fix ref

The numa flag may be having a performance impact on multi-socket systems with GPU loads

.github/workflows/release.yaml

@@ -269,7 +269,7 @@ jobs:
             dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
-          name: windows-cuda-deps
+          name: windows-cuda-deps-${{ matrix.cuda.version }}
           path: dist/deps/*
 
 
@@ -328,7 +328,10 @@ jobs:
           name: generate-windows-cuda-12
       - uses: actions/download-artifact@v4
         with:
-          name: windows-cuda-deps
+          name: windows-cuda-deps-11
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-12
       - uses: actions/download-artifact@v4
         with:
           name: windows-rocm-deps
@@ -469,7 +472,10 @@ jobs:
           merge-multiple: true
       - run: |
           ls -lh dist/
-          (cd dist; sha256sum * > sha256sum.txt)
+          (cd dist; find . -type f | xargs sha256sum > ../sha256sum.txt)
+          mv sha256sum.txt dist/
+          mv dist/linux-???64 .
+          mv dist/linux-amd64-rocm .
           cat dist/sha256sum.txt
       - name: Create or update Release
         run: |

Dockerfile

@@ -95,8 +95,8 @@ ARG AMDGPU_TARGETS
 ENV GOARCH amd64 
 RUN --mount=type=cache,target=/root/.ccache \
     OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
-RUN mkdir -p ../../dist/linux-amd64/lib/ollama && \
+RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
-    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64/lib/ollama && tar xf - )
+    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
 
 FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
 ARG CMAKE_VERSION

cmd/cmd.go

@@ -204,6 +204,12 @@ func tempZipFiles(path string) (string, error) {
 		// safetensors files might be unresolved git lfs references; skip if they are
 		// covers model-x-of-y.safetensors, model.fp32-x-of-y.safetensors, model.safetensors
 		files = append(files, st...)
+	} else if st, _ := glob(filepath.Join(path, "adapters.safetensors"), "application/octet-stream"); len(st) > 0 {
+		// covers adapters.safetensors
+		files = append(files, st...)
+	} else if st, _ := glob(filepath.Join(path, "adapter_model.safetensors"), "application/octet-stream"); len(st) > 0 {
+		// covers adapter_model.safetensors
+		files = append(files, st...)
 	} else if pt, _ := glob(filepath.Join(path, "pytorch_model*.bin"), "application/zip"); len(pt) > 0 {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers pytorch_model-x-of-y.bin, pytorch_model.fp32-x-of-y.bin, pytorch_model.bin
@@ -223,6 +229,14 @@ func tempZipFiles(path string) (string, error) {
 	}
 	files = append(files, js...)
 
+	// bert models require a nested config.json
+	// TODO(mxyng): merge this with the glob above
+	js, err = glob(filepath.Join(path, "**/*.json"), "text/plain")
+	if err != nil {
+		return "", err
+	}
+	files = append(files, js...)
+
 	if tks, _ := glob(filepath.Join(path, "tokenizer.model"), "application/octet-stream"); len(tks) > 0 {
 		// add tokenizer.model if it exists, tokenizer.json is automatically picked up by the previous glob
 		// tokenizer.model might be a unresolved git lfs reference; error if it is
@@ -252,6 +266,11 @@ func tempZipFiles(path string) (string, error) {
 			return "", err
 		}
 
+		zfi.Name, err = filepath.Rel(path, file)
+		if err != nil {
+			return "", err
+		}
+
 		zf, err := zipfile.CreateHeader(zfi)
 		if err != nil {
 			return "", err

convert/convert.go

@@ -7,16 +7,27 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
+	"strings"
 
 	"github.com/ollama/ollama/llm"
 )
 
-type Parameters struct {
+type ModelParameters struct {
 	Architectures []string `json:"architectures"`
 	VocabSize     uint32   `json:"vocab_size"`
 }
 
-func (Parameters) KV(t *Tokenizer) llm.KV {
+type AdapterParameters struct {
+	Alpha          uint32 `json:"lora_alpha"`
+	LoraLayers     uint32 `json:"lora_layers"`
+	LoraParameters struct {
+		Rank  uint32  `json:"rank"`
+		Alpha float32 `json:"alpha"`
+		Scale float32 `json:"scale"`
+	} `json:"lora_parameters"`
+}
+
+func (ModelParameters) KV(t *Tokenizer) llm.KV {
 	kv := llm.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
@@ -43,40 +54,119 @@ func (Parameters) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (Parameters) specialTokenTypes() []string {
+func (p AdapterParameters) KV() llm.KV {
+	var alpha float32
+	if p.LoraParameters.Alpha == 0 {
+		alpha = float32(p.Alpha)
+	} else {
+		alpha = p.LoraParameters.Alpha
+	}
+
+	kv := llm.KV{
+		"adapter.lora.alpha": alpha,
+		"adapter.type":       "lora",
+		"general.file_type":  uint32(1),
+		"general.type":       "adapter",
+		"general.version":    "v0.2",
+	}
+
+	return kv
+}
+
+func (ModelParameters) specialTokenTypes() []string {
 	return []string{
 		"bos", "eos", "unk", "sep", "pad", "cls", "mask",
 	}
 }
 
-func (Parameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
 	return llm.WriteGGUF(ws, kv, ts)
 }
 
-type Converter interface {
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
+}
+
+type ModelConverter 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
+	// 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
 
-	// 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 writes the model to the provided io.WriteSeeker
 	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }
 
+type moreParser interface {
+	parseMore(fs.FS) error
+}
+
+type AdapterConverter interface {
+	// KV maps parameters to LLM key-values
+	KV(llm.KV) llm.KV
+	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
+	Tensors([]Tensor) []llm.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
+}
+
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV llm.KV) error {
+	bts, err := fs.ReadFile(fsys, "adapter_config.json")
+	if err != nil {
+		return err
+	}
+
+	var p AdapterParameters
+	if err := json.Unmarshal(bts, &p); err != nil {
+		return err
+	}
+
+	arch, ok := baseKV["general.architecture"]
+	if !ok {
+		return errors.New("architecture not set for the base model")
+	}
+
+	var conv AdapterConverter
+	switch arch {
+	case "llama":
+		conv = &llamaAdapter{}
+	case "gemma2":
+		conv = &gemma2Adapter{}
+	default:
+		return errors.New("unsupported architecture")
+	}
+
+	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
+	if err != nil {
+		return err
+	}
+
+	if err := json.Unmarshal(bts, conv); err != nil {
+		return err
+	}
+
+	return conv.writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
+}
+
 // 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 {
+func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
 		return err
 	}
 
-	var p Parameters
+	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
 		return err
 	}
@@ -85,16 +175,20 @@ func Convert(fsys fs.FS, ws io.WriteSeeker) error {
 		return errors.New("unknown architecture")
 	}
 
-	var conv Converter
+	var conv ModelConverter
 	switch p.Architectures[0] {
 	case "LlamaForCausalLM", "MistralForCausalLM":
-		conv = &llama{}
+		conv = &llamaModel{}
 	case "MixtralForCausalLM":
-		conv = &mixtral{}
+		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
-		conv = &gemma{}
+		conv = &gemmaModel{}
+	case "Gemma2ForCausalLM":
+		conv = &gemma2Model{}
 	case "Phi3ForCausalLM":
-		conv = &phi3{}
+		conv = &phi3Model{}
+	case "BertModel":
+		conv = &bertModel{}
 	default:
 		return errors.New("unsupported architecture")
 	}
@@ -103,6 +197,12 @@ func Convert(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}
 
+	if t, ok := conv.(moreParser); ok {
+		if err := t.parseMore(fsys); err != nil {
+			return err
+		}
+	}
+
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
 		return err
@@ -119,7 +219,7 @@ func Convert(fsys fs.FS, ws io.WriteSeeker) error {
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
 
-	ts, err := parseTensors(fsys)
+	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
 		return err
 	}

convert/convert_bert.go

@@ -0,0 +1,174 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"io/fs"
+	"path/filepath"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type bertModel struct {
+	ModelParameters
+	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"`
+	LayerNormEPS          float32 `json:"layer_norm_eps"`
+	LayerNormEpsilon      float32 `json:"layer_norm_epsilon"`
+	NormEpsilon           float32 `json:"norm_epsilon"`
+
+	PoolingType uint32
+}
+
+var (
+	_ ModelConverter = (*bertModel)(nil)
+	_ moreParser     = (*bertModel)(nil)
+)
+
+func (p *bertModel) parseMore(fsys fs.FS) error {
+	bts, err := fs.ReadFile(fsys, "modules.json")
+	if err != nil {
+		return err
+	}
+
+	var modules []struct {
+		Type string `json:"type"`
+		Path string `json:"path"`
+	}
+
+	if err := json.Unmarshal(bts, &modules); err != nil {
+		return err
+	}
+
+	var pooling string
+	for _, m := range modules {
+		if m.Type == "sentence_transformers.models.Pooling" {
+			pooling = m.Path
+			break
+		}
+	}
+
+	if pooling != "" {
+		bts, err := fs.ReadFile(fsys, filepath.Join(pooling, "config.json"))
+		if err != nil {
+			return err
+		}
+
+		var pc struct {
+			PoolingModeCLSToken   bool `json:"pooling_mode_cls_token"`
+			PoolingModeMeanTokens bool `json:"pooling_mode_mean_tokens"`
+		}
+
+		if err := json.Unmarshal(bts, &pc); err != nil {
+			return err
+		}
+
+		if pc.PoolingModeMeanTokens {
+			p.PoolingType = 1
+		} else if pc.PoolingModeCLSToken {
+			p.PoolingType = 2
+		}
+	}
+
+	return nil
+}
+
+func (p *bertModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "bert"
+	kv["bert.attention.causal"] = false
+	kv["bert.pooling_type"] = p.PoolingType
+
+	kv["bert.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
+
+	if contextLength := cmp.Or(p.MaxPositionEmbeddings, p.NCtx); contextLength > 0 {
+		kv["bert.context_length"] = contextLength
+	}
+
+	if embeddingLength := cmp.Or(p.HiddenSize, p.NEmbd); embeddingLength > 0 {
+		kv["bert.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
+	}
+
+	if feedForwardLength := cmp.Or(p.IntermediateSize, p.NInner); feedForwardLength > 0 {
+		kv["bert.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner)
+	}
+
+	if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 {
+		kv["bert.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
+	}
+
+	if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon, p.NormEpsilon); layerNormEpsilon > 0 {
+		kv["bert.attention.layer_norm_epsilon"] = layerNormEpsilon
+	}
+
+	kv["tokenizer.ggml.model"] = "bert"
+	kv["tokenizer.ggml.token_type_count"] = uint32(2)
+
+	// convert to phantom space tokens
+	for i, e := range t.Tokens {
+		if strings.HasPrefix(e, "[") && strings.HasSuffix(e, "]") {
+			// noop
+		} else if strings.HasPrefix(e, "##") {
+			t.Tokens[i] = e[2:]
+		} else {
+			t.Tokens[i] = "\u2581" + e
+		}
+	}
+
+	kv["tokenizer.ggml.tokens"] = t.Tokens
+
+	return kv
+}
+
+func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		if slices.Contains([]string{
+			"embeddings.position_ids",
+			"pooler.dense.weight",
+			"pooler.dense.bias",
+		}, t.Name()) {
+			continue
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (bertModel) Replacements() []string {
+	return []string{
+		"encoder.layer", "blk",
+		"encoder.layers", "blk",
+		"embeddings.word_embeddings", "token_embd",
+		"embeddings.token_type_embeddings", "token_types",
+		"embeddings.LayerNorm", "token_embd_norm",
+		"embeddings.position_embeddings", "position_embd",
+		"attention.self.query", "attn_q",
+		"attention.self.key", "attn_k",
+		"attention.self.value", "attn_v",
+		"attention.output.dense", "attn_output",
+		"attention.output.LayerNorm", "attn_output_norm",
+		"intermediate.dense", "ffn_up",
+		"output.dense", "ffn_down",
+		"output.LayerNorm", "layer_output_norm",
+	}
+}

