next ollama runner (#7913)

feat: add new Ollama engine using ggml through cgo

This change introduces a new way to run pretrained models. It introduces 3 high level interfaces and a bunch of smaller helper interfaces to facilitate this.

- `model.Model` defines the interface for a model architecture. Models such as `llama` and `mllama`, which are provided as examples, can implement the model's forward propagation in the `Forward` method. This method will be called to generate completions. This interface can be found in `model/model.go`
- `ml.Backend` defines the interface for a backend tensor library, in this case `ggml`. Among other things, a Backend is responsible for loading a pretrained model into hardware (GPU, CPU, etc) and providing an interface for Models to access loaded tensors. This interface can be found in `ml/backend.go`
- `ml.Tensor` defines the interface for a tensor and tensor operations

This is the first implementation of the new engine. Follow up PRs will implement more features:

- non-greedy sampling (#8410)
- integration with Ollama and KV caching (#8301)
- more model support (#9080) with more coming soon
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>

cache/cache.go

+package cache
+
+import (
+	"github.com/ollama/ollama/ml"
+)
+
+type Options struct {
+	Position int
+}
+
+type Cache interface {
+	Sub(i int) Cache
+	Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor)
+}
+
+type Simple struct {
+	DType    ml.DType
+	Capacity int
+
+	keys, values []ml.Tensor
+}
+
+func (c *Simple) Sub(i int) Cache {
+	if i >= len(c.keys) {
+		c.keys = append(c.keys, make([]ml.Tensor, i-len(c.keys)+1)...)
+		c.values = append(c.values, make([]ml.Tensor, i-len(c.values)+1)...)
+	}
+
+	return &Simple{
+		keys:     c.keys[i : i+1],
+		values:   c.values[i : i+1],
+		Capacity: c.Capacity,
+		DType:    c.DType,
+	}
+}
+
+func (c *Simple) Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor) {
+	if c.keys[0] == nil || c.values[0] == nil {
+		c.keys[0] = ctx.Zeros(c.DType, int(key.Dim(0)*key.Dim(1))*c.Capacity)
+		c.values[0] = ctx.Zeros(c.DType, int(value.Dim(0)*value.Dim(1))*c.Capacity)
+	}
+
+	ctx.Forward(key.Copy(ctx, c.keys[0].View(ctx, int(key.Stride(2))*opts.Position, int(key.Dim(0)*key.Dim(1)*key.Dim(2)))))
+	ctx.Forward(value.Copy(ctx, c.values[0].View(ctx, int(value.Stride(2))*opts.Position, int(value.Dim(0)*value.Dim(1)*value.Dim(2)))))
+
+	n := min(c.Capacity, int(key.Dim(2))+opts.Position)
+
+	key = c.keys[0].View(ctx, 0,
+		int(key.Dim(0)), int(key.Stride(1)),
+		int(key.Dim(1)), int(key.Stride(2)),
+		n,
+	)
+
+	value = c.values[0].View(ctx, 0,
+		int(value.Dim(0)), int(value.Stride(1)),
+		int(value.Dim(1)), int(value.Stride(2)),
+		n,
+	)
+
+	// TODO shift context if necessary
+
+	return key, value
+}

convert/convert.go

@@ -9,7 +9,7 @@ import (
 	"log/slog"
 	"strings"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type ModelParameters struct {
@@ -27,8 +27,8 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) llm.KV {
-	kv := llm.KV{
+func (ModelParameters) KV(t *Tokenizer) ggml.KV {
+	kv := ggml.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -54,7 +54,7 @@ func (ModelParameters) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() llm.KV {
+func (p AdapterParameters) KV() ggml.KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -62,7 +62,7 @@ func (p AdapterParameters) KV() llm.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := llm.KV{
+	kv := ggml.KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -79,19 +79,19 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
-	return llm.WriteGGUF(ws, kv, ts)
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
+	return ggml.WriteGGUF(ws, kv, ts)
 }
 
-func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
-	return llm.WriteGGUF(ws, kv, ts)
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
+	return ggml.WriteGGUF(ws, kv, ts)
 }
 
 type ModelConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) llm.KV
+	KV(*Tokenizer) ggml.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
-	Tensors([]Tensor) []llm.Tensor
+	Tensors([]Tensor) []ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
@@ -99,7 +99,7 @@ type ModelConverter interface {
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
 	// writeFile writes the model to the provided io.WriteSeeker
-	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
+	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
 type moreParser interface {
@@ -108,17 +108,17 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(llm.KV) llm.KV
+	KV(ggml.KV) ggml.KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
-	Tensors([]Tensor) []llm.Tensor
+	Tensors([]Tensor) []ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
 
