGML_CALL static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+GGML_CALL static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        memcpy(dst->data, src->data, ggml_nbytes(src));
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_name,
+    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// for buffers from ptr, free is not called
+static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_name,
+    /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+GGML_CALL static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU";
+
+    GGML_UNUSED(buft);
+}
+
+GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
+}
+
+GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
+        /* .iface = */ {
+            /* .get_name         = */ ggml_backend_cpu_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type;
+}
+
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include 
+
+GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buf);
+}
+
+GGML_CALL static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    //void * ptr = hbw_malloc(size);
+    void * ptr;
+    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name;
+    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
+struct ggml_backend_cpu_context {
+    int n_threads;
+    void * work_data;
+    size_t work_size;
+
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
+};
+
+GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
+    return "CPU";
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    free(cpu_ctx->work_data);
+    free(cpu_ctx);
+    free(backend);
+}
+
+GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
+        if (cpu_plan->cplan.work_data == NULL) {
+            free(cpu_plan);
+            return NULL;
+        }
+    }
+
+    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
+    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return cpu_plan;
+}
+
+GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    free(cpu_plan->cplan.work_data);
+    free(cpu_plan);
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        free(cpu_ctx->work_data);
+        cpu_ctx->work_data = malloc(cplan.work_size);
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+        cpu_ctx->work_size = cplan.work_size;
+    }
+    cplan.work_data = cpu_ctx->work_data;
+
+    cplan.abort_callback      = cpu_ctx->abort_callback;
+    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return ggml_graph_compute(cgraph, &cplan);
+}
+
+GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_CPY:
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+        case GGML_OP_MUL_MAT:
+            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
+        default:
+            return true;
+    }
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static bool ggml_backend_cpu_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i cpu_backend_i = {
+    /* .get_name                = */ ggml_backend_cpu_name,
+    /* .free                    = */ ggml_backend_cpu_free,
+    /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op             = */ ggml_backend_cpu_supports_op,
+    /* .supports_buft           = */ ggml_backend_cpu_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_cpu_guid(void) {
+    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
+    if (ctx == NULL) {
+        return NULL;
+    }
+
+    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->work_data           = NULL;
+    ctx->work_size           = 0;
+    ctx->abort_callback      = NULL;
+    ctx->abort_callback_data = NULL;
+
+    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
+    if (cpu_backend == NULL) {
+        free(ctx);
+        return NULL;
+    }
+
+    *cpu_backend = (struct ggml_backend) {
+        /* .guid      = */ ggml_backend_cpu_guid(),
+        /* .interface = */ cpu_backend_i,
+        /* .context   = */ ctx
+    };
+    return cpu_backend;
+}
+
+GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
+GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
+    GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
+    return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
+}
+
+GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
+    return ggml_backend_cpu_init();
+
+    GGML_UNUSED(params);
+    GGML_UNUSED(user_data);
+}
+
+// multi-buffer buffer
+
+struct ggml_backend_multi_buffer_context {
+    ggml_backend_buffer_t * buffers;
+    size_t n_buffers;
+};
+
+typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t;
+
+GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+
+    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
+}
+
+GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_free(ctx->buffers[i]);
+    }
+
+    free(ctx->buffers);
+    free(ctx);
+}
+
+GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_clear(ctx->buffers[i], value);
+    }
+}
+
+static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) {
+    static struct ggml_backend_buffer_i multi_backend_buffer_i = {
+        /* .get_name        = */ ggml_backend_multi_buffer_get_name,
+        /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
+        /* .get_base        = */ NULL,
+        /* .init_tensor     = */ NULL,
+        /* .set_tensor      = */ NULL,
+        /* .get_tensor      = */ NULL,
+        /* .cpy_tensor      = */ NULL,
+        /* .clear           = */ ggml_backend_multi_buffer_clear,
+        /* .reset           = */ NULL,
+    };
+
+    return multi_backend_buffer_i;
+}
+
+GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context));
+    ctx->n_buffers = n_buffers;
+    ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
+
+    GGML_ASSERT(ctx->buffers != NULL);
+
+    size_t total_size = 0;
+    for (size_t i = 0; i < n_buffers; i++) {
+        ctx->buffers[i] = buffers[i];
+        total_size += ggml_backend_buffer_get_size(buffers[i]);
+    }
+
+    return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size);
+}
+
+GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
+}
+
+GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
+    }
+}
+
+// creates a copy of the tensor with the same memory layout
+static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
+    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        dup->nb[i] = tensor->nb[i];
+    }
+    return dup;
+}
+
+static bool ggml_is_view_op(enum ggml_op op) {
+    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
+}
+
+// scheduler
+
+#ifndef GGML_SCHED_MAX_BACKENDS
+#define GGML_SCHED_MAX_BACKENDS 16
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLITS
+#define GGML_SCHED_MAX_SPLITS 2048
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLIT_INPUTS
+#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
+#endif
+
+#ifndef GGML_SCHED_MAX_COPIES
+#define GGML_SCHED_MAX_COPIES 4
+#endif
+
+struct ggml_backend_sched_split {
+    int backend_id;
+    int i_start;
+    int i_end;
+    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_inputs;
+    // graph view of this split
+    struct ggml_cgraph graph;
+};
+
+struct ggml_backend_sched {
+    bool is_reset; // true if the scheduler has been reset since the last graph split
+    bool is_alloc;
+
+    int n_backends;
+
+    ggml_backend_t backends[GGML_SCHED_MAX_BACKENDS];
+    ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
+    ggml_gallocr_t galloc;
+
+    // hash map of the nodes in the graph
+    struct ggml_hash_set  hash_set;
+    int                 * hv_tensor_backend_ids; // [hash_set.size]
+    struct ggml_tensor ** hv_tensor_copies;      // [hash_set.size][n_backends][n_copies]
+
+    int * node_backend_ids; // [graph_size]
+    int * leaf_backend_ids; // [graph_size]
+
+    int * prev_node_backend_ids; // [graph_size]
+    int * prev_leaf_backend_ids; // [graph_size]
+
+    // copy of the graph with modified inputs
+    struct ggml_cgraph graph;
+
+    // graph splits
+    struct ggml_backend_sched_split * splits;
+    int n_splits;
+    int splits_capacity;
+
+    // pipeline parallelism support
+    int n_copies;
+    int cur_copy;
+    ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
+    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_graph_inputs;
+
+    struct ggml_context * ctx;
+
+    ggml_backend_sched_eval_callback callback_eval;
+    void * callback_eval_user_data;
+
+    char * context_buffer;
+    size_t context_buffer_size;
+
+    bool debug;
+};
+
+#define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
+#define tensor_backend_id(tensor) sched->hv_tensor_backend_ids[hash_id(tensor)]
+#define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
+#define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
+
+// returns the priority of the backend, lower id is higher priority
+static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->backends[i] == backend) {
+            return i;
+        }
+    }
+    return -1;
+}
+
+static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
+    ggml_backend_buffer_t buffer = tensor->buffer;
+    if (buffer == NULL) {
+        return -1;
+    }
+
+    // find highest prio backend that supports the buffer type and the op
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (ggml_backend_supports_buft(sched->backends[i], buffer->buft) &&
+            ggml_backend_supports_op(sched->backends[i], op)) {
+            return i;
+        }
+    }
+
+#ifndef NDEBUG
+    fprintf(stderr, "%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
+        __func__, ggml_op_desc(tensor), ggml_backend_buffer_name(buffer), tensor->name);
+#endif
+
+    return -1;
+}
+
+#if 0
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
+#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
+#define GET_CAUSE(node) causes[hash_id(node)]
+#else
+#define SET_CAUSE(node, ...)
+#define GET_CAUSE(node) ""
+#endif
+
+// returns the backend that should be used for the node based on the current locations
+static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
+    // TODO: use supports_op to check if the backend supports the op
+
+    // assign pre-allocated nodes to their backend
+    int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
+    if (cur_backend_id != -1) {
+        SET_CAUSE(tensor, "1.dst");
+        return cur_backend_id;
+    }
+
+    // view_src
+    if (tensor->view_src != NULL) {
+        cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src, tensor);
+        if (cur_backend_id != -1) {
+            SET_CAUSE(tensor, "1.vsrc");
+            return cur_backend_id;
+        }
+    }
+
+    // graph input
+    if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
+        cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU)
+        SET_CAUSE(tensor, "1.inp");
+        return cur_backend_id;
+    }
+
+    // operations with weights are preferably run on the same backend as the weights
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        const struct ggml_tensor * src = tensor->src[i];
+        if (src == NULL) {
+            continue;
+        }
+        if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
+            // check if a backend with higher prio wants to offload the op
+            if (src_backend_id == sched->n_backends - 1) {
+                for (int b = 0; b < src_backend_id; b++) {
+                    if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
+                        SET_CAUSE(tensor, "1.off");
+                        return b;
+                    }
+                }
+            }
+            SET_CAUSE(tensor, "1.wgt%d", i);
+            return src_backend_id;
+        }
+    }
+
+    return -1;
+}
+
+static char * fmt_size(size_t size) {
+    static char buffer[128];
+    if (size >= 1024*1024) {
+        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
+    } else {
+        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
+    }
+    return buffer;
+}
+
+static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    int cur_split = 0;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
+            ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
+            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
+                sched->splits[cur_split].n_inputs);
+            for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
+                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
+                    fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+            }
+            fprintf(stderr, "\n");
+            cur_split++;
+        }
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
+        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
+            fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
+            fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
+                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+static bool ggml_backend_sched_buffer_supported(ggml_backend_sched_t sched, struct ggml_tensor * t, int backend_id) {
+    ggml_backend_buffer_t buf = t->view_src ? t->view_src->buffer : t->buffer;
+    ggml_backend_buffer_type_t buft = NULL;
+
+    if (buf) {
+        // the tensor is already allocated
+        buft = buf->buft;
+    } else {
+        // see if the tensor already has a backend assigned, and use the buffer type of that backend
+        int tensor_backend_id = tensor_backend_id(t);
+        if (tensor_backend_id == -1 && t->view_src) {
+            tensor_backend_id = tensor_backend_id(t->view_src);
+        }
+        if (tensor_backend_id != -1) {
+            buft = sched->bufts[tensor_backend_id];
+        }
+    }
+
+    return buft != NULL && ggml_backend_supports_buft(sched->backends[backend_id], buft);
+}
+
+static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, struct ggml_tensor * node, int cur_backend_id, int * node_backend_id) {
+    if (ggml_backend_supports_op(sched->backends[cur_backend_id], node)) {
+        *node_backend_id = cur_backend_id;
+        SET_CAUSE(node, "2.sup");
+    }
+}
+
+// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
+static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    // reset splits
+    sched->n_splits = 0;
+    sched->n_graph_inputs = 0;
+    sched->is_reset = false;
+
+    struct ggml_init_params params = {
+        /* .mem_size =   */ sched->context_buffer_size,
+        /* .mem_buffer = */ sched->context_buffer,
+        /* .no_alloc =   */ true
+    };
+
+    ggml_free(sched->ctx);
+
+    sched->ctx = ggml_init(params);
+    if (sched->ctx == NULL) {
+        GGML_ABORT("%s: failed to initialize context\n", __func__);
+    }
+
+    // pass 1: assign backends to ops with pre-allocated inputs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        int * leaf_backend_id = &tensor_backend_id(leaf);
+        // do not overwrite user assignments
+        if (*leaf_backend_id == -1) {
+            *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
+        }
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int * node_backend_id = &tensor_backend_id(node);
+        // do not overwrite user assignments
+        if (*node_backend_id == -1) {
+            *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
+
+#if 0
+            // src
+            if (node->op == GGML_OP_NONE) {
+                continue;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+                int * src_backend_id = &tensor_backend_id(src);
+                if (*src_backend_id == -1) {
+                    *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+                }
+            }
+#endif
+        }
+    }
+
+    // pass 2: expand current backend assignments
+    // assign the same backend to adjacent nodes
+    // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
+    // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops
+    // ops unsupported by the backend being expanded will be left unassigned so that they can be assigned later when the locations of its inputs are known
+    // expand gpu down
+    {
+        int cur_backend_id = -1;
+        for (int i = 0; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
+                    // skip cpu (lowest prio backend)
+                    cur_backend_id = -1;
+                } else {
+                    cur_backend_id = *node_backend_id;
+                }
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand gpu up
+    {
+        int cur_backend_id = -1;
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
+                    // skip cpu (lowest prio backend)
+                    cur_backend_id = -1;
+                } else {
+                    cur_backend_id = *node_backend_id;
+                }
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand rest down
+    {
+        int cur_backend_id = -1;
+        for (int i = 0; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand rest up
+    {
+        int cur_backend_id = -1;
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+
+    // pass 3: upgrade nodes to higher prio backends with compatible buffer types
+    // if the tensor is already in the same buffer type (*) as another higher priority backend, we should move it there
+    // however, we also need to verify that the sources are in compatible buffer types
+    // (*) the actual requirement is more relaxed, the buffer type of the backend should be supported by all the users of this tensor further down the graph
+    // however, this is slow to verify, so we have a more strict requirement that the buffer type is the same
+    // this is not uncommon since multiple backends can use host memory, with the same buffer type (eg. BLAS and CPU)
+    // additionally, set remaining unassigned nodes to the backend with the most supported inputs
+    // only nodes that could not be assigned during expansion due to the backend not supporting the op should be unassigned at this point
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        int * node_backend_id = &tensor_backend_id(node);
+        if (*node_backend_id == -1) {
+            // unassigned node: find the backend with the most supported inputs
+            int n_supported_best = -1;
+            for (int b = 0; b < sched->n_backends; b++) {
+                if (ggml_backend_supports_op(sched->backends[b], node)) {
+                    int n_supported = 0;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if ((tensor_backend_id(src) != -1 || tensor_backend_id(src->view_src) != -1) && ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            n_supported++;
+                        }
+                    }
+                    if (n_supported > n_supported_best) {
+                        n_supported_best = n_supported;
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.best");
+                    }
+                }
+            }
+        } else {
+            // assigned node: upgrade to higher prio backend if possible
+            for (int b = 0; b < *node_backend_id; b++) {
+                if (sched->bufts[b] == sched->bufts[*node_backend_id] && ggml_backend_supports_op(sched->backends[b], node)) {
+                    bool supported = true;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if (!ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            supported = false;
+                            break;
+                        }
+                    }
+                    if (supported) {
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.upg");
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    // pass 4: assign backends to remaining src from dst and view_src
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int * cur_backend_id = &tensor_backend_id(node);
+        if (node->view_src != NULL && *cur_backend_id == -1) {
+            *cur_backend_id = tensor_backend_id(node->view_src);
+            SET_CAUSE(node, "4.vsrc");
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            int * src_backend_id = &tensor_backend_id(src);
+            if (*src_backend_id == -1) {
+                if (src->view_src != NULL) {
+                    // views are always on the same backend as the source
+                    *src_backend_id = tensor_backend_id(src->view_src);
+                    SET_CAUSE(src, "4.vsrc");
+                } else {
+                    *src_backend_id = *cur_backend_id;
+                    SET_CAUSE(src, "4.cur");
+                }
+            }
+        }
+    }
+
+    // pass 5: split graph, find tensors that need to be copied
+    {
+        int i_split = 0;
+        struct ggml_backend_sched_split * split = &sched->splits[0];
+        // find the backend of the first split, skipping view ops
+        int i = 0;
+        for (; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (!ggml_is_view_op(node->op)) {
+                split->backend_id = tensor_backend_id(node);
+                break;
+            }
+        }
+        split->i_start = 0;
+        split->n_inputs = 0;
+        int cur_backend_id = split->backend_id;
+        for (; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+
+            const int node_backend_id = tensor_backend_id(node);
+
+            assert(node_backend_id != -1); // all nodes should be assigned by now
+
+            // check if we should start a new split based on the sources of the current node
+            bool need_new_split = false;
+            if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * src = node->src[j];
+                    if (src == NULL) {
+                        continue;
+                    }
+                    // check if a weight is on a different backend
+                    // by starting a new split, the memory of the previously offloaded weights can be reused
+                    if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                        int src_backend_id = tensor_backend_id(src);
+                        if (src_backend_id != cur_backend_id) {
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                    // check if the split has too many inputs
+                    // FIXME: count the number of inputs instead of only checking when full
+                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                        const size_t id = hash_id(src);
+                        int src_backend_id = sched->hv_tensor_backend_ids[id];
+                        bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
+                        if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
+                            //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                }
+            }
+
+            if (node_backend_id != cur_backend_id || need_new_split) {
+                split->i_end = i;
+                i_split++;
+                if (i_split >= sched->splits_capacity) {
+                    sched->splits_capacity *= 2;
+                    sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
+                    GGML_ASSERT(sched->splits != NULL);
+                }
+                GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
+                split = &sched->splits[i_split];
+                split->backend_id = node_backend_id;
+                split->i_start = i;
+                split->n_inputs = 0;
+                cur_backend_id = node_backend_id;
+            }
+
+            // find inputs that are not on the same backend
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+
+                size_t src_id = hash_id(src);
+                const int src_backend_id = sched->hv_tensor_backend_ids[src_id];
+                assert(src_backend_id != -1); // all inputs should be assigned by now
+
+                if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
+                    if (tensor_id_copy(src_id, src_backend_id, 0) == NULL) {
+                        ggml_backend_t backend = sched->backends[src_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy;
+                            if (c == sched->cur_copy) {
+                                tensor_copy = src; // use the original tensor as the current copy
+                            } else {
+                                tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                                ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            }
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_graph_inputs = sched->n_graph_inputs++;
+                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        sched->graph_inputs[n_graph_inputs] = src;
+                    }
+                }
+
+                if (src_backend_id != cur_backend_id && !ggml_backend_sched_buffer_supported(sched, src, cur_backend_id)) {
+                    // create a copy of the input in the split's backend
+                    if (tensor_id_copy(src_id, cur_backend_id, 0) == NULL) {
+                        ggml_backend_t backend = sched->backends[cur_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                            ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            tensor_id_copy(src_id, cur_backend_id, c) = tensor_copy;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_inputs = split->n_inputs++;
+                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        split->inputs[n_inputs] = src;
+                    }
+                    node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
+                }
+            }
+        }
+        split->i_end = graph->n_nodes;
+        sched->n_splits = i_split + 1;
+    }
+
+    if (sched->debug) {
+        ggml_backend_sched_print_assignments(sched, graph);
+    }
+
+    // swap node_backend_ids and leaf _backend_ids with prevs
+    {
+        int * tmp = sched->node_backend_ids;
+        sched->node_backend_ids = sched->prev_node_backend_ids;
+        sched->prev_node_backend_ids = tmp;
+
+        tmp = sched->leaf_backend_ids;
+        sched->leaf_backend_ids = sched->prev_leaf_backend_ids;
+        sched->prev_leaf_backend_ids = tmp;
+    }
+
+    int graph_size = graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    if (sched->graph.size < graph_size) {
+        sched->graph.size = graph_size;
+        sched->graph.nodes = realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
+        sched->graph.leafs = realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *));
+        GGML_ASSERT(sched->graph.nodes != NULL);
+        GGML_ASSERT(sched->graph.leafs != NULL);
+    }
+    sched->graph.n_nodes = 0;
+    sched->graph.n_leafs = 0;
+
+    struct ggml_cgraph * graph_copy = &sched->graph;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &sched->splits[i];
+        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
+
+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
+        for (int j = 0; j < split->n_inputs; j++) {
+            assert(graph_copy->size > (graph_copy->n_nodes + 1));
+
+            struct ggml_tensor * input = split->inputs[j];