convert/convert_gemma.go

@@ -9,8 +9,8 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
-type gemma struct {
+type gemmaModel struct {
-	Parameters
+	ModelParameters
 	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
 	HiddenSize            uint32  `json:"hidden_size"`
 	HiddenLayers          uint32  `json:"num_hidden_layers"`
@@ -21,12 +21,11 @@ type gemma struct {
 	HeadDim               uint32  `json:"head_dim"`
 }
 
-var _ Converter = (*gemma)(nil)
+var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemma) KV(t *Tokenizer) llm.KV {
+func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
-	kv := p.Parameters.KV(t)
+	kv := p.ModelParameters.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
@@ -43,16 +42,15 @@ func (p *gemma) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *gemma) Tensors(ts []Tensor) []llm.Tensor {
+func (p *gemmaModel) Tensors(ts []Tensor) []llm.Tensor {
 	var out []llm.Tensor
 	for _, t := range ts {
-		name := p.tensorName(t.Name())
+		if strings.HasSuffix(t.Name(), "_norm.weight") {
-		if strings.HasSuffix(name, "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 
 		out = append(out, llm.Tensor{
-			Name:     name,
+			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
 			WriterTo: t,
@@ -62,8 +60,8 @@ func (p *gemma) Tensors(ts []Tensor) []llm.Tensor {
 	return out
 }
 
-func (p *gemma) tensorName(n string) string {
+func (p *gemmaModel) Replacements() []string {
-	return strings.NewReplacer(
+	return []string{
 		"model.embed_tokens", "token_embd",
 		"model.norm", "output_norm",
 		"model.layers", "blk",
@@ -76,11 +74,10 @@ func (p *gemma) tensorName(n string) string {
 		"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) {
+func (*gemmaModel) 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]))
 

convert/convert_gemma2.go

@@ -0,0 +1,43 @@
+package convert
+
+import (
+	"github.com/ollama/ollama/llm"
+)
+
+type gemma2Model struct {
+	gemmaModel
+	SlidingWindow         uint32  `json:"sliding_window"`
+	AttentionLogitSoftcap float32 `json:"attn_logit_softcapping"`
+	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
+}
+
+func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma2"
+	kv["gemma2.context_length"] = p.MaxPositionEmbeddings
+	kv["gemma2.embedding_length"] = p.HiddenSize
+	kv["gemma2.block_count"] = p.HiddenLayers
+	kv["gemma2.feed_forward_length"] = p.IntermediateSize
+	kv["gemma2.attention.head_count"] = p.NumAttentionHeads
+	kv["gemma2.attention.head_count_kv"] = p.NumKeyValueHeads
+	kv["gemma2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["gemma2.attention.key_length"] = p.HeadDim
+	kv["gemma2.attention.value_length"] = p.HeadDim
+	kv["gemma2.attention.sliding_window"] = p.SlidingWindow
+	kv["gemma2.attn_logit_softcapping"] = p.AttentionLogitSoftcap
+	kv["gemma2.final_logit_softcapping"] = p.FinalLogitSoftcap
+	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 *gemma2Model) Replacements() []string {
+	return append(
+		p.gemmaModel.Replacements(),
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+	)
+}

convert/convert_gemma2_adapter.go

@@ -0,0 +1,91 @@
+package convert
+
+import (
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type gemma2Adapter struct {
+	AdapterParameters
+}
+
+var _ AdapterConverter = (*gemma2Adapter)(nil)
+
+func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
+	kv := p.AdapterParameters.KV()
+	kv["general.architecture"] = "gemma2"
+	return kv
+}
+
+func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		shape := t.Shape()
+		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
+			(strings.HasSuffix(t.Name(), "weight.lora_b") && shape[0] < shape[1]) {
+			shape[0], shape[1] = shape[1], shape[0]
+			t.SetRepacker(p.repack)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *gemma2Adapter) Replacements() []string {
+	return []string{
+		"base_model.model.", "",
+		"model.layers", "blk",
+		"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",
+		"lora_A.weight", "weight.lora_a",
+		"lora_B.weight", "weight.lora_b",
+		"lora_a", "weight.lora_a",
+		"lora_b", "weight.lora_b",
+	}
+}
+
+func (p *gemma2Adapter) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	if err := n.T(1, 0); 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_llama.go

@@ -3,6 +3,7 @@ package convert
 import (
 	"cmp"
 	"fmt"
+	"math"
 	"strings"
 
 	"github.com/pdevine/tensor"
@@ -11,8 +12,8 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
-type llama struct {
+type llamaModel struct {
-	Parameters
+	ModelParameters
 	NLayers               uint32  `json:"n_layers"`
 	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
 	NLayer                uint32  `json:"n_layer"`
@@ -27,8 +28,14 @@ type llama struct {
 	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
 	RopeTheta             float32 `json:"rope_theta"`
 	RopeScaling           struct {
-		Type   string  `json:"type"`
+		Type                            string  `json:"type"`
-		Factor float32 `json:"factor"`
+		RopeType                        string  `json:"rope_type"`
+		Factor                          float32 `json:"factor"`
+		LowFrequencyFactor              float32 `json:"low_freq_factor"`
+		HighFrequencyFactor             float32 `json:"high_freq_factor"`
+		OriginalMaxPositionalEmbeddings uint32  `json:"original_max_positional_embeddings"`
+
+		factors ropeFactor
 	} `json:"rope_scaling"`
 	RMSNormEPS       float32 `json:"rms_norm_eps"`
 	LayerNormEPS     float32 `json:"layer_norm_eps"`
@@ -37,12 +44,11 @@ type llama struct {
 	HeadDim          uint32  `json:"head_dim"`
 }
 
-var _ Converter = (*llama)(nil)
+var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llama) KV(t *Tokenizer) llm.KV {
+func (p *llamaModel) KV(t *Tokenizer) llm.KV {
-	kv := p.Parameters.KV(t)
+	kv := p.ModelParameters.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)
@@ -71,6 +77,27 @@ func (p *llama) KV(t *Tokenizer) llm.KV {
 	if p.RopeScaling.Type == "linear" {
 		kv["llama.rope.scaling.type"] = p.RopeScaling.Type
 		kv["llama.rope.scaling.factor"] = p.RopeScaling.Factor
+	} else if p.RopeScaling.RopeType == "llama3" {
+		dim := p.HiddenSize / p.NumAttentionHeads
+		for i := uint32(0); i < dim; i += 2 {
+			factor := cmp.Or(p.RopeScaling.Factor, 8.0)
+			factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
+			factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
+
+			original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
+			lambdaLow := float32(original) / factorLow
+			lambdaHigh := float32(original) / factorHigh
+
+			lambda := 2 * math.Pi * math.Pow(float64(p.RopeTheta), float64(i)/float64(dim))
+			if lambda < float64(lambdaHigh) {
+				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0)
+			} else if lambda > float64(lambdaLow) {
+				p.RopeScaling.factors = append(p.RopeScaling.factors, factor)
+			} else {
+				smooth := (float32(original)/float32(lambda) - factorLow) / (factorHigh - factorLow)
+				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0/((1-smooth)/factor+smooth))
+			}
+		}
 	}
 
 	if p.NumKeyValueHeads > 0 {
@@ -93,17 +120,26 @@ func (p *llama) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *llama) Tensors(ts []Tensor) []llm.Tensor {
+func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
 	var out []llm.Tensor
+
+	if p.RopeScaling.factors != nil {
+		out = append(out, llm.Tensor{
+			Name:     "rope_freqs.weight",
+			Kind:     0,
+			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
+			WriterTo: p.RopeScaling.factors,
+		})
+	}
+
 	for _, t := range ts {
-		name := p.tensorName(t.Name())
+		if strings.HasSuffix(t.Name(), "attn_q.weight") ||
-		if strings.HasSuffix(name, "attn_q.weight") ||
+			strings.HasSuffix(t.Name(), "attn_k.weight") {
-			strings.HasSuffix(name, "attn_k.weight") {
 			t.SetRepacker(p.repack)
 		}
 
 		out = append(out, llm.Tensor{
-			Name:     name,
+			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
 			WriterTo: t,
@@ -113,8 +149,8 @@ func (p *llama) Tensors(ts []Tensor) []llm.Tensor {
 	return out
 }
 
