add examples of streaming in python and node

Signed-off-by: Matt Williams <m@technovangelist.com>

.dockerignore

@@ -2,3 +2,4 @@
 ollama
 app
 llm/llama.cpp/ggml
+llm/llama.cpp/gguf

.gitmodules

@@ -1,4 +1,9 @@
 [submodule "llm/llama.cpp/ggml"]
-	path = llm/llama.cpp/ggml
-	url = https://github.com/ggerganov/llama.cpp.git
-	ignore = dirty
+    path = llm/llama.cpp/ggml
+    url = https://github.com/ggerganov/llama.cpp.git
+    ignore = dirty
+    shallow = true
+[submodule "llm/llama.cpp/gguf"]
+    path = llm/llama.cpp/gguf
+    url = https://github.com/ggerganov/llama.cpp.git
+    shallow = true

Dockerfile.cuda

@@ -0,0 +1,22 @@
+FROM nvidia/cuda:12.2.0-devel-ubuntu22.04
+
+WORKDIR /go/src/github.com/jmorganca/ollama
+RUN apt-get update && apt-get install -y git build-essential cmake
+ADD https://dl.google.com/go/go1.21.1.linux-amd64.tar.gz /tmp/go1.21.1.tar.gz
+RUN mkdir -p /usr/local && tar xz -C /usr/local </tmp/go1.21.1.tar.gz
+
+COPY . .
+RUN /usr/local/go/bin/go generate ./... && /usr/local/go/bin/go build -ldflags '-linkmode external -extldflags "-static"' .
+
+FROM nvidia/cuda:12.2.0-runtime-ubuntu22.04
+ENV OLLAMA_HOST 0.0.0.0
+
+ARG USER=ollama
+ARG GROUP=ollama
+RUN groupadd $GROUP && useradd -m -g $GROUP $USER
+
+COPY --from=0 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+
+USER $USER:$GROUP
+ENTRYPOINT ["/bin/ollama"]
+CMD ["serve"]

api/types.go

@@ -291,7 +291,7 @@ func DefaultOptions() Options {
 		NumCtx:             2048,
 		NumKeep:            -1,
 		NumBatch:           512,
-		NumGPU:             1,
+		NumGPU:             -1, // -1 here indicates that NumGPU should be set dynamically
 		NumGQA:             1,
 		LowVRAM:            false,
 		F16KV:              true,

cmd/cmd.go

@@ -672,6 +672,12 @@ func RunServer(cmd *cobra.Command, _ []string) error {
 		origins = strings.Split(o, ",")
 	}
 
+	if noprune := os.Getenv("OLLAMA_NOPRUNE"); noprune == "" {
+		if err := server.PruneLayers(); err != nil {
+			return err
+		}
+	}
+
 	return server.Serve(ln, origins)
 }
 

docs/api.md

@@ -20,6 +20,10 @@ Model names follow a `model:tag` format. Some examples are `orca-mini:3b-q4_1` a
 
 All durations are returned in nanoseconds.
 