+            const size_t input_id = hash_id(input);
+            struct ggml_tensor * input_cpy = tensor_id_copy(input_id, split->backend_id, sched->cur_copy);
+
+            // add a dependency to the input source so that it is not freed before the copy is done
+            struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
+            input_dep->src[0] = input;
+            sched->node_backend_ids[graph_copy->n_nodes] = sched->hv_tensor_backend_ids[input_id];
+            graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
+
+            // add a dependency to the input copy so that it is allocated at the start of the split
+            sched->node_backend_ids[graph_copy->n_nodes] = split->backend_id;
+            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
+        }
+
+        for (int j = split->i_start; j < split->i_end; j++) {
+            assert(graph_copy->size > graph_copy->n_nodes);
+            sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]);
+            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
+        }
+    }
+
+    if (sched->n_copies > 1) {
+        // add input copies as leafs so that they are allocated first
+        for (int i = 0; i < sched->n_graph_inputs; i++) {
+            struct ggml_tensor * input = sched->graph_inputs[i];
+            size_t id = hash_id(input);
+            int backend_id = tensor_backend_id(input);
+            for (int c = 0; c < sched->n_copies; c++) {
+                struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
+                sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+            }
+        }
+
+        for (int i = 0; i < sched->n_splits; i++) {
+            struct ggml_backend_sched_split * split = &sched->splits[i];
+            int backend_id = split->backend_id;
+            for (int j = 0; j < split->n_inputs; j++) {
+                struct ggml_tensor * input = split->inputs[j];
+                size_t id = hash_id(input);
+                for (int c = 0; c < sched->n_copies; c++) {
+                    struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
+                    sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                    graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+                }
+            }
+        }
+    }
+
+    // add leafs from the original graph
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
+        graph_copy->leafs[graph_copy->n_leafs++] = leaf;
+    }
+}
+
+static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
+    bool backend_ids_changed = false;
+    for (int i = 0; i < sched->graph.n_nodes; i++) {
+        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
+            sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
+            backend_ids_changed = true;
+            break;
+        }
+    }
+    if (!backend_ids_changed) {
+        for (int i = 0; i < sched->graph.n_leafs; i++) {
+            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
+                sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
+                backend_ids_changed = true;
+                break;
+            }
+        }
+    }
+
+    // allocate graph
+    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+        // the re-allocation may cause the split inputs to be moved to a different address
+        ggml_backend_sched_synchronize(sched);
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+#endif
+        ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
+        if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+            fprintf(stderr, "%s: failed to allocate graph\n", __func__);
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
+    struct ggml_backend_sched_split * splits = sched->splits;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &splits[i];
+        int split_backend_id = split->backend_id;
+        ggml_backend_t split_backend = sched->backends[split_backend_id];
+
+        // copy the input tensors to the split backend
+        for (int j = 0; j < split->n_inputs; j++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
+            struct ggml_tensor * input = split->inputs[j];
+            struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
+
+            if (input->flags & GGML_TENSOR_FLAG_INPUT) {
+                // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                ggml_backend_tensor_copy(input, input_cpy);
+            } else {
+                // wait for the split backend to finish using the input before overwriting it
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+            }
+        }
+
+        if (!sched->callback_eval) {
+            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
+            if (ec != GGML_STATUS_SUCCESS) {
+                return ec;
+            }
+        } else {
+            // similar to ggml_backend_compare_graph_backend
+            for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {
+                struct ggml_tensor * t = split->graph.nodes[j0];
+
+                // check if the user needs data from this node
+                bool need = sched->callback_eval(t, true, sched->callback_eval_user_data);
+
+                int j1 = j0;
+
+                // determine the range [j0, j1] of nodes that can be computed together
+                while (!need && j1 < split->graph.n_nodes - 1) {
+                    t = split->graph.nodes[++j1];
+                    need = sched->callback_eval(t, true, sched->callback_eval_user_data);
+                }
+
+                struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
+
+                enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &gv);
+                if (ec != GGML_STATUS_SUCCESS) {
+                    return ec;
+                }
+
+                // TODO: pass backend to the callback, then the user can decide if they want to synchronize
+                ggml_backend_synchronize(split_backend);
+
+                if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
+                    break;
+                }
+
+                j0 = j1;
+            }
+        }
+
+        // record the event of this copy
+        if (split->n_inputs > 0) {
+            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]);
+            }
+        }
+    }
+
+    sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
+
+    return GGML_STATUS_SUCCESS;
+}
+
+ggml_backend_sched_t ggml_backend_sched_new(
+        ggml_backend_t * backends,
+        ggml_backend_buffer_type_t * bufts,
+        int n_backends,
+        size_t graph_size,
+        bool parallel) {
+    GGML_ASSERT(n_backends > 0);
+    GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
+    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
+
+    struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
+
+    sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
+    sched->n_backends = n_backends;
+    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
+
+    // initialize hash table
+    // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead)
+    sched->hash_set    = ggml_hash_set_new(graph_size);
+    sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+    sched->hv_tensor_copies      = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
+
+    const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
+    sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
+
+    sched->context_buffer_size = GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer = malloc(sched->context_buffer_size);
+
+    const int initial_splits_capacity = 16;
+    sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
+    sched->splits_capacity = initial_splits_capacity;
+
+    for (int b = 0; b < n_backends; b++) {
+        sched->backends[b] = backends[b];
+        sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
+        GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b]));
+        if (sched->n_copies > 1) {
+            for (int c = 0; c < sched->n_copies; c++) {
+                sched->events[b][c] = ggml_backend_event_new(backends[b]);
+            }
+        }
+    }
+
+    sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
+
+    ggml_backend_sched_reset(sched);
+
+    return sched;
+}
+
+void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+    if (sched == NULL) {
+        return;
+    }
+    for (int b = 0; b < sched->n_backends; b++) {
+        for (int c = 0; c < sched->n_copies; c++) {
+            ggml_backend_event_free(sched->events[b][c]);
+        }
+    }
+    ggml_gallocr_free(sched->galloc);
+    ggml_free(sched->ctx);
+    ggml_hash_set_free(&sched->hash_set);
+    free(sched->splits);
+    free(sched->hv_tensor_backend_ids);
+    free(sched->hv_tensor_copies);
+    free(sched->node_backend_ids);
+    free(sched->leaf_backend_ids);
+    free(sched->prev_node_backend_ids);
+    free(sched->prev_leaf_backend_ids);
+    free(sched->context_buffer);
+    free(sched->graph.nodes);
+    free(sched->graph.leafs);
+    free(sched);
+}
+
+void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
+    // reset state for the next run
+    if (!sched->is_reset) {
+        ggml_hash_set_reset(&sched->hash_set);
+        memset(sched->hv_tensor_backend_ids, -1, sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+        memset(sched->hv_tensor_copies,       0, sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
+        sched->is_reset = true;
+    }
+    sched->is_alloc = false;
+}
+
+bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
+
+    ggml_backend_sched_split_graph(sched, measure_graph);
+
+    if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
+        return false;
+    }
+
+    ggml_backend_sched_reset(sched);
+    ggml_backend_sched_synchronize(sched);
+
+    return true;
+}
+
+bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
+
+    ggml_backend_sched_split_graph(sched, graph);
+
+
+    if (!ggml_backend_sched_alloc_splits(sched)) {
+        return false;
+    }
+
+    sched->is_alloc = true;
+
+    return true;
+}
+
+enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    enum ggml_status err = ggml_backend_sched_graph_compute_async(sched, graph);
+    ggml_backend_sched_synchronize(sched);
+    return err;
+}
+
+enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    if (!sched->is_reset && !sched->is_alloc) {
+        ggml_backend_sched_reset(sched);
+    }
+
+    if (!sched->is_alloc) {
+        if (!ggml_backend_sched_alloc_graph(sched, graph)) {
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+    }
+
+    return ggml_backend_sched_compute_splits(sched);
+}
+
+void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_backend_synchronize(sched->backends[i]);
+    }
+}
+
+void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
+    sched->callback_eval = callback;
+    sched->callback_eval_user_data = user_data;
+}
+
+int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
+    return sched->n_splits;
+}
+
+int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
+    return sched->n_copies;
+}
+
+int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched) {
+    return sched->n_backends;
+}
+
+ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i) {
+    GGML_ASSERT(i >= 0 && i < sched->n_backends);
+    return sched->backends[i];
+}
+
+size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+
+    return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
+}
+
+void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+    tensor_backend_id(node) = backend_index;
+    SET_CAUSE(node, "usr");
+    sched->is_reset = false;
+}
+
+ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+    int backend_index = tensor_backend_id(node);
+    if (backend_index == -1) {
+        return NULL;
+    }
+    return sched->backends[backend_index];
+}
+
+// utils
+
+void ggml_backend_view_init(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    GGML_ASSERT(tensor->view_src != NULL);
+    GGML_ASSERT(tensor->view_src->buffer != NULL);
+    GGML_ASSERT(tensor->view_src->data != NULL);
+
+    tensor->buffer = tensor->view_src->buffer;
+    tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
+    ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
+}
+
+void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    GGML_ASSERT(tensor->data == NULL);
+    GGML_ASSERT(tensor->view_src == NULL);
+    GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
+    GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
+                (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
+
+    tensor->buffer = buffer;
+    tensor->data = addr;
+    ggml_backend_buffer_init_tensor(buffer, tensor);
+}
+
+static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies,
+    struct ggml_context * ctx_allocated, struct ggml_context * ctx_unallocated, struct ggml_tensor * src) {
+
+    GGML_ASSERT(src != NULL);
+    GGML_ASSERT(src->data && "graph must be allocated");
+
+    size_t id = ggml_hash_insert(&hash_set, src);
+    if (id == GGML_HASHSET_ALREADY_EXISTS) {
+        return node_copies[ggml_hash_find(&hash_set, src)];
+    }
+
+    struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
+    if (src->view_src != NULL) {
+        dst->view_src = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, src->view_src);
+        dst->view_offs = src->view_offs;
+    }
+    dst->op = src->op;
+    memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
+    ggml_set_name(dst, src->name);
+
+    // copy src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            continue;
+        }
+        dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
+    }
+
+    node_copies[id] = dst;
+    return dst;
+}
+
+static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
+    size_t id = ggml_hash_find(hash_set, src);
+    if (node_init[id]) {
+        return;
+    }
+    node_init[id] = true;
+
+    struct ggml_tensor * dst = node_copies[id];
+    if (dst->view_src != NULL) {
+        graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
+        ggml_backend_view_init(dst);
+    }
+    else {
+        ggml_backend_tensor_copy(src, dst);
+    }
+
+    // init src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            continue;
+        }
+        graph_copy_init_tensor(hash_set, node_copies, node_init, s);
+    }
+}
+
+struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
+    struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size);
+    struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
+    bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
+
+    struct ggml_init_params params = {
+        /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true
+    };
+
+    struct ggml_context * ctx_allocated = ggml_init(params);
+    struct ggml_context * ctx_unallocated = ggml_init(params);
+
+    if (ctx_allocated == NULL || ctx_unallocated == NULL) {
+        fprintf(stderr, "failed to allocate context for graph copy\n");
+        ggml_hash_set_free(&hash_set);
+        free(node_copies);
+        free(node_init);
+        ggml_free(ctx_allocated);
+        ggml_free(ctx_unallocated);
+        return (struct ggml_backend_graph_copy) {
+            /* .buffer           = */ NULL,
+            /* .ctx_allocated    = */ NULL,
+            /* .ctx_unallocated  = */ NULL,
+            /* .graph            = */ NULL,
+        };
+    }
+
+    // dup nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, node);
+    }
+
+    // allocate nodes
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
+    if (buffer == NULL) {
+        fprintf(stderr, "failed to allocate buffer for graph copy\n");
+        ggml_hash_set_free(&hash_set);
+        free(node_copies);
+        free(node_init);
+        ggml_free(ctx_allocated);
+        ggml_free(ctx_unallocated);
+        return (struct ggml_backend_graph_copy) {
+            /* .buffer           = */ NULL,
+            /* .ctx_allocated    = */ NULL,
+            /* .ctx_unallocated  = */ NULL,
+            /* .graph            = */ NULL,
+        };
+    }
+
+    //printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024);
+
+    // copy data and init views
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_copy_init_tensor(&hash_set, node_copies, node_init, node);
+    }
+
+    // build graph copy
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(&hash_set, node)];
+        graph_copy->nodes[i] = node_copy;
+    }
+    graph_copy->n_nodes = graph->n_nodes;
+
+    ggml_hash_set_free(&hash_set);
+    free(node_copies);
+    free(node_init);
+
+    return (struct ggml_backend_graph_copy) {
+        /* .buffer           = */ buffer,
+        /* .ctx_allocated    = */ ctx_allocated,
+        /* .ctx_unallocated  = */ ctx_unallocated,
+        /* .graph            = */ graph_copy,
+    };
+}
+
+void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
+    ggml_backend_buffer_free(copy.buffer);
+    ggml_free(copy.ctx_allocated);
+    ggml_free(copy.ctx_unallocated);
+}
+
+bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
+    struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
+    if (copy.buffer == NULL) {
+        return false;
+    }
+
+    struct ggml_cgraph * g1 = graph;
+    struct ggml_cgraph * g2 = copy.graph;
+
+    assert(g1->n_nodes == g2->n_nodes);
+
+    for (int i = 0; i < g1->n_nodes; i++) {
+        //printf("eval %d/%d\n", i, g1->n_nodes);
+        struct ggml_tensor * t1 = g1->nodes[i];
+        struct ggml_tensor * t2 = g2->nodes[i];
+
+        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+        ggml_backend_graph_compute(backend1, &g1v);
+        ggml_backend_graph_compute(backend2, &g2v);
+
+        if (ggml_is_view_op(t1->op)) {
+            continue;
+        }
+
+        // compare results, calculate rms etc
+        if (!callback(i, t1, t2, user_data)) {
+            break;
+        }
+    }
+
+    ggml_backend_graph_copy_free(copy);
+
+    return true;
+}
diff --git a/llm/llama.cpp/ggml/src/ggml-blas.cpp b/llm/llama.cpp/ggml/src/ggml-blas.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..71373173598c7743cb6e2cf52c1f844dab8d0b39
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-blas.cpp
@@ -0,0 +1,367 @@
+#include "ggml-blas.h"
+#include "ggml-backend-impl.h"
+
+#include 
+#include 
+
+#if defined(GGML_USE_ACCELERATE)
+#   include 
+#elif defined(GGML_BLAS_USE_MKL)
+#   include 
+#elif defined(GGML_BLAS_USE_BLIS)
+#   include 
+#elif defined(GGML_BLAS_USE_NVPL)
+#   include 
+#else
+#   include 
+#endif
+
+struct ggml_backend_blas_context {
+    int n_threads = GGML_DEFAULT_N_THREADS;
+    std::unique_ptr work_data;
+    size_t work_size = 0;
+#ifndef GGML_USE_OPENMP
+    std::vector> tasks;
+#endif
+};
+
+// helper function to determine if it is better to use BLAS or not
+// for large matrices, BLAS is faster
+static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    // TODO: find the optimal values for these
+    if (ggml_is_contiguous(src0) &&
+        ggml_is_contiguous(src1) &&
+        src1->type == GGML_TYPE_F32 &&
+        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
+
+        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
+        return true;
+    }
+
+    return false;
+}
+
+static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const int64_t ne_plane      = ne01*ne00;
+    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
+
+    if (ctx->work_size < desired_wsize) {
+        ctx->work_data.reset(new char[desired_wsize]);
+        ctx->work_size = desired_wsize;
+    }
+    void * wdata = ctx->work_data.get();
+
+    // convert src0 to float
+    if (type != GGML_TYPE_F32) {
+        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits.to_float;
+
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
+                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;
+
+                const int min_cols_per_thread = 4096;
+                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
+                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
+
+#ifdef GGML_USE_OPENMP
+                #pragma omp parallel for num_threads(n_threads)
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                }
+#else
+                for (int i = 1; i < n_threads; i++) {
+                    const int64_t start =       i*ne01/n_threads;
+                    const int64_t end   = (i + 1)*ne01/n_threads;
+                    if (start < end) {
+                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
+                            for (int64_t i01 = start; i01 < end; i01++) {
+                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                            }
+                        }));
+                    }
+                }
+                {
+                    // reuse the current thread for the first task
+                    const int64_t start = 0;
+                    const int64_t end   = ne01/n_threads;
+                    for (int64_t i01 = start; i01 < end; i01++) {
+                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                    }
+                }
+#endif
+            }
+        }
+
+#ifndef GGML_USE_OPENMP
+        // wait for all tasks to finish
+        for (auto & task : ctx->tasks) {
+            task.get();
+        }
+        ctx->tasks.clear();
+#endif
+    }
+
+#if defined(OPENBLAS_VERSION)
+    openblas_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(GGML_BLAS_USE_BLIS)
+    bli_thread_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(GGML_BLAS_USE_NVPL)
+    nvpl_blas_set_num_threads(ctx->n_threads);
+#endif
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            const int64_t i03 = i13/r3;
+            const int64_t i02 = i12/r2;
+
+            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
+            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
+                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
+
+            if (type != GGML_TYPE_F32) {
+                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
+            }
+
+            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
+                        ne1, ne01, ne10,
+                        1.0f,   y, ne10,
+                                x, ne00,
+                        0.0f,   d, ne01);
+        }
+    }
+}
+
+static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(ne0  == ne00);
+    GGML_ASSERT(ne1  == ne10);
+    GGML_ASSERT(ne2  == ne02);
+    GGML_ASSERT(ne02 == ne12);
+    GGML_ASSERT(ne3  == ne13);
+    GGML_ASSERT(ne03 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    // GGML_ASSERT(nb0 <= nb1);
+    // GGML_ASSERT(nb1 <= nb2);
+    // GGML_ASSERT(nb2 <= nb3);
+
+    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
+    // src0: (k,n)
+    // src1: (k,m)
+    // dst:  (m,n)
+    //
+    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
+    // Also expressed as (major,minor)
+    // a: (m,k): so src1 transposed
+    // b: (k,n): so src0
+    // c: (m,n)
+    //
+    // However, if ggml_is_transposed(src1) is true, then
+    // src1->data already contains a transposed version, so sgemm mustn't
+    // transpose it further.
+
+    int n = src0->ne[0];
+    int k = src0->ne[1];
+    int m = src1->ne[0];
+
+    CBLAS_TRANSPOSE transposeA;
+    int lda;
+
+    if (!ggml_is_transposed(src1)) {
+        transposeA = CblasTrans;
+        lda = m;
+    } else {
+        transposeA = CblasNoTrans;
+        lda = k;
+    }
+
+    float * a = (float *) ((char *) src1->data);
+    float * b = (float *) ((char *) src0->data);
+    float * c = (float *) ((char *) dst->data);
+
+    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
+
+    GGML_UNUSED(ctx);
+}
+
+// backend interface
+
+GGML_CALL static const char * ggml_backend_blas_name(ggml_backend_t backend) {
+    return "BLAS";
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static void ggml_backend_blas_free(ggml_backend_t backend) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+GGML_CALL static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_backend_blas_mul_mat(ctx, node);
+                break;
+
+            case GGML_OP_OUT_PROD:
+                ggml_backend_blas_out_prod(ctx, node);
+                break;
+
+            case GGML_OP_NONE:
+            case GGML_OP_RESHAPE:
+            case GGML_OP_VIEW:
+            case GGML_OP_PERMUTE:
+            case GGML_OP_TRANSPOSE:
+                break;
+
+            default:
+                GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
+           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
+                                          op->src[1]->type == GGML_TYPE_F32 &&
+                                          ggml_is_matrix(src0) &&
+                                          ggml_is_matrix(src1) &&
+                                          ggml_is_contiguous(src0) &&
+                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i blas_backend_i = {
+    /* .get_name                = */ ggml_backend_blas_name,
+    /* .free                    = */ ggml_backend_blas_free,
+    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
+    /* .supports_op             = */ ggml_backend_blas_supports_op,
+    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_blas_guid(void) {
+    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_blas_init(void) {
+    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_blas_guid(),
+        /* .interface = */ blas_backend_i,
+        /* .context   = */ ctx,
+    };
+
+#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
+        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
+    }
+#endif
+
+#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
+    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
+#endif
+
+    return backend;
+}
+
+GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
+}
+
+void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_blas(backend_blas));
+
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
+    ctx->n_threads = n_threads;
+}
diff --git a/llm/llama.cpp/ggml/src/ggml-cann.cpp b/llm/llama.cpp/ggml/src/ggml-cann.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..06930ba2e5bee04870bbc5165a097f42e0ce7bd1
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-cann.cpp
@@ -0,0 +1,2020 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "ggml-cann.h"
+
+#include 
+#include 
+
+#include 
+#include 
+#include 
+#include 
+
+#include "ggml-backend-impl.h"
+#include "ggml-cann/aclnn_ops.h"
+#include "ggml-cann/common.h"
+
+#define GGML_COMMON_DECL_C
+
+#include "ggml-common.h"
+
+/**
+ * @brief Default logging callback for GGML.
+ *
+ * This function is the default logging callback that logs messages to stderr.
+ *
+ * @param level The log level.
+ * @param msg The log message.