-	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
+	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
-func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV llm.KV) error {
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err

convert/convert_bert.go

@@ -8,7 +8,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type bertModel struct {
@@ -85,7 +85,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) llm.KV {
+func (p *bertModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false
@@ -132,8 +132,8 @@ func (p *bertModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *bertModel) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
 	for _, t := range ts {
 		if slices.Contains([]string{
 			"embeddings.position_ids",
@@ -143,7 +143,7 @@ func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
 			continue
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_commandr.go

@@ -3,7 +3,7 @@ package convert
 import (
 	"cmp"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type commandrModel struct {
@@ -24,7 +24,7 @@ type commandrModel struct {
 
 var _ ModelConverter = (*commandrModel)(nil)
 
-func (p *commandrModel) KV(t *Tokenizer) llm.KV {
+func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "command-r"
 	kv["general.name"] = "command-r"
@@ -43,10 +43,10 @@ func (p *commandrModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *commandrModel) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *commandrModel) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
 	for _, t := range ts {
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemmaModel struct {
@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
+func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings
@@ -42,14 +42,14 @@ func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *gemmaModel) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *gemmaModel) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
 	for _, t := range ts {
 		if strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma2.go

 package convert
 
-import (
-	"github.com/ollama/ollama/llm"
-)
+import "github.com/ollama/ollama/fs/ggml"
 
 type gemma2Model struct {
 	gemmaModel
@@ -11,7 +9,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
+func (p *gemma2Model) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemma2Adapter struct {
@@ -15,14 +15,14 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
+func (p *gemma2Adapter) KV(baseKV ggml.KV) ggml.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv
 }
 
-func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *gemma2Adapter) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -31,7 +31,7 @@ func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama.go

@@ -9,7 +9,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type llamaModel struct {
@@ -46,7 +46,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) llm.KV {
+func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize
@@ -120,11 +120,11 @@ func (p *llamaModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *llamaModel) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
 
 	if p.RopeScaling.factors != nil {
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     "rope_freqs.weight",
 			Kind:     0,
 			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
@@ -138,7 +138,7 @@ func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama_adapter.go

@@ -7,7 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type llamaAdapter struct {
@@ -18,7 +18,7 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
+func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
 	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
@@ -29,8 +29,8 @@ func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
 	return kv
 }
 
-func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *llamaAdapter) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -41,7 +41,7 @@ func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    shape,

convert/convert_mixtral.go

@@ -6,7 +6,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type mixtralModel struct {
@@ -15,7 +15,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
+func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {
@@ -29,7 +29,7 @@ func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []ggml.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -56,10 +56,10 @@ func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
 		return true
 	})
 
-	var out []llm.Tensor
+	var out []ggml.Tensor
 	for n, e := range experts {
 		// TODO(mxyng): sanity check experts
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     n,
 			Kind:     e[0].Kind(),
 			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),

convert/convert_phi3.go

@@ -8,7 +8,7 @@ import (
 	"strings"
 	"sync"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type phi3Model struct {
@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) llm.KV {
+func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings
@@ -68,19 +68,19 @@ func (p *phi3Model) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []ggml.Tensor {
 	var addRopeFactors sync.Once
 
-	out := make([]llm.Tensor, 0, len(ts)+2)
+	out := make([]ggml.Tensor, 0, len(ts)+2)
 	for _, t := range ts {
 		if strings.HasPrefix(t.Name(), "blk.0.") {
 			addRopeFactors.Do(func() {
-				out = append(out, llm.Tensor{
+				out = append(out, ggml.Tensor{
 					Name:     "rope_factors_long.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
 					WriterTo: p.RopeScaling.LongFactor,
-				}, llm.Tensor{
+				}, ggml.Tensor{
 					Name:     "rope_factors_short.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
@@ -89,7 +89,7 @@ func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
 			})
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_qwen2.go

 package convert
 
-import "github.com/ollama/ollama/llm"
+import "github.com/ollama/ollama/fs/ggml"
 
 type qwen2Model struct {
 	ModelParameters
@@ -21,7 +21,7 @@ type qwen2Model struct {
 
 var _ ModelConverter = (*qwen2Model)(nil)
 
