Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "2a2d3478afe8cdb336888f2e6faa3775ac40254e"
Unverified Commit 0ff42e84 authored by Michael Yang's avatar Michael Yang Committed by GitHub
Browse files

Merge pull request #4756 from ollama/mxyng/convert2 

refactor convert
parents 3e614260 d8e2664c
Expand all Show whitespace changes
Inline Side-by-side
Showing with 2325 additions and 1007 deletions
convert/convert.go
View file @ 0ff42e84
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 {
type Parameters 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
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:
var conv Converter
switch p.Architectures[0] {
case "LlamaForCausalLM", "MistralForCausalLM":
conv = &llama{}
case "MixtralForCausalLM":
conv = &mixtral{}
case "GemmaForCausalLM":
conv = &gemma{}
default:
t = sentencepiece.ModelProto_SentencePiece_NORMAL
}
v.Types = append(v.Types, int32(t))
return errors.New("unsupported architecture")
}
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
if err := json.Unmarshal(bts, conv); err != nil {
return err
}
slog.Info("reading user defined tokens")
var extraTokenData map[string]int
if err := json.Unmarshal(addIn, &extraTokenData); err != nil {
return nil, err
t, err := parseTokenizer(fsys, conv.specialTokenTypes())
if err != nil {
return err
}
type token struct {
key string
pos int
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)
}
extraTokens := make([]token, 0)
for k, id := range extraTokenData {
extraTokens = append(extraTokens, token{k, id})
} else {
slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
}
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)
}
v.Tokens = append(v.Tokens, t.key)
v.Scores = append(v.Scores, -1000.0)
v.Types = append(v.Types, tokenTypeUserDefined)
}
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 := range missingTokens {
v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", 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))
}
convert/convert_gemma.go 0 → 100644
View file @ 0ff42e84
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
}
convert/convert_llama.go 0 → 100644
View file @ 0ff42e84
package convert
import (
"cmp"
"fmt"
"strings"
"github.com/ollama/ollama/llm"
"github.com/pdevine/tensor"
"github.com/pdevine/tensor/native"
)
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
}
convert/convert_mixtral.go 0 → 100644
View file @ 0ff42e84
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
}
convert/convert_test.go
View file @ 0ff42e84
//go:build slow
package convert
import (
"crypto/sha256"
"encoding/json"
"flag"
"fmt"
"io"
"io/fs"
"log/slog"
"math"
"os"
"path/filepath"
"slices"
"testing"
"github.com/ollama/ollama/llm"
"golang.org/x/exp/maps"
)
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 {
m, _, err := llm.DecodeGGML(r, math.MaxInt)
if err != nil {
t.Fatal(err)
}
if err := arch.GetTensors(); err != nil {
if _, err := r.Seek(0, io.SeekStart); err != nil {
t.Fatal(err)
}
f, err := os.CreateTemp(t.TempDir(), "f16")
if err != nil {
t.Fatal(err)
}
defer f.Close()
return r, m.KV(), m.Tensors()
}
if err := arch.WriteGGUF(f); err != nil {
t.Fatal(err)
}
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())
}
r, err := os.Open(f.Name())
if err != nil {
t.Fatal(err)
func TestConvertFull(t *testing.T) {
cases := []string{
"Meta-Llama-3-8B-Instruct",
"Mistral-7B-Instruct-v0.2",
"Mixtral-8x7B-Instruct-v0.1",
"gemma-2b-it",
}
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)
}
defer r.Close()
m, _, err := llm.DecodeGGML(r)
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)
}
return m.KV(), m.Tensors()
}
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},
}
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 {
t.Skipf("%s not found", p)
actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts))
}
}
kv, tensors := convertFull(t, p)
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] = fmt.Sprintf("%x", 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)
}
}
})
}
......