-func (p *llama) tensorName(n string) string {
+func (p *llamaModel) Replacements() []string {
-	return strings.NewReplacer(
+	return []string{
 		"lm_head", "output",
 		"model.embed_tokens", "token_embd",
 		"model.norm", "output_norm",
@@ -128,21 +164,19 @@ func (p *llama) tensorName(n string) string {
 		"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) {
+func (p *llamaModel) 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") {
+	if strings.HasSuffix(name, "attn_q.weight") {
 		heads = p.NumAttentionHeads
-	} else if strings.HasSuffix(name, "k_proj.weight") {
+	} else if strings.HasSuffix(name, "attn_k.weight") {
 		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
 	} else {
 		return nil, fmt.Errorf("unknown tensor for repack: %s", name)

convert/convert_llama_adapter.go

@@ -0,0 +1,169 @@
+package convert
+
+import (
+	"cmp"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type llamaAdapter struct {
+	AdapterParameters
+	NumAttentionHeads uint32 `json:"num_attention_heads"`
+	NumKeyValueHeads  uint32 `json:"num_key_value_heads"`
+}
+
+var _ AdapterConverter = (*llamaAdapter)(nil)
+
+func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
+	kv := p.AdapterParameters.KV()
+	kv["general.architecture"] = "llama"
+	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
+	kv["llama.attention.head_count_kv"] = baseKV["llama.attention.head_count_kv"]
+
+	p.NumAttentionHeads = baseKV["llama.attention.head_count"].(uint32)
+
+	return kv
+}
+
+func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		shape := t.Shape()
+		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
+			(strings.HasSuffix(t.Name(), "weight.lora_b") && shape[0] < shape[1]) {
+			shape[0], shape[1] = shape[1], shape[0]
+			t.SetRepacker(p.repackAndTranspose)
+		} else {
+			t.SetRepacker(p.repack)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    shape,
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *llamaAdapter) Replacements() []string {
+	return []string{
+		"base_model.model.", "",
+		"model.layers", "blk",
+		"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",
+		"lora_A.weight", "weight.lora_a",
+		"lora_B.weight", "weight.lora_b",
+		"lora_a", "weight.lora_a",
+		"lora_b", "weight.lora_b",
+	}
+}
+
+func (p *llamaAdapter) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	var heads uint32
+	if strings.HasSuffix(name, "attn_q.weight.lora_a") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight.lora_a") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	} else {
+		return data, nil
+	}
+
+	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
+}
+
+func (p *llamaAdapter) repackAndTranspose(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	var heads uint32
+	if strings.HasSuffix(name, "attn_q.weight.lora_a") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight.lora_a") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	}
+
+	if heads > 0 {
+		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
+		}
+	}
+
+	if err := n.T(1, 0); 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

@@ -9,16 +9,14 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
-type mixtral struct {
+type mixtralModel struct {
-	llama
+	llamaModel
 	NumLocalExperts    uint32 `json:"num_local_experts"`
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-var _ Converter = (*mixtral)(nil)
+func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
-
+	kv := p.llamaModel.KV(t)
-func (p *mixtral) KV(t *Tokenizer) llm.KV {
-	kv := p.llama.KV(t)
 
 	if p.NumLocalExperts > 0 {
 		kv["llama.expert_count"] = p.NumLocalExperts
@@ -31,7 +29,7 @@ func (p *mixtral) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *mixtral) Tensors(ts []Tensor) []llm.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -69,7 +67,14 @@ func (p *mixtral) Tensors(ts []Tensor) []llm.Tensor {
 		})
 	}
 
-	return append(out, p.llama.Tensors(ts)...)
+	return append(out, p.llamaModel.Tensors(ts)...)
+}
+
+func (p *mixtralModel) Replacements() []string {
+	return append(
+		p.llamaModel.Replacements(),
+		"block_sparse_moe.gate", "ffn_gate_inp",
+	)
 }
 
 type experts []Tensor

convert/convert_phi3.go

@@ -11,8 +11,8 @@ import (
 	"github.com/ollama/ollama/llm"
 )
 
-type phi3 struct {
+type phi3Model struct {
-	Parameters
+	ModelParameters
 	NumHiddenLayers   uint32  `json:"num_hidden_layers"`
 	NLayers           uint32  `json:"n_layers"`
 	HiddenSize        uint32  `json:"hidden_size"`
@@ -35,12 +35,11 @@ type phi3 struct {
 	SlidingWindow                 uint32  `json:"sliding_window"`
 }
 
-var _ Converter = (*phi3)(nil)
+var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3) KV(t *Tokenizer) llm.KV {
+func (p *phi3Model) KV(t *Tokenizer) llm.KV {
-	kv := p.Parameters.KV(t)
+	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
-	kv["general.name"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings
 	kv["phi3.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
 	kv["phi3.feed_forward_length"] = p.IntermediateSize
@@ -69,13 +68,12 @@ func (p *phi3) KV(t *Tokenizer) llm.KV {
 	return kv
 }
 
-func (p *phi3) Tensors(ts []Tensor) []llm.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
 	var addRopeFactors sync.Once
 
 	out := make([]llm.Tensor, 0, len(ts)+2)
 	for _, t := range ts {
-		name := p.tensorName(t.Name())
+		if strings.HasPrefix(t.Name(), "blk.0.") {
-		if strings.HasPrefix(name, "blk.0.") {
 			addRopeFactors.Do(func() {
 				out = append(out, llm.Tensor{
 					Name:     "rope_factors_long.weight",
@@ -92,7 +90,7 @@ func (p *phi3) Tensors(ts []Tensor) []llm.Tensor {
 		}
 
 		out = append(out, llm.Tensor{
-			Name:     name,
+			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
 			WriterTo: t,
@@ -102,8 +100,8 @@ func (p *phi3) Tensors(ts []Tensor) []llm.Tensor {
 	return out
 }
 
-func (p *phi3) tensorName(n string) string {
+func (p *phi3Model) Replacements() []string {
-	return strings.NewReplacer(
+	return []string{
 		"lm_head", "output",
 		"model.embed_tokens", "token_embd",
 		"model.norm", "output_norm",
@@ -114,7 +112,7 @@ func (p *phi3) tensorName(n string) string {
 		"mlp.down_proj", "ffn_down",
 		"mlp.gate_up_proj", "ffn_up",
 		"post_attention_layernorm", "ffn_norm",
-	).Replace(n)
+	}
 }
 
 type ropeFactor []float32

convert/convert_test.go

@@ -1,7 +1,9 @@
 package convert
 
 import (
+	"bytes"
 	"crypto/sha256"
+	"encoding/binary"
 	"encoding/hex"
 	"encoding/json"
 	"flag"
@@ -29,7 +31,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	}
 	defer f.Close()
 
-	if err := Convert(fsys, f); err != nil {
+	if err := ConvertModel(fsys, f); err != nil {
 		t.Fatal(err)
 	}
 
@@ -51,6 +53,34 @@ 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 {
+	actual := make(map[string]string)
+	for k, v := range kv {
+		if s, ok := v.(json.Marshaler); !ok {
+			actual[k] = fmt.Sprintf("%v", v)
+		} else {
+			bts, err := json.Marshal(s)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts))
+		}
+	}
+
+	for _, tensor := range tensors.Items {
+		sha256sum := sha256.New()
+		sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
+		if _, err := io.Copy(sha256sum, sr); err != nil {
+			t.Fatal(err)
+		}
+
+		actual[tensor.Name] = hex.EncodeToString(sha256sum.Sum(nil))
+	}
+
+	return actual
+}
+
 func TestMain(m *testing.M) {
 	var level slog.Level
 	flag.TextVar(&level, "level", slog.LevelInfo, "log level")
@@ -62,11 +92,14 @@ func TestMain(m *testing.M) {
 func TestConvertFull(t *testing.T) {
 	cases := []string{
 		"Meta-Llama-3-8B-Instruct",
+		"Meta-Llama-3.1-8B-Instruct",
 		"Mistral-7B-Instruct-v0.2",
 		"Mixtral-8x7B-Instruct-v0.1",
 		"gemma-2b-it",
 		// microsoft/Phi-3-mini-128-instruct@d548c233192db00165d842bf8edff054bb3212f8
 		"Phi-3-mini-128k-instruct",
+		"all-MiniLM-L6-v2",
+		"gemma-2-9b-it",
 	}
 
 	for i := range cases {
@@ -82,29 +115,7 @@ func TestConvertFull(t *testing.T) {
 			}
 
 			f, kv, tensors := convertFull(t, os.DirFS(p))
-			actual := make(map[string]string)
+			actual := generateResultsJSON(t, f, kv, tensors)
-			for k, v := range kv {
-				if s, ok := v.(json.Marshaler); !ok {
-					actual[k] = fmt.Sprintf("%v", v)
-				} else {
-					bts, err := json.Marshal(s)
-					if err != nil {
-						t.Fatal(err)
-					}
-
-					actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts))
-				}
-			}
-
-			for _, tensor := range tensors.Items {
-				sha256sum := sha256.New()
-				sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
-				if _, err := io.Copy(sha256sum, sr); err != nil {
-					t.Fatal(err)
-				}
-
-				actual[tensor.Name] = hex.EncodeToString(sha256sum.Sum(nil))
-			}
 
