Merge https://github.com/ollama/ollama

Signed-off-by: baalajimaestro <me@baalajimaestro.me>
2024-08-25 22:02:07 +05:30 · 2024-08-25 22:02:07 +05:30 · 0c61920bc9
commit 0c61920bc9
parent 99dfb67553 0f92b19bec
59 changed files with 1647 additions and 471 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@ -187,6 +187,13 @@ jobs:
  generate-windows-cuda:
    environment: release
    runs-on: windows
+    strategy:
+      matrix:
+        cuda:
+          - version: "11"
+            url: 'https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe'
+          - version: "12"
+            url: 'https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe'
    env:
      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
    steps:
@ -220,11 +227,11 @@ jobs:
        with:
          go-version-file: go.mod
          cache: true
-      - name: 'Install CUDA'
+      - name: 'Install CUDA ${{ matrix.cuda.version }}'
        run: |
          $ErrorActionPreference = "Stop"
          write-host "downloading CUDA Installer"
-          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          Invoke-WebRequest -Uri "${{ matrix.cuda.url }}" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
          write-host "Installing CUDA"
          Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
          write-host "Completed CUDA"
@ -256,15 +263,16 @@ jobs:
          cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
      - uses: actions/upload-artifact@v4
        with:
-          name: generate-windows-cuda
+          name: generate-windows-cuda-${{ matrix.cuda.version }}
          path: |
            llm/build/**/bin/*
            dist/windows-amd64/**
      - uses: actions/upload-artifact@v4
        with:
-          name: windows-cuda-deps
+          name: windows-cuda-deps-${{ matrix.cuda.version }}
          path: dist/deps/*

+
  # Import the prior generation steps and build the final windows assets
  build-windows:
    environment: release
@ -314,10 +322,16 @@ jobs:
          name: generate-windows-cpu
      - uses: actions/download-artifact@v4
        with:
-          name: generate-windows-cuda
+          name: generate-windows-cuda-11
      - uses: actions/download-artifact@v4
        with:
-          name: windows-cuda-deps
+          name: generate-windows-cuda-12
+      - uses: actions/download-artifact@v4
+        with:
+          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
@ -363,7 +377,6 @@ jobs:
      - run: |
          ./scripts/build_linux.sh
          ./scripts/build_docker.sh
-          mv dist/deps/* dist/
      - uses: actions/upload-artifact@v4
        with:
          name: dist-linux-amd64
@ -459,7 +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: |
--- a/app/ollama.iss
+++ b/app/ollama.iss
@ -87,20 +87,11 @@ DialogFontSize=12

 [Files]
 Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
-Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs
+Source: "..\ollama.exe"; DestDir: "{app}\bin"; Flags: ignoreversion 64bit
+Source: "..\dist\windows-{#ARCH}\lib\ollama\runners\*"; DestDir: "{app}\lib\ollama\runners"; Flags: ignoreversion 64bit recursesubdirs
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-#if DirExists("..\dist\windows-amd64\cuda")
-  Source: "..\dist\windows-amd64\cuda\*"; DestDir: "{app}\cuda\"; Flags: ignoreversion recursesubdirs
-#endif
-#if DirExists("..\dist\windows-amd64\oneapi")
-  Source: "..\dist\windows-amd64\oneapi\*"; DestDir: "{app}\oneapi\"; Flags: ignoreversion recursesubdirs
-#endif
-#if DirExists("..\dist\windows-amd64\rocm")
-  Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
-#endif
-
+Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Flags: ignoreversion recursesubdirs

 [Icons]
 Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
@ -108,7 +99,7 @@ Name: "{userstartup}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilen
 Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"

 [Run]
-Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
+Filename: "{cmd}"; Parameters: "/C set PATH={app}\bin;%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden

 [UninstallRun]
 ; Filename: "{cmd}"; Parameters: "/C ""taskkill /im ''{#MyAppExeName}'' /f /t"; Flags: runhidden
@ -143,8 +134,8 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi

 [Registry]
 Root: HKCU; Subkey: "Environment"; \
-    ValueType: expandsz; ValueName: "Path"; ValueData: "{olddata};{app}"; \
-    Check: NeedsAddPath('{app}')
+    ValueType: expandsz; ValueName: "Path"; ValueData: "{olddata};{app}\bin"; \
+    Check: NeedsAddPath('{app}\bin')

 [Code]

--- a/app/tray/wintray/tray.go
+++ b/app/tray/wintray/tray.go
@ -11,6 +11,7 @@ import (
 	"path/filepath"
 	"sort"
 	"sync"
+	"syscall"
 	"unsafe"

 	"golang.org/x/sys/windows"
@ -433,7 +434,12 @@ func (t *winTray) setIcon(src string) error {
 	t.muNID.Lock()
 	defer t.muNID.Unlock()
 	t.nid.Icon = h
-	t.nid.Flags |= NIF_ICON
+	t.nid.Flags |= NIF_ICON | NIF_TIP
+	if toolTipUTF16, err := syscall.UTF16FromString(commontray.ToolTip); err == nil {
+		copy(t.nid.Tip[:], toolTipUTF16)
+	} else {
+		return err
+	}
 	t.nid.Size = uint32(unsafe.Sizeof(*t.nid))

 	return t.nid.modify()
--- a/app/tray/wintray/w32api.go
+++ b/app/tray/wintray/w32api.go
@ -61,6 +61,7 @@ const (
 	MIIM_SUBMENU        = 0x00000004
 	MIM_APPLYTOSUBMENUS = 0x80000000
 	NIF_ICON            = 0x00000002
+	NIF_TIP             = 0x00000004
 	NIF_INFO            = 0x00000010
 	NIF_MESSAGE         = 0x00000001
 	SW_HIDE             = 0
--- a/cmd/cmd.go
+++ b/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
--- a/convert/convert.go
+++ b/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
 	}
--- a/convert/convert_bert.go
+++ b/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",
+	}
+}
--- a/convert/convert_gemma.go
+++ b/convert/convert_gemma.go
@ -9,8 +9,8 @@ import (
 	"github.com/ollama/ollama/llm"
 )

-type gemma struct {
-	Parameters
+type gemmaModel struct {
+	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 {
-	kv := p.Parameters.KV(t)
+func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
+	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(name, "_norm.weight") {
+		if strings.HasSuffix(t.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 {
-	return strings.NewReplacer(
+func (p *gemmaModel) Replacements() []string {
+	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]))