-func (q *qwen2Model) KV(t *Tokenizer) llm.KV {
+func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen2"
 	kv["qwen2.block_count"] = q.HiddenLayers
@@ -45,10 +45,10 @@ func (q *qwen2Model) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (q *qwen2Model) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (q *qwen2Model) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
 	for _, t := range ts {
-		out = append(out, llm.Tensor{
+		out = append(out, ggml.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_test.go

@@ -20,7 +20,7 @@ import (
 
 	"golang.org/x/exp/maps"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type tensorData struct {
@@ -29,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, *llm.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -48,7 +48,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, *llm.Tensors) {
 	}
 	t.Cleanup(func() { r.Close() })
 
-	m, _, err := llm.DecodeGGML(r, math.MaxInt)
+	m, _, err := ggml.Decode(r, math.MaxInt)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -60,7 +60,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, *llm.Tensors) {
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv llm.KV, tensors *llm.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv ggml.KV, tensors ggml.Tensors) map[string]string {
 	actual := make(map[string]string)
 	for k, v := range kv {
 		if s, ok := v.(json.Marshaler); !ok {
@@ -75,7 +75,7 @@ func generateResultsJSON(t *testing.T, f *os.File, kv llm.KV, tensors *llm.Tenso
 		}
 	}
 
-	for _, tensor := range tensors.Items {
+	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 {
@@ -332,7 +332,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := llm.DecodeGGML(r, math.MaxInt)
+			m, _, err := ggml.Decode(r, math.MaxInt)
 			if err != nil {
 				t.Fatal(err)
 			}

llm/ggml.go → fs/ggml/ggml.go

-package llm
+package ggml
 
 import (
 	"encoding/binary"
 	"errors"
 	"fmt"
 	"io"
+	"log/slog"
 	"slices"
 	"strings"
-	"sync"
 
-	"github.com/ollama/ollama/util/bufioutil"
+	"github.com/ollama/ollama/fs/util/bufioutil"
 )
 
 type GGML struct {
@@ -19,145 +19,166 @@ type GGML struct {
 
 type model interface {
 	KV() KV
-	Tensors() *Tensors
+	Tensors() Tensors
 }
 
 type KV map[string]any
 
-func (kv KV) u64(key string) uint64 {
-	switch v := kv[key].(type) {
-	case uint64:
-		return v
-	case uint32:
-		return uint64(v)
-	case float64:
-		return uint64(v)
-	default:
-		return 0
-	}
-}
-
 func (kv KV) Architecture() string {
-	if s, ok := kv["general.architecture"].(string); ok {
-		return s
-	}
-
-	return "unknown"
+	return kv.String("general.architecture", "unknown")
 }
 
 func (kv KV) Kind() string {
-	if s, ok := kv["general.type"].(string); ok {
-		return s
-	}
-
-	return "unknown"
+	return kv.String("general.type", "unknown")
 }
 
 func (kv KV) ParameterCount() uint64 {
-	return kv.u64("general.parameter_count")
+	return keyValue[uint64](kv, "general.parameter_count")
 }
 
 func (kv KV) FileType() fileType {
-	if u64 := kv.u64("general.file_type"); u64 > 0 {
-		return fileType(uint32(u64))
+	if t := kv.Uint("general.file_type"); t > 0 {
+		return fileType(t)
 	}
 
 	return fileTypeUnknown
 }
 
 func (kv KV) BlockCount() uint64 {
-	return kv.u64(fmt.Sprintf("%s.block_count", kv.Architecture()))
+	return uint64(kv.Uint("block_count"))
+}
+
+func (kv KV) EmbeddingLength() uint64 {
+	return uint64(kv.Uint("embedding_length"))
 }
 
 func (kv KV) HeadCount() uint64 {
-	return kv.u64(fmt.Sprintf("%s.attention.head_count", kv.Architecture()))
+	return uint64(kv.Uint("attention.head_count"))
 }
 
 func (kv KV) HeadCountKV() uint64 {
-	if headCountKV := kv.u64(fmt.Sprintf("%s.attention.head_count_kv", kv.Architecture())); headCountKV > 0 {
-		return headCountKV
-	}
-
-	return 1
+	return uint64(kv.Uint("attention.head_count_kv", 1))
 }
 
 func (kv KV) EmbeddingHeadCount() uint64 {
 	if heads := kv.HeadCount(); heads > 0 {
-		return kv.EmbeddingLength() / kv.HeadCount()
+		return kv.EmbeddingLength() / heads
 	}
 