convert/fs.go 0 → 100644
View file @ 0ff42e84
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)
}
convert/gemma.go deleted 100644 → 0
View file @ 3e614260
package convert
import (
"fmt"
"io"
"log/slog"
"strings"
"github.com/pdevine/tensor"
"github.com/pdevine/tensor/native"
"github.com/ollama/ollama/llm"
)
type GemmaModel struct {
ModelData
}
func addOnes(data []float32, vectorSize int) ([]float32, error) {
n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
ones := tensor.Ones(tensor.Float32, vectorSize)
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
}
func (m *GemmaModel) GetTensors() error {
t, err := m.Format.GetTensors(m.Path, m.Params)
if err != nil {
return err
}
slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
for _, l := range t {
if strings.HasSuffix(l.Name, "norm.weight") {
wt := l.WriterTo.(safetensorWriterTo)
wt.repacker = m.Repack
l.WriterTo = wt
}
m.Tensors = append(m.Tensors, l)
}
return nil
}
func (m *GemmaModel) LoadVocab() error {
v, err := LoadSentencePieceTokens(m.Path, m.Params)
if err != nil {
return err
}
m.Vocab = v
return nil
}
func (m *GemmaModel) Repack(_ string, data []float32, shape []uint64) ([]float32, error) {
return addOnes(data, int(shape[0]))
}
func (m *GemmaModel) WriteGGUF(ws io.WriteSeeker) error {
kv := llm.KV{
"general.architecture": "gemma",
"general.name": m.Name,
"gemma.context_length": uint32(m.Params.ContextSize),
"gemma.embedding_length": uint32(m.Params.HiddenSize),
"gemma.block_count": uint32(m.Params.HiddenLayers),
"gemma.feed_forward_length": uint32(m.Params.IntermediateSize),
"gemma.attention.head_count": uint32(m.Params.AttentionHeads),
"gemma.attention.head_count_kv": uint32(m.Params.KeyValHeads),
"gemma.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
"gemma.attention.key_length": uint32(m.Params.HeadDimension),
"gemma.attention.value_length": uint32(m.Params.HeadDimension),
"general.file_type": uint32(1),
"tokenizer.ggml.model": "llama",
"tokenizer.ggml.tokens": m.Vocab.Tokens,
"tokenizer.ggml.scores": m.Vocab.Scores,
"tokenizer.ggml.token_type": m.Vocab.Types,
"tokenizer.ggml.bos_token_id": uint32(m.Params.BoSTokenID),
"tokenizer.ggml.eos_token_id": uint32(m.Params.EoSTokenID),
"tokenizer.ggml.padding_token_id": uint32(m.Params.PaddingTokenID),
"tokenizer.ggml.unknown_token_id": uint32(3),
"tokenizer.ggml.add_bos_token": true,
"tokenizer.ggml.add_eos_token": false,
}
return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
}
convert/llama.go deleted 100644 → 0
View file @ 3e614260
package convert
import (
"cmp"
"errors"
"fmt"
"io"
"os"
"path/filepath"
"regexp"
"strings"
"github.com/pdevine/tensor"
"github.com/pdevine/tensor/native"
"github.com/ollama/ollama/llm"
)
type LlamaModel struct {
ModelData
}
func (m *LlamaModel) GetTensors() error {
t, err := m.Format.GetTensors(m.Path, m.Params)
if err != nil {
return err
}
pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
re, err := regexp.Compile(pattern)
if err != nil {
return err
}
for _, l := range t {
matches := re.FindAllStringSubmatch(l.Name, -1)
if len(matches) > 0 {
switch m.Format.(type) {
case *TorchFormat:
wt := l.WriterTo.(torchWriterTo)
wt.repacker = m.Repack
l.WriterTo = wt
case *SafetensorFormat:
wt := l.WriterTo.(safetensorWriterTo)
wt.repacker = m.Repack
l.WriterTo = wt
}
}
m.Tensors = append(m.Tensors, l)
}
return nil
}
func (m *LlamaModel) LoadVocab() (err error) {
pre, ts, merges, err := parseTokens(filepath.Join(m.Path, "tokenizer.json"))
if errors.Is(err, os.ErrNotExist) {
return nil
} else if err != nil {
return err
}
m.Vocab = &Vocab{}
for _, t := range ts {
m.Vocab.Tokens = append(m.Vocab.Tokens, t.Content)
m.Vocab.Types = append(m.Vocab.Types, t.Type())
}
m.Vocab.Merges = merges
m.Params.PreTokenizer = pre
return nil
}
func (m *LlamaModel) WriteGGUF(ws io.WriteSeeker) error {
kv := llm.KV{
"general.architecture": "llama",
"general.name": m.Name,