+ * @param user_data User data passed to the callback.
+ */
+static void ggml_cann_default_log_callback(enum ggml_log_level level,
+                                           const char* msg, void* user_data) {
+    GGML_UNUSED(level);
+    GGML_UNUSED(user_data);
+    fprintf(stderr, "%s", msg);
+}
+
+ggml_log_callback ggml_cann_log_callback = ggml_cann_default_log_callback;
+void* ggml_cann_log_user_data = NULL;
+
+GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback,
+                                                 void* user_data) {
+    ggml_cann_log_callback = log_callback;
+    ggml_cann_log_user_data = user_data;
+}
+
+#define GGML_CANN_LOG_INFO(...) ggml_cann_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
+#define GGML_CANN_LOG_WARN(...) ggml_cann_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_CANN_LOG_ERROR(...) \
+    ggml_cann_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+GGML_ATTRIBUTE_FORMAT(2, 3)
+
+/**
+ * @brief Log a message using the current logging callback.
+ *
+ * This function formats a log message and passes it to the current logging
+ * callback.
+ *
+ * @param level The log level.
+ * @param format The format string for the log message.
+ * @param ... The arguments for the format string.
+ */
+static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) {
+    if (ggml_cann_log_callback != NULL) {
+        va_list args;
+        va_start(args, format);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            ggml_cann_log_callback(level, buffer, ggml_cann_log_user_data);
+        } else {
+             // vsnprintf adds a null terminator
+            std::vector buffer2(len + 1);
+            va_end(args);
+            va_start(args, format);
+            vsnprintf(&buffer2[0], buffer2.size(), format, args);
+            ggml_cann_log_callback(level, buffer2.data(),
+                                   ggml_cann_log_user_data);
+        }
+        va_end(args);
+    }
+}
+
+/**
+ * @brief Handles CANN errors by printing an error message and aborting.
+ *
+ * @param stmt The statement that caused the error.
+ * @param func The function in which the error occurred.
+ * @param file The file in which the error occurred.
+ * @param line The line number where the error occurred.
+ * @param msg The error message.
+ */
+[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
+                                  const char* file, int line, const char* msg) {
+    int32_t id = -1;
+    aclrtGetDevice(&id);
+
+    GGML_CANN_LOG_ERROR("CANN error: %s\n", msg);
+    GGML_CANN_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func,
+            file, line);
+    GGML_CANN_LOG_ERROR("  %s\n", stmt);
+    // abort with GGML_ASSERT to get a stack trace
+    GGML_ABORT("CANN error");
+}
+
+/**
+ * @brief Sets the device to be used by CANN.
+ *
+ * @param device The device ID to set.
+ */
+void ggml_cann_set_device(const int32_t device) {
+    // TODO: uncomment these lines after empty context has fixed.
+    // int current_device;
+    // ACL_CHECK(aclrtGetDevice(¤t_device));
+
+    // if (device == current_device) {
+    //   return;
+    // }
+    ACL_CHECK(aclrtSetDevice(device));
+}
+
+/**
+ * @brief Retrieves the current device ID.
+ *
+ * @return The current device ID.
+ */
+int32_t ggml_cann_get_device() {
+    int32_t id;
+    ACL_CHECK(aclrtGetDevice(&id));
+    return id;
+}
+
+/**
+ * @brief Initialize the CANN device information.
+ *
+ * This function initializes the CANN device information by obtaining the
+ * device count and setting the memory allocation granularity for each device.
+ *
+ * @return A structure containing the device information.
+ */
+static ggml_cann_device_info ggml_cann_init() {
+    ggml_cann_device_info info = {};
+
+    aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count);
+
+    if (err != ACL_SUCCESS) {
+        GGML_CANN_LOG_ERROR("%s: failed to initialize CANN: %s\n",
+                __func__, aclGetRecentErrMsg());
+        return info;
+    }
+
+    GGML_ASSERT(info.device_count <= GGML_CANN_MAX_DEVICES);
+
+    for (int id = 0; id < info.device_count; ++id) {
+        aclrtPhysicalMemProp prop = {};
+        prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
+        prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
+        prop.memAttr = ACL_HBM_MEM_HUGE;
+        prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
+        prop.location.id = id;
+        prop.reserve = 0;
+        ACL_CHECK(aclrtMemGetAllocationGranularity(
+            &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
+            &info.devices[id].vmm_granularity));
+    }
+
+    // TODO: add more device info later.
+    return info;
+}
+
+/**
+ * @brief Retrieve the CANN device information.
+ *
+ * This function returns a reference to a structure containing the CANN device
+ * information. The device information is initialized once and reused on
+ * subsequent calls.
+ *
+ * @return A reference to the structure containing the device information.
+ */
+const ggml_cann_device_info& ggml_cann_info() {
+    static ggml_cann_device_info info = ggml_cann_init();
+    return info;
+}
+
+//#define DEBUG_CANN_MALLOC
+/**
+ * @brief A pool of CANN buffers(legacy).
+ *
+ * This class manages a pool of CANN buffers for a specific device.
+ */
+struct ggml_cann_pool_leg : public ggml_cann_pool {
+    /**
+     * @brief The maximum number of buffers in the pool.
+     */
+    static const int MAX_BUFFERS = 256;
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer memory.
+        size_t size = 0;      ///< Size of the buffer.
+    };
+
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    ggml_cann_buffer buffer_pool[MAX_BUFFERS] = {};
+
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_leg(int device) : device(device) {}
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_leg() {
+        ggml_cann_set_device(device);
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+            }
+        }
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+#ifdef DEBUG_CANN_MALLOC
+        int nnz = 0;
+        size_t max_size = 0;
+#endif
+        size_t best_diff = 1ull << 36;
+        int ibest = -1;
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+#ifdef DEBUG_CANN_MALLOC
+                ++nnz;
+                if (b.size > max_size) max_size = b.size;
+#endif
+                if (b.size >= size) {
+                    size_t diff = b.size - size;
+                    if (diff < best_diff) {
+                        best_diff = diff;
+                        ibest = i;
+                        if (!best_diff) {
+                            void* ptr = b.ptr;
+                            *actual_size = b.size;
+                            b.ptr = nullptr;
+                            b.size = 0;
+                            return ptr;
+                        }
+                    }
+                }
+            }
+        }
+        if (ibest >= 0) {
+            ggml_cann_buffer& b = buffer_pool[ibest];
+            void* ptr = b.ptr;
+            *actual_size = b.size;
+            b.ptr = nullptr;
+            b.size = 0;
+            return ptr;
+        }
+        void* ptr;
+        size_t look_ahead_size = (size_t)(1.05 * size);
+        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
+        ggml_cann_set_device(device);
+        ACL_CHECK(
+            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = look_ahead_size;
+        pool_size += look_ahead_size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_CANN_LOG_INFO(
+            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
+            "requested %u MB\n",
+            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
+            (uint32_t)(pool_size / 1024 / 1024),
+            (uint32_t)(size / 1024 / 1024));
+#endif
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr == nullptr) {
+                b.ptr = ptr;
+                b.size = size;
+                return;
+            }
+        }
+        // memory should always buffered. these memory may still needed by
+        // tasks in stream.
+        // TODO, fix me.
+        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+    }
+};
+
+/**
+ * @brief A pool of CANN buffers with virtual memory.
+ *
+ * This class manages a pool of CANN buffers with virtual memory for a specific
+ * device.
+ */
+struct ggml_cann_pool_vmm : public ggml_cann_pool {
+    /**
+     * @brief The maximum size of the virtual memory pool (32 GB).
+     */
+    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Pointer to the start of the virtual memory pool.
+     */
+    void* pool_addr = 0;
+
+    /**
+     * @brief Amount of virtual memory used in the pool.
+     */
+    size_t pool_used = 0;
+
+    /**
+     * @brief Total size of the virtual memory pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Allocation granularity for the virtual memory pool.
+     */
+    size_t granularity;
+
+    /**
+     * @brief Handles for the physical memory allocated.
+     */
+    std::vector handles;
+
+    /**
+     * @brief Offsets for the mapped memory regions.
+     */
+    std::vector map_offsets;
+
+    /**
+     * @brief Constructor to initialize the buffer pool with virtual memory for
+     * a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_vmm(int device)
+        : device(device),
+          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
+
+    /**
+     * @brief Destructor to free all buffers in the virtual memory pool.
+     */
+    ~ggml_cann_pool_vmm() {
+        if (pool_addr != 0) {
+            for (auto& offset : map_offsets) {
+                ACL_CHECK(aclrtUnmapMem(offset));
+            }
+            for (auto& handle : handles) {
+                ACL_CHECK(aclrtFreePhysical(handle));
+            }
+            ACL_CHECK(aclrtReleaseMemAddress(pool_addr));
+        }
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size in the virtual memory pool.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        // round up the allocation size to the alignment to ensure that all
+        // allocations are aligned for all data types
+        const size_t alignment = 128;
+        size = alignment * ((size + alignment - 1) / alignment);
+
+        size_t avail = pool_size - pool_used;
+
+        if (size > avail) {
+            // round up to the next multiple of the granularity
+            size_t reserve_size = size - avail;
+            reserve_size =
+                granularity * ((reserve_size + granularity - 1) / granularity);
+
+            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
+
+            // allocate more physical memory
+            aclrtPhysicalMemProp prop = {};
+            prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
+            prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
+            prop.memAttr = ACL_HBM_MEM_HUGE;
+            prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
+            prop.location.id = device;
+            prop.reserve = 0;
+            aclrtDrvMemHandle handle;
+            ACL_CHECK(aclrtMallocPhysical(&handle, reserve_size, &prop, 0));
+
+            // reserve virtual address space (if not already reserved)
+            if (pool_addr == 0) {
+                ACL_CHECK(aclrtReserveMemAddress(
+                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
+            }
+
+            // map at the end of the pool
+            ACL_CHECK(aclrtMapMem((char*)pool_addr + pool_size, reserve_size, 0,
+                                  handle, 0));
+
+            handles.push_back(handle);
+            map_offsets.push_back((char*)pool_addr + pool_size);
+
+            // add to the pool
+            pool_size += reserve_size;
+
+            // GGML_CANN_LOG_INFO("cann pool[%d]: size increased to %llu MB (
+            // reserved %llu MB)\n",
+            //       device, (unsigned long long) (pool_size/1024/1024),
+            //       (unsigned long long) (reserve_size/1024/1024));
+        }
+
+        GGML_ASSERT(pool_addr != 0);
+
+        void* ptr = (void*)((char*)pool_addr + pool_used);
+        *actual_size = size;
+        pool_used += size;
+
+#ifdef DEBUG_CANN_MALLOC
+        GGML_CANN_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device,
+               (unsigned long long)size, (unsigned long long)ptr);
+#endif
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the virtual memory pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+#ifdef DEBUG_CANN_MALLOC
+        GGML_CANN_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device,
+               (unsigned long long)size, (unsigned long long)ptr);
+#endif
+
+        pool_used -= size;
+
+        // all deallocations must be in reverse order of the allocations
+        GGML_ASSERT(ptr == (void*)((char*)pool_addr + pool_used));
+    }
+};
+
+/**
+ * @brief Create a new CANN pool for a specific device.
+ *
+ * Factory method to create a new CANN pool object based on the device type.
+ *
+ * @param device The device ID for which to create the pool.
+ * @return A unique pointer to the created CANN pool.
+ */
+std::unique_ptr ggml_backend_cann_context::new_pool_for_device(
+    int device) {
+    // return std::unique_ptr(new ggml_cann_pool_leg(device));
+    return std::unique_ptr(new ggml_cann_pool_vmm(device));
+}
+
+// cann buffer
+/**
+ * @brief Context for managing a CANN buffer associated with a specific device.
+ *
+ * This structure holds information about a CANN buffer, including the device
+ * ID, device pointer, and a name derived from GGML_CANN_NAME and the device ID.
+ */
+struct ggml_backend_cann_buffer_context {
+    int32_t device;  ///< The device ID associated with this buffer context.
+    void* dev_ptr =
+        nullptr;  ///< Pointer to the device memory allocated for the buffer.
+
+    /**
+     * @brief Constructor to initialize the CANN buffer context.
+     *
+     * @param device The device ID associated with this buffer context.
+     * @param dev_ptr Pointer to the device memory allocated for the buffer.
+     */
+    ggml_backend_cann_buffer_context(int32_t device, void* dev_ptr)
+        : device(device),
+          dev_ptr(dev_ptr) {}
+
+    /**
+     * @brief Destructor to free the device memory allocated for the buffer.
+     */
+    ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
+};
+
+/**
+ * @brief Retrieve the name associated with a CANN buffer.
+ *
+ * This function returns the name of a CANN buffer, which is stored in the
+ * context of the buffer.
+ *
+ * @param buffer The CANN buffer whose name is to be retrieved.
+ * @return A pointer to a C-string containing the name of the buffer.
+ */
+
+GGML_CALL static const char* ggml_backend_cann_buffer_get_name(
+    ggml_backend_buffer_t buffer) {
+    return "CANN";
+
+    GGML_UNUSED(buffer);
+}
+
+/**
+ * @brief Check if a buffer is a CANN buffer.
+ *
+ * This function checks if a given buffer is a CANN buffer by comparing its
+ * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
+ *
+ * @param buffer The buffer to check.
+ * @return true if the buffer is a CANN buffer, false otherwise.
+ */
+GGML_CALL static bool ggml_backend_buffer_is_cann(
+    ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_cann_buffer_get_name;
+}
+
+/**
+ * @brief Free resources associated with a CANN buffer.
+ *
+ * This function frees the resources associated with a CANN buffer, including
+ * its context.
+ *
+ * @param buffer The CANN buffer to free.
+ */
+GGML_CALL static void ggml_backend_cann_buffer_free_buffer(
+    ggml_backend_buffer_t buffer) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+    delete ctx;
+}
+
+/**
+ * @brief Retrieve the base pointer of a CANN buffer.
+ *
+ * This function returns the base pointer of a CANN buffer, which points to the
+ * device memory allocated for the buffer.
+ *
+ * @param buffer The CANN buffer whose base pointer is to be retrieved.
+ * @return A pointer to the base of the device memory allocated for the buffer.
+ */
+GGML_CALL static void* ggml_backend_cann_buffer_get_base(
+    ggml_backend_buffer_t buffer) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+    return ctx->dev_ptr;
+}
+
+/**
+ * @brief Transform quantized Q4.0 tensor data into a format suitable for CANN
+ * processing.
+ *
+ * This function transforms quantized Q4.0 tensor data into a format suitable
+ * for CANN processing. It extracts quantization values and scales from the
+ * source data and prepares them in a format expected by CANN operations.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data in Q4.0 format.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
+                                                       const void* src,
+                                                       void* dst) {
+
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK4_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
+
+    uint8_t* quant_offset = (uint8_t*)dst;
+    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        const block_q4_0* group =
+            (const block_q4_0*)((const char*)src + i * sizeof(block_q4_0));
+        *scale_offset = group->d;
+        scale_offset++;
+
+        // 0-15
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            (*quant_offset) = (group->qs[j] & 0x0F);
+            (*quant_offset) |= ((group->qs[j + 1] << 4));
+            quant_offset++;
+        }
+
+        // 16-31
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            (*quant_offset) = (group->qs[j] >> 4);
+            (*quant_offset) |= (group->qs[j + 1] & 0xF0);
+            quant_offset++;
+        }
+    }
+
+    // put (uint4b_t -8) into int4b_t
+    for (quant_offset = (uint8_t*)dst;
+         quant_offset < (uint8_t*)dst + quant_bytes; quant_offset++) {
+        (*quant_offset) ^= 0x88;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into quantized Q4.0 format.
+ *
+ * This function transforms CANN processed data back into quantized Q4.0 format.
+ * It reverses the transformation performed by
+ * ggml_backend_cann_transform_q4_0(), converting the data back into its
+ * original quantized form.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source buffer containing transformed data.
+ * @param dst Pointer to the destination buffer where the Q4.0 formatted data
+ * will be stored.
+ */
+GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
+    const ggml_tensor* tensor, void* src, void* dst) {
+
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK4_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
+
+    uint8_t* quant_offset = (uint8_t*)src;
+    uint16_t* scale_offset = (uint16_t*)((char*)src + quant_bytes);
+
+    for (; quant_offset < (uint8_t*)src + quant_bytes; quant_offset++) {
+        (*quant_offset) ^= 0x88;
+    }
+    quant_offset = (uint8_t*)src;
+
+    for (int i = 0; i < groups; i++) {
+        block_q4_0* group = (block_q4_0*)((char*)dst + i * sizeof(block_q4_0));
+        group->d = *scale_offset;
+        scale_offset++;
+
+        // 0-15
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            group->qs[j] = ((*quant_offset) & 0x0F);
+            group->qs[j + 1] = ((*quant_offset) >> 4);
+            quant_offset++;
+        }
+
+        // 16-31
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            group->qs[j] |= ((*quant_offset) << 4);
+            group->qs[j + 1] |= ((*quant_offset) & 0xF0);
+            quant_offset++;
+        }
+    }
+}
+
+/**
+ * @brief Transform quantized Q8.0 tensor data into a format suitable for CANN
+ * processing.
+ *
+ * This function transforms quantized Q8.0 tensor data into a format suitable
+ * for CANN processing. It extracts quantization values and scales from the
+ * source data and prepares them in a format expected by CANN operations.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data in Q8.0 format.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+GGML_CALL static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor,
+                                                       const void* src,
+                                                       void* dst) {
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK8_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t);
+
+    uint8_t* quant_offset = (uint8_t*)dst;
+    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        const block_q8_0* group =
+            (const block_q8_0*)((const char*)src + i * sizeof(block_q8_0));
+        *scale_offset = group->d;
+        scale_offset++;
+        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
+        memcpy(quant_offset, group->qs, group_quant_size);
+        quant_offset += group_quant_size;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into quantized Q8.0 format.
+ *
+ * This function transforms CANN processed data back into quantized Q8.0 format.
+ * It reverses the transformation performed by
+ * ggml_backend_cann_transform_q8_0(), converting the data back into its
+ * original quantized form.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source buffer containing transformed data.
+ * @param dst Pointer to the destination buffer where the Q8.0 formatted data
+ * will be stored.
+ */
+GGML_CALL static void ggml_backend_cann_transform_back_q8_0(
+    const ggml_tensor* tensor, const void* src, void* dst) {
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK8_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t);
+
+    const uint8_t* quant_offset = (const uint8_t*)src;
+    const uint16_t* scale_offset =
+        (const uint16_t*)((const char*)src + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        block_q8_0* group = (block_q8_0*)((char*)dst + i * sizeof(block_q8_0));
+        group->d = *scale_offset;
+        scale_offset++;
+        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
+        memcpy(group->qs, quant_offset, group_quant_size);
+        quant_offset += group_quant_size;
+    }
+}
+
+/**
+ * @brief Transform tensor data based on its type for CANN processing.
+ *
+ * This function transforms tensor data based on its quantization type for CANN
+ * processing. It dispatches the transformation based on the tensor's type to
+ * specialized functions handling Q4.0 and Q8.0 formats.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data to be transformed.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+GGML_CALL static void ggml_backend_cann_transform(ggml_tensor* tensor,
+                                                  const void* src, void* dst) {
+    switch (tensor->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_backend_cann_transform_q4_0(tensor, src, dst);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_backend_cann_transform_q8_0(tensor, src, dst);
+            break;
+        default:
+            break;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into tensor data based on its type.
+ *
+ * This function transforms CANN processed data back into tensor data based on
+ * its quantization type for Q4.0 and Q8.0 formats. It dispatches the
+ * transformation based on the tensor's type to specialized functions.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data containing CANN processed data.
+ * @param dst Pointer to the destination buffer where transformed tensor data
+ * will be stored.
+ */
+GGML_CALL static void ggml_backend_cann_transform_back(
+    const ggml_tensor* tensor, void* src, void* dst) {
+    switch (tensor->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_backend_cann_transform_back_q4_0(tensor, src, dst);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_backend_cann_transform_back_q8_0(tensor, src, dst);
+            break;
+        default:
+            break;
+    }
+}
+
+/**
+ * @brief Check if transformation is needed for a given tensor type.
+ *
+ * This function checks if transformation is needed for a given tensor type
+ * to prepare data for CANN processing.
+ *
+ * @param type The tensor type to check.
+ * @return true if transformation is needed, false otherwise.