 	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()
+	return uint64(kv.Uint("attention.key_length", uint32(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()
+	return uint64(kv.Uint("attention.value_length", uint32(kv.EmbeddingHeadCount())))
 }
 
 func (kv KV) GQA() uint64 {
 	return kv.HeadCount() / kv.HeadCountKV()
 }
 
-func (kv KV) EmbeddingLength() uint64 {
-	return kv.u64(fmt.Sprintf("%s.embedding_length", kv.Architecture()))
-}
-
 func (kv KV) ContextLength() uint64 {
-	return kv.u64(fmt.Sprintf("%s.context_length", kv.Architecture()))
+	return uint64(kv.Uint("context_length"))
 }
 
 func (kv KV) ChatTemplate() string {
-	s, _ := kv["tokenizer.chat_template"].(string)
+	return kv.String("tokenizer.chat_template")
+}
+
+func (kv KV) String(key string, defaultValue ...string) string {
+	return keyValue(kv, key, append(defaultValue, "")...)
+}
+
+func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
+	return keyValue(kv, key, append(defaultValue, 0)...)
+}
+
+func (kv KV) Float(key string, defaultValue ...float32) float32 {
+	return keyValue(kv, key, append(defaultValue, 0)...)
+}
+
+func (kv KV) Strings(key string, defaultValue ...[]string) []string {
+	r := keyValue(kv, key, &array{})
+	s := make([]string, r.size)
+	for i := range r.size {
+		s[i] = r.values[i].(string)
+	}
+
+	return s
+}
+
+func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
+	r := keyValue(kv, key, &array{})
+	s := make([]uint32, r.size)
+	for i := range r.size {
+		s[i] = uint32(r.values[i].(int32))
+	}
+
 	return s
 }
 
+func keyValue[T string | uint32 | uint64 | float32 | *array](kv KV, key string, defaultValue ...T) T {
+	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
+		key = kv.Architecture() + "." + key
+	}
+
+	if val, ok := kv[key]; ok {
+		return val.(T)
+	}
+
+	slog.Warn("key not found", "key", key, "default", defaultValue[0])
+	return defaultValue[0]
+}
+
 type Tensors struct {
-	Items  []*Tensor
+	items  []*Tensor
 	Offset uint64
+}
 
-	layers     map[string]Layer
-	layersOnce sync.Once
-}
-
-func (ts *Tensors) Layers() map[string]Layer {
-	ts.layersOnce.Do(func() {
-		ts.layers = make(map[string]Layer)
-		for _, t := range ts.Items {
-			parts := strings.Split(t.Name, ".")
-			if index := slices.IndexFunc(parts, func(s string) bool { return s == "blk" || s == "mm" }); index != -1 {
-				if len(parts) > index+2 {
-					// blk and mm should have a number after them, join it
-					parts = append(
-						[]string{strings.Join(parts[:index+2], ".")},
-						parts[index+2:]...)
-				}
-			}
+func (s Tensors) Items(prefix ...string) []*Tensor {
+	if len(prefix) == 0 {
+		return s.items
+	}
 
-			if _, ok := ts.layers[parts[0]]; !ok {
-				ts.layers[parts[0]] = make(Layer)
+	var items []*Tensor
+	for _, t := range s.items {
+		if strings.HasPrefix(t.Name, prefix[0]) {
+			items = append(items, t)
+		}
+	}
+
+	return items
+}
+
+func (ts Tensors) GroupLayers() map[string]Layer {
+	layers := make(map[string]Layer)
+	for _, t := range ts.items {
+		parts := strings.Split(t.Name, ".")
+		if index := slices.IndexFunc(parts, func(s string) bool { return s == "blk" || s == "mm" }); index != -1 {
+			if len(parts) > index+2 {
+				// blk and mm should have a number after them, join it
+				parts = append(
+					[]string{strings.Join(parts[:index+2], ".")},
+					parts[index+2:]...)
 			}
+		}
 