--- a/convert/convert_gemma2.go
+++ b/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",
+	)
+}
--- a/convert/convert_gemma2_adapter.go
+++ b/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
+}
--- a/convert/convert_llama.go
+++ b/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 {
-	Parameters
+type llamaModel struct {
+	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"`
-		Factor float32 `json:"factor"`
+		Type                            string  `json:"type"`
+		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 {
-	kv := p.Parameters.KV(t)
+func (p *llamaModel) KV(t *Tokenizer) llm.KV {
+	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(name, "attn_q.weight") ||
-			strings.HasSuffix(name, "attn_k.weight") {
+		if strings.HasSuffix(t.Name(), "attn_q.weight") ||
+			strings.HasSuffix(t.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 {
-	return strings.NewReplacer(
+func (p *llamaModel) Replacements() []string {
+	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)
--- a/convert/convert_llama_adapter.go
+++ b/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
+}
--- a/convert/convert_mixtral.go
+++ b/convert/convert_mixtral.go
@ -9,16 +9,14 @@ import (
 	"github.com/ollama/ollama/llm"
 )

-type mixtral struct {
-	llama
+type mixtralModel struct {
+	llamaModel
 	NumLocalExperts    uint32 `json:"num_local_experts"`
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }

-var _ Converter = (*mixtral)(nil)
-
-func (p *mixtral) KV(t *Tokenizer) llm.KV {
-	kv := p.llama.KV(t)
+func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
+	kv := p.llamaModel.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
--- a/convert/convert_phi3.go
+++ b/convert/convert_phi3.go
@ -11,8 +11,8 @@ import (
 	"github.com/ollama/ollama/llm"
 )

-type phi3 struct {
-	Parameters
+type phi3Model struct {
+	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 {
-	kv := p.Parameters.KV(t)
+func (p *phi3Model) KV(t *Tokenizer) llm.KV {
+	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(name, "blk.0.") {
+		if strings.HasPrefix(t.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 {
-	return strings.NewReplacer(
+func (p *phi3Model) Replacements() []string {
+	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
--- a/convert/convert_test.go
+++ b/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)
-			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))
-			}
+			actual := generateResultsJSON(t, f, kv, tensors)

 			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)
+	}
+}
--- a/convert/reader.go
+++ b/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...)
 		}
 	}

--- a/convert/reader_safetensors.go
+++ b/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,
 					},
 				})
--- a/convert/reader_torch.go
+++ b/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,
 				},
 			})
--- a/convert/testdata/Meta-Llama-3.1-8B-Instruct.json
+++ b/convert/testdata/Meta-Llama-3.1-8B-Instruct.json
@ -0,0 +1,3 @@
+{
+  "rope_freqs.weight": "80fd5efb2f729381785b293a091a268cfeceb0079167f6ece9b07070e662b222"
+}
--- a/convert/testdata/all-MiniLM-L6-v2.json
+++ b/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"
+}
--- a/convert/testdata/gemma-2-9b-it.json
+++ b/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"
+}
--- a/convert/tokenizer.go
+++ b/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 {
-		t.UserDefined = true
-		tokens = append(tokens, t)
+	for _, token := range t.AddedTokens {
+		token.UserDefined = true
+		tokens[token.ID] = token
 	}

-	slices.SortFunc(tokens, func(i, j token) int {
-		return cmp.Compare(i.ID, j.ID)
-	})
+	keys := maps.Keys(tokens)
+	slices.Sort(keys)

 	v := Vocabulary{Model: "gpt2"}
-	for _, t := range tokens {
-		v.Tokens = append(v.Tokens, t.Content)
-		v.Scores = append(v.Scores, float32(t.ID))
+	for _, k := range keys {
+		token := tokens[k]
+		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)
--- a/convert/tokenizer_spm.go
+++ b/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
+}
--- a/docs/faq.md
+++ b/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?

--- a/docs/linux.md
+++ b/docs/linux.md
@ -20,13 +20,12 @@ GPU.

 ## Manual install

-### Download the `ollama` binary
+### Download `ollama`

-Ollama is distributed as a self-contained binary. Download it to a directory in your PATH:
+Download and extract the Linux package:

 ```bash
-sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
-sudo chmod +x /usr/bin/ollama
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
 ```

 ### Adding Ollama as a startup service (recommended)
@ -96,8 +95,7 @@ curl -fsSL https://ollama.com/install.sh | sh
 Or by downloading the ollama binary:

 ```bash
-sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
-sudo chmod +x /usr/bin/ollama
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
 ```

 ## Installing specific versions
--- a/envconfig/config.go
+++ b/envconfig/config.go
@ -174,7 +174,7 @@ func RunnersDir() (p string) {

 	defer func() {
 		if p == "" {
-			slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
+			slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama/runners'")
 		}
 	}()

@ -190,17 +190,17 @@ func RunnersDir() (p string) {
 	}

 	var paths []string
-	for _, root := range []string{filepath.Dir(exe), cwd} {
+	for _, root := range []string{filepath.Dir(exe), filepath.Join(filepath.Dir(exe), ".."), cwd} {
 		paths = append(paths,
 			root,
-			filepath.Join(root, "windows-"+runtime.GOARCH),
-			filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
+			filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH),
+			filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH),
 		)
 	}

 	// Try a few variations to improve developer experience when building from source in the local tree
 	for _, path := range paths {
-		candidate := filepath.Join(path, "ollama_runners")
+		candidate := filepath.Join(path, "lib", "ollama", "runners")
 		if _, err := os.Stat(candidate); err == nil {
 			p = candidate
 			break
--- a/gpu/amd_common.go
+++ b/gpu/amd_common.go
@ -54,7 +54,7 @@ func commonAMDValidateLibDir() (string, error) {
 	// Installer payload location if we're running the installed binary
 	exe, err := os.Executable()
 	if err == nil {
-		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), "..", "lib", "ollama")
 		if rocmLibUsable(rocmTargetDir) {
 			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
 			return rocmTargetDir, nil
--- a/gpu/amd_windows.go
+++ b/gpu/amd_windows.go
@ -153,7 +153,7 @@ func AMDValidateLibDir() (string, error) {
 	// Installer payload (if we're running from some other location)
 	localAppData := os.Getenv("LOCALAPPDATA")
 	appDir := filepath.Join(localAppData, "Programs", "Ollama")
-	rocmTargetDir := filepath.Join(appDir, "rocm")
+	rocmTargetDir := filepath.Join(appDir, "..", "lib", "ollama")
 	if rocmLibUsable(rocmTargetDir) {
 		slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
 		return rocmTargetDir, nil
--- a/gpu/cuda_common.go
+++ b/gpu/cuda_common.go
@ -4,9 +4,17 @@ package gpu

 import (
 	"log/slog"
+	"os"
+	"regexp"
+	"runtime"
+	"strconv"
 	"strings"
 )

+// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
+// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
+var CudaTegra string = os.Getenv("JETSON_JETPACK")
+
 func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	ids := []string{}
 	for _, info := range gpuInfo {
@ -19,3 +27,38 @@ func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	}
 	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
 }
+
+func cudaVariant(gpuInfo CudaGPUInfo) string {
+	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
+		if CudaTegra != "" {
+			ver := strings.Split(CudaTegra, ".")
+			if len(ver) > 0 {
+				return "jetpack" + ver[0]
+			}
+		} else if data, err := os.ReadFile("/etc/nv_tegra_release"); err == nil {
+			r := regexp.MustCompile(` R(\d+) `)
+			m := r.FindSubmatch(data)
+			if len(m) != 2 {
+				slog.Info("Unexpected format for /etc/nv_tegra_release.  Set JETSON_JETPACK to select version")
+			} else {
+				if l4t, err := strconv.Atoi(string(m[1])); err == nil {
+					// Note: mapping from L4t -> JP is inconsistent (can't just subtract 30)
+					// https://developer.nvidia.com/embedded/jetpack-archive
+					switch l4t {
+					case 35:
+						return "jetpack5"
+					case 36:
+						return "jetpack6"
+					default:
+						slog.Info("unsupported L4T version", "nv_tegra_release", string(data))
+					}
+				}
+			}
+		}
+	}
+
+	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 {
+		return "v11"
+	}
+	return "v12"
+}
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@ -64,10 +64,6 @@ var RocmComputeMin = 9
 // TODO find a better way to detect iGPU instead of minimum memory
 const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU

-// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
-// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
-var CudaTegra string = os.Getenv("JETSON_JETPACK")
-
 // Note: gpuMutex must already be held
 func initCudaHandles() *cudaHandles {
 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
@ -215,7 +211,7 @@ func GetGPUInfo() GpuInfoList {
 				GpuInfo: GpuInfo{
 					memInfo: mem,
 					Library: "cpu",
-					Variant: cpuCapability,
+					Variant: cpuCapability.String(),
 					ID:      "0",
 				},
 			},
@ -229,11 +225,7 @@ func GetGPUInfo() GpuInfoList {
 			return GpuInfoList{cpus[0].GpuInfo}
 		}

-		// On windows we bundle the nvidia library one level above the runner dir
-		depPath := ""
-		if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
-			depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "cuda")
-		}
+		depPath := LibraryDir()

 		// Load ALL libraries
 		cHandles = initCudaHandles()
@ -269,11 +261,23 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.FreeMemory = uint64(memInfo.free)
 				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 				gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
+				gpuInfo.computeMajor = int(memInfo.major)
+				gpuInfo.computeMinor = int(memInfo.minor)
 				gpuInfo.MinimumMemory = cudaMinimumMemory
-				gpuInfo.DependencyPath = depPath
-				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
+				variant := cudaVariant(gpuInfo)
+				if depPath != "" {
+					gpuInfo.DependencyPath = depPath
+					// Check for variant specific directory
+					if variant != "" {
+						if _, err := os.Stat(filepath.Join(depPath, "cuda_"+variant)); err == nil {
+							gpuInfo.DependencyPath = filepath.Join(depPath, "cuda_"+variant)
+						}
+					}
+				}
+				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+				gpuInfo.Variant = variant

 				// query the management library as well so we can record any skew between the two
 				// which represents overhead on the GPU we must set aside on subsequent updates
@ -306,13 +310,6 @@ func GetGPUInfo() GpuInfoList {
 		if envconfig.IntelGPU() {
 			oHandles = initOneAPIHandles()
 			if oHandles != nil && oHandles.oneapi != nil {
-
-				// On windows we bundle the oneapi library one level above the runner dir
-				depPath = ""
-				if runtime.GOOS == "windows" && envconfig.RunnersDir() != "" {
-					depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir()), "oneapi")
-				}
-
 				for d := range oHandles.oneapi.num_drivers {
 					if oHandles.oneapi == nil {
 						// shouldn't happen
@ -467,10 +464,12 @@ func GetGPUInfo() GpuInfoList {
 func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
 	var ldPaths []string
-	var patterns []string
 	gpuLibPaths := []string{}
 	slog.Debug("Searching for GPU library", "name", baseLibName)

+	// Start with our bundled libraries
+	patterns := []string{filepath.Join(LibraryDir(), baseLibName)}
+
 	switch runtime.GOOS {
 	case "windows":
 		ldPaths = strings.Split(os.Getenv("PATH"), ";")
@ -479,13 +478,14 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	default:
 		return gpuLibPaths
 	}
-	// Start with whatever we find in the PATH/LD_LIBRARY_PATH
+
+	// Then with whatever we find in the PATH/LD_LIBRARY_PATH
 	for _, ldPath := range ldPaths {
 		d, err := filepath.Abs(ldPath)
 		if err != nil {
 			continue
 		}
-		patterns = append(patterns, filepath.Join(d, baseLibName+"*"))
+		patterns = append(patterns, filepath.Join(d, baseLibName))
 	}
 	patterns = append(patterns, defaultPatterns...)
 	slog.Debug("gpu library search", "globs", patterns)
@ -641,3 +641,31 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 		return "", ""
 	}
 }
+
+func LibraryDir() string {
+	// On Windows/linux we bundle the dependencies at the same level as the executable
+	appExe, err := os.Executable()
+	if err != nil {
+		slog.Warn("failed to lookup executable path", "error", err)
+	}
+	cwd, err := os.Getwd()
+	if err != nil {
+		slog.Warn("failed to lookup working directory", "error", err)
+	}
+	// Scan for any of our dependeices, and pick first match
+	for _, root := range []string{filepath.Dir(appExe), filepath.Join(filepath.Dir(appExe), ".."), cwd} {
+		libDep := filepath.Join("lib", "ollama")
+		if _, err := os.Stat(filepath.Join(root, libDep)); err == nil {
+			return filepath.Join(root, libDep)
+		}
+		// Developer mode, local build
+		if _, err := os.Stat(filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
+			return filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)
+		}
+		if _, err := os.Stat(filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
+			return filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)
+		}
+	}
+	slog.Warn("unable to locate gpu dependency libraries")
+	return ""
+}
--- a/gpu/gpu_darwin.go
+++ b/gpu/gpu_darwin.go
@ -25,7 +25,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
-				Variant: GetCPUCapability(),
+				Variant: GetCPUCapability().String(),
 				memInfo: mem,
 			},
 		}
@ -48,7 +48,7 @@ func GetCPUInfo() GpuInfoList {
 	return []GpuInfo{
 		{
 			Library: "cpu",
-			Variant: GetCPUCapability(),
+			Variant: GetCPUCapability().String(),
 			memInfo: mem,
 		},
 	}
--- a/gpu/gpu_linux.go
+++ b/gpu/gpu_linux.go
@ -47,7 +47,7 @@ var (
 	CudartMgmtName = "libcudart.so*"
 	NvcudaMgmtName = "libcuda.so*"
 	NvmlMgmtName   = "" // not currently wired on linux
-	OneapiMgmtName = "libze_intel_gpu.so"
+	OneapiMgmtName = "libze_intel_gpu.so*"
 )

 func GetCPUMem() (memInfo, error) {
--- a/gpu/gpu_test.go
+++ b/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
--- a/gpu/types.go
+++ b/gpu/types.go
@ -19,7 +19,7 @@ type GpuInfo struct {
 	Library string `json:"library,omitempty"`