+ */
+GGML_CALL static bool need_transform(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+/**
+ * @brief Initialize a tensor using data from a CANN buffer.
+ *
+ * This function initializes a tensor using data from a CANN buffer.
+ * It handles special cases such as views and quantization.
+ *
+ * @param buffer The CANN buffer from which to initialize the tensor.
+ * @param tensor Pointer to the tensor to be initialized.
+ */
+GGML_CALL static void ggml_backend_cann_buffer_init_tensor(
+    ggml_backend_buffer_t buffer, ggml_tensor* tensor) {
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
+        return;
+    }
+
+    // TODO: can backend doesn't support quantized yet. Just leave the code
+    // here.
+    if (ggml_is_quantized(tensor->type)) {
+        // Initialize padding to 0 to avoid possible NaN values
+        size_t original_size = ggml_nbytes(tensor);
+        size_t padded_size =
+            ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
+
+        if (padded_size > original_size && tensor->view_src == nullptr) {
+            size_t memset_size = padded_size - original_size;
+            ACL_CHECK(aclrtMemset((char*)tensor->data + original_size,
+                                  memset_size, 0, memset_size));
+        }
+    }
+}
+
+// TODO: need handle tensor which has paddings.
+/**
+ * @brief Set tensor data in a CANN buffer.
+ *
+ * This function sets tensor data in a CANN buffer, handling transformations
+ * if needed based on the tensor's type.
+ *
+ * @param buffer The CANN buffer where the tensor data will be set.
+ * @param tensor Pointer to the tensor whose data will be set.
+ * @param data Pointer to the source data to be copied into the tensor.
+ * @param offset Offset in the source data from where to start copying.
+ * @param size Size of the data to be copied, in bytes.
+ */
+GGML_CALL static void ggml_backend_cann_buffer_set_tensor(
+    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context *ctx =
+        (ggml_backend_cann_buffer_context *)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    // TODO: refer to cann(#6017), it use thread's default stream.
+    // For acl, synchronous functions use this default stream.
+    // Why aclrtSynchronizeDevice?
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
+    } else {
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
+
+#ifndef NDEBUG
+        void *check_buffer = malloc(size);
+        ggml_backend_cann_transform_back(tensor, transform_buffer,
+                                         check_buffer);
+        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
+        free(check_buffer);
+#endif
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
+                              transform_buffer, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Get tensor data from a CANN buffer.
+ *
+ * This function retrieves tensor data from a CANN buffer, handling
+ * transformations if needed based on the tensor's type.
+ *
+ * @param buffer The CANN buffer from which to retrieve tensor data.
+ * @param tensor Pointer to the tensor whose data will be retrieved.
+ * @param data Pointer to the destination buffer where the tensor data will be
+ * copied.
+ * @param offset Offset in the destination buffer where to start copying.
+ * @param size Size of the data to be copied, in bytes.
+ */
+GGML_CALL static void ggml_backend_cann_buffer_get_tensor(
+    ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
+                              ACL_MEMCPY_DEVICE_TO_HOST));
+    } else {
+        void* transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
+                              (char*)tensor->data + offset, size,
+                              ACL_MEMCPY_DEVICE_TO_HOST));
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Copy tensor data between CANN buffers if possible.
+ *
+ * This function copies tensor data between CANN buffers if the source and
+ * destination buffers are CANN buffers and they meet the necessary conditions
+ * (same device or devices can access each other).
+ *
+ * @param buffer The destination CANN buffer where the tensor data will be
+ * copied.
+ * @param src Pointer to the source tensor whose data will be copied.
+ * @param dst Pointer to the destination tensor where the data will be copied.
+ * @return true if the copy operation succeeded, false otherwise.
+ */
+GGML_CALL static bool ggml_backend_cann_buffer_cpy_tensor(
+    ggml_backend_buffer_t buffer, const ggml_tensor* src, ggml_tensor* dst) {
+    if (ggml_backend_buffer_is_cann(src->buffer)) {
+        ggml_backend_cann_buffer_context* src_ctx =
+            (ggml_backend_cann_buffer_context*)src->buffer->context;
+        ggml_backend_cann_buffer_context* dst_ctx =
+            (ggml_backend_cann_buffer_context*)buffer->context;
+
+        size_t memcpy_size = ggml_nbytes(src);
+        // Same device.
+        if (src_ctx->device == dst_ctx->device) {
+            ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
+                                  (const char*)src->data, memcpy_size,
+                                  ACL_MEMCPY_DEVICE_TO_DEVICE));
+            return true;
+        } else {
+            // Different device but can access by peer.
+            int32_t canAccessPeer = 0;
+            ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device,
+                                               dst_ctx->device));
+            if (canAccessPeer) {
+                ggml_cann_set_device(src_ctx->device);
+                ACL_CHECK(aclrtDeviceEnablePeerAccess(dst_ctx->device, 0));
+                ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
+                                      (const char*)src->data, memcpy_size,
+                                      ACL_MEMCPY_DEVICE_TO_DEVICE));
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+/**
+ * @brief Clear a CANN buffer by setting all its memory to a specified value.
+ *
+ * This function clears a CANN buffer by setting all its memory to a specified
+ * value.
+ *
+ * @param buffer The CANN buffer to be cleared.
+ * @param value The value to which each byte in the buffer will be set.
+ */
+GGML_CALL static void ggml_backend_cann_buffer_clear(
+    ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    ACL_CHECK(aclrtMemset(ctx->dev_ptr, buffer->size, value, buffer->size));
+}
+
+/**
+ * @brief Interface for a CANN buffer in the backend.
+ *
+ * This structure defines function pointers to operations that can be performed
+ * on a CANN buffer within the backend.
+ */
+static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
+    /* .get_name        = */ ggml_backend_cann_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cann_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
+    /* .set_tensor      = */ ggml_backend_cann_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cann_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cann_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cann_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// cann buffer type
+/**
+ * @brief Structure representing context information for a specific backend
+ * buffer type.
+ */
+struct ggml_backend_cann_buffer_type_context {
+    int32_t
+        device; /**< Device identifier associated with the buffer context. */
+    std::string name; /**< Name associated with the buffer context. */
+};
+
+/**
+ * @brief Retrieves the name associated with a CANN buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN buffer type context.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+GGML_CALL static const char* ggml_backend_cann_buffer_type_name(
+    ggml_backend_buffer_type_t buft) {
+    return "CANN";
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Allocates a new CANN buffer of the specified type and size.
+ *
+ * This function allocates a new CANN buffer on the specified device with the
+ * given size.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @param size Size in bytes of the buffer to allocate.
+ * @return Pointer to the allocated buffer, or nullptr if allocation fails.
+ */
+GGML_CALL static ggml_backend_buffer_t
+ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                           size_t size) {
+    ggml_backend_cann_buffer_type_context* buft_ctx =
+        (ggml_backend_cann_buffer_type_context*)buft->context;
+
+    ggml_cann_set_device(buft_ctx->device);
+
+    size = std::max(size, (size_t)1);
+
+    void* dev_ptr;
+    aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
+    if (err != ACL_SUCCESS) {
+        GGML_CANN_LOG_ERROR(
+            "%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n",
+            __func__, size / 1024.0 / 1024.0, buft_ctx->device,
+            aclGetRecentErrMsg());
+        return nullptr;
+    }
+
+    ggml_backend_cann_buffer_context* ctx =
+        new ggml_backend_cann_buffer_context(buft_ctx->device, dev_ptr);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cann_buffer_interface,
+                                    ctx, size);
+}
+
+/**
+ * @brief Retrieves the memory alignment requirement for CANN buffers of this
+ * type.
+ *
+ * This function returns the alignment requirement in bytes for memory allocated
+ * by the CANN buffer type.
+ *
+ * @param buft Pointer to the buffer type context (unused in this
+ * implementation).
+ * @return The alignment requirement in bytes (fixed at 128 bytes for CANN
+ * buffers).
+ */
+GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alignment(
+    ggml_backend_buffer_type_t buft) {
+    return 128;
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Calculates the allocation size required for a tensor in a CANN buffer.
+ *
+ * Computes the total allocation size needed for storing the tensor's data in a
+ * CANN buffer, considering any necessary padding or adjustments for quantized
+ * types.
+ *
+ * @param buft Pointer to the buffer type context (unused in this
+ * implementation).
+ * @param tensor Pointer to the tensor for which the allocation size is
+ * calculated.
+ * @return The total allocation size in bytes required for the tensor in the
+ * CANN buffer.
+ */
+GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alloc_size(
+    ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+    size_t size = ggml_nbytes(tensor);
+    int64_t ne0 = tensor->ne[0];
+
+    // last line must bigger than 32, because every single op deal at
+    // least 32 bytes.
+    // TODO: quantized type?
+    // int64_t line_size = ne0 * ggml_element_size(tensor);
+    // int64_t line_size_align_32 = (line_size + 31) & ~31;
+    // size += (line_size_align_32 - line_size);
+
+    // TODO: not support quantized yet.
+    // TODO: consider un-continue tensor.
+    if (ggml_is_quantized(tensor->type)) {
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(
+                tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return size;
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Interface for managing CANN buffer types in the GGML backend.
+ *
+ * Provides function pointers for allocating, querying properties, and managing
+ * memory for CANN buffer types in the GGML backend.
+ */
+static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_cann_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_cann_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_cann_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL,  // defaults to SIZE_MAX
+    /* .get_alloc_size   = */ ggml_backend_cann_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+/**
+ * @brief Retrieves the CANN buffer type for a specified device.
+ *
+ * This function initializes and returns the buffer type interface associated
+ * with the given device. It ensures thread-safe access using a mutex.
+ *
+ * @param device The device index for which to retrieve the buffer type.
+ * @return A pointer to the buffer type interface for the specified device, or
+ * nullptr if the device index is out of range.
+ */
+GGML_CALL ggml_backend_buffer_type_t
+ggml_backend_cann_buffer_type(int32_t device) {
+    static std::mutex mutex;
+    std::lock_guard lock(mutex);
+
+    if (device >= ggml_backend_cann_get_device_count()) {
+        return nullptr;
+    }
+
+    static ggml_backend_buffer_type
+        ggml_backend_cann_buffer_types[GGML_CANN_MAX_DEVICES];
+
+    static bool ggml_backend_cann_buffer_type_initialized = false;
+
+    if (!ggml_backend_cann_buffer_type_initialized) {
+        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
+            ggml_backend_cann_buffer_types[i] = {
+                /* .iface    = */ ggml_backend_cann_buffer_type_interface,
+                /* .context  = */
+                 new ggml_backend_cann_buffer_type_context{
+                    i, "CANN" + std::to_string(i)},
+            };
+        }
+        ggml_backend_cann_buffer_type_initialized = true;
+    }
+
+    return &ggml_backend_cann_buffer_types[device];
+}
+
+/**
+ * @brief Computes the forward operation for a given tensor using CANN
+ * operations.
+ *
+ * This function selects the appropriate CANN operation based on the type of
+ * operation specified in the tensor and performs the computation.
+ *
+ * @param ctx The CANN context containing necessary resources and
+ * configurations.
+ * @param dst The destination tensor where the result of the computation will be
+ * stored.
+ * @return true if the computation was successful; false otherwise.
+ */
+static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
+                                      struct ggml_tensor* dst) {
+    switch (dst->op) {
+        case GGML_OP_REPEAT:
+            ggml_cann_repeat(ctx, dst);
+            break;
+        case GGML_OP_GET_ROWS:
+            ggml_cann_get_rows(ctx, dst);
+            break;
+        case GGML_OP_DUP:
+            ggml_cann_dup(ctx, dst);
+            break;
+        case GGML_OP_ADD:
+            ggml_cann_add(ctx, dst);
+            break;
+        case GGML_OP_ACC:
+            ggml_cann_acc(ctx, dst);
+            break;
+        case GGML_OP_MUL:
+            ggml_cann_mul_div(ctx, dst);
+            break;
+        case GGML_OP_DIV:
+            ggml_cann_mul_div(ctx, dst);
+            break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(dst)) {
+                case GGML_UNARY_OP_GELU:
+                    ggml_cann_activation(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SILU:
+                    ggml_cann_activation(
+                        ctx, dst);
+                    break;
+                // TODO: Use faster gelu??
+                case GGML_UNARY_OP_GELU_QUICK:
+                    ggml_cann_activation(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    ggml_cann_activation(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_RELU:
+                    ggml_cann_activation(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSIGMOID:
+                    ggml_cann_activation(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSWISH:
+                    ggml_cann_activation(ctx, dst);
+                    break;
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_NORM:
+            ggml_cann_norm(ctx, dst);
+            break;
+        case GGML_OP_GROUP_NORM:
+            ggml_cann_group_norm(ctx, dst);
+            break;
+        case GGML_OP_CONCAT:
+            ggml_cann_concat(ctx, dst);
+            break;
+        case GGML_OP_UPSCALE:
+            ggml_cann_upsample_nearest2d(ctx, dst);
+            break;
+        case GGML_OP_PAD:
+            ggml_cann_pad(ctx, dst);
+            break;
+        case GGML_OP_ARANGE:
+            ggml_cann_arange(ctx, dst);
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            ggml_cann_timestep_embedding(ctx, dst);
+            break;
+        case GGML_OP_LEAKY_RELU:
+            ggml_cann_leaky_relu(ctx, dst);
+            break;
+        case GGML_OP_RMS_NORM:
+            ggml_cann_rms_norm(ctx, dst);
+            break;
+        case GGML_OP_MUL_MAT:
+            ggml_cann_mul_mat(ctx, dst);
+            break;
+        case GGML_OP_MUL_MAT_ID:
+            return false;
+        case GGML_OP_SCALE:
+            ggml_cann_scale(ctx, dst);
+            break;
+        case GGML_OP_SQR:
+            ggml_cann_sqr(ctx, dst);
+            break;
+        case GGML_OP_CLAMP:
+            ggml_cann_clamp(ctx, dst);
+            break;
+        case GGML_OP_CPY:
+            ggml_cann_cpy(ctx, dst);
+            break;
+        case GGML_OP_CONT:
+            ggml_cann_dup(ctx, dst);
+            break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            break;
+        case GGML_OP_DIAG_MASK_INF:
+            ggml_cann_diag_mask(ctx, dst, -INFINITY);
+            break;
+        case GGML_OP_SOFT_MAX:
+            ggml_cann_softmax(ctx, dst);
+            break;
+        case GGML_OP_ROPE:
+            ggml_cann_rope(ctx, dst);
+            break;
+        case GGML_OP_IM2COL:
+            ggml_cann_im2col(ctx, dst);
+            break;
+        case GGML_OP_POOL_2D:
+            ggml_cann_pool2d(ctx, dst);
+            break;
+        case GGML_OP_SUM_ROWS:
+            ggml_cann_sum_rows(ctx, dst);
+            break;
+        case GGML_OP_ARGSORT:
+            ggml_cann_argsort(ctx, dst);
+            break;
+        default:
+            return false;
+    }
+
+    return true;
+}
+
+// backend
+/**
+ * @brief Retrieves the name associated with the CANN backend.
+ *
+ * This function returns the name assigned to the CANN backend, which is stored
+ * in the context of the provided backend structure.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @return A pointer to a constant string representing the backend name.
+ */
+GGML_CALL static const char* ggml_backend_cann_name(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    return cann_ctx->name.c_str();
+}
+
+/**
+ * @brief Frees resources associated with the CANN backend.
+ *
+ * This function releases resources associated with the CANN backend context
+ * and resets the device associated with the backend to its initial state.
+ *
+ * @param backend Pointer to the CANN backend structure to be freed.
+ */
+GGML_CALL static void ggml_backend_cann_free(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+    ACL_CHECK(aclrtSynchronizeDevice());
+    ACL_CHECK(aclrtResetDevice(cann_ctx->device));
+
+    // finalize when last backend freed.
+    if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
+        ACL_CHECK(aclFinalize());
+    }
+
+    delete cann_ctx;
+    delete backend;
+}
+
+/**
+ * @brief Retrieves the default buffer type associated with the CANN backend.
+ *
+ * This function returns the buffer type specific to the device associated
+ * with the CANN backend. It is used to allocate buffers for computations
+ * performed by the backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @return Pointer to the buffer type structure for the CANN backend.
+ */
+GGML_CALL static ggml_backend_buffer_type_t
+ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    return ggml_backend_cann_buffer_type(cann_ctx->device);
+}
+
+/**
+ * @brief Sets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously sets tensor data in the CANN backend. Depending
+ * on the tensor type, it may perform data transformations before copying data
+ * to the device.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to set data for.
+ * @param data Pointer to the host data to copy to the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
+GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
+                                                         ggml_tensor *tensor,
+                                                         const void *data,
+                                                         size_t offset,
+                                                         size_t size) {
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
+                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
+                                   cann_ctx->stream()));
+    } else {
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
+
+#ifndef NDEBUG
+        void *check_buffer = malloc(size);
+        ggml_backend_cann_transform_back(tensor, transform_buffer,
+                                         check_buffer);
+        GGML_ASSERT(memcmp(data, check_buffer, size));
+        free(check_buffer);
+#endif
+        ACL_CHECK(aclrtMemcpyAsync(
+            (char *)tensor->data + offset, size, transform_buffer, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+        free(transform_buffer);
+    }
+}
+
+GGML_CALL static void ggml_backend_cann_get_tensor_async(
+    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+                "unsupported buffer type");
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
+                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
+                                   cann_ctx->stream()));
+    } else {
+        void *transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(
+            transform_buffer, size, (char *)tensor->data + offset, size,
+            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Asynchronously copies tensor data between CANN backends.
+ *
+ * This function copies tensor data asynchronously between two CANN backends. It
+ * checks if both tensors reside in CANN buffers and whether the devices support
+ * peer-to-peer access for direct copying. If not, it returns false.
+ *
+ * @param backend_src Pointer to the source CANN backend structure.
+ * @param backend_dst Pointer to the destination CANN backend structure.
+ * @param src Pointer to the source tensor to copy data from.
+ * @param dst Pointer to the destination tensor to copy data to.
+ * @return true if the copy operation succeeds, false otherwise.
+ */
+GGML_CALL static bool ggml_backend_cann_cpy_tensor_async(
+    ggml_backend_t backend_src, ggml_backend_t backend_dst,
+    const ggml_tensor* src, ggml_tensor* dst) {
+    GGML_ASSERT(ggml_backend_is_cann(backend_src) ||
+                ggml_backend_is_cann(backend_dst));
+
+    if (!ggml_backend_buffer_is_cann(src->buffer) ||
+        !ggml_backend_buffer_is_cann(dst->buffer)) {
+        return false;
+    }
+
+    ggml_backend_buffer_t buf_src =
+        src->view_src ? src->view_src->buffer : src->buffer;
+    ggml_backend_buffer_t buf_dst =
+        dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    ggml_backend_cann_context* cann_ctx_src =
+        (ggml_backend_cann_context*)backend_src->context;
+    ggml_backend_cann_context* cann_ctx_dst =
+        (ggml_backend_cann_context*)backend_dst->context;
+
+    size_t copy_size = ggml_nbytes(dst);
+    if (backend_src != backend_dst) {
+        ggml_backend_cann_buffer_context* buf_ctx_src =
+            (ggml_backend_cann_buffer_context*)buf_src->context;
+        ggml_backend_cann_buffer_context* buf_ctx_dst =
+            (ggml_backend_cann_buffer_context*)buf_dst->context;
+
+        GGML_ASSERT(cann_ctx_src->device == buf_ctx_src->device);
+        GGML_ASSERT(cann_ctx_dst->device == buf_ctx_dst->device);
+
+        int32_t canAccessPeer = 0;
+        ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, cann_ctx_src->device,
+                                           cann_ctx_dst->device));
+        if (!canAccessPeer) {
+            return false;
+        }
+
+        // need open both directions for memcpyasync between devices.
+        ggml_cann_set_device(cann_ctx_dst->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0));
+        ggml_cann_set_device(cann_ctx_src->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
+
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                   cann_ctx_src->stream()));
+
+        //TODO: workaround for Event didn`t work here.
+        aclrtSynchronizeStream(cann_ctx_src->stream());
+    } else {
+        // src and dst are on the same backend
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                   cann_ctx_dst->stream()));
+    }
+
+    return true;
+}
+
+/**
+ * @brief Synchronizes a CANN backend.
+ *
+ * This function synchronizes the specified CANN backend by waiting for all
+ * operations in its associated stream to complete.