+### Streams
+
+Many API responses are streams of JSON objects showing the current status. For examples of working with streams in various languages, see [streaming.md](./streaming.md)
+
 ## Generate a completion
 
 ```

docs/development.md

@@ -6,6 +6,10 @@
 
 Install required tools:
 
+- cmake version 3.24 or higher
+- go version 1.20 or higher
+- gcc version 11.4.0 or higher
+
 ```
 brew install go cmake gcc
 ```
@@ -27,3 +31,9 @@ Now you can run `ollama`:
 ```
 ./ollama
 ```
+
+## Building on Linux with GPU support
+
+- Install cmake and nvidia-cuda-toolkit
+- run `go generate ./...`
+- run `go build .`

docs/streaming.md

@@ -0,0 +1,35 @@
+# Streaming responses in the Ollama Client API
+
+## JavaScript / TypeScript / Deno
+
+```javascript
+const pull = async () => {
+  const request = await fetch("http://localhost:11434/api/pull", {
+    method: "POST",
+    body: JSON.stringify({ name: "llama2:7b-q5_0" }),
+  });
+
+  const reader = await request.body?.pipeThrough(new TextDecoderStream());
+  if (!reader) throw new Error("No reader");
+  for await (const chunk of reader) {
+    const out = JSON.parse(chunk);
+    if (out.status.startsWith("downloading")) {
+      console.log(`${out.status} - ${(out.completed / out.total) * 100}%`);
+    }
+  }
+}
+
+pull();
+```
+
+## Python
+
+```python
+import requests
+import json
+response = requests.post("http://localhost:11434/api/pull", json={"name": "llama2:7b-q5_0"}, stream=True)
+for data in response.iter_lines():
+  out = json.loads(data)
+  if "completed" in out:
+    print(out["completed"] / out["total"] * 100)
+```

llm/falcon.go

@@ -0,0 +1,22 @@
+package llm
+
+const ModelFamilyFalcon = "falcon"
+
+const (
+	falconModelType7B   = 32
+	falconModelType40B  = 60
+	falconModelType180B = 80
+)
+
+func falconModelType(numLayer uint32) string {
+	switch numLayer {
+	case 32:
+		return "7B"
+	case 60:
+		return "40B"
+	case 80:
+		return "180B"
+	default:
+		return "Unknown"
+	}
+}

llm/ggml.go

@@ -3,72 +3,97 @@ package llm
 import (
 	"encoding/binary"
 	"errors"
-	"fmt"
 	"io"
+	"path"
+	"sync"
 )
 
-type ModelFamily string
-
-type ModelType uint32
-
-const (
-	ModelType3B  ModelType = 26
-	ModelType7B  ModelType = 32
-	ModelType13B ModelType = 40
-	ModelType34B ModelType = 48
-	ModelType30B ModelType = 60
-	ModelType65B ModelType = 80
-)
-
-func (mt ModelType) String() string {
-	switch mt {
-	case ModelType3B:
-		return "3B"
-	case ModelType7B:
-		return "7B"
-	case ModelType13B:
-		return "13B"
-	case ModelType34B:
-		return "34B"
-	case ModelType30B:
-		return "30B"
-	case ModelType65B:
-		return "65B"
-	default:
-		return "Unknown"
-	}
-}
-
-type FileType interface {
-	String() string
-}
-
 type GGML struct {
 	magic uint32
 	container
 	model
 }
 
+const (
+	fileTypeF32 uint32 = iota
+	fileTypeF16
+	fileTypeQ4_0
+	fileTypeQ4_1
+	fileTypeQ4_1_F16
+	fileTypeQ8_0 uint32 = iota + 2
+	fileTypeQ5_0
+	fileTypeQ5_1
+	fileTypeQ2_K
+	fileTypeQ3_K_S
+	fileTypeQ3_K_M
+	fileTypeQ3_K_L
+	fileTypeQ4_K_S
+	fileTypeQ4_K_M
+	fileTypeQ5_K_S
+	fileTypeQ5_K_M
+	fileTypeQ6_K
+)
+
+func fileType(fileType uint32) string {
+	switch fileType {
+	case fileTypeF32:
+		return "F32"
+	case fileTypeF16:
+		return "F16"
+	case fileTypeQ4_0:
+		return "Q4_0"
+	case fileTypeQ4_1:
+		return "Q4_1"
+	case fileTypeQ4_1_F16:
+		return "Q4_1_F16"
+	case fileTypeQ8_0:
+		return "Q8_0"
+	case fileTypeQ5_0:
+		return "Q5_0"
+	case fileTypeQ5_1:
+		return "Q5_1"
+	case fileTypeQ2_K:
+		return "Q2_K"
+	case fileTypeQ3_K_S:
+		return "Q3_K_S"
+	case fileTypeQ3_K_M:
+		return "Q3_K_M"
+	case fileTypeQ3_K_L:
+		return "Q3_K_L"
+	case fileTypeQ4_K_S:
+		return "Q4_K_S"
+	case fileTypeQ4_K_M:
+		return "Q4_K_M"
+	case fileTypeQ5_K_S:
+		return "Q5_K_S"
+	case fileTypeQ5_K_M:
+		return "Q5_K_M"
+	case fileTypeQ6_K:
+		return "Q6_K"
+	default:
+		return "Unknown"
+	}
+}
+
 type model interface {
-	ModelFamily() ModelFamily
-	ModelType() ModelType
-	FileType() FileType
+	ModelFamily() string
+	ModelType() string
+	FileType() string
 }
 
 type container interface {
 	Name() string
-	Decode(io.Reader) error
+	Decode(io.Reader) (model, error)
 }
 
-type containerGGML struct {
-}
+type containerGGML struct{}
 
 func (c *containerGGML) Name() string {
 	return "ggml"
 }
 
-func (c *containerGGML) Decode(r io.Reader) error {
-	return nil
+func (c *containerGGML) Decode(r io.Reader) (model, error) {
+	return nil, nil
 }
 
 type containerGGMF struct {
@@ -79,18 +104,18 @@ func (c *containerGGMF) Name() string {
 	return "ggmf"
 }
 
-func (c *containerGGMF) Decode(r io.Reader) error {
+func (c *containerGGMF) Decode(r io.Reader) (model, error) {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)
 
 	switch version {
 	case 1:
 	default:
-		return errors.New("invalid version")
+		return nil, errors.New("invalid version")
 	}
 
 	c.version = version
-	return nil
+	return nil, nil
 }
 
 type containerGGJT struct {
@@ -101,18 +126,22 @@ func (c *containerGGJT) Name() string {
 	return "ggjt"
 }
 
-func (c *containerGGJT) Decode(r io.Reader) error {
+func (c *containerGGJT) Decode(r io.Reader) (model, error) {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)
 
 	switch version {
 	case 1, 2, 3:
 	default:
-		return errors.New("invalid version")
+		return nil, errors.New("invalid version")
 	}
 
 	c.version = version
-	return nil
+
+	// different model types may have different layouts for hyperparameters
+	var llama llamaModel
+	binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
+	return &llama, nil
 }
 
 type containerLORA struct {
@@ -123,32 +152,51 @@ func (c *containerLORA) Name() string {
 	return "ggla"
 }
 
-func (c *containerLORA) Decode(r io.Reader) error {
+func (c *containerLORA) Decode(r io.Reader) (model, error) {
 	var version uint32
 	binary.Read(r, binary.LittleEndian, &version)
 
 	switch version {
 	case 1:
 	default:
-		return errors.New("invalid version")
+		return nil, errors.New("invalid version")
 	}
 
 	c.version = version
-	return nil
+	return nil, nil
+}
+
+var (
+	ggmlGPU = path.Join("llama.cpp", "ggml", "build", "gpu", "bin")
+	ggmlCPU = path.Join("llama.cpp", "ggml", "build", "cpu", "bin")
+)
+
+var (
+	ggmlInit       sync.Once
+	ggmlRunnerPath string
+)
+
+func ggmlRunner() ModelRunner {
+	ggmlInit.Do(func() {
+		ggmlRunnerPath = chooseRunner(ggmlGPU, ggmlCPU)
+	})
+	return ModelRunner{Path: ggmlRunnerPath}
 }
 
 const (
-	// / Magic constant for `ggml` files (unversioned).
+	// Magic constant for `ggml` files (unversioned).
 	FILE_MAGIC_GGML = 0x67676d6c
-	// / Magic constant for `ggml` files (versioned, ggmf).
+	// Magic constant for `ggml` files (versioned, ggmf).
 	FILE_MAGIC_GGMF = 0x67676d66
-	// / Magic constant for `ggml` files (versioned, ggjt).
+	// Magic constant for `ggml` files (versioned, ggjt).
 	FILE_MAGIC_GGJT = 0x67676a74
-	// / Magic constant for `ggla` files (LoRA adapter).
+	// Magic constant for `ggla` files (LoRA adapter).
 	FILE_MAGIC_GGLA = 0x67676C61
+	// Magic constant for `gguf` files (versioned, gguf)
+	FILE_MAGIC_GGUF = 0x46554747