-			ts.layers[parts[0]][strings.Join(parts[1:], ".")] = t
+		if _, ok := layers[parts[0]]; !ok {
+			layers[parts[0]] = make(Layer)
 		}
-	})
 
-	return ts.layers
+		layers[parts[0]][strings.Join(parts[1:], ".")] = t
+	}
+
+	return layers
 }
 
 type Layer map[string]*Tensor
 
-func (l Layer) size() (size uint64) {
+func (l Layer) Size() (size uint64) {
 	for _, t := range l {
 		size += t.Size()
 	}
@@ -255,8 +276,6 @@ func (t Tensor) typeSize() uint64 {
 		return 8
 	case 29: // IQ1_M
 		return blockSize/8 + blockSize/16 + blockSize/32
-	case 30: // BF16
-		return 2
 	default:
 		return 0
 	}
@@ -295,7 +314,7 @@ const (
 
 var ErrUnsupportedFormat = errors.New("unsupported model format")
 
-func DetectGGMLType(b []byte) string {
+func DetectContentType(b []byte) string {
 	switch binary.LittleEndian.Uint32(b[:4]) {
 	case FILE_MAGIC_GGML:
 		return "ggml"
@@ -312,12 +331,12 @@ func DetectGGMLType(b []byte) string {
 	}
 }
 
-// DecodeGGML decodes a GGML model from the given reader.
+// Decode 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) {
+func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	if maxArraySize == 0 {
 		maxArraySize = 1024
 	}
@@ -331,10 +350,6 @@ func DecodeGGML(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 
 	var c container
 	switch magic {
-	case FILE_MAGIC_GGML, FILE_MAGIC_GGMF, FILE_MAGIC_GGJT:
-		return nil, 0, ErrUnsupportedFormat
-	case FILE_MAGIC_GGLA:
-		c = &containerGGLA{}
 	case FILE_MAGIC_GGUF_LE:
 		c = &containerGGUF{ByteOrder: binary.LittleEndian, maxArraySize: maxArraySize}
 	case FILE_MAGIC_GGUF_BE:
@@ -360,22 +375,22 @@ func DecodeGGML(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }
 
-func (llm GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
-	embedding := llm.KV().EmbeddingLength()
-	heads := llm.KV().HeadCount()
-	headsKV := llm.KV().HeadCountKV()
-	vocab := uint64(llm.KV()["tokenizer.ggml.tokens"].(*array).size)
+func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+	embedding := f.KV().EmbeddingLength()
+	heads := f.KV().HeadCount()
+	headsKV := f.KV().HeadCountKV()
+	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array).size)
 
-	embeddingHeads := llm.KV().EmbeddingHeadCount()
-	embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
-	embeddingHeadsV := llm.KV().EmbeddingHeadCountV()
+	embeddingHeads := f.KV().EmbeddingHeadCount()
+	embeddingHeadsK := f.KV().EmbeddingHeadCountK()
+	embeddingHeadsV := f.KV().EmbeddingHeadCountV()
 
-	layers := llm.Tensors().Layers()
+	layers := f.Tensors().GroupLayers()
 
 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = uint64(float64(context*llm.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
 
-	switch llm.KV().Architecture() {
+	switch f.KV().Architecture() {
 	case "llama":
 		fullOffload = max(
 			4*batch*(1+4*embedding+context*(1+heads)),
@@ -390,7 +405,7 @@ func (llm GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partia
 
 		if ffnGateExpsWeight, ok := layers["blk.0"]["ffn_gate_exps.weight"]; ok {
 			// mixtral 8x22b
-			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
+			ff := uint64(f.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
 				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
 				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
@@ -407,11 +422,11 @@ func (llm GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partia
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4
 
-		if crossAttentionLayers, ok := llm.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
+		if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
 			kv = headsKV *
 				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
 				(2* // sizeof(float16)
-					(llm.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
+					(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
 					context +
 					4* // sizeof(float32)
 						uint64(crossAttentionLayers.size)* // num cross attention layers
@@ -426,7 +441,7 @@ func (llm GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partia
 		)
 
 		var ropeFreqsCount uint64
-		if ropeFreqs, ok := llm.Tensors().Layers()["rope_freqs"]; ok {
+		if ropeFreqs, ok := f.Tensors().GroupLayers()["rope_freqs"]; ok {
 			if ropeFreqsWeights, ok := ropeFreqs["weights"]; ok {
 				ropeFreqsCount = ropeFreqsWeights.parameters()
 			}
@@ -530,21 +545,20 @@ func (llm GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partia
 }
 