+ *
+ * @param backend Pointer to the CANN backend structure to synchronize.
+ */
+GGML_CALL static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+}
+
+/**
+ * @brief Computes a computational graph using a CANN backend.
+ *
+ * This function computes the operations defined in the computational graph
+ * using the specified CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure to use for computation.
+ * @param cgraph Pointer to the computational graph structure containing nodes
+ *               representing operations to be computed.
+ * @return enum ggml_status Returns GGML_STATUS_SUCCESS if computation
+ *         completes successfully, otherwise an appropriate error status.
+ */
+GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute(
+    ggml_backend_t backend, ggml_cgraph* cgraph) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_tensor* node = cgraph->nodes[i];
+
+        if (ggml_is_empty(node) || node->op == GGML_OP_NONE) {
+            continue;
+        }
+
+        bool ok = ggml_cann_compute_forward(*cann_ctx, node);
+
+        if (!ok) {
+            GGML_CANN_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
+                    node->name, ggml_op_name(node->op));
+        }
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific operation.
+ *
+ * This function checks whether the specified operation is supported by the
+ * CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure to check support for
+ *                the operation.
+ * @param op Pointer to the tensor representing the operation to check.
+ * @return bool Returns true if the operation is supported by the backend,
+ *              otherwise false.
+ */
+GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
+                                                    const ggml_tensor* op) {
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_TANH:
+                    return true;
+                default:
+                    return false;
+            }
+        case GGML_OP_MUL_MAT: {
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F16:
+                case GGML_TYPE_F32:
+                case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_MUL_MAT_ID:
+            return false;
+        // embedding
+        case GGML_OP_GET_ROWS: {
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q8_0:
+                    return true;
+                default:
+                    return false;
+            }
+        } break;
+        case GGML_OP_CPY: {
+            switch (op->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_DUP:
+        case GGML_OP_REPEAT:
+        case GGML_OP_CONCAT:
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_NORM:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_CLAMP:
+        case GGML_OP_CONT:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_ROPE:
+        case GGML_OP_IM2COL:
+        case GGML_OP_POOL_2D:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_ACC:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_ARANGE:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_LEAKY_RELU:
+            return true;
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(backend);
+}
+
+/**
+ * @brief Checks if the backend buffer type is associated with the CANN backend.
+ *
+ * This function checks whether the provided backend buffer type is associated
+ * with the CANN backend based on the comparison of its name retrieval function
+ * pointer.
+ *
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the buffer type is associated with the CANN
+ * backend, otherwise false.
+ */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific backend buffer type.
+ *
+ * This function determines whether the CANN backend supports the given backend
+ * buffer type by comparing the device context of the backend and buffer type.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the CANN backend supports the buffer type,
+ *              otherwise false.
+ */
+GGML_CALL static bool ggml_backend_cann_supports_buft(
+    ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_context * cann_ctx =
+                        (ggml_backend_cann_context *)backend->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == cann_ctx->device;
+    }
+    return false;
+}
+
+/**
+ * @brief Determines if a tensor operation should be offloaded to the CANN
+ * backend.
+ *
+ * This function checks if a given tensor operation should be offloaded to the
+ * CANN backend based on the operation type and the size of the tensor. It
+ * returns true if the second dimension (ne[1]) of the tensor is greater than or
+ * equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param op Pointer to the tensor operation to check.
+ * @return bool Returns true if the operation should be offloaded, otherwise
+ * false.
+ */
+GGML_CALL static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
+                                                   const ggml_tensor* op) {
+    const int min_batch_size = 32;
+    GGML_UNUSED(backend);
+
+    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
+}
+
+/**
+ * @brief Creates a new event for the CANN backend.
+ *
+ * This function initializes a new event for the CANN backend by setting the
+ * device and creating an ACL runtime event. The created event is then wrapped
+ * in a ggml_backend_event structure and returned.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @return ggml_backend_event_t Returns a pointer to the new event structure.
+ */
+static ggml_backend_event_t ggml_backend_cann_event_new(
+    ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    aclrtEvent event;
+    ACL_CHECK(aclrtCreateEvent(&event));
+
+    return new ggml_backend_event{
+        /* .backend = */ backend,
+        /* .context = */ event,
+    };
+}
+
+/**
+ * @brief Frees a CANN backend event.
+ *
+ * This function destroys the ACL runtime event associated with the given CANN
+ * backend event and then deletes the event structure itself.
+ *
+ * @param event Pointer to the event structure to be freed.
+ */
+static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
+    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
+
+    delete event;
+}
+
+/**
+ * @brief Records an event on the CANN backend stream.
+ *
+ * This function records the given event on the ACL runtime stream associated
+ * with the backend context.
+ *
+ * @param event Pointer to the event structure to be recorded.
+ */
+static void ggml_backend_cann_event_record(ggml_backend_event_t event) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)event->backend->context;
+
+    ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream()));
+}
+
+/**
+ * @brief Waits for a recorded event to complete on the CANN backend stream.
+ *
+ * This function makes the given backend wait for the event to complete on its
+ * ACL runtime stream.
+ *
+ * @param backend Pointer to the backend structure.
+ * @param event Pointer to the event structure that the backend needs to wait
+ * for.
+ */
+static void ggml_backend_cann_event_wait(ggml_backend_t backend,
+                                         ggml_backend_event_t event) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    if (ggml_backend_is_cann(event->backend)) {
+        ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
+                                       (aclrtEvent)event->context));
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+/**
+ * @brief Synchronizes the given event on the CANN backend.
+ *
+ * This function waits for the specified event to complete on the ACL runtime.
+ *
+ * @param event Pointer to the event structure to be synchronized.
+ */
+static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
+    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
+}
+
+/**
+ * @brief Structure defining the interface for the CANN backend.
+ *
+ * This structure contains function pointers for various operations
+ * supported by the CANN backend, including name retrieval, memory
+ * management, tensor operations, synchronization, and event handling.
+ */
+static ggml_backend_i ggml_backend_cann_interface = {
+    /* .get_name                = */ ggml_backend_cann_name,
+    /* .free                    = */ ggml_backend_cann_free,
+    /* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type,
+    /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
+    /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
+    /* .synchronize             = */ ggml_backend_cann_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_cann_graph_compute,
+    /* .supports_op             = */ ggml_backend_cann_supports_op,
+    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
+    /* .offload_op              = */ ggml_backend_cann_offload_op,
+    /* .event_new               = */ ggml_backend_cann_event_new,
+    /* .event_free              = */ ggml_backend_cann_event_free,
+    /* .event_record            = */ ggml_backend_cann_event_record,
+    /* .event_wait              = */ ggml_backend_cann_event_wait,
+    /* .event_synchronize       = */ ggml_backend_cann_event_synchronize,
+};
+
+/**
+ * @brief Return the hardcoded GUID for the CANN backend.
+ *
+ * This function returns a static GUID which uniquely identifies the CANN
+ * backend.
+ *
+ * @return A pointer to the static GUID.
+ */
+static ggml_guid_t ggml_backend_cann_guid() {
+    static ggml_guid guid = {0xa1, 0x94, 0xaf, 0xac, 0xbd, 0x4f, 0x47, 0x34,
+                             0xbe, 0x1a, 0x9e, 0x71, 0x1f, 0x9e, 0xed, 0x64};
+    return &guid;
+}
+
+GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) {
+    aclInit(nullptr);
+    if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
+        GGML_CANN_LOG_ERROR("%s: error: invalid device %d\n", __func__, device);
+        return nullptr;
+    }
+
+    ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device);
+    if (ctx == nullptr) {
+        GGML_CANN_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return nullptr;
+    }
+
+    ggml_backend_t cann_backend =
+        new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
+                         /* .interface = */ ggml_backend_cann_interface,
+                         /* .context   = */ ctx};
+
+    return cann_backend;
+}
+
+GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend) {
+    return backend != NULL &&
+           ggml_guid_matches(backend->guid, ggml_backend_cann_guid());
+}
+
+GGML_CALL int32_t ggml_backend_cann_get_device_count() {
+    return ggml_cann_info().device_count;
+}
+
+GGML_CALL void ggml_backend_cann_get_device_description(
+    int32_t device, char* description, size_t description_size) {
+    ggml_cann_set_device(device);
+    const char* soc_name = aclrtGetSocName();
+    snprintf(description, description_size, "%s", soc_name);
+}
+
+GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
+                                                   size_t* total) {
+    ggml_cann_set_device(device);
+    ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
+}
+
+// backend registry
+/**
+ * @brief Initializes a CANN backend based on the provided parameters.
+ *
+ * This function initializes a CANN backend using the device index and then
+ * initializes the backend using `ggml_backend_cann_init`.
+ *
+ * @param params Parameters for initialization (unused in this implementation).
+ * @param user_data User data containing the device index to initialize the
+ * backend.
+ * @return ggml_backend_t The initialized CANN backend.
+ */
+GGML_CALL static ggml_backend_t ggml_backend_reg_cann_init(const char* params,
+                                                           void* user_data) {
+    ggml_backend_t cann_backend =
+        ggml_backend_cann_init((int)(intptr_t)user_data);
+    return cann_backend;
+
+    GGML_UNUSED(params);
+}
+
+extern "C" GGML_CALL int ggml_backend_cann_reg_devices();
+
+/**
+ * @brief Registers CANN (Ascend) devices as backend options.
+ *
+ * This function initializes ACL, retrieves the number of available CANN
+ * devices, and registers each device as a backend option using
+ * `ggml_backend_register`. Each device is given a unique name based on
+ * `GGML_CANN_NAME` followed by its index.
+ *
+ * @return int The number of CANN devices registered.
+ */
+GGML_CALL int ggml_backend_cann_reg_devices() {
+    uint32_t device_count = ggml_backend_cann_get_device_count();
+    // initialization
+    for (uint32_t i = 0; i < device_count; i++) {
+        char name[128];
+        snprintf(name, sizeof(name), "CANN%d", i);
+        ggml_backend_register(name, ggml_backend_reg_cann_init,
+                              ggml_backend_cann_buffer_type(i),
+                              (void*)(intptr_t)i);
+    }
+    return device_count;
+}
diff --git a/llm/llama.cpp/ggml/src/ggml-cann/.clang-format b/llm/llama.cpp/ggml/src/ggml-cann/.clang-format
new file mode 100644
index 0000000000000000000000000000000000000000..2ad03d73919364e45ccdf57e2c877b4469994dde
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-cann/.clang-format
@@ -0,0 +1,168 @@
+---
+Language:        Cpp
+# BasedOnStyle:  Google
+AccessModifierOffset: -1
+AlignAfterOpenBracket: Align
+AlignConsecutiveMacros: false
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+AlignEscapedNewlines: Left
+AlignOperands:   true
+AlignTrailingComments: true
+AllowAllArgumentsOnNextLine: true
+AllowAllConstructorInitializersOnNextLine: true
+AllowAllParametersOfDeclarationOnNextLine: true
+AllowShortBlocksOnASingleLine: Never
+AllowShortCaseLabelsOnASingleLine: false
+AllowShortFunctionsOnASingleLine: All
+AllowShortLambdasOnASingleLine: All
+AllowShortIfStatementsOnASingleLine: WithoutElse
+AllowShortLoopsOnASingleLine: true
+AlwaysBreakAfterDefinitionReturnType: None
+AlwaysBreakAfterReturnType: None
+AlwaysBreakBeforeMultilineStrings: true
+AlwaysBreakTemplateDeclarations: Yes
+BinPackArguments: true
+BinPackParameters: true
+BraceWrapping:
+  AfterCaseLabel:  false
+  AfterClass:      false
+  AfterControlStatement: false
+  AfterEnum:       false
+  AfterFunction:   false
+  AfterNamespace:  false
+  AfterObjCDeclaration: false
+  AfterStruct:     false
+  AfterUnion:      false
+  AfterExternBlock: false
+  BeforeCatch:     false
+  BeforeElse:      false
+  IndentBraces:    false
+  SplitEmptyFunction: true
+  SplitEmptyRecord: true
+  SplitEmptyNamespace: true
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Attach
+BreakBeforeInheritanceComma: false
+BreakInheritanceList: BeforeColon
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: false
+BreakConstructorInitializers: BeforeColon
+BreakAfterJavaFieldAnnotations: false
+BreakStringLiterals: true
+ColumnLimit:     80
+CommentPragmas:  '^ IWYU pragma:'
+CompactNamespaces: false
+ConstructorInitializerAllOnOneLineOrOnePerLine: true
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: true
+DeriveLineEnding: true
+DerivePointerAlignment: true
+DisableFormat:   false
+ExperimentalAutoDetectBinPacking: false
+FixNamespaceComments: true
+ForEachMacros:
+  - foreach
+  - Q_FOREACH
+  - BOOST_FOREACH
+IncludeBlocks:   Regroup
+IncludeCategories:
+  - Regex:           '^'
+    Priority:        2
+    SortPriority:    0
+  - Regex:           '^<.*\.h>'
+    Priority:        1
+    SortPriority:    0
+  - Regex:           '^<.*'
+    Priority:        2
+    SortPriority:    0
+  - Regex:           '.*'
+    Priority:        3
+    SortPriority:    0
+IncludeIsMainRegex: '([-_](test|unittest))?$'
+IncludeIsMainSourceRegex: ''
+IndentCaseLabels: true
+IndentGotoLabels: true
+IndentPPDirectives: None
+IndentWidth:     4
+IndentWrappedFunctionNames: false
+JavaScriptQuotes: Leave
+JavaScriptWrapImports: true
+KeepEmptyLinesAtTheStartOfBlocks: false
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBinPackProtocolList: Never
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: true
+PenaltyBreakAssignment: 2
+PenaltyBreakBeforeFirstCallParameter: 1
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyBreakTemplateDeclaration: 10
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Left
+RawStringFormats:
+  - Language:        Cpp
+    Delimiters:
+      - cc
+      - CC
+      - cpp
+      - Cpp
+      - CPP
+      - 'c++'
+      - 'C++'
+    CanonicalDelimiter: ''
+    BasedOnStyle:    google
+  - Language:        TextProto
+    Delimiters:
+      - pb
+      - PB
+      - proto
+      - PROTO
+    EnclosingFunctions:
+      - EqualsProto
+      - EquivToProto
+      - PARSE_PARTIAL_TEXT_PROTO
+      - PARSE_TEST_PROTO
+      - PARSE_TEXT_PROTO
+      - ParseTextOrDie
+      - ParseTextProtoOrDie
+    CanonicalDelimiter: ''
+    BasedOnStyle:    google
+ReflowComments:  true
+SortIncludes:    true
+SortUsingDeclarations: true
+SpaceAfterCStyleCast: false
+SpaceAfterLogicalNot: false
+SpaceAfterTemplateKeyword: true
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeCpp11BracedList: false
+SpaceBeforeCtorInitializerColon: true
+SpaceBeforeInheritanceColon: true
+SpaceBeforeParens: ControlStatements
+SpaceBeforeRangeBasedForLoopColon: true
+SpaceInEmptyBlock: false
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 2
+SpacesInAngles:  false
+SpacesInConditionalStatement: false
+SpacesInContainerLiterals: true
+SpacesInCStyleCastParentheses: false
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+SpaceBeforeSquareBrackets: false
+Standard:        Auto
+StatementMacros:
+  - Q_UNUSED
+  - QT_REQUIRE_VERSION
+TabWidth:        8
+UseCRLF:         false
+UseTab:          Never
+...
+
diff --git a/llm/llama.cpp/ggml/src/ggml-cann/Doxyfile b/llm/llama.cpp/ggml/src/ggml-cann/Doxyfile
new file mode 100644
index 0000000000000000000000000000000000000000..2b009e8f9bd26acad24f58b43ec97cc4d7d6e879
--- /dev/null
+++ b/llm/llama.cpp/ggml/src/ggml-cann/Doxyfile
@@ -0,0 +1,2579 @@
+# Doxyfile 1.8.17
+
+# This file describes the settings to be used by the documentation system
+# doxygen (www.doxygen.org) for a project.
+#
+# All text after a double hash (##) is considered a comment and is placed in
+# front of the TAG it is preceding.
+#
+# All text after a single hash (#) is considered a comment and will be ignored.
+# The format is:
+# TAG = value [value, ...]
+# For lists, items can also be appended using:
+# TAG += value [value, ...]
+# Values that contain spaces should be placed between quotes (\" \").
+
+#---------------------------------------------------------------------------
+# Project related configuration options
+#---------------------------------------------------------------------------
+
+# This tag specifies the encoding used for all characters in the configuration
+# file that follow. The default is UTF-8 which is also the encoding used for all
+# text before the first occurrence of this tag. Doxygen uses libiconv (or the
+# iconv built into libc) for the transcoding. See
+# https://www.gnu.org/software/libiconv/ for the list of possible encodings.
+# The default value is: UTF-8.
+
+DOXYFILE_ENCODING      = UTF-8
+
+# The PROJECT_NAME tag is a single word (or a sequence of words surrounded by
+# double-quotes, unless you are using Doxywizard) that should identify the
+# project for which the documentation is generated. This name is used in the
+# title of most generated pages and in a few other places.
+# The default value is: My Project.
+
+PROJECT_NAME           = "llama.cpp"
+
+# The PROJECT_NUMBER tag can be used to enter a project or revision number. This
+# could be handy for archiving the generated documentation or if some version
+# control system is used.
+
+PROJECT_NUMBER         =
+
+# Using the PROJECT_BRIEF tag one can provide an optional one line description
+# for a project that appears at the top of each page and should give viewer a
+# quick idea about the purpose of the project. Keep the description short.
+
+PROJECT_BRIEF          = "llama inference engine"
+
+# With the PROJECT_LOGO tag one can specify a logo or an icon that is included
+# in the documentation. The maximum height of the logo should not exceed 55
+# pixels and the maximum width should not exceed 200 pixels. Doxygen will copy
+# the logo to the output directory.
+
+PROJECT_LOGO           =
+
+# The OUTPUT_DIRECTORY tag is used to specify the (relative or absolute) path
+# into which the generated documentation will be written. If a relative path is
+# entered, it will be relative to the location where doxygen was started. If
+# left blank the current directory will be used.
+
+OUTPUT_DIRECTORY       = docs
+
+# If the CREATE_SUBDIRS tag is set to YES then doxygen will create 4096 sub-
+# directories (in 2 levels) under the output directory of each output format and
+# will distribute the generated files over these directories. Enabling this
+# option can be useful when feeding doxygen a huge amount of source files, where
+# putting all generated files in the same directory would otherwise causes
+# performance problems for the file system.
+# The default value is: NO.
+
+CREATE_SUBDIRS         = NO
+
+# If the ALLOW_UNICODE_NAMES tag is set to YES, doxygen will allow non-ASCII
+# characters to appear in the names of generated files. If set to NO, non-ASCII
+# characters will be escaped, for example _xE3_x81_x84 will be used for Unicode
+# U+3044.
+# The default value is: NO.
+
+ALLOW_UNICODE_NAMES    = NO
+
+# The OUTPUT_LANGUAGE tag is used to specify the language in which all
+# documentation generated by doxygen is written. Doxygen will use this
+# information to generate all constant output in the proper language.
+# Possible values are: Afrikaans, Arabic, Armenian, Brazilian, Catalan, Chinese,
+# Chinese-Traditional, Croatian, Czech, Danish, Dutch, English (United States),
+# Esperanto, Farsi (Persian), Finnish, French, German, Greek, Hungarian,
+# Indonesian, Italian, Japanese, Japanese-en (Japanese with English messages),
+# Korean, Korean-en (Korean with English messages), Latvian, Lithuanian,
+# Macedonian, Norwegian, Persian (Farsi), Polish, Portuguese, Romanian, Russian,
+# Serbian, Serbian-Cyrillic, Slovak, Slovene, Spanish, Swedish, Turkish,
+# Ukrainian and Vietnamese.
+# The default value is: English.
+
+OUTPUT_LANGUAGE        = English
+
+# The OUTPUT_TEXT_DIRECTION tag is used to specify the direction in which all
+# documentation generated by doxygen is written. Doxygen will use this
+# information to generate all generated output in the proper direction.
+# Possible values are: None, LTR, RTL and Context.
+# The default value is: None.
+
+OUTPUT_TEXT_DIRECTION  = None
+
+# If the BRIEF_MEMBER_DESC tag is set to YES, doxygen will include brief member
+# descriptions after the members that are listed in the file and class
+# documentation (similar to Javadoc). Set to NO to disable this.