 	// Optional variant to select (e.g. versions, cpu feature flags)
-	Variant CPUCapability `json:"variant"`
+	Variant string `json:"variant"`

 	// MinimumMemory represents the minimum memory required to use the GPU
 	MinimumMemory uint64 `json:"-"`
@ -53,8 +53,10 @@ type CPUInfo struct {

 type CudaGPUInfo struct {
 	GpuInfo
-	OSOverhead uint64 // Memory overhead between the driver library and management library
-	index      int    //nolint:unused,nolintlint
+	OSOverhead   uint64 // Memory overhead between the driver library and management library
+	index        int    //nolint:unused,nolintlint
+	computeMajor int    //nolint:unused,nolintlint
+	computeMinor int    //nolint:unused,nolintlint
 }
 type CudaGPUInfoList []CudaGPUInfo

@ -81,8 +83,8 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != CPUCapabilityNone {
-			requested += "_" + info.Variant.String()
+		if info.Variant != CPUCapabilityNone.String() {
+			requested += "_" + info.Variant
 		}
 		for i, lib := range libs {
 			if lib == requested {
@ -92,7 +94,7 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 			}
 		}
 		if !found {
-			libs = append(libs, info.Library)
+			libs = append(libs, requested)
 			resp = append(resp, []GpuInfo{info})
 		}
 	}
@ -105,6 +107,7 @@ func (l GpuInfoList) LogDetails() {
 		slog.Info("inference compute",
 			"id", g.ID,
 			"library", g.Library,
+			"variant", g.Variant,
 			"compute", g.Compute,
 			"driver", fmt.Sprintf("%d.%d", g.DriverMajor, g.DriverMinor),
 			"name", g.Name,
--- a/integration/embed_test.go
+++ b/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 {
-		t.Fatalf("expected 8 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 6 {
+		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 {
-		t.Fatalf("expected 16 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 12 {
+		t.Fatalf("expected 12 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }

--- a/llm/ext_server/CMakeLists.txt
+++ b/llm/ext_server/CMakeLists.txt
@ -1,12 +1,13 @@
 set(TARGET ollama_llama_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+set(LLAMA_SERVER_LDFLAGS $ENV{LLAMA_SERVER_LDFLAGS})
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_SERVER_LDFLAGS})
 if (WIN32)
    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
--- a/llm/ext_server/server.cpp
+++ b/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
--- a/llm/generate/gen_common.sh
+++ b/llm/generate/gen_common.sh
@ -9,11 +9,14 @@ init_vars() {
        ARCH="arm64"
        ;;
    *)
-        ARCH=$(uname -m | sed -e "s/aarch64/arm64/g")
+        echo "GOARCH must be set"
+        echo "this script is meant to be run from within go generate"
+        exit 1
+        ;;
    esac

    LLAMACPP_DIR=../llama.cpp
-    CMAKE_DEFS=""
+    CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on"
    CMAKE_TARGETS="--target ollama_llama_server"
    if echo "${CGO_CFLAGS}" | grep -- '-g' >/dev/null; then
        CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on ${CMAKE_DEFS}"
@ -27,6 +30,7 @@ init_vars() {
        WHOLE_ARCHIVE="-Wl,-force_load"
        NO_WHOLE_ARCHIVE=""
        GCC_ARCH="-arch ${ARCH}"
+        DIST_BASE=../../dist/darwin-${GOARCH}/
        ;;
    "Linux")
        LIB_EXT="so"
@ -35,6 +39,7 @@ init_vars() {

        # Cross compiling not supported on linux - Use docker
        GCC_ARCH=""
+        DIST_BASE=../../dist/linux-${GOARCH}/
        ;;
    *)
        ;;
@ -42,6 +47,7 @@ init_vars() {
    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
        CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
    fi
+    GZIP=$(which pigz 2>/dev/null || echo "gzip")
 }

 git_module_setup() {
@ -85,26 +91,36 @@ build() {

 compress() {
    echo "Compressing payloads to reduce overall binary size..."
-    pids=""
    rm -rf ${BUILD_DIR}/bin/*.gz
    for f in ${BUILD_DIR}/bin/* ; do
-        gzip -n --best -f ${f} &
-        pids+=" $!"
+        ${GZIP} -n --best -f ${f} &
+        compress_pids+=" $!"
    done
    # check for lib directory
    if [ -d ${BUILD_DIR}/lib ]; then
        for f in ${BUILD_DIR}/lib/* ; do
-            gzip -n --best -f ${f} &
-            pids+=" $!"
+            ${GZIP} -n --best -f ${f} &
+            compress_pids+=" $!"
        done
    fi
    echo
-    for pid in ${pids}; do
+}
+
+wait_for_compress() {
+    for pid in ${compress_pids}; do
        wait $pid
    done
    echo "Finished compression"
 }

+install() {
+    echo "Installing libraries to bin dir ${BUILD_DIR}/bin/"
+    for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT}); do
+        rm -f "${BUILD_DIR}/bin/$(basename ${lib})"
+        cp -af "${lib}" "${BUILD_DIR}/bin/"
+    done
+}
+
 # Keep the local tree clean after we're done with the build
 cleanup() {
    (cd ${LLAMACPP_DIR}/ && git checkout CMakeLists.txt)
--- a/llm/generate/gen_darwin.sh
+++ b/llm/generate/gen_darwin.sh
@ -6,6 +6,7 @@

 set -ex
 set -o pipefail
+compress_pids=""
 echo "Starting darwin generate script"
 source $(dirname $0)/gen_common.sh
 init_vars
@ -98,4 +99,5 @@ case "${GOARCH}" in
 esac

 cleanup
+wait_for_compress
 echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
--- a/llm/generate/gen_linux.sh
+++ b/llm/generate/gen_linux.sh
@ -13,6 +13,7 @@

 set -ex
 set -o pipefail
+compress_pids=""

 # See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
 amdGPUs() {
@ -51,7 +52,7 @@ if [ -z "${CUDACXX}" ]; then
        export CUDACXX=$(command -v nvcc)
    fi
 fi
-COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
+COMMON_CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
 source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
@ -77,10 +78,11 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
    if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
        init_vars
        echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
        BUILD_DIR="../build/linux/${ARCH}/cpu"
        echo "Building custom CPU"
        build
+        install
        compress
    else
        # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
@ -93,7 +95,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
        # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake

-        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
+        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
            #
            # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
@ -103,6 +105,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
            BUILD_DIR="../build/linux/${ARCH}/cpu"
            echo "Building LCD CPU"
            build
+            install
            compress
        fi

@ -120,6 +123,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
                echo "Building AVX CPU"
                build
+                install
                compress
            fi