 )
 
-func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
+func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
 	var ggml GGML
 	binary.Read(r, binary.LittleEndian, &ggml.magic)
 
@@ -161,24 +209,18 @@ func DecodeGGML(r io.ReadSeeker, hint ModelFamily) (*GGML, error) {
 		ggml.container = &containerGGJT{}
 	case FILE_MAGIC_GGLA:
 		ggml.container = &containerLORA{}
+	case FILE_MAGIC_GGUF:
+		ggml.container = &containerGGUF{}
 	default:
 		return nil, errors.New("invalid file magic")
 	}
 
-	if err := ggml.Decode(r); err != nil {
+	model, err := ggml.Decode(r)
+	if err != nil {
 		return nil, err
 	}
 
-	// different model types may have different layouts for hyperparameters
-	switch hint {
-	case ModelFamilyLlama:
-		var llama llamaModel
-		binary.Read(r, binary.LittleEndian, &llama.hyperparameters)
-		ggml.model = &llama
-		// TODO: sanity check hyperparameters
-	default:
-		return nil, fmt.Errorf("unsupported model type: %s", hint)
-	}
+	ggml.model = model
 
 	// final model type
 	return &ggml, nil

llm/gguf.go

@@ -0,0 +1,389 @@
+package llm
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"path"
+	"sync"
+)
+
+type containerGGUF struct {
+	Version uint32
+
+	V1 struct {
+		NumTensor uint32
+		NumKV     uint32
+	}
+
+	V2 struct {
+		NumTensor uint64
+		NumKV     uint64
+	}
+}
+
+func (c *containerGGUF) Name() string {
+	return "gguf"
+}
+
+func (c *containerGGUF) Decode(r io.Reader) (model, error) {
+	binary.Read(r, binary.LittleEndian, &c.Version)
+
+	switch c.Version {
+	case 1:
+		binary.Read(r, binary.LittleEndian, &c.V1)
+	case 2:
+		binary.Read(r, binary.LittleEndian, &c.V2)
+	default:
+		return nil, errors.New("invalid version")
+	}
+
+	model := newGGUFModel(c)
+	if err := model.Decode(r); err != nil {
+		return nil, err
+	}
+
+	return model, nil
+}
+
+const (
+	ggufTypeUint8 uint32 = iota
+	ggufTypeInt8
+	ggufTypeUint16
+	ggufTypeInt16
+	ggufTypeUint32
+	ggufTypeInt32
+	ggufTypeFloat32
+	ggufTypeBool
+	ggufTypeString
+	ggufTypeArray
+	ggufTypeUint64
+	ggufTypeInt64
+	ggufTypeFloat64
+)
+
+type kv map[string]any
+
+type ggufModel struct {
+	*containerGGUF
+	kv
+}
+
+func newGGUFModel(container *containerGGUF) *ggufModel {
+	return &ggufModel{
+		containerGGUF: container,
+		kv:            make(kv),
+	}
+}
+
+func (llm *ggufModel) NumKV() uint64 {
+	if llm.Version == 1 {
+		return uint64(llm.V1.NumKV)
+	}
+
+	return llm.V2.NumKV
+}
+
+func (llm *ggufModel) ModelFamily() string {
+	t, ok := llm.kv["general.architecture"].(string)
+	if ok {
+		return t
+	}
+
+	return "unknown"
+}
+
+func (llm *ggufModel) ModelType() string {
+	switch llm.ModelFamily() {
+	case "llama":
+		if blocks, ok := llm.kv["llama.block_count"].(uint32); ok {
+			heads, headsOK := llm.kv["llama.head_count"].(uint32)
+			headKVs, headsKVsOK := llm.kv["llama.head_count_kv"].(uint32)
+			if headsOK && headsKVsOK && heads/headKVs == 8 {
+				return "70B"
+			}
+
+			return llamaModelType(blocks)
+		}
+	case "falcon":
+		if blocks, ok := llm.kv["falcon.block_count"].(uint32); ok {
+			return falconModelType(blocks)
+		}
+	}
+
+	return "Unknown"
+}
+
+func (llm *ggufModel) FileType() string {
+	t, ok := llm.kv["general.file_type"].(uint32)
+	if ok {
+		return fileType(t)
+	}
+
+	return "Unknown"
+}
+
+func (llm *ggufModel) Decode(r io.Reader) error {
+	read := llm.readString
+	if llm.Version == 1 {
+		read = llm.readStringV1
+	}
+
+	for i := 0; uint64(i) < llm.NumKV(); i++ {
+		k, err := read(r)
+		if err != nil {
+			return err
+		}
+
+		vtype := llm.readU32(r)
+
+		var v any
+		switch vtype {
+		case ggufTypeUint8:
+			v = llm.readU8(r)
+		case ggufTypeInt8:
+			v = llm.readI8(r)
+		case ggufTypeUint16:
+			v = llm.readU16(r)
+		case ggufTypeInt16:
+			v = llm.readI16(r)
+		case ggufTypeUint32:
+			v = llm.readU32(r)
+		case ggufTypeInt32:
+			v = llm.readI32(r)
+		case ggufTypeUint64:
+			v = llm.readU64(r)
+		case ggufTypeInt64:
+			v = llm.readI64(r)
+		case ggufTypeFloat32:
+			v = llm.readF32(r)
+		case ggufTypeFloat64:
+			v = llm.readF64(r)
+		case ggufTypeBool:
+			v = llm.readBool(r)
+		case ggufTypeString:
+			fn := llm.readString
+			if llm.Version == 1 {
+				fn = llm.readStringV1
+			}
+
+			s, err := fn(r)
+			if err != nil {
+				return err
+			}
+
+			v = s
+		case ggufTypeArray:
+			fn := llm.readArray
+			if llm.Version == 1 {
+				fn = llm.readArrayV1
+			}
+
+			a, err := fn(r)
+			if err != nil {
+				return err
+			}
+
+			v = a
+		default:
+			return fmt.Errorf("invalid type: %d", vtype)
+		}
+
+		llm.kv[k] = v
+	}
+
+	return nil
+}
+
+func (ggufModel) readU8(r io.Reader) uint8 {
+	var u8 uint8
+	binary.Read(r, binary.LittleEndian, &u8)
+	return u8
+}
+
+func (ggufModel) readI8(r io.Reader) int8 {
+	var i8 int8
+	binary.Read(r, binary.LittleEndian, &i8)
+	return i8
+}
+
+func (ggufModel) readU16(r io.Reader) uint16 {
+	var u16 uint16
+	binary.Read(r, binary.LittleEndian, &u16)
+	return u16
+}
+
+func (ggufModel) readI16(r io.Reader) int16 {
+	var i16 int16
+	binary.Read(r, binary.LittleEndian, &i16)
+	return i16
+}
+
+func (ggufModel) readU32(r io.Reader) uint32 {
+	var u32 uint32
+	binary.Read(r, binary.LittleEndian, &u32)
+	return u32
+}
+
+func (ggufModel) readI32(r io.Reader) int32 {
+	var i32 int32
+	binary.Read(r, binary.LittleEndian, &i32)
+	return i32
+}
+
+func (ggufModel) readU64(r io.Reader) uint64 {
+	var u64 uint64
+	binary.Read(r, binary.LittleEndian, &u64)
+	return u64
+}
+
+func (ggufModel) readI64(r io.Reader) int64 {
+	var i64 int64
+	binary.Read(r, binary.LittleEndian, &i64)
+	return i64
+}
+
+func (ggufModel) readF32(r io.Reader) float32 {
+	var f32 float32
+	binary.Read(r, binary.LittleEndian, &f32)
+	return f32
+}
+
+func (ggufModel) readF64(r io.Reader) float64 {
+	var f64 float64
+	binary.Read(r, binary.LittleEndian, &f64)
+	return f64
+}
+
+func (ggufModel) readBool(r io.Reader) bool {
+	var b bool
+	binary.Read(r, binary.LittleEndian, &b)
+	return b
+}
+
+func (ggufModel) readStringV1(r io.Reader) (string, error) {
+	var nameLength uint32
+	binary.Read(r, binary.LittleEndian, &nameLength)
+
+	var b bytes.Buffer
+	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {
+		return "", err
+	}
+
+	// gguf v1 strings are null-terminated
+	b.Truncate(b.Len() - 1)
+
+	return b.String(), nil
+}
+
+func (llm ggufModel) readString(r io.Reader) (string, error) {
+	var nameLength uint64
+	binary.Read(r, binary.LittleEndian, &nameLength)
+
+	var b bytes.Buffer
+	if _, err := io.CopyN(&b, r, int64(nameLength)); err != nil {
+		return "", err
+	}
+
+	return b.String(), nil
+}
+
+func (llm *ggufModel) readArrayV1(r io.Reader) (arr []any, err error) {
+	atype := llm.readU32(r)
+	n := llm.readU32(r)
+
+	for i := 0; uint32(i) < n; i++ {
+		switch atype {
+		case ggufTypeUint8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeInt8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeUint16:
+			arr = append(arr, llm.readU16(r))
+		case ggufTypeInt16:
+			arr = append(arr, llm.readI16(r))
+		case ggufTypeUint32:
+			arr = append(arr, llm.readU32(r))
+		case ggufTypeInt32:
+			arr = append(arr, llm.readI32(r))
+		case ggufTypeFloat32:
+			arr = append(arr, llm.readF32(r))
+		case ggufTypeBool:
+			arr = append(arr, llm.readBool(r))
+		case ggufTypeString:
+			s, err := llm.readStringV1(r)
+			if err != nil {
+				return nil, err
+			}
+
+			arr = append(arr, s)
+		default:
+			return nil, fmt.Errorf("invalid array type: %d", atype)
+		}
+	}
+
+	return
+}
+
+func (llm *ggufModel) readArray(r io.Reader) (arr []any, err error) {
+	atype := llm.readU32(r)
+	n := llm.readU64(r)
+
+	for i := 0; uint64(i) < n; i++ {
+		switch atype {
+		case ggufTypeUint8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeInt8:
+			arr = append(arr, llm.readU8(r))
+		case ggufTypeUint16:
+			arr = append(arr, llm.readU16(r))
+		case ggufTypeInt16:
+			arr = append(arr, llm.readI16(r))
+		case ggufTypeUint32:
+			arr = append(arr, llm.readU32(r))
+		case ggufTypeInt32:
+			arr = append(arr, llm.readI32(r))
+		case ggufTypeUint64:
+			arr = append(arr, llm.readU64(r))
+		case ggufTypeInt64:
+			arr = append(arr, llm.readI64(r))
+		case ggufTypeFloat32:
+			arr = append(arr, llm.readF32(r))
+		case ggufTypeFloat64:
+			arr = append(arr, llm.readF64(r))
+		case ggufTypeBool:
+			arr = append(arr, llm.readBool(r))
+		case ggufTypeString:
+			s, err := llm.readString(r)
+			if err != nil {
+				return nil, err
+			}
+
+			arr = append(arr, s)
+		default:
+			return nil, fmt.Errorf("invalid array type: %d", atype)
+		}
+	}
+
+	return
+}
+
+var (
+	ggufGPU = path.Join("llama.cpp", "gguf", "build", "gpu", "bin")
+	ggufCPU = path.Join("llama.cpp", "gguf", "build", "cpu", "bin")
+)
+
+var (
+	ggufInit       sync.Once
+	ggufRunnerPath string
+)
+
+func ggufRunner() ModelRunner {
+	ggufInit.Do(func() {
+		ggufRunnerPath = chooseRunner(ggufGPU, ggufCPU)
+	})
+
+	return ModelRunner{Path: ggufRunnerPath}
+}

llm/llama.cpp/generate.go

@@ -4,10 +4,14 @@
 package llm
 
 //go:generate git submodule init
+
 //go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_K_QUANTS=on
 //go:generate cmake --build ggml/build/cpu --target server --config Release
+
+//go:generate git submodule update --force gguf
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_K_QUANTS=on
+//go:generate cmake --build gguf/build/cpu --target server --config Release

llm/llama.cpp/generate_darwin_amd64.go

@@ -1,10 +1,16 @@
 package llm
 
 //go:generate git submodule init
+
 //go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate git-apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git-apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake -S ggml -B ggml/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build ggml/build/cpu --target server --config Release
+
+//go:generate git submodule update --force gguf
+//go:generate cmake -S gguf -B gguf/build/cpu -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=x86_64 -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake --build gguf/build/cpu --target server --config Release

llm/llama.cpp/generate_darwin_arm64.go

@@ -1,10 +1,16 @@
 package llm
 
 //go:generate git submodule init
+
 //go:generate git submodule update --force ggml
-//go:generate git -C ggml apply ../ggml_patch/0001-add-detokenize-endpoint.patch
-//go:generate git -C ggml apply ../ggml_patch/0002-34B-model-support.patch
-//go:generate git -C ggml apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
-//go:generate git -C ggml apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
-//go:generate cmake --fresh -S ggml -B ggml/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate git-apply ../ggml_patch/0003-metal-fix-synchronization-in-new-matrix-multiplicati.patch
+//go:generate git-apply ../ggml_patch/0004-metal-add-missing-barriers-for-mul-mat-2699.patch
+//go:generate cmake -S ggml -B ggml/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
 //go:generate cmake --build ggml/build/gpu --target server --config Release
+
+//go:generate git submodule update --force gguf
+//go:generate cmake -S gguf -B gguf/build/gpu -DLLAMA_METAL=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on -DCMAKE_SYSTEM_PROCESSOR=arm64 -DCMAKE_OSX_ARCHITECTURES=arm64 -DCMAKE_OSX_DEPLOYMENT_TARGET=11.0
+//go:generate cmake --build gguf/build/gpu --target server --config Release

llm/llama.cpp/generate_linux.go

@@ -0,0 +1,15 @@
+package llm
+
+//go:generate git submodule init
+
+//go:generate git submodule update --force ggml
+//go:generate -command git-apply git -C ggml apply
+//go:generate git-apply ../ggml_patch/0001-add-detokenize-endpoint.patch
+//go:generate git-apply ../ggml_patch/0002-34B-model-support.patch
+//go:generate git-apply ../ggml_patch/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch
+//go:generate cmake -S ggml -B ggml/build/gpu -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
+//go:generate cmake --build ggml/build/gpu --target server --config Release
+
+//go:generate git submodule update --force gguf
+//go:generate cmake -S gguf -B gguf/build/gpu -DLLAMA_CUBLAS=on -DLLAMA_ACCELERATE=on -DLLAMA_K_QUANTS=on
+//go:generate cmake --build gguf/build/gpu --target server --config Release

llm/llama.cpp/ggml_patch/0005-ggml-support-CUDA-s-half-type-for-aarch64-1455-2670.patch

@@ -0,0 +1,32 @@
+From 1e3bc523d8053a77df3ac7126a84d0297ee97ef6 Mon Sep 17 00:00:00 2001
+From: Kylin <56434533+KyL0N@users.noreply.github.com>
+Date: Tue, 22 Aug 2023 15:14:23 +0800
+Subject: [PATCH] ggml : support CUDA's half type for aarch64(#1455) (#2670)
+
+* ggml: support CUDA's half type for aarch64(#1455)
+support CUDA's half type for aarch64 in ggml_fp16_t definition
+
+* ggml: use __CUDACC__ to recognise nvcc compiler
+---
+ ggml.h | 5 +++--
+ 1 file changed, 3 insertions(+), 2 deletions(-)
+
+diff --git a/ggml.h b/ggml.h
+index 544ad2d..0ec7ec5 100644
+--- a/ggml.h
++++ b/ggml.h
+@@ -259,8 +259,9 @@
+ extern "C" {
+ #endif
+ 
+-#ifdef __ARM_NEON
+-    // we use the built-in 16-bit float type
++#if defined(__ARM_NEON) && defined(__CUDACC__)
++    typedef half ggml_fp16_t;
++#elif defined(__ARM_NEON)
+     typedef __fp16 ggml_fp16_t;
+ #else
+     typedef uint16_t ggml_fp16_t;
+-- 
+2.39.2 (Apple Git-143)
+

llm/llama.go

@@ -20,27 +20,14 @@ import (
 	"runtime"
 	"strconv"
 	"strings"
-	"sync"
 	"time"
 
 	"github.com/jmorganca/ollama/api"
 )
 
-const ModelFamilyLlama ModelFamily = "llama"
-
-//go:embed llama.cpp/ggml/build/*/bin/*
+//go:embed llama.cpp/*/build/*/bin/*
 var llamaCppEmbed embed.FS
 
-var (
-	ggmlGPU = path.Join("llama.cpp", "ggml", "build", "gpu", "bin")
-	ggmlCPU = path.Join("llama.cpp", "ggml", "build", "cpu", "bin")
-)
-
-var (
-	ggmlInit       sync.Once
-	ggmlRunnerPath string
-)
-
 func osPath(llamaPath string) string {
 	if runtime.GOOS == "windows" {
 		return path.Join(llamaPath, "Release")
@@ -49,98 +36,98 @@ func osPath(llamaPath string) string {
 	return llamaPath
 }
 
-func initGGML() {
-	ggmlInit.Do(func() {
-		tmpDir, err := os.MkdirTemp("", "llama-*")
-		if err != nil {
-			log.Fatalf("llama.cpp: failed to create temp dir: %v", err)
-		}