"llama.vocab_size": uint32(len(m.Vocab.Tokens)),
"llama.context_length": uint32(m.Params.ContextSize),
"llama.embedding_length": uint32(m.Params.HiddenSize),
"llama.block_count": uint32(m.Params.HiddenLayers),
"llama.feed_forward_length": uint32(m.Params.IntermediateSize),
"llama.rope.freq_base": float32(m.Params.RopeFrequencyBase),
"llama.rope.dimension_count": uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
"llama.attention.head_count": uint32(m.Params.AttentionHeads),
"llama.attention.head_count_kv": uint32(m.Params.KeyValHeads),
"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
"general.file_type": uint32(1),
"tokenizer.ggml.model": "gpt2",
"tokenizer.ggml.pre": m.Params.PreTokenizer,
"tokenizer.ggml.tokens": m.Vocab.Tokens,
"tokenizer.ggml.token_type": m.Vocab.Types,
"tokenizer.ggml.bos_token_id": uint32(m.Params.BoSTokenID),
"tokenizer.ggml.eos_token_id": uint32(m.Params.EoSTokenID),
"tokenizer.ggml.unknown_token_id": uint32(0),
}
if len(m.Vocab.Merges) > 0 {
kv["tokenizer.ggml.merges"] = m.Vocab.Merges
} else {
kv["tokenizer.ggml.scores"] = m.Vocab.Scores
}
return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
}
func (m *LlamaModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
return llamaRepack(name, m.Params, data, shape)
}
func llamaRepack(name string, params *Params, data []float32, shape []uint64) ([]float32, error) {
var dims []int
for _, dim := range shape {
if dim != 0 {
dims = append(dims, int(dim))
}
}
var heads int
switch {
case strings.HasSuffix(name, "attn_q.weight"):
heads = params.AttentionHeads
case strings.HasSuffix(name, "attn_k.weight"):
heads = cmp.Or(params.KeyValHeads, params.AttentionHeads)
default:
return nil, fmt.Errorf("unknown tensor name: %s", name)
}
n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
if err := n.Reshape(append([]int{heads, 2, dims[0] / 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
}
convert/mistral.go deleted 100644 → 0
View file @ 3e614260
package convert
import (
"io"
"regexp"
"github.com/ollama/ollama/llm"
)
type MistralModel struct {
ModelData
}
func (m *MistralModel) GetTensors() error {
t, err := m.Format.GetTensors(m.Path, m.Params)
if err != nil {
return err
}
pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
re, err := regexp.Compile(pattern)
if err != nil {
return err
}
for _, l := range t {
matches := re.FindAllStringSubmatch(l.Name, -1)
if len(matches) > 0 {
wt := l.WriterTo.(safetensorWriterTo)
wt.repacker = m.Repack
l.WriterTo = wt
}
m.Tensors = append(m.Tensors, l)
}
return nil
}
func (m *MistralModel) LoadVocab() error {
v, err := LoadSentencePieceTokens(m.Path, m.Params)
if err != nil {
return err
}
m.Vocab = v
return nil
}
func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
kv := llm.KV{
"general.architecture": "llama",
"general.name": m.Name,
"llama.context_length": uint32(m.Params.ContextSize),
"llama.embedding_length": uint32(m.Params.HiddenSize),
"llama.block_count": uint32(m.Params.HiddenLayers),
"llama.feed_forward_length": uint32(m.Params.IntermediateSize),
"llama.rope.dimension_count": uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
"llama.attention.head_count": uint32(m.Params.AttentionHeads),
"llama.attention.head_count_kv": uint32(m.Params.KeyValHeads),
"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
"general.file_type": uint32(1),
"tokenizer.ggml.model": "llama",
"tokenizer.ggml.tokens": m.Vocab.Tokens,
"tokenizer.ggml.scores": m.Vocab.Scores,
"tokenizer.ggml.token_type": m.Vocab.Types,
"tokenizer.ggml.bos_token_id": uint32(m.Params.BoSTokenID),
"tokenizer.ggml.eos_token_id": uint32(m.Params.EoSTokenID),
"tokenizer.ggml.add_bos_token": true,
"tokenizer.ggml.add_eos_token": false,
"tokenizer.ggml.unknown_token_id": uint32(0),
}
if m.Params.HeadDimension > 0 {
kv["llama.attention.key_length"] = uint32(m.Params.HeadDimension)
kv["llama.attention.value_length"] = uint32(m.Params.HeadDimension)
}