 // SupportsKVCacheType checks if the requested cache type is supported
-func (ggml GGML) SupportsKVCacheType(cacheType string) bool {
-	validKVCacheTypes := []string{"f16", "q8_0", "q4_0"}
-	return slices.Contains(validKVCacheTypes, cacheType)
+func (f GGML) SupportsKVCacheType(cacheType string) bool {
+	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
 }
 
 // SupportsFlashAttention checks if the model supports flash attention
-func (ggml GGML) SupportsFlashAttention() bool {
-	_, isEmbedding := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]
+func (f GGML) SupportsFlashAttention() bool {
+	_, isEmbedding := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]
 	if isEmbedding {
 		return false
 	}
 
 	// Check head counts match and are non-zero
-	headCountK := ggml.KV().EmbeddingHeadCountK()
-	headCountV := ggml.KV().EmbeddingHeadCountV()
+	headCountK := f.KV().EmbeddingHeadCountK()
+	headCountV := f.KV().EmbeddingHeadCountV()
 	return headCountK != 0 && headCountV != 0 && headCountK == headCountV
 }
 

fs/ggml/ggml_test.go

+package ggml
+
+import (
+	"maps"
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+func TestTensorLayers(t *testing.T) {
+	tensors := make(map[string]*Tensor)
+	for _, name := range []string{
+		"token_embd.weight",
+		"blk.0.attn_k.weight",
+		"blk.0.attn_output.weight",
+		"blk.0.attn_q.weight",
+		"blk.0.attn_v.weight",
+		"blk.0.attn_norm.weight",
+		"blk.0.ffn_down.weight",
+		"blk.0.ffn_gate.weight",
+		"blk.0.ffn_up.weight",
+		"blk.0.ffn_norm.weight",
+		"output_norm.weight",
+		"mm.0.bias",
+		"mm.0.weight",
+		"v.blk.0.attn_k.weight",
+		"v.blk.0.attn_output.weight",
+		"v.blk.0.attn_q.weight",
+		"v.blk.0.attn_v.weight",
+		"v.blk.0.attn_norm.weight",
+		"v.blk.0.ffn_down.weight",
+		"v.blk.0.ffn_gate.weight",
+		"v.blk.0.ffn_up.weight",
+		"v.blk.0.ffn_norm.weight",
+		"v.patch_embd.weight",
+		"v.position_embd.gate",
+		"v.position_embd.weight",
+	} {
+		tensors[name] = &Tensor{Name: name}
+	}
+
+	cases := []struct {
+		name  string
+		items []*Tensor
+		want  map[string]Layer
+	}{
+		{
+			name: "text",
+			items: slices.Collect(func(yield func(*Tensor) bool) {
+				for k, v := range tensors {
+					if !strings.HasPrefix(k, "mm.") && !strings.HasPrefix(k, "v.") {
+						if !yield(v) {
+							return
+						}
+					}
+				}
+			}),
+			want: map[string]Layer{
+				"blk.0": {
+					"attn_k.weight":      tensors["blk.0.attn_k.weight"],
+					"attn_q.weight":      tensors["blk.0.attn_q.weight"],
+					"attn_v.weight":      tensors["blk.0.attn_v.weight"],
+					"attn_output.weight": tensors["blk.0.attn_output.weight"],
+					"attn_norm.weight":   tensors["blk.0.attn_norm.weight"],
+					"ffn_down.weight":    tensors["blk.0.ffn_down.weight"],
+					"ffn_gate.weight":    tensors["blk.0.ffn_gate.weight"],
+					"ffn_up.weight":      tensors["blk.0.ffn_up.weight"],
+					"ffn_norm.weight":    tensors["blk.0.ffn_norm.weight"],
+				},
+				"token_embd":  {"weight": tensors["token_embd.weight"]},
+				"output_norm": {"weight": tensors["output_norm.weight"]},
+			},
+		},
+		{
+			name: "vision",
+			items: slices.Collect(func(yield func(*Tensor) bool) {
+				for k, v := range tensors {
+					if strings.HasPrefix(k, "mm.") || strings.HasPrefix(k, "v.") {
+						if !yield(v) {