@ -133,6 +137,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
                echo "Building AVX2 CPU"
                build
+                install
                compress
            fi
        fi
@ -160,7 +165,7 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
    echo "CUDA libraries detected - building dynamic CUDA library"
    init_vars
    CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
-    if [ -n "${CUDA_MAJOR}" ]; then
+    if [ -n "${CUDA_MAJOR}" -a -z "${CUDA_VARIANT}" ]; then
        CUDA_VARIANT=_v${CUDA_MAJOR}
    fi
    if [ "${ARCH}" == "arm64" ]; then
@ -178,29 +183,19 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
        echo "Building custom CUDA GPU"
    else
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
    fi
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
+    export CUDAFLAGS="-t8"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
    BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
-    EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
+    export LLAMA_SERVER_LDFLAGS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
+    CUDA_DIST_DIR="${CUDA_DIST_DIR:-${DIST_BASE}/lib/ollama}"
    build
-
-    # Carry the CUDA libs as payloads to help reduce dependency burden on users
-    #
-    # TODO - in the future we may shift to packaging these separately and conditionally
-    #        downloading them in the install script.
-    DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
-    for lib in libcudart.so libcublas.so libcublasLt.so ; do
-        DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
-        if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
-            cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
-        elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
-            cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
-        elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
-            cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
-        else
-            cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
-        fi
+    install
+    echo "Installing CUDA dependencies in ${CUDA_DIST_DIR}"
+    mkdir -p "${CUDA_DIST_DIR}"
+    for lib in ${CUDA_LIB_DIR}/libcudart.so* ${CUDA_LIB_DIR}/libcublas.so* ${CUDA_LIB_DIR}/libcublasLt.so* ; do
+        cp -a "${lib}" "${CUDA_DIST_DIR}"
    done
    compress

@ -218,21 +213,24 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
    CC=icx
    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
    BUILD_DIR="../build/linux/${ARCH}/oneapi"
-    EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
+    ONEAPI_DIST_DIR="${DIST_BASE}/lib/ollama"
+    export LLAMA_SERVER_LDFLAGS="-fsycl -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
    DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
    build

    # copy oneAPI dependencies
+    mkdir -p "${ONEAPI_DIST_DIR}"
    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e sycl -e mkl -e tbb); do
-        cp "${dep}" "${BUILD_DIR}/bin/"
+        cp -a "${dep}" "${ONEAPI_DIST_DIR}"
    done
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${ONEAPI_DIST_DIR}"
+    install
    compress
 fi

@ -254,7 +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=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
@ -262,23 +260,22 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
        echo "Building custom ROCM GPU"
    fi
    BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
-    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
+    # 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"
    build

-    # Record the ROCM dependencies
-    rm -f "${BUILD_DIR}/bin/deps.txt"
-    touch "${BUILD_DIR}/bin/deps.txt"
-    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
-        echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
+    # copy the ROCM dependencies
+    mkdir -p "${ROCM_DIST_DIR}"
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -v "${ARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo ); do
+        cp -a "${dep}"* "${ROCM_DIST_DIR}"
    done
-    # bomb out if for some reason we didn't get a few deps
-    if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
-        cat "${BUILD_DIR}/bin/deps.txt"
-        echo "ERROR: deps file short"
-        exit 1
-    fi
+    install
    compress
 fi

 cleanup
+wait_for_compress
 echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
--- a/llm/generate/gen_windows.ps1
+++ b/llm/generate/gen_windows.ps1
@ -35,7 +35,7 @@ function init_vars {
        )
    $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
-    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\ollama_runners"
+    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\runners"
    md "$script:DIST_BASE" -ea 0 > $null
    if ($env:CGO_CFLAGS -contains "-g") {
        $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
@ -117,7 +117,7 @@ function build {
    if ($cmakeDefs -contains "-G") {
        $extra=@("-j8")
    } else {
-        $extra= @("--", "/p:CL_MPcount=8")
+        $extra= @("--", "/maxCpuCount:8")
    }
    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ }) $extra"
    & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ }) $extra
@ -261,7 +261,7 @@ function build_cuda() {
    if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${script:CUDA_LIB_DIR}")) {
        # Then build cuda as a dynamically loaded library
        $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
-        $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
+        $script:CUDA_VERSION=((get-item ($nvcc | split-path | split-path)).Basename -Split "\.")[0]
        if ($null -ne $script:CUDA_VERSION) {
            $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
        }
@ -273,9 +273,9 @@ function build_cuda() {
            "-DGGML_CUDA=ON",
            "-DGGML_AVX=on",
            "-DGGML_AVX2=off",
-            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
-            "-DCMAKE_CUDA_FLAGS=-t8",
-            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
+            "-DCMAKE_CUDA_FLAGS=-t6",
+            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}",
+            "-DCMAKE_CUDA_COMPILER_TOOLKIT_ROOT=$env:CUDA_PATH"
            )
        if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
            write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
@ -286,12 +286,11 @@ function build_cuda() {
        sign
        install

-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\" -ea 0 > $null
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
+        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
    } else {
        write-host "Skipping CUDA generation step"
    }
@ -325,18 +324,17 @@ function build_oneapi() {
    sign
    install

-    rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\" -ea 0 > $null
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
  } else {
    Write-Host "Skipping oneAPI generation step"
  }
@ -357,7 +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",
@ -386,12 +384,11 @@ function build_rocm() {
        sign
        install

-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
-        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\" -ea 0 > $null
+        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
        # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
-        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\"
+        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\"
    } else {
        write-host "Skipping ROCm generation step"
    }
--- a/llm/ggml.go
+++ b/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")
 }
--- a/llm/memory_test.go
+++ b/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),
--- a/llm/patches/08-pooling.diff
+++ b/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;
-             }
--- a/llm/payload.go
+++ b/llm/payload.go
@ -82,8 +82,8 @@ func serversForGpu(info gpu.GpuInfo) []string {
 	// glob workDir for files that start with ollama_
 	availableServers := getAvailableServers()
 	requested := info.Library
-	if info.Variant != gpu.CPUCapabilityNone {
-		requested += "_" + info.Variant.String()
+	if info.Variant != gpu.CPUCapabilityNone.String() {
+		requested += "_" + info.Variant
 	}

 	servers := []string{}
--- a/llm/server.go
+++ b/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 {
@ -306,20 +306,18 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if runtime.GOOS == "windows" {
 			pathEnv = "PATH"
 		}
-		// prepend the server directory to LD_LIBRARY_PATH/PATH and the parent dir for common dependencies
-		libraryPaths := []string{dir, filepath.Dir(dir)}
+		// Start with the server directory for the LD_LIBRARY_PATH/PATH
+		libraryPaths := []string{dir}

 		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
-			// Append our runner directory to the path
-			// This will favor system libraries over our bundled library dependencies
+			// favor our bundled library dependencies over system libraries
 			libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
 		}