+# The default value is: YES.
+
+BRIEF_MEMBER_DESC      = YES
+
+# If the REPEAT_BRIEF tag is set to YES, doxygen will prepend the brief
+# description of a member or function before the detailed description
+#
+# Note: If both HIDE_UNDOC_MEMBERS and BRIEF_MEMBER_DESC are set to NO, the
+# brief descriptions will be completely suppressed.
+# The default value is: YES.
+
+REPEAT_BRIEF           = YES
+
+# This tag implements a quasi-intelligent brief description abbreviator that is
+# used to form the text in various listings. Each string in this list, if found
+# as the leading text of the brief description, will be stripped from the text
+# and the result, after processing the whole list, is used as the annotated
+# text. Otherwise, the brief description is used as-is. If left blank, the
+# following values are used ($name is automatically replaced with the name of
+# the entity):The $name class, The $name widget, The $name file, is, provides,
+# specifies, contains, represents, a, an and the.
+
+ABBREVIATE_BRIEF       = "The $name class" \
+                         "The $name widget" \
+                         "The $name file" \
+                         is \
+                         provides \
+                         specifies \
+                         contains \
+                         represents \
+                         a \
+                         an \
+                         the
+
+# If the ALWAYS_DETAILED_SEC and REPEAT_BRIEF tags are both set to YES then
+# doxygen will generate a detailed section even if there is only a brief
+# description.
+# The default value is: NO.
+
+ALWAYS_DETAILED_SEC    = NO
+
+# If the INLINE_INHERITED_MEMB tag is set to YES, doxygen will show all
+# inherited members of a class in the documentation of that class as if those
+# members were ordinary class members. Constructors, destructors and assignment
+# operators of the base classes will not be shown.
+# The default value is: NO.
+
+INLINE_INHERITED_MEMB  = NO
+
+# If the FULL_PATH_NAMES tag is set to YES, doxygen will prepend the full path
+# before files name in the file list and in the header files. If set to NO the
+# shortest path that makes the file name unique will be used
+# The default value is: YES.
+
+FULL_PATH_NAMES        = YES
+
+# The STRIP_FROM_PATH tag can be used to strip a user-defined part of the path.
+# Stripping is only done if one of the specified strings matches the left-hand
+# part of the path. The tag can be used to show relative paths in the file list.
+# If left blank the directory from which doxygen is run is used as the path to
+# strip.
+#
+# Note that you can specify absolute paths here, but also relative paths, which
+# will be relative from the directory where doxygen is started.
+# This tag requires that the tag FULL_PATH_NAMES is set to YES.
+
+STRIP_FROM_PATH        =
+
+# The STRIP_FROM_INC_PATH tag can be used to strip a user-defined part of the
+# path mentioned in the documentation of a class, which tells the reader which
+# header file to include in order to use a class. If left blank only the name of
+# the header file containing the class definition is used. Otherwise one should
+# specify the list of include paths that are normally passed to the compiler
+# using the -I flag.
+
+STRIP_FROM_INC_PATH    =
+
+# If the SHORT_NAMES tag is set to YES, doxygen will generate much shorter (but
+# less readable) file names. This can be useful is your file systems doesn't
+# support long names like on DOS, Mac, or CD-ROM.
+# The default value is: NO.
+
+SHORT_NAMES            = NO
+
+# If the JAVADOC_AUTOBRIEF tag is set to YES then doxygen will interpret the
+# first line (until the first dot) of a Javadoc-style comment as the brief
+# description. If set to NO, the Javadoc-style will behave just like regular Qt-
+# style comments (thus requiring an explicit @brief command for a brief
+# description.)
+# The default value is: NO.
+
+JAVADOC_AUTOBRIEF      = NO
+
+# If the JAVADOC_BANNER tag is set to YES then doxygen will interpret a line
+# such as
+# /***************
+# as being the beginning of a Javadoc-style comment "banner". If set to NO, the
+# Javadoc-style will behave just like regular comments and it will not be
+# interpreted by doxygen.
+# The default value is: NO.
+
+JAVADOC_BANNER         = NO
+
+# If the QT_AUTOBRIEF tag is set to YES then doxygen will interpret the first
+# line (until the first dot) of a Qt-style comment as the brief description. If
+# set to NO, the Qt-style will behave just like regular Qt-style comments (thus
+# requiring an explicit \brief command for a brief description.)
+# The default value is: NO.
+
+QT_AUTOBRIEF           = NO
+
+# The MULTILINE_CPP_IS_BRIEF tag can be set to YES to make doxygen treat a
+# multi-line C++ special comment block (i.e. a block of //! or /// comments) as
+# a brief description. This used to be the default behavior. The new default is
+# to treat a multi-line C++ comment block as a detailed description. Set this
+# tag to YES if you prefer the old behavior instead.
+#
+# Note that setting this tag to YES also means that rational rose comments are
+# not recognized any more.
+# The default value is: NO.
+
+MULTILINE_CPP_IS_BRIEF = NO
+
+# If the INHERIT_DOCS tag is set to YES then an undocumented member inherits the
+# documentation from any documented member that it re-implements.
+# The default value is: YES.
+
+INHERIT_DOCS           = YES
+
+# If the SEPARATE_MEMBER_PAGES tag is set to YES then doxygen will produce a new
+# page for each member. If set to NO, the documentation of a member will be part
+# of the file/class/namespace that contains it.
+# The default value is: NO.
+
+SEPARATE_MEMBER_PAGES  = NO
+
+# The TAB_SIZE tag can be used to set the number of spaces in a tab. Doxygen
+# uses this value to replace tabs by spaces in code fragments.
+# Minimum value: 1, maximum value: 16, default value: 4.
+
+TAB_SIZE               = 4
+
+# This tag can be used to specify a number of aliases that act as commands in
+# the documentation. An alias has the form:
+# name=value
+# For example adding
+# "sideeffect=@par Side Effects:\n"
+# will allow you to put the command \sideeffect (or @sideeffect) in the
+# documentation, which will result in a user-defined paragraph with heading
+# "Side Effects:". You can put \n's in the value part of an alias to insert
+# newlines (in the resulting output). You can put ^^ in the value part of an
+# alias to insert a newline as if a physical newline was in the original file.
+# When you need a literal { or } or , in the value part of an alias you have to
+# escape them by means of a backslash (\), this can lead to conflicts with the
+# commands \{ and \} for these it is advised to use the version @{ and @} or use
+# a double escape (\\{ and \\})
+
+ALIASES                =
+
+# This tag can be used to specify a number of word-keyword mappings (TCL only).
+# A mapping has the form "name=value". For example adding "class=itcl::class"
+# will allow you to use the command class in the itcl::class meaning.
+
+TCL_SUBST              =
+
+# Set the OPTIMIZE_OUTPUT_FOR_C tag to YES if your project consists of C sources
+# only. Doxygen will then generate output that is more tailored for C. For
+# instance, some of the names that are used will be different. The list of all
+# members will be omitted, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_FOR_C  = NO
+
+# Set the OPTIMIZE_OUTPUT_JAVA tag to YES if your project consists of Java or
+# Python sources only. Doxygen will then generate output that is more tailored
+# for that language. For instance, namespaces will be presented as packages,
+# qualified scopes will look different, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_JAVA   = NO
+
+# Set the OPTIMIZE_FOR_FORTRAN tag to YES if your project consists of Fortran
+# sources. Doxygen will then generate output that is tailored for Fortran.
+# The default value is: NO.
+
+OPTIMIZE_FOR_FORTRAN   = NO
+
+# Set the OPTIMIZE_OUTPUT_VHDL tag to YES if your project consists of VHDL
+# sources. Doxygen will then generate output that is tailored for VHDL.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_VHDL   = NO
+
+# Set the OPTIMIZE_OUTPUT_SLICE tag to YES if your project consists of Slice
+# sources only. Doxygen will then generate output that is more tailored for that
+# language. For instance, namespaces will be presented as modules, types will be
+# separated into more groups, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_SLICE  = NO
+
+# Doxygen selects the parser to use depending on the extension of the files it
+# parses. With this tag you can assign which parser to use for a given
+# extension. Doxygen has a built-in mapping, but you can override or extend it
+# using this tag. The format is ext=language, where ext is a file extension, and
+# language is one of the parsers supported by doxygen: IDL, Java, JavaScript,
+# Csharp (C#), C, C++, D, PHP, md (Markdown), Objective-C, Python, Slice,
+# Fortran (fixed format Fortran: FortranFixed, free formatted Fortran:
+# FortranFree, unknown formatted Fortran: Fortran. In the later case the parser
+# tries to guess whether the code is fixed or free formatted code, this is the
+# default for Fortran type files), VHDL, tcl. For instance to make doxygen treat
+# .inc files as Fortran files (default is PHP), and .f files as C (default is
+# Fortran), use: inc=Fortran f=C.
+#
+# Note: For files without extension you can use no_extension as a placeholder.
+#
+# Note that for custom extensions you also need to set FILE_PATTERNS otherwise
+# the files are not read by doxygen.
+
+EXTENSION_MAPPING      =
+
+# If the MARKDOWN_SUPPORT tag is enabled then doxygen pre-processes all comments
+# according to the Markdown format, which allows for more readable
+# documentation. See https://daringfireball.net/projects/markdown/ for details.
+# The output of markdown processing is further processed by doxygen, so you can
+# mix doxygen, HTML, and XML commands with Markdown formatting. Disable only in
+# case of backward compatibilities issues.
+# The default value is: YES.
+
+MARKDOWN_SUPPORT       = YES
+
+# When the TOC_INCLUDE_HEADINGS tag is set to a non-zero value, all headings up
+# to that level are automatically included in the table of contents, even if
+# they do not have an id attribute.
+# Note: This feature currently applies only to Markdown headings.
+# Minimum value: 0, maximum value: 99, default value: 5.
+# This tag requires that the tag MARKDOWN_SUPPORT is set to YES.
+
+TOC_INCLUDE_HEADINGS   = 5
+
+# When enabled doxygen tries to link words that correspond to documented
+# classes, or namespaces to their corresponding documentation. Such a link can
+# be prevented in individual cases by putting a % sign in front of the word or
+# globally by setting AUTOLINK_SUPPORT to NO.
+# The default value is: YES.
+
+AUTOLINK_SUPPORT       = YES
+
+# If you use STL classes (i.e. std::string, std::vector, etc.) but do not want
+# to include (a tag file for) the STL sources as input, then you should set this
+# tag to YES in order to let doxygen match functions declarations and
+# definitions whose arguments contain STL classes (e.g. func(std::string);
+# versus func(std::string) {}). This also make the inheritance and collaboration
+# diagrams that involve STL classes more complete and accurate.
+# The default value is: NO.
+
+BUILTIN_STL_SUPPORT    = NO
+
+# If you use Microsoft's C++/CLI language, you should set this option to YES to
+# enable parsing support.
+# The default value is: NO.
+
+CPP_CLI_SUPPORT        = NO
+
+# Set the SIP_SUPPORT tag to YES if your project consists of sip (see:
+# https://www.riverbankcomputing.com/software/sip/intro) sources only. Doxygen
+# will parse them like normal C++ but will assume all classes use public instead
+# of private inheritance when no explicit protection keyword is present.
+# The default value is: NO.
+
+SIP_SUPPORT            = NO
+
+# For Microsoft's IDL there are propget and propput attributes to indicate
+# getter and setter methods for a property. Setting this option to YES will make
+# doxygen to replace the get and set methods by a property in the documentation.
+# This will only work if the methods are indeed getting or setting a simple
+# type. If this is not the case, or you want to show the methods anyway, you
+# should set this option to NO.
+# The default value is: YES.
+
+IDL_PROPERTY_SUPPORT   = YES
+
+# If member grouping is used in the documentation and the DISTRIBUTE_GROUP_DOC
+# tag is set to YES then doxygen will reuse the documentation of the first
+# member in the group (if any) for the other members of the group. By default
+# all members of a group must be documented explicitly.
+# The default value is: NO.
+
+DISTRIBUTE_GROUP_DOC   = NO
+
+# If one adds a struct or class to a group and this option is enabled, then also
+# any nested class or struct is added to the same group. By default this option
+# is disabled and one has to add nested compounds explicitly via \ingroup.
+# The default value is: NO.
+
+GROUP_NESTED_COMPOUNDS = NO
+
+# Set the SUBGROUPING tag to YES to allow class member groups of the same type
+# (for instance a group of public functions) to be put as a subgroup of that
+# type (e.g. under the Public Functions section). Set it to NO to prevent
+# subgrouping. Alternatively, this can be done per class using the
+# \nosubgrouping command.
+# The default value is: YES.
+
+SUBGROUPING            = YES
+
+# When the INLINE_GROUPED_CLASSES tag is set to YES, classes, structs and unions
+# are shown inside the group in which they are included (e.g. using \ingroup)
+# instead of on a separate page (for HTML and Man pages) or section (for LaTeX
+# and RTF).
+#
+# Note that this feature does not work in combination with
+# SEPARATE_MEMBER_PAGES.
+# The default value is: NO.
+
+INLINE_GROUPED_CLASSES = NO
+
+# When the INLINE_SIMPLE_STRUCTS tag is set to YES, structs, classes, and unions
+# with only public data fields or simple typedef fields will be shown inline in
+# the documentation of the scope in which they are defined (i.e. file,
+# namespace, or group documentation), provided this scope is documented. If set
+# to NO, structs, classes, and unions are shown on a separate page (for HTML and
+# Man pages) or section (for LaTeX and RTF).
+# The default value is: NO.
+
+INLINE_SIMPLE_STRUCTS  = NO
+
+# When TYPEDEF_HIDES_STRUCT tag is enabled, a typedef of a struct, union, or
+# enum is documented as struct, union, or enum with the name of the typedef. So
+# typedef struct TypeS {} TypeT, will appear in the documentation as a struct
+# with name TypeT. When disabled the typedef will appear as a member of a file,
+# namespace, or class. And the struct will be named TypeS. This can typically be
+# useful for C code in case the coding convention dictates that all compound
+# types are typedef'ed and only the typedef is referenced, never the tag name.
+# The default value is: NO.
+
+TYPEDEF_HIDES_STRUCT   = NO
+
+# The size of the symbol lookup cache can be set using LOOKUP_CACHE_SIZE. This
+# cache is used to resolve symbols given their name and scope. Since this can be
+# an expensive process and often the same symbol appears multiple times in the
+# code, doxygen keeps a cache of pre-resolved symbols. If the cache is too small
+# doxygen will become slower. If the cache is too large, memory is wasted. The
+# cache size is given by this formula: 2^(16+LOOKUP_CACHE_SIZE). The valid range
+# is 0..9, the default is 0, corresponding to a cache size of 2^16=65536
+# symbols. At the end of a run doxygen will report the cache usage and suggest
+# the optimal cache size from a speed point of view.
+# Minimum value: 0, maximum value: 9, default value: 0.
+
+LOOKUP_CACHE_SIZE      = 0
+
+#---------------------------------------------------------------------------
+# Build related configuration options
+#---------------------------------------------------------------------------
+
+# If the EXTRACT_ALL tag is set to YES, doxygen will assume all entities in
+# documentation are documented, even if no documentation was available. Private
+# class members and static file members will be hidden unless the
+# EXTRACT_PRIVATE respectively EXTRACT_STATIC tags are set to YES.
+# Note: This will also disable the warnings about undocumented members that are
+# normally produced when WARNINGS is set to YES.
+# The default value is: NO.
+
+EXTRACT_ALL            = YES
+
+# If the EXTRACT_PRIVATE tag is set to YES, all private members of a class will
+# be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PRIVATE        = YES
+
+# If the EXTRACT_PRIV_VIRTUAL tag is set to YES, documented private virtual
+# methods of a class will be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PRIV_VIRTUAL   = YES
+
+# If the EXTRACT_PACKAGE tag is set to YES, all members with package or internal
+# scope will be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PACKAGE        = YES
+
+# If the EXTRACT_STATIC tag is set to YES, all static members of a file will be
+# included in the documentation.
+# The default value is: NO.
+
+EXTRACT_STATIC         = YES
+
+# If the EXTRACT_LOCAL_CLASSES tag is set to YES, classes (and structs) defined
+# locally in source files will be included in the documentation. If set to NO,
+# only classes defined in header files are included. Does not have any effect
+# for Java sources.
+# The default value is: YES.
+
+EXTRACT_LOCAL_CLASSES  = YES
+
+# This flag is only useful for Objective-C code. If set to YES, local methods,
+# which are defined in the implementation section but not in the interface are
+# included in the documentation. If set to NO, only methods in the interface are
+# included.
+# The default value is: NO.
+
+EXTRACT_LOCAL_METHODS  = YES
+
+# If this flag is set to YES, the members of anonymous namespaces will be
+# extracted and appear in the documentation as a namespace called
+# 'anonymous_namespace{file}', where file will be replaced with the base name of
+# the file that contains the anonymous namespace. By default anonymous namespace
+# are hidden.
+# The default value is: NO.
+
+EXTRACT_ANON_NSPACES   = NO
+
+# If the HIDE_UNDOC_MEMBERS tag is set to YES, doxygen will hide all
+# undocumented members inside documented classes or files. If set to NO these
+# members will be included in the various overviews, but no documentation
+# section is generated. This option has no effect if EXTRACT_ALL is enabled.
+# The default value is: NO.
+
+HIDE_UNDOC_MEMBERS     = NO
+
+# If the HIDE_UNDOC_CLASSES tag is set to YES, doxygen will hide all
+# undocumented classes that are normally visible in the class hierarchy. If set
+# to NO, these classes will be included in the various overviews. This option
+# has no effect if EXTRACT_ALL is enabled.
+# The default value is: NO.
+
+HIDE_UNDOC_CLASSES     = NO
+
+# If the HIDE_FRIEND_COMPOUNDS tag is set to YES, doxygen will hide all friend
+# declarations. If set to NO, these declarations will be included in the
+# documentation.
+# The default value is: NO.
+
+HIDE_FRIEND_COMPOUNDS  = NO
+
+# If the HIDE_IN_BODY_DOCS tag is set to YES, doxygen will hide any
+# documentation blocks found inside the body of a function. If set to NO, these
+# blocks will be appended to the function's detailed documentation block.
+# The default value is: NO.
+
+HIDE_IN_BODY_DOCS      = NO
+
+# The INTERNAL_DOCS tag determines if documentation that is typed after a
+# \internal command is included. If the tag is set to NO then the documentation
+# will be excluded. Set it to YES to include the internal documentation.
+# The default value is: NO.
+
+INTERNAL_DOCS          = NO
+
+# If the CASE_SENSE_NAMES tag is set to NO then doxygen will only generate file
+# names in lower-case letters. If set to YES, upper-case letters are also
+# allowed. This is useful if you have classes or files whose names only differ
+# in case and if your file system supports case sensitive file names. Windows
+# (including Cygwin) ands Mac users are advised to set this option to NO.
+# The default value is: system dependent.
+
+CASE_SENSE_NAMES       = YES
+
+# If the HIDE_SCOPE_NAMES tag is set to NO then doxygen will show members with
+# their full class and namespace scopes in the documentation. If set to YES, the
+# scope will be hidden.
+# The default value is: NO.
+
+HIDE_SCOPE_NAMES       = NO
+
+# If the HIDE_COMPOUND_REFERENCE tag is set to NO (default) then doxygen will
+# append additional text to a page's title, such as Class Reference. If set to
+# YES the compound reference will be hidden.
+# The default value is: NO.
+
+HIDE_COMPOUND_REFERENCE= NO
+
+# If the SHOW_INCLUDE_FILES tag is set to YES then doxygen will put a list of
+# the files that are included by a file in the documentation of that file.
+# The default value is: YES.
+
+SHOW_INCLUDE_FILES     = YES
+
+# If the SHOW_GROUPED_MEMB_INC tag is set to YES then Doxygen will add for each
+# grouped member an include statement to the documentation, telling the reader
+# which file to include in order to use the member.
+# The default value is: NO.
+
+SHOW_GROUPED_MEMB_INC  = NO
+
+# If the FORCE_LOCAL_INCLUDES tag is set to YES then doxygen will list include
+# files with double quotes in the documentation rather than with sharp brackets.
+# The default value is: NO.
+
+FORCE_LOCAL_INCLUDES   = NO
+
+# If the INLINE_INFO tag is set to YES then a tag [inline] is inserted in the
+# documentation for inline members.