+							return
+						}
+					}
+				}
+			}),
+			want: map[string]Layer{
+				"mm.0": {
+					"bias":   tensors["mm.0.bias"],
+					"weight": tensors["mm.0.weight"],
+				},
+				"v.blk.0": {
+					"attn_k.weight":      tensors["v.blk.0.attn_k.weight"],
+					"attn_q.weight":      tensors["v.blk.0.attn_q.weight"],
+					"attn_v.weight":      tensors["v.blk.0.attn_v.weight"],
+					"attn_output.weight": tensors["v.blk.0.attn_output.weight"],
+					"attn_norm.weight":   tensors["v.blk.0.attn_norm.weight"],
+					"ffn_down.weight":    tensors["v.blk.0.ffn_down.weight"],
+					"ffn_gate.weight":    tensors["v.blk.0.ffn_gate.weight"],
+					"ffn_up.weight":      tensors["v.blk.0.ffn_up.weight"],
+					"ffn_norm.weight":    tensors["v.blk.0.ffn_norm.weight"],
+				},
+				"v": {
+					"patch_embd.weight":    tensors["v.patch_embd.weight"],
+					"position_embd.gate":   tensors["v.position_embd.gate"],
+					"position_embd.weight": tensors["v.position_embd.weight"],
+				},
+			},
+		},
+		{
+			name:  "vision and text",
+			items: slices.Collect(maps.Values(tensors)),
+			want: map[string]Layer{
+				"blk.0": {
+					"attn_k.weight":      tensors["blk.0.attn_k.weight"],
+					"attn_q.weight":      tensors["blk.0.attn_q.weight"],
+					"attn_v.weight":      tensors["blk.0.attn_v.weight"],
+					"attn_output.weight": tensors["blk.0.attn_output.weight"],
+					"attn_norm.weight":   tensors["blk.0.attn_norm.weight"],
+					"ffn_down.weight":    tensors["blk.0.ffn_down.weight"],
+					"ffn_gate.weight":    tensors["blk.0.ffn_gate.weight"],
+					"ffn_up.weight":      tensors["blk.0.ffn_up.weight"],
+					"ffn_norm.weight":    tensors["blk.0.ffn_norm.weight"],
+				},
+				"token_embd":  {"weight": tensors["token_embd.weight"]},
+				"output_norm": {"weight": tensors["output_norm.weight"]},
+				"mm.0": {
+					"bias":   tensors["mm.0.bias"],
+					"weight": tensors["mm.0.weight"],
+				},
+				"v.blk.0": {
+					"attn_k.weight":      tensors["v.blk.0.attn_k.weight"],
+					"attn_q.weight":      tensors["v.blk.0.attn_q.weight"],
+					"attn_v.weight":      tensors["v.blk.0.attn_v.weight"],
+					"attn_output.weight": tensors["v.blk.0.attn_output.weight"],
+					"attn_norm.weight":   tensors["v.blk.0.attn_norm.weight"],
+					"ffn_down.weight":    tensors["v.blk.0.ffn_down.weight"],
+					"ffn_gate.weight":    tensors["v.blk.0.ffn_gate.weight"],
+					"ffn_up.weight":      tensors["v.blk.0.ffn_up.weight"],
+					"ffn_norm.weight":    tensors["v.blk.0.ffn_norm.weight"],
+				},
+				"v": {
+					"patch_embd.weight":    tensors["v.patch_embd.weight"],
+					"position_embd.gate":   tensors["v.position_embd.gate"],
+					"position_embd.weight": tensors["v.position_embd.weight"],
+				},
+			},
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			got := Tensors{items: tt.items}.GroupLayers()
+			if diff := cmp.Diff(got, tt.want); diff != "" {
+				t.Errorf("unexpected layers (-got +want):\n%s", diff)
+			}
+		})
+	}
+}

llm/gguf.go → fs/ggml/gguf.go

-package llm
+package ggml
 
 import (
 	"bytes"
@@ -8,10 +8,9 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"maps"
 	"slices"
 	"strings"
-
-	"golang.org/x/exp/maps"
 )
 
 type containerGGUF struct {
@@ -110,9 +109,9 @@ func (llm *gguf) KV() KV {
 	return llm.kv
 }
 