return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
}
func (m *MistralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
return llamaRepack(name, m.Params, data, shape)
}
convert/mixtral.go deleted 100644 → 0
View file @ 3e614260
package convert
import (
"io"
"regexp"
"github.com/ollama/ollama/llm"
)
type MixtralModel struct {
ModelData
}
func (m *MixtralModel) GetTensors() error {
t, err := m.Format.GetTensors(m.Path, m.Params)
if err != nil {
return err
}
pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
re, err := regexp.Compile(pattern)
if err != nil {
return err
}
for _, l := range t {
matches := re.FindAllStringSubmatch(l.Name, -1)
if len(matches) > 0 {
wt := l.WriterTo.(safetensorWriterTo)
wt.repacker = m.Repack
l.WriterTo = wt
}
m.Tensors = append(m.Tensors, l)
}
return nil
}
func (m *MixtralModel) LoadVocab() error {
v, err := LoadSentencePieceTokens(m.Path, m.Params)
if err != nil {
return err
}
m.Vocab = v
return nil
}
func (m *MixtralModel) WriteGGUF(ws io.WriteSeeker) error {
kv := llm.KV{
"general.architecture": "llama",
"general.name": m.Name,
"llama.block_count": uint32(m.Params.HiddenLayers),
"llama.context_length": uint32(m.Params.ContextSize),
"llama.embedding_length": uint32(m.Params.HiddenSize),
"llama.feed_forward_length": uint32(m.Params.IntermediateSize),
"llama.attention.head_count": uint32(m.Params.AttentionHeads),
"llama.attention.head_count_kv": uint32(m.Params.KeyValHeads),
"llama.rope.freq_base": float32(m.Params.RopeFrequencyBase),
"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
"llama.expert_count": uint32(m.Params.Experts),
"llama.expert_used_count": uint32(m.Params.ExpertsUsed),
"llama.vocab_size": uint32(len(m.Vocab.Tokens)),
"llama.rope.dimension_count": uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
"general.file_type": uint32(1),
"tokenizer.ggml.model": "llama",
"tokenizer.ggml.tokens": m.Vocab.Tokens,
"tokenizer.ggml.scores": m.Vocab.Scores,
"tokenizer.ggml.token_type": m.Vocab.Types,
"tokenizer.ggml.bos_token_id": uint32(m.Params.BoSTokenID),
"tokenizer.ggml.eos_token_id": uint32(m.Params.EoSTokenID),
"tokenizer.ggml.unknown_token_id": uint32(0),
"tokenizer.ggml.add_bos_token": true,
"tokenizer.ggml.add_eos_token": false,
}
return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
}
func (m *MixtralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
return llamaRepack(name, m.Params, data, shape)
}
convert/reader.go 0 → 100644
View file @ 0ff42e84
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")
}
convert/reader_safetensors.go 0 → 100644
View file @ 0ff42e84
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
}
for _, b := range u16s {
f32s = append(f32s, float16.Frombits(b).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())
}
}
convert/reader_torch.go 0 → 100644
View file @ 0ff42e84
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
}
convert/safetensors.go deleted 100644 → 0
View file @ 3e614260
package convert
import (
"bytes"
"encoding/binary"
"encoding/json"
"fmt"
"io"
"os"
"path/filepath"
"regexp"
"slices"
"strings"
"github.com/d4l3k/go-bfloat16"
"github.com/x448/float16"
"github.com/ollama/ollama/llm"
)
type safetensorWriterTo struct {
t *llm.Tensor
params *Params
bo ByteOrder
filename string
dtype string
offset, size int64
repacker func(string, []float32, []uint64) ([]float32, error)
}
type safetensorMetadata struct {
Type string `json:"dtype"`
Shape []uint64 `json:"shape"`
Offsets []int64 `json:"data_offsets"`
}
type SafetensorFormat struct{}
func (m *SafetensorFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
var tensors []llm.Tensor
matches, err := filepath.Glob(filepath.Join(dirpath, "*.safetensors"))
if err != nil {
return nil, err
}
var offset uint64
for _, f := range matches {
var t []llm.Tensor
var err error
t, offset, err = m.readTensors(f, offset, params)
if err != nil {
return nil, err
}
tensors = append(tensors, t...)