+func chooseRunner(gpuPath, cpuPath string) string {
+	tmpDir, err := os.MkdirTemp("", "llama-*")
+	if err != nil {
+		log.Fatalf("llama.cpp: failed to create temp dir: %v", err)
+	}
 
-		llamaPath := osPath(ggmlGPU)
+	llamaPath := osPath(gpuPath)
+	if _, err := fs.Stat(llamaCppEmbed, llamaPath); err != nil {
+		llamaPath = osPath(cpuPath)
 		if _, err := fs.Stat(llamaCppEmbed, llamaPath); err != nil {
-			llamaPath = osPath(ggmlCPU)
-			if _, err := fs.Stat(llamaCppEmbed, llamaPath); err != nil {
-				log.Fatalf("llama.cpp executable not found")
-			}
+			log.Fatalf("llama.cpp executable not found")
 		}
+	}
 
-		files := []string{"server"}
-		switch runtime.GOOS {
-		case "windows":
-			files = []string{"server.exe"}
-		case "darwin":
-			if llamaPath == osPath(ggmlGPU) {
-				files = append(files, "ggml-metal.metal")
-			}
+	files := []string{"server"}
+	switch runtime.GOOS {
+	case "windows":
+		files = []string{"server.exe"}
+	case "darwin":
+		if llamaPath == osPath(gpuPath) {
+			files = append(files, "ggml-metal.metal")
 		}
-
-		for _, f := range files {
-			srcPath := path.Join(llamaPath, f)
-			destPath := filepath.Join(tmpDir, f)
-
-			srcFile, err := llamaCppEmbed.Open(srcPath)
-			if err != nil {
-				log.Fatalf("read llama.cpp %s: %v", f, err)
-			}
-			defer srcFile.Close()
-
-			destFile, err := os.OpenFile(destPath, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-			if err != nil {
-				log.Fatalf("write llama.cpp %s: %v", f, err)
-			}
-			defer destFile.Close()
-
-			if _, err := io.Copy(destFile, srcFile); err != nil {
-				log.Fatalf("copy llama.cpp %s: %v", f, err)
-			}
+	case "linux":
+		// check if there is a GPU available
+		if _, err := CheckVRAM(); errors.Is(err, errNoGPU) {
+			// this error was logged on start-up, so we don't need to log it again
+			llamaPath = osPath(cpuPath)
 		}
+	}
 
-		ggmlRunnerPath = filepath.Join(tmpDir, "server")
-		if runtime.GOOS == "windows" {
-			ggmlRunnerPath = filepath.Join(tmpDir, "server.exe")
+	for _, f := range files {
+		srcPath := path.Join(llamaPath, f)
+		destPath := filepath.Join(tmpDir, f)
+
+		srcFile, err := llamaCppEmbed.Open(srcPath)
+		if err != nil {
+			log.Fatalf("read llama.cpp %s: %v", f, err)
 		}
-	})
-}
+		defer srcFile.Close()
 
-type ModelRunner struct {
-	Path string // path to the model runner executable
-}
+		destFile, err := os.OpenFile(destPath, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+		if err != nil {
+			log.Fatalf("write llama.cpp %s: %v", f, err)
+		}
+		defer destFile.Close()
 
-func ggmlRunner() ModelRunner {
-	initGGML()
-	return ModelRunner{Path: ggmlRunnerPath}
+		if _, err := io.Copy(destFile, srcFile); err != nil {
+			log.Fatalf("copy llama.cpp %s: %v", f, err)
+		}
+	}
+
+	runPath := filepath.Join(tmpDir, "server")
+	if runtime.GOOS == "windows" {
+		runPath = filepath.Join(tmpDir, "server.exe")
+	}
+
+	return runPath
 }
 
 type llamaModel struct {
 	hyperparameters llamaHyperparameters
 }
 
-func (llm *llamaModel) ModelFamily() ModelFamily {
-	return ModelFamilyLlama
+func (llm *llamaModel) ModelFamily() string {
+	return "llama"
 }
 
-func (llm *llamaModel) ModelType() ModelType {
-	switch llm.hyperparameters.NumLayer {
+func llamaModelType(numLayer uint32) string {
+	switch numLayer {
 	case 26:
-		return ModelType3B
+		return "3B"
 	case 32:
-		return ModelType7B
+		return "7B"
 	case 40:
-		return ModelType13B
+		return "13B"
 	case 48:
-		return ModelType34B
+		return "34B"
 	case 60:
-		return ModelType30B
+		return "30B"
 	case 80:
-		return ModelType65B
+		return "65B"
+	default:
+		return "Unknown"
 	}
-
-	// TODO: find a better default
-	return ModelType7B
 }
 
-func (llm *llamaModel) FileType() FileType {
-	return llm.hyperparameters.FileType
+func (llm *llamaModel) ModelType() string {
+	return llamaModelType(llm.hyperparameters.NumLayer)
+}
+
+func (llm *llamaModel) FileType() string {
+	return fileType(llm.hyperparameters.FileType)
 }
 
 type llamaHyperparameters struct {
@@ -157,70 +144,7 @@ type llamaHyperparameters struct {
 	NumRot   uint32
 
 	// FileType describes the quantization level of the model, e.g. Q4_0, Q5_K, etc.
-	FileType llamaFileType
-}
-
-type llamaFileType uint32
-
-const (
-	llamaFileTypeF32 llamaFileType = iota
-	llamaFileTypeF16
-	llamaFileTypeQ4_0
-	llamaFileTypeQ4_1
-	llamaFileTypeQ4_1_F16
-	llamaFileTypeQ8_0 llamaFileType = iota + 2
-	llamaFileTypeQ5_0
-	llamaFileTypeQ5_1
-	llamaFileTypeQ2_K
-	llamaFileTypeQ3_K_S
-	llamaFileTypeQ3_K_M
-	llamaFileTypeQ3_K_L
-	llamaFileTypeQ4_K_S
-	llamaFileTypeQ4_K_M
-	llamaFileTypeQ5_K_S
-	llamaFileTypeQ5_K_M
-	llamaFileTypeQ6_K
-)
-
-func (ft llamaFileType) String() string {
-	switch ft {
-	case llamaFileTypeF32:
-		return "F32"
-	case llamaFileTypeF16:
-		return "F16"
-	case llamaFileTypeQ4_0:
-		return "Q4_0"
-	case llamaFileTypeQ4_1:
-		return "Q4_1"
-	case llamaFileTypeQ4_1_F16:
-		return "Q4_1_F16"
-	case llamaFileTypeQ8_0:
-		return "Q8_0"
-	case llamaFileTypeQ5_0:
-		return "Q5_0"
-	case llamaFileTypeQ5_1:
-		return "Q5_1"
-	case llamaFileTypeQ2_K:
-		return "Q2_K"
-	case llamaFileTypeQ3_K_S:
-		return "Q3_K_S"
-	case llamaFileTypeQ3_K_M:
-		return "Q3_K_M"
-	case llamaFileTypeQ3_K_L:
-		return "Q3_K_L"
-	case llamaFileTypeQ4_K_S:
-		return "Q4_K_S"
-	case llamaFileTypeQ4_K_M:
-		return "Q4_K_M"
-	case llamaFileTypeQ5_K_S:
-		return "Q5_K_S"
-	case llamaFileTypeQ5_K_M:
-		return "Q5_K_M"
-	case llamaFileTypeQ6_K:
-		return "Q6_K"
-	default:
-		return "Unknown"
-	}
+	FileType uint32
 }
 
 type Running struct {
@@ -229,11 +153,81 @@ type Running struct {
 	Cancel context.CancelFunc
 }
 
+type ModelRunner struct {
+	Path string // path to the model runner executable
+}
+
 type llama struct {
 	api.Options
 	Running
 }
 
+var errNoGPU = errors.New("nvidia-smi command failed")
+
+// CheckVRAM returns the available VRAM in MiB on Linux machines with NVIDIA GPUs
+func CheckVRAM() (int, error) {
+	cmd := exec.Command("nvidia-smi", "--query-gpu=memory.total", "--format=csv,noheader,nounits")
+	var stdout bytes.Buffer
+	cmd.Stdout = &stdout
+	err := cmd.Run()
+	if err != nil {
+		return 0, errNoGPU
+	}
+
+	var total int
+	scanner := bufio.NewScanner(&stdout)
+	for scanner.Scan() {
+		line := scanner.Text()
+		vram, err := strconv.Atoi(line)
+		if err != nil {
+			return 0, fmt.Errorf("failed to parse available VRAM: %v", err)
+		}
+
+		total += vram
+	}
+
+	return total, nil
+}
+
+func NumGPU(opts api.Options) int {
+	if opts.NumGPU != -1 {
+		return opts.NumGPU
+	}
+	n := 1 // default to enable metal on macOS
+	if runtime.GOOS == "linux" {
+		vram, err := CheckVRAM()
+		if err != nil {
+			if err.Error() != "nvidia-smi command failed" {
+				log.Print(err.Error())
+			}
+			// nvidia driver not installed or no nvidia GPU found
+			return 0
+		}
+		// TODO: this is a very rough heuristic, better would be to calculate this based on number of layers and context size
+		switch {
+		case vram < 500:
+			log.Printf("WARNING: Low VRAM detected, disabling GPU")
+			n = 0
+		case vram < 1000:
+			n = 4
+		case vram < 2000:
+			n = 8
+		case vram < 4000:
+			n = 12
+		case vram < 8000:
+			n = 16
+		case vram < 12000:
+			n = 24
+		case vram < 16000:
+			n = 32
+		default:
+			n = 48
+		}
+		log.Printf("%d MB VRAM available, loading %d GPU layers", vram, n)
+	}
+	return n
+}
+
 func newLlama(model string, adapters []string, runner ModelRunner, opts api.Options) (*llama, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
@@ -250,14 +244,17 @@ func newLlama(model string, adapters []string, runner ModelRunner, opts api.Opti
 	params := []string{
 		"--model", model,
 		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
-		"--gqa", fmt.Sprintf("%d", opts.NumGQA),
 		"--rope-freq-base", fmt.Sprintf("%f", opts.RopeFrequencyBase),
 		"--rope-freq-scale", fmt.Sprintf("%f", opts.RopeFrequencyScale),
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
-		"--n-gpu-layers", fmt.Sprintf("%d", opts.NumGPU),
+		"--n-gpu-layers", fmt.Sprintf("%d", NumGPU(opts)),
 		"--embedding",
 	}
 
+	if opts.NumGQA > 0 {
+		params = append(params, "--gqa", fmt.Sprintf("%d", opts.NumGQA))
+	}
+
 	if len(adapters) > 0 {
 		// TODO: applying multiple adapters is not supported by the llama.cpp server yet
 		params = append(params, "--lora", adapters[0])
@@ -289,17 +286,25 @@ func newLlama(model string, adapters []string, runner ModelRunner, opts api.Opti
 			runner.Path,
 			append(params, "--port", strconv.Itoa(port))...,
 		)
+
 		cmd.Stdout = os.Stderr
 		cmd.Stderr = os.Stderr
 
 		llm := &llama{Options: opts, Running: Running{Port: port, Cmd: cmd, Cancel: cancel}}
 
+		log.Print("starting llama.cpp server")
+		if err := llm.Cmd.Start(); err != nil {
+			log.Printf("error starting the external llama.cpp server: %v", err)
+			continue
+		}
+
 		if err := waitForServer(llm); err != nil {
 			log.Printf("error starting llama.cpp server: %v", err)
 			llm.Close()
 			// try again
 			continue
 		}
+
 		// server started successfully
 		return llm, nil
 	}
@@ -308,48 +313,31 @@ func newLlama(model string, adapters []string, runner ModelRunner, opts api.Opti
 }
 
 func waitForServer(llm *llama) error {
-	log.Print("starting llama.cpp server")
-	var stderr bytes.Buffer
-	llm.Cmd.Stderr = &stderr
-	err := llm.Cmd.Start()
-	if err != nil {
-		return fmt.Errorf("error starting the external llama.cpp server: %w", err)
-	}
-
-	exitChan := make(chan error, 1)
-
-	// the server is a long running process, watch for it exiting to keep track of something going wrong
-	go func() {
-		err := llm.Cmd.Wait()
-		log.Print(stderr.String())
-		exitChan <- err
-	}()
-
 	// wait for the server to start responding
 	start := time.Now()
-	expiresAt := time.Now().Add(30 * time.Second)
-	ticker := time.NewTicker(100 * time.Millisecond)
+	expiresAt := time.Now().Add(45 * time.Second)
+	ticker := time.NewTicker(200 * time.Millisecond)
 
 	log.Print("waiting for llama.cpp server to start responding")
+	for range ticker.C {
+		if time.Now().After(expiresAt) {
+			return fmt.Errorf("llama.cpp server did not start within alloted time, retrying")
+		}
 
-	for {
-		select {
-		case <-ticker.C:
-			if time.Now().After(expiresAt) {
-				return fmt.Errorf("llama.cpp server did not start responding within 30 seconds, retrying")
-			}
-			if err := llm.Ping(context.Background()); err == nil {
-				log.Printf("llama.cpp server started in %f seconds", time.Since(start).Seconds())
-				return nil
-			}
-		case err := <-exitChan:
-			return fmt.Errorf("llama.cpp server exited unexpectedly: %w", err)
+		if err := llm.Ping(context.Background()); err == nil {
+			break
 		}
 	}
+
+	log.Printf("llama.cpp server started in %f seconds", time.Since(start).Seconds())
+	return nil
 }
 
 func (llm *llama) Close() {
-	llm.Running.Cmd.Cancel()
+	llm.Cancel()
+	if err := llm.Cmd.Wait(); err != nil {
+		log.Printf("llama.cpp server exited with error: %v", err)
+	}
 }
 
 func (llm *llama) SetOptions(opts api.Options) {
@@ -676,7 +664,7 @@ func (llm *llama) Embedding(ctx context.Context, input string) ([]float64, error
 
 // Ping checks that the server subprocess is still running and responding to requests
 func (llm *llama) Ping(ctx context.Context) error {
-	resp, err := http.Head(fmt.Sprintf("http://127.0.0.1:%d", llm.Running.Port))
+	resp, err := http.Head(fmt.Sprintf("http://127.0.0.1:%d", llm.Port))
 	if err != nil {
 		return fmt.Errorf("ping resp: %w", err)
 	}

llm/llm.go

@@ -32,15 +32,22 @@ func New(model string, adapters []string, opts api.Options) (LLM, error) {
 	}
 	defer f.Close()
 
-	ggml, err := DecodeGGML(f, ModelFamilyLlama)
+	ggml, err := DecodeGGML(f)
 	if err != nil {
 		return nil, err
 	}
 
-	switch ggml.FileType().String() {
-	case "F32", "Q5_0", "Q5_1", "Q8_0":
+	switch ggml.FileType() {
+	case "Q8_0":
+		if ggml.Name() != "gguf" && opts.NumGPU != 0 {
+			// GGML Q8_0 do not support Metal API and will
+			// cause the runner to segmentation fault so disable GPU
+			log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
+			opts.NumGPU = 0
+		}
+	case "F32", "Q5_0", "Q5_1":
 		if opts.NumGPU != 0 {
-			// F32, F16, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
+			// F32, Q5_0, Q5_1, and Q8_0 do not support Metal API and will
 			// cause the runner to segmentation fault so disable GPU
 			log.Printf("WARNING: GPU disabled for F32, Q5_0, Q5_1, and Q8_0")
 			opts.NumGPU = 0
@@ -49,34 +56,43 @@ func New(model string, adapters []string, opts api.Options) (LLM, error) {
 
 	totalResidentMemory := memory.TotalMemory()
 	switch ggml.ModelType() {
-	case ModelType3B, ModelType7B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 16*1024*1024 {
+	case "3B", "7B":
+		if ggml.FileType() == "F16" && totalResidentMemory < 16*1024*1024 {
 			return nil, fmt.Errorf("F16 model requires at least 16GB of memory")
 		} else if totalResidentMemory < 8*1024*1024 {
 			return nil, fmt.Errorf("model requires at least 8GB of memory")
 		}
-	case ModelType13B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 32*1024*1024 {
+	case "13B":
+		if ggml.FileType() == "F16" && totalResidentMemory < 32*1024*1024 {
 			return nil, fmt.Errorf("F16 model requires at least 32GB of memory")
 		} else if totalResidentMemory < 16*1024*1024 {
 			return nil, fmt.Errorf("model requires at least 16GB of memory")
 		}
-	case ModelType30B, ModelType34B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 64*1024*1024 {
+	case "30B", "34B", "40B":
+		if ggml.FileType() == "F16" && totalResidentMemory < 64*1024*1024 {
 			return nil, fmt.Errorf("F16 model requires at least 64GB of memory")
 		} else if totalResidentMemory < 32*1024*1024 {
 			return nil, fmt.Errorf("model requires at least 32GB of memory")
 		}
-	case ModelType65B:
-		if ggml.FileType().String() == "F16" && totalResidentMemory < 128*1024*1024 {
+	case "65B", "70B":
+		if ggml.FileType() == "F16" && totalResidentMemory < 128*1024*1024 {
 			return nil, fmt.Errorf("F16 model requires at least 128GB of memory")
 		} else if totalResidentMemory < 64*1024*1024 {
 			return nil, fmt.Errorf("model requires at least 64GB of memory")