-func (llm *gguf) Tensors() *Tensors {
-	return &Tensors{
-		Items:  llm.tensors,
+func (llm *gguf) Tensors() Tensors {
+	return Tensors{
+		items:  llm.tensors,
 		Offset: llm.tensorOffset,
 	}
 }
@@ -523,7 +522,7 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 		return err
 	}
 
-	keys := maps.Keys(kv)
+	keys := slices.Collect(maps.Keys(kv))
 	slices.Sort(keys)
 
 	for _, key := range keys {

llm/filetype.go → fs/ggml/type.go

-package llm
+package ggml
 
 import "fmt"
 
@@ -98,10 +98,10 @@ func ParseFileType(s string) (fileType, error) {
 		return fileTypeIQ3_M, nil
 	case "IQ2_S":
 		return fileTypeIQ2_S, nil
-	case "IQ4_XS":
-		return fileTypeIQ4_XS, nil
 	case "IQ2_M":
 		return fileTypeIQ2_M, nil
+	case "IQ4_XS":
+		return fileTypeIQ4_XS, nil
 	case "IQ1_M":
 		return fileTypeIQ1_M, nil
 	case "BF16":

llm/ggla.go

-package llm
-
-import (
-	"encoding/binary"
-	"errors"
-	"io"
-	"slices"
-)
-
-type containerGGLA struct {
-	version uint32
-}
-
-func (c *containerGGLA) Name() string {
-	return "ggla"
-}
-
-func (c *containerGGLA) Decode(rs io.ReadSeeker) (model, error) {
-	if err := binary.Read(rs, binary.LittleEndian, &c.version); err != nil {
-		return nil, err
-	}
-
-	switch c.version {
-	case 1:
-	default:
-		return nil, errors.New("invalid version")
-	}
-
-	model := newGGLA(c)
-	err := model.decode(rs)
-	return model, err
-}
-
-type ggla struct {
-	*containerGGLA
-
-	kv      KV
-	tensors []*Tensor
-
-	tensorOffset uint64
-}
-
-func newGGLA(container *containerGGLA) *ggla {
-	return &ggla{
-		containerGGLA: container,
-		kv:            make(KV),
-	}
-}
-
-func (llm *ggla) KV() KV {
-	return llm.kv
-}
-
-func (llm *ggla) Tensors() *Tensors {
-	return &Tensors{
-		Items:  llm.tensors,
-		Offset: llm.tensorOffset,
-	}
-}
-
-func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
-	var r uint32
-	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
-		return err
-	}
-	llm.kv["r"] = r
-
-	var alpha uint32
-	if err := binary.Read(rs, binary.LittleEndian, &alpha); err != nil {
-		return err
-	}
-	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
-		}
-
-		var t Tensor
-		if err := binary.Read(rs, binary.LittleEndian, &t.Kind); err != nil {
-			return err
-		}
-
-		t.Shape = make([]uint64, dims)
-		for i := 0; uint32(i) < dims; i++ {
-			var shape32 uint32
-			if err := binary.Read(rs, binary.LittleEndian, &shape32); err != nil {
-				return err
-			}
-
-			t.Shape[i] = uint64(shape32)
-		}
-
-		// ggla tensor shape is reversed
-		// ref: https://github.com/ggerganov/llama.cpp/blob/29ae62d2ae163e2b68aa0ad3bf2ab4636de0c957/convert-lora-to-ggml.py#L44
-		slices.Reverse(t.Shape)
-
-		name := make([]byte, namesize)
-		if err := binary.Read(rs, binary.LittleEndian, &name); err != nil {
-			return err
-		}
-
-		t.Name = string(name)
-
-		offset, err := rs.Seek(0, io.SeekCurrent)
-		if err != nil {
-			return err
-		}
-
-		if _, err := rs.Seek((offset+31)&-32-offset, io.SeekCurrent); err != nil {
-			return err
-		}
-
-		offset, err = rs.Seek(0, io.SeekCurrent)
-		if err != nil {
-			return err
-		}
-
-		t.Offset = uint64(offset)
-
-		if _, err := rs.Seek(int64(t.Size()), io.SeekCurrent); err != nil {
-			return err
-		}
-
-		llm.tensors = append(llm.tensors, &t)
-	}
-}