+# The default value is: YES.
+
+INLINE_INFO            = YES
+
+# If the SORT_MEMBER_DOCS tag is set to YES then doxygen will sort the
+# (detailed) documentation of file and class members alphabetically by member
+# name. If set to NO, the members will appear in declaration order.
+# The default value is: YES.
+
+SORT_MEMBER_DOCS       = YES
+
+# If the SORT_BRIEF_DOCS tag is set to YES then doxygen will sort the brief
+# descriptions of file, namespace and class members alphabetically by member
+# name. If set to NO, the members will appear in declaration order. Note that
+# this will also influence the order of the classes in the class list.
+# The default value is: NO.
+
+SORT_BRIEF_DOCS        = NO
+
+# If the SORT_MEMBERS_CTORS_1ST tag is set to YES then doxygen will sort the
+# (brief and detailed) documentation of class members so that constructors and
+# destructors are listed first. If set to NO the constructors will appear in the
+# respective orders defined by SORT_BRIEF_DOCS and SORT_MEMBER_DOCS.
+# Note: If SORT_BRIEF_DOCS is set to NO this option is ignored for sorting brief
+# member documentation.
+# Note: If SORT_MEMBER_DOCS is set to NO this option is ignored for sorting
+# detailed member documentation.
+# The default value is: NO.
+
+SORT_MEMBERS_CTORS_1ST = NO
+
+# If the SORT_GROUP_NAMES tag is set to YES then doxygen will sort the hierarchy
+# of group names into alphabetical order. If set to NO the group names will
+# appear in their defined order.
+# The default value is: NO.
+
+SORT_GROUP_NAMES       = NO
+
+# If the SORT_BY_SCOPE_NAME tag is set to YES, the class list will be sorted by
+# fully-qualified names, including namespaces. If set to NO, the class list will
+# be sorted only by class name, not including the namespace part.
+# Note: This option is not very useful if HIDE_SCOPE_NAMES is set to YES.
+# Note: This option applies only to the class list, not to the alphabetical
+# list.
+# The default value is: NO.
+
+SORT_BY_SCOPE_NAME     = NO
+
+# If the STRICT_PROTO_MATCHING option is enabled and doxygen fails to do proper
+# type resolution of all parameters of a function it will reject a match between
+# the prototype and the implementation of a member function even if there is
+# only one candidate or it is obvious which candidate to choose by doing a
+# simple string match. By disabling STRICT_PROTO_MATCHING doxygen will still
+# accept a match between prototype and implementation in such cases.
+# The default value is: NO.
+
+STRICT_PROTO_MATCHING  = NO
+
+# The GENERATE_TODOLIST tag can be used to enable (YES) or disable (NO) the todo
+# list. This list is created by putting \todo commands in the documentation.
+# The default value is: YES.
+
+GENERATE_TODOLIST      = YES
+
+# The GENERATE_TESTLIST tag can be used to enable (YES) or disable (NO) the test
+# list. This list is created by putting \test commands in the documentation.
+# The default value is: YES.
+
+GENERATE_TESTLIST      = YES
+
+# The GENERATE_BUGLIST tag can be used to enable (YES) or disable (NO) the bug
+# list. This list is created by putting \bug commands in the documentation.
+# The default value is: YES.
+
+GENERATE_BUGLIST       = YES
+
+# The GENERATE_DEPRECATEDLIST tag can be used to enable (YES) or disable (NO)
+# the deprecated list. This list is created by putting \deprecated commands in
+# the documentation.
+# The default value is: YES.
+
+GENERATE_DEPRECATEDLIST= YES
+
+# The ENABLED_SECTIONS tag can be used to enable conditional documentation
+# sections, marked by \if  ... \endif and \cond 
+# ... \endcond blocks.
+
+ENABLED_SECTIONS       =
+
+# The MAX_INITIALIZER_LINES tag determines the maximum number of lines that the
+# initial value of a variable or macro / define can have for it to appear in the
+# documentation. If the initializer consists of more lines than specified here
+# it will be hidden. Use a value of 0 to hide initializers completely. The
+# appearance of the value of individual variables and macros / defines can be
+# controlled using \showinitializer or \hideinitializer command in the
+# documentation regardless of this setting.
+# Minimum value: 0, maximum value: 10000, default value: 30.
+
+MAX_INITIALIZER_LINES  = 30
+
+# Set the SHOW_USED_FILES tag to NO to disable the list of files generated at
+# the bottom of the documentation of classes and structs. If set to YES, the
+# list will mention the files that were used to generate the documentation.
+# The default value is: YES.
+
+SHOW_USED_FILES        = YES
+
+# Set the SHOW_FILES tag to NO to disable the generation of the Files page. This
+# will remove the Files entry from the Quick Index and from the Folder Tree View
+# (if specified).
+# The default value is: YES.
+
+SHOW_FILES             = YES
+
+# Set the SHOW_NAMESPACES tag to NO to disable the generation of the Namespaces
+# page. This will remove the Namespaces entry from the Quick Index and from the
+# Folder Tree View (if specified).
+# The default value is: YES.
+
+SHOW_NAMESPACES        = YES
+
+# The FILE_VERSION_FILTER tag can be used to specify a program or script that
+# doxygen should invoke to get the current version for each file (typically from
+# the version control system). Doxygen will invoke the program by executing (via
+# popen()) the command command input-file, where command is the value of the
+# FILE_VERSION_FILTER tag, and input-file is the name of an input file provided
+# by doxygen. Whatever the program writes to standard output is used as the file
+# version. For an example see the documentation.
+
+FILE_VERSION_FILTER    =
+
+# The LAYOUT_FILE tag can be used to specify a layout file which will be parsed
+# by doxygen. The layout file controls the global structure of the generated
+# output files in an output format independent way. To create the layout file
+# that represents doxygen's defaults, run doxygen with the -l option. You can
+# optionally specify a file name after the option, if omitted DoxygenLayout.xml
+# will be used as the name of the layout file.
+#
+# Note that if you run doxygen from a directory containing a file called
+# DoxygenLayout.xml, doxygen will parse it automatically even if the LAYOUT_FILE
+# tag is left empty.
+
+LAYOUT_FILE            =
+
+# The CITE_BIB_FILES tag can be used to specify one or more bib files containing
+# the reference definitions. This must be a list of .bib files. The .bib
+# extension is automatically appended if omitted. This requires the bibtex tool
+# to be installed. See also https://en.wikipedia.org/wiki/BibTeX for more info.
+# For LaTeX the style of the bibliography can be controlled using
+# LATEX_BIB_STYLE. To use this feature you need bibtex and perl available in the
+# search path. See also \cite for info how to create references.
+
+CITE_BIB_FILES         =
+
+#---------------------------------------------------------------------------
+# Configuration options related to warning and progress messages
+#---------------------------------------------------------------------------
+
+# The QUIET tag can be used to turn on/off the messages that are generated to
+# standard output by doxygen. If QUIET is set to YES this implies that the
+# messages are off.
+# The default value is: NO.
+
+QUIET                  = NO
+
+# The WARNINGS tag can be used to turn on/off the warning messages that are
+# generated to standard error (stderr) by doxygen. If WARNINGS is set to YES
+# this implies that the warnings are on.
+#
+# Tip: Turn warnings on while writing the documentation.
+# The default value is: YES.
+
+WARNINGS               = YES
+
+# If the WARN_IF_UNDOCUMENTED tag is set to YES then doxygen will generate
+# warnings for undocumented members. If EXTRACT_ALL is set to YES then this flag
+# will automatically be disabled.
+# The default value is: YES.
+
+WARN_IF_UNDOCUMENTED   = YES
+
+# If the WARN_IF_DOC_ERROR tag is set to YES, doxygen will generate warnings for
+# potential errors in the documentation, such as not documenting some parameters
+# in a documented function, or documenting parameters that don't exist or using
+# markup commands wrongly.
+# The default value is: YES.
+
+WARN_IF_DOC_ERROR      = YES
+
+# This WARN_NO_PARAMDOC option can be enabled to get warnings for functions that
+# are documented, but have no documentation for their parameters or return
+# value. If set to NO, doxygen will only warn about wrong or incomplete
+# parameter documentation, but not about the absence of documentation. If
+# EXTRACT_ALL is set to YES then this flag will automatically be disabled.
+# The default value is: NO.
+
+WARN_NO_PARAMDOC       = NO
+
+# If the WARN_AS_ERROR tag is set to YES then doxygen will immediately stop when
+# a warning is encountered.
+# The default value is: NO.
+
+WARN_AS_ERROR          = NO
+
+# The WARN_FORMAT tag determines the format of the warning messages that doxygen
+# can produce. The string should contain the $file, $line, and $text tags, which
+# will be replaced by the file and line number from which the warning originated
+# and the warning text. Optionally the format may contain $version, which will
+# be replaced by the version of the file (if it could be obtained via
+# FILE_VERSION_FILTER)
+# The default value is: $file:$line: $text.
+
+WARN_FORMAT            = "$file:$line: $text"
+
+# The WARN_LOGFILE tag can be used to specify a file to which warning and error
+# messages should be written. If left blank the output is written to standard
+# error (stderr).
+
+WARN_LOGFILE           =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the input files
+#---------------------------------------------------------------------------
+
+# The INPUT tag is used to specify the files and/or directories that contain
+# documented source files. You may enter file names like myfile.cpp or
+# directories like /usr/src/myproject. Separate the files or directories with
+# spaces. See also FILE_PATTERNS and EXTENSION_MAPPING
+# Note: If this tag is empty the current directory is searched.
+
+INPUT                  =
+
+# This tag can be used to specify the character encoding of the source files
+# that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses
+# libiconv (or the iconv built into libc) for the transcoding. See the libiconv
+# documentation (see: https://www.gnu.org/software/libiconv/) for the list of
+# possible encodings.
+# The default value is: UTF-8.
+
+INPUT_ENCODING         = UTF-8
+
+# If the value of the INPUT tag contains directories, you can use the
+# FILE_PATTERNS tag to specify one or more wildcard patterns (like *.cpp and
+# *.h) to filter out the source-files in the directories.
+#
+# Note that for custom extensions or not directly supported extensions you also
+# need to set EXTENSION_MAPPING for the extension otherwise the files are not
+# read by doxygen.
+#
+# If left blank the following patterns are tested:*.c, *.cc, *.cxx, *.cpp,
+# *.c++, *.java, *.ii, *.ixx, *.ipp, *.i++, *.inl, *.idl, *.ddl, *.odl, *.h,
+# *.hh, *.hxx, *.hpp, *.h++, *.cs, *.d, *.php, *.php4, *.php5, *.phtml, *.inc,
+# *.m, *.markdown, *.md, *.mm, *.dox (to be provided as doxygen C comment),
+# *.doc (to be provided as doxygen C comment), *.txt (to be provided as doxygen
+# C comment), *.py, *.pyw, *.f90, *.f95, *.f03, *.f08, *.f, *.for, *.tcl, *.vhd,
+# *.vhdl, *.ucf, *.qsf and *.ice.
+
+FILE_PATTERNS          = *.c \
+                         *.cc \
+                         *.cxx \
+                         *.cpp \
+                         *.c++ \
+                         *.java \
+                         *.ii \
+                         *.ixx \
+                         *.ipp \
+                         *.i++ \
+                         *.inl \
+                         *.idl \
+                         *.ddl \
+                         *.odl \
+                         *.h \
+                         *.hh \
+                         *.hxx \
+                         *.hpp \
+                         *.h++ \
+                         *.cs \
+                         *.d \
+                         *.php \
+                         *.php4 \
+                         *.php5 \
+                         *.phtml \
+                         *.inc \
+                         *.m \
+                         *.markdown \
+                         *.md \
+                         *.mm \
+                         *.dox \
+                         *.doc \
+                         *.txt \
+                         *.py \
+                         *.pyw \
+                         *.f90 \
+                         *.f95 \
+                         *.f03 \
+                         *.f08 \
+                         *.f \
+                         *.for \
+                         *.tcl \
+                         *.vhd \
+                         *.vhdl \
+                         *.ucf \
+                         *.qsf \
+                         *.ice
+
+# The RECURSIVE tag can be used to specify whether or not subdirectories should
+# be searched for input files as well.
+# The default value is: NO.
+
+RECURSIVE              = YES
+
+# The EXCLUDE tag can be used to specify files and/or directories that should be
+# excluded from the INPUT source files. This way you can easily exclude a
+# subdirectory from a directory tree whose root is specified with the INPUT tag.
+#
+# Note that relative paths are relative to the directory from which doxygen is
+# run.
+
+EXCLUDE                =
+
+# The EXCLUDE_SYMLINKS tag can be used to select whether or not files or
+# directories that are symbolic links (a Unix file system feature) are excluded
+# from the input.
+# The default value is: NO.
+
+EXCLUDE_SYMLINKS       = NO
+
+# If the value of the INPUT tag contains directories, you can use the
+# EXCLUDE_PATTERNS tag to specify one or more wildcard patterns to exclude
+# certain files from those directories.
+#
+# Note that the wildcards are matched against the file with absolute path, so to
+# exclude all test directories for example use the pattern */test/*
+
+EXCLUDE_PATTERNS       =
+
+# The EXCLUDE_SYMBOLS tag can be used to specify one or more symbol names
+# (namespaces, classes, functions, etc.) that should be excluded from the
+# output. The symbol name can be a fully qualified name, a word, or if the
+# wildcard * is used, a substring. Examples: ANamespace, AClass,
+# AClass::ANamespace, ANamespace::*Test
+#
+# Note that the wildcards are matched against the file with absolute path, so to
+# exclude all test directories use the pattern */test/*
+
+EXCLUDE_SYMBOLS        =
+
+# The EXAMPLE_PATH tag can be used to specify one or more files or directories
+# that contain example code fragments that are included (see the \include
+# command).
+
+EXAMPLE_PATH           =
+
+# If the value of the EXAMPLE_PATH tag contains directories, you can use the
+# EXAMPLE_PATTERNS tag to specify one or more wildcard pattern (like *.cpp and
+# *.h) to filter out the source-files in the directories. If left blank all
+# files are included.
+
+EXAMPLE_PATTERNS       = *
+
+# If the EXAMPLE_RECURSIVE tag is set to YES then subdirectories will be
+# searched for input files to be used with the \include or \dontinclude commands
+# irrespective of the value of the RECURSIVE tag.
+# The default value is: NO.
+
+EXAMPLE_RECURSIVE      = NO
+
+# The IMAGE_PATH tag can be used to specify one or more files or directories
+# that contain images that are to be included in the documentation (see the
+# \image command).
+
+IMAGE_PATH             =
+
+# The INPUT_FILTER tag can be used to specify a program that doxygen should
+# invoke to filter for each input file. Doxygen will invoke the filter program
+# by executing (via popen()) the command:
+#
+#  
+#
+# where  is the value of the INPUT_FILTER tag, and  is the
+# name of an input file. Doxygen will then use the output that the filter
+# program writes to standard output. If FILTER_PATTERNS is specified, this tag
+# will be ignored.
+#
+# Note that the filter must not add or remove lines; it is applied before the
+# code is scanned, but not when the output code is generated. If lines are added
+# or removed, the anchors will not be placed correctly.
+#
+# Note that for custom extensions or not directly supported extensions you also
+# need to set EXTENSION_MAPPING for the extension otherwise the files are not
+# properly processed by doxygen.
+
+INPUT_FILTER           =
+
+# The FILTER_PATTERNS tag can be used to specify filters on a per file pattern
+# basis. Doxygen will compare the file name with each pattern and apply the
+# filter if there is a match. The filters are a list of the form: pattern=filter
+# (like *.cpp=my_cpp_filter). See INPUT_FILTER for further information on how
+# filters are used. If the FILTER_PATTERNS tag is empty or if none of the
+# patterns match the file name, INPUT_FILTER is applied.
+#
+# Note that for custom extensions or not directly supported extensions you also
+# need to set EXTENSION_MAPPING for the extension otherwise the files are not
+# properly processed by doxygen.
+
+FILTER_PATTERNS        =
+
+# If the FILTER_SOURCE_FILES tag is set to YES, the input filter (if set using
+# INPUT_FILTER) will also be used to filter the input files that are used for
+# producing the source files to browse (i.e. when SOURCE_BROWSER is set to YES).
+# The default value is: NO.
+
+FILTER_SOURCE_FILES    = NO
+
+# The FILTER_SOURCE_PATTERNS tag can be used to specify source filters per file
+# pattern. A pattern will override the setting for FILTER_PATTERN (if any) and
+# it is also possible to disable source filtering for a specific pattern using
+# *.ext= (so without naming a filter).
+# This tag requires that the tag FILTER_SOURCE_FILES is set to YES.
+
+FILTER_SOURCE_PATTERNS =
+
+# If the USE_MDFILE_AS_MAINPAGE tag refers to the name of a markdown file that
+# is part of the input, its contents will be placed on the main page
+# (index.html). This can be useful if you have a project on for instance GitHub
+# and want to reuse the introduction page also for the doxygen output.
+
+USE_MDFILE_AS_MAINPAGE =
+
+#---------------------------------------------------------------------------
+# Configuration options related to source browsing
+#---------------------------------------------------------------------------
+
+# If the SOURCE_BROWSER tag is set to YES then a list of source files will be
+# generated. Documented entities will be cross-referenced with these sources.
+#
+# Note: To get rid of all source code in the generated output, make sure that
+# also VERBATIM_HEADERS is set to NO.
+# The default value is: NO.
+
+SOURCE_BROWSER         = NO
+
+# Setting the INLINE_SOURCES tag to YES will include the body of functions,
+# classes and enums directly into the documentation.
+# The default value is: NO.
+
+INLINE_SOURCES         = NO
+
+# Setting the STRIP_CODE_COMMENTS tag to YES will instruct doxygen to hide any
+# special comment blocks from generated source code fragments. Normal C, C++ and
+# Fortran comments will always remain visible.
+# The default value is: YES.
+
+STRIP_CODE_COMMENTS    = YES
+
+# If the REFERENCED_BY_RELATION tag is set to YES then for each documented
+# entity all documented functions referencing it will be listed.
+# The default value is: NO.
+
+REFERENCED_BY_RELATION = NO
+
+# If the REFERENCES_RELATION tag is set to YES then for each documented function
+# all documented entities called/used by that function will be listed.
+# The default value is: NO.
+
+REFERENCES_RELATION    = NO
+
+# If the REFERENCES_LINK_SOURCE tag is set to YES and SOURCE_BROWSER tag is set
+# to YES then the hyperlinks from functions in REFERENCES_RELATION and
+# REFERENCED_BY_RELATION lists will link to the source code. Otherwise they will
+# link to the documentation.
+# The default value is: YES.
+
+REFERENCES_LINK_SOURCE = YES
+
+# If SOURCE_TOOLTIPS is enabled (the default) then hovering a hyperlink in the
+# source code will show a tooltip with additional information such as prototype,
+# brief description and links to the definition and documentation. Since this
+# will make the HTML file larger and loading of large files a bit slower, you
+# can opt to disable this feature.
+# The default value is: YES.
+# This tag requires that the tag SOURCE_BROWSER is set to YES.
+
+SOURCE_TOOLTIPS        = YES
+
+# If the USE_HTAGS tag is set to YES then the references to source code will
+# point to the HTML generated by the htags(1) tool instead of doxygen built-in
+# source browser. The htags tool is part of GNU's global source tagging system
+# (see https://www.gnu.org/software/global/global.html). You will need version
+# 4.8.6 or higher.
+#
+# To use it do the following:
+# - Install the latest version of global
+# - Enable SOURCE_BROWSER and USE_HTAGS in the configuration file
+# - Make sure the INPUT points to the root of the source tree
+# - Run doxygen as normal
+#
+# Doxygen will invoke htags (and that will in turn invoke gtags), so these
+# tools must be available from the command line (i.e. in the search path).
+#
+# The result: instead of the source browser generated by doxygen, the links to
+# source code will now point to the output of htags.
+# The default value is: NO.
+# This tag requires that the tag SOURCE_BROWSER is set to YES.
+
+USE_HTAGS              = NO
+
+# If the VERBATIM_HEADERS tag is set the YES then doxygen will generate a
+# verbatim copy of the header file for each class for which an include is
+# specified. Set to NO to disable this.
+# See also: Section \class.
+# The default value is: YES.