 		// Note: we always put the dependency path first
-		// since this was the exact version we verified for AMD GPUs
-		// and we favor what the user had in their path
+		// since this was the exact version we compiled/linked against
 		if gpus[0].DependencyPath != "" {
-			// TODO refine for multi-gpu support
+			// assume gpus from the same library have the same dependency path
 			libraryPaths = append([]string{gpus[0].DependencyPath}, libraryPaths...)
 		}

--- a/scripts/build_linux.sh
+++ b/scripts/build_linux.sh
@ -4,6 +4,7 @@ set -eu

 export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
 export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
+GZIP=$(which pigz 2>/dev/null || echo "gzip")

 BUILD_ARCH=${BUILD_ARCH:-"amd64 arm64"}
 export AMDGPU_TARGETS=${AMDGPU_TARGETS:=""}
@ -21,11 +22,16 @@ for TARGETARCH in ${BUILD_ARCH}; do
        -t builder:$TARGETARCH \
        .
    docker create --platform linux/$TARGETARCH --name builder-$TARGETARCH builder:$TARGETARCH
-    docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/ollama ./dist/ollama-linux-$TARGETARCH
-
-    if [ "$TARGETARCH" = "amd64" ]; then
-        docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/deps/ ./dist/
+    rm -rf ./dist/linux-$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
--- a/scripts/build_windows.ps1
+++ b/scripts/build_windows.ps1
@ -7,6 +7,7 @@
 $ErrorActionPreference = "Stop"

 function checkEnv() {
+    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
    $script:TARGET_ARCH=$Env:PROCESSOR_ARCHITECTURE.ToLower()
    Write-host "Building for ${script:TARGET_ARCH}"
    write-host "Locating required tools and paths"
@ -15,26 +16,23 @@ function checkEnv() {
        $MSVC_INSTALL=(Get-CimInstance MSFT_VSInstance -Namespace root/cimv2/vs)[0].InstallLocation
        $env:VCToolsRedistDir=(get-item "${MSVC_INSTALL}\VC\Redist\MSVC\*")[0]
    }
-    # Try to find the CUDA dir
-    if ($null -eq $env:NVIDIA_DIR) {
+    # Locate CUDA versions
+    # Note: this assumes every version found will be built
+    $cudaList=(get-item "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*\bin\" -ea 'silentlycontinue')
+    if ($cudaList.length -eq 0) {
        $d=(get-command -ea 'silentlycontinue' nvcc).path
-        if ($d -ne $null) {
-            $script:NVIDIA_DIR=($d| split-path -parent)
-        } else {
-            $cudaList=(get-item "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*\bin\" -ea 'silentlycontinue')
-            if ($cudaList.length > 0) {
-                $script:NVIDIA_DIR=$cudaList[0]
-            }
+        if ($null -ne $d) {
+            $script:CUDA_DIRS=@($d| split-path -parent)
        }
    } else {
-        $script:NVIDIA_DIR=$env:NVIDIA_DIR
+        $script:CUDA_DIRS=$cudaList
    }
    
    $script:INNO_SETUP_DIR=(get-item "C:\Program Files*\Inno Setup*\")[0]

    $script:DEPS_DIR="${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}"
    $env:CGO_ENABLED="1"
-    echo "Checking version"
+    Write-Output "Checking version"
    if (!$env:VERSION) {
        $data=(git describe --tags --first-parent --abbrev=7 --long --dirty --always)
        $pattern="v(.+)"
@ -71,7 +69,48 @@ function checkEnv() {
 function buildOllama() {
    write-host "Building ollama CLI"
    if ($null -eq ${env:OLLAMA_SKIP_GENERATE}) {
-        & go generate ./...
+        Remove-Item -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}"
+
+        # TODO - consider trying to parallelize this with Start-ThreadJob, but env vars can't be used to toggle
+        #        which targets to build
+
+        # Start by skipping CUDA to build everything else
+        pwsh -Command { $env:OLLAMA_SKIP_CUDA_GENERATE="1"; & go generate ./... }
+        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}    
+
+        # Then skip everyhting else and build all the CUDA variants
+        foreach ($env:CUDA_LIB_DIR in $script:CUDA_DIRS) {
+            write-host "Building CUDA ${env:CUDA_LIB_DIR}"
+
+            if ($env:CUDA_LIB_DIR.Contains("v12")) {
+                pwsh -Command {
+                    $env:OLLAMA_SKIP_CUDA_GENERATE=""
+                    $env:OLLAMA_SKIP_STATIC_GENERATE="1"
+                    $env:OLLAMA_SKIP_CPU_GENERATE="1"
+                    $env:OLLAMA_SKIP_ONEAPI_GENERATE="1"
+                    $env:OLLAMA_SKIP_ROCM_GENERATE="1"
+                    $env:CMAKE_CUDA_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
+                    $env:OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on"
+                    $env:CUDA_PATH=split-path -path $env:CUDA_LIB_DIR -parent
+                    $env:PATH="$envs:CUDA_LIB_DIR;$env:PATH"
+                    & go generate ./...
+                }
+            } else {
+                pwsh -Command {
+                    $env:OLLAMA_SKIP_CUDA_GENERATE=""
+                    $env:OLLAMA_SKIP_STATIC_GENERATE="1"
+                    $env:OLLAMA_SKIP_CPU_GENERATE="1"
+                    $env:OLLAMA_SKIP_ONEAPI_GENERATE="1"
+                    $env:OLLAMA_SKIP_ROCM_GENERATE="1"
+                    $env:CMAKE_CUDA_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+                    $env:OLLAMA_CUSTOM_CUDA_DEFS=""
+                    $env:CUDA_PATH=split-path -path $env:CUDA_LIB_DIR -parent
+                    $env:PATH="$envs:CUDA_LIB_DIR;$env:PATH"
+                    & go generate ./...
+                }
+            }
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}    
    } else {
        write-host "Skipping generate step with OLLAMA_SKIP_GENERATE set"
@ -83,8 +122,8 @@ function buildOllama() {
            /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} ollama.exe
        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
    }
-    New-Item -ItemType Directory -Path .\dist\windows-${script:TARGET_ARCH}\ -Force
-    cp .\ollama.exe .\dist\windows-${script:TARGET_ARCH}\
+    New-Item -ItemType Directory -Path .\dist\windows-${script:TARGET_ARCH}\bin\ -Force
+    cp .\ollama.exe .\dist\windows-${script:TARGET_ARCH}\bin\
 }

 function buildApp() {
@ -103,22 +142,22 @@ function buildApp() {
 function gatherDependencies() {
    write-host "Gathering runtime dependencies"
    cd "${script:SRC_DIR}"
-    md "${script:DEPS_DIR}\ollama_runners" -ea 0 > $null
+    md "${script:DEPS_DIR}\lib\ollama" -ea 0 > $null