 			expectFile, err := os.Open(filepath.Join("testdata", fmt.Sprintf("%s.json", tt)))
 			if err != nil {
@@ -128,3 +139,209 @@ func TestConvertFull(t *testing.T) {
 		})
 	}
 }
+
+func TestConvertAdapter(t *testing.T) {
+	type AdapterCase struct {
+		Name     string
+		BaseKV   map[string]any
+		Expected map[string]string
+	}
+
+	cases := []AdapterCase{
+		{
+			Name: "discollama",
+			BaseKV: map[string]any{
+				"general.architecture":          "llama",
+				"llama.attention.head_count":    uint32(32),
+				"llama.attention.head_count_kv": uint32(8),
+			},
+			Expected: map[string]string{
+				"general.architecture":          "llama",
+				"general.file_type":             "1",
+				"general.parameter_count":       "106496",
+				"general.type":                  "adapter",
+				"general.version":               "v0.2",
+				"adapter.lora.alpha":            "16",
+				"adapter.type":                  "lora",
+				"llama.attention.head_count":    "32",
+				"llama.attention.head_count_kv": "8",
+				"blk.31.attn_q.weight.lora_a":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_q.weight.lora_b":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_v.weight.lora_a":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_v.weight.lora_b":   "071dcafe89df065d6e1c935ecb8fdf6479b3c202eb912e7da938597673ff5857",
+			},
+		},
+	}
+
+	for _, c := range cases {
+		t.Run(c.Name, func(t *testing.T) {
+			t.Parallel()
+
+			f, err := os.CreateTemp(t.TempDir(), "f16")
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer f.Close()
+
+			tempDir := t.TempDir()
+			generateLoraTestData(t, tempDir)
+
+			if err = ConvertAdapter(os.DirFS(tempDir), f, c.BaseKV); err != nil {
+				t.Fatal(err)
+			}
+
+			r, err := os.Open(f.Name())
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer r.Close()
+
+			m, _, err := llm.DecodeGGML(r, math.MaxInt)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if _, err := r.Seek(0, io.SeekStart); err != nil {
+				t.Fatal(err)
+			}
+
+			actual := generateResultsJSON(t, r, m.KV(), m.Tensors())
+
+			keys := maps.Keys(c.Expected)
+			slices.Sort(keys)
+			for _, k := range keys {
+				if v, ok := actual[k]; !ok {
+					t.Errorf("missing %s", k)
+				} else if v != c.Expected[k] {
+					t.Errorf("unexpected %s: want %s, got %s", k, c.Expected[k], v)
+				}
+			}
+		})
+	}
+}
+
+func generateLoraTestData(t *testing.T, tempDir string) {
+	type tensorData struct {
+		Offsets []int  `json:"data_offsets"`
+		Type    string `json:"dtype"`
+		Shape   []int  `json:"shape"`
+	}
+	offset := 4096 * 8 * 4
+
+	td := map[string]*tensorData{"__metadata__": nil}
+	td["model.layers.31.self_attn.q_proj.lora_a"] = &tensorData{
+		Offsets: []int{0, offset},
+		Type:    "F32",
+		Shape:   []int{4096, 8},
+	}
+	td["model.layers.31.self_attn.q_proj.lora_b"] = &tensorData{
+		Offsets: []int{offset, offset * 2},
+		Type:    "F32",
+		Shape:   []int{8, 4096},
+	}
+	td["model.layers.31.self_attn.v_proj.lora_a"] = &tensorData{
+		Offsets: []int{offset * 2, offset * 3},
+		Type:    "F32",
+		Shape:   []int{4096, 8},
+	}
+	td["model.layers.31.self_attn.v_proj.lora_b"] = &tensorData{
+		Offsets: []int{offset * 3, offset*3 + 8*1024*4},
+		Type:    "F32",
+		Shape:   []int{8, 1024},
+	}
+
+	data, err := json.Marshal(td)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var buf bytes.Buffer
+
+	l := int64(len(data))
+	err = binary.Write(&buf, binary.LittleEndian, l)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = buf.Write(data)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// write some data for the tensors
+
+	ones := make([]float32, 4096*8)
+	for i := range ones {
+		ones[i] = float32(1)
+	}
+
+	for range 3 {
+		err = binary.Write(&buf, binary.LittleEndian, ones)
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	ones = make([]float32, 1024*8)
+	for i := range ones {
+		ones[i] = float32(1)
+	}
+
+	err = binary.Write(&buf, binary.LittleEndian, ones)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	fdata, err := os.Create(filepath.Join(tempDir, "adapters.safetensors"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer fdata.Close()
+
+	_, err = fdata.Write(buf.Bytes())
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	configData := `
+{
+    "adapter_path": "adapters-test",
+    "batch_size": 8,
+    "config": "config-tiny.json",
+    "data": "../discollama-completion",
+    "grad_checkpoint": null,
+    "iters": 1000,
+    "learning_rate": 1e-05,
+    "lora_layers": 1,
+    "lora_parameters": {
+        "rank": 8,
+        "alpha": 16,
+        "dropout": 0.0,
+        "scale": 2.0
+    },
+    "lr_schedule": null,
+    "max_seq_length": 2048,
+    "model": "/Users/pdevine/git/Meta-Llama-3-8B-Instruct",
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 200,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "use_dora": false,
+    "val_batches": 25
+}
+`
+	f, err := os.Create(filepath.Join(tempDir, "adapter_config.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	_, err = f.WriteString(configData)
+	if err != nil {
+		t.Fatal(err)
+	}
+}

convert/reader.go

@@ -35,7 +35,9 @@ const (
 )
 
 func (t tensorBase) Kind() uint32 {
-	if strings.HasSuffix(t.name, ".block_sparse_moe.gate.weight") {
+	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
+		t.name == "token_types.weight" {
+		// these tensors are always F32
 		return 0
 	}
 
@@ -55,13 +57,15 @@ func (t *tensorBase) SetRepacker(fn repacker) {
 
 type repacker func(string, []float32, []uint64) ([]float32, error)
 
-func parseTensors(fsys fs.FS) ([]Tensor, error) {
+func parseTensors(fsys fs.FS, replacer *strings.Replacer) ([]Tensor, error) {
 	patterns := []struct {
 		Pattern string
-		Func    func(fs.FS, ...string) ([]Tensor, error)
+		Func    func(fs.FS, *strings.Replacer, ...string) ([]Tensor, error)
 	}{
 		{"model-*-of-*.safetensors", parseSafetensors},
 		{"model.safetensors", parseSafetensors},
+		{"adapters.safetensors", parseSafetensors},
+		{"adapter_model.safetensors", parseSafetensors},
 		{"pytorch_model-*-of-*.bin", parseTorch},
 		{"pytorch_model.bin", parseTorch},
 		{"consolidated.*.pth", parseTorch},
@@ -74,7 +78,7 @@ func parseTensors(fsys fs.FS) ([]Tensor, error) {
 		}
 
 		if len(matches) > 0 {
-			return pattern.Func(fsys, matches...)
+			return pattern.Func(fsys, replacer, matches...)
 		}
 	}
 

convert/reader_safetensors.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"io/fs"
 	"slices"
+	"strings"
 
 	"github.com/d4l3k/go-bfloat16"
 	"github.com/x448/float16"
@@ -20,7 +21,7 @@ type safetensorMetadata struct {
 	Offsets []int64  `json:"data_offsets"`
 }
 
-func parseSafetensors(fsys fs.FS, ps ...string) ([]Tensor, error) {
+func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]Tensor, error) {
 	var ts []Tensor
 	for _, p := range ps {
 		f, err := fsys.Open(p)
@@ -56,7 +57,7 @@ func parseSafetensors(fsys fs.FS, ps ...string) ([]Tensor, error) {
 					offset: safetensorsPad(n, value.Offsets[0]),
 					size:   safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
 					tensorBase: &tensorBase{
-						name:  key,
+						name:  replacer.Replace(key),
 						shape: value.Shape,
 					},
 				})

convert/reader_torch.go

@@ -3,12 +3,13 @@ package convert
 import (
 	"io"
 	"io/fs"
+	"strings"
 
 	"github.com/nlpodyssey/gopickle/pytorch"
 	"github.com/nlpodyssey/gopickle/types"
 )
 
-func parseTorch(fsys fs.FS, ps ...string) ([]Tensor, error) {
+func parseTorch(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]Tensor, error) {
 	var ts []Tensor
 	for _, p := range ps {
 		pt, err := pytorch.Load(p)
@@ -27,7 +28,7 @@ func parseTorch(fsys fs.FS, ps ...string) ([]Tensor, error) {
 			ts = append(ts, torch{
 				storage: t.(*pytorch.Tensor).Source,
 				tensorBase: &tensorBase{
-					name:  k.(string),
+					name:  replacer.Replace(k.(string)),
 					shape: shape,
 				},
 			})

convert/testdata/Meta-Llama-3.1-8B-Instruct.json

@@ -0,0 +1,3 @@
+{
+  "rope_freqs.weight": "80fd5efb2f729381785b293a091a268cfeceb0079167f6ece9b07070e662b222"
+}