+
+VERBATIM_HEADERS       = YES
+
+# If the CLANG_ASSISTED_PARSING tag is set to YES then doxygen will use the
+# clang parser (see: http://clang.llvm.org/) for more accurate parsing at the
+# cost of reduced performance. This can be particularly helpful with template
+# rich C++ code for which doxygen's built-in parser lacks the necessary type
+# information.
+# Note: The availability of this option depends on whether or not doxygen was
+# generated with the -Duse_libclang=ON option for CMake.
+# The default value is: NO.
+
+CLANG_ASSISTED_PARSING = NO
+
+# If clang assisted parsing is enabled you can provide the compiler with command
+# line options that you would normally use when invoking the compiler. Note that
+# the include paths will already be set by doxygen for the files and directories
+# specified with INPUT and INCLUDE_PATH.
+# This tag requires that the tag CLANG_ASSISTED_PARSING is set to YES.
+
+CLANG_OPTIONS          =
+
+# If clang assisted parsing is enabled you can provide the clang parser with the
+# path to the compilation database (see:
+# http://clang.llvm.org/docs/HowToSetupToolingForLLVM.html) used when the files
+# were built. This is equivalent to specifying the "-p" option to a clang tool,
+# such as clang-check. These options will then be passed to the parser.
+# Note: The availability of this option depends on whether or not doxygen was
+# generated with the -Duse_libclang=ON option for CMake.
+
+CLANG_DATABASE_PATH    =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the alphabetical class index
+#---------------------------------------------------------------------------
+
+# If the ALPHABETICAL_INDEX tag is set to YES, an alphabetical index of all
+# compounds will be generated. Enable this if the project contains a lot of
+# classes, structs, unions or interfaces.
+# The default value is: YES.
+
+ALPHABETICAL_INDEX     = YES
+
+# The COLS_IN_ALPHA_INDEX tag can be used to specify the number of columns in
+# which the alphabetical index list will be split.
+# Minimum value: 1, maximum value: 20, default value: 5.
+# This tag requires that the tag ALPHABETICAL_INDEX is set to YES.
+
+COLS_IN_ALPHA_INDEX    = 5
+
+# In case all classes in a project start with a common prefix, all classes will
+# be put under the same header in the alphabetical index. The IGNORE_PREFIX tag
+# can be used to specify a prefix (or a list of prefixes) that should be ignored
+# while generating the index headers.
+# This tag requires that the tag ALPHABETICAL_INDEX is set to YES.
+
+IGNORE_PREFIX          =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the HTML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_HTML tag is set to YES, doxygen will generate HTML output
+# The default value is: YES.
+
+GENERATE_HTML          = YES
+
+# The HTML_OUTPUT tag is used to specify where the HTML docs will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: html.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_OUTPUT            = html
+
+# The HTML_FILE_EXTENSION tag can be used to specify the file extension for each
+# generated HTML page (for example: .htm, .php, .asp).
+# The default value is: .html.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_FILE_EXTENSION    = .html
+
+# The HTML_HEADER tag can be used to specify a user-defined HTML header file for
+# each generated HTML page. If the tag is left blank doxygen will generate a
+# standard header.
+#
+# To get valid HTML the header file that includes any scripts and style sheets
+# that doxygen needs, which is dependent on the configuration options used (e.g.
+# the setting GENERATE_TREEVIEW). It is highly recommended to start with a
+# default header using
+# doxygen -w html new_header.html new_footer.html new_stylesheet.css
+# YourConfigFile
+# and then modify the file new_header.html. See also section "Doxygen usage"
+# for information on how to generate the default header that doxygen normally
+# uses.
+# Note: The header is subject to change so you typically have to regenerate the
+# default header when upgrading to a newer version of doxygen. For a description
+# of the possible markers and block names see the documentation.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_HEADER            =
+
+# The HTML_FOOTER tag can be used to specify a user-defined HTML footer for each
+# generated HTML page. If the tag is left blank doxygen will generate a standard
+# footer. See HTML_HEADER for more information on how to generate a default
+# footer and what special commands can be used inside the footer. See also
+# section "Doxygen usage" for information on how to generate the default footer
+# that doxygen normally uses.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_FOOTER            =
+
+# The HTML_STYLESHEET tag can be used to specify a user-defined cascading style
+# sheet that is used by each HTML page. It can be used to fine-tune the look of
+# the HTML output. If left blank doxygen will generate a default style sheet.
+# See also section "Doxygen usage" for information on how to generate the style
+# sheet that doxygen normally uses.
+# Note: It is recommended to use HTML_EXTRA_STYLESHEET instead of this tag, as
+# it is more robust and this tag (HTML_STYLESHEET) will in the future become
+# obsolete.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_STYLESHEET        =
+
+# The HTML_EXTRA_STYLESHEET tag can be used to specify additional user-defined
+# cascading style sheets that are included after the standard style sheets
+# created by doxygen. Using this option one can overrule certain style aspects.
+# This is preferred over using HTML_STYLESHEET since it does not replace the
+# standard style sheet and is therefore more robust against future updates.
+# Doxygen will copy the style sheet files to the output directory.
+# Note: The order of the extra style sheet files is of importance (e.g. the last
+# style sheet in the list overrules the setting of the previous ones in the
+# list). For an example see the documentation.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_EXTRA_STYLESHEET  =
+
+# The HTML_EXTRA_FILES tag can be used to specify one or more extra images or
+# other source files which should be copied to the HTML output directory. Note
+# that these files will be copied to the base HTML output directory. Use the
+# $relpath^ marker in the HTML_HEADER and/or HTML_FOOTER files to load these
+# files. In the HTML_STYLESHEET file, use the file name only. Also note that the
+# files will be copied as-is; there are no commands or markers available.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_EXTRA_FILES       =
+
+# The HTML_COLORSTYLE_HUE tag controls the color of the HTML output. Doxygen
+# will adjust the colors in the style sheet and background images according to
+# this color. Hue is specified as an angle on a colorwheel, see
+# https://en.wikipedia.org/wiki/Hue for more information. For instance the value
+# 0 represents red, 60 is yellow, 120 is green, 180 is cyan, 240 is blue, 300
+# purple, and 360 is red again.
+# Minimum value: 0, maximum value: 359, default value: 220.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_HUE    = 220
+
+# The HTML_COLORSTYLE_SAT tag controls the purity (or saturation) of the colors
+# in the HTML output. For a value of 0 the output will use grayscales only. A
+# value of 255 will produce the most vivid colors.
+# Minimum value: 0, maximum value: 255, default value: 100.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_SAT    = 100
+
+# The HTML_COLORSTYLE_GAMMA tag controls the gamma correction applied to the
+# luminance component of the colors in the HTML output. Values below 100
+# gradually make the output lighter, whereas values above 100 make the output
+# darker. The value divided by 100 is the actual gamma applied, so 80 represents
+# a gamma of 0.8, The value 220 represents a gamma of 2.2, and 100 does not
+# change the gamma.
+# Minimum value: 40, maximum value: 240, default value: 80.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_GAMMA  = 80
+
+# If the HTML_TIMESTAMP tag is set to YES then the footer of each generated HTML
+# page will contain the date and time when the page was generated. Setting this
+# to YES can help to show when doxygen was last run and thus if the
+# documentation is up to date.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_TIMESTAMP         = NO
+
+# If the HTML_DYNAMIC_MENUS tag is set to YES then the generated HTML
+# documentation will contain a main index with vertical navigation menus that
+# are dynamically created via JavaScript. If disabled, the navigation index will
+# consists of multiple levels of tabs that are statically embedded in every HTML
+# page. Disable this option to support browsers that do not have JavaScript,
+# like the Qt help browser.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_DYNAMIC_MENUS     = YES
+
+# If the HTML_DYNAMIC_SECTIONS tag is set to YES then the generated HTML
+# documentation will contain sections that can be hidden and shown after the
+# page has loaded.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_DYNAMIC_SECTIONS  = NO
+
+# With HTML_INDEX_NUM_ENTRIES one can control the preferred number of entries
+# shown in the various tree structured indices initially; the user can expand
+# and collapse entries dynamically later on. Doxygen will expand the tree to
+# such a level that at most the specified number of entries are visible (unless
+# a fully collapsed tree already exceeds this amount). So setting the number of
+# entries 1 will produce a full collapsed tree by default. 0 is a special value
+# representing an infinite number of entries and will result in a full expanded
+# tree by default.
+# Minimum value: 0, maximum value: 9999, default value: 100.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_INDEX_NUM_ENTRIES = 100
+
+# If the GENERATE_DOCSET tag is set to YES, additional index files will be
+# generated that can be used as input for Apple's Xcode 3 integrated development
+# environment (see: https://developer.apple.com/xcode/), introduced with OSX
+# 10.5 (Leopard). To create a documentation set, doxygen will generate a
+# Makefile in the HTML output directory. Running make will produce the docset in
+# that directory and running make install will install the docset in
+# ~/Library/Developer/Shared/Documentation/DocSets so that Xcode will find it at
+# startup. See https://developer.apple.com/library/archive/featuredarticles/Doxy
+# genXcode/_index.html for more information.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_DOCSET        = NO
+
+# This tag determines the name of the docset feed. A documentation feed provides
+# an umbrella under which multiple documentation sets from a single provider
+# (such as a company or product suite) can be grouped.
+# The default value is: Doxygen generated docs.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_FEEDNAME        = "Doxygen generated docs"
+
+# This tag specifies a string that should uniquely identify the documentation
+# set bundle. This should be a reverse domain-name style string, e.g.
+# com.mycompany.MyDocSet. Doxygen will append .docset to the name.
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_BUNDLE_ID       = org.doxygen.Project
+
+# The DOCSET_PUBLISHER_ID tag specifies a string that should uniquely identify
+# the documentation publisher. This should be a reverse domain-name style
+# string, e.g. com.mycompany.MyDocSet.documentation.
+# The default value is: org.doxygen.Publisher.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_PUBLISHER_ID    = org.doxygen.Publisher
+
+# The DOCSET_PUBLISHER_NAME tag identifies the documentation publisher.
+# The default value is: Publisher.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_PUBLISHER_NAME  = Publisher
+
+# If the GENERATE_HTMLHELP tag is set to YES then doxygen generates three
+# additional HTML index files: index.hhp, index.hhc, and index.hhk. The
+# index.hhp is a project file that can be read by Microsoft's HTML Help Workshop
+# (see: https://www.microsoft.com/en-us/download/details.aspx?id=21138) on
+# Windows.
+#
+# The HTML Help Workshop contains a compiler that can convert all HTML output
+# generated by doxygen into a single compiled HTML file (.chm). Compiled HTML
+# files are now used as the Windows 98 help format, and will replace the old
+# Windows help format (.hlp) on all Windows platforms in the future. Compressed
+# HTML files also contain an index, a table of contents, and you can search for
+# words in the documentation. The HTML workshop also contains a viewer for
+# compressed HTML files.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_HTMLHELP      = NO
+
+# The CHM_FILE tag can be used to specify the file name of the resulting .chm
+# file. You can add a path in front of the file if the result should not be
+# written to the html output directory.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+CHM_FILE               =
+
+# The HHC_LOCATION tag can be used to specify the location (absolute path
+# including file name) of the HTML help compiler (hhc.exe). If non-empty,
+# doxygen will try to run the HTML help compiler on the generated index.hhp.
+# The file has to be specified with full path.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+HHC_LOCATION           =
+
+# The GENERATE_CHI flag controls if a separate .chi index file is generated
+# (YES) or that it should be included in the master .chm file (NO).
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+GENERATE_CHI           = NO
+
+# The CHM_INDEX_ENCODING is used to encode HtmlHelp index (hhk), content (hhc)
+# and project file content.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+CHM_INDEX_ENCODING     =
+
+# The BINARY_TOC flag controls whether a binary table of contents is generated
+# (YES) or a normal table of contents (NO) in the .chm file. Furthermore it
+# enables the Previous and Next buttons.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+BINARY_TOC             = NO
+
+# The TOC_EXPAND flag can be set to YES to add extra items for group members to
+# the table of contents of the HTML help documentation and to the tree view.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+TOC_EXPAND             = NO
+
+# If the GENERATE_QHP tag is set to YES and both QHP_NAMESPACE and
+# QHP_VIRTUAL_FOLDER are set, an additional index file will be generated that
+# can be used as input for Qt's qhelpgenerator to generate a Qt Compressed Help
+# (.qch) of the generated HTML documentation.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_QHP           = NO
+
+# If the QHG_LOCATION tag is specified, the QCH_FILE tag can be used to specify
+# the file name of the resulting .qch file. The path specified is relative to
+# the HTML output folder.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QCH_FILE               =
+
+# The QHP_NAMESPACE tag specifies the namespace to use when generating Qt Help
+# Project output. For more information please see Qt Help Project / Namespace
+# (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#namespace).
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_NAMESPACE          = org.doxygen.Project
+
+# The QHP_VIRTUAL_FOLDER tag specifies the namespace to use when generating Qt
+# Help Project output. For more information please see Qt Help Project / Virtual
+# Folders (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#virtual-
+# folders).
+# The default value is: doc.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_VIRTUAL_FOLDER     = doc
+
+# If the QHP_CUST_FILTER_NAME tag is set, it specifies the name of a custom
+# filter to add. For more information please see Qt Help Project / Custom
+# Filters (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#custom-
+# filters).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_CUST_FILTER_NAME   =
+
+# The QHP_CUST_FILTER_ATTRS tag specifies the list of the attributes of the
+# custom filter to add. For more information please see Qt Help Project / Custom
+# Filters (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#custom-
+# filters).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_CUST_FILTER_ATTRS  =
+
+# The QHP_SECT_FILTER_ATTRS tag specifies the list of the attributes this
+# project's filter section matches. Qt Help Project / Filter Attributes (see:
+# https://doc.qt.io/archives/qt-4.8/qthelpproject.html#filter-attributes).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_SECT_FILTER_ATTRS  =
+
+# The QHG_LOCATION tag can be used to specify the location of Qt's
+# qhelpgenerator. If non-empty doxygen will try to run qhelpgenerator on the
+# generated .qhp file.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHG_LOCATION           =
+
+# If the GENERATE_ECLIPSEHELP tag is set to YES, additional index files will be
+# generated, together with the HTML files, they form an Eclipse help plugin. To
+# install this plugin and make it available under the help contents menu in
+# Eclipse, the contents of the directory containing the HTML and XML files needs
+# to be copied into the plugins directory of eclipse. The name of the directory
+# within the plugins directory should be the same as the ECLIPSE_DOC_ID value.
+# After copying Eclipse needs to be restarted before the help appears.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_ECLIPSEHELP   = NO
+
+# A unique identifier for the Eclipse help plugin. When installing the plugin
+# the directory name containing the HTML and XML files should also have this
+# name. Each documentation set should have its own identifier.
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_ECLIPSEHELP is set to YES.
+
+ECLIPSE_DOC_ID         = org.doxygen.Project
+
+# If you want full control over the layout of the generated HTML pages it might
+# be necessary to disable the index and replace it with your own. The
+# DISABLE_INDEX tag can be used to turn on/off the condensed index (tabs) at top
+# of each HTML page. A value of NO enables the index and the value YES disables
+# it. Since the tabs in the index contain the same information as the navigation
+# tree, you can set this option to YES if you also set GENERATE_TREEVIEW to YES.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+DISABLE_INDEX          = NO
+
+# The GENERATE_TREEVIEW tag is used to specify whether a tree-like index
+# structure should be generated to display hierarchical information. If the tag
+# value is set to YES, a side panel will be generated containing a tree-like
+# index structure (just like the one that is generated for HTML Help). For this
+# to work a browser that supports JavaScript, DHTML, CSS and frames is required
+# (i.e. any modern browser). Windows users are probably better off using the
+# HTML help feature. Via custom style sheets (see HTML_EXTRA_STYLESHEET) one can
+# further fine-tune the look of the index. As an example, the default style
+# sheet generated by doxygen has an example that shows how to put an image at
+# the root of the tree instead of the PROJECT_NAME. Since the tree basically has
+# the same information as the tab index, you could consider setting
+# DISABLE_INDEX to YES when enabling this option.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_TREEVIEW      = NO
+
+# The ENUM_VALUES_PER_LINE tag can be used to set the number of enum values that
+# doxygen will group on one line in the generated HTML documentation.
+#
+# Note that a value of 0 will completely suppress the enum values from appearing
+# in the overview section.
+# Minimum value: 0, maximum value: 20, default value: 4.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+ENUM_VALUES_PER_LINE   = 4
+
+# If the treeview is enabled (see GENERATE_TREEVIEW) then this tag can be used
+# to set the initial width (in pixels) of the frame in which the tree is shown.
+# Minimum value: 0, maximum value: 1500, default value: 250.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+TREEVIEW_WIDTH         = 250
+
+# If the EXT_LINKS_IN_WINDOW option is set to YES, doxygen will open links to
+# external symbols imported via tag files in a separate window.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+EXT_LINKS_IN_WINDOW    = NO
+
+# Use this tag to change the font size of LaTeX formulas included as images in
+# the HTML documentation. When you change the font size after a successful
+# doxygen run you need to manually remove any form_*.png images from the HTML
+# output directory to force them to be regenerated.
+# Minimum value: 8, maximum value: 50, default value: 10.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+FORMULA_FONTSIZE       = 10
+
+# Use the FORMULA_TRANSPARENT tag to determine whether or not the images
+# generated for formulas are transparent PNGs. Transparent PNGs are not
+# supported properly for IE 6.0, but are supported on all modern browsers.
+#
+# Note that when changing this option you need to delete any form_*.png files in
+# the HTML output directory before the changes have effect.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+FORMULA_TRANSPARENT    = YES
+
+# The FORMULA_MACROFILE can contain LaTeX \newcommand and \renewcommand commands
+# to create new LaTeX commands to be used in formulas as building blocks. See
+# the section "Including formulas" for details.
+
+FORMULA_MACROFILE      =
+
+# Enable the USE_MATHJAX option to render LaTeX formulas using MathJax (see
+# https://www.mathjax.org) which uses client side JavaScript for the rendering
+# instead of using pre-rendered bitmaps. Use this if you do not have LaTeX
+# installed or if you want to formulas look prettier in the HTML output. When
+# enabled you may also need to install MathJax separately and configure the path
+# to it using the MATHJAX_RELPATH option.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+USE_MATHJAX            = YES
+
+# When MathJax is enabled you can set the default output format to be used for
+# the MathJax output. See the MathJax site (see:
+# http://docs.mathjax.org/en/latest/output.html) for more details.
+# Possible values are: HTML-CSS (which is slower, but has the best
+# compatibility), NativeMML (i.e. MathML) and SVG.
+# The default value is: HTML-CSS.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_FORMAT         = HTML-CSS
+
+# When MathJax is enabled you need to specify the location relative to the HTML
+# output directory using the MATHJAX_RELPATH option. The destination directory
+# should contain the MathJax.js script. For instance, if the mathjax directory
+# is located at the same level as the HTML output directory, then
+# MATHJAX_RELPATH should be ../mathjax. The default value points to the MathJax
+# Content Delivery Network so you can quickly see the result without installing
+# MathJax. However, it is strongly recommended to install a local copy of
+# MathJax from https://www.mathjax.org before deployment.
+# The default value is: https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.5/.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_RELPATH        = https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.5/
+
+# The MATHJAX_EXTENSIONS tag can be used to specify one or more MathJax
+# extension names that should be enabled during MathJax rendering. For example
+# MATHJAX_EXTENSIONS = TeX/AMSmath TeX/AMSsymbols
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_EXTENSIONS     =
+
+# The MATHJAX_CODEFILE tag can be used to specify a file with javascript pieces
+# of code that will be used on startup of the MathJax code. See the MathJax site
+# (see: http://docs.mathjax.org/en/latest/output.html) for more details. For an
+# example see the documentation.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_CODEFILE       =
+
+# When the SEARCHENGINE tag is enabled doxygen will generate a search box for
+# the HTML output. The underlying search engine uses javascript and DHTML and
+# should work on any modern browser. Note that when using HTML help
+# (GENERATE_HTMLHELP), Qt help (GENERATE_QHP), or docsets (GENERATE_DOCSET)
+# there is already a search function so this one should typically be disabled.
+# For large projects the javascript based search engine can be slow, then
+# enabling SERVER_BASED_SEARCH may provide a better solution. It is possible to
+# search using the keyboard; to jump to the search box use  + S
+# (what the  is depends on the OS and browser, but it is typically
+# , /