    # TODO - this varies based on host build system and MSVC version - drive from dumpbin output
    # currently works for Win11 + MSVC 2019 + Cuda V11
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140*.dll" "${script:DEPS_DIR}\ollama_runners\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\ollama_runners\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\ollama_runners\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140*.dll" "${script:DEPS_DIR}\lib\ollama\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\lib\ollama\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\lib\ollama\"
    foreach ($part in $("runtime", "stdio", "filesystem", "math", "convert", "heap", "string", "time", "locale", "environment")) {
-        cp "$env:VCToolsRedistDir\..\..\..\Tools\Llvm\x64\bin\api-ms-win-crt-${part}*.dll" "${script:DEPS_DIR}\ollama_runners\"
+        cp "$env:VCToolsRedistDir\..\..\..\Tools\Llvm\x64\bin\api-ms-win-crt-${part}*.dll" "${script:DEPS_DIR}\lib\ollama\"
    }


    cp "${script:SRC_DIR}\app\ollama_welcome.ps1" "${script:SRC_DIR}\dist\"
    if ("${env:KEY_CONTAINER}") {
        write-host "about to sign"
-        foreach ($file in (get-childitem "${script:DEPS_DIR}\cuda\cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
+        foreach ($file in (get-childitem "${script:DEPS_DIR}\lib\ollama\cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
            write-host "signing $file"
            & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
                /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} $file
--- a/scripts/install.sh
+++ b/scripts/install.sh
@ -63,16 +63,36 @@ if [ -n "$NEEDS" ]; then
    exit 1
 fi

-status "Downloading ollama..."
-curl --fail --show-error --location --progress-bar -o $TEMP_DIR/ollama "https://ollama.com/download/ollama-linux-${ARCH}${VER_PARAM}"
-
 for BINDIR in /usr/local/bin /usr/bin /bin; do
    echo $PATH | grep -q $BINDIR && break || continue
 done
+OLLAMA_INSTALL_DIR=$(dirname ${BINDIR})

-status "Installing ollama to $BINDIR..."
+status "Installing ollama to $OLLAMA_INSTALL_DIR"
 $SUDO install -o0 -g0 -m755 -d $BINDIR
-$SUDO install -o0 -g0 -m755 $TEMP_DIR/ollama $BINDIR/ollama
+$SUDO install -o0 -g0 -m755 -d "$OLLAMA_INSTALL_DIR"
+if curl -I --silent --fail --location "https://ollama.com/download/ollama-linux-${ARCH}.tgz${VER_PARAM}" >/dev/null ; then
+    status "Downloading Linux ${ARCH} bundle"
+    curl --fail --show-error --location --progress-bar \
+        "https://ollama.com/download/ollama-linux-${ARCH}.tgz${VER_PARAM}" | \
+        $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+    BUNDLE=1
+    if [ "$OLLAMA_INSTALL_DIR/bin/ollama" != "$BINDIR/ollama" ] ; then
+        status "Making ollama accessible in the PATH in $BINDIR"
+        $SUDO ln -sf "$OLLAMA_INSTALL_DIR/ollama" "$BINDIR/ollama"
+    fi
+else
+    status "Downloading Linux ${ARCH} CLI"
+    curl --fail --show-error --location --progress-bar -o "$TEMP_DIR/ollama"\
+    "https://ollama.com/download/ollama-linux-${ARCH}${VER_PARAM}"
+    $SUDO install -o0 -g0 -m755 $TEMP_DIR/ollama $OLLAMA_INSTALL_DIR/ollama
+    BUNDLE=0
+    if [ "$OLLAMA_INSTALL_DIR/ollama" != "$BINDIR/ollama" ] ; then
+        status "Making ollama accessible in the PATH in $BINDIR"
+        $SUDO ln -sf "$OLLAMA_INSTALL_DIR/ollama" "$BINDIR/ollama"
+    fi
+fi
+

 install_success() {
    status 'The Ollama API is now available at 127.0.0.1:11434.'
@ -178,6 +198,16 @@ if ! check_gpu lspci nvidia && ! check_gpu lshw nvidia && ! check_gpu lspci amdg
 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
+    fi
    # Look for pre-existing ROCm v6 before downloading the dependencies
    for search in "${HIP_PATH:-''}" "${ROCM_PATH:-''}" "/opt/rocm" "/usr/lib64"; do
        if [ -n "${search}" ] && [ -e "${search}/libhipblas.so.2" -o -e "${search}/lib/libhipblas.so.2" ]; then
--- a/scripts/rh_linux_deps.sh
+++ b/scripts/rh_linux_deps.sh
@ -3,6 +3,7 @@
 # Script for common Dockerfile dependency installation in redhat linux based images

 set -ex
+set -o pipefail
 MACHINE=$(uname -m)

 if grep -i "centos" /etc/system-release >/dev/null; then
@ -29,7 +30,7 @@ if grep -i "centos" /etc/system-release >/dev/null; then
        dnf install -y rh-git227-git
        ln -s /opt/rh/rh-git227/root/usr/bin/git /usr/local/bin/git
    fi
-    dnf install -y devtoolset-10-gcc devtoolset-10-gcc-c++
+    dnf install -y devtoolset-10-gcc devtoolset-10-gcc-c++ pigz
 elif grep -i "rocky" /etc/system-release >/dev/null; then
    # Temporary workaround until rocky 8 AppStream ships GCC 10.4 (10.3 is incompatible with NVCC)
    cat << EOF > /etc/yum.repos.d/Rocky-Vault.repo
@ -43,12 +44,21 @@ gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-rockyofficial
 EOF
    dnf install -y git \
        gcc-toolset-10-gcc-10.2.1-8.2.el8 \
-        gcc-toolset-10-gcc-c++-10.2.1-8.2.el8
+        gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 \
+        pigz
 else
    echo "ERROR Unexpected distro"
    exit 1
 fi

+if [ "${MACHINE}" = "x86_64" ] ; then
+    curl -s -L https://github.com/ccache/ccache/releases/download/v4.10.2/ccache-4.10.2-linux-x86_64.tar.xz | tar -Jx -C /tmp --strip-components 1 && \
+    mv /tmp/ccache /usr/local/bin/
+else
+    yum -y install epel-release
+    yum install -y ccache
+fi
+
 if [ -n "${CMAKE_VERSION}" ]; then
    curl -s -L https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-$(uname -m).tar.gz | tar -zx -C /usr --strip-components 1
 fi
--- a/server/images.go
+++ b/server/images.go
@ -215,25 +215,20 @@ func GetManifest(mp ModelPath) (*Manifest, string, error) {
 		return nil, "", err
 	}

-	if _, err = os.Stat(fp); err != nil {
-		return nil, "", err
-	}
-
-	var manifest *Manifest
-
-	bts, err := os.ReadFile(fp)
+	f, err := os.Open(fp)
 	if err != nil {
-		return nil, "", fmt.Errorf("couldn't open file '%s'", fp)
+		return nil, "", err
 	}
+	defer f.Close()