convert/testdata/all-MiniLM-L6-v2.json

@@ -0,0 +1,124 @@
+{
+  "general.architecture": "bert",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "bert.attention.causal": "false",
+  "bert.attention.head_count": "12",
+  "bert.attention.layer_norm_epsilon": "1e-12",
+  "bert.block_count": "6",
+  "bert.context_length": "512",
+  "bert.embedding_length": "384",
+  "bert.feed_forward_length": "1536",
+  "bert.pooling_type": "1",
+  "tokenizer.ggml.model": "bert",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "100",
+  "tokenizer.ggml.cls_token_id": "101",
+  "tokenizer.ggml.seperator_token_id": "102",
+  "tokenizer.ggml.mask_token_id": "103",
+  "tokenizer.ggml.token_type_count": "2",
+  "tokenizer.ggml.scores": "6db964fe67338aca57790481a390121ff3dd643eebe49f7dd308029ad99abb6f",
+  "tokenizer.ggml.token_type": "98d247c5404b6b18f05f133b92dd56edf6efefefac326794b00d7b351f6c5aa1",
+  "tokenizer.ggml.tokens": "9efe405e229a45ff9916f54c475d151d2200cd2ab0006f347abfb069cf096c86",
+  "token_embd.weight": "8c1ee80a9ea4f65aa385ba30112010068af3d209bebc6e149d3d4589c2cd0a5a",
+  "position_embd.weight": "6c516f0b1c4e2388ab90394dd80ad69e4e4509b890982fc3408108ae66210eb6",
+  "token_types.weight": "f879f8e422ed211948f28b560d3c5e17aae7993f063b51196a28cf5c0fb3da21",
+  "token_embd_norm.weight": "75076e095d717aab96f8b6beeee503c27940d9a76f2b891a0e3de72f8a6043e4",
+  "token_embd_norm.bias": "298735285ffe944e1bf03e5d35c7280326b85cf121bde9874f1af5dc51ab939d",
+  "blk.0.attn_q.weight": "ab0923ce4c1549175112dcdfcc860fe30137f991e03ea6857fb5993670adaf6c",
+  "blk.0.attn_q.bias": "a3ec29551dabf976e1d34256b8ab5ab7b758f3ed9742c3cafdbd984d5441df62",
+  "blk.0.attn_k.weight": "4c1038a6d035c3e9ffed7fa672b614627814752503755fbad0cfb76a41ad71ba",
+  "blk.0.attn_k.bias": "e0363930eb588d91816aa3d230bb03b6e2551c165117b80b8d60397413819ef9",
+  "blk.0.attn_v.weight": "425e2e53e3f00ce98d29c3e6a161eb55d3e6ae0d96fdb9f6242d1c4fd6eef4b3",
+  "blk.0.attn_v.bias": "6579173a1e65ee124fbd0bd53cbdca4225515b4f2c5f18fb1bfd000f5978f9bb",
+  "blk.0.attn_output.weight": "a6d70a08cd7164de5d12af65d86d657c3db35aaecde778b2b3fda9193c4c9802",
+  "blk.0.attn_output.bias": "2b8d12c4f9a9c5bfaa29c597839568f6e0525cb41eeaf64ddeb6bd84dfeb9701",
+  "blk.0.attn_output_norm.weight": "bbe6e502a473228b525aeed26cc31b7db123ad63bdc5a6eebac6ea70b8b51d62",
+  "blk.0.attn_output_norm.bias": "36eaacaf0007c5c62daea97aab0115390c0682914f78482e37eb76885f4b7a50",
+  "blk.0.ffn_up.weight": "24654561c76ce387d125759ba843f06b904ef721fcceaeff6ccc62180a48e874",
+  "blk.0.ffn_up.bias": "fd3f0126aa1d95768fa60eb6f4ab8a2763cfcb7e5405f35b92353031d86f4d34",
+  "blk.0.ffn_down.weight": "97a829763a6a5bf3329ceb4d39c424ba4787d61653a5b0bbd1f84782e4d4e0ca",
+  "blk.0.ffn_down.bias": "7aa980c30ae8b4ee7f69df28808dbf5c431f56ccc4a80340f644a0419f16c054",
+  "blk.0.layer_output_norm.weight": "ef30dad4c2a083ae1ff5039a2a6cda60ecc89bf1e486a6f8c0d15f50589603f8",
+  "blk.0.layer_output_norm.bias": "8b1b77e67568b1bce43fc476de1b177c53ff688d66beb66995e8eb3dc290da8a",
+  "blk.1.attn_q.weight": "284331622a1f6f9b87ccee4f652bd66a394ca493c4d93be4d1844e4f6159ad10",
+  "blk.1.attn_q.bias": "e24ebd4860330e08f6bfdd077a82db0bee33f4c8846cf1db26327a34754c7069",
+  "blk.1.attn_k.weight": "729dd0d555544b5bd0f7580b3c8b384256b974605f0e7487b95f295aa032997d",
+  "blk.1.attn_k.bias": "2aa51a828a858f35473f54477583fea54ce2ccc34ea60fbd1d228fbe9bca827f",
+  "blk.1.attn_v.weight": "6be304671cc311d5ca5c103f2b51467ee800c589bc5b8101e09ff5aed1f68c21",
+  "blk.1.attn_v.bias": "43bcbab78a8819e07f723bc9e5b737b71e87a7594f15234e882b63e327a64199",
+  "blk.1.attn_output.weight": "15ec8a1a12b26c9976445308a09f748ab0e4bef0f583d13ab08c3129f8738d73",
+  "blk.1.attn_output.bias": "dac2146f4baa6ed16f6c0dc7443831fb7ec79bedcceafd80d1a4b628a1bb072d",
+  "blk.1.attn_output_norm.weight": "d2151eb33bffac536787a4c9a5d2b31c7a80b17c4611877842a3cce2cd6e98d8",
+  "blk.1.attn_output_norm.bias": "31e1b779716dafb855d2cf5631ee168a0ccf372eb9c6ea6091f66fa97a9b9d2d",
+  "blk.1.ffn_up.weight": "a57547fc3fc3b77406f5cdcb0c87af9bc184701f175c39c1f35297826fce3cc7",
+  "blk.1.ffn_up.bias": "123be6d541d086202913c75d878c54d59a749f3af7b58f7ef9eb9e7c62a24c9a",
+  "blk.1.ffn_down.weight": "cfdb79788377e5cbded8790cd41b9e66c397ecab75474071fcd7cf32d30f9613",
+  "blk.1.ffn_down.bias": "bcb58315519a573097960891c9ae41cf4c685ab78c3e0e77471471758a7eae88",
+  "blk.1.layer_output_norm.weight": "819b554271452bfb1d84c2603b90377b2e41a0ac1e3aa8b417ccf9dce63375bd",
+  "blk.1.layer_output_norm.bias": "47a3433ac27f5ce8947fb38dd491f3706df4ef6adb0ddf74612bf0f54b19e164",
+  "blk.2.attn_q.weight": "1557a9ea852b1880551f7290e00aded4f35e6c4180fdcbed1b0039bf805f639e",
+  "blk.2.attn_q.bias": "c3bfe5f3066f655fd36b055530997b59ff33ef013563aaeb3cb8ff07dabd59a9",
+  "blk.2.attn_k.weight": "cfd08eb69c61ae2f9f14f9b7ff5c5394ca264b1a9f3d48156677f90dd1766289",
+  "blk.2.attn_k.bias": "9b839bc0e79974a0b3f5d1895972bc6f5c9a1bc16052e1af786e6a530758152d",
+  "blk.2.attn_v.weight": "02b26b1208480eaeeb00e7b4cf8b690006ca14759357fc44ed4a2a8924ead993",
+  "blk.2.attn_v.bias": "e7e6f0089fded1659a867ab736c220d9653ea7da6b1b94baf5c8d30a748b63ab",
+  "blk.2.attn_output.weight": "a1db121c7d33806b349cadd050300a57db49fdc91224fd07c9ac43bf4299dc79",
+  "blk.2.attn_output.bias": "7675128b6a92555cd955c820311e91e9417d31f48848f45d047b4100c62148b3",
+  "blk.2.attn_output_norm.weight": "5b4595e0fbcba67a700c4331adf746d2fba3546364a4db5607ae241947bb1a21",
+  "blk.2.attn_output_norm.bias": "7b8e16826ea30e5a2ba0b02e0095a901775981a296e98819625320e983060d08",
+  "blk.2.ffn_up.weight": "a0d815d946ac07a65095c4ae4df77b818845e6d97795c7d82f55e689d944db59",
+  "blk.2.ffn_up.bias": "ce37c0a4174d6bf773ded7bd016ede627ad3bdb8bc99b9992a18dc8e8898f252",
+  "blk.2.ffn_down.weight": "f6231d2a25426fbd45b9f1160aa484220eb227ceef0348c4a6a6de890606e5ef",
+  "blk.2.ffn_down.bias": "429e00556e8dc63a785238b309b9d83738500c1ef6d736fe6526ad88ea496d27",
+  "blk.2.layer_output_norm.weight": "651457a573adf3f7dd9ee5dfe1c8e89389e94443993aab77ec6a0b05aa621e35",
+  "blk.2.layer_output_norm.bias": "41fbbeda7fd89b0cef5f945ae44011c316982390401d6f75ba8c6d365e185247",
+  "blk.3.attn_q.weight": "95a43f32949d2cb8d22815bb27a44abfc6665ba96221af817dfe058cb6ca72c6",
+  "blk.3.attn_q.bias": "f4e34385e75d8108b6b3bd336106e2133a8c9be0cc343dfe5dc48c32a823c7cb",
+  "blk.3.attn_k.weight": "6b892da6a17d4d3265265a15f695864a31813ee8c8e710ae9bc9e1adbc6c9a18",
+  "blk.3.attn_k.bias": "40b8067b641a56014cee42548240aa8930820958b1933004892b5f04fbaef39e",
+  "blk.3.attn_v.weight": "9fcd5922319dd2a461082a5ce040c1dfe65d87d70ca6547dd0b46eeecc3eeb2b",
+  "blk.3.attn_v.bias": "b528c56212e66931fdbe267ac327a9c2f87cd03baff3ea719e30afe681da15f1",
+  "blk.3.attn_output.weight": "e3b178c1b03981e75510e0d277af23ea59cc404b5394e61bd32291825719b502",
+  "blk.3.attn_output.bias": "712c84d39a6a5a9c06a09da8fd9939ba0d5525524a4bba61ea4de09b48f45cae",
+  "blk.3.attn_output_norm.weight": "d1ffac88e675592ff72f8a617be32b4a381d443b2f8f2645dbe44a1e5745aac0",
+  "blk.3.attn_output_norm.bias": "ea31a1c73146234c50e0e43f485c458413714867b8e2703af66482f7db2d6c40",
+  "blk.3.ffn_up.weight": "4ef4f3b9a1ea6ab2ef2eb6e8b008e06a44790d099d97482a05a51e39a29afac0",
+  "blk.3.ffn_up.bias": "06a4296dda16f452675c51f108079fe7722552d6521c737d97734943818b9a2b",
+  "blk.3.ffn_down.weight": "f114b2bebe392c7d80433bb880c6730293aa4561b0b0370dcdaf7472daebd847",
+  "blk.3.ffn_down.bias": "2c8e67831d28a3bf613fc7912ae3259b63d72abcaf4d30efd8800758400158de",
+  "blk.3.layer_output_norm.weight": "a1dfeb7b5a51dd56447312ca41e2ad2f361a3ea12ddc355127f5f4219fb0a482",
+  "blk.3.layer_output_norm.bias": "1ed630021b25c6c6fc93fd32988b9907df966d4982a93081f639aac3044618ab",
+  "blk.4.attn_q.weight": "b5fae4c1f9a5f33a2a2e816ac0c01c25f422e4efdd59ef1ed93da2610e5370fc",
+  "blk.4.attn_q.bias": "c2e376524ea98ac3b10d9eee19ecb1b1e261fa5149efe0232844c923dfb428fb",
+  "blk.4.attn_k.weight": "a4632f5ebf9321d9d08f9112a4e5dda2efe5671df4a4e67fee24845f5b14af16",
+  "blk.4.attn_k.bias": "a9a02ffb8b8b4f6dfe487a7e0341f1d5318c9d2b793a688f34cb1b22fc66ef60",
+  "blk.4.attn_v.weight": "10ad8deb81d9fa093b1e5c0f24ea82aa7df43e6aca49e260fcbea56eab8cc86a",
+  "blk.4.attn_v.bias": "7326813e181e021130bd33ac136293fcffccce2d1d8cb59041e5b13a8cceacf6",
+  "blk.4.attn_output.weight": "c92573088c7437c2b3cda51490e152c27fb19e5468df591eabba5a49d5398d44",
+  "blk.4.attn_output.bias": "14e10b419e5859af1eb685af5c330aee67048cd704dcead9217840c6f5393222",
+  "blk.4.attn_output_norm.weight": "02b6831c0e0fb0edbc579a92812a1dd972cb15d14fcd382d4427c5a7b300ac44",
+  "blk.4.attn_output_norm.bias": "7eed5cd503bb6bb6ceb1bc8b07cc077903a4f14fb8b9d6cdf39644815ecf1374",
+  "blk.4.ffn_up.weight": "8d0c91d62e74d6431321116a37cf3339e630bd50ba164d3304fc4fe8dd831223",
+  "blk.4.ffn_up.bias": "d325f07f73c005a273c484c7be8e7abb4d6e8a5c4fd093f5869133b97629d017",
+  "blk.4.ffn_down.weight": "7ba7bd81143f40537b84f938e403e19f30e4928625eb371de052b9025beb4d21",
+  "blk.4.ffn_down.bias": "2853d9c2a75288214a4bf4907dc19d04d01926f4913d302b1aa7bdbfcce0f7a1",
+  "blk.4.layer_output_norm.weight": "a4ed1885fa77b90fed5300c355ef0aa0c876a8c747151d9d790939d464d57d4f",
+  "blk.4.layer_output_norm.bias": "62142a81e813a9e636333b2b805d6bc3b17c5e7cd4b15adce1ada6bc9a32563c",
+  "blk.5.attn_q.weight": "afc1dff080a72c3daad01384b1448d476aaf789871017c8ff8e144788887995d",
+  "blk.5.attn_q.bias": "748a820371c1d4f872c84545b36358d239c35bf6c99e2812c237d88c3292763b",
+  "blk.5.attn_k.weight": "59e30c1ed8acd2cbb01de5f62e7804015b9ecf98ba157d98cab016344639eda5",
+  "blk.5.attn_k.bias": "f839520078f9e589496e982e86d0126c7aa14196047339abffcf49a696229f77",
+  "blk.5.attn_v.weight": "3e21fb874e21b90308e1f46af034a3c32d3eba1628d62ae5f2246d6af5818923",
+  "blk.5.attn_v.bias": "5cd4852bf95c1444d10d756750f6bf49f842c0b39e9953c7f408bb67c325ac8c",
+  "blk.5.attn_output.weight": "636ce6a7752895f204b9d01ba0aedd9a294f908b42f372c22a16d9dd590d7471",
+  "blk.5.attn_output.bias": "82d924d4b0d2b94f2bbff91619216d6967a3541ce9b1531a6a60457a67b5d219",
+  "blk.5.attn_output_norm.weight": "5e7bd0a8d3396080f3360d7c4700bf094a06216431bd014c4479eef72ecf4271",
+  "blk.5.attn_output_norm.bias": "66c6de5edda5466d029c6753780be81ccd4218bf8bc00680000e0f06856ab712",
+  "blk.5.ffn_up.weight": "5bbf6e7ea380e216e33f8bee06d25f2265359d3876a300e92bc6e41d48e33430",
+  "blk.5.ffn_up.bias": "9d795388bb36fb33ad3a37fea3ccb4937838e02800a608fb47d363cd06b47370",
+  "blk.5.ffn_down.weight": "2fd628974e7f075479dd227b46fbd48ae8d3ca34d735b36f391ac06410730368",
+  "blk.5.ffn_down.bias": "cd213ba9eaa75fa541648097fbe9c96e58077e6c3ad6ad2fb1f21f8350f44291",
+  "blk.5.layer_output_norm.weight": "159a9df41d15b7022d136f86a2a2631c4635f9816e957472217077b522bcf52a",
+  "blk.5.layer_output_norm.bias": "24c1f27ffd1eb4e5be7e3a2909943e6f0980635d761fa1efdd0c19645da23766"
+}