}
return tensors, nil
}
func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params) ([]llm.Tensor, uint64, error) {
f, err := os.Open(fn)
if err != nil {
return nil, 0, err
}
defer f.Close()
var n int64
if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
return nil, 0, err
}
b := bytes.NewBuffer(make([]byte, 0, n))
if _, err = io.CopyN(b, f, n); err != nil {
return nil, 0, err
}
var headers map[string]safetensorMetadata
if err := json.NewDecoder(b).Decode(&headers); err != nil {
return nil, 0, err
}
var keys []string
for key := range headers {
if !strings.HasSuffix(key, "self_attn.rotary_embd.inv_freq") {
keys = append(keys, key)
}
}
slices.Sort(keys)
var tensors []llm.Tensor
for _, key := range keys {
value := headers[key]
var kind uint32
switch len(value.Shape) {
case 0:
// valuedata
continue
case 2:
kind = 1
}
name, err := m.GetLayerName(key)
if err != nil {
return nil, 0, err
}
shape := make([]uint64, len(value.Shape))
copy(shape, value.Shape)
pad := func(s int64) int64 {
return 8 + n + s
}
t := llm.Tensor{
Name: name,
Kind: kind,
Offset: offset,
Shape: shape,
}
t.WriterTo = safetensorWriterTo{
t: &t,
params: params,
bo: params.ByteOrder,
filename: fn,
dtype: value.Type,
offset: pad(value.Offsets[0]),
size: pad(value.Offsets[1]) - pad(value.Offsets[0]),
}
offset += t.Size()
tensors = append(tensors, t)
}
return tensors, offset, nil
}
func (m *SafetensorFormat) GetParams(dirpath string) (*Params, error) {
f, err := os.Open(filepath.Join(dirpath, "config.json"))
if err != nil {
return nil, err
}
defer f.Close()
var params Params
if err := json.NewDecoder(f).Decode(&params); err != nil {
return nil, err
}
params.ByteOrder = binary.LittleEndian
return &params, nil
}
func (m *SafetensorFormat) GetLayerName(n string) (string, error) {
directMap := map[string]string{
"model.embed_tokens.weight": "token_embd.weight",
"lm_head.weight": "output.weight",
"model.norm.weight": "output_norm.weight",
}
tMap := map[string]string{
"model.layers.(\\d+).input_layernorm.weight": "blk.$1.attn_norm.weight",
"model.layers.(\\d+).mlp.down_proj.weight": "blk.$1.ffn_down.weight",
"model.layers.(\\d+).mlp.gate_proj.weight": "blk.$1.ffn_gate.weight",
"model.layers.(\\d+).mlp.up_proj.weight": "blk.$1.ffn_up.weight",
"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
"model.layers.(\\d+).self_attn.k_proj.weight": "blk.$1.attn_k.weight",
"model.layers.(\\d+).self_attn.o_proj.weight": "blk.$1.attn_output.weight",
"model.layers.(\\d+).self_attn.q_proj.weight": "blk.$1.attn_q.weight",
"model.layers.(\\d+).self_attn.v_proj.weight": "blk.$1.attn_v.weight",
"model.layers.(\\d+).block_sparse_moe.gate.weight": "blk.$1.ffn_gate_inp.weight",
"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w1.weight": "blk.$1.ffn_gate.$2.weight",
"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w2.weight": "blk.$1.ffn_down.$2.weight",
"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w3.weight": "blk.$1.ffn_up.$2.weight",
}
v, ok := directMap[n]
if ok {
return v, nil
}
// quick hack to rename the layers to gguf format
for k, v := range tMap {
re := regexp.MustCompile(k)
newName := re.ReplaceAllString(n, v)
if newName != n {
return newName, nil
}
}
return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
}
func (r safetensorWriterTo) WriteTo(w io.Writer) (n int64, err error) {