-	shaSum := sha256.Sum256(bts)
-	shaStr := hex.EncodeToString(shaSum[:])
+	sha256sum := sha256.New()

-	if err := json.Unmarshal(bts, &manifest); err != nil {
+	var manifest Manifest
+	if err := json.NewDecoder(io.TeeReader(f, sha256sum)).Decode(&manifest); err != nil {
 		return nil, "", err
 	}

-	return manifest, shaStr, nil
+	return &manifest, hex.EncodeToString(sha256sum.Sum(nil)), nil
 }

 func GetModel(name string) (*Model, error) {
@ -374,13 +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() {
+			if name := model.ParseName(c.Args); name.IsValid() && command == "model" {
 				baseLayers, err = parseFromModel(ctx, name, fn)
 				if err != nil {
 					return err
@ -414,14 +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
 				}
@ -692,43 +688,18 @@ func CopyModel(src, dst model.Name) error {
 	return err
 }

-func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]struct{}) error {
-	fp, err := GetManifestPath()
+func deleteUnusedLayers(deleteMap map[string]struct{}) error {
+	manifests, err := Manifests()
 	if err != nil {
 		return err
 	}

-	walkFunc := func(path string, info os.FileInfo, _ error) error {
-		if info.IsDir() {
-			return nil
-		}
-
-		dir, file := filepath.Split(path)
-		dir = strings.Trim(strings.TrimPrefix(dir, fp), string(os.PathSeparator))
-		tag := strings.Join([]string{dir, file}, ":")
-		fmp := ParseModelPath(tag)
-
-		// skip the manifest we're trying to delete
-		if skipModelPath != nil && skipModelPath.GetFullTagname() == fmp.GetFullTagname() {
-			return nil
-		}
-
-		// save (i.e. delete from the deleteMap) any files used in other manifests
-		manifest, _, err := GetManifest(fmp)
-		if err != nil {
-			return err
-		}
-
+	for _, manifest := range manifests {
 		for _, layer := range manifest.Layers {
 			delete(deleteMap, layer.Digest)
 		}

 		delete(deleteMap, manifest.Config.Digest)
-		return nil
-	}
-
-	if err := filepath.Walk(fp, walkFunc); err != nil {
-		return err
 	}

 	// only delete the files which are still in the deleteMap
@ -781,8 +752,7 @@ func PruneLayers() error {

 	slog.Info(fmt.Sprintf("total blobs: %d", len(deleteMap)))

-	err = deleteUnusedLayers(nil, deleteMap)
-	if err != nil {
+	if err := deleteUnusedLayers(deleteMap); err != nil {
 		slog.Error(fmt.Sprintf("couldn't remove unused layers: %v", err))
 		return nil
 	}
@ -877,26 +847,19 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	mp := ParseModelPath(name)

-	var manifest *Manifest
-	var err error
-	var noprune string
-
 	// build deleteMap to prune unused layers
 	deleteMap := make(map[string]struct{})
-
-	if !envconfig.NoPrune() {
-		manifest, _, err = GetManifest(mp)
-		if err != nil && !errors.Is(err, os.ErrNotExist) {
-			return err
+	manifest, _, err := GetManifest(mp)
+	if errors.Is(err, os.ErrNotExist) {
+		// noop
+	} else if err != nil && !errors.Is(err, os.ErrNotExist) {
+		return err
+	} else {
+		for _, l := range manifest.Layers {
+			deleteMap[l.Digest] = struct{}{}
 		}
-
-		if manifest != nil {
-			for _, l := range manifest.Layers {
-				deleteMap[l.Digest] = struct{}{}
-			}
-			if manifest.Config.Digest != "" {
-				deleteMap[manifest.Config.Digest] = struct{}{}
-			}
+		if manifest.Config.Digest != "" {
+			deleteMap[manifest.Config.Digest] = struct{}{}
 		}
 	}

@ -975,11 +938,9 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		return err
 	}

-	if noprune == "" {
-		fn(api.ProgressResponse{Status: "removing any unused layers"})
-		err = deleteUnusedLayers(nil, deleteMap)
-		if err != nil {
-			slog.Error(fmt.Sprintf("couldn't remove unused layers: %v", err))
+	if !envconfig.NoPrune() && len(deleteMap) > 0 {
+		fn(api.ProgressResponse{Status: "removing unused layers"})
+		if err := deleteUnusedLayers(deleteMap); err != nil {
 			fn(api.ProgressResponse{Status: fmt.Sprintf("couldn't remove unused layers: %v", err)})
 		}
 	}
@ -1000,12 +961,12 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptio
 	}
 	defer resp.Body.Close()

-	var m *Manifest
+	var m Manifest
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
 	}

-	return m, err
+	return &m, err
 }

 // GetSHA256Digest returns the SHA256 hash of a given buffer and returns it, and the size of buffer
--- a/server/layer.go
+++ b/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{
--- a/server/manifest.go
+++ b/server/manifest.go
@ -5,6 +5,7 @@ import (
 	"encoding/hex"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"log/slog"
 	"os"
@ -150,14 +151,16 @@ func Manifests() (map[model.Name]*Manifest, error) {

 			n := model.ParseNameFromFilepath(rel)
 			if !n.IsValid() {
-				slog.Warn("bad manifest name", "path", rel, "error", err)
+				slog.Warn("bad manifest name", "path", rel)
 				continue
 			}

 			m, err := ParseNamedManifest(n)
-			if err != nil {
+			if syntax := &(json.SyntaxError{}); errors.As(err, &syntax) {
 				slog.Warn("bad manifest", "name", n, "error", err)
 				continue
+			} else if err != nil {
+				return nil, fmt.Errorf("%s: %w", n, err)
 			}

 			ms[n] = m
--- a/server/model.go
+++ b/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"})
-	if err := convert.Convert(convert.NewZipReader(r, p, 32<<20), t); err != nil {
-		return nil, err
+	var layerType string
+
+	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"
--- a/server/model_test.go
+++ b/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)
 	}
--- a/server/sched.go
+++ b/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
@ -734,7 +739,10 @@ func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoL

 // If multiple Libraries are detected, pick the Library which loads the most layers for the model
 func pickBestPartialFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
-	*numParallel = 1
+	if *numParallel <= 0 {
+		*numParallel = 1
+		req.opts.NumCtx = req.origNumCtx
+	}
 	byLibrary := gpus.ByLibrary()
 	if len(byLibrary) <= 1 {
 		return gpus
--- a/server/sched_test.go
+++ b/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),
--- a/server/upload.go
+++ b/server/upload.go
@ -45,7 +45,7 @@ type blobUpload struct {
 }

 const (
-	numUploadParts          = 64
+	numUploadParts          = 16
 	minUploadPartSize int64 = 100 * format.MegaByte
 	maxUploadPartSize int64 = 1000 * format.MegaByte
 )