convert/testdata/gemma-2-9b-it.json

@@ -0,0 +1,6 @@
+{
+  "general.architecture": "gemma2",
+  "gemma2.attention.sliding_window": "4096",
+  "gemma2.attn_logit_softcapping": "50",
+  "gemma2.final_logit_softcapping": "30"
+}

convert/tokenizer.go

@@ -1,7 +1,6 @@
 package convert
 
 import (
-	"cmp"
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
@@ -11,6 +10,8 @@ import (
 	"log/slog"
 	"os"
 	"slices"
+
+	"golang.org/x/exp/maps"
 )
 
 const (
@@ -184,32 +185,32 @@ func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) {
 		return nil, err
 	}
 
-	var tokens []token
+	tokens := make(map[int]token, len(t.Model.Vocab))
 	for k, v := range t.Model.Vocab {
-		tokens = append(tokens, token{
+		tokens[v] = token{
 			ID:      v,
 			Content: k,
-		})
+		}
 	}
 
-	for _, t := range t.AddedTokens {
+	for _, token := range t.AddedTokens {
-		t.UserDefined = true
+		token.UserDefined = true
-		tokens = append(tokens, t)
+		tokens[token.ID] = token
 	}
 
-	slices.SortFunc(tokens, func(i, j token) int {
+	keys := maps.Keys(tokens)
-		return cmp.Compare(i.ID, j.ID)
+	slices.Sort(keys)
-	})
 
 	v := Vocabulary{Model: "gpt2"}
-	for _, t := range tokens {
+	for _, k := range keys {
-		v.Tokens = append(v.Tokens, t.Content)
+		token := tokens[k]
-		v.Scores = append(v.Scores, float32(t.ID))
+		v.Tokens = append(v.Tokens, token.Content)
+		v.Scores = append(v.Scores, float32(token.ID))
 
 		switch {
-		case t.Special:
+		case token.Special:
 			v.Types = append(v.Types, tokenTypeControl)
-		case t.UserDefined:
+		case token.UserDefined:
 			v.Types = append(v.Types, tokenTypeUserDefined)
 		default:
 			v.Types = append(v.Types, tokenTypeNormal)

convert/tokenizer_spm.go

@@ -15,6 +15,11 @@ import (
 )
 
 func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
+	ast, err := parseAdditionalSpecialTokens(fsys)
+	if err != nil {
+		return nil, err
+	}
+
 	bts, err := fs.ReadFile(fsys, "tokenizer.model")
 	if err != nil {
 		return nil, err
@@ -37,7 +42,12 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 			sentencepiece.ModelProto_SentencePiece_BYTE:
 			v.Types = append(v.Types, int32(t))
 		default:
-			v.Types = append(v.Types, int32(sentencepiece.ModelProto_SentencePiece_NORMAL))
+			tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
+			if slices.Contains(ast, piece.GetPiece()) {
+				tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
+			}
+
+			v.Types = append(v.Types, tt)
 		}
 	}
 
@@ -81,3 +91,23 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 
 	return &v, nil
 }
+
+func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
+	f, err := fsys.Open("special_tokens_map.json")
+	if errors.Is(err, os.ErrNotExist) {
+		return nil, nil
+	} else if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var m struct {
+		AdditionalSpecialTokens []string `json:"additional_special_tokens"`
+	}
+
+	if err := json.NewDecoder(f).Decode(&m); err != nil {
+		return nil, err
+	}
+
+	return m.AdditionalSpecialTokens, nil
+}

docs/faq.md

@@ -111,7 +111,10 @@ On Windows, Ollama inherits your user and system environment variables.
 
 ## How do I use Ollama behind a proxy?
 
-Ollama is compatible with proxy servers if `HTTP_PROXY` or `HTTPS_PROXY` are configured. When using either variables, ensure it is set where `ollama serve` can access the values. When using `HTTPS_PROXY`, ensure the proxy certificate is installed as a system certificate. Refer to the section above for how to use environment variables on your platform.
+Ollama pulls models from the Internet and may require a proxy server to access the models. Use `HTTPS_PROXY` to redirect outbound requests through the proxy. Ensure the proxy certificate is installed as a system certificate. Refer to the section above for how to use environment variables on your platform.
+
+> [!NOTE]
+> Avoid setting `HTTP_PROXY`. Ollama does not use HTTP for model pulls, only HTTPS. Setting `HTTP_PROXY` may interrupt client connections to the server.
 
 ### How do I use Ollama behind a proxy in Docker?
 
@@ -276,4 +279,4 @@ Note: Windows with Radeon GPUs currently default to 1 model maximum due to limit
 
 ## How does Ollama load models on multiple GPUs?
 
-Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.
+Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.

docs/import.md

@@ -1,44 +1,129 @@
-# Import
+# Importing a model
 
-GGUF models and select Safetensors models can be imported directly into Ollama.
+## Table of Contents
 
-## Import GGUF
+  * [Importing a Safetensors adapter](#Importing-a-fine-tuned-adapter-from-Safetensors-weights)
+  * [Importing a Safetensors model](#Importing-a-model-from-Safetensors-weights)
+  * [Importing a GGUF file](#Importing-a-GGUF-based-model-or-adapter)
+  * [Sharing models on ollama.com](#Sharing-your-model-on-ollamacom)
 
-A binary GGUF file can be imported directly into Ollama through a Modelfile.
+## Importing a fine tuned adapter from Safetensors weights
+
+First, create a `Modelfile` with a `FROM` command pointing at the base model you used for fine tuning, and an `ADAPTER` command which points to the directory with your Safetensors adapter:
 
 ```dockerfile
-FROM /path/to/file.gguf
+FROM <base model name>
+ADAPTER /path/to/safetensors/adapter/directory
 ```
 
-## Import Safetensors
+Make sure that you use the same base model in the `FROM` command as you used to create the adapter otherwise you will get erratic results. Most frameworks use different quantization methods, so it's best to use non-quantized (i.e. non-QLoRA) adapters. If your adapter is in the same directory as your `Modelfile`, use `ADAPTER .` to specify the adapter path.
 
-If the model being imported is one of these architectures, it can be imported directly into Ollama through a Modelfile:
+Now run `ollama create` from the directory where the `Modelfile` was created:
 
- - LlamaForCausalLM
+```bash
- - MistralForCausalLM
+ollama create my-model
- - MixtralForCausalLM
+```
- - GemmaForCausalLM
+
- - Phi3ForCausalLM
+Lastly, test the model:
+
+```bash
+ollama run my-model
+```
+
+Ollama supports importing adapters based on several different model architectures including:
+
+  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral); and
+  * Gemma (including Gemma 1 and Gemma 2)
+
+You can create the adapter using a fine tuning framework or tool which can output adapters in the Safetensors format, such as:
+
+  * Hugging Face [fine tuning framework] (https://huggingface.co/docs/transformers/en/training)
+  * [Unsloth](https://github.com/unslothai/unsloth)
+  * [MLX](https://github.com/ml-explore/mlx)
+
+
+## Importing a model from Safetensors weights
+
+First, create a `Modelfile` with a `FROM` command which points to the directory containing your Safetensors weights:
 
 ```dockerfile
 FROM /path/to/safetensors/directory
 ```
 
-For architectures not directly convertable by Ollama, see llama.cpp's [guide](https://github.com/ggerganov/llama.cpp/blob/master/README.md#prepare-and-quantize) on conversion. After conversion, see [Import GGUF](#import-gguf).
+If you create the Modelfile in the same directory as the weights, you can use the command `FROM .`.
 
-## Automatic Quantization
+Now run the `ollama create` command from the directory where you created the `Modelfile`:
 
-> [!NOTE]
+```shell
-> Automatic quantization requires v0.1.35 or higher.
+ollama create my-model
+```
 
-Ollama is capable of quantizing FP16 or FP32 models to any of the supported quantizations with the `-q/--quantize` flag in `ollama create`.
+Lastly, test the model:
+
+```shell
+ollama run my-model
+```
+
+Ollama supports importing models for several different architectures including:
+
+  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral);
+  * Gemma (including Gemma 1 and Gemma 2); and
+  * Phi3
+
+This includes importing foundation models as well as any fine tuned models which which have been _fused_ with a foundation model.
+
+
+## Importing a GGUF based model or adapter
+
+If you have a GGUF based model or adapter it is possible to import it into Ollama. You can obtain a GGUF model or adapter by:
+
+  * converting a Safetensors model with the `convert_hf_to_gguf.py` from Llama.cpp; 
+  * converting a Safetensors adapter with the `convert_lora_to_gguf.py` from Llama.cpp; or
+  * downloading a model or adapter from a place such as HuggingFace
+
+To import a GGUF model, create a `Modelfile` containg:
+
+```dockerfile
+FROM /path/to/file.gguf
+```
+
+For a GGUF adapter, create the `Modelfile` with:
+
+```dockerfile
+FROM <model name>
+ADAPTER /path/to/file.gguf
+```
+
+When importing a GGUF adapter, it's important to use the same base model as the base model that the adapter was created with. You can use:
+
+ * a model from Ollama
+ * a GGUF file
+ * a Safetensors based model 
+
+Once you have created your `Modelfile`, use the `ollama create` command to build the model.
+
+```shell
+ollama create my-model
+```
+
+## Quantizing a Model
+
+Quantizing a model allows you to run models faster and with less memory consumption but at reduced accuracy. This allows you to run a model on more modest hardware.
+
+Ollama can quantize FP16 and FP32 based models into different quantization levels using the `-q/--quantize` flag with the `ollama create` command.
+
+First, create a Modelfile with the FP16 or FP32 based model you wish to quantize.
 
 ```dockerfile
 FROM /path/to/my/gemma/f16/model
 ```
 
+Use `ollama create` to then create the quantized model.
+
 ```shell
-$ ollama create -q Q4_K_M mymodel
+$ ollama create --quantize q4_K_M mymodel
 transferring model data
 quantizing F16 model to Q4_K_M
 creating new layer sha256:735e246cc1abfd06e9cdcf95504d6789a6cd1ad7577108a70d9902fef503c1bd
@@ -49,42 +134,53 @@ success
 
 ### Supported Quantizations
 
-- `Q4_0`
+- `q4_0`
-- `Q4_1`
+- `q4_1`
-- `Q5_0`
+- `q5_0`
-- `Q5_1`
+- `q5_1`
-- `Q8_0`
+- `q8_0`
 
 #### K-means Quantizations
 
-- `Q3_K_S`
+- `q3_K_S`
-- `Q3_K_M`
+- `q3_K_M`
-- `Q3_K_L`
+- `q3_K_L`
-- `Q4_K_S`
+- `q4_K_S`
-- `Q4_K_M`
+- `q4_K_M`
-- `Q5_K_S`
+- `q5_K_S`
-- `Q5_K_M`
+- `q5_K_M`
-- `Q6_K`
+- `q6_K`
 
-## Template Detection
 
-> [!NOTE]
+## Sharing your model on ollama.com
-> Template detection requires v0.1.42 or higher.
 
-Ollama uses model metadata, specifically `tokenizer.chat_template`, to automatically create a template appropriate for the model you're importing.
+You can share any model you have created by pushing it to [ollama.com](https://ollama.com) so that other users can try it out.
 
-```dockerfile
+First, use your browser to go to the [Ollama Sign-Up](https://ollama.com/signup) page. If you already have an account, you can skip this step.
-FROM /path/to/my/gemma/model
+
-```
+![Sign-Up](images/signup.png)
+
+The `Username` field will be used as part of your model's name (e.g. `jmorganca/mymodel`), so make sure you are comfortable with the username that you have selected.
+
+Now that you have created an account and are signed-in, go to the [Ollama Keys Settings](https://ollama.com/settings/keys) page.
+
+Follow the directions on the page to determine where your Ollama Public Key is located.
+
+![Ollama Key](images/ollama-keys.png)
+
+Click on the `Add Ollama Public Key` button, and copy and paste the contents of your Ollama Public Key into the text field.
+
+To push a model to [ollama.com](https://ollama.com), first make sure that it is named correctly with your username. You may have to use the `ollama cp` command to copy
+your model to give it the correct name. Once you're happy with your model's name, use the `ollama push` command to push it to [ollama.com](https://ollama.com).
 
 ```shell
-$ ollama create mymodel
+ollama cp mymodel myuser/mymodel
-transferring model data
+ollama push myuser/mymodel
-using autodetected template gemma-instruct
+```
-creating new layer sha256:baa2a0edc27d19cc6b7537578a9a7ba1a4e3214dc185ed5ae43692b319af7b84
+
-creating new layer sha256:ba66c3309914dbef07e5149a648fd1877f030d337a4f240d444ea335008943cb
+Once your model has been pushed, other users can pull and run it by using the command:
-writing manifest
+
-success
+```shell
+ollama run myuser/mymodel
 ```
 
-Defining a template in the Modelfile will disable this feature which may be useful if you want to use a different template than the autodetected one.

gpu/gpu.go

@@ -264,6 +264,8 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.computeMajor = int(memInfo.major)
 				gpuInfo.computeMinor = int(memInfo.minor)
 				gpuInfo.MinimumMemory = cudaMinimumMemory
+				gpuInfo.DriverMajor = driverMajor
+				gpuInfo.DriverMinor = driverMinor
 				variant := cudaVariant(gpuInfo)
 				if depPath != "" {
 					gpuInfo.DependencyPath = depPath
@@ -275,8 +277,6 @@ func GetGPUInfo() GpuInfoList {
 					}
 				}
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-				gpuInfo.DriverMajor = driverMajor
-				gpuInfo.DriverMinor = driverMinor
 				gpuInfo.Variant = variant
 
 				// query the management library as well so we can record any skew between the two

gpu/gpu_test.go

@@ -32,4 +32,29 @@ func TestCPUMemInfo(t *testing.T) {
 	}
 }
 
+func TestByLibrary(t *testing.T) {
+	type testCase struct {
+		input  []GpuInfo
+		expect int
+	}
+
+	testCases := map[string]*testCase{
+		"empty":                    {input: []GpuInfo{}, expect: 0},
+		"cpu":                      {input: []GpuInfo{{Library: "cpu"}}, expect: 1},
+		"cpu + GPU":                {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}}, expect: 2},
+		"cpu + 2 GPU no variant":   {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}, {Library: "cuda"}}, expect: 2},
+		"cpu + 2 GPU same variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v11"}}, expect: 2},
+		"cpu + 2 GPU diff variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v12"}}, expect: 3},
+	}
+
+	for k, v := range testCases {
+		t.Run(k, func(t *testing.T) {
+			resp := (GpuInfoList)(v.input).ByLibrary()
+			if len(resp) != v.expect {
+				t.Fatalf("expected length %d, got %d => %+v", v.expect, len(resp), resp)
+			}
+		})
+	}
+}
+
 // TODO - add some logic to figure out card type through other means and actually verify we got back what we expected

gpu/types.go

@@ -94,7 +94,7 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 			}
 		}
 		if !found {
-			libs = append(libs, info.Library)
+			libs = append(libs, requested)
 			resp = append(resp, []GpuInfo{info})
 		}
 	}

integration/embed_test.go

@@ -70,8 +70,8 @@ func TestAllMiniLMEmbed(t *testing.T) {
 		t.Fatalf("expected 0.010071031, got %.8f", res.Embeddings[0][0])
 	}
 
-	if res.PromptEvalCount != 8 {
+	if res.PromptEvalCount != 6 {
-		t.Fatalf("expected 8 prompt tokens, got %d", res.PromptEvalCount)
+		t.Fatalf("expected 6 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 
@@ -102,8 +102,8 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 		t.Fatalf("expected 0.010071031 and -0.009802706, got %.8f and %.8f", res.Embeddings[0][0], res.Embeddings[1][0])
 	}
 
-	if res.PromptEvalCount != 16 {
+	if res.PromptEvalCount != 12 {
-		t.Fatalf("expected 16 prompt tokens, got %d", res.PromptEvalCount)
+		t.Fatalf("expected 12 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 

llm/ext_server/server.cpp

@@ -1429,7 +1429,13 @@ struct llama_server_context
         switch (task.type)
         {
             case TASK_TYPE_COMPLETION: {
-                server_slot *slot = prefix_slot(task.data["prompt"]);
+                server_slot *slot = nullptr;
+                if (task.embedding_mode) {
+                    // Embedding seq_id (aka slot id) must always be <= token length, so always use slot 0
+                    slot = slots[0].available() ? &slots[0] : nullptr;
+                } else {
+                    slot = prefix_slot(task.data["prompt"]);
+                }
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later

llm/generate/gen_linux.sh

@@ -252,7 +252,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
     fi
     init_vars
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DLLAMA_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DGGML_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
@@ -260,7 +260,8 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         echo "Building custom ROCM GPU"
     fi
     BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
-    ROCM_DIST_DIR="${DIST_BASE}/lib/ollama"
+    # ROCm dependencies are too large to fit into a unified bundle
+    ROCM_DIST_DIR="${DIST_BASE}/../linux-${GOARCH}-rocm/lib/ollama"
     # TODO figure out how to disable runpath (rpath)
     # export CMAKE_HIP_FLAGS="-fno-rtlib-add-rpath" # doesn't work
     export LLAMA_SERVER_LDFLAGS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"

llm/generate/gen_windows.ps1

@@ -355,7 +355,7 @@ function build_rocm() {
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
             "-DGGML_HIPBLAS=on",
-            "-DLLAMA_CUDA_NO_PEER_COPY=on",
+            "-DGGML_CUDA_NO_PEER_COPY=on",
             "-DHIP_PLATFORM=amd",
             "-DGGML_AVX=on",
             "-DGGML_AVX2=off",

llm/ggml.go

@@ -43,6 +43,14 @@ func (kv KV) Architecture() string {
 	return "unknown"
 }
 
+func (kv KV) Kind() string {
+	if s, ok := kv["general.type"].(string); ok {
+		return s
+	}
+
+	return "unknown"
+}
+
 func (kv KV) ParameterCount() uint64 {
 	return kv.u64("general.parameter_count")
 }

llm/memory_test.go

@@ -33,7 +33,6 @@ func TestEstimateGPULayers(t *testing.T) {
 	assert.Len(t, tensors, inputLayerCount+1)
 	err = WriteGGUF(f, KV{
 		"general.architecture":          "llama",
-		"general.name":                  "name",
 		"llama.context_length":          uint32(32),
 		"llama.embedding_length":        uint32(4096),
 		"llama.block_count":             uint32(inputLayerCount),

llm/patches/08-pooling.diff

@@ -1,60 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 721b8f4e..cfe7ac40 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -8420,14 +8420,14 @@ struct llm_build_context {
-     }
- 
-     struct ggml_tensor * build_inp_mean() {
--        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
-+        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, cparams.n_seq_max);
-         cb(lctx.inp_mean, "inp_mean", -1);
-         ggml_set_input(lctx.inp_mean);
-         return lctx.inp_mean;
-     }
- 
-     struct ggml_tensor * build_inp_cls() {
--        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
-+        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_seq_max);
-         cb(lctx.inp_cls, "inp_cls", -1);
-         ggml_set_input(lctx.inp_cls);
-         return lctx.inp_cls;
-@@ -13847,19 +13847,16 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
-         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
- 
-         float * data = (float *) lctx.inp_mean->data;
--        memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean));
-+        memset(lctx.inp_mean->data, 0, n_tokens * cparams.n_seq_max * ggml_element_size(lctx.inp_mean));
- 
-         std::vector<uint64_t> sum(n_tokens, 0);
-         for (int i = 0; i < n_tokens; ++i) {
-             const llama_seq_id seq_id = batch.seq_id[i][0];
--
--            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
--
-             sum[seq_id] += 1;
-         }
- 
--        std::vector<float> div(n_tokens, 0.0f);
--        for (int i = 0; i < n_tokens; ++i) {
-+        std::vector<float> div(cparams.n_seq_max, 0.0f);
-+        for (uint32_t i = 0; i < cparams.n_seq_max; ++i) {
-             const uint64_t s = sum[i];
-             if (s > 0) {
-                 div[i] = 1.0f/float(s);
-@@ -13879,14 +13876,11 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
-         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
- 
-         uint32_t * data = (uint32_t *) lctx.inp_cls->data;
--        memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
-+        memset(lctx.inp_cls->data, 0, cparams.n_seq_max * ggml_element_size(lctx.inp_cls));
- 
-         for (int i = 0; i < n_tokens; ++i) {
-             const llama_seq_id seq_id = batch.seq_id[i][0];
-             const llama_pos    pos    = batch.pos[i];
--
--            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS");
--
-             if (pos == 0) {
-                 data[seq_id] = i;
-             }

llm/server.go

@@ -258,7 +258,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--mlock")
 	}
 
-	if gpu.IsNUMA() {
+	if gpu.IsNUMA() && gpus[0].Library == "cpu" {
 		numaMode := "distribute"
 		if runtime.GOOS == "linux" {
 			if _, err := exec.LookPath("numactl"); err == nil {

scripts/build_linux.sh

@@ -24,8 +24,14 @@ for TARGETARCH in ${BUILD_ARCH}; do
     docker create --platform linux/$TARGETARCH --name builder-$TARGETARCH builder:$TARGETARCH
     rm -rf ./dist/linux-$TARGETARCH
     docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/linux-$TARGETARCH ./dist
+    if echo ${TARGETARCH} | grep "amd64" > /dev/null; then
+        docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/linux-$TARGETARCH-rocm ./dist
+    fi
     docker rm builder-$TARGETARCH
     echo "Compressing final linux bundle..."
     rm -f ./dist/ollama-linux-$TARGETARCH.tgz
     (cd dist/linux-$TARGETARCH && tar cf - . | ${GZIP} --best > ../ollama-linux-$TARGETARCH.tgz )
+    if [ -d dist/linux-$TARGETARCH-rocm ]; then
+        (cd dist/linux-$TARGETARCH-rocm && tar cf - . | ${GZIP} --best > ../ollama-linux-$TARGETARCH-rocm.tgz )
+    fi
 done

scripts/install.sh

@@ -199,6 +199,11 @@ fi
 
 if check_gpu lspci amdgpu || check_gpu lshw amdgpu; then
     if [ $BUNDLE -ne 0 ]; then
+        status "Downloading Linux ROCm ${ARCH} bundle"
+        curl --fail --show-error --location --progress-bar \
+            "https://ollama.com/download/ollama-linux-${ARCH}-rocm.tgz${VER_PARAM}" | \
+            $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+
         install_success
         status "AMD GPU ready."
         exit 0

server/images.go

@@ -369,13 +369,14 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 	parameters := make(map[string]any)
 
 	var layers []Layer
+	var baseLayers []*layerGGML
 	for _, c := range modelfile.Commands {
 		mediatype := fmt.Sprintf("application/vnd.ollama.image.%s", c.Name)
+		command := c.Name
 
-		switch c.Name {
+		switch command {
 		case "model", "adapter":
-			var baseLayers []*layerGGML
+			if name := model.ParseName(c.Args); name.IsValid() && command == "model" {
-			if name := model.ParseName(c.Args); name.IsValid() {
 				baseLayers, err = parseFromModel(ctx, name, fn)
 				if err != nil {
 					return err
@@ -409,14 +410,14 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 				}
 				defer blob.Close()
 
-				baseLayers, err = parseFromFile(ctx, blob, digest, fn)
+				baseLayers, err = parseFromFile(ctx, command, baseLayers, blob, digest, fn)
 				if err != nil {
 					return err
 				}
 			} else if file, err := os.Open(realpath(modelFileDir, c.Args)); err == nil {
 				defer file.Close()
 
-				baseLayers, err = parseFromFile(ctx, file, "", fn)
+				baseLayers, err = parseFromFile(ctx, command, baseLayers, file, "", fn)
 				if err != nil {
 					return err
 				}

server/layer.go

@@ -51,6 +51,9 @@ func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 		if err := os.Rename(temp.Name(), blob); err != nil {
 			return Layer{}, err
 		}
+		if err := os.Chmod(blob, 0o644); err != nil {
+			return Layer{}, err
+		}
 	}
 
 	return Layer{

server/model.go

@@ -81,7 +81,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	return layers, nil
 }
 
-func parseFromZipFile(_ context.Context, f *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+func parseFromZipFile(_ context.Context, command string, baseLayers []*layerGGML, f *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
 	fi, err := f.Stat()
 	if err != nil {
 		return nil, err
@@ -108,16 +108,38 @@ func parseFromZipFile(_ context.Context, f *os.File, digest string, fn func(api.
 	defer t.Close()
 	defer os.Remove(t.Name())
 
-	fn(api.ProgressResponse{Status: "converting model"})
+	var layerType string
-	if err := convert.Convert(convert.NewZipReader(r, p, 32<<20), t); err != nil {
+
-		return nil, err
+	switch command {
+	case "adapter":
+		var baseModel *llm.GGML
+		for _, l := range baseLayers {
+			if l.GGML != nil {
+				baseModel = l.GGML
+				break
+			}
+		}
+
+		if baseModel == nil {
+			return nil, fmt.Errorf("no base model specified for the adapter")
+		}
+
+		if err := convert.ConvertAdapter(convert.NewZipReader(r, p, 32<<20), t, baseModel.KV()); err != nil {
+			return nil, err
+		}
+		layerType = "application/vnd.ollama.image.adapter"
+	case "model":
+		if err := convert.ConvertModel(convert.NewZipReader(r, p, 32<<20), t); err != nil {
+			return nil, err
+		}
+		layerType = "application/vnd.ollama.image.model"
 	}
 
 	if _, err := t.Seek(0, io.SeekStart); err != nil {
 		return nil, err
 	}
 
-	layer, err := NewLayer(t, "application/vnd.ollama.image.model")
+	layer, err := NewLayer(t, layerType)
 	if err != nil {
 		return nil, err
 	}
@@ -139,7 +161,7 @@ func parseFromZipFile(_ context.Context, f *os.File, digest string, fn func(api.
 	return detectChatTemplate(layers)
 }
 
-func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+func parseFromFile(ctx context.Context, command string, baseLayers []*layerGGML, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
 	sr := io.NewSectionReader(file, 0, 512)
 	contentType, err := detectContentType(sr)
 	if err != nil {
@@ -150,7 +172,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 	case "gguf", "ggla":
 		// noop
 	case "application/zip":
-		return parseFromZipFile(ctx, file, digest, fn)
+		return parseFromZipFile(ctx, command, baseLayers, file, digest, fn)
 	default:
 		return nil, fmt.Errorf("unsupported content type: %s", contentType)
 	}
@@ -170,7 +192,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 		}
 
 		mediatype := "application/vnd.ollama.image.model"
-		if ggml.Name() == "ggla" {
+		if ggml.Name() == "ggla" || ggml.KV().Kind() == "adapter" {
 			mediatype = "application/vnd.ollama.image.adapter"
 		} else if ggml.KV().Architecture() == "clip" {
 			mediatype = "application/vnd.ollama.image.projector"

server/model_test.go

@@ -153,7 +153,7 @@ func TestParseFromFileFromLayer(t *testing.T) {
 		t.Fatalf("failed to seek to start: %v", err)
 	}
 
-	layers, err := parseFromFile(context.Background(), file, "", func(api.ProgressResponse) {})
+	layers, err := parseFromFile(context.Background(), "model", []*layerGGML{}, file, "", func(api.ProgressResponse) {})
 	if err != nil {
 		t.Fatalf("failed to parse from file: %v", err)
 	}
@@ -166,7 +166,7 @@ func TestParseFromFileFromLayer(t *testing.T) {
 		t.Fatalf("failed to seek to start: %v", err)
 	}
 
-	layers2, err := parseFromFile(context.Background(), file, layers[0].Digest, func(api.ProgressResponse) {})
+	layers2, err := parseFromFile(context.Background(), "model", []*layerGGML{}, file, layers[0].Digest, func(api.ProgressResponse) {})
 	if err != nil {
 		t.Fatalf("failed to parse from file: %v", err)
 	}
@@ -206,7 +206,7 @@ func TestParseLayerFromCopy(t *testing.T) {
 		t.Fatalf("failed to seek to start: %v", err)
 	}
 
-	layers, err := parseFromFile(context.Background(), file2, "", func(api.ProgressResponse) {})
+	layers, err := parseFromFile(context.Background(), "model", []*layerGGML{}, file2, "", func(api.ProgressResponse) {})
 	if err != nil {
 		t.Fatalf("failed to parse from file: %v", err)
 	}

server/sched.go

@@ -193,6 +193,11 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						break
 					}
 
+					// Embedding models should always be loaded with parallel=1
+					if pending.model.CheckCapabilities(CapabilityCompletion) != nil {
+						numParallel = 1
+					}
+
 					// Evaluate if the model will fit in the available system memory, or if we should unload a model first
 					if len(gpus) == 1 && gpus[0].Library == "cpu" {
 						// simplifying assumption of defaultParallel when in CPU mode

server/sched_test.go

@@ -117,7 +117,6 @@ func newScenarioRequest(t *testing.T, ctx context.Context, modelName string, est
 
 	require.NoError(t, llm.WriteGGUF(f, llm.KV{
 		"general.architecture":          "llama",
-		"general.name":                  "name",
 		"llama.context_length":          uint32(32),
 		"llama.embedding_length":        uint32(4096),
 		"llama.block_count":             uint32(1),