 		}
+	case "180B":
+		if ggml.FileType() == "F16" && totalResidentMemory < 512*1024*1024 {
+			return nil, fmt.Errorf("F16 model requires at least 512GB of memory")
+		} else if totalResidentMemory < 128*1024*1024 {
+			return nil, fmt.Errorf("model requires at least 128GB of memory")
+		}
 	}
 
-	switch ggml.ModelFamily() {
-	case ModelFamilyLlama:
+	switch ggml.Name() {
+	case "gguf":
+		opts.NumGQA = 0 // TODO: remove this when llama.cpp runners differ enough to need separate newLlama functions
+		return newLlama(model, adapters, ggufRunner(), opts)
+	case "ggml", "ggmf", "ggjt", "ggla":
 		return newLlama(model, adapters, ggmlRunner(), opts)
 	default:
 		return nil, fmt.Errorf("unknown ggml type: %s", ggml.ModelFamily())

scripts/build_darwin.sh

@@ -8,7 +8,7 @@ GO_LDFLAGS="$GO_LDFLAGS -X github.com/jmorganca/ollama/server.mode=release"
 # build universal binary
 GOARCH=arm64 go generate ./...
 GOARCH=arm64 go build -ldflags "$GO_LDFLAGS" -o dist/ollama-darwin-arm64
-rm -rf llm/llama.cpp/ggml/build/*/bin
+rm -rf llm/llama.cpp/*/build/*/bin
 GOARCH=amd64 go generate ./...
 GOARCH=amd64 go build -ldflags "$GO_LDFLAGS" -o dist/ollama-darwin-amd64
 lipo -create -output dist/ollama dist/ollama-darwin-arm64 dist/ollama-darwin-amd64

server/auth.go

@@ -103,7 +103,7 @@ func getAuthToken(ctx context.Context, redirData AuthRedirect, regOpts *Registry
 
 	headers := make(http.Header)
 	headers.Set("Authorization", sig)
-	resp, err := makeRequest(ctx, "GET", redirectURL, headers, nil, regOpts)
+	resp, err := makeRequest(ctx, "GET", redirectURL, headers, nil, nil)
 	if err != nil {
 		log.Printf("couldn't get token: %q", err)
 	}

server/images.go

@@ -114,10 +114,11 @@ type LayerReader struct {
 }
 
 type ConfigV2 struct {
-	ModelFamily llm.ModelFamily `json:"model_family"`
-	ModelType   string          `json:"model_type"`
-	FileType    string          `json:"file_type"`
-	RootFS      RootFS          `json:"rootfs"`
+	ModelFormat string `json:"model_format"`
+	ModelFamily string `json:"model_family"`
+	ModelType   string `json:"model_type"`
+	FileType    string `json:"file_type"`
+	RootFS      RootFS `json:"rootfs"`
 
 	// required by spec
 	Architecture string `json:"architecture"`
@@ -268,6 +269,29 @@ func filenameWithPath(path, f string) (string, error) {
 }
 
 func CreateModel(ctx context.Context, name string, path string, fn func(resp api.ProgressResponse)) error {
+	mp := ParseModelPath(name)
+
+	var manifest *ManifestV2
+	var err error
+	var noprune string
+
+	// build deleteMap to prune unused layers
+	deleteMap := make(map[string]bool)
+
+	if noprune = os.Getenv("OLLAMA_NOPRUNE"); noprune == "" {
+		manifest, _, err = GetManifest(mp)
+		if err != nil && !errors.Is(err, os.ErrNotExist) {
+			return err
+		}
+
+		if manifest != nil {
+			for _, l := range manifest.Layers {
+				deleteMap[l.Digest] = true
+			}
+			deleteMap[manifest.Config.Digest] = true
+		}
+	}
+
 	mf, err := os.Open(path)
 	if err != nil {
 		fn(api.ProgressResponse{Status: fmt.Sprintf("couldn't open modelfile '%s'", path)})
@@ -328,14 +352,15 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 					}
 					defer file.Close()
 
-					ggml, err := llm.DecodeGGML(file, llm.ModelFamilyLlama)
+					ggml, err := llm.DecodeGGML(file)
 					if err != nil {
 						return err
 					}
 
+					config.ModelFormat = ggml.Name()
 					config.ModelFamily = ggml.ModelFamily()
-					config.ModelType = ggml.ModelType().String()
-					config.FileType = ggml.FileType().String()
+					config.ModelType = ggml.ModelType()
+					config.FileType = ggml.FileType()
 
 					// reset the file
 					file.Seek(0, io.SeekStart)
@@ -369,6 +394,7 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 				// copie the model metadata
 				config.ModelFamily = source.ModelFamily
 				config.ModelType = source.ModelType
+				config.ModelFormat = source.ModelFormat
 				config.FileType = source.FileType
 
 				for _, l := range mf.Layers {
@@ -472,6 +498,12 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 			}
 		}
 
+		if config.ModelType == "65B" {
+			if numGQA, ok := formattedParams["num_gqa"].(int); ok && numGQA == 8 {
+				config.ModelType = "70B"
+			}
+		}
+
 		bts, err := json.Marshal(formattedParams)
 		if err != nil {
 			return err
@@ -503,6 +535,7 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 	var manifestLayers []*Layer
 	for _, l := range layers {
 		manifestLayers = append(manifestLayers, &l.Layer)
+		delete(deleteMap, l.Layer.Digest)
 	}
 
 	// Create a layer for the config object
@@ -512,6 +545,7 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 		return err
 	}
 	layers = append(layers, cfg)
+	delete(deleteMap, cfg.Layer.Digest)
 
 	if err := SaveLayers(layers, fn, false); err != nil {
 		return err
@@ -524,6 +558,14 @@ func CreateModel(ctx context.Context, name string, path string, fn func(resp api
 		return err
 	}
 
+	if noprune == "" {
+		fn(api.ProgressResponse{Status: "removing any unused layers"})
+		err = deleteUnusedLayers(nil, deleteMap, false)
+		if err != nil {
+			return err
+		}
+	}
+
 	fn(api.ProgressResponse{Status: "success"})
 	return nil
 }
@@ -779,14 +821,14 @@ func formatParams(params map[string][]string) (map[string]interface{}, error) {
 						return nil, fmt.Errorf("invalid float value %s", vals)
 					}
 
-					out[key] = floatVal
+					out[key] = float32(floatVal)
 				case reflect.Int:
-					intVal, err := strconv.ParseInt(vals[0], 10, 0)
+					intVal, err := strconv.ParseInt(vals[0], 10, 64)
 					if err != nil {
 						return nil, fmt.Errorf("invalid int value %s", vals)
 					}
 
-					out[key] = intVal
+					out[key] = int(intVal)
 				case reflect.Bool:
 					boolVal, err := strconv.ParseBool(vals[0])
 					if err != nil {
@@ -866,18 +908,7 @@ func CopyModel(src, dest string) error {
 	return nil
 }
 
-func DeleteModel(name string) error {
-	mp := ParseModelPath(name)
-	manifest, _, err := GetManifest(mp)
-	if err != nil {
-		return err
-	}
-	deleteMap := make(map[string]bool)
-	for _, layer := range manifest.Layers {
-		deleteMap[layer.Digest] = true
-	}
-	deleteMap[manifest.Config.Digest] = true
-
+func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]bool, dryRun bool) error {
 	fp, err := GetManifestPath()
 	if err != nil {
 		return err
@@ -894,14 +925,13 @@ func DeleteModel(name string) error {
 		fmp := ParseModelPath(tag)
 
 		// skip the manifest we're trying to delete
-		if mp.GetFullTagname() == fmp.GetFullTagname() {
+		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 {
-			log.Printf("skipping file: %s", fp)
 			return nil
 		}
 
@@ -925,14 +955,72 @@ func DeleteModel(name string) error {
 				log.Printf("couldn't get file path for '%s': %v", k, err)
 				continue
 			}
-			if err := os.Remove(fp); err != nil {
-				log.Printf("couldn't remove file '%s': %v", fp, err)
-				continue
+			if !dryRun {
+				if err := os.Remove(fp); err != nil {
+					log.Printf("couldn't remove file '%s': %v", fp, err)
+					continue
+				}
+			} else {
+				log.Printf("wanted to remove: %s", fp)
 			}
 		}
 	}
 
-	fp, err = mp.GetManifestPath(false)
+	return nil
+}
+
+func PruneLayers() error {
+	deleteMap := make(map[string]bool)
+	p, err := GetBlobsPath("")
+	if err != nil {
+		return err
+	}
+
+	blobs, err := os.ReadDir(p)
+	if err != nil {
+		log.Printf("couldn't read dir '%s': %v", p, err)
+		return err
+	}
+
+	for _, blob := range blobs {
+		name := blob.Name()
+		if runtime.GOOS == "windows" {
+			name = strings.ReplaceAll(name, "-", ":")
+		}
+		deleteMap[name] = true
+	}
+
+	log.Printf("total blobs: %d", len(deleteMap))
+
+	err = deleteUnusedLayers(nil, deleteMap, false)
+	if err != nil {
+		return err
+	}
+
+	log.Printf("total unused blobs removed: %d", len(deleteMap))
+
+	return nil
+}
+
+func DeleteModel(name string) error {
+	mp := ParseModelPath(name)
+	manifest, _, err := GetManifest(mp)
+	if err != nil {
+		return err
+	}
+
+	deleteMap := make(map[string]bool)
+	for _, layer := range manifest.Layers {
+		deleteMap[layer.Digest] = true
+	}
+	deleteMap[manifest.Config.Digest] = true
+
+	err = deleteUnusedLayers(&mp, deleteMap, false)
+	if err != nil {
+		return err
+	}
+
+	fp, err := mp.GetManifestPath(false)
 	if err != nil {
 		return err
 	}
@@ -1111,13 +1199,34 @@ 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 *ManifestV2
+	var err error
+	var noprune string
+
+	// build deleteMap to prune unused layers
+	deleteMap := make(map[string]bool)
+
+	if noprune = os.Getenv("OLLAMA_NOPRUNE"); noprune == "" {
+		manifest, _, err = GetManifest(mp)
+		if err != nil && !errors.Is(err, os.ErrNotExist) {
+			return err
+		}
+
+		if manifest != nil {
+			for _, l := range manifest.Layers {
+				deleteMap[l.Digest] = true
+			}
+			deleteMap[manifest.Config.Digest] = true
+		}
+	}
+
 	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
 		return fmt.Errorf("insecure protocol http")
 	}
 
 	fn(api.ProgressResponse{Status: "pulling manifest"})
 
-	manifest, err := pullModelManifest(ctx, mp, regOpts)
+	manifest, err = pullModelManifest(ctx, mp, regOpts)
 	if err != nil {
 		return fmt.Errorf("pull model manifest: %s", err)
 	}
@@ -1137,7 +1246,9 @@ func PullModel(ctx context.Context, name string, regOpts *RegistryOptions, fn fu
 			}); err != nil {
 			return err
 		}
+		delete(deleteMap, layer.Digest)
 	}
+	delete(deleteMap, manifest.Config.Digest)
 
 	fn(api.ProgressResponse{Status: "verifying sha256 digest"})
 	for _, layer := range layers {
@@ -1175,6 +1286,14 @@ 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, false)
+		if err != nil {
+			return err
+		}
+	}
+
 	fn(api.ProgressResponse{Status: "success"})
 
 	return nil
@@ -1300,7 +1419,7 @@ func makeRequestWithRetry(ctx context.Context, method string, requestURL *url.UR
 }
 
 func makeRequest(ctx context.Context, method string, requestURL *url.URL, headers http.Header, body io.Reader, regOpts *RegistryOptions) (*http.Response, error) {
-	if requestURL.Scheme != "http" && regOpts.Insecure {
+	if requestURL.Scheme != "http" && regOpts != nil && regOpts.Insecure {
 		requestURL.Scheme = "http"
 	}
 
@@ -1313,10 +1432,12 @@ func makeRequest(ctx context.Context, method string, requestURL *url.URL, header
 		req.Header = headers
 	}
 
-	if regOpts.Token != "" {
-		req.Header.Set("Authorization", "Bearer "+regOpts.Token)
-	} else if regOpts.Username != "" && regOpts.Password != "" {
-		req.SetBasicAuth(regOpts.Username, regOpts.Password)
+	if regOpts != nil {
+		if regOpts.Token != "" {
+			req.Header.Set("Authorization", "Bearer "+regOpts.Token)
+		} else if regOpts.Username != "" && regOpts.Password != "" {
+			req.SetBasicAuth(regOpts.Username, regOpts.Password)
+		}
 	}
 
 	req.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))

server/modelpath.go

@@ -133,7 +133,12 @@ func GetBlobsPath(digest string) (string, error) {
 	}
 
 	path := filepath.Join(home, ".ollama", "models", "blobs", digest)
-	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+	dirPath := filepath.Dir(path)
+	if digest == "" {
+		dirPath = path
+	}
+
+	if err := os.MkdirAll(dirPath, 0o755); err != nil {
 		return "", err
 	}
 

server/routes.go

@@ -12,6 +12,7 @@ import (
 	"os/signal"
 	"path/filepath"
 	"reflect"
+	"runtime"
 	"strconv"
 	"strings"
 	"sync"
@@ -363,6 +364,7 @@ func DeleteModelHandler(c *gin.Context) {
 		}
 		return
 	}
+	c.JSON(http.StatusOK, nil)
 }
 
 func ShowModelHandler(c *gin.Context) {
@@ -547,6 +549,13 @@ func Serve(ln net.Listener, origins []string) error {
 		os.Exit(0)
 	}()
 
+	if runtime.GOOS == "linux" {
+		// check compatibility to log warnings
+		if _, err := llm.CheckVRAM(); err != nil {
+			log.Printf("Warning: GPU support not enabled, you may need to install GPU drivers: %v", err)
+		}
+	}
+
 	return s.Serve(ln)
 }
 

server/upload.go

@@ -66,31 +66,39 @@ func uploadBlobChunked(ctx context.Context, requestURL *url.URL, layer *Layer, r
 
 		sectionReader := io.NewSectionReader(f, int64(offset), chunk)
 		for try := 0; try < MaxRetries; try++ {
+			ch := make(chan error, 1)
+
 			r, w := io.Pipe()
 			defer r.Close()
 			go func() {
 				defer w.Close()
 
 				for chunked := int64(0); chunked < chunk; {
-					n, err := io.CopyN(w, sectionReader, 1024*1024)
-					if err != nil && !errors.Is(err, io.EOF) {
+					select {
+					case err := <-ch:
+						log.Printf("chunk interrupted: %v", err)
+						return
+					default:
+						n, err := io.CopyN(w, sectionReader, 1024*1024)
+						if err != nil && !errors.Is(err, io.EOF) {
+							fn(api.ProgressResponse{
+								Status:    fmt.Sprintf("error reading chunk: %v", err),
+								Digest:    layer.Digest,
+								Total:     layer.Size,
+								Completed: int(offset),
+							})
+
+							return
+						}
+
+						chunked += n
 						fn(api.ProgressResponse{
-							Status:    fmt.Sprintf("error reading chunk: %v", err),
+							Status:    fmt.Sprintf("uploading %s", layer.Digest),
 							Digest:    layer.Digest,
 							Total:     layer.Size,
-							Completed: int(offset),
+							Completed: int(offset) + int(chunked),
 						})
-
-						return
 					}
-
-					chunked += n
-					fn(api.ProgressResponse{
-						Status:    fmt.Sprintf("uploading %s", layer.Digest),
-						Digest:    layer.Digest,
-						Total:     layer.Size,
-						Completed: int(offset) + int(chunked),
-					})
 				}
 			}()
 
@@ -113,6 +121,8 @@ func uploadBlobChunked(ctx context.Context, requestURL *url.URL, layer *Layer, r
 
 			switch {
 			case resp.StatusCode == http.StatusUnauthorized:
+				ch <- errors.New("unauthorized")
+
 				auth := resp.Header.Get("www-authenticate")
 				authRedir := ParseAuthRedirectString(auth)
 				token, err := getAuthToken(ctx, authRedir, regOpts)
@@ -121,10 +131,7 @@ func uploadBlobChunked(ctx context.Context, requestURL *url.URL, layer *Layer, r
 				}
 
 				regOpts.Token = token
-				if _, err := sectionReader.Seek(0, io.SeekStart); err != nil {
-					return err
-				}
-
+				sectionReader = io.NewSectionReader(f, int64(offset), chunk)
 				continue
 			case resp.StatusCode >= http.StatusBadRequest:
 				body, _ := io.ReadAll(resp.Body)