f, err := os.Open(r.filename)
if err != nil {
return 0, err
}
defer f.Close()
if _, err = f.Seek(r.offset, io.SeekStart); err != nil {
return 0, err
}
var f32s []float32
switch r.dtype {
case "F32":
f32s = make([]float32, r.size/4)
if err = binary.Read(f, r.bo, f32s); err != nil {
return 0, err
}
case "F16":
u16s := make([]uint16, r.size/2)
if err = binary.Read(f, r.bo, u16s); err != nil {
return 0, err
}
for _, b := range u16s {
f32s = append(f32s, float16.Frombits(b).Float32())
}
case "BF16":
u8s := make([]uint8, r.size)
if err = binary.Read(f, r.bo, u8s); err != nil {
return 0, err
}
f32s = bfloat16.DecodeFloat32(u8s)
default:
return 0, fmt.Errorf("unknown data type: %s", r.dtype)
}
if r.repacker != nil {
f32s, err = r.repacker(r.t.Name, f32s, r.t.Shape)
if err != nil {
return 0, err
}
}
switch r.t.Kind {
case 0:
return 0, binary.Write(w, r.bo, f32s)
case 1:
f16s := make([]uint16, len(f32s))
for i := range f32s {
f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
}
return 0, binary.Write(w, r.bo, f16s)
default:
return 0, fmt.Errorf("unknown storage type: %d", r.t.Kind)
}
}
func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
switch len(params.Architectures) {
case 0:
return nil, fmt.Errorf("No architecture specified to convert")
case 1:
switch params.Architectures[0] {
case "LlamaForCausalLM":
return &LlamaModel{
ModelData{
Name: name,
Path: dirPath,
Params: params,
Format: m,
},
}, nil
case "MistralForCausalLM":
return &MistralModel{
ModelData{
Name: name,
Path: dirPath,
Params: params,
Format: m,
},
}, nil
case "MixtralForCausalLM":
return &MixtralModel{
ModelData{
Name: name,
Path: dirPath,
Params: params,
Format: m,
},
}, nil
case "GemmaForCausalLM":
return &GemmaModel{
ModelData{
Name: name,
Path: dirPath,
Params: params,
Format: m,
},
}, nil
default:
return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
}
}
return nil, fmt.Errorf("Unknown error")
}
convert/testdata/Meta-Llama-3-8B-Instruct.json 0 → 100644
View file @ 0ff42e84
This diff is collapsed.
convert/testdata/Mistral-7B-Instruct-v0.2.json 0 → 100644
View file @ 0ff42e84
This diff is collapsed.
convert/testdata/Mixtral-8x7B-Instruct-v0.1.json 0 → 100644
View file @ 0ff42e84
This diff is collapsed.
convert/testdata/gemma-2b-it.json 0 → 100644
View file @ 0ff42e84
{
"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"
}
convert/tokenizer.go
View file @ 0ff42e84
......@@ -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 {
*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 TokenizerModel `json:"model"`
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 {
......@@ -27,80 +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
}
}
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,
)
type Vocabulary struct {
Model string
Tokens []string
Scores []float32
Types []int32
}
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 {
if pt.Type == "Split" && 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)
}
}
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 &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)
}
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")
}
convert/tokenizer_spm.go 0 → 100644
View file @ 0ff42e84
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
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment