api embed docs (#5282)

Refine error reporting for subprocess crash

.github/workflows/release.yaml

@@ -147,7 +147,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP"
@@ -437,6 +437,7 @@ jobs:
     env:
       OLLAMA_SKIP_IMAGE_BUILD: '1'
       PUSH: '1'
+      GH_TOKEN: ${{ github.token }}
     steps:
       - uses: actions/checkout@v4
       - name: Set Version
@@ -460,15 +461,20 @@ jobs:
           ls -lh dist/
           (cd dist; sha256sum * > sha256sum.txt)
           cat dist/sha256sum.txt
-      - uses: ncipollo/release-action@v1
-        with:
-          name: ${{ env.RELEASE_VERSION }}
-          allowUpdates: true
-          artifacts: 'dist/*'
-          draft: true
-          prerelease: true
-          omitBodyDuringUpdate: true
-          generateReleaseNotes: true
-          omitDraftDuringUpdate: true
-          omitPrereleaseDuringUpdate: true
-          replacesArtifacts: true
+      - name: Create or update Release
+        run: |
+          echo "Looking for existing release for ${{ env.RELEASE_VERSION }}"
+          OLD_TAG=$(gh release ls --json name,tagName | jq -r ".[] | select(.name == \"${{ env.RELEASE_VERSION }}\") | .tagName")
+          if [ -n "$OLD_TAG" ]; then
+            echo "Updating release ${{ env.RELEASE_VERSION }} to point to new tag ${GITHUB_REF_NAME}"
+            gh release edit ${OLD_TAG} --tag ${GITHUB_REF_NAME}
+          else
+            echo "Creating new release ${{ env.RELEASE_VERSION }} pointing to tag ${GITHUB_REF_NAME}"
+            gh release create ${GITHUB_REF_NAME} \
+              --title ${{ env.RELEASE_VERSION }} \
+              --draft \
+              --generate-notes \
+              --prerelease
+          fi
+          echo "Uploading artifacts for tag ${GITHUB_REF_NAME}"
+          gh release upload ${GITHUB_REF_NAME} dist/* --clobber

.github/workflows/test.yaml

@@ -58,6 +58,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     env:
       GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: '1'
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
@@ -79,6 +80,7 @@ jobs:
       - run: go generate -x ./...
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
         name: 'Unix Go Generate'
+      - run: go build .
       - uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
@@ -124,7 +126,7 @@ jobs:
     strategy:
       matrix:
         rocm-version:
-          - '6.1.1'
+          - '6.1.2'
     runs-on: linux
     container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
     steps:
@@ -167,7 +169,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP"

Dockerfile

@@ -2,7 +2,7 @@ ARG GOLANG_VERSION=1.22.1
 ARG CMAKE_VERSION=3.22.1
 # this CUDA_VERSION corresponds with the one specified in docs/gpu.md
 ARG CUDA_VERSION=11.3.1
-ARG ROCM_VERSION=6.1.1
+ARG ROCM_VERSION=6.1.2
 
 # Copy the minimal context we need to run the generate scripts
 FROM scratch AS llm-code
@@ -70,12 +70,12 @@ RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
 RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
 
-FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64
+FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS

README.md

@@ -53,8 +53,8 @@ Here are some example models that can be downloaded:
 | Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
 | Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
 | Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
-| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
-| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
+| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`            |
+| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`        |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
 | Moondream 2        | 1.4B       | 829MB | `ollama run moondream`         |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
@@ -182,6 +182,12 @@ $ ollama run llama3 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
+### Show model information
+
+```
+ollama show llama3
+```
+
 ### List models on your computer
 
 ```
@@ -286,6 +292,10 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
+- [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama)
+- [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
+- [AI Studio](https://github.com/MindWorkAI/AI-Studio)
+- [Sidellama](https://github.com/gyopak/sidellama) (browser-based LLM client)
 
 ### Terminal
 

api/client.go

@@ -347,7 +347,16 @@ func (c *Client) Heartbeat(ctx context.Context) error {
 	return nil
 }
 
-// Embeddings generates embeddings from a model.
+// Embed generates embeddings from a model.
+func (c *Client) Embed(ctx context.Context, req *EmbedRequest) (*EmbedResponse, error) {
+	var resp EmbedResponse
+	if err := c.do(ctx, http.MethodPost, "/api/embed", req, &resp); err != nil {
+		return nil, err
+	}
+	return &resp, nil
+}
+
+// Embeddings generates an embedding from a model.
 func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*EmbeddingResponse, error) {
 	var resp EmbeddingResponse
 	if err := c.do(ctx, http.MethodPost, "/api/embeddings", req, &resp); err != nil {

api/types.go

@@ -47,6 +47,9 @@ type GenerateRequest struct {
 	// Prompt is the textual prompt to send to the model.
 	Prompt string `json:"prompt"`
 
+	// Suffix is the text that comes after the inserted text.
+	Suffix string `json:"suffix"`
+
 	// System overrides the model's default system message/prompt.
 	System string `json:"system"`
 
@@ -97,17 +100,80 @@ type ChatRequest struct {
 	// followin the request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
+	// Tools is an optional list of tools the model has access to.
+	Tools `json:"tools,omitempty"`
+
 	// Options lists model-specific options.
 	Options map[string]interface{} `json:"options"`
 }
 
+type Tools []Tool
+
+func (t Tools) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
 // Message is a single message in a chat sequence. The message contains the
 // role ("system", "user", or "assistant"), the content and an optional list
 // of images.
 type Message struct {
-	Role    string      `json:"role"`
-	Content string      `json:"content"`
-	Images  []ImageData `json:"images,omitempty"`
+	Role      string      `json:"role"`
+	Content   string      `json:"content"`
+	Images    []ImageData `json:"images,omitempty"`
+	ToolCalls []ToolCall  `json:"tool_calls,omitempty"`
+}
+
+func (m *Message) UnmarshalJSON(b []byte) error {
+	type Alias Message
+	var a Alias
+	if err := json.Unmarshal(b, &a); err != nil {
+		return err
+	}
+
+	*m = Message(a)
+	m.Role = strings.ToLower(m.Role)
+	return nil
+}
+
+type ToolCall struct {
+	Function ToolCallFunction `json:"function"`
+}
+
+type ToolCallFunction struct {
+	Name      string                    `json:"name"`
+	Arguments ToolCallFunctionArguments `json:"arguments"`
+}
+
+type ToolCallFunctionArguments map[string]any
+
+func (t *ToolCallFunctionArguments) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
+type Tool struct {
+	Type     string       `json:"type"`
+	Function ToolFunction `json:"function"`
+}
+
+type ToolFunction struct {
+	Name        string `json:"name"`
+	Description string `json:"description"`
+	Parameters  struct {
+		Type       string   `json:"type"`
+		Required   []string `json:"required"`
+		Properties map[string]struct {
+			Type        string   `json:"type"`
+			Description string   `json:"description"`
+			Enum        []string `json:"enum,omitempty"`
+		} `json:"properties"`
+	} `json:"parameters"`
+}
+
+func (t *ToolFunction) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
 }
 
 // ChatResponse is the response returned by [Client.Chat]. Its fields are
@@ -159,18 +225,42 @@ type Options struct {
 
 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	UseNUMA   bool `json:"numa,omitempty"`
-	NumCtx    int  `json:"num_ctx,omitempty"`
-	NumBatch  int  `json:"num_batch,omitempty"`
-	NumGPU    int  `json:"num_gpu,omitempty"`
-	MainGPU   int  `json:"main_gpu,omitempty"`
-	LowVRAM   bool `json:"low_vram,omitempty"`
-	F16KV     bool `json:"f16_kv,omitempty"`
-	LogitsAll bool `json:"logits_all,omitempty"`
-	VocabOnly bool `json:"vocab_only,omitempty"`
-	UseMMap   bool `json:"use_mmap,omitempty"`
-	UseMLock  bool `json:"use_mlock,omitempty"`
-	NumThread int  `json:"num_thread,omitempty"`
+	UseNUMA   bool  `json:"numa,omitempty"`
+	NumCtx    int   `json:"num_ctx,omitempty"`
+	NumBatch  int   `json:"num_batch,omitempty"`
+	NumGPU    int   `json:"num_gpu,omitempty"`
+	MainGPU   int   `json:"main_gpu,omitempty"`
+	LowVRAM   bool  `json:"low_vram,omitempty"`
+	F16KV     bool  `json:"f16_kv,omitempty"`
+	LogitsAll bool  `json:"logits_all,omitempty"`
+	VocabOnly bool  `json:"vocab_only,omitempty"`
+	UseMMap   *bool `json:"use_mmap,omitempty"`
+	UseMLock  bool  `json:"use_mlock,omitempty"`
+	NumThread int   `json:"num_thread,omitempty"`
+}
+
+// EmbedRequest is the request passed to [Client.Embed].
+type EmbedRequest struct {
+	// Model is the model name.
+	Model string `json:"model"`
+
+	// Input is the input to embed.
+	Input any `json:"input"`
+
+	// KeepAlive controls how long the model will stay loaded in memory following
+	// this request.
+	KeepAlive *Duration `json:"keep_alive,omitempty"`
+
+	Truncate *bool `json:"truncate,omitempty"`
+
+	// Options lists model-specific options.
+	Options map[string]interface{} `json:"options"`
+}
+
+// EmbedResponse is the response from [Client.Embed].
+type EmbedResponse struct {
+	Model      string      `json:"model"`
+	Embeddings [][]float32 `json:"embeddings"`
 }
 
 // EmbeddingRequest is the request passed to [Client.Embeddings].
@@ -219,9 +309,12 @@ type DeleteRequest struct {
 
 // ShowRequest is the request passed to [Client.Show].
 type ShowRequest struct {
-	Model    string `json:"model"`
-	System   string `json:"system"`
+	Model  string `json:"model"`
+	System string `json:"system"`
+
+	// Template is deprecated
 	Template string `json:"template"`
+	Verbose  bool   `json:"verbose"`
 
 	Options map[string]interface{} `json:"options"`
 
@@ -231,13 +324,16 @@ type ShowRequest struct {
 
 // ShowResponse is the response returned from [Client.Show].
 type ShowResponse struct {
-	License    string       `json:"license,omitempty"`
-	Modelfile  string       `json:"modelfile,omitempty"`
-	Parameters string       `json:"parameters,omitempty"`
-	Template   string       `json:"template,omitempty"`
-	System     string       `json:"system,omitempty"`
-	Details    ModelDetails `json:"details,omitempty"`
-	Messages   []Message    `json:"messages,omitempty"`
+	License       string         `json:"license,omitempty"`
+	Modelfile     string         `json:"modelfile,omitempty"`
+	Parameters    string         `json:"parameters,omitempty"`
+	Template      string         `json:"template,omitempty"`
+	System        string         `json:"system,omitempty"`
+	Details       ModelDetails   `json:"details,omitempty"`
+	Messages      []Message      `json:"messages,omitempty"`
+	ModelInfo     map[string]any `json:"model_info,omitempty"`
+	ProjectorInfo map[string]any `json:"projector_info,omitempty"`
+	ModifiedAt    time.Time      `json:"modified_at,omitempty"`
 }
 
 // CopyRequest is the request passed to [Client.Copy].
@@ -310,6 +406,13 @@ type ProcessModelResponse struct {
 	SizeVRAM  int64        `json:"size_vram"`
 }
 
+type RetrieveModelResponse struct {
+	Id      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	OwnedBy string `json:"owned_by"`
+}
+
 type TokenResponse struct {
 	Token string `json:"token"`
 }
@@ -448,6 +551,17 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					slice[i] = str
 				}
 				field.Set(reflect.ValueOf(slice))
+			case reflect.Pointer:
+				var b bool
+				if field.Type() == reflect.TypeOf(&b) {
+					val, ok := val.(bool)
+					if !ok {
+						return fmt.Errorf("option %q must be of type boolean", key)
+					}
+					field.Set(reflect.ValueOf(&val))
+				} else {
+					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
+				}
 			default:
 				return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 			}
@@ -490,7 +604,7 @@ func DefaultOptions() Options {
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,
-			UseMMap:   true,
+			UseMMap:   nil,
 			UseNUMA:   false,
 		},
 	}
@@ -587,6 +701,17 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 				case reflect.Slice:
 					// TODO: only string slices are supported right now
 					out[key] = vals
+				case reflect.Pointer:
+					var b bool
+					if field.Type() == reflect.TypeOf(&b) {
+						boolVal, err := strconv.ParseBool(vals[0])
+						if err != nil {
+							return nil, fmt.Errorf("invalid bool value %s", vals)
+						}
+						out[key] = &boolVal
+					} else {
+						return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
+					}
 				default:
 					return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
 				}

api/types_test.go

@@ -2,6 +2,7 @@ package api
 
 import (
 	"encoding/json"
+	"fmt"
 	"math"
 	"testing"
 	"time"
@@ -105,3 +106,128 @@ func TestDurationMarshalUnmarshal(t *testing.T) {
 		})
 	}
 }
+
+func TestUseMmapParsingFromJSON(t *testing.T) {
+	tr := true
+	fa := false
+	tests := []struct {
+		name string
+		req  string
+		exp  *bool
+	}{
+		{
+			name: "Undefined",
+			req:  `{ }`,
+			exp:  nil,
+		},
+		{
+			name: "True",
+			req:  `{ "use_mmap": true }`,
+			exp:  &tr,
+		},
+		{
+			name: "False",
+			req:  `{ "use_mmap": false }`,
+			exp:  &fa,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var oMap map[string]interface{}
+			err := json.Unmarshal([]byte(test.req), &oMap)
+			require.NoError(t, err)
+			opts := DefaultOptions()
+			err = opts.FromMap(oMap)
+			require.NoError(t, err)
+			assert.Equal(t, test.exp, opts.UseMMap)
+		})
+	}
+}
+
+func TestUseMmapFormatParams(t *testing.T) {
+	tr := true
+	fa := false
+	tests := []struct {
+		name string
+		req  map[string][]string
+		exp  *bool
+		err  error
+	}{
+		{
+			name: "True",
+			req: map[string][]string{
+				"use_mmap": {"true"},
+			},
+			exp: &tr,
+			err: nil,
+		},
+		{
+			name: "False",
+			req: map[string][]string{
+				"use_mmap": {"false"},
+			},
+			exp: &fa,
+			err: nil,
+		},
+		{
+			name: "Numeric True",
+			req: map[string][]string{
+				"use_mmap": {"1"},
+			},
+			exp: &tr,
+			err: nil,
+		},
+		{
+			name: "Numeric False",
+			req: map[string][]string{
+				"use_mmap": {"0"},
+			},
+			exp: &fa,
+			err: nil,
+		},
+		{
+			name: "invalid string",
+			req: map[string][]string{
+				"use_mmap": {"foo"},
+			},
+			exp: nil,
+			err: fmt.Errorf("invalid bool value [foo]"),
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			resp, err := FormatParams(test.req)
+			require.Equal(t, test.err, err)
+			respVal, ok := resp["use_mmap"]
+			if test.exp != nil {
+				assert.True(t, ok, "resp: %v", resp)
+				assert.Equal(t, *test.exp, *respVal.(*bool))
+			}
+		})
+	}
+}
+
+func TestMessage_UnmarshalJSON(t *testing.T) {
+	tests := []struct {
+		input    string
+		expected string
+	}{
+		{`{"role": "USER", "content": "Hello!"}`, "user"},
+		{`{"role": "System", "content": "Initialization complete."}`, "system"},
+		{`{"role": "assistant", "content": "How can I help you?"}`, "assistant"},
+		{`{"role": "TOOl", "content": "Access granted."}`, "tool"},
+	}
+
+	for _, test := range tests {
+		var msg Message
+		if err := json.Unmarshal([]byte(test.input), &msg); err != nil {
+			t.Errorf("Unexpected error: %v", err)
+		}
+
+		if msg.Role != test.expected {
+			t.Errorf("role not lowercased: got %v, expected %v", msg.Role, test.expected)
+		}
+	}
+}

app/lifecycle/logging.go

@@ -5,6 +5,8 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
+	"strconv"
+	"strings"
 
 	"github.com/ollama/ollama/envconfig"
 )
@@ -24,6 +26,7 @@ func InitLogging() {
 		logFile = os.Stderr
 		// TODO - write one-line to the app.log file saying we're running in console mode to help avoid confusion
 	} else {
+		rotateLogs(AppLogFile)
 		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
 		if err != nil {
 			slog.Error(fmt.Sprintf("failed to create server log %v", err))
@@ -46,3 +49,32 @@ func InitLogging() {
 
 	slog.Info("ollama app started")
 }
+
+func rotateLogs(logFile string) {
+	if _, err := os.Stat(logFile); os.IsNotExist(err) {
+		return
+	}
+	index := strings.LastIndex(logFile, ".")
+	pre := logFile[:index]
+	post := "." + logFile[index+1:]
+	for i := LogRotationCount; i > 0; i-- {
+		older := pre + "-" + strconv.Itoa(i) + post
+		newer := pre + "-" + strconv.Itoa(i-1) + post
+		if i == 1 {
+			newer = pre + post
+		}
+		if _, err := os.Stat(newer); err == nil {
+			if _, err := os.Stat(older); err == nil {
+				err := os.Remove(older)
+				if err != nil {
+					slog.Warn("Failed to remove older log", "older", older, "error", err)
+					continue
+				}
+			}
+			err := os.Rename(newer, older)
+			if err != nil {
+				slog.Warn("Failed to rotate log", "older", older, "newer", newer, "error", err)
+			}
+		}
+	}
+}

app/lifecycle/logging_test.go

@@ -0,0 +1,44 @@
+package lifecycle
+
+import (
+	"os"
+	"path/filepath"
+	"strconv"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestRotateLogs(t *testing.T) {
+	logDir := t.TempDir()
+	logFile := filepath.Join(logDir, "testlog.log")
+
+	// No log exists
+	rotateLogs(logFile)
+
+	require.NoError(t, os.WriteFile(logFile, []byte("1"), 0644))
+	assert.FileExists(t, logFile)
+	// First rotation
+	rotateLogs(logFile)
+	assert.FileExists(t, filepath.Join(logDir, "testlog-1.log"))
+	assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log"))
+	assert.NoFileExists(t, logFile)
+
+	// Should be a no-op without a new log
+	rotateLogs(logFile)
+	assert.FileExists(t, filepath.Join(logDir, "testlog-1.log"))
+	assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log"))
+	assert.NoFileExists(t, logFile)
+
+	for i := 2; i <= LogRotationCount+1; i++ {
+		require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0644))
+		assert.FileExists(t, logFile)
+		rotateLogs(logFile)
+		assert.NoFileExists(t, logFile)
+		for j := 1; j < i; j++ {
+			assert.FileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(j)+".log"))
+		}
+		assert.NoFileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(i+1)+".log"))
+	}
+}

app/lifecycle/paths.go

@@ -16,11 +16,12 @@ var (
 	AppDir     = "/opt/Ollama"
 	AppDataDir = "/opt/Ollama"
 	// TODO - should there be a distinct log dir?
-	UpdateStageDir = "/tmp"
-	AppLogFile     = "/tmp/ollama_app.log"
-	ServerLogFile  = "/tmp/ollama.log"
-	UpgradeLogFile = "/tmp/ollama_update.log"
-	Installer      = "OllamaSetup.exe"
+	UpdateStageDir   = "/tmp"
+	AppLogFile       = "/tmp/ollama_app.log"
+	ServerLogFile    = "/tmp/ollama.log"
+	UpgradeLogFile   = "/tmp/ollama_update.log"
+	Installer        = "OllamaSetup.exe"
+	LogRotationCount = 5
 )
 
 func init() {

app/lifecycle/server.go

@@ -54,7 +54,7 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) {
 		return nil, fmt.Errorf("failed to spawn server stderr pipe: %w", err)
 	}
 
-	// TODO - rotation
+	rotateLogs(ServerLogFile)
 	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
 	if err != nil {
 		return nil, fmt.Errorf("failed to create server log: %w", err)

app/ollama.iss

@@ -88,10 +88,15 @@ 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}\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
 Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\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
@@ -122,6 +127,10 @@ Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\models"
 Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\history"
 ; NOTE: if the user has a custom OLLAMA_MODELS it will be preserved
 
+[InstallDelete]
+Type: filesandordirs; Name: "{%TEMP}\ollama*"
+Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama"
+
 [Messages]
 WizardReady=Ollama Windows Preview
 ReadyLabel1=%nLet's get you up and running with your own large language models.

cmd/cmd.go

@@ -162,9 +162,6 @@ func tempZipFiles(path string) (string, error) {
 	}
 	defer tempfile.Close()
 
-	zipfile := zip.NewWriter(tempfile)
-	defer zipfile.Close()
-
 	detectContentType := func(path string) (string, error) {
 		f, err := os.Open(path)
 		if err != nil {
@@ -233,6 +230,9 @@ func tempZipFiles(path string) (string, error) {
 		files = append(files, tks...)
 	}
 
+	zipfile := zip.NewWriter(tempfile)
+	defer zipfile.Close()
+
 	for _, file := range files {
 		f, err := os.Open(file)
 		if err != nil {
@@ -287,38 +287,12 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er
 }
 
 func RunHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	name := args[0]
-
-	// check if the model exists on the server
-	show, err := client.Show(cmd.Context(), &api.ShowRequest{Name: name})
-	var statusError api.StatusError
-	switch {
-	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
-		if err := PullHandler(cmd, []string{name}); err != nil {
-			return err
-		}
-
-		show, err = client.Show(cmd.Context(), &api.ShowRequest{Name: name})
-		if err != nil {
-			return err
-		}
-	case err != nil:
-		return err
-	}
-
 	interactive := true
 
 	opts := runOptions{
-		Model:       args[0],
-		WordWrap:    os.Getenv("TERM") == "xterm-256color",
-		Options:     map[string]interface{}{},
-		MultiModal:  slices.Contains(show.Details.Families, "clip"),
-		ParentModel: show.Details.ParentModel,
+		Model:    args[0],
+		WordWrap: os.Getenv("TERM") == "xterm-256color",
+		Options:  map[string]interface{}{},
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -362,11 +336,38 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.WordWrap = !nowrap
 
-	if !interactive {
-		return generate(cmd, opts)
+	// Fill out the rest of the options based on information about the
+	// model.
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
 	}
 
-	return generateInteractive(cmd, opts)
+	name := args[0]
+	info, err := func() (*api.ShowResponse, error) {
+		showReq := &api.ShowRequest{Name: name}
+		info, err := client.Show(cmd.Context(), showReq)
+		var se api.StatusError
+		if errors.As(err, &se) && se.StatusCode == http.StatusNotFound {
+			if err := PullHandler(cmd, []string{name}); err != nil {
+				return nil, err
+			}
+			return client.Show(cmd.Context(), &api.ShowRequest{Name: name})
+		}
+		return info, err
+	}()
+	if err != nil {
+		return err
+	}
+
+	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
+	opts.ParentModel = info.Details.ParentModel
+	opts.Messages = append(opts.Messages, info.Messages...)
+
+	if interactive {
+		return generateInteractive(cmd, opts)
+	}
+	return generate(cmd, opts)
 }
 
 func errFromUnknownKey(unknownKeyErr error) error {
@@ -579,10 +580,6 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	if len(args) != 1 {
-		return errors.New("missing model name")
-	}
-
 	license, errLicense := cmd.Flags().GetBool("license")
 	modelfile, errModelfile := cmd.Flags().GetBool("modelfile")
 	parameters, errParams := cmd.Flags().GetBool("parameters")
@@ -625,8 +622,6 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 
 	if flagsSet > 1 {
 		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
-	} else if flagsSet == 0 {
-		return errors.New("one of '--license', '--modelfile', '--parameters', '--system', or '--template' must be specified")
 	}
 
 	req := api.ShowRequest{Name: args[0]}
@@ -635,22 +630,141 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	switch showType {
-	case "license":
-		fmt.Println(resp.License)
-	case "modelfile":
-		fmt.Println(resp.Modelfile)
-	case "parameters":
-		fmt.Println(resp.Parameters)
-	case "system":
-		fmt.Println(resp.System)
-	case "template":
-		fmt.Println(resp.Template)
+	if flagsSet == 1 {
+		switch showType {
+		case "license":
+			fmt.Println(resp.License)
+		case "modelfile":
+			fmt.Println(resp.Modelfile)
+		case "parameters":
+			fmt.Println(resp.Parameters)
+		case "system":
+			fmt.Println(resp.System)
+		case "template":
+			fmt.Println(resp.Template)
+		}
+
+		return nil
 	}
 
+	showInfo(resp)
+
 	return nil
 }
 
+func showInfo(resp *api.ShowResponse) {
+	arch := resp.ModelInfo["general.architecture"].(string)
+
+	modelData := [][]string{
+		{"arch", arch},
+		{"parameters", resp.Details.ParameterSize},
+		{"quantization", resp.Details.QuantizationLevel},
+		{"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))},
+		{"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))},
+	}
+
+	mainTableData := [][]string{
+		{"Model"},
+		{renderSubTable(modelData, false)},
+	}
+
+	if resp.ProjectorInfo != nil {
+		projectorData := [][]string{
+			{"arch", "clip"},
+			{"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))},
+		}
+
+		if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok {
+			projectorData = append(projectorData, []string{"projector type", projectorType.(string)})
+		}
+
+		projectorData = append(projectorData,
+			[]string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))},
+			[]string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))},
+		)
+
+		mainTableData = append(mainTableData,
+			[]string{"Projector"},
+			[]string{renderSubTable(projectorData, false)},
+		)
+	}
+
+	if resp.Parameters != "" {
+		mainTableData = append(mainTableData, []string{"Parameters"}, []string{formatParams(resp.Parameters)})
+	}
+
+	if resp.System != "" {
+		mainTableData = append(mainTableData, []string{"System"}, []string{renderSubTable(twoLines(resp.System), true)})
+	}
+
+	if resp.License != "" {
+		mainTableData = append(mainTableData, []string{"License"}, []string{renderSubTable(twoLines(resp.License), true)})
+	}
+
+	table := tablewriter.NewWriter(os.Stdout)
+	table.SetAutoWrapText(false)
+	table.SetBorder(false)
+	table.SetAlignment(tablewriter.ALIGN_LEFT)
+
+	for _, v := range mainTableData {
+		table.Append(v)
+	}
+
+	table.Render()
+}
+
+func renderSubTable(data [][]string, file bool) string {
+	var buf bytes.Buffer
+	table := tablewriter.NewWriter(&buf)
+	table.SetAutoWrapText(!file)
+	table.SetBorder(false)
+	table.SetNoWhiteSpace(true)
+	table.SetTablePadding("\t")
+	table.SetAlignment(tablewriter.ALIGN_LEFT)
+
+	for _, v := range data {
+		table.Append(v)
+	}
+
+	table.Render()
+
+	renderedTable := buf.String()
+	lines := strings.Split(renderedTable, "\n")
+	for i, line := range lines {
+		lines[i] = "\t" + line
+	}
+
+	return strings.Join(lines, "\n")
+}
+
+func twoLines(s string) [][]string {
+	lines := strings.Split(s, "\n")
+	res := [][]string{}
+
+	count := 0
+	for _, line := range lines {
+		line = strings.TrimSpace(line)
+		if line != "" {
+			count++
+			res = append(res, []string{line})
+			if count == 2 {
+				return res
+			}
+		}
+	}
+	return res
+}
+
+func formatParams(s string) string {
+	lines := strings.Split(s, "\n")
+	table := [][]string{}
+
+	for _, line := range lines {
+		table = append(table, strings.Fields(line))
+	}
+	return renderSubTable(table, false)
+}
+
 func CopyHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -729,7 +843,6 @@ type runOptions struct {
 	WordWrap    bool
 	Format      string
 	System      string
-	Template    string
 	Images      []api.ImageData
 	Options     map[string]interface{}
 	MultiModal  bool
@@ -923,7 +1036,6 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		Images:    opts.Images,
 		Format:    opts.Format,
 		System:    opts.System,
-		Template:  opts.Template,
 		Options:   opts.Options,
 		KeepAlive: opts.KeepAlive,
 	}
@@ -1232,7 +1344,6 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_LLM_LIBRARY"],
-				envVars["OLLAMA_MAX_VRAM"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

cmd/interactive.go

@@ -27,69 +27,43 @@ const (
 	MultilineNone MultilineState = iota
 	MultilinePrompt
 	MultilineSystem
-	MultilineTemplate
 )
 
 func loadModel(cmd *cobra.Command, opts *runOptions) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
 	p := progress.NewProgress(os.Stderr)
 	defer p.StopAndClear()
 
 	spinner := progress.NewSpinner("")
 	p.Add("", spinner)
 
-	showReq := api.ShowRequest{Name: opts.Model}
-	showResp, err := client.Show(cmd.Context(), &showReq)
+	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		return err
 	}
-	opts.MultiModal = slices.Contains(showResp.Details.Families, "clip")
-	opts.ParentModel = showResp.Details.ParentModel
-
-	if len(showResp.Messages) > 0 {
-		opts.Messages = append(opts.Messages, showResp.Messages...)
-	}
 
 	chatReq := &api.ChatRequest{
-		Model:    opts.Model,
-		Messages: []api.Message{},
+		Model:     opts.Model,
+		KeepAlive: opts.KeepAlive,
 	}
 
-	if opts.KeepAlive != nil {
-		chatReq.KeepAlive = opts.KeepAlive
-	}
-
-	err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
+	return client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
 		p.StopAndClear()
-		if len(opts.Messages) > 0 {
-			for _, msg := range opts.Messages {
-				switch msg.Role {
-				case "user":
-					fmt.Printf(">>> %s\n", msg.Content)
-				case "assistant":
-					state := &displayResponseState{}
-					displayResponse(msg.Content, opts.WordWrap, state)
-					fmt.Println()
-					fmt.Println()
-				}
+		for _, msg := range opts.Messages {
+			switch msg.Role {
+			case "user":
+				fmt.Printf(">>> %s\n", msg.Content)
+			case "assistant":
+				state := &displayResponseState{}
+				displayResponse(msg.Content, opts.WordWrap, state)
+				fmt.Println()
+				fmt.Println()
 			}
 		}
 		return nil
 	})
-	if err != nil {
-		return err
-	}
-
-	return nil
 }
 
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
-	opts.Messages = make([]api.Message, 0)
-
 	err := loadModel(cmd, &opts)
 	if err != nil {
 		return err
@@ -119,7 +93,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
 		fmt.Fprintln(os.Stderr, "  /set parameter ...     Set a parameter")
 		fmt.Fprintln(os.Stderr, "  /set system <string>   Set system message")
-		fmt.Fprintln(os.Stderr, "  /set template <string> Set prompt template")
 		fmt.Fprintln(os.Stderr, "  /set history           Enable history")
 		fmt.Fprintln(os.Stderr, "  /set nohistory         Disable history")
 		fmt.Fprintln(os.Stderr, "  /set wordwrap          Enable wordwrap")
@@ -229,10 +202,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				opts.Messages = append(opts.Messages, api.Message{Role: "system", Content: opts.System})
 				fmt.Println("Set system message.")
 				sb.Reset()
-			case MultilineTemplate:
-				opts.Template = sb.String()
-				fmt.Println("Set prompt template.")
-				sb.Reset()
 			}
 
 			multiline = MultilineNone
@@ -351,17 +320,13 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					}
 					fmt.Printf("Set parameter '%s' to '%s'\n", args[2], strings.Join(params, ", "))
 					opts.Options[args[2]] = fp[args[2]]
-				case "system", "template":
+				case "system":
 					if len(args) < 3 {
 						usageSet()
 						continue
 					}
 
-					if args[1] == "system" {
-						multiline = MultilineSystem
-					} else if args[1] == "template" {
-						multiline = MultilineTemplate
-					}
+					multiline = MultilineSystem
 
 					line := strings.Join(args[2:], " ")
 					line, ok := strings.CutPrefix(line, `"""`)
@@ -381,23 +346,17 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						continue
 					}
 
-					if args[1] == "system" {
-						opts.System = sb.String() // for display in modelfile
-						newMessage := api.Message{Role: "system", Content: sb.String()}
-						// Check if the slice is not empty and the last message is from 'system'
-						if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" {
-							// Replace the last message
-							opts.Messages[len(opts.Messages)-1] = newMessage
-						} else {
-							opts.Messages = append(opts.Messages, newMessage)
-						}
-						fmt.Println("Set system message.")
-						sb.Reset()
-					} else if args[1] == "template" {
-						opts.Template = sb.String()
-						fmt.Println("Set prompt template.")
-						sb.Reset()
+					opts.System = sb.String() // for display in modelfile
+					newMessage := api.Message{Role: "system", Content: sb.String()}
+					// Check if the slice is not empty and the last message is from 'system'
+					if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" {
+						// Replace the last message
+						opts.Messages[len(opts.Messages)-1] = newMessage
+					} else {
+						opts.Messages = append(opts.Messages, newMessage)
 					}
+					fmt.Println("Set system message.")
+					sb.Reset()
 
 					sb.Reset()
 					continue
@@ -418,7 +377,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				req := &api.ShowRequest{
 					Name:     opts.Model,
 					System:   opts.System,
-					Template: opts.Template,
 					Options:  opts.Options,
 				}
 				resp, err := client.Show(cmd.Context(), req)
@@ -429,15 +387,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					fmt.Println("Model details:")
-					if len(resp.Details.Families) > 0 {
-						fmt.Printf("Family              %s\n", strings.Join(resp.Details.Families, ", "))
-					} else if resp.Details.Family != "" {
-						fmt.Printf("Family              %s\n", resp.Details.Family)
-					}
-					fmt.Printf("Parameter Size      %s\n", resp.Details.ParameterSize)
-					fmt.Printf("Quantization Level  %s\n", resp.Details.QuantizationLevel)
-					fmt.Println("")
+					showInfo(resp)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")
@@ -470,12 +420,9 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						fmt.Println("No system message was specified for this model.")
 					}
 				case "template":
-					switch {
-					case opts.Template != "":
-						fmt.Println(opts.Template + "\n")
-					case resp.Template != "":
+					if resp.Template != "" {
 						fmt.Println(resp.Template)
-					default:
+					} else {
 						fmt.Println("No prompt template was specified for this model.")
 					}
 				default:
@@ -569,10 +516,6 @@ func buildModelfile(opts runOptions) string {
 		fmt.Fprintf(&mf, "SYSTEM \"\"\"%s\"\"\"\n", opts.System)
 	}
 
-	if opts.Template != "" {
-		fmt.Fprintf(&mf, "TEMPLATE \"\"\"%s\"\"\"\n", opts.Template)
-	}
-
 	keys := make([]string, 0)
 	for k := range opts.Options {
 		keys = append(keys, k)

cmd/interactive_test.go

@@ -59,7 +59,6 @@ func TestModelfileBuilder(t *testing.T) {
 	opts := runOptions{
 		Model:    "hork",
 		System:   "You are part horse and part shark, but all hork. Do horklike things",
-		Template: "This is a template.",
 		Messages: []api.Message{
 			{Role: "user", Content: "Hey there hork!"},
 			{Role: "assistant", Content: "Yes it is true, I am half horse, half shark."},
@@ -75,7 +74,6 @@ func TestModelfileBuilder(t *testing.T) {
 	mf := buildModelfile(opts)
 	expectedModelfile := `FROM {{.Model}}
 SYSTEM """{{.System}}"""
-TEMPLATE """{{.Template}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
 PARAMETER stop [hi there]
@@ -97,7 +95,6 @@ MESSAGE assistant """Yes it is true, I am half horse, half shark."""
 	mf = buildModelfile(opts)
 	expectedModelfile = `FROM {{.ParentModel}}
 SYSTEM """{{.System}}"""
-TEMPLATE """{{.Template}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
 PARAMETER stop [hi there]

convert/mistral.go

@@ -71,6 +71,11 @@ func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
 		"tokenizer.ggml.unknown_token_id": uint32(0),
 	}
 
+	if m.Params.HeadDimension > 0 {
+		kv["llama.attention.key_length"] = uint32(m.Params.HeadDimension)
+		kv["llama.attention.value_length"] = uint32(m.Params.HeadDimension)
+	}
+
 	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }
 

docs/api.md

@@ -26,7 +26,7 @@ All durations are returned in nanoseconds.
 
 ### Streaming responses
 
-Certain endpoints stream responses as JSON objects and can optional return non-streamed responses.
+Certain endpoints stream responses as JSON objects. Streaming can be disabled by providing `{"stream": false}` for these endpoints.
 
 ## Generate a completion
 
@@ -777,11 +777,12 @@ A single JSON object will be returned.
 POST /api/show
 ```
 
-Show information about a model including details, modelfile, template, parameters, license, and system prompt.
+Show information about a model including details, modelfile, template, parameters, license, system prompt.
 
 ### Parameters
 
 - `name`: name of the model to show
+- `verbose`: (optional) if set to `true`, returns full data for verbose response fields
 
 ### Examples
 
@@ -798,14 +799,40 @@ curl http://localhost:11434/api/show -d '{
 ```json
 {
   "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"",
-  "parameters": "num_ctx                        4096\nstop                           \u003c/s\u003e\nstop                           USER:\nstop                           ASSISTANT:",
-  "template": "{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: ",
+  "parameters": "num_keep                       24\nstop                           \"<|start_header_id|>\"\nstop                           \"<|end_header_id|>\"\nstop                           \"<|eot_id|>\"",
+  "template": "{{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>",
   "details": {
+    "parent_model": "",
     "format": "gguf",
     "family": "llama",
-    "families": ["llama", "clip"],
-    "parameter_size": "7B",
+    "families": [
+      "llama"
+    ],
+    "parameter_size": "8.0B",
     "quantization_level": "Q4_0"
+  },
+  "model_info": {
+    "general.architecture": "llama",
+    "general.file_type": 2,
+    "general.parameter_count": 8030261248,
+    "general.quantization_version": 2,
+    "llama.attention.head_count": 32,
+    "llama.attention.head_count_kv": 8,
+    "llama.attention.layer_norm_rms_epsilon": 0.00001,
+    "llama.block_count": 32,
+    "llama.context_length": 8192,
+    "llama.embedding_length": 4096,
+    "llama.feed_forward_length": 14336,
+    "llama.rope.dimension_count": 128,
+    "llama.rope.freq_base": 500000,
+    "llama.vocab_size": 128256,
+    "tokenizer.ggml.bos_token_id": 128000,
+    "tokenizer.ggml.eos_token_id": 128009,
+    "tokenizer.ggml.merges": [],            // populates if `verbose=true`
+    "tokenizer.ggml.model": "gpt2",
+    "tokenizer.ggml.pre": "llama-bpe",
+    "tokenizer.ggml.token_type": [],        // populates if `verbose=true`
+    "tokenizer.ggml.tokens": []             // populates if `verbose=true`
   }
 }
 ```
@@ -999,7 +1026,7 @@ If `stream` is set to `false`, then the response is a single JSON object:
 ## Generate Embeddings
 
 ```shell
-POST /api/embeddings
+POST /api/embed
 ```
 
 Generate embeddings from a model
@@ -1007,10 +1034,11 @@ Generate embeddings from a model
 ### Parameters
 
 - `model`: name of model to generate embeddings from
-- `prompt`: text to generate embeddings for
+- `input`: text or list of text to generate embeddings for
 
 Advanced parameters:
 
+- `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
@@ -1019,9 +1047,9 @@ Advanced parameters:
 #### Request
 
 ```shell
-curl http://localhost:11434/api/embeddings -d '{
+curl http://localhost:11434/api/embed -d '{
   "model": "all-minilm",
-  "prompt": "Here is an article about llamas..."
+  "input": "Why is the sky blue?"
 }'
 ```
 
@@ -1029,10 +1057,35 @@ curl http://localhost:11434/api/embeddings -d '{
 
 ```json
 {
-  "embedding": [
-    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
-    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
-  ]
+  "model": "all-minilm",
+  "embeddings": [[
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ]]
+}
+```
+
+#### Request (Multiple input)
+
+```shell
+curl http://localhost:11434/api/embed -d '{
+  "model": "all-minilm",
+  "input": ["Why is the sky blue?", "Why is the grass green?"]
+}'
+```
+
+#### Response
+
+```json
+{
+  "model": "all-minilm",
+  "embeddings": [[
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ],[
+    -0.0098027075, 0.06042469, 0.025257962, -0.006364387, 0.07272725,
+    0.017194884, 0.09032035, -0.051705178, 0.09951512, 0.09072481
+  ]]
 }
 ```
 
@@ -1079,3 +1132,45 @@ A single JSON object will be returned.
   ]
 }
 ```
+
+## Generate Embedding
+
+> Note: this endpoint has been superseded by `/api/embed`
+
+```shell
+POST /api/embeddings
+```
+
+Generate embeddings from a model
+
+### Parameters
+
+- `model`: name of model to generate embeddings from
+- `prompt`: text to generate embeddings for
+
+Advanced parameters:
+
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+
+### Examples
+
+#### Request
+
+```shell
+curl http://localhost:11434/api/embeddings -d '{
+  "model": "all-minilm",
+  "prompt": "Here is an article about llamas..."
+}'
+```
+
+#### Response
+
+```json
+{
+  "embedding": [
+    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
+    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
+  ]
+}
+```

docs/development.md

@@ -104,7 +104,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
 you might use:
 
 ```
-OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
+OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
 go build .
 ```
 
@@ -114,15 +114,18 @@ If you have Docker available, you can build linux binaries with `./scripts/build
 
 ### Windows
 
-Note: The windows build for Ollama is still under development.
+Note: The Windows build for Ollama is still under development.
 
-Install required tools:
+First, install required tools:
 
 - MSVC toolchain - C/C++ and cmake as minimal requirements
 - Go version 1.22 or higher
 - MinGW (pick one variant) with GCC.
   - [MinGW-w64](https://www.mingw-w64.org/)
   - [MSYS2](https://www.msys2.org/)
+- The `ThreadJob` Powershell module: `Install-Module -Name ThreadJob -Scope CurrentUser`
+
+Then, build the `ollama` binary:
 
 ```powershell
 $env:CGO_ENABLED="1"

docs/faq.md

@@ -257,3 +257,19 @@ If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` AP
 ## How do I manage the maximum number of requests the Ollama server can queue?
 
 If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded.  You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`.
+
+## How does Ollama handle concurrent requests?
+
+Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing.
+
+If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded.  As prior models become idle, one or more will be unloaded to make room for the new model.  Queued requests will be processed in order.  When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads.
+
+Parallel request processing for a given model results in increasing the context size by the number of parallel requests.  For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation.
+
+The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
+
+- `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory.  The default is 3 * the number of GPUs or 3 for CPU inference.
+- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
+- `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
+
+Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.

docs/gpu.md

@@ -18,7 +18,7 @@ Check your compute compatibility to see if your card is supported:
 |                    | Quadro              | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000`                                                                 |
 | 7.0                | NVIDIA              | `TITAN V` `V100` `Quadro GV100`                                                                             |
 | 6.1                | NVIDIA TITAN        | `TITAN Xp` `TITAN X`                                                                                        |
-|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050`                                     |
+|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050 Ti` `GTX 1050`                       |
 |                    | Quadro              | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` |
 |                    | Tesla               | `P40` `P4`                                                                                                  |
 | 6.0                | NVIDIA              | `Tesla P100` `Quadro GP100`                                                                                 |
@@ -46,13 +46,24 @@ sudo modprobe nvidia_uvm`
 
 ## AMD Radeon
 Ollama supports the following AMD GPUs:
+
+### Linux Support
 | Family         | Cards and accelerators                                                                                                               |
 | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
 | AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
 | AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
 | AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
 
-### Overrides
+### Windows Support
+With ROCm v6.1, the following GPUs are supported on Windows.
+
+| Family         | Cards and accelerators                                                                                                               |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800`    |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` |
+
+
+### Overrides on Linux
 Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
 some cases you can force the system to try to use a similar LLVM target that is
 close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
@@ -63,7 +74,7 @@ would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
 server.  If you have an unsupported AMD GPU you can experiment using the list of
 supported types below.
 
-At this time, the known supported GPU types are the following LLVM Targets.
+At this time, the known supported GPU types on linux are the following LLVM Targets.
 This table shows some example GPUs that map to these LLVM targets:
 | **LLVM Target** | **An Example GPU** |
 |-----------------|---------------------|

docs/import.md

@@ -47,19 +47,13 @@ success
 
 ### Supported Quantizations
 
-<details>
-<summary>Legacy Quantization</summary>
-
 - `Q4_0`
 - `Q4_1`
 - `Q5_0`
 - `Q5_1`
 - `Q8_0`
 
-</details>
-
-<details>
-<summary>K-means Quantization</summary>`
+#### K-means Quantizations
 
 - `Q3_K_S`
 - `Q3_K_M`
@@ -70,11 +64,6 @@ success
 - `Q5_K_M`
 - `Q6_K`
 
-</details>
-
-> [!NOTE]
-> Activation-aware Weight Quantization (i.e. IQ) are not currently supported for automatic quantization however you can still import the quantized model into Ollama, see [Import GGUF](#import-gguf).
-
 ## Template Detection
 
 > [!NOTE]

docs/openai.md

@@ -65,6 +65,7 @@ curl http://localhost:11434/v1/chat/completions \
             }
         ]
     }'
+
 ```
 
 ## Endpoints
@@ -102,11 +103,6 @@ curl http://localhost:11434/v1/chat/completions \
 - [ ] `user`
 - [ ] `n`
 
-#### Notes
-
-- `finish_reason` will always be `stop`
-- `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached
-
 ## Models
 
 Before using a model, pull it locally `ollama pull`:

docs/troubleshooting.md

@@ -22,7 +22,7 @@ docker logs <container-name>
 If manually running `ollama serve` in a terminal, the logs will be on that terminal.
 
 When you run Ollama on **Windows**, there are a few different locations. You can view them in the explorer window by hitting `<cmd>+R` and type in:
-- `explorer %LOCALAPPDATA%\Ollama` to view logs
+- `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
 - `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
@@ -70,14 +70,18 @@ curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 
 If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
 
-## Container fails to run on NVIDIA GPU
+## NVIDIA GPU Discovery
 
-Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+When Ollama starts up, it takes inventory of the GPUs present in the system to determine compatibility and how much VRAM is available.  Sometimes this discovery can fail to find your GPUs.  In general, running the latest driver will yield the best results.
 
-Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+### Linux NVIDIA Troubleshooting
 
-- Is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
-- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
+If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+
+Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+
+- If you are using a container, is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
+- Is the uvm driver loaded? `sudo nvidia-modprobe -u`
 - Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
 - Try rebooting
 - Make sure you're running the latest nvidia drivers
@@ -85,3 +89,8 @@ Sometimes the container runtime can have difficulties initializing the GPU. When
 If none of those resolve the problem, gather additional information and file an issue:
 - Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`
+
+
+## Windows Terminal Errors
+
+Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.

docs/windows.md

@@ -19,7 +19,7 @@ Logs will often be helpful in diagnosing the problem (see
 
 ## System Requirements
 
-* Windows 10 or newer, Home or Pro
+* Windows 10 22H2 or newer, Home or Pro
 * NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
 * AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
 
@@ -39,8 +39,8 @@ server.
 Ollama on Windows stores files in a few different locations.  You can view them in
 the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
-    - *app.log* contains logs from the GUI application
-    - *server.log* contains the server logs
+    - *app.log* contains most resent logs from the GUI application
+    - *server.log* contains the most recent server logs
     - *upgrade.log* contains log output for upgrades
 - `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` contains models and configuration

envconfig/config.go

@@ -4,12 +4,14 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"math"
 	"net"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
+	"time"
 )
 
 type OllamaHost struct {
@@ -34,7 +36,7 @@ var (
 	// Set via OLLAMA_HOST in the environment
 	Host *OllamaHost
 	// Set via OLLAMA_KEEP_ALIVE in the environment
-	KeepAlive string
+	KeepAlive time.Duration
 	// Set via OLLAMA_LLM_LIBRARY in the environment
 	LLMLibrary string
 	// Set via OLLAMA_MAX_LOADED_MODELS in the environment
@@ -43,8 +45,6 @@ var (
 	MaxQueuedRequests int
 	// Set via OLLAMA_MODELS in the environment
 	ModelsDir string
-	// Set via OLLAMA_MAX_VRAM in the environment
-	MaxVRAM uint64
 	// Set via OLLAMA_NOHISTORY in the environment
 	NoHistory bool
 	// Set via OLLAMA_NOPRUNE in the environment
@@ -57,6 +57,8 @@ var (
 	SchedSpread bool
 	// Set via OLLAMA_TMPDIR in the environment
 	TmpDir string
+	// Set via OLLAMA_INTEL_GPU in the environment
+	IntelGpu bool
 
 	// Set via CUDA_VISIBLE_DEVICES in the environment
 	CudaVisibleDevices string
@@ -83,13 +85,12 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_HOST":              {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"},
 		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"},
 		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"},
-		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models (default 1)"},
+		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU"},
 		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
-		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"},
 		"OLLAMA_MODELS":            {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"},
 		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"},
 		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"},
-		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default 1)"},
+		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"},
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"},
 		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"},
@@ -101,6 +102,7 @@ func AsMap() map[string]EnvVar {
 		ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices, "Set which AMD devices are visible"}
 		ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal, "Set which AMD devices are visible"}
 		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion, "Override the gfx used for all detected AMD GPUs"}
+		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGpu, "Enable experimental Intel GPU detection"}
 	}
 	return ret
 }
@@ -126,9 +128,10 @@ func clean(key string) string {
 
 func init() {
 	// default values
-	NumParallel = 1
-	MaxRunners = 1
+	NumParallel = 0 // Autoselect
+	MaxRunners = 0  // Autoselect
 	MaxQueuedRequests = 512
+	KeepAlive = 5 * time.Minute
 
 	LoadConfig()
 }
@@ -188,22 +191,12 @@ func LoadConfig() {
 
 	TmpDir = clean("OLLAMA_TMPDIR")
 
-	userLimit := clean("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseUint(userLimit, 10, 64)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_VRAM", userLimit, "error", err)
-		} else {
-			MaxVRAM = avail
-		}
-	}
-
 	LLMLibrary = clean("OLLAMA_LLM_LIBRARY")
 
 	if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {
 		val, err := strconv.Atoi(onp)
-		if err != nil || val <= 0 {
-			slog.Error("invalid setting must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err)
+		if err != nil {
+			slog.Error("invalid setting, ignoring", "OLLAMA_NUM_PARALLEL", onp, "error", err)
 		} else {
 			NumParallel = val
 		}
@@ -248,7 +241,7 @@ func LoadConfig() {
 	if maxRunners != "" {
 		m, err := strconv.Atoi(maxRunners)
 		if err != nil {
-			slog.Error("invalid setting", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
+			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
 		} else {
 			MaxRunners = m
 		}
@@ -257,13 +250,16 @@ func LoadConfig() {
 	if onp := os.Getenv("OLLAMA_MAX_QUEUE"); onp != "" {
 		p, err := strconv.Atoi(onp)
 		if err != nil || p <= 0 {
-			slog.Error("invalid setting", "OLLAMA_MAX_QUEUE", onp, "error", err)
+			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_QUEUE", onp, "error", err)
 		} else {
 			MaxQueuedRequests = p
 		}
 	}
 
-	KeepAlive = clean("OLLAMA_KEEP_ALIVE")
+	ka := clean("OLLAMA_KEEP_ALIVE")
+	if ka != "" {
+		loadKeepAlive(ka)
+	}
 
 	var err error
 	ModelsDir, err = getModelsDir()
@@ -276,6 +272,10 @@ func LoadConfig() {
 		slog.Error("invalid setting", "OLLAMA_HOST", Host, "error", err, "using default port", Host.Port)
 	}
 
+	if set, err := strconv.ParseBool(clean("OLLAMA_INTEL_GPU")); err == nil {
+		IntelGpu = set
+	}
+
 	CudaVisibleDevices = clean("CUDA_VISIBLE_DEVICES")
 	HipVisibleDevices = clean("HIP_VISIBLE_DEVICES")
 	RocrVisibleDevices = clean("ROCR_VISIBLE_DEVICES")
@@ -337,3 +337,24 @@ func getOllamaHost() (*OllamaHost, error) {
 		Port:   port,
 	}, nil
 }
+
+func loadKeepAlive(ka string) {
+	v, err := strconv.Atoi(ka)
+	if err != nil {
+		d, err := time.ParseDuration(ka)
+		if err == nil {
+			if d < 0 {
+				KeepAlive = time.Duration(math.MaxInt64)
+			} else {
+				KeepAlive = d
+			}
+		}
+	} else {
+		d := time.Duration(v) * time.Second
+		if d < 0 {
+			KeepAlive = time.Duration(math.MaxInt64)
+		} else {
+			KeepAlive = d
+		}
+	}
+}

envconfig/config_test.go

@@ -2,8 +2,10 @@ package envconfig
 
 import (
 	"fmt"
+	"math"
 	"net"
 	"testing"
+	"time"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -23,6 +25,21 @@ func TestConfig(t *testing.T) {
 	t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
 	LoadConfig()
 	require.True(t, FlashAttention)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "")
+	LoadConfig()
+	require.Equal(t, 5*time.Minute, KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "3")
+	LoadConfig()
+	require.Equal(t, 3*time.Second, KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "1h")
+	LoadConfig()
+	require.Equal(t, 1*time.Hour, KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "-1s")
+	LoadConfig()
+	require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "-1")
+	LoadConfig()
+	require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
 }
 
 func TestClientFromEnvironment(t *testing.T) {

go.mod

@@ -18,6 +18,7 @@ require (
 require (
 	github.com/agnivade/levenshtein v1.1.1
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
+	github.com/google/go-cmp v0.6.0
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
@@ -71,7 +72,7 @@ require (
 	golang.org/x/net v0.25.0 // indirect
 	golang.org/x/sys v0.20.0
 	golang.org/x/term v0.20.0
-	golang.org/x/text v0.15.0 // indirect
+	golang.org/x/text v0.15.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

gpu/amd_common.go

@@ -49,9 +49,17 @@ func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 }
 
 func commonAMDValidateLibDir() (string, error) {
-	// We try to favor system paths first, so that we can wire up the subprocess to use
-	// the system version.  Only use our bundled version if the system version doesn't work
-	// This gives users a more recovery options if versions have subtle problems at runtime
+	// Favor our bundled version
+
+	// Installer payload location if we're running the installed binary
+	exe, err := os.Executable()
+	if err == nil {
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
+	}
 
 	// Prefer explicit HIP env var
 	hipPath := os.Getenv("HIP_PATH")
@@ -87,14 +95,5 @@ 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")
-		if rocmLibUsable(rocmTargetDir) {
-			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-			return rocmTargetDir, nil
-		}
-	}
 	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
 }

gpu/amd_hip_windows.go

@@ -33,9 +33,10 @@ type HipLib struct {
 }
 
 func NewHipLib() (*HipLib, error) {
-	h, err := windows.LoadLibrary("amdhip64.dll")
+	// At runtime we depend on v6, so discover GPUs with the same library for a consistent set of GPUs
+	h, err := windows.LoadLibrary("amdhip64_6.dll")
 	if err != nil {
-		return nil, fmt.Errorf("unable to load amdhip64.dll: %w", err)
+		return nil, fmt.Errorf("unable to load amdhip64_6.dll, please make sure to upgrade to the latest amd driver: %w", err)
 	}
 	hl := &HipLib{}
 	hl.dll = h
@@ -84,9 +85,8 @@ func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	}
 
 	slog.Debug("hipDriverGetVersion", "version", version)
-	// TODO - this isn't actually right, but the docs claim hipDriverGetVersion isn't accurate anyway...
-	driverMajor = version / 1000
-	driverMinor = (version - (driverMajor * 1000)) / 10
+	driverMajor = version / 10000000
+	driverMinor = (version - (driverMajor * 10000000)) / 100000
 
 	return driverMajor, driverMinor, nil
 }

gpu/amd_windows.go

@@ -22,8 +22,8 @@ const (
 
 var (
 	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // TODO - probably include more coverage of files here...
-	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\5.7\\bin"} // TODO glob?
+	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // This is not sufficient to discern v5 vs v6
+	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\6.1\\bin"} // TODO glob?
 )
 
 func AMDGetGPUInfo() []RocmGPUInfo {
@@ -35,12 +35,11 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	}
 	defer hl.Release()
 
-	// TODO - this reports incorrect version information, so omitting for now
-	// driverMajor, driverMinor, err := hl.AMDDriverVersion()
-	// if err != nil {
-	// 	// For now this is benign, but we may eventually need to fail compatibility checks
-	// 	slog.Debug("error looking up amd driver version", "error", err)
-	// }
+	driverMajor, driverMinor, err := hl.AMDDriverVersion()
+	if err != nil {
+		// For now this is benign, but we may eventually need to fail compatibility checks
+		slog.Debug("error looking up amd driver version", "error", err)
+	}
 
 	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
@@ -93,7 +92,8 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			continue
 		}
 		if gfxOverride == "" {
-			if !slices.Contains[[]string, string](supported, gfx) {
+			// Strip off Target Features when comparing
+			if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
 				slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
 				// TODO - consider discrete markdown just for ROCM troubleshooting?
 				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")
@@ -115,8 +115,6 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			continue
 		}
 
-		// TODO revisit this once ROCm v6 is available on windows.
-		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
 		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
 		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
 		gpuInfo := RocmGPUInfo{
@@ -126,15 +124,16 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 					TotalMemory: totalMemory,
 					FreeMemory:  freeMemory,
 				},
+				// Free memory reporting on Windows is not reliable until we bump to ROCm v6.2
+				UnreliableFreeMemory: true,
+
 				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
 				DependencyPath: libDir,
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
 				Compute:        gfx,
-
-				// TODO - this information isn't accurate on windows, so don't report it until we find the right way to retrieve
-				// DriverMajor:    driverMajor,
-				// DriverMinor:    driverMinor,
+				DriverMajor:    driverMajor,
+				DriverMinor:    driverMinor,
 			},
 			index: i,
 		}

gpu/assets.go

@@ -77,20 +77,27 @@ func cleanupTmpDirs() {
 			continue
 		}
 		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
-		if err == nil {
-			pid, err := strconv.Atoi(string(raw))
-			if err == nil {
-				if proc, err := os.FindProcess(pid); err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
-					// Another running ollama, ignore this tmpdir
-					continue
-				}
-			}
-		} else {
-			slog.Debug("failed to open ollama.pid", "path", d, "error", err)
-		}
-		err = os.RemoveAll(d)
 		if err != nil {
-			slog.Debug("unable to cleanup stale tmpdir", "path", d, "error", err)
+			slog.Warn("failed to read ollama.pid", "path", d, "error", err)
+			// No pid, ignore this tmpdir
+			continue
+		}
+
+		pid, err := strconv.Atoi(string(raw))
+		if err != nil {
+			slog.Warn("failed to parse pid", "path", d, "error", err)
+			continue
+		}
+
+		proc, err := os.FindProcess(pid)
+		if err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+			slog.Warn("found running ollama", "pid", pid, "path", d)
+			// Another running ollama, ignore this tmpdir
+			continue
+		}
+
+		if err := os.Remove(d); err != nil {
+			slog.Warn("unable to cleanup stale tmpdir", "path", d, "error", err)
 		}
 	}
 }

gpu/gpu.go

@@ -202,7 +202,7 @@ func GetGPUInfo() GpuInfoList {
 	}()
 
 	if !bootstrapped {
-		slog.Debug("Detecting GPUs")
+		slog.Info("looking for compatible GPUs")
 		needRefresh = false
 		cpuCapability = GetCPUCapability()
 		var memInfo C.mem_info_t
@@ -231,7 +231,7 @@ func GetGPUInfo() GpuInfoList {
 		// On windows we bundle the nvidia library one level above the runner dir
 		depPath := ""
 		if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
-			depPath = filepath.Dir(envconfig.RunnersDir)
+			depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir), "cuda")
 		}
 
 		// Load ALL libraries
@@ -274,44 +274,77 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
 
+				// 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
+				if cHandles.nvml != nil {
+					C.nvml_get_free(*cHandles.nvml, C.int(gpuInfo.index), &memInfo.free, &memInfo.total, &memInfo.used)
+					if memInfo.err != nil {
+						slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+						C.free(unsafe.Pointer(memInfo.err))
+					} else {
+						if memInfo.free != 0 && uint64(memInfo.free) > gpuInfo.FreeMemory {
+							gpuInfo.OSOverhead = uint64(memInfo.free) - gpuInfo.FreeMemory
+							slog.Info("detected OS VRAM overhead",
+								"id", gpuInfo.ID,
+								"library", gpuInfo.Library,
+								"compute", gpuInfo.Compute,
+								"driver", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor),
+								"name", gpuInfo.Name,
+								"overhead", format.HumanBytes2(gpuInfo.OSOverhead),
+							)
+						}
+					}
+				}
+
 				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
 				cudaGPUs = append(cudaGPUs, gpuInfo)
 			}
 		}
 
 		// Intel
-		oHandles = initOneAPIHandles()
-		for d := 0; oHandles.oneapi != nil && d < int(oHandles.oneapi.num_drivers); d++ {
-			if oHandles.oneapi == nil {
-				// shouldn't happen
-				slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
-				continue
+		if envconfig.IntelGpu {
+			oHandles = initOneAPIHandles()
+			// 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")
 			}
-			devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
-			for i := range devCount {
-				gpuInfo := OneapiGPUInfo{
-					GpuInfo: GpuInfo{
-						Library: "oneapi",
-					},
-					driverIndex: d,
-					gpuIndex:    int(i),
+
+			for d := range oHandles.oneapi.num_drivers {
+				if oHandles.oneapi == nil {
+					// shouldn't happen
+					slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
+					continue
+				}
+				devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
+				for i := range devCount {
+					gpuInfo := OneapiGPUInfo{
+						GpuInfo: GpuInfo{
+							Library: "oneapi",
+						},
+						driverIndex: int(d),
+						gpuIndex:    int(i),
+					}
+					// TODO - split bootstrapping from updating free memory
+					C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
+					// TODO - convert this to MinimumMemory based on testing...
+					var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
+					memInfo.free = C.uint64_t(totalFreeMem)
+					gpuInfo.TotalMemory = uint64(memInfo.total)
+					gpuInfo.FreeMemory = uint64(memInfo.free)
+					gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+					gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+					gpuInfo.DependencyPath = depPath
+					oneapiGPUs = append(oneapiGPUs, gpuInfo)
 				}
-				// TODO - split bootstrapping from updating free memory
-				C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
-				// TODO - convert this to MinimumMemory based on testing...
-				var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
-				memInfo.free = C.uint64_t(totalFreeMem)
-				gpuInfo.TotalMemory = uint64(memInfo.total)
-				gpuInfo.FreeMemory = uint64(memInfo.free)
-				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-				// TODO dependency path?
-				oneapiGPUs = append(oneapiGPUs, gpuInfo)
 			}
 		}
 
 		rocmGPUs = AMDGetGPUInfo()
 		bootstrapped = true
+		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
+			slog.Info("no compatible GPUs were discovered")
+		}
 	}
 
 	// For detected GPUs, load library if not loaded
@@ -327,14 +360,17 @@ func GetGPUInfo() GpuInfoList {
 					"before",
 					"total", format.HumanBytes2(cpus[0].TotalMemory),
 					"free", format.HumanBytes2(cpus[0].FreeMemory),
+					"free_swap", format.HumanBytes2(cpus[0].FreeSwap),
 				),
 				slog.Group(
 					"now",
 					"total", format.HumanBytes2(mem.TotalMemory),
 					"free", format.HumanBytes2(mem.FreeMemory),
+					"free_swap", format.HumanBytes2(mem.FreeSwap),
 				),
 			)
 			cpus[0].FreeMemory = mem.FreeMemory
+			cpus[0].FreeSwap = mem.FreeSwap
 		}
 
 		var memInfo C.mem_info_t
@@ -363,9 +399,14 @@ func GetGPUInfo() GpuInfoList {
 				slog.Warn("error looking up nvidia GPU memory")
 				continue
 			}
+			if cHandles.nvml != nil && gpu.OSOverhead > 0 {
+				// When using the management library update based on recorded overhead
+				memInfo.free -= C.uint64_t(gpu.OSOverhead)
+			}
 			slog.Debug("updating cuda memory data",
 				"gpu", gpu.ID,
 				"name", gpu.Name,
+				"overhead", format.HumanBytes2(gpu.OSOverhead),
 				slog.Group(
 					"before",
 					"total", format.HumanBytes2(gpu.TotalMemory),
@@ -506,7 +547,23 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 		defer C.free(unsafe.Pointer(lib))
 		C.nvcuda_init(lib, &resp)
 		if resp.err != nil {
-			slog.Debug("Unable to load nvcuda", "library", libPath, "error", C.GoString(resp.err))
+			// Decide what log level based on the type of error message to help users understand why
+			msg := C.GoString(resp.err)
+			switch resp.cudaErr {
+			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
+				slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
+			case C.CUDA_ERROR_NO_DEVICE:
+				slog.Info("no nvidia devices detected", "library", libPath)
+			case C.CUDA_ERROR_UNKNOWN:
+				slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
+				slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
+			default:
+				if strings.Contains(msg, "wrong ELF class") {
+					slog.Debug("skipping 32bit library", "library", libPath)
+				} else {
+					slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
+				}
+			}
 			C.free(unsafe.Pointer(resp.err))
 		} else {
 			return int(resp.num_devices), &resp.ch, libPath

gpu/gpu_darwin.go

@@ -56,7 +56,8 @@ func GetCPUInfo() GpuInfoList {
 func GetCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
-		FreeMemory:  0,
+		FreeMemory:  uint64(C.getFreeMemory()),
+		// FreeSwap omitted as Darwin uses dynamic paging
 	}, nil
 }
 

gpu/gpu_info_cudart.c

@@ -40,7 +40,7 @@ void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
 
   for (i = 0; l[i].s != NULL; i++) {
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!l[i].p) {
+    if (!*(l[i].p)) {
       char *msg = LOAD_ERR();
       LOG(resp->ch.verbose, "dlerr: %s\n", msg);
       UNLOAD_LIBRARY(resp->ch.handle);

gpu/gpu_info_darwin.h

@@ -2,3 +2,4 @@
 #include <stdint.h>
 uint64_t getRecommendedMaxVRAM();
 uint64_t getPhysicalMemory();
+uint64_t getFreeMemory();

gpu/gpu_info_darwin.m

@@ -1,4 +1,5 @@
-// go:build darwin
+#import <Foundation/Foundation.h>
+#import <mach/mach.h>
 #include "gpu_info_darwin.h"
 
 uint64_t getRecommendedMaxVRAM() {
@@ -8,6 +9,27 @@ uint64_t getRecommendedMaxVRAM() {
   return result;
 }
 
+// getPhysicalMemory returns the total physical memory in bytes
 uint64_t getPhysicalMemory() {
-  return [[NSProcessInfo processInfo] physicalMemory];
+  return [NSProcessInfo processInfo].physicalMemory;
+}
+
+// getFreeMemory returns the total free memory in bytes, including inactive
+// memory that can be reclaimed by the system.
+uint64_t getFreeMemory() {
+  mach_port_t host_port = mach_host_self();
+  mach_msg_type_number_t host_size = sizeof(vm_statistics64_data_t) / sizeof(integer_t);
+  vm_size_t pagesize;
+  vm_statistics64_data_t vm_stat;
+
+  host_page_size(host_port, &pagesize);
+  if (host_statistics64(host_port, HOST_VM_INFO64, (host_info64_t)&vm_stat, &host_size) != KERN_SUCCESS) {
+    return 0;
+  }
+
+  uint64_t free_memory = (uint64_t)vm_stat.free_count * pagesize;
+  free_memory += (uint64_t)vm_stat.speculative_count * pagesize;
+  free_memory += (uint64_t)vm_stat.inactive_count * pagesize;
+
+  return free_memory;
 }

gpu/gpu_info_nvcuda.c

@@ -7,6 +7,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
   CUresult ret;
   resp->err = NULL;
   resp->num_devices = 0;
+  resp->cudaErr = CUDA_SUCCESS;
   const int buflen = 256;
   char buf[buflen + 1];
   int i;
@@ -38,12 +39,13 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
             nvcuda_lib_path, msg);
     free(msg);
     resp->err = strdup(buf);
+    resp->cudaErr = -1;
     return;
   }
 
   for (i = 0; l[i].s != NULL; i++) {
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!*l[i].p) {
+    if (!*(l[i].p)) {
       char *msg = LOAD_ERR();
       LOG(resp->ch.verbose, "dlerr: %s\n", msg);
       UNLOAD_LIBRARY(resp->ch.handle);
@@ -52,6 +54,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
               msg);
       free(msg);
       resp->err = strdup(buf);
+      resp->cudaErr = -1;
       return;
     }
   }
@@ -61,12 +64,9 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
-    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
-      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
-      return;
-    }
-    snprintf(buf, buflen, "nvcuda init failure: %d", ret);
+    snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
     resp->err = strdup(buf);
+    resp->cudaErr = ret;
     return;
   }
 
@@ -91,6 +91,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     resp->ch.handle = NULL;
     snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
+    resp->cudaErr = ret;
     return;
   }
 }
@@ -106,13 +107,13 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 
   if (h.handle == NULL) {
-    resp->err = strdup("nvcuda handle isn't initialized");
+    resp->err = strdup("cuda driver library handle isn't initialized");
     return;
   }
 
   ret = (*h.cuDeviceGet)(&device, i);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device failed to initialize");
+    snprintf(buf, buflen, "cuda driver library device failed to initialize");
     resp->err = strdup(buf);
     return;
   }
@@ -168,14 +169,14 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   // To get memory we have to set (and release) a context
   ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda failed to get device context %d", ret);
+    snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
     resp->err = strdup(buf);
     return;
   }
 
   ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device memory info lookup failure %d", ret);
+    snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
     resp->err = strdup(buf);
     // Best effort on failure...
     (*h.cuCtxDestroy)(ctx);
@@ -193,7 +194,7 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
 
   ret = (*h.cuCtxDestroy)(ctx);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to release device context %d", ret);
+    LOG(1, "cuda driver library failed to release device context %d", ret);
   }
 }
 
@@ -206,7 +207,7 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
 
   ret = (*h.cuDeviceGet)(&device, i);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda device failed to initialize");
+    LOG(1, "cuda driver library device failed to initialize");
     return;
   }
 
@@ -214,13 +215,13 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
   // To get memory we have to set (and release) a context
   ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to get device context %d", ret);
+    LOG(1, "cuda driver library failed to get device context %d", ret);
     return;
   }
 
   ret = (*h.cuMemGetInfo_v2)(free, total);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda device memory info lookup failure %d", ret);
+    LOG(1, "cuda driver library device memory info lookup failure %d", ret);
     // Best effort on failure...
     (*h.cuCtxDestroy)(ctx);
     return;
@@ -228,12 +229,12 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
 
   ret = (*h.cuCtxDestroy)(ctx);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to release device context %d", ret);
+    LOG(1, "cuda driver library failed to release device context %d", ret);
   }
 }
 
 void nvcuda_release(nvcuda_handle_t h) {
-  LOG(h.verbose, "releasing nvcuda library\n");
+  LOG(h.verbose, "releasing cuda driver library\n");
   UNLOAD_LIBRARY(h.handle);
   // TODO and other context release logic?
   h.handle = NULL;

gpu/gpu_info_nvcuda.h

@@ -7,9 +7,12 @@
 typedef enum cudaError_enum {
   CUDA_SUCCESS = 0,
   CUDA_ERROR_INVALID_VALUE = 1,
-  CUDA_ERROR_MEMORY_ALLOCATION = 2,
+  CUDA_ERROR_OUT_OF_MEMORY = 2,
   CUDA_ERROR_NOT_INITIALIZED = 3,
   CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  CUDA_ERROR_NO_DEVICE = 100,
+  CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
+  CUDA_ERROR_UNKNOWN = 999,
   // Other values omitted for now...
 } CUresult;
 
@@ -64,6 +67,7 @@ typedef struct nvcuda_init_resp {
   char *err;  // If err is non-null handle is invalid
   nvcuda_handle_t ch;
   int num_devices;
+  CUresult cudaErr;
 } nvcuda_init_resp_t;
 
 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);

gpu/gpu_info_nvml.c

@@ -42,7 +42,7 @@ void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
     // LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
 
     *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!l[i].p) {
+    if (!*(l[i].p)) {
       resp->ch.handle = NULL;
       char *msg = LOAD_ERR();
       LOG(resp->ch.verbose, "dlerr: %s\n", msg);

gpu/gpu_info_oneapi.c

@@ -13,7 +13,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
   resp->oh.num_drivers = 0;
   const int buflen = 256;
   char buf[buflen + 1];
-  int i, d, count;
+  int i, d;
   struct lookup {
     char *s;
     void **p;
@@ -50,7 +50,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
     LOG(resp->oh.verbose, "dlsym: %s\n", l[i].s);
 
     *l[i].p = LOAD_SYMBOL(resp->oh.handle, l[i].s);
-    if (!l[i].p) {
+    if (!*(l[i].p)) {
       resp->oh.handle = NULL;
       char *msg = LOAD_ERR();
       LOG(resp->oh.verbose, "dlerr: %s\n", msg);
@@ -62,6 +62,8 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
     }
   }
 
+  LOG(resp->oh.verbose, "calling zesInit\n");
+
   ret = (*resp->oh.zesInit)(0);
   if (ret != ZE_RESULT_SUCCESS) {
     LOG(resp->oh.verbose, "zesInit err: %x\n", ret);
@@ -71,7 +73,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
     return;
   }
 
-  count = 0;
+  LOG(resp->oh.verbose, "calling zesDriverGet\n");
   ret = (*resp->oh.zesDriverGet)(&resp->oh.num_drivers, NULL);
   if (ret != ZE_RESULT_SUCCESS) {
     LOG(resp->oh.verbose, "zesDriverGet err: %x\n", ret);
@@ -96,6 +98,7 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
   }
 
   for (d = 0; d < resp->oh.num_drivers; d++) {
+    LOG(resp->oh.verbose, "calling zesDeviceGet count %d: %p\n", d, resp->oh.drivers[d]);
     ret = (*resp->oh.zesDeviceGet)(resp->oh.drivers[d],
                                    &resp->oh.num_devices[d], NULL);
     if (ret != ZE_RESULT_SUCCESS) {
@@ -116,7 +119,6 @@ void oneapi_init(char *oneapi_lib_path, oneapi_init_resp_t *resp) {
       oneapi_release(resp->oh);
       return;
     }
-    count += resp->oh.num_devices[d];
   }
 
   return;
@@ -160,7 +162,7 @@ void oneapi_check_vram(oneapi_handle_t h, int driver, int device,
     return;
   }
 
-  snprintf(&resp->gpu_name[0], GPU_NAME_LEN, props.modelName);
+  snprintf(&resp->gpu_name[0], GPU_NAME_LEN, "%s", props.modelName);
 
   // TODO this needs to map to ONEAPI_DEVICE_SELECTOR syntax
   // (this is probably wrong...)

gpu/gpu_linux.go

@@ -50,7 +50,7 @@ var OneapiMgmtName = "libze_intel_gpu.so"
 
 func GetCPUMem() (memInfo, error) {
 	var mem memInfo
-	var total, available, free, buffers, cached uint64
+	var total, available, free, buffers, cached, freeSwap uint64
 	f, err := os.Open("/proc/meminfo")
 	if err != nil {
 		return mem, err
@@ -70,20 +70,21 @@ func GetCPUMem() (memInfo, error) {
 			_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
 		case strings.HasPrefix(line, "Cached:"):
 			_, err = fmt.Sscanf(line, "Cached:%d", &cached)
+		case strings.HasPrefix(line, "SwapFree:"):
+			_, err = fmt.Sscanf(line, "SwapFree:%d", &freeSwap)
 		default:
 			continue
 		}
 		if err != nil {
 			return mem, err
 		}
-
-		if total > 0 && available > 0 {
-			mem.TotalMemory = total * format.KibiByte
-			mem.FreeMemory = available * format.KibiByte
-			return mem, nil
-		}
 	}
 	mem.TotalMemory = total * format.KibiByte
-	mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	mem.FreeSwap = freeSwap * format.KibiByte
+	if available > 0 {
+		mem.FreeMemory = available * format.KibiByte
+	} else {
+		mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	}
 	return mem, nil
 }

gpu/gpu_windows.go

@@ -51,5 +51,5 @@ func GetCPUMem() (memInfo, error) {
 	if r1 == 0 {
 		return memInfo{}, fmt.Errorf("GlobalMemoryStatusEx failed: %w", err)
 	}
-	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys}, nil
+	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys, FreeSwap: memStatus.AvailPageFile}, nil
 }

gpu/types.go

@@ -10,6 +10,7 @@ import (
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
+	FreeSwap    uint64 `json:"free_swap,omitempty"`
 }
 
 // Beginning of an `ollama info` command
@@ -29,6 +30,11 @@ type GpuInfo struct {
 	// Extra environment variables specific to the GPU as list of [key,value]
 	EnvWorkarounds [][2]string `json:"envs,omitempty"`
 
+	// Set to true if we can NOT reliably discover FreeMemory.  A value of true indicates
+	// the FreeMemory is best effort, and may over or under report actual memory usage
+	// False indicates FreeMemory can generally be trusted on this GPU
+	UnreliableFreeMemory bool
+
 	// GPU information
 	ID      string `json:"gpu_id"`  // string to use for selection of this specific GPU
 	Name    string `json:"name"`    // user friendly name if available
@@ -47,7 +53,8 @@ type CPUInfo struct {
 
 type CudaGPUInfo struct {
 	GpuInfo
-	index int //nolint:unused,nolintlint
+	OSOverhead uint64 // Memory overhead between the driver library and management library
+	index      int    //nolint:unused,nolintlint
 }
 type CudaGPUInfoList []CudaGPUInfo
 

integration/concurrency_test.go

@@ -69,7 +69,7 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 	reqLimit := len(req)
 	iterLimit := 5
 
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
+	vram := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
 	if vram != "" {
 		max, err := strconv.ParseUint(vram, 10, 64)
 		require.NoError(t, err)
@@ -106,7 +106,7 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 
 // Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
 func TestMultiModelStress(t *testing.T) {
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
+	vram := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
 	if vram == "" {
 		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
 	}

integration/context_test.go

@@ -12,7 +12,7 @@ import (
 
 func TestContextExhaustion(t *testing.T) {
 	// Longer needed for small footprint GPUs
-	ctx, cancel := context.WithTimeout(context.Background(), 6*time.Minute)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
@@ -25,5 +25,10 @@ func TestContextExhaustion(t *testing.T) {
 			"num_ctx":     128,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"once", "upon", "lived"})
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("PullIfMissing failed: %v", err)
+	}
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
 }

integration/embed_test.go

@@ -0,0 +1,152 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestAllMiniLMEmbed(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbedRequest{
+		Model: "all-minilm",
+		Input: "why is the sky blue?",
+	}
+
+	res, err := embedTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embeddings) != 1 {
+		t.Fatalf("expected 1 embedding, got %d", len(res.Embeddings))
+	}
+
+	if len(res.Embeddings[0]) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
+	}
+
+	if res.Embeddings[0][0] != 0.010071031 {
+		t.Fatalf("expected 0.010071031, got %f", res.Embeddings[0][0])
+	}
+}
+
+func TestAllMiniLMBatchEmbed(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbedRequest{
+		Model: "all-minilm",
+		Input: []string{"why is the sky blue?", "why is the grass green?"},
+	}
+
+	res, err := embedTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embeddings) != 2 {
+		t.Fatalf("expected 2 embeddings, got %d", len(res.Embeddings))
+	}
+
+	if len(res.Embeddings[0]) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
+	}
+
+	if res.Embeddings[0][0] != 0.010071031 || res.Embeddings[1][0] != -0.009802706 {
+		t.Fatalf("expected 0.010071031 and -0.009802706, got %f and %f", res.Embeddings[0][0], res.Embeddings[1][0])
+	}
+}
+
+func TestAllMiniLmEmbedTruncate(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	truncTrue, truncFalse := true, false
+
+	type testReq struct {
+		Name    string
+		Request api.EmbedRequest
+	}
+
+	reqs := []testReq{
+		{
+			Name: "Target Truncation",
+			Request: api.EmbedRequest{
+				Model: "all-minilm",
+				Input: "why",
+			},
+		},
+		{
+			Name: "Default Truncate",
+			Request: api.EmbedRequest{
+				Model:   "all-minilm",
+				Input:   "why is the sky blue?",
+				Options: map[string]any{"num_ctx": 1},
+			},
+		},
+		{
+			Name: "Explicit Truncate",
+			Request: api.EmbedRequest{
+				Model:    "all-minilm",
+				Input:    "why is the sky blue?",
+				Truncate: &truncTrue,
+				Options:  map[string]any{"num_ctx": 1},
+			},
+		},
+	}
+
+	res := make(map[string]*api.EmbedResponse)
+
+	for _, req := range reqs {
+		response, err := embedTestHelper(ctx, t, req.Request)
+		if err != nil {
+			t.Fatalf("error: %v", err)
+		}
+		res[req.Name] = response
+	}
+
+	if res["Target Truncation"].Embeddings[0][0] != res["Default Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request to truncate correctly")
+	}
+
+	if res["Default Truncate"].Embeddings[0][0] != res["Explicit Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request and truncate true request to be the same")
+	}
+
+	// check that truncate set to false returns an error if context length is exceeded
+	_, err := embedTestHelper(ctx, t, api.EmbedRequest{
+		Model:    "all-minilm",
+		Input:    "why is the sky blue?",
+		Truncate: &truncFalse,
+		Options:  map[string]any{"num_ctx": 1},
+	})
+
+	if err == nil {
+		t.Fatal("expected error, got nil")
+	}
+}
+
+func embedTestHelper(ctx context.Context, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+	}
+
+	response, err := client.Embed(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}

llm/ext_server/CMakeLists.txt

@@ -1,14 +1,13 @@
-
-set(TARGET ollama_llama_server)
-option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-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 common llava ${CMAKE_THREAD_LIBS_INIT})
-if (WIN32)
-    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
-endif()
+set(TARGET ollama_llama_server)
+option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+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 ${CMAKE_THREAD_LIBS_INIT})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

llm/ext_server/server.cpp

@@ -56,7 +56,6 @@ struct server_params {
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
-    std::string chat_template = "";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
@@ -427,16 +426,6 @@ struct llama_server_context
         return true;
     }
 
-    void validate_model_chat_template(server_params & sparams) {
-        llama_chat_message chat[] = {{"user", "test"}};
-        std::vector<char> buf(1);
-        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
-        if (res < 0) {
-            LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
-            sparams.chat_template = "chatml";
-        }
-    }
-
     void initialize() {
         // create slots
         all_slots_are_idle = true;
@@ -1393,12 +1382,50 @@ struct llama_server_context
         }
     }
 
+    std::string common_prefix(const std::string& str1, const std::string& str2) {
+        auto mismatch_pair = std::mismatch(str1.begin(), str1.end(), str2.begin());
+        return std::string(str1.begin(), mismatch_pair.first);
+    }
+
+    // Find the slot that has the greatest common prefix
+    server_slot *prefix_slot(const json &prompt) {
+        if (!prompt.is_string()) {
+            return nullptr;
+        }
+
+        std::string prompt_str = prompt.get<std::string>();
+        server_slot *slot = nullptr;
+        size_t longest = 0;
+
+        for (server_slot &s : slots) {
+            if (s.available() && s.prompt.is_string()) {
+                std::string s_prompt = s.prompt.get<std::string>();
+                std::string prefix = common_prefix(s_prompt, prompt_str);
+
+                if (prefix.size() > longest) {
+                    slot = &s;
+                    longest = prefix.size();
+                }
+            }
+        }
+
+        if (!slot) {
+            return get_slot(-1);
+        }
+
+        LOG_DEBUG("slot with common prefix found", {{
+            "slot_id", slot->id,
+            "characters", longest
+        }});
+        return slot;
+    }
+
     void process_single_task(task_server& task)
     {
         switch (task.type)
         {
             case TASK_TYPE_COMPLETION: {
-                server_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
+                server_slot *slot = prefix_slot(task.data["prompt"]);
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later
@@ -1665,22 +1692,23 @@ struct llama_server_context
                     if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx)
                     {
                         const int n_left = slot.n_ctx - slot.params.n_keep;
-                        const int n_block_size = n_left / 2;
-                        const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
+                        const int n_shift = n_left / 2;
+                        const int n_erase = slot.n_prompt_tokens - slot.params.n_keep - n_shift;
 
                         std::vector<llama_token> new_tokens(
                             prompt_tokens.begin(),
                             prompt_tokens.begin() + slot.params.n_keep);
                         new_tokens.insert(
                             new_tokens.end(),
-                            prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size,
+                            prompt_tokens.begin() + slot.params.n_keep + n_erase,
                             prompt_tokens.end());
 
-                        LOG_VERBOSE("input truncated", {
-                            {"n_ctx",      slot.n_ctx},
-                            {"n_keep",     slot.params.n_keep},
-                            {"n_left",     n_left},
-                            {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                        LOG_INFO("input truncated", {
+                            {"n_ctx",        slot.n_ctx},
+                            {"n_keep",       slot.params.n_keep},
+                            {"n_left",       n_left},
+                            {"n_shift",      n_shift},
+                            {"n_erase",      n_erase},
                         });
                         slot.truncated = true;
                         prompt_tokens = new_tokens;
@@ -1715,7 +1743,7 @@ struct llama_server_context
                             slot.n_past -= 1;
                         }
 
-                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens - slot.n_past;
+                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens;
 
                         if (slot.ga_n != 1)
                         {
@@ -2535,7 +2563,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            sparams.chat_template = argv[i];
         }
         else if (arg == "--override-kv")
         {
@@ -3008,11 +3035,6 @@ int main(int argc, char **argv) {
     }
     const auto model_meta = llama.model_meta();
 
-    if (sparams.chat_template.empty()) { // custom chat template is not supplied
-        // check if the template comes with the model is supported by us
-        llama.validate_model_chat_template(sparams);
-    }
-
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation
@@ -3166,26 +3188,33 @@ int main(int argc, char **argv) {
                     prompt = "";
                 }
 
-                json image_data;
-                if (body.count("image_data") != 0) {
-                    image_data = body["image_data"];
-                }
-                else
-                {
-                    image_data = "";
+                if (prompt.size() == 1) {
+                    prompt = prompt[0];
                 }
 
                 // create and queue the task
-                const int task_id = llama.queue_tasks.get_new_id();
-                llama.queue_results.add_waiting_task_id(task_id);
-                llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, true, -1);
+                json responses;
+                {
+                    const int id_task = llama.queue_tasks.get_new_id();
+                    llama.queue_results.add_waiting_task_id(id_task);
+                    llama.request_completion(id_task, {{"prompt", prompt}}, true, -1);
 
-                // get the result
-                task_result result = llama.queue_results.recv(task_id);
-                llama.queue_results.remove_waiting_task_id(task_id);
+                    // get the result
+                    task_result result = llama.queue_results.recv(id_task);
+                    llama.queue_results.remove_waiting_task_id(id_task);
+                    if (result.error) {
+                        return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
+                    }
 
-                // send the result
-                return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
+                    responses = result.result_json.value("results", std::vector<json>{result.result_json});
+                    json embeddings = json::array();
+                    for (auto & elem : responses) {
+                        embeddings.push_back(elem.at("embedding"));
+                    }
+                    // send the result
+                    json embedding_res = json{{"embedding", embeddings}};
+                    return res.set_content(embedding_res.dump(), "application/json; charset=utf-8");
+                }
             });
 
     // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!?

llm/generate/gen_darwin.sh

@@ -18,16 +18,16 @@ sign() {
     fi
 }
 
-COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on"
+COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
 
 case "${GOARCH}" in
 "amd64")
-    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
+    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DGGML_METAL=off -DGGML_NATIVE=off"
 
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_BLAS=off -DGGML_ACCELERATE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
@@ -37,7 +37,7 @@ case "${GOARCH}" in
         # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/cpu"
         echo "Building LCD CPU"
         build
@@ -49,7 +49,7 @@ case "${GOARCH}" in
         # Approximately 400% faster than LCD on same CPU
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
         echo "Building AVX CPU"
         build
@@ -61,7 +61,7 @@ case "${GOARCH}" in
         # Approximately 10% faster than AVX on same CPU
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
         echo "Building AVX2 CPU"
         EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
@@ -75,14 +75,14 @@ case "${GOARCH}" in
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
 
     if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
         init_vars
-        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
         BUILD_DIR="../build/darwin/${ARCH}/metal"
         EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
         build

llm/generate/gen_linux.sh

@@ -51,7 +51,7 @@ if [ -z "${CUDACXX}" ]; then
         export CUDACXX=$(command -v nvcc)
     fi
 fi
-COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off"
+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"
 source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
@@ -64,7 +64,7 @@ if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ];
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DGGML_NATIVE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}_static"
     echo "Building static library"
     build
@@ -77,29 +77,29 @@ 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} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
         BUILD_DIR="../build/linux/${ARCH}/cpu"
         echo "Building custom CPU"
         build
         compress
     else
         # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
-        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
-        # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+        # -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
         # Note: the following seem to yield slower results than AVX2 - ymmv
-        # -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
-        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
-        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
+        # -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
+        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
 
-        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off"
+        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=off -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)
             #
             init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
             BUILD_DIR="../build/linux/${ARCH}/cpu"
             echo "Building LCD CPU"
             build
@@ -116,7 +116,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 400% faster than LCD on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
                 echo "Building AVX CPU"
                 build
@@ -129,7 +129,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 10% faster than AVX on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
                 BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
                 echo "Building AVX2 CPU"
                 build
@@ -170,15 +170,15 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
         #
         # CUDA compute < 6.0 lacks proper FP16 support on ARM.
         # Disabling has minimal performance effect while maintaining compatibility.
-        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
+        ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
     fi
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
         echo "Building custom CUDA GPU"
     else
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
     fi
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
@@ -216,7 +216,7 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
     init_vars
     source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
     CC=icx
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
+    CMAKE_DEFS="${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"
     DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@@ -254,7 +254,7 @@ 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} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DLLAMA_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""

llm/generate/gen_windows.ps1

@@ -1,21 +1,14 @@
 #!powershell
 
+$ErrorActionPreference = "Stop"
+
 function amdGPUs {
     if ($env:AMDGPU_TARGETS) {
         return $env:AMDGPU_TARGETS
     }
-    # TODO - load from some common data file for linux + windows build consistency
+    # Current supported rocblas list from ROCm v6.1.2 on windows
+    # https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html#windows-supported-gpus
     $GPU_LIST = @(
-        "gfx900"
-        "gfx906:xnack-"
-        "gfx908:xnack-"
-        "gfx90a:xnack+"
-        "gfx90a:xnack-"
-        "gfx940"
-        "gfx941"
-        "gfx942"
-        "gfx1010"
-        "gfx1012"
         "gfx1030"
         "gfx1100"
         "gfx1101"
@@ -37,7 +30,8 @@ function init_vars {
     }
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
-        "-DLLAMA_NATIVE=off"
+        "-DGGML_NATIVE=off",
+        "-DGGML_OPENMP=off"
         )
     $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
     $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
@@ -83,9 +77,9 @@ function init_vars {
 function git_module_setup {
     # TODO add flags to skip the init/patch logic to make it easier to mod llama.cpp code in-repo
     & git submodule init
-    if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     & git submodule update --force "${script:llamacppDir}"
-    if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
 }
 
 function apply_patches {
@@ -119,7 +113,7 @@ function build {
     write-host "generating config with: cmake -S ${script:llamacppDir} -B $script:buildDir $script:cmakeDefs"
     & cmake --version
     & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
-    if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     if ($cmakeDefs -contains "-G") {
         $extra=@("-j8")
     } else {
@@ -127,7 +121,7 @@ function build {
     }
     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", $_ }) $extra
-    if ($LASTEXITCODE -ne 0) { write-host "cmake build exit status $LASTEXITCODE"; throw($LASTEXITCODE)}
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     # Rearrange output to be consistent between different generators
     if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
         mv -force "${script:buildDir}/bin/${script:config}/*" "${script:buildDir}/bin/"
@@ -141,7 +135,7 @@ function sign {
         foreach ($file in @(get-childitem "${script:buildDir}/bin/*.exe") + @(get-childitem "${script:buildDir}/bin/*.dll")){
             & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
                 /csp "Google Cloud KMS Provider" /kc "${env:KEY_CONTAINER}" $file
-            if ($LASTEXITCODE -ne 0) { throw($LASTEXITCODE)}
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
         }
     }
 }
@@ -179,9 +173,9 @@ function cleanup {
 }
 
 
-# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
 
 function build_static() {
@@ -201,12 +195,13 @@ function build_static() {
             "-DCMAKE_C_COMPILER=gcc.exe",
             "-DCMAKE_CXX_COMPILER=g++.exe",
             "-DBUILD_SHARED_LIBS=off",
-            "-DLLAMA_NATIVE=off",
-            "-DLLAMA_AVX=off",
-            "-DLLAMA_AVX2=off",
-            "-DLLAMA_AVX512=off",
-            "-DLLAMA_F16C=off",
-            "-DLLAMA_FMA=off")
+            "-DGGML_NATIVE=off",
+            "-DGGML_AVX=off",
+            "-DGGML_AVX2=off",
+            "-DGGML_AVX512=off",
+            "-DGGML_F16C=off",
+            "-DGGML_FMA=off",
+            "-DGGML_OPENMP=off")
         $script:buildDir="../build/windows/${script:ARCH}_static"
         write-host "Building static library"
         build
@@ -216,17 +211,11 @@ function build_static() {
     }
 }
 
-function build_cpu() {
-    if ($script:ARCH -eq "arm64") {
-        $gen_arch = "ARM64"
-    } else { # amd64
-        $gen_arch = "x64"
-    }
-
+function build_cpu($gen_arch) {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
         # remaining llama.cpp builds use MSVC 
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu"
         $script:distDir="$script:DIST_BASE\cpu"
         write-host "Building LCD CPU"
@@ -241,7 +230,7 @@ function build_cpu() {
 function build_cpu_avx() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
         $script:distDir="$script:DIST_BASE\cpu_avx"
         write-host "Building AVX CPU"
@@ -256,7 +245,7 @@ function build_cpu_avx() {
 function build_cpu_avx2() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=on", "-DGGML_AVX512=off", "-DGGML_FMA=on", "-DGGML_F16C=on") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
         $script:distDir="$script:DIST_BASE\cpu_avx2"
         write-host "Building AVX2 CPU"
@@ -281,11 +270,11 @@ function build_cuda() {
         $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
         $script:cmakeDefs += @(
             "-A", "x64",
-            "-DLLAMA_CUDA=ON",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_CUDA=ON",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
             "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
-            "-DCMAKE_CUDA_FLAGS=-t8"
+            "-DCMAKE_CUDA_FLAGS=-t8",
             "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
             )
         if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
@@ -297,10 +286,12 @@ function build_cuda() {
         sign
         install
 
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+        rm -ea 0 -recurse -force -path "${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\"
     } else {
         write-host "Skipping CUDA generation step"
     }
@@ -319,7 +310,7 @@ function build_oneapi() {
     $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
     $script:cmakeDefs += @(
       "-G", "MinGW Makefiles",
-      "-DLLAMA_SYCL=ON",
+      "-DGGML_SYCL=ON",
       "-DCMAKE_C_COMPILER=icx",
       "-DCMAKE_CXX_COMPILER=icx",
       "-DCMAKE_BUILD_TYPE=Release"
@@ -334,16 +325,18 @@ function build_oneapi() {
     sign
     install
 
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:distDir}"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:distDir}"
+    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-${script:ARCH}\oneapi\"
   } else {
     Write-Host "Skipping oneAPI generation step"
   }
@@ -363,10 +356,11 @@ function build_rocm() {
             "-G", "Ninja", 
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DLLAMA_HIPBLAS=on",
+            "-DGGML_HIPBLAS=on",
+            "-DLLAMA_CUDA_NO_PEER_COPY=on",
             "-DHIP_PLATFORM=amd",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
             "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
             "-DAMDGPU_TARGETS=$(amdGPUs)",
             "-DGPU_TARGETS=$(amdGPUs)"
@@ -392,7 +386,6 @@ function build_rocm() {
         sign
         install
 
-        # Assumes v5.7, may need adjustments for v6
         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\"
@@ -408,29 +401,16 @@ init_vars
 if ($($args.count) -eq 0) {
     git_module_setup
     apply_patches
-
-    $tasks = @("build_static", "build_cpu")
-    $jobs = @()
-    if ($script:ARCH -ne "arm64") {
-        $tasks += $("build_cpu_avx", "build_cpu_avx2", "build_cuda", "build_oneapi", "build_rocm")
-    }
-    foreach ($t in $tasks) {
-        $jobs += @(Start-ThreadJob -ThrottleLimit 12 -FilePath .\gen_windows.ps1 -ArgumentList $t -Name $t)
-    }
-    get-job
-    foreach ($job in $jobs) {
-        write-host "----" $job.Name output follows
-        receive-job -wait -job $job
-        write-host "----" $job.Name $job.State
-        write-host ""
-        if ($job.State -contains 'Failed') {
-            cleanup
-            write-host "Terminating remaining jobs (this takes a while, you can ^C)"
-            # TODO find some way to kill the spawned cmake processes faster
-            remove-job -force -job $jobs
-            exit(-1)
-        }
-        get-job
+    build_static
+    if ($script:ARCH -eq "arm64") {
+        build_cpu("ARM64")
+    } else { # amd64
+        build_cpu("x64")
+        build_cpu_avx
+        build_cpu_avx2
+        build_cuda
+        build_oneapi
+        build_rocm
     }
 
     cleanup

llm/ggla.go

@@ -53,7 +53,7 @@ func (llm *ggla) Tensors() Tensors {
 	return llm.tensors
 }
 
-func (llm *ggla) decode(rs io.ReadSeeker) error {
+func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
 	var r uint32
 	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
 		return err
@@ -69,9 +69,18 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 	for {
 		var dims uint32
 		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
+			if errors.Is(err, io.EOF) {
+				return nil
+			}
 			return err
 		}
 
+		defer func() {
+			if errors.Is(retErr, io.EOF) {
+				retErr = io.ErrUnexpectedEOF
+			}
+		}()
+
 		var namesize uint32
 		if err := binary.Read(rs, binary.LittleEndian, &namesize); err != nil {
 			return err
@@ -108,7 +117,7 @@ func (llm *ggla) decode(rs io.ReadSeeker) error {
 			return err
 		}
 
-		if _, err := rs.Seek((offset+31)&-32, io.SeekStart); err != nil {
+		if _, err := rs.Seek((offset+31)&-32-offset, io.SeekCurrent); err != nil {
 			return err
 		}
 

llm/ggml.go

@@ -6,6 +6,8 @@ import (
 	"fmt"
 	"io"
 	"strings"
+
+	"github.com/ollama/ollama/util/bufioutil"
 )
 
 type GGML struct {
@@ -69,6 +71,30 @@ func (kv KV) HeadCountKV() uint64 {
 	return 1
 }
 
+func (kv KV) EmbeddingHeadCount() uint64 {
+	if heads := kv.HeadCount(); heads > 0 {
+		return kv.EmbeddingLength() / kv.HeadCount()
+	}
+
+	return 0
+}
+
+func (kv KV) EmbeddingHeadCountK() uint64 {
+	if k := kv.u64(fmt.Sprintf("%s.attention.key_length", kv.Architecture())); k > 0 {
+		return k
+	}
+
+	return kv.EmbeddingHeadCount()
+}
+
+func (kv KV) EmbeddingHeadCountV() uint64 {
+	if v := kv.u64(fmt.Sprintf("%s.attention.value_length", kv.Architecture())); v > 0 {
+		return v
+	}
+
+	return kv.EmbeddingHeadCount()
+}
+
 func (kv KV) GQA() uint64 {
 	return kv.HeadCount() / kv.HeadCountKV()
 }
@@ -254,7 +280,18 @@ func DetectGGMLType(b []byte) string {
 	}
 }
 
-func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
+// DecodeGGML decodes a GGML model from the given reader.
+//
+// It collects array values for arrays with a size less than or equal to
+// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
+// the maxArraySize is negative, all arrays are collected.
+func DecodeGGML(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
+	if maxArraySize == 0 {
+		maxArraySize = 1024
+	}
+
+	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
+
 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
 		return nil, 0, err
@@ -267,17 +304,15 @@ func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
 	case FILE_MAGIC_GGLA:
 		c = &containerGGLA{}
 	case FILE_MAGIC_GGUF_LE:
-		c = &containerGGUF{ByteOrder: binary.LittleEndian}
+		c = &containerGGUF{ByteOrder: binary.LittleEndian, maxArraySize: maxArraySize}
 	case FILE_MAGIC_GGUF_BE:
-		c = &containerGGUF{ByteOrder: binary.BigEndian}
+		c = &containerGGUF{ByteOrder: binary.BigEndian, maxArraySize: maxArraySize}
 	default:
 		return nil, 0, errors.New("invalid file magic")
 	}
 
 	model, err := c.Decode(rs)
-	if errors.Is(err, io.EOF) {
-		// noop
-	} else if err != nil {
+	if err != nil {
 		return nil, 0, err
 	}
 
@@ -297,7 +332,10 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 	embedding := llm.KV().EmbeddingLength()
 	heads := llm.KV().HeadCount()
 	headsKV := llm.KV().HeadCountKV()
-	vocab := uint64(len(llm.KV()["tokenizer.ggml.tokens"].([]any)))
+	vocab := uint64(llm.KV()["tokenizer.ggml.tokens"].(*array).size)
+
+	embeddingHeads := llm.KV().EmbeddingHeadCount()
+	embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
 
 	layers := llm.Tensors().Layers()
 
@@ -308,7 +346,7 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 		partialOffload = 4 * batch * embedding
 		partialOffload += max(
 			// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
-			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embedding/heads*headsKV),
+			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
 
@@ -316,21 +354,30 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			// mixtral 8x22b
 			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
-				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embedding/heads*headsKV),
-				4*(context*batch*heads+context*embedding/heads*headsKV+batch*1024+embedding/heads*headsKV*batch),
+				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
+				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
 			)
 		} else if ffnGateWeight, ok := layers["blk.0"]["ffn_gate.0.weight"]; ok {
 			// mixtral 8x7b
 			ffnGateWeight1 := ffnGateWeight.Shape[1]
 			fullOffload = 4 * batch * (2 + 3*embedding + context*(1+heads) + 2*headsKV + ffnGateWeight1)
 			partialOffload = max(
-				4*batch*(3+embedding/heads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
+				4*batch*(3+embeddingHeads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
 				4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
 			)
 		}
-	case "gemma":
-		fullOffload = 4 * batch * (embedding + vocab)
-		partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128
+	case "gemma", "gemma2":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
+		)
+
+		partialOffload = max(
+			4*embedding*batch+embedding*vocab*105/128+4*vocab*batch,
+			4*batch*(2*embedding+1+2*embeddingHeadsK*heads+context+context*heads)+
+				4*embeddingHeadsK*context*8+
+				embedding*embeddingHeadsK*heads*9/16,
+		)
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
@@ -367,6 +414,42 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			4*batch*(vocab+2*embedding),
 			fullOffload,
 		)
+	case "deepseek2":
+		fullOffload = max(
+			4*batch*(3*embedding+vocab),
+			4*batch*(3*embedding+2+context*(1+headsKV)+2*embeddingHeadsK*headsKV),
+		)
+
+		partialOffload = max(
+			4*batch*(3*embedding+vocab)+embedding*vocab*105/128,
+			4*batch*(2*embedding+1+2*embeddingHeadsK*headsKV+context+context*headsKV)+4*embeddingHeadsK*context*headsKV+embedding*embeddingHeadsK*headsKV*9/16,
+		)
+	case "chatglm":
+		fullOffload = 4 * batch * (embedding + vocab)
+		partialOffload = 4*batch*(embedding+vocab) + embedding*vocab*105/128
+		if qkvBias, ok := layers["blk.0"]["attn_qkv.bias"]; ok {
+			fullOffload = max(
+				fullOffload,
+				4*batch*(2+
+					2*embedding+
+					context+
+					context*heads+
+					embeddingHeadsK*heads+
+					qkvBias.Shape[0]),
+			)
+
+			partialOffload = max(
+				partialOffload,
+				4*batch*(1+
+					2*embedding+
+					embeddingHeadsK*heads+
+					context+
+					context*heads)+
+					4*embeddingHeadsK*context+
+					4*context*embeddingHeadsK+
+					4*qkvBias.Shape[0],
+			)
+		}
 	}
 
 	return

llm/ggml_test.go

@@ -0,0 +1 @@
+package llm

llm/gguf.go

@@ -3,11 +3,10 @@ package llm
 import (
 	"bytes"
 	"encoding/binary"
+	"encoding/json"
 	"fmt"
 	"io"
 	"strings"
-
-	"log/slog"
 )
 
 type containerGGUF struct {
@@ -29,6 +28,12 @@ type containerGGUF struct {
 		NumTensor uint64
 		NumKV     uint64
 	}
+
+	maxArraySize int
+}
+
+func (c *containerGGUF) canCollectArray(size int) bool {
+	return c.maxArraySize < 0 || size <= c.maxArraySize
 }
 
 func (c *containerGGUF) Name() string {
@@ -54,7 +59,6 @@ func (c *containerGGUF) Decode(rs io.ReadSeeker) (model, error) {
 	}
 
 	model := newGGUF(c)
-	slog.Debug(fmt.Sprintf("model = %#v", model))
 	if err := model.Decode(rs); err != nil {
 		return nil, err
 	}
@@ -85,6 +89,8 @@ type gguf struct {
 	tensors []*Tensor
 
 	parameters uint64
+
+	scratch [16 << 10]byte
 }
 
 func newGGUF(container *containerGGUF) *gguf {
@@ -181,34 +187,34 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	}
 
 	// decode tensors
-	for i := 0; uint64(i) < llm.numTensor(); i++ {
+	for range llm.numTensor() {
 		name, err := readGGUFString(llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor name: %w", err)
 		}
 
 		// dims is the number of dimensions in the tensor
 		dims, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor dimensions: %w", err)
 		}
 
 		shape := [4]uint64{1, 1, 1, 1}
 		for i := 0; uint32(i) < dims; i++ {
 			shape[i], err = readGGUF[uint64](llm, rs)
 			if err != nil {
-				return err
+				return fmt.Errorf("failed to read tensor shape: %w", err)
 			}
 		}
 
 		kind, err := readGGUF[uint32](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor kind: %w", err)
 		}
 
 		offset, err := readGGUF[uint64](llm, rs)
 		if err != nil {
-			return err
+			return fmt.Errorf("failed to read tensor offset: %w", err)
 		}
 
 		tensor := Tensor{
@@ -230,24 +236,19 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		alignment = 32
 	}
 
-	offset, err := rs.Seek(0, io.SeekCurrent)
-	if err != nil {
-		return err
-	}
-
-	padding := llm.padding(offset, int64(alignment))
-	if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
-		return err
-	}
-
 	for _, tensor := range llm.tensors {
-		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
-			return err
+		offset, err := rs.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return fmt.Errorf("failed to get current offset: %w", err)
 		}
 
-		padding := llm.padding(int64(tensor.Size()), int64(alignment))
+		padding := llm.padding(offset, int64(alignment))
 		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
-			return err
+			return fmt.Errorf("failed to seek to init padding: %w", err)
+		}
+
+		if _, err := rs.Seek(int64(tensor.Size()), io.SeekCurrent); err != nil {
+			return fmt.Errorf("failed to seek to tensor: %w", err)
 		}
 	}
 
@@ -285,22 +286,48 @@ func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
 	return b.String(), nil
 }
 
+func discardGGUFString(llm *gguf, r io.Reader) error {
+	buf := llm.scratch[:8]
+	_, err := io.ReadFull(r, buf)
+	if err != nil {
+		return err
+	}
+
+	size := int(llm.ByteOrder.Uint64(buf))
+	for size > 0 {
+		n, err := r.Read(llm.scratch[:min(size, cap(llm.scratch))])
+		if err != nil {
+			return err
+		}
+		size -= n
+	}
+	return nil
+}
+
 func readGGUFString(llm *gguf, r io.Reader) (string, error) {
 	if llm.Version == 1 {
 		return readGGUFV1String(llm, r)
 	}
 
-	var length uint64
-	if err := binary.Read(r, llm.ByteOrder, &length); err != nil {
+	buf := llm.scratch[:8]
+	_, err := io.ReadFull(r, buf)
+	if err != nil {
 		return "", err
 	}
 
-	var b bytes.Buffer
-	if _, err := io.CopyN(&b, r, int64(length)); err != nil {
+	length := int(llm.ByteOrder.Uint64(buf))
+	if length > len(llm.scratch) {
+		buf = make([]byte, length)
+	} else {
+		buf = llm.scratch[:length]
+	}
+	clear(buf)
+
+	_, err = io.ReadFull(r, buf)
+	if err != nil {
 		return "", err
 	}
-
-	return b.String(), nil
+	return string(buf), nil
 }
 
 func writeGGUFString(llm *gguf, w io.Writer, s string) error {
@@ -316,7 +343,16 @@ func writeGGUFString(llm *gguf, w io.Writer, s string) error {
 	return err
 }
 
-func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
+type array struct {
+	size   int
+	values []any
+}
+
+func (a *array) MarshalJSON() ([]byte, error) {
+	return json.Marshal(a.values)
+}
+
+func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
 	t, err := readGGUF[uint32](llm, r)
 	if err != nil {
 		return nil, err
@@ -327,7 +363,12 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
 		return nil, err
 	}
 
-	for i := 0; uint32(i) < n; i++ {
+	a := &array{size: int(n)}
+	if llm.canCollectArray(int(n)) {
+		a.values = make([]any, 0, int(n))
+	}
+
+	for i := range n {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -361,13 +402,15 @@ func readGGUFV1Array(llm *gguf, r io.Reader) (a []any, err error) {
 			return nil, err
 		}
 
-		a = append(a, e)
+		if a.values != nil {
+			a.values[i] = e
+		}
 	}
 
-	return
+	return a, nil
 }
 
-func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
+func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 	if llm.Version == 1 {
 		return readGGUFV1Array(llm, r)
 	}
@@ -382,7 +425,12 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 		return nil, err
 	}
 
-	for i := 0; uint64(i) < n; i++ {
+	a := &array{size: int(n)}
+	if llm.canCollectArray(int(n)) {
+		a.values = make([]any, int(n))
+	}
+
+	for i := range n {
 		var e any
 		switch t {
 		case ggufTypeUint8:
@@ -408,7 +456,11 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 		case ggufTypeBool:
 			e, err = readGGUF[bool](llm, r)
 		case ggufTypeString:
-			e, err = readGGUFString(llm, r)
+			if a.values != nil {
+				e, err = readGGUFString(llm, r)
+			} else {
+				err = discardGGUFString(llm, r)
+			}
 		default:
 			return nil, fmt.Errorf("invalid array type: %d", t)
 		}
@@ -416,10 +468,12 @@ func readGGUFArray(llm *gguf, r io.Reader) (a []any, err error) {
 			return nil, err
 		}
 
-		a = append(a, e)
+		if a.values != nil {
+			a.values[i] = e
+		}
 	}
 
-	return
+	return a, nil
 }
 
 func writeGGUFArray[S ~[]E, E any](llm *gguf, w io.Writer, t uint32, s S) error {
@@ -483,6 +537,7 @@ var ggufKVOrder = map[string][]string{
 		"tokenizer.ggml.add_bos_token",
 		"tokenizer.ggml.add_eos_token",
 		"tokenizer.chat_template",
+		"bert.pooling_type",
 	},
 }
 

llm/llm.go

@@ -1,12 +1,13 @@
 package llm
 
-// #cgo CFLAGS: -Illama.cpp
-// #cgo darwin,arm64 LDFLAGS: ${SRCDIR}/build/darwin/arm64_static/libllama.a -lstdc++
-// #cgo darwin,amd64 LDFLAGS: ${SRCDIR}/build/darwin/x86_64_static/libllama.a -lstdc++
-// #cgo windows,amd64 LDFLAGS: ${SRCDIR}/build/windows/amd64_static/libllama.a -static -lstdc++
-// #cgo windows,arm64 LDFLAGS: ${SRCDIR}/build/windows/arm64_static/libllama.a -static -lstdc++
-// #cgo linux,amd64 LDFLAGS: ${SRCDIR}/build/linux/x86_64_static/libllama.a -lstdc++
-// #cgo linux,arm64 LDFLAGS: ${SRCDIR}/build/linux/arm64_static/libllama.a -lstdc++
+// #cgo CFLAGS: -Illama.cpp -Illama.cpp/include -Illama.cpp/ggml/include
+// #cgo LDFLAGS: -lllama -lggml -lstdc++ -lpthread
+// #cgo darwin,arm64 LDFLAGS: -L${SRCDIR}/build/darwin/arm64_static -L${SRCDIR}/build/darwin/arm64_static/src -L${SRCDIR}/build/darwin/arm64_static/ggml/src -framework Accelerate -framework Metal
+// #cgo darwin,amd64 LDFLAGS: -L${SRCDIR}/build/darwin/x86_64_static -L${SRCDIR}/build/darwin/x86_64_static/src -L${SRCDIR}/build/darwin/x86_64_static/ggml/src
+// #cgo windows,amd64 LDFLAGS: -static-libstdc++ -static-libgcc -static -L${SRCDIR}/build/windows/amd64_static -L${SRCDIR}/build/windows/amd64_static/src -L${SRCDIR}/build/windows/amd64_static/ggml/src
+// #cgo windows,arm64 LDFLAGS: -static-libstdc++ -static-libgcc -static -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_static/ggml/src
+// #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/linux/x86_64_static -L${SRCDIR}/build/linux/x86_64_static/src -L${SRCDIR}/build/linux/x86_64_static/ggml/src
+// #cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/linux/arm64_static -L${SRCDIR}/build/linux/arm64_static/src -L${SRCDIR}/build/linux/arm64_static/ggml/src
 // #include <stdlib.h>
 // #include "llama.h"
 import "C"
@@ -32,7 +33,7 @@ func Quantize(infile, outfile string, ftype fileType) error {
 	params.ftype = ftype.Value()
 
 	if rc := C.llama_model_quantize(cinfile, coutfile, &params); rc != 0 {
-		return fmt.Errorf("llama_model_quantize: %d", rc)
+		return fmt.Errorf("failed to quantize model. This model architecture may not be supported, or you may need to upgrade Ollama to the latest version")
 	}
 
 	return nil

llm/memory.go

@@ -1,6 +1,7 @@
 package llm
 
 import (
+	"fmt"
 	"log/slog"
 	"strconv"
 	"strings"
@@ -49,6 +50,18 @@ type MemoryEstimate struct {
 
 	// For multi-GPU scenarios, this is the size in bytes per GPU
 	GPUSizes []uint64
+
+	// internal fields for logging purposes
+	inferenceLibrary    string
+	layersRequested     int
+	layersModel         int
+	availableList       []string
+	kv                  uint64
+	allocationsList     []string
+	memoryWeights       uint64
+	memoryLayerOutput   uint64
+	graphFullOffload    uint64
+	graphPartialOffload uint64
 }
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
@@ -102,8 +115,8 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		slog.Warn("model missing blk.0 layer size")
 	}
 
-	// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
-	var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()
+	// fp16 k,v = sizeof(float16) * n_ctx * n_layer * (n_embd_head_k + n_embd_head_v) * n_head_kv
+	var kv uint64 = 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * (ggml.KV().EmbeddingHeadCountK() + ggml.KV().EmbeddingHeadCountV()) * ggml.KV().HeadCountKV()
 
 	// KV is proportional to the number of layers
 	layerSize += kv / ggml.KV().BlockCount()
@@ -167,6 +180,11 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 
 	// For all the layers, find where they can fit on the GPU(s)
 	for i := range int(ggml.KV().BlockCount()) {
+		// Some models have inconsistent layer sizes
+		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
+			layerSize = blk.size()
+			layerSize += kv / ggml.KV().BlockCount()
+		}
 		memoryWeights += layerSize
 
 		if opts.NumGPU >= 0 && layerCount >= opts.NumGPU {
@@ -252,78 +270,86 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		allocationsList = append(allocationsList, format.HumanBytes2(a))
 	}
 
+	estimate := MemoryEstimate{
+		TotalSize: memoryRequiredTotal,
+		Layers:    0,
+		Graph:     0,
+		VRAMSize:  0,
+		GPUSizes:  []uint64{},
+
+		inferenceLibrary:    gpus[0].Library,
+		layersRequested:     opts.NumGPU,
+		layersModel:         int(ggml.KV().BlockCount()) + 1,
+		availableList:       availableList,
+		kv:                  kv,
+		allocationsList:     allocationsList,
+		memoryWeights:       memoryWeights,
+		memoryLayerOutput:   memoryLayerOutput,
+		graphFullOffload:    graphFullOffload,
+		graphPartialOffload: graphPartialOffload,
+	}
+
+	if gpus[0].Library == "cpu" {
+		return estimate
+	}
+	if layerCount == 0 {
+		slog.Debug("insufficient VRAM to load any model layers")
+		return estimate
+	}
+	estimate.Layers = layerCount
+	estimate.Graph = graphOffload
+	estimate.VRAMSize = memoryRequiredPartial
+	estimate.TotalSize = memoryRequiredTotal
+	estimate.TensorSplit = tensorSplit
+	estimate.GPUSizes = gpuAllocations
+	return estimate
+}
+
+func (m MemoryEstimate) log() {
 	slog.Info(
-		"offload to gpu",
+		"offload to "+m.inferenceLibrary,
 		slog.Group(
 			"layers",
 			// requested number of layers to offload
-			"requested", opts.NumGPU,
+			"requested", m.layersRequested,
 			// The number of layers the model has (including output)
-			"model", int(ggml.KV().BlockCount())+1,
+			"model", m.layersModel,
 			// estimated number of layers that can be offloaded
-			"offload", layerCount,
-			// multi-gpu split for tesnors
-			"split", tensorSplit,
+			"offload", m.Layers,
+			// multi-gpu split for tensors
+			"split", m.TensorSplit,
 		),
 		slog.Group(
 			"memory",
 			// memory available by GPU for offloading
-			"available", availableList,
+			"available", m.availableList,
 			slog.Group(
 				"required",
 				// memory required for full offloading
-				"full", format.HumanBytes2(memoryRequiredTotal),
+				"full", format.HumanBytes2(m.TotalSize),
 				// memory required to offload layers.estimate layers
-				"partial", format.HumanBytes2(memoryRequiredPartial),
+				"partial", format.HumanBytes2(m.VRAMSize),
 				// memory of KV cache
-				"kv", format.HumanBytes2(kv),
+				"kv", format.HumanBytes2(m.kv),
 				// Allocations across the GPUs
-				"allocations", allocationsList,
+				"allocations", m.allocationsList,
 			),
 			slog.Group(
 				"weights",
 				// memory of the weights
-				"total", format.HumanBytes2(memoryWeights),
+				"total", format.HumanBytes2(m.memoryWeights),
 				// memory of repeating layers
-				"repeating", format.HumanBytes2(memoryWeights-memoryLayerOutput),
+				"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
 				// memory of non-repeating layers
-				"nonrepeating", format.HumanBytes2(memoryLayerOutput),
+				"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
 			),
 			slog.Group(
 				"graph",
 				// memory of graph when fully offloaded
-				"full", format.HumanBytes2(graphFullOffload),
+				"full", format.HumanBytes2(m.graphFullOffload),
 				// memory of graph when not fully offloaded
-				"partial", format.HumanBytes2(graphPartialOffload),
+				"partial", format.HumanBytes2(m.graphPartialOffload),
 			),
 		),
 	)
-	if gpus[0].Library == "cpu" {
-		return MemoryEstimate{
-			Layers:    0,
-			Graph:     0,
-			VRAMSize:  0,
-			TotalSize: memoryRequiredTotal,
-			GPUSizes:  []uint64{},
-		}
-	}
-	if layerCount == 0 {
-		slog.Debug("insufficient VRAM to load any model layers")
-		return MemoryEstimate{
-			Layers:    0,
-			Graph:     0,
-			VRAMSize:  0,
-			TotalSize: memoryRequiredTotal,
-			GPUSizes:  []uint64{},
-		}
-	}
-
-	return MemoryEstimate{
-		Layers:      layerCount,
-		Graph:       graphOffload,
-		VRAMSize:    memoryRequiredPartial,
-		TotalSize:   memoryRequiredTotal,
-		TensorSplit: tensorSplit,
-		GPUSizes:    gpuAllocations,
-	}
 }

llm/memory_test.go

@@ -22,13 +22,14 @@ func TestEstimateGPULayers(t *testing.T) {
 	defer f.Close()
 	gguf := NewGGUFV3(binary.LittleEndian)
 	inputLayerCount := 5
+
 	tensors := []Tensor{
-		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.1.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.2.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.3.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "blk.4.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.1.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.2.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.3.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "blk.4.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
 	}
 	assert.Len(t, tensors, inputLayerCount+1)
 	err = gguf.Encode(f, KV{
@@ -45,8 +46,10 @@ func TestEstimateGPULayers(t *testing.T) {
 	}, tensors)
 	require.NoError(t, err)
 
-	ggml, err := LoadModel(f.Name())
-	require.NoError(t, err)
+	ggml, err := LoadModel(f.Name(), 0)
+	if err != nil {
+		t.Fatal(err)
+	}
 
 	// Simple CPU scenario
 	gpus := []gpu.GpuInfo{

llm/patches/01-load-progress.diff

@@ -1,8 +1,8 @@
 diff --git a/common/common.cpp b/common/common.cpp
-index ba1ecf0e..cead57cc 100644
+index 2c05a4d4..927f0e3d 100644
 --- a/common/common.cpp
 +++ b/common/common.cpp
-@@ -1836,6 +1836,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
+@@ -2093,6 +2093,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
      mparams.use_mmap        = params.use_mmap;
      mparams.use_mlock       = params.use_mlock;
      mparams.check_tensors   = params.check_tensors;
@@ -12,20 +12,20 @@ index ba1ecf0e..cead57cc 100644
          mparams.kv_overrides = NULL;
      } else {
 diff --git a/common/common.h b/common/common.h
-index d80344f2..71e84834 100644
+index 65c0ef81..ebca2c77 100644
 --- a/common/common.h
 +++ b/common/common.h
-@@ -174,6 +174,13 @@ struct gpt_params {
-     // multimodal models (see examples/llava)
+@@ -184,6 +184,13 @@ struct gpt_params {
      std::string mmproj = "";        // path to multimodal projector
      std::vector<std::string> image; // path to image file(s)
-+
+ 
 +    // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
 +    // If the provided progress_callback returns true, model loading continues.
 +    // If it returns false, model loading is immediately aborted.
 +    llama_progress_callback progress_callback = NULL;
 +    // context pointer passed to the progress callback
 +    void * progress_callback_user_data;
- };
- 
- void gpt_params_handle_model_default(gpt_params & params);
++
+     // embedding
+     bool embedding         = false; // get only sentence embedding
+     int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)

llm/patches/03-load_exception.diff

@@ -1,17 +1,8 @@
-From 544a2d2e646d39e878d87dfbb3398a356bc560ab Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Thu, 23 May 2024 11:18:45 -0700
-Subject: [PATCH] throw exception on load errors
-
----
- llama.cpp | 25 ++++++++++++++++---------
- 1 file changed, 16 insertions(+), 9 deletions(-)
-
-diff --git a/llama.cpp b/llama.cpp
-index 15c66077..8ba90b6a 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -6346,7 +6346,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 73f52435..58a00fb1 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -7241,7 +7241,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
          }
      } catch (const std::exception & err) {
          LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
@@ -20,7 +11,7 @@ index 15c66077..8ba90b6a 100644
      }
  
      return 0;
-@@ -15600,16 +15600,23 @@ struct llama_model * llama_load_model_from_file(
+@@ -17564,16 +17564,23 @@ struct llama_model * llama_load_model_from_file(
          }
          model->rpc_servers.push_back(servers);
      }
@@ -52,6 +43,3 @@ index 15c66077..8ba90b6a 100644
      }
  
      return model;
--- 
-2.45.1
-

llm/patches/04-metal.diff

@@ -1,7 +1,7 @@
-diff --git a/ggml-metal.m b/ggml-metal.m
+diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
 index 0207b787..b5e9884b 100644
---- a/ggml-metal.m
-+++ b/ggml-metal.m
+--- a/ggml/src/ggml-metal.m
++++ b/ggml/src/ggml-metal.m
 @@ -1396,27 +1396,23 @@ static enum ggml_status ggml_metal_graph_compute(
                          // to the matrix-vector kernel
                          int ne11_mm_min = 1;

llm/patches/05-default-pretokenizer.diff

@@ -1,11 +1,11 @@
-diff --git a/llama.cpp b/llama.cpp
-index 40d2ec2c..74f3ee9c 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -4642,16 +4642,7 @@ static void llm_load_vocab(
- 
-         // for now, only BPE models have pre-tokenizers
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 8fe51971..7113ba64 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -5433,16 +5433,7 @@ static void llm_load_vocab(
          if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
+             vocab.tokenizer_add_space_prefix = false;
+             vocab.tokenizer_clean_spaces = true;
 -            if (tokenizer_pre.empty()) {
 -                LLAMA_LOG_WARN("%s: missing pre-tokenizer type, using: 'default'\n", __func__);
 -                LLAMA_LOG_WARN("%s:                                             \n", __func__);
@@ -15,18 +15,18 @@ index 40d2ec2c..74f3ee9c 100644
 -                LLAMA_LOG_WARN("%s: ************************************        \n", __func__);
 -                LLAMA_LOG_WARN("%s:                                             \n", __func__);
 -                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
--            } else if (
-+            if (
-                     tokenizer_pre == "default") {
+-            } else if (tokenizer_pre == "default") {
++            if (tokenizer_pre == "default") {
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
-@@ -4703,7 +4694,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "smaug-bpe") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG;
+                     tokenizer_pre == "llama3"   ||
+@@ -5526,7 +5517,8 @@ static void llm_load_vocab(
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMOLLM;
+                 vocab.tokenizer_clean_spaces = false;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
 +                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
 +                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              }
-         } else {
+         } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
              vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;

llm/patches/06-embeddings.diff

@@ -0,0 +1,45 @@
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 1fe2b9f7..a43312a7 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -13689,7 +13689,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
+     const auto n_embd  = hparams.n_embd;
+ 
+     // TODO: use a per-batch flag for logits presence instead
+-    const bool has_logits = !cparams.embeddings;
++    const bool has_logits =  cparams.causal_attn;
+     const bool has_embd   =  lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE));
+ 
+     const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
+@@ -13959,17 +13959,25 @@ static int llama_decode_internal(
+             // no output
+             res  = nullptr;
+             embd = nullptr;
+-        } else if (cparams.embeddings) {
+-            res = nullptr; // do not extract logits for embedding case
+-            embd = gf->nodes[gf->n_nodes - 1];
+-            if (strcmp(embd->name, "result_embd_pooled") != 0) {
+-                embd = gf->nodes[gf->n_nodes - 2];
++        }
++
++        if (cparams.embeddings) {
++            for (int i = gf->n_nodes - 1; i >= 0; --i) {
++                embd = gf->nodes[i];
++                if (strcmp(embd->name, "result_embd_pooled") == 0) {
++                    break;
++                }
+             }
+             GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+-        } else {
++         } else {
+             embd = nullptr; // do not extract embeddings when not needed
+             GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
+         }
++
++        if (!cparams.causal_attn) {
++            res = nullptr; // do not extract logits when not needed
++        }
++
+         // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
+ 
+         ggml_backend_sched_alloc_graph(lctx.sched, gf);

llm/patches/06-qwen2.diff

@@ -1,13 +0,0 @@
-diff --git a/llama.cpp b/llama.cpp
-index 40d2ec2c..f34eb79a 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -6943,7 +6943,7 @@ static struct ggml_tensor * llm_build_kqv(
-         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
-         cb(kq, "kq", il);
- 
--        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
-+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
-             // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
-             // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
-             ggml_mul_mat_set_prec(kq, GGML_PREC_F32);

llm/patches/07-clip-unicode.diff

@@ -0,0 +1,42 @@
+diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
+index 95fbe3d0..5a02a6ec 100644
+--- a/examples/llava/clip.cpp
++++ b/examples/llava/clip.cpp
+@@ -32,6 +33,14 @@
+ #include <cinttypes>
+ #include <limits>
+ 
++#if defined(_WIN32)
++#define WIN32_LEAN_AND_MEAN
++#ifndef NOMINMAX
++    #define NOMINMAX
++#endif
++#include <windows.h>
++#endif
++
+ //#define CLIP_DEBUG_FUNCTIONS
+ 
+ // RGB uint8 image
+@@ -1055,7 +1064,22 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+             return nullptr;
+         }
+ 
++#ifdef _WIN32
++        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
++        if (!wlen) {
++            return NULL;
++        }
++        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
++        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
++        if (!wlen) {
++            free(wbuf);
++            return NULL;
++        }
++        auto fin = std::ifstream(wbuf, std::ios::binary);
++        free(wbuf);
++#else
+         auto fin = std::ifstream(fname, std::ios::binary);
++#endif
+         if (!fin) {
+             LOG_TEE("cannot open model file for loading tensors\n");
+             clip_free(new_clip);

llm/patches/08-pooling.diff

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

llm/patches/09-lora.diff

@@ -0,0 +1,360 @@
+diff --git a/common/common.cpp b/common/common.cpp
+index dbb724fb..c26fe6ee 100644
+--- a/common/common.cpp
++++ b/common/common.cpp
+@@ -2087,14 +2087,29 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
+     for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
+         const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
+         float lora_scale = std::get<1>(params.lora_adapter[i]);
++
++        // try to load as gguf
+         auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
+         if (adapter == nullptr) {
+-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+-            llama_free(lctx);
+-            llama_free_model(model);
+-            return std::make_tuple(nullptr, nullptr);
++            fprintf(stderr, "%s: error: failed to apply lora adapter, trying ggla\n", __func__);
++
++            // if that fails, try loading as ggla for compatibility
++            int err = llama_model_apply_lora_from_file(model,
++                                                    lora_adapter.c_str(),
++                                                    lora_scale,
++                                                    ((i > 0) || params.lora_base.empty())
++                                                        ? NULL
++                                                        : params.lora_base.c_str(),
++                                                    params.n_threads);
++            if (err != 0) {
++                fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
++                llama_free(lctx);
++                llama_free_model(model);
++                return std::make_tuple(nullptr, nullptr);
++            }
++        } else {
++            llama_lora_adapter_set(lctx, adapter, lora_scale);
+         }
+-        llama_lora_adapter_set(lctx, adapter, lora_scale);
+     }
+ 
+     if (params.ignore_eos) {
+diff --git a/include/llama.h b/include/llama.h
+index 93fd77ca..b0fb37a6 100644
+--- a/include/llama.h
++++ b/include/llama.h
+@@ -1160,6 +1160,20 @@ extern "C" {
+ 
+     LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
+ 
++    // Apply a LoRA adapter to a loaded model
++    // path_base_model is the path to a higher quality model to use as a base for
++    // the layers modified by the adapter. Can be NULL to use the current loaded model.
++    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
++    // will be applied on top of the previous one
++    // Returns 0 on success
++    LLAMA_API int32_t llama_model_apply_lora_from_file(
++            const struct llama_model * model,
++                            const char * path_lora,
++                                float   scale,
++                            const char * path_base_model,
++                                int32_t   n_threads);
++
++
+ #ifdef __cplusplus
+ }
+ #endif
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 80a0dd0f..9d7b0e17 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -21880,3 +21880,290 @@ static void llama_log_callback_default(ggml_log_level level, const char * text,
+     fputs(text, stderr);
+     fflush(stderr);
+ }
++
++static int llama_apply_lora_from_file_internal(
++    const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads
++) {
++    LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
++
++    const int64_t t_start_lora_us = ggml_time_us();
++
++    llama_file fin(path_lora, "rb");
++
++    // verify magic and version
++    {
++        uint32_t magic = fin.read_u32();
++        if (magic != LLAMA_FILE_MAGIC_GGLA) {
++            LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
++            return 1;
++        }
++
++        uint32_t format_version = fin.read_u32();
++        if (format_version != 1) {
++            LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
++            return 1;
++        }
++    }
++
++    int32_t lora_r = fin.read_u32();
++    int32_t lora_alpha = fin.read_u32();
++    float scaling = scale * (float)lora_alpha / (float)lora_r;
++
++    LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
++
++    // load base model
++    std::unique_ptr<llama_model_loader> ml;
++    if (path_base_model) {
++        LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
++        ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*check_tensors*/ false, /*kv_overrides*/ nullptr));
++        ml->init_mappings(/*prefetch*/ false); // no prefetching
++    }
++
++    struct tensor_meta {
++        std::string name;
++        ggml_type type;
++        int32_t ne[2];
++        size_t offset;
++    };
++    std::map<std::string, tensor_meta> tensor_meta_map;
++
++    // load all tensor meta
++    while (true) {
++        if (fin.tell() == fin.size) {
++            // eof
++            break;
++        }
++
++        int32_t n_dims;
++        int32_t name_len;
++        int32_t ftype;
++
++        fin.read_raw(&n_dims, sizeof(n_dims));
++        fin.read_raw(&name_len, sizeof(name_len));
++        fin.read_raw(&ftype, sizeof(ftype));
++
++        if (n_dims != 1 && n_dims != 2) {
++            LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
++            return 1;
++        }
++
++        int32_t ne[2] = { 1, 1 };
++        for (int i = 0; i < n_dims; ++i) {
++            fin.read_raw(&ne[i], sizeof(ne[i]));
++        }
++
++        std::string name;
++        {
++            GGML_ASSERT(name_len < GGML_MAX_NAME);
++            char buf[GGML_MAX_NAME];
++            fin.read_raw(buf, name_len);
++            name = std::string(buf, name_len);
++        }
++
++        // check for lora suffix
++        std::string lora_suffix;
++        if (name.length() > 6) {
++            lora_suffix = name.substr(name.length() - 6);
++        }
++        if (lora_suffix != ".loraA" && lora_suffix != ".loraB") {
++            LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
++            return 1;
++        }
++
++        // tensor type
++        ggml_type wtype;
++        switch (ftype) {
++            case 0: wtype = GGML_TYPE_F32;  break;
++            case 1: wtype = GGML_TYPE_F16;  break;
++            default:
++                    {
++                        LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
++                                __func__, ftype);
++                        return 1;
++                    }
++        }
++
++        // data offset
++        size_t offset = fin.tell();
++        offset = (offset + 31) & -32;
++
++        // skip tensor data
++        fin.seek(offset + ggml_row_size(wtype, ne[0]) * ne[1], SEEK_SET);
++
++        tensor_meta_map.emplace(name, tensor_meta{ name, wtype, { ne[0], ne[1] }, offset });
++    }
++
++    bool warned = false;
++    int n_tensors = 0;
++
++    // apply
++    ggml_backend_t backend_cpu = ggml_backend_cpu_init();
++    if (backend_cpu == nullptr) {
++        LLAMA_LOG_ERROR("%s: error: failed to initialize cpu backend\n", __func__);
++        return 1;
++    }
++    ggml_backend_cpu_set_n_threads(backend_cpu, n_threads);
++
++    std::vector<no_init<uint8_t>> read_buf;
++    for (const auto & it : model.tensors_by_name) {
++        const std::string & base_name = it.first;
++        ggml_tensor * model_t = it.second;
++
++        if (tensor_meta_map.find(base_name + ".loraA") == tensor_meta_map.end() ||
++            tensor_meta_map.find(base_name + ".loraB") == tensor_meta_map.end()) {
++            continue;
++        }
++
++        tensor_meta & metaA = tensor_meta_map.at(base_name + ".loraA");
++        tensor_meta & metaB = tensor_meta_map.at(base_name + ".loraB");
++
++        ggml_init_params lora_init_params = {
++            /* .mem_size   */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
++            /* .mem_buffer */ nullptr,
++            /* .no_alloc   */ true,
++        };
++        ggml_context * lora_ctx = ggml_init(lora_init_params);
++        if (lora_ctx == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to initialize lora context\n", __func__);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        // create tensors
++        ggml_tensor * loraA = ggml_new_tensor_2d(lora_ctx, metaA.type, metaA.ne[0], metaA.ne[1]);
++        ggml_tensor * loraB = ggml_new_tensor_2d(lora_ctx, metaB.type, metaB.ne[0], metaB.ne[1]);
++        ggml_set_name(loraA, metaA.name.c_str());
++        ggml_set_name(loraB, metaB.name.c_str());
++
++        ggml_tensor * base_t;
++        if (ml) {
++            if (!ml->get_tensor_meta(base_name.c_str())) {
++                LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
++                return 1;
++            }
++            base_t = ggml_dup_tensor(lora_ctx, ml->get_tensor_meta(base_name.c_str()));
++        } else {
++            base_t = ggml_dup_tensor(lora_ctx, model_t);
++        }
++        ggml_set_name(base_t, base_name.c_str());
++
++        // allocate in backend buffer
++        ggml_backend_buffer_t lora_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
++        if (lora_buf == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to allocate lora tensors\n", __func__);
++            return 1;
++        }
++
++        // load tensor data
++        auto load_tensor = [&read_buf, &fin](const tensor_meta & tensor_meta, ggml_tensor * tensor) {
++            read_buf.resize(ggml_nbytes(tensor));
++            fin.seek(tensor_meta.offset, SEEK_SET);
++            fin.read_raw(read_buf.data(), ggml_nbytes(tensor));
++            ggml_backend_tensor_set(tensor, read_buf.data(), 0, read_buf.size());
++        };
++        load_tensor(metaA, loraA);
++        load_tensor(metaB, loraB);
++
++        // load base model tensor data
++        if (ml) {
++            ml->load_data_for(base_t);
++        } else {
++            ggml_backend_tensor_copy(model_t, base_t);
++        }
++
++        if (ggml_is_quantized(base_t->type) && !warned) {
++            LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
++                            "use a f16 or f32 base model with --lora-base\n", __func__);
++            warned = true;
++        }
++
++        if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
++            LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
++                            " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
++            ggml_free(lora_ctx);
++            ggml_backend_buffer_free(lora_buf);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        auto build_lora_graph = [&]() {
++            // w = w + BA*s
++            ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
++            ggml_set_name(BA, "BA");
++
++            if (scaling != 1.0f) {
++                BA = ggml_scale(lora_ctx, BA, scaling);
++                ggml_set_name(BA, "BA_scaled");
++            }
++
++            ggml_tensor * r;
++            r = ggml_add_inplace(lora_ctx, base_t, BA);
++            ggml_set_name(r, "r_add");
++
++            if (base_t->type != model_t->type) {
++                // convert the result to the model type
++                r = ggml_cast(lora_ctx, r, model_t->type);
++                ggml_set_name(r, "r_cast");
++            }
++
++            return r;
++        };
++
++        ggml_cgraph * gf = ggml_new_graph(lora_ctx);
++        ggml_tensor * r = build_lora_graph();
++        ggml_build_forward_expand(gf, r);
++
++        ggml_backend_buffer_t graph_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
++        if (graph_buf == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to allocate graph tensors\n", __func__);
++            ggml_free(lora_ctx);
++            ggml_backend_buffer_free(lora_buf);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        ggml_backend_graph_compute(backend_cpu, gf);
++
++        ggml_backend_tensor_set(model_t, r->data, 0, ggml_nbytes(r));
++
++#if 0
++        // TODO: use scheduler with fallback to CPU for less copies between CPU and GPU
++        //ggml_backend_sched_t sched = ggml_backend_sched_new(backends.data(), backends.size(), GGML_DEFAULT_GRAPH_SIZE);
++
++        // sched compute
++        ggml_build_forward_expand(gf, build_graph());
++        ggml_backend_sched_init_measure(sched, gf);
++
++        // create the graph again, since the previous one was destroyed by the measure
++        ggml_graph_clear(gf);
++        ggml_build_forward_expand(gf, build_graph());
++        ggml_backend_sched_graph_compute(sched, gf);
++        ggml_backend_sched_free(sched);
++#endif
++
++        ggml_backend_buffer_free(lora_buf);
++        ggml_backend_buffer_free(graph_buf);
++        ggml_free(lora_ctx);
++
++        n_tensors++;
++        if (n_tensors % 4 == 0) {
++            LLAMA_LOG_INFO(".");
++        }
++    }
++
++    ggml_backend_free(backend_cpu);
++
++    const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
++    LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
++
++    return 0;
++}
++
++int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) {
++    try {
++        return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads);
++    } catch (const std::exception & err) {
++        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
++        return 1;
++    }
++}
+\ No newline at end of file

llm/payload.go

@@ -38,7 +38,7 @@ func Init() error {
 	}
 
 	var variants []string
-	for v := range availableServers() {
+	for v := range getAvailableServers() {
 		variants = append(variants, v)
 	}
 	slog.Info(fmt.Sprintf("Dynamic LLM libraries %v", variants))
@@ -50,7 +50,7 @@ func Init() error {
 // binary names may contain an optional variant separated by '_'
 // For example, "ollama_rocm_v6" and "ollama_rocm_v5" or "ollama_cpu" and "ollama_cpu_avx2"
 // Any library without a variant is the lowest common denominator
-func availableServers() map[string]string {
+func getAvailableServers() map[string]string {
 	payloadsDir, err := gpu.PayloadsDir()
 	if err != nil {
 		slog.Error("payload lookup error", "error", err)
@@ -58,7 +58,7 @@ func availableServers() map[string]string {
 	}
 
 	// glob payloadsDir for files that start with ollama_
-	pattern := filepath.Join(payloadsDir, "*")
+	pattern := filepath.Join(payloadsDir, "*", "ollama_*")
 
 	files, err := filepath.Glob(pattern)
 	if err != nil {
@@ -69,7 +69,7 @@ func availableServers() map[string]string {
 	servers := make(map[string]string)
 	for _, file := range files {
 		slog.Debug("availableServers : found", "file", file)
-		servers[filepath.Base(file)] = file
+		servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
 	}
 
 	return servers
@@ -80,7 +80,7 @@ func availableServers() map[string]string {
 // TODO - switch to metadata based mapping
 func serversForGpu(info gpu.GpuInfo) []string {
 	// glob workDir for files that start with ollama_
-	availableServers := availableServers()
+	availableServers := getAvailableServers()
 	requested := info.Library
 	if info.Variant != gpu.CPUCapabilityNone {
 		requested += "_" + info.Variant.String()
@@ -115,27 +115,29 @@ func serversForGpu(info gpu.GpuInfo) []string {
 		servers = append(servers, alt...)
 	}
 
-	// Load up the best CPU variant if not primary requested
-	if info.Library != "cpu" {
-		variant := gpu.GetCPUCapability()
-		// If no variant, then we fall back to default
-		// If we have a variant, try that if we find an exact match
-		// Attempting to run the wrong CPU instructions will panic the
-		// process
-		if variant != gpu.CPUCapabilityNone {
-			for cmp := range availableServers {
-				if cmp == "cpu_"+variant.String() {
-					servers = append(servers, cmp)
-					break
+	if !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
+		// Load up the best CPU variant if not primary requested
+		if info.Library != "cpu" {
+			variant := gpu.GetCPUCapability()
+			// If no variant, then we fall back to default
+			// If we have a variant, try that if we find an exact match
+			// Attempting to run the wrong CPU instructions will panic the
+			// process
+			if variant != gpu.CPUCapabilityNone {
+				for cmp := range availableServers {
+					if cmp == "cpu_"+variant.String() {
+						servers = append(servers, cmp)
+						break
+					}
 				}
+			} else {
+				servers = append(servers, "cpu")
 			}
-		} else {
-			servers = append(servers, "cpu")
 		}
-	}
 
-	if len(servers) == 0 {
-		servers = []string{"cpu"}
+		if len(servers) == 0 {
+			servers = []string{"cpu"}
+		}
 	}
 
 	return servers
@@ -147,7 +149,7 @@ func serverForCpu() string {
 		return "metal"
 	}
 	variant := gpu.GetCPUCapability()
-	availableServers := availableServers()
+	availableServers := getAvailableServers()
 	if variant != gpu.CPUCapabilityNone {
 		for cmp := range availableServers {
 			if cmp == "cpu_"+variant.String() {

llm/server.go

@@ -33,7 +33,7 @@ type LlamaServer interface {
 	Ping(ctx context.Context) error
 	WaitUntilRunning(ctx context.Context) error
 	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
-	Embedding(ctx context.Context, prompt string) ([]float64, error)
+	Embed(ctx context.Context, input []string) ([][]float32, error)
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
@@ -60,7 +60,12 @@ type llmServer struct {
 	sem *semaphore.Weighted
 }
 
-func LoadModel(model string) (*GGML, error) {
+// LoadModel will load a model from disk. The model must be in the GGML format.
+//
+// It collects array values for arrays with a size less than or equal to
+// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
+// the maxArraySize is negative, all arrays are collected.
+func LoadModel(model string, maxArraySize int) (*GGML, error) {
 	if _, err := os.Stat(model); err != nil {
 		return nil, err
 	}
@@ -71,17 +76,29 @@ func LoadModel(model string) (*GGML, error) {
 	}
 	defer f.Close()
 
-	ggml, _, err := DecodeGGML(f)
+	ggml, _, err := DecodeGGML(f, maxArraySize)
 	return ggml, err
 }
 
 // NewLlamaServer will run a server for the given GPUs
 // The gpu list must be a single family.
-func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options) (LlamaServer, error) {
+func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
 	var err error
 	var cpuRunner string
 	var estimate MemoryEstimate
-	var systemMemory uint64
+	var systemTotalMemory uint64
+	var systemFreeMemory uint64
+	var systemSwapFreeMemory uint64
+
+	systemMemInfo, err := gpu.GetCPUMem()
+	if err != nil {
+		slog.Error("failed to lookup system memory", "error", err)
+	} else {
+		systemTotalMemory = systemMemInfo.TotalMemory
+		systemFreeMemory = systemMemInfo.FreeMemory
+		systemSwapFreeMemory = systemMemInfo.FreeSwap
+		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
+	}
 
 	// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
 	if opts.NumGPU == 0 {
@@ -91,19 +108,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		cpuRunner = serverForCpu()
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
-		if gpus[0].Library == "metal" {
-			memInfo, err := gpu.GetCPUMem()
-			if err != nil {
-				slog.Error("failed to lookup system memory", "error", err)
-			} else {
-				systemMemory = memInfo.TotalMemory
-				slog.Debug("system memory", "total", format.HumanBytes2(systemMemory))
-			}
-		}
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
 
 		switch {
-		case gpus[0].Library == "metal" && estimate.VRAMSize > systemMemory:
+		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:
 			// disable partial offloading when model is greater than total system memory as this
 			// can lead to locking up the system
 			opts.NumGPU = 0
@@ -116,6 +124,18 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 	}
 
+	// On linux, over-allocating CPU memory will almost always result in an error
+	if runtime.GOOS == "linux" {
+		systemMemoryRequired := estimate.TotalSize - estimate.VRAMSize
+		available := systemFreeMemory + systemSwapFreeMemory
+		if systemMemoryRequired > available {
+			slog.Warn("model request too large for system", "requested", format.HumanBytes2(systemMemoryRequired), "available", available, "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "swap", format.HumanBytes2(systemSwapFreeMemory))
+			return nil, fmt.Errorf("model requires more system memory (%s) than is available (%s)", format.HumanBytes2(systemMemoryRequired), format.HumanBytes2(available))
+		}
+	}
+
+	estimate.log()
+
 	// Loop through potential servers
 	finalErr := errors.New("no suitable llama servers found")
 
@@ -123,7 +143,20 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}
 
-	availableServers := availableServers()
+	availableServers := getAvailableServers()
+	if len(availableServers) == 0 {
+		if runtime.GOOS != "windows" {
+			slog.Warn("llama server binary disappeared, reinitializing payloads")
+			err = Init()
+			if err != nil {
+				slog.Warn("failed to reinitialize payloads", "error", err)
+				return nil, err
+			}
+			availableServers = getAvailableServers()
+		} else {
+			return nil, finalErr
+		}
+	}
 	var servers []string
 	if cpuRunner != "" {
 		servers = []string{cpuRunner}
@@ -200,7 +233,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if g.Library == "metal" &&
 			uint64(opts.NumGPU) > 0 &&
 			uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
-			opts.UseMMap = false
+			opts.UseMMap = new(bool)
+			*opts.UseMMap = false
 		}
 	}
 
@@ -208,7 +242,13 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--flash-attn")
 	}
 
-	if !opts.UseMMap {
+	// Windows CUDA should not use mmap for best performance
+	// Linux  with a model larger than free space, mmap leads to thrashing
+	// For CPU loads we want the memory to be allocated, not FS cache
+	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == nil) ||
+		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == nil) ||
+		(gpus[0].Library == "cpu" && opts.UseMMap == nil) ||
+		(opts.UseMMap != nil && !*opts.UseMMap) {
 		params = append(params, "--no-mmap")
 	}
 
@@ -220,25 +260,12 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--numa")
 	}
 
-	numParallel := envconfig.NumParallel
-
-	// TODO (jmorganca): multimodal models don't support parallel yet
-	// see https://github.com/ollama/ollama/issues/4165
-	if len(projectors) > 0 {
-		numParallel = 1
-		slog.Warn("multimodal models don't support parallel requests yet")
-	}
-
 	params = append(params, "--parallel", fmt.Sprintf("%d", numParallel))
 
 	if estimate.TensorSplit != "" {
 		params = append(params, "--tensor-split", estimate.TensorSplit)
 	}
 
-	if estimate.TensorSplit != "" {
-		params = append(params, "--tensor-split", estimate.TensorSplit)
-	}
-
 	for i := range len(servers) {
 		dir := availableServers[servers[i]]
 		if dir == "" {
@@ -271,8 +298,8 @@ 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
-		libraryPaths := []string{dir}
+		// prepend the server directory to LD_LIBRARY_PATH/PATH and the parent dir for common dependencies
+		libraryPaths := []string{dir, filepath.Dir(dir)}
 
 		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
 			// Append our runner directory to the path
@@ -358,8 +385,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			filteredEnv := []string{}
 			for _, ev := range s.cmd.Env {
 				if strings.HasPrefix(ev, "CUDA_") ||
+					strings.HasPrefix(ev, "ROCR_") ||
 					strings.HasPrefix(ev, "ROCM_") ||
 					strings.HasPrefix(ev, "HIP_") ||
+					strings.HasPrefix(ev, "GPU_") ||
 					strings.HasPrefix(ev, "HSA_") ||
 					strings.HasPrefix(ev, "GGML_") ||
 					strings.HasPrefix(ev, "PATH=") ||
@@ -388,7 +417,17 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 
 		// reap subprocess when it exits
 		go func() {
-			s.done <- s.cmd.Wait()
+			err := s.cmd.Wait()
+			// Favor a more detailed message over the process exit status
+			if err != nil && s.status != nil && s.status.LastErrMsg != "" {
+				slog.Debug("llama runner terminated", "error", err)
+				if strings.Contains(s.status.LastErrMsg, "unknown model") {
+					s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
+				}
+				s.done <- fmt.Errorf(s.status.LastErrMsg)
+			} else {
+				s.done <- err
+			}
 		}()
 
 		return s, nil
@@ -405,7 +444,7 @@ func projectorMemoryRequirements(filename string) uint64 {
 	}
 	defer file.Close()
 
-	ggml, _, err := DecodeGGML(file)
+	ggml, _, err := DecodeGGML(file, 0)
 	if err != nil {
 		return 0
 	}
@@ -551,11 +590,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 			slog.Warn("client connection closed before server finished loading, aborting load")
 			return fmt.Errorf("timed out waiting for llama runner to start: %w", ctx.Err())
 		case err := <-s.done:
-			msg := ""
-			if s.status != nil && s.status.LastErrMsg != "" {
-				msg = s.status.LastErrMsg
-			}
-			return fmt.Errorf("llama runner process has terminated: %v %s", err, msg)
+			return fmt.Errorf("llama runner process has terminated: %w", err)
 		default:
 		}
 		if time.Now().After(stallTimer) {
@@ -657,7 +692,7 @@ type CompletionRequest struct {
 	Prompt  string
 	Format  string
 	Images  []ImageData
-	Options api.Options
+	Options *api.Options
 }
 
 type CompletionResponse struct {
@@ -677,10 +712,9 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	}
 	defer s.sem.Release(1)
 
-	// only allow maximum 10 "context shifts" to avoid infinite generation
+	// put an upper limit on num_predict to avoid the model running on forever
 	if req.Options.NumPredict < 0 || req.Options.NumPredict > 10*s.options.NumCtx {
 		req.Options.NumPredict = 10 * s.options.NumCtx
-		slog.Debug("setting token limit to 10x num_ctx", "num_ctx", s.options.NumCtx, "num_predict", req.Options.NumPredict)
 	}
 
 	request := map[string]any{
@@ -838,15 +872,15 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	return nil
 }
 
-type EmbeddingRequest struct {
-	Content string `json:"content"`
+type EmbedRequest struct {
+	Content []string `json:"content"`
 }
 
-type EmbeddingResponse struct {
-	Embedding []float64 `json:"embedding"`
+type EmbedResponse struct {
+	Embedding [][]float32 `json:"embedding"`
 }
 
-func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, error) {
+func (s *llmServer) Embed(ctx context.Context, input []string) ([][]float32, error) {
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		slog.Error("Failed to acquire semaphore", "error", err)
 		return nil, err
@@ -861,7 +895,7 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
 		return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
 	}
 
-	data, err := json.Marshal(TokenizeRequest{Content: prompt})
+	data, err := json.Marshal(EmbedRequest{Content: input})
 	if err != nil {
 		return nil, fmt.Errorf("error marshaling embed data: %w", err)
 	}
@@ -888,7 +922,7 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
 		return nil, fmt.Errorf("%s", body)
 	}
 
-	var embedding EmbeddingResponse
+	var embedding EmbedResponse
 	if err := json.Unmarshal(body, &embedding); err != nil {
 		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
 	}

llm/status.go

@@ -25,6 +25,7 @@ var errorPrefixes = []string{
 	"CUDA error",
 	"cudaMalloc failed",
 	"\"ERR\"",
+	"error loading model",
 }
 
 func (w *StatusWriter) Write(b []byte) (int, error) {

openai/openai.go

@@ -3,15 +3,19 @@ package openai
 
 import (
 	"bytes"
+	"encoding/base64"
 	"encoding/json"
 	"fmt"
 	"io"
+	"log/slog"
 	"math/rand"
 	"net/http"
+	"strings"
 	"time"
 
 	"github.com/gin-gonic/gin"
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
 )
 
 type Error struct {
@@ -26,8 +30,9 @@ type ErrorResponse struct {
 }
 
 type Message struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
+	Role      string     `json:"role"`
+	Content   any        `json:"content"`
+	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
 }
 
 type Choice struct {
@@ -42,6 +47,12 @@ type ChunkChoice struct {
 	FinishReason *string `json:"finish_reason"`
 }
 
+type CompleteChunkChoice struct {
+	Text         string  `json:"text"`
+	Index        int     `json:"index"`
+	FinishReason *string `json:"finish_reason"`
+}
+
 type Usage struct {
 	PromptTokens     int `json:"prompt_tokens"`
 	CompletionTokens int `json:"completion_tokens"`
@@ -52,6 +63,11 @@ type ResponseFormat struct {
 	Type string `json:"type"`
 }
 
+type EmbedRequest struct {
+	Input any    `json:"input"`
+	Model string `json:"model"`
+}
+
 type ChatCompletionRequest struct {
 	Model            string          `json:"model"`
 	Messages         []Message       `json:"messages"`
@@ -64,6 +80,7 @@ type ChatCompletionRequest struct {
 	PresencePenalty  *float64        `json:"presence_penalty_penalty"`
 	TopP             *float64        `json:"top_p"`
 	ResponseFormat   *ResponseFormat `json:"response_format"`
+	Tools            []api.Tool      `json:"tools"`
 }
 
 type ChatCompletion struct {
@@ -85,6 +102,73 @@ type ChatCompletionChunk struct {
 	Choices           []ChunkChoice `json:"choices"`
 }
 
+// TODO (https://github.com/ollama/ollama/issues/5259): support []string, []int and [][]int
+type CompletionRequest struct {
+	Model            string   `json:"model"`
+	Prompt           string   `json:"prompt"`
+	FrequencyPenalty float32  `json:"frequency_penalty"`
+	MaxTokens        *int     `json:"max_tokens"`
+	PresencePenalty  float32  `json:"presence_penalty"`
+	Seed             *int     `json:"seed"`
+	Stop             any      `json:"stop"`
+	Stream           bool     `json:"stream"`
+	Temperature      *float32 `json:"temperature"`
+	TopP             float32  `json:"top_p"`
+	Suffix           string   `json:"suffix"`
+}
+
+type Completion struct {
+	Id                string                `json:"id"`
+	Object            string                `json:"object"`
+	Created           int64                 `json:"created"`
+	Model             string                `json:"model"`
+	SystemFingerprint string                `json:"system_fingerprint"`
+	Choices           []CompleteChunkChoice `json:"choices"`
+	Usage             Usage                 `json:"usage,omitempty"`
+}
+
+type CompletionChunk struct {
+	Id                string                `json:"id"`
+	Object            string                `json:"object"`
+	Created           int64                 `json:"created"`
+	Choices           []CompleteChunkChoice `json:"choices"`
+	Model             string                `json:"model"`
+	SystemFingerprint string                `json:"system_fingerprint"`
+}
+
+type ToolCall struct {
+	ID       string `json:"id"`
+	Type     string `json:"type"`
+	Function struct {
+		Name      string `json:"name"`
+		Arguments string `json:"arguments"`
+	} `json:"function"`
+}
+
+type Model struct {
+	Id      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	OwnedBy string `json:"owned_by"`
+}
+
+type Embedding struct {
+	Object    string    `json:"object"`
+	Embedding []float32 `json:"embedding"`
+	Index     int       `json:"index"`
+}
+
+type ListCompletion struct {
+	Object string  `json:"object"`
+	Data   []Model `json:"data"`
+}
+
+type EmbeddingList struct {
+	Object string      `json:"object"`
+	Data   []Embedding `json:"data"`
+	Model  string      `json:"model"`
+}
+
 func NewError(code int, message string) ErrorResponse {
 	var etype string
 	switch code {
@@ -99,7 +183,31 @@ func NewError(code int, message string) ErrorResponse {
 	return ErrorResponse{Error{Type: etype, Message: message}}
 }
 
+func toolCallId() string {
+	const letterBytes = "abcdefghijklmnopqrstuvwxyz0123456789"
+	b := make([]byte, 8)
+	for i := range b {
+		b[i] = letterBytes[rand.Intn(len(letterBytes))]
+	}
+	return "call_" + strings.ToLower(string(b))
+}
+
 func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
+	toolCalls := make([]ToolCall, len(r.Message.ToolCalls))
+	for i, tc := range r.Message.ToolCalls {
+		toolCalls[i].ID = toolCallId()
+		toolCalls[i].Type = "function"
+		toolCalls[i].Function.Name = tc.Function.Name
+
+		args, err := json.Marshal(tc.Function.Arguments)
+		if err != nil {
+			slog.Error("could not marshall function arguments to json", "error", err)
+			continue
+		}
+
+		toolCalls[i].Function.Arguments = string(args)
+	}
+
 	return ChatCompletion{
 		Id:                id,
 		Object:            "chat.completion",
@@ -108,7 +216,7 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 		SystemFingerprint: "fp_ollama",
 		Choices: []Choice{{
 			Index:   0,
-			Message: Message{Role: r.Message.Role, Content: r.Message.Content},
+			Message: Message{Role: r.Message.Role, Content: r.Message.Content, ToolCalls: toolCalls},
 			FinishReason: func(reason string) *string {
 				if len(reason) > 0 {
 					return &reason
@@ -117,7 +225,6 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 			}(r.DoneReason),
 		}},
 		Usage: Usage{
-			// TODO: ollama returns 0 for prompt eval if the prompt was cached, but openai returns the actual count
 			PromptTokens:     r.PromptEvalCount,
 			CompletionTokens: r.EvalCount,
 			TotalTokens:      r.PromptEvalCount + r.EvalCount,
@@ -145,10 +252,169 @@ func toChunk(id string, r api.ChatResponse) ChatCompletionChunk {
 	}
 }
 
-func fromRequest(r ChatCompletionRequest) api.ChatRequest {
+func toCompletion(id string, r api.GenerateResponse) Completion {
+	return Completion{
+		Id:                id,
+		Object:            "text_completion",
+		Created:           r.CreatedAt.Unix(),
+		Model:             r.Model,
+		SystemFingerprint: "fp_ollama",
+		Choices: []CompleteChunkChoice{{
+			Text:  r.Response,
+			Index: 0,
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
+					return &reason
+				}
+				return nil
+			}(r.DoneReason),
+		}},
+		Usage: Usage{
+			PromptTokens:     r.PromptEvalCount,
+			CompletionTokens: r.EvalCount,
+			TotalTokens:      r.PromptEvalCount + r.EvalCount,
+		},
+	}
+}
+
+func toCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
+	return CompletionChunk{
+		Id:                id,
+		Object:            "text_completion",
+		Created:           time.Now().Unix(),
+		Model:             r.Model,
+		SystemFingerprint: "fp_ollama",
+		Choices: []CompleteChunkChoice{{
+			Text:  r.Response,
+			Index: 0,
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
+					return &reason
+				}
+				return nil
+			}(r.DoneReason),
+		}},
+	}
+}
+
+func toListCompletion(r api.ListResponse) ListCompletion {
+	var data []Model
+	for _, m := range r.Models {
+		data = append(data, Model{
+			Id:      m.Name,
+			Object:  "model",
+			Created: m.ModifiedAt.Unix(),
+			OwnedBy: model.ParseName(m.Name).Namespace,
+		})
+	}
+
+	return ListCompletion{
+		Object: "list",
+		Data:   data,
+	}
+}
+
+func toEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
+	if r.Embeddings != nil {
+		var data []Embedding
+		for i, e := range r.Embeddings {
+			data = append(data, Embedding{
+				Object:    "embedding",
+				Embedding: e,
+				Index:     i,
+			})
+		}
+
+		return EmbeddingList{
+			Object: "list",
+			Data:   data,
+			Model:  model,
+		}
+	}
+
+	return EmbeddingList{}
+}
+
+func toModel(r api.ShowResponse, m string) Model {
+	return Model{
+		Id:      m,
+		Object:  "model",
+		Created: r.ModifiedAt.Unix(),
+		OwnedBy: model.ParseName(m).Namespace,
+	}
+}
+
+func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 	var messages []api.Message
 	for _, msg := range r.Messages {
-		messages = append(messages, api.Message{Role: msg.Role, Content: msg.Content})
+		switch content := msg.Content.(type) {
+		case string:
+			messages = append(messages, api.Message{Role: msg.Role, Content: content})
+		case []any:
+			for _, c := range content {
+				data, ok := c.(map[string]any)
+				if !ok {
+					return nil, fmt.Errorf("invalid message format")
+				}
+				switch data["type"] {
+				case "text":
+					text, ok := data["text"].(string)
+					if !ok {
+						return nil, fmt.Errorf("invalid message format")
+					}
+					messages = append(messages, api.Message{Role: msg.Role, Content: text})
+				case "image_url":
+					var url string
+					if urlMap, ok := data["image_url"].(map[string]any); ok {
+						if url, ok = urlMap["url"].(string); !ok {
+							return nil, fmt.Errorf("invalid message format")
+						}
+					} else {
+						if url, ok = data["image_url"].(string); !ok {
+							return nil, fmt.Errorf("invalid message format")
+						}
+					}
+
+					types := []string{"jpeg", "jpg", "png"}
+					valid := false
+					for _, t := range types {
+						prefix := "data:image/" + t + ";base64,"
+						if strings.HasPrefix(url, prefix) {
+							url = strings.TrimPrefix(url, prefix)
+							valid = true
+							break
+						}
+					}
+
+					if !valid {
+						return nil, fmt.Errorf("invalid image input")
+					}
+
+					img, err := base64.StdEncoding.DecodeString(url)
+					if err != nil {
+						return nil, fmt.Errorf("invalid message format")
+					}
+
+					messages = append(messages, api.Message{Role: msg.Role, Images: []api.ImageData{img}})
+				default:
+					return nil, fmt.Errorf("invalid message format")
+				}
+			}
+		default:
+			if msg.ToolCalls == nil {
+				return nil, fmt.Errorf("invalid message content type: %T", content)
+			}
+
+			toolCalls := make([]api.ToolCall, len(msg.ToolCalls))
+			for i, tc := range msg.ToolCalls {
+				toolCalls[i].Function.Name = tc.Function.Name
+				err := json.Unmarshal([]byte(tc.Function.Arguments), &toolCalls[i].Function.Arguments)
+				if err != nil {
+					return nil, fmt.Errorf("invalid tool call arguments")
+				}
+			}
+			messages = append(messages, api.Message{Role: msg.Role, ToolCalls: toolCalls})
+		}
 	}
 
 	options := make(map[string]interface{})
@@ -156,7 +422,7 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 	switch stop := r.Stop.(type) {
 	case string:
 		options["stop"] = []string{stop}
-	case []interface{}:
+	case []any:
 		var stops []string
 		for _, s := range stop {
 			if str, ok := s.(string); ok {
@@ -199,22 +465,98 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 		format = "json"
 	}
 
-	return api.ChatRequest{
+	return &api.ChatRequest{
 		Model:    r.Model,
 		Messages: messages,
 		Format:   format,
 		Options:  options,
 		Stream:   &r.Stream,
-	}
+		Tools:    r.Tools,
+	}, nil
 }
 
-type writer struct {
-	stream bool
-	id     string
+func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
+	options := make(map[string]any)
+
+	switch stop := r.Stop.(type) {
+	case string:
+		options["stop"] = []string{stop}
+	case []any:
+		var stops []string
+		for _, s := range stop {
+			if str, ok := s.(string); ok {
+				stops = append(stops, str)
+			} else {
+				return api.GenerateRequest{}, fmt.Errorf("invalid type for 'stop' field: %T", s)
+			}
+		}
+		options["stop"] = stops
+	}
+
+	if r.MaxTokens != nil {
+		options["num_predict"] = *r.MaxTokens
+	}
+
+	if r.Temperature != nil {
+		options["temperature"] = *r.Temperature * 2.0
+	} else {
+		options["temperature"] = 1.0
+	}
+
+	if r.Seed != nil {
+		options["seed"] = *r.Seed
+	}
+
+	options["frequency_penalty"] = r.FrequencyPenalty * 2.0
+
+	options["presence_penalty"] = r.PresencePenalty * 2.0
+
+	if r.TopP != 0.0 {
+		options["top_p"] = r.TopP
+	} else {
+		options["top_p"] = 1.0
+	}
+
+	return api.GenerateRequest{
+		Model:   r.Model,
+		Prompt:  r.Prompt,
+		Options: options,
+		Stream:  &r.Stream,
+		Suffix:  r.Suffix,
+	}, nil
+}
+
+type BaseWriter struct {
 	gin.ResponseWriter
 }
 
-func (w *writer) writeError(code int, data []byte) (int, error) {
+type ChatWriter struct {
+	stream bool
+	id     string
+	BaseWriter
+}
+
+type CompleteWriter struct {
+	stream bool
+	id     string
+	BaseWriter
+}
+
+type ListWriter struct {
+	BaseWriter
+}
+
+type RetrieveWriter struct {
+	BaseWriter
+	model string
+}
+
+type EmbedWriter struct {
+	BaseWriter
+	model string
+}
+
+func (w *BaseWriter) writeError(code int, data []byte) (int, error) {
 	var serr api.StatusError
 	err := json.Unmarshal(data, &serr)
 	if err != nil {
@@ -230,7 +572,7 @@ func (w *writer) writeError(code int, data []byte) (int, error) {
 	return len(data), nil
 }
 
-func (w *writer) writeResponse(data []byte) (int, error) {
+func (w *ChatWriter) writeResponse(data []byte) (int, error) {
 	var chatResponse api.ChatResponse
 	err := json.Unmarshal(data, &chatResponse)
 	if err != nil {
@@ -270,7 +612,7 @@ func (w *writer) writeResponse(data []byte) (int, error) {
 	return len(data), nil
 }
 
-func (w *writer) Write(data []byte) (int, error) {
+func (w *ChatWriter) Write(data []byte) (int, error) {
 	code := w.ResponseWriter.Status()
 	if code != http.StatusOK {
 		return w.writeError(code, data)
@@ -279,7 +621,244 @@ func (w *writer) Write(data []byte) (int, error) {
 	return w.writeResponse(data)
 }
 
-func Middleware() gin.HandlerFunc {
+func (w *CompleteWriter) writeResponse(data []byte) (int, error) {
+	var generateResponse api.GenerateResponse
+	err := json.Unmarshal(data, &generateResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// completion chunk
+	if w.stream {
+		d, err := json.Marshal(toCompleteChunk(w.id, generateResponse))
+		if err != nil {
+			return 0, err
+		}
+
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+		if err != nil {
+			return 0, err
+		}
+
+		if generateResponse.Done {
+			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		return len(data), nil
+	}
+
+	// completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toCompletion(w.id, generateResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *CompleteWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *ListWriter) writeResponse(data []byte) (int, error) {
+	var listResponse api.ListResponse
+	err := json.Unmarshal(data, &listResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toListCompletion(listResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ListWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *RetrieveWriter) writeResponse(data []byte) (int, error) {
+	var showResponse api.ShowResponse
+	err := json.Unmarshal(data, &showResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// retrieve completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toModel(showResponse, w.model))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *RetrieveWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *EmbedWriter) writeResponse(data []byte) (int, error) {
+	var embedResponse api.EmbedResponse
+	err := json.Unmarshal(data, &embedResponse)
+
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toEmbeddingList(w.model, embedResponse))
+
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *EmbedWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func ListMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		w := &ListWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func RetrieveMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		// response writer
+		w := &RetrieveWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      c.Param("model"),
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func CompletionsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req CompletionRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		genReq, err := fromCompleteRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &CompleteWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         fmt.Sprintf("cmpl-%d", rand.Intn(999)),
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}
+
+func EmbeddingsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req EmbedRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Input == "" {
+			req.Input = []string{""}
+		}
+
+		if req.Input == nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		if v, ok := req.Input.([]any); ok && len(v) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &EmbedWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      req.Model,
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func ChatMiddleware() gin.HandlerFunc {
 	return func(c *gin.Context) {
 		var req ChatCompletionRequest
 		err := c.ShouldBindJSON(&req)
@@ -294,17 +873,24 @@ func Middleware() gin.HandlerFunc {
 		}
 
 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(fromRequest(req)); err != nil {
+
+		chatReq, err := fromChatRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
 			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
 			return
 		}
 
 		c.Request.Body = io.NopCloser(&b)
 
-		w := &writer{
-			ResponseWriter: c.Writer,
-			stream:         req.Stream,
-			id:             fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
+		w := &ChatWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
 		}
 
 		c.Writer = w

openai/openai_test.go

@@ -0,0 +1,496 @@
+package openai
+
+import (
+	"bytes"
+	"encoding/base64"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/ollama/ollama/api"
+	"github.com/stretchr/testify/assert"
+)
+
+const prefix = `data:image/jpeg;base64,`
+const image = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
+const imageURL = prefix + image
+
+func prepareRequest(req *http.Request, body any) {
+	bodyBytes, _ := json.Marshal(body)
+	req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+	req.Header.Set("Content-Type", "application/json")
+}
+
+func captureRequestMiddleware(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		err := json.Unmarshal(bodyBytes, capturedRequest)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, "failed to unmarshal request")
+		}
+		c.Next()
+	}
+}
+
+func TestChatMiddleware(t *testing.T) {
+	type testCase struct {
+		Name     string
+		Setup    func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder)
+	}
+
+	var capturedRequest *api.ChatRequest
+
+	testCases := []testCase{
+		{
+			Name: "chat handler",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := ChatCompletionRequest{
+					Model:    "test-model",
+					Messages: []Message{{Role: "user", Content: "Hello"}},
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusOK {
+					t.Fatalf("expected 200, got %d", resp.Code)
+				}
+
+				if req.Messages[0].Role != "user" {
+					t.Fatalf("expected 'user', got %s", req.Messages[0].Role)
+				}
+
+				if req.Messages[0].Content != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Messages[0].Content)
+				}
+			},
+		},
+		{
+			Name: "chat handler with image content",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := ChatCompletionRequest{
+					Model: "test-model",
+					Messages: []Message{
+						{
+							Role: "user", Content: []map[string]any{
+								{"type": "text", "text": "Hello"},
+								{"type": "image_url", "image_url": map[string]string{"url": imageURL}},
+							},
+						},
+					},
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusOK {
+					t.Fatalf("expected 200, got %d", resp.Code)
+				}
+
+				if req.Messages[0].Role != "user" {
+					t.Fatalf("expected 'user', got %s", req.Messages[0].Role)
+				}
+
+				if req.Messages[0].Content != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Messages[0].Content)
+				}
+
+				img, _ := base64.StdEncoding.DecodeString(imageURL[len(prefix):])
+
+				if req.Messages[1].Role != "user" {
+					t.Fatalf("expected 'user', got %s", req.Messages[1].Role)
+				}
+
+				if !bytes.Equal(req.Messages[1].Images[0], img) {
+					t.Fatalf("expected image encoding, got %s", req.Messages[1].Images[0])
+				}
+			},
+		},
+		{
+			Name: "chat handler with tools",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := ChatCompletionRequest{
+					Model: "test-model",
+					Messages: []Message{
+						{Role: "user", Content: "What's the weather like in Paris Today?"},
+						{Role: "assistant", ToolCalls: []ToolCall{{
+							ID:   "id",
+							Type: "function",
+							Function: struct {
+								Name      string `json:"name"`
+								Arguments string `json:"arguments"`
+							}{
+								Name:      "get_current_weather",
+								Arguments: "{\"location\": \"Paris, France\", \"format\": \"celsius\"}",
+							},
+						}}},
+					},
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != 200 {
+					t.Fatalf("expected 200, got %d", resp.Code)
+				}
+
+				if req.Messages[0].Content != "What's the weather like in Paris Today?" {
+					t.Fatalf("expected What's the weather like in Paris Today?, got %s", req.Messages[0].Content)
+				}
+
+				if req.Messages[1].ToolCalls[0].Function.Arguments["location"] != "Paris, France" {
+					t.Fatalf("expected 'Paris, France', got %v", req.Messages[1].ToolCalls[0].Function.Arguments["location"])
+				}
+
+				if req.Messages[1].ToolCalls[0].Function.Arguments["format"] != "celsius" {
+					t.Fatalf("expected celsius, got %v", req.Messages[1].ToolCalls[0].Function.Arguments["format"])
+				}
+			},
+		},
+		{
+			Name: "chat handler error forwarding",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := ChatCompletionRequest{
+					Model:    "test-model",
+					Messages: []Message{{Role: "user", Content: 2}},
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusBadRequest {
+					t.Fatalf("expected 400, got %d", resp.Code)
+				}
+
+				if !strings.Contains(resp.Body.String(), "invalid message content type") {
+					t.Fatalf("error was not forwarded")
+				}
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ChatMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/chat", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.Name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/chat", nil)
+
+			tc.Setup(t, req)
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			tc.Expected(t, capturedRequest, resp)
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestCompletionsMiddleware(t *testing.T) {
+	type testCase struct {
+		Name     string
+		Setup    func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, req *api.GenerateRequest, resp *httptest.ResponseRecorder)
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	testCases := []testCase{
+		{
+			Name: "completions handler",
+			Setup: func(t *testing.T, req *http.Request) {
+				temp := float32(0.8)
+				body := CompletionRequest{
+					Model:       "test-model",
+					Prompt:      "Hello",
+					Temperature: &temp,
+					Stop:        []string{"\n", "stop"},
+					Suffix:      "suffix",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.GenerateRequest, resp *httptest.ResponseRecorder) {
+				if req.Prompt != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Prompt)
+				}
+
+				if req.Options["temperature"] != 1.6 {
+					t.Fatalf("expected 1.6, got %f", req.Options["temperature"])
+				}
+
+				stopTokens, ok := req.Options["stop"].([]any)
+
+				if !ok {
+					t.Fatalf("expected stop tokens to be a list")
+				}
+
+				if stopTokens[0] != "\n" || stopTokens[1] != "stop" {
+					t.Fatalf("expected ['\\n', 'stop'], got %v", stopTokens)
+				}
+
+				if req.Suffix != "suffix" {
+					t.Fatalf("expected 'suffix', got %s", req.Suffix)
+				}
+			},
+		},
+		{
+			Name: "completions handler error forwarding",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := CompletionRequest{
+					Model:       "test-model",
+					Prompt:      "Hello",
+					Temperature: nil,
+					Stop:        []int{1, 2},
+					Suffix:      "suffix",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.GenerateRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusBadRequest {
+					t.Fatalf("expected 400, got %d", resp.Code)
+				}
+
+				if !strings.Contains(resp.Body.String(), "invalid type for 'stop' field") {
+					t.Fatalf("error was not forwarded")
+				}
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(CompletionsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.Name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", nil)
+
+			tc.Setup(t, req)
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			tc.Expected(t, capturedRequest, resp)
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestEmbeddingsMiddleware(t *testing.T) {
+	type testCase struct {
+		Name     string
+		Setup    func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder)
+	}
+
+	var capturedRequest *api.EmbedRequest
+
+	testCases := []testCase{
+		{
+			Name: "embed handler single input",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := EmbedRequest{
+					Input: "Hello",
+					Model: "test-model",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder) {
+				if req.Input != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Input)
+				}
+
+				if req.Model != "test-model" {
+					t.Fatalf("expected 'test-model', got %s", req.Model)
+				}
+			},
+		},
+		{
+			Name: "embed handler batch input",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := EmbedRequest{
+					Input: []string{"Hello", "World"},
+					Model: "test-model",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder) {
+				input, ok := req.Input.([]any)
+
+				if !ok {
+					t.Fatalf("expected input to be a list")
+				}
+
+				if input[0].(string) != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", input[0])
+				}
+
+				if input[1].(string) != "World" {
+					t.Fatalf("expected 'World', got %s", input[1])
+				}
+
+				if req.Model != "test-model" {
+					t.Fatalf("expected 'test-model', got %s", req.Model)
+				}
+			},
+		},
+		{
+			Name: "embed handler error forwarding",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := EmbedRequest{
+					Model: "test-model",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusBadRequest {
+					t.Fatalf("expected 400, got %d", resp.Code)
+				}
+
+				if !strings.Contains(resp.Body.String(), "invalid input") {
+					t.Fatalf("error was not forwarded")
+				}
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(EmbeddingsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/embed", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.Name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/embed", nil)
+
+			tc.Setup(t, req)
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			tc.Expected(t, capturedRequest, resp)
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestMiddlewareResponses(t *testing.T) {
+	type testCase struct {
+		Name     string
+		Method   string
+		Path     string
+		TestPath string
+		Handler  func() gin.HandlerFunc
+		Endpoint func(c *gin.Context)
+		Setup    func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, resp *httptest.ResponseRecorder)
+	}
+
+	testCases := []testCase{
+		{
+			Name:     "list handler",
+			Method:   http.MethodGet,
+			Path:     "/api/tags",
+			TestPath: "/api/tags",
+			Handler:  ListMiddleware,
+			Endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ListResponse{
+					Models: []api.ListModelResponse{
+						{
+							Name: "Test Model",
+						},
+					},
+				})
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				var listResp ListCompletion
+				if err := json.NewDecoder(resp.Body).Decode(&listResp); err != nil {
+					t.Fatal(err)
+				}
+
+				if listResp.Object != "list" {
+					t.Fatalf("expected list, got %s", listResp.Object)
+				}
+
+				if len(listResp.Data) != 1 {
+					t.Fatalf("expected 1, got %d", len(listResp.Data))
+				}
+
+				if listResp.Data[0].Id != "Test Model" {
+					t.Fatalf("expected Test Model, got %s", listResp.Data[0].Id)
+				}
+			},
+		},
+		{
+			Name:     "retrieve model",
+			Method:   http.MethodGet,
+			Path:     "/api/show/:model",
+			TestPath: "/api/show/test-model",
+			Handler:  RetrieveMiddleware,
+			Endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ShowResponse{
+					ModifiedAt: time.Date(2024, 6, 17, 13, 45, 0, 0, time.UTC),
+				})
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				var retrieveResp Model
+				if err := json.NewDecoder(resp.Body).Decode(&retrieveResp); err != nil {
+					t.Fatal(err)
+				}
+
+				if retrieveResp.Object != "model" {
+					t.Fatalf("Expected object to be model, got %s", retrieveResp.Object)
+				}
+
+				if retrieveResp.Id != "test-model" {
+					t.Fatalf("Expected id to be test-model, got %s", retrieveResp.Id)
+				}
+			},
+		},
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+
+	for _, tc := range testCases {
+		t.Run(tc.Name, func(t *testing.T) {
+			router = gin.New()
+			router.Use(tc.Handler())
+			router.Handle(tc.Method, tc.Path, tc.Endpoint)
+			req, _ := http.NewRequest(tc.Method, tc.TestPath, nil)
+
+			if tc.Setup != nil {
+				tc.Setup(t, req)
+			}
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			assert.Equal(t, http.StatusOK, resp.Code)
+
+			tc.Expected(t, resp)
+		})
+	}
+}

parser/parser.go

@@ -124,7 +124,7 @@ func ParseFile(r io.Reader) (*File, error) {
 			case stateComment, stateNil:
 				// pass
 			case stateValue:
-				s, ok := unquote(b.String())
+				s, ok := unquote(strings.TrimSpace(b.String()))
 				if !ok || isSpace(r) {
 					if _, err := b.WriteRune(r); err != nil {
 						return nil, err
@@ -158,7 +158,7 @@ func ParseFile(r io.Reader) (*File, error) {
 	case stateComment, stateNil:
 		// pass; nothing to flush
 	case stateValue:
-		s, ok := unquote(b.String())
+		s, ok := unquote(strings.TrimSpace(b.String()))
 		if !ok {
 			return nil, io.ErrUnexpectedEOF
 		}

parser/parser_test.go

@@ -22,7 +22,13 @@ ADAPTER adapter1
 LICENSE MIT
 PARAMETER param1 value1
 PARAMETER param2 value2
-TEMPLATE template1
+TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
+
+{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ .Response }}<|eot_id|>"""    
 `
 
 	reader := strings.NewReader(input)
@@ -36,7 +42,40 @@ TEMPLATE template1
 		{Name: "license", Args: "MIT"},
 		{Name: "param1", Args: "value1"},
 		{Name: "param2", Args: "value2"},
-		{Name: "template", Args: "template1"},
+		{Name: "template", Args: "{{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>"},
+	}
+
+	assert.Equal(t, expectedCommands, modelfile.Commands)
+}
+
+func TestParseFileTrimSpace(t *testing.T) {
+	input := `
+FROM "     model 1"
+ADAPTER      adapter3
+LICENSE "MIT       "
+PARAMETER param1        value1
+PARAMETER param2    value2
+TEMPLATE """   {{ if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
+
+{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ .Response }}<|eot_id|>   """    
+`
+
+	reader := strings.NewReader(input)
+
+	modelfile, err := ParseFile(reader)
+	require.NoError(t, err)
+
+	expectedCommands := []Command{
+		{Name: "model", Args: "     model 1"},
+		{Name: "adapter", Args: "adapter3"},
+		{Name: "license", Args: "MIT       "},
+		{Name: "param1", Args: "value1"},
+		{Name: "param2", Args: "value2"},
+		{Name: "template", Args: "   {{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>   "},
 	}
 
 	assert.Equal(t, expectedCommands, modelfile.Commands)
@@ -48,6 +87,26 @@ func TestParseFileFrom(t *testing.T) {
 		expected []Command
 		err      error
 	}{
+		{
+			"FROM \"FOO  BAR  \"",
+			[]Command{{Name: "model", Args: "FOO  BAR  "}},
+			nil,
+		},
+		{
+			"FROM \"FOO BAR\"\nPARAMETER param1 value1",
+			[]Command{{Name: "model", Args: "FOO BAR"}, {Name: "param1", Args: "value1"}},
+			nil,
+		},
+		{
+			"FROM     FOOO BAR    ",
+			[]Command{{Name: "model", Args: "FOOO BAR"}},
+			nil,
+		},
+		{
+			"FROM /what/is/the path ",
+			[]Command{{Name: "model", Args: "/what/is/the path"}},
+			nil,
+		},
 		{
 			"FROM foo",
 			[]Command{{Name: "model", Args: "foo"}},
@@ -86,6 +145,11 @@ func TestParseFileFrom(t *testing.T) {
 			[]Command{{Name: "param1", Args: "value1"}, {Name: "model", Args: "foo"}},
 			nil,
 		},
+		{
+			"PARAMETER what the \nFROM lemons make lemonade ",
+			[]Command{{Name: "what", Args: "the"}, {Name: "model", Args: "lemons make lemonade"}},
+			nil,
+		},
 	}
 
 	for _, c := range cases {
@@ -399,7 +463,7 @@ func TestParseFileParameters(t *testing.T) {
 		"mirostat_eta 1.0":             {"mirostat_eta", "1.0"},
 		"penalize_newline true":        {"penalize_newline", "true"},
 		"stop ### User:":               {"stop", "### User:"},
-		"stop ### User: ":              {"stop", "### User: "},
+		"stop ### User: ":              {"stop", "### User:"},
 		"stop \"### User:\"":           {"stop", "### User:"},
 		"stop \"### User: \"":          {"stop", "### User: "},
 		"stop \"\"\"### User:\"\"\"":   {"stop", "### User:"},

scripts/build_windows.ps1

@@ -103,19 +103,22 @@ function buildApp() {
 function gatherDependencies() {
     write-host "Gathering runtime dependencies"
     cd "${script:SRC_DIR}"
-    md "${script:DEPS_DIR}" -ea 0 > $null
+    md "${script:DEPS_DIR}\ollama_runners" -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}\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\"
+    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\"
+    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 "${script:SRC_DIR}\app\ollama_welcome.ps1" "${script:SRC_DIR}\dist\"
     if ("${env:KEY_CONTAINER}") {
         write-host "about to sign"
-        foreach ($file in (get-childitem "${script:DEPS_DIR}/cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
+        foreach ($file in (get-childitem "${script:DEPS_DIR}\cuda\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

scripts/install.sh

@@ -279,7 +279,7 @@ if ! check_gpu nvidia-smi || [ -z "$(nvidia-smi | grep -o "CUDA Version: [0-9]*\
     case $OS_NAME in
         centos|rhel) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -d '.' -f 1) ;;
         rocky) install_cuda_driver_yum 'rhel' $(echo $OS_VERSION | cut -c1) ;;
-        fedora) [ $OS_VERSION -lt '37' ] && install_cuda_driver_yum $OS_NAME $OS_VERSION || install_cuda_driver_yum $OS_NAME '37';;
+        fedora) [ $OS_VERSION -lt '39' ] && install_cuda_driver_yum $OS_NAME $OS_VERSION || install_cuda_driver_yum $OS_NAME '39';;
         amzn) install_cuda_driver_yum 'fedora' '37' ;;
         debian) install_cuda_driver_apt $OS_NAME $OS_VERSION ;;
         ubuntu) install_cuda_driver_apt $OS_NAME $(echo $OS_VERSION | sed 's/\.//') ;;

scripts/rh_linux_deps.sh

@@ -6,10 +6,21 @@ set -ex
 MACHINE=$(uname -m)
 
 if grep -i "centos" /etc/system-release >/dev/null; then
+    # As of 7/1/2024 mirrorlist.centos.org has been taken offline, so adjust accordingly
+    sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo
+    sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo
+    sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo
+
     # Centos 7 derivatives have too old of a git version to run our generate script
     # uninstall and ignore failures
     yum remove -y git
     yum -y install epel-release centos-release-scl
+
+    # The release packages reinstate the mirrors, undo that again
+    sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo
+    sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo
+    sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo
+
     yum -y install dnf
     if [ "${MACHINE}" = "x86_64" ]; then
         yum -y install https://repo.ius.io/ius-release-el7.rpm

server/images.go

@@ -28,11 +28,27 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
 
+var (
+	errCapabilities         = errors.New("does not support")
+	errCapabilityCompletion = errors.New("completion")
+	errCapabilityTools      = errors.New("tools")
+	errCapabilityInsert     = errors.New("insert")
+)
+
+type Capability string
+
+const (
+	CapabilityCompletion = Capability("completion")
+	CapabilityTools      = Capability("tools")
+	CapabilityInsert     = Capability("insert")
+)
+
 type registryOptions struct {
 	Insecure bool
 	Username string
@@ -48,16 +64,59 @@ type Model struct {
 	ParentModel    string
 	AdapterPaths   []string
 	ProjectorPaths []string
-	Template       string
 	System         string
 	License        []string
 	Digest         string
 	Options        map[string]interface{}
 	Messages       []Message
+
+	Template *template.Template
 }
 
-func (m *Model) IsEmbedding() bool {
-	return slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert")
+// CheckCapabilities checks if the model has the specified capabilities returning an error describing
+// any missing or unknown capabilities
+func (m *Model) CheckCapabilities(caps ...Capability) error {
+	var errs []error
+	for _, cap := range caps {
+		switch cap {
+		case CapabilityCompletion:
+			f, err := os.Open(m.ModelPath)
+			if err != nil {
+				slog.Error("couldn't open model file", "error", err)
+				continue
+			}
+			defer f.Close()
+
+			// TODO(mxyng): decode the GGML into model to avoid doing this multiple times
+			ggml, _, err := llm.DecodeGGML(f, 0)
+			if err != nil {
+				slog.Error("couldn't decode ggml", "error", err)
+				continue
+			}
+
+			if _, ok := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]; ok {
+				errs = append(errs, errCapabilityCompletion)
+			}
+		case CapabilityTools:
+			if !slices.Contains(m.Template.Vars(), "tools") {
+				errs = append(errs, errCapabilityTools)
+			}
+		case CapabilityInsert:
+			vars := m.Template.Vars()
+			if !slices.Contains(vars, "suffix") {
+				errs = append(errs, errCapabilityInsert)
+			}
+		default:
+			slog.Error("unknown capability", "capability", cap)
+			return fmt.Errorf("unknown capability: %s", cap)
+		}
+	}
+
+	if err := errors.Join(errs...); err != nil {
+		return fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
+	}
+
+	return nil
 }
 
 func (m *Model) String() string {
@@ -82,10 +141,10 @@ func (m *Model) String() string {
 		})
 	}
 
-	if m.Template != "" {
+	if m.Template != nil {
 		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "template",
-			Args: m.Template,
+			Args: m.Template.String(),
 		})
 	}
 
@@ -135,13 +194,6 @@ type Message struct {
 	Content string `json:"content"`
 }
 
-type ManifestV2 struct {
-	SchemaVersion int      `json:"schemaVersion"`
-	MediaType     string   `json:"mediaType"`
-	Config        *Layer   `json:"config"`
-	Layers        []*Layer `json:"layers"`
-}
-
 type ConfigV2 struct {
 	ModelFormat   string   `json:"model_format"`
 	ModelFamily   string   `json:"model_family"`
@@ -160,7 +212,7 @@ type RootFS struct {
 	DiffIDs []string `json:"diff_ids"`
 }
 
-func GetManifest(mp ModelPath) (*ManifestV2, string, error) {
+func GetManifest(mp ModelPath) (*Manifest, string, error) {
 	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return nil, "", err
@@ -170,7 +222,7 @@ func GetManifest(mp ModelPath) (*ManifestV2, string, error) {
 		return nil, "", err
 	}
 
-	var manifest *ManifestV2
+	var manifest *Manifest
 
 	bts, err := os.ReadFile(fp)
 	if err != nil {
@@ -198,8 +250,7 @@ func GetModel(name string) (*Model, error) {
 		Name:      mp.GetFullTagname(),
 		ShortName: mp.GetShortTagname(),
 		Digest:    digest,
-		Template:  "{{ .Prompt }}",
-		License:   []string{},
+		Template:  template.DefaultTemplate,
 	}
 
 	filename, err := GetBlobsPath(manifest.Config.Digest)
@@ -235,13 +286,17 @@ func GetModel(name string) (*Model, error) {
 			model.AdapterPaths = append(model.AdapterPaths, filename)
 		case "application/vnd.ollama.image.projector":
 			model.ProjectorPaths = append(model.ProjectorPaths, filename)
-		case "application/vnd.ollama.image.template":
+		case "application/vnd.ollama.image.prompt",
+			"application/vnd.ollama.image.template":
 			bts, err := os.ReadFile(filename)
 			if err != nil {
 				return nil, err
 			}
 
-			model.Template = string(bts)
+			model.Template, err = template.Parse(string(bts))
+			if err != nil {
+				return nil, err
+			}
 		case "application/vnd.ollama.image.system":
 			bts, err := os.ReadFile(filename)
 			if err != nil {
@@ -249,13 +304,6 @@ func GetModel(name string) (*Model, error) {
 			}
 
 			model.System = string(bts)
-		case "application/vnd.ollama.image.prompt":
-			bts, err := os.ReadFile(filename)
-			if err != nil {
-				return nil, err
-			}
-
-			model.Template = string(bts)
 		case "application/vnd.ollama.image.params":
 			params, err := os.Open(filename)
 			if err != nil {
@@ -414,17 +462,22 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 							return err
 						}
 
-						layers, err := parseFromFile(ctx, temp, "", fn)
+						layer, err := NewLayer(temp, baseLayer.MediaType)
 						if err != nil {
 							return err
 						}
 
-						if len(layers) != 1 {
-							return errors.New("quantization failed")
+						if _, err := temp.Seek(0, io.SeekStart); err != nil {
+							return err
 						}
 
-						baseLayer.Layer = layers[0].Layer
-						baseLayer.GGML = layers[0].GGML
+						ggml, _, err := llm.DecodeGGML(temp, 0)
+						if err != nil {
+							return err
+						}
+
+						baseLayer.Layer = layer
+						baseLayer.GGML = ggml
 					}
 				}
 
@@ -439,6 +492,12 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 				layers = append(layers, baseLayer.Layer)
 			}
 		case "license", "template", "system":
+			if c.Name == "template" {
+				if _, err := template.Parse(c.Args); err != nil {
+					return fmt.Errorf("%w: %s", errBadTemplate, err)
+				}
+			}
+
 			if c.Name != "license" {
 				// replace
 				layers = slices.DeleteFunc(layers, func(layer *Layer) bool {
@@ -817,7 +876,7 @@ 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 manifest *Manifest
 	var err error
 	var noprune string
 
@@ -924,7 +983,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	return nil
 }
 
-func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*ManifestV2, error) {
+func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
 	requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
 
 	headers := make(http.Header)
@@ -935,7 +994,7 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptio
 	}
 	defer resp.Body.Close()
 
-	var m *ManifestV2
+	var m *Manifest
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
 	}

server/manifest.go

@@ -14,7 +14,10 @@ import (
 )
 
 type Manifest struct {
-	ManifestV2
+	SchemaVersion int      `json:"schemaVersion"`
+	MediaType     string   `json:"mediaType"`
+	Config        *Layer   `json:"config"`
+	Layers        []*Layer `json:"layers"`
 
 	filepath string
 	fi       os.FileInfo
@@ -66,7 +69,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 
 	p := filepath.Join(manifests, n.Filepath())
 
-	var m ManifestV2
+	var m Manifest
 	f, err := os.Open(p)
 	if err != nil {
 		return nil, err
@@ -83,12 +86,11 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, err
 	}
 
-	return &Manifest{
-		ManifestV2: m,
-		filepath:   p,
-		fi:         fi,
-		digest:     fmt.Sprintf("%x", sha256sum.Sum(nil)),
-	}, nil
+	m.filepath = p
+	m.fi = fi
+	m.digest = fmt.Sprintf("%x", sha256sum.Sum(nil))
+
+	return &m, nil
 }
 
 func WriteManifest(name model.Name, config *Layer, layers []*Layer) error {
@@ -108,7 +110,7 @@ func WriteManifest(name model.Name, config *Layer, layers []*Layer) error {
 	}
 	defer f.Close()
 
-	m := ManifestV2{
+	m := Manifest{
 		SchemaVersion: 2,
 		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
 		Config:        config,

server/manifest_test.go

@@ -25,7 +25,7 @@ func createManifest(t *testing.T, path, name string) {
 	}
 	defer f.Close()
 
-	if err := json.NewEncoder(f).Encode(ManifestV2{}); err != nil {
+	if err := json.NewEncoder(f).Encode(Manifest{}); err != nil {
 		t.Fatal(err)
 	}
 }

server/model.go

@@ -4,6 +4,7 @@ import (
 	"archive/zip"
 	"bytes"
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -11,11 +12,14 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
+	"slices"
+	"strings"
+	"text/template/parse"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/llm"
-	"github.com/ollama/ollama/templates"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -63,7 +67,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 			}
 			defer blob.Close()
 
-			ggml, _, err := llm.DecodeGGML(blob)
+			ggml, _, err := llm.DecodeGGML(blob, 0)
 			if err != nil {
 				return nil, err
 			}
@@ -77,62 +81,79 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	return layers, nil
 }
 
-func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+func extractFromZipFile(p string, file *os.File, fn func(api.ProgressResponse)) error {
 	stat, err := file.Stat()
 	if err != nil {
-		return nil, err
+		return err
 	}
 
 	r, err := zip.NewReader(file, stat.Size())
 	if err != nil {
-		return nil, err
+		return err
 	}
 
-	tempdir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
-	if err != nil {
-		return nil, err
-	}
-	defer os.RemoveAll(tempdir)
-
 	fn(api.ProgressResponse{Status: "unpacking model metadata"})
 	for _, f := range r.File {
+		if !filepath.IsLocal(f.Name) {
+			return fmt.Errorf("%w: %s", zip.ErrInsecurePath, f.Name)
+		}
+
+		n := filepath.Join(p, f.Name)
+		if err := os.MkdirAll(filepath.Dir(n), 0o750); err != nil {
+			return err
+		}
+
 		// TODO(mxyng): this should not write out all files to disk
-		outfile, err := os.Create(filepath.Join(tempdir, f.Name))
+		outfile, err := os.Create(n)
 		if err != nil {
-			return nil, err
+			return err
 		}
 		defer outfile.Close()
 
 		infile, err := f.Open()
 		if err != nil {
-			return nil, err
+			return err
 		}
 		defer infile.Close()
 
 		if _, err = io.Copy(outfile, infile); err != nil {
-			return nil, err
+			return err
 		}
 
 		if err := outfile.Close(); err != nil {
-			return nil, err
+			return err
 		}
 
 		if err := infile.Close(); err != nil {
-			return nil, err
+			return err
 		}
 	}
 
-	mf, err := convert.GetModelFormat(tempdir)
+	return nil
+}
+
+func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+	tempDir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
+	if err != nil {
+		return nil, err
+	}
+	defer os.RemoveAll(tempDir)
+
+	if err := extractFromZipFile(tempDir, file, fn); err != nil {
+		return nil, err
+	}
+
+	mf, err := convert.GetModelFormat(tempDir)
 	if err != nil {
 		return nil, err
 	}
 
-	params, err := mf.GetParams(tempdir)
+	params, err := mf.GetParams(tempDir)
 	if err != nil {
 		return nil, err
 	}
 
-	mArch, err := mf.GetModelArch("", tempdir, params)
+	mArch, err := mf.GetModelArch("", tempDir, params)
 	if err != nil {
 		return nil, err
 	}
@@ -150,7 +171,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a
 
 	// TODO(mxyng): this should write directly into a layer
 	// e.g. NewLayer(arch.Reader(), "application/vnd.ollama.image.model")
-	temp, err := os.CreateTemp(tempdir, "fp16")
+	temp, err := os.CreateTemp(tempDir, "fp16")
 	if err != nil {
 		return nil, err
 	}
@@ -176,7 +197,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a
 	}
 	defer bin.Close()
 
-	ggml, _, err := llm.DecodeGGML(bin)
+	ggml, _, err := llm.DecodeGGML(bin, 0)
 	if err != nil {
 		return nil, err
 	}
@@ -210,7 +231,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 
 	var offset int64
 	for offset < stat.Size() {
-		ggml, n, err := llm.DecodeGGML(file)
+		ggml, n, err := llm.DecodeGGML(file, 0)
 		if errors.Is(err, io.EOF) {
 			break
 		} else if err != nil {
@@ -239,7 +260,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 	for _, layer := range layers {
 		if s := layer.GGML.KV().ChatTemplate(); s != "" {
-			if t, err := templates.NamedTemplate(s); err != nil {
+			if t, err := template.Named(s); err != nil {
 				slog.Debug("template detection", "error", err)
 			} else {
 				tmpl, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
@@ -272,3 +293,113 @@ func detectContentType(r io.Reader) (string, error) {
 
 	return "unknown", nil
 }
+
+// parseToolCalls attempts to parse a JSON string into a slice of ToolCalls.
+// mxyng: this only really works if the input contains tool calls in some JSON format
+func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
+	// create a subtree from the node that ranges over .ToolCalls
+	tmpl := m.Template.Subtree(func(n parse.Node) bool {
+		if t, ok := n.(*parse.RangeNode); ok {
+			return slices.Contains(template.Identifiers(t.Pipe), "ToolCalls")
+		}
+
+		return false
+	})
+
+	if tmpl == nil {
+		return nil, false
+	}
+
+	var b bytes.Buffer
+	if err := tmpl.Execute(&b, map[string][]api.ToolCall{
+		"ToolCalls": {
+			{
+				Function: api.ToolCallFunction{
+					Name: "@@name@@",
+					Arguments: api.ToolCallFunctionArguments{
+						"@@argument@@": 1,
+					},
+				},
+			},
+		},
+	}); err != nil {
+		return nil, false
+	}
+
+	var kv map[string]any
+	// execute the subtree with placeholders to identify the keys
+	// trim any commands that might exist in the template
+	if err := json.Unmarshal(bytes.TrimSuffix(b.Bytes(), []byte(",")), &kv); err != nil {
+		return nil, false
+	}
+
+	// find the keys that correspond to the name and arguments fields
+	var name, arguments string
+	for k, v := range kv {
+		switch v.(type) {
+		case string:
+			name = k
+		case map[string]any:
+			arguments = k
+		}
+	}
+
+	if name == "" || arguments == "" {
+		return nil, false
+	}
+
+	var objs []map[string]any
+	for offset := 0; offset < len(s); {
+		var obj map[string]any
+		decoder := json.NewDecoder(strings.NewReader(s[offset:]))
+		if err := decoder.Decode(&obj); errors.Is(err, io.EOF) || errors.Is(err, io.ErrUnexpectedEOF) {
+			break
+		} else if syntax := &(json.SyntaxError{}); errors.As(err, &syntax) {
+			// skip over any syntax errors
+			offset += int(syntax.Offset)
+		} else if unmarshalType := &(json.UnmarshalTypeError{}); errors.As(err, &unmarshalType) {
+			// skip over any unmarshalable types
+			offset += int(unmarshalType.Offset)
+		} else if err != nil {
+			slog.Error("parseToolCalls", "error", err)
+			return nil, false
+		} else {
+			offset += int(decoder.InputOffset())
+
+			// collect all nested objects
+			var collect func(any) []map[string]any
+			collect = func(obj any) (all []map[string]any) {
+				switch o := obj.(type) {
+				case map[string]any:
+					all = append(all, o)
+					for _, v := range o {
+						all = append(all, collect(v)...)
+					}
+				case []any:
+					for _, v := range o {
+						all = append(all, collect(v)...)
+					}
+				}
+
+				return all
+			}
+			objs = append(objs, collect(obj)...)
+		}
+	}
+
+	var toolCalls []api.ToolCall
+	for _, kv := range objs {
+		n, nok := kv[name].(string)
+		a, aok := kv[arguments].(map[string]any)
+		if nok && aok {
+			toolCalls = append(toolCalls, api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      n,
+					Arguments: a,
+				},
+			})
+		}
+	}
+
+	return toolCalls, len(toolCalls) > 0
+}

server/model_test.go

@@ -0,0 +1,236 @@
+package server
+
+import (
+	"archive/zip"
+	"bytes"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/template"
+)
+
+func createZipFile(t *testing.T, name string) *os.File {
+	t.Helper()
+
+	f, err := os.CreateTemp(t.TempDir(), "")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	zf := zip.NewWriter(f)
+	defer zf.Close()
+
+	zh, err := zf.CreateHeader(&zip.FileHeader{Name: name})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if _, err := io.Copy(zh, bytes.NewReader([]byte(""))); err != nil {
+		t.Fatal(err)
+	}
+
+	return f
+}
+
+func TestExtractFromZipFile(t *testing.T) {
+	cases := []struct {
+		name   string
+		expect []string
+		err    error
+	}{
+		{
+			name:   "good",
+			expect: []string{"good"},
+		},
+		{
+			name:   strings.Join([]string{"path", "..", "to", "good"}, string(os.PathSeparator)),
+			expect: []string{filepath.Join("to", "good")},
+		},
+		{
+			name:   strings.Join([]string{"path", "..", "to", "..", "good"}, string(os.PathSeparator)),
+			expect: []string{"good"},
+		},
+		{
+			name:   strings.Join([]string{"path", "to", "..", "..", "good"}, string(os.PathSeparator)),
+			expect: []string{"good"},
+		},
+		{
+			name: strings.Join([]string{"..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"}, string(os.PathSeparator)),
+			err:  zip.ErrInsecurePath,
+		},
+		{
+			name: strings.Join([]string{"path", "..", "..", "to", "bad"}, string(os.PathSeparator)),
+			err:  zip.ErrInsecurePath,
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			f := createZipFile(t, tt.name)
+			defer f.Close()
+
+			tempDir := t.TempDir()
+			if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); !errors.Is(err, tt.err) {
+				t.Fatal(err)
+			}
+
+			var matches []string
+			if err := filepath.Walk(tempDir, func(p string, fi os.FileInfo, err error) error {
+				if err != nil {
+					return err
+				}
+
+				if !fi.IsDir() {
+					matches = append(matches, p)
+				}
+
+				return nil
+			}); err != nil {
+				t.Fatal(err)
+			}
+
+			var actual []string
+			for _, match := range matches {
+				rel, err := filepath.Rel(tempDir, match)
+				if err != nil {
+					t.Error(err)
+				}
+
+				actual = append(actual, rel)
+			}
+
+			if !slices.Equal(actual, tt.expect) {
+				t.Fatalf("expected %d files, got %d", len(tt.expect), len(matches))
+			}
+		})
+	}
+}
+
+func readFile(t *testing.T, base, name string) *bytes.Buffer {
+	t.Helper()
+
+	bts, err := os.ReadFile(filepath.Join(base, name))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	return bytes.NewBuffer(bts)
+}
+
+func TestExecuteWithTools(t *testing.T) {
+	p := filepath.Join("testdata", "tools")
+	cases := []struct {
+		model  string
+		output string
+		ok     bool
+	}{
+		{"mistral", `[TOOL_CALLS]  [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
+		{"mistral", `[TOOL_CALLS]  [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]
+
+The temperature in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.`, true},
+		{"mistral", `I'm not aware of that information. However, I can suggest searching for the weather using the "get_current_weather" function:
+
+		[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
+		{"mistral", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
+		{"command-r-plus", "Action: ```json" + `
+[
+    {
+        "tool_name": "get_current_weather",
+        "parameters": {
+            "format": "fahrenheit",
+            "location": "San Francisco, CA"
+        }
+    },
+    {
+        "tool_name": "get_current_weather",
+        "parameters": {
+            "format": "celsius",
+            "location": "Toronto, Canada"
+        }
+    }
+]
+` + "```", true},
+		{"command-r-plus", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
+		{"firefunction", ` functools[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
+		{"firefunction", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
+		{"llama3-groq-tool-use", `<tool_call>
+{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}}
+{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}
+</tool_call>`, true},
+		{"xlam", `{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]}`, true},
+	}
+
+	var tools []api.Tool
+	if err := json.Unmarshal(readFile(t, p, "tools.json").Bytes(), &tools); err != nil {
+		t.Fatal(err)
+	}
+
+	var messages []api.Message
+	if err := json.Unmarshal(readFile(t, p, "messages.json").Bytes(), &messages); err != nil {
+		t.Fatal(err)
+	}
+
+	calls := []api.ToolCall{
+		{
+			Function: api.ToolCallFunction{
+				Name: "get_current_weather",
+				Arguments: api.ToolCallFunctionArguments{
+					"format":   "fahrenheit",
+					"location": "San Francisco, CA",
+				},
+			},
+		},
+		{
+			Function: api.ToolCallFunction{
+				Name: "get_current_weather",
+				Arguments: api.ToolCallFunctionArguments{
+					"format":   "celsius",
+					"location": "Toronto, Canada",
+				},
+			},
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.model, func(t *testing.T) {
+			tmpl, err := template.Parse(readFile(t, p, fmt.Sprintf("%s.gotmpl", tt.model)).String())
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			t.Run("template", func(t *testing.T) {
+				var actual bytes.Buffer
+				if err := tmpl.Execute(&actual, template.Values{Tools: tools, Messages: messages}); err != nil {
+					t.Fatal(err)
+				}
+
+				if diff := cmp.Diff(actual.String(), readFile(t, p, fmt.Sprintf("%s.out", tt.model)).String()); diff != "" {
+					t.Errorf("mismatch (-got +want):\n%s", diff)
+				}
+			})
+
+			t.Run("parse", func(t *testing.T) {
+				m := &Model{Template: tmpl}
+				actual, ok := m.parseToolCalls(tt.output)
+				if ok != tt.ok {
+					t.Fatalf("expected %t, got %t", tt.ok, ok)
+				}
+
+				if tt.ok {
+					if diff := cmp.Diff(actual, calls); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				}
+			})
+		})
+	}
+}

server/modelpath.go

@@ -103,18 +103,9 @@ func (mp ModelPath) GetShortTagname() string {
 	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
 }
 
-// modelsDir returns the value of the OLLAMA_MODELS environment variable or the user's home directory if OLLAMA_MODELS is not set.
-// The models directory is where Ollama stores its model files and manifests.
-func modelsDir() (string, error) {
-	return envconfig.ModelsDir, nil
-}
-
 // GetManifestPath returns the path to the manifest file for the given model path, it is up to the caller to create the directory if it does not exist.
 func (mp ModelPath) GetManifestPath() (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
+	dir := envconfig.ModelsDir
 
 	return filepath.Join(dir, "manifests", mp.Registry, mp.Namespace, mp.Repository, mp.Tag), nil
 }
@@ -127,10 +118,7 @@ func (mp ModelPath) BaseURL() *url.URL {
 }
 
 func GetManifestPath() (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
+	dir := envconfig.ModelsDir
 
 	path := filepath.Join(dir, "manifests")
 	if err := os.MkdirAll(path, 0o755); err != nil {
@@ -141,10 +129,7 @@ func GetManifestPath() (string, error) {
 }
 
 func GetBlobsPath(digest string) (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
+	dir := envconfig.ModelsDir
 
 	// only accept actual sha256 digests
 	pattern := "^sha256[:-][0-9a-fA-F]{64}$"

server/prompt.go

@@ -1,221 +1,74 @@
 package server
 
 import (
-	"fmt"
+	"bytes"
+	"context"
 	"log/slog"
-	"strings"
-	"text/template"
-	"text/template/parse"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/template"
 )
 
-// isResponseNode checks if the node contains .Response
-func isResponseNode(node *parse.ActionNode) bool {
-	for _, cmd := range node.Pipe.Cmds {
-		for _, arg := range cmd.Args {
-			if fieldNode, ok := arg.(*parse.FieldNode); ok && len(fieldNode.Ident) > 0 {
-				if fieldNode.Ident[0] == "Response" {
-					return true
-				}
+type tokenizeFunc func(context.Context, string) ([]int, error)
+
+// chatPrompt accepts a list of messages and returns the prompt and images that should be used for the next chat turn.
+// chatPrompt truncates any messages that exceed the context window of the model, making sure to always include 1) the
+// latest message and 2) system messages
+func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool) (prompt string, images []llm.ImageData, _ error) {
+	var system []api.Message
+	// always include the last message
+	n := len(msgs) - 1
+	// in reverse, find all messages that fit into context window
+	for i := n - 1; i >= 0; i-- {
+		system = make([]api.Message, 0)
+		for j := range i {
+			if msgs[j].Role == "system" {
+				system = append(system, msgs[j])
 			}
 		}
-	}
-	return false
-}
 
-// formatTemplateForResponse formats the template AST to:
-// 1. remove all nodes after the first .Response (if generate=true)
-// 2. add a .Response node to the end if it doesn't exist
-// TODO(jmorganca): this should recursively cut the template before the first .Response
-func formatTemplateForResponse(tmpl *template.Template, generate bool) {
-	var found bool
-	for i, node := range tmpl.Tree.Root.Nodes {
-		if actionNode, ok := node.(*parse.ActionNode); ok {
-			if isResponseNode(actionNode) {
-				found = true
-				if generate {
-					tmpl.Tree.Root.Nodes = tmpl.Tree.Root.Nodes[:i+1]
-					break
-				}
-			}
+		var b bytes.Buffer
+		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools}); err != nil {
+			return "", nil, err
 		}
-	}
 
-	if !found {
-		// add the response node if it doesn't exist
-		responseFieldNode := &parse.FieldNode{NodeType: parse.NodeField, Ident: []string{"Response"}}
-		responsePipeNode := &parse.PipeNode{NodeType: parse.NodePipe, Cmds: []*parse.CommandNode{{NodeType: parse.NodeCommand, Args: []parse.Node{responseFieldNode}}}}
-		responseActionNode := &parse.ActionNode{NodeType: parse.NodeAction, Pipe: responsePipeNode}
-		tmpl.Tree.Root.Nodes = append(tmpl.Tree.Root.Nodes, responseActionNode)
-	}
-}
-
-// Prompt renders a prompt from a template. If generate is set to true,
-// the response and parts of the template following it are not rendered
-func Prompt(tmpl, system, prompt, response string, generate bool) (string, error) {
-	parsed, err := template.New("").Option("missingkey=zero").Parse(tmpl)
-	if err != nil {
-		return "", err
-	}
-
-	formatTemplateForResponse(parsed, generate)
-
-	vars := map[string]any{
-		"System":   system,
-		"Prompt":   prompt,
-		"Response": response,
-	}
-
-	var sb strings.Builder
-	if err := parsed.Execute(&sb, vars); err != nil {
-		return "", err
-	}
-
-	return sb.String(), nil
-}
-
-func countTokens(tmpl string, system string, prompt string, response string, encode func(string) ([]int, error)) (int, error) {
-	rendered, err := Prompt(tmpl, system, prompt, response, false)
-	if err != nil {
-		return 0, err
-	}
-
-	tokens, err := encode(rendered)
-	if err != nil {
-		slog.Error("failed to encode prompt", "err", err)
-		return 0, err
-	}
-
-	return len(tokens), err
-}
-
-// ChatPrompt builds up a prompt from a series of messages, truncating based on context window size
-func ChatPrompt(tmpl string, messages []api.Message, window int, encode func(string) ([]int, error)) (string, error) {
-	type prompt struct {
-		System   string
-		Prompt   string
-		Response string
-
-		images []int
-		tokens int
-	}
-
-	var p prompt
-
-	// iterate through messages to build up {system,user,response} prompts
-	var imgId int
-	var prompts []prompt
-	for _, msg := range messages {
-		switch strings.ToLower(msg.Role) {
-		case "system":
-			if p.System != "" || p.Prompt != "" || p.Response != "" {
-				prompts = append(prompts, p)
-				p = prompt{}
-			}
-
-			p.System = msg.Content
-		case "user":
-			if p.Prompt != "" || p.Response != "" {
-				prompts = append(prompts, p)
-				p = prompt{}
-			}
-
-			var sb strings.Builder
-			for range msg.Images {
-				fmt.Fprintf(&sb, "[img-%d] ", imgId)
-				p.images = append(p.images, imgId)
-				imgId += 1
-			}
-
-			sb.WriteString(msg.Content)
-			p.Prompt = sb.String()
-		case "assistant":
-			if p.Response != "" {
-				prompts = append(prompts, p)
-				p = prompt{}
-			}
-
-			p.Response = msg.Content
-		default:
-			return "", fmt.Errorf("invalid role: %s, role must be one of [system, user, assistant]", msg.Role)
-		}
-	}
-
-	// add final prompt
-	if p.System != "" || p.Prompt != "" || p.Response != "" {
-		prompts = append(prompts, p)
-	}
-
-	// calculate token lengths for each prompt, estimating 768 tokens per images
-	for i, p := range prompts {
-		tokens, err := countTokens(tmpl, p.System, p.Prompt, p.Response, encode)
+		s, err := tokenize(ctx, b.String())
 		if err != nil {
-			return "", err
+			return "", nil, err
 		}
 
-		prompts[i].tokens = tokens + len(prompts[i].images)*768
-	}
-
-	// truncate images and prompts starting from the beginning of the list
-	// until either one prompt remains or the total tokens fits the context window
-	// TODO (jmorganca): this doesn't account for the context window room required for the response
-	for {
-		var required int
-		for _, p := range prompts {
-			required += p.tokens
+		c := len(s)
+		if m.ProjectorPaths != nil {
+			for _, m := range msgs[i:] {
+				// images are represented as 768 sized embeddings
+				// TODO: get embedding length from project metadata
+				c += 768 * len(m.Images)
+			}
 		}
 
-		required += 1 // for bos token
-
-		if required <= window {
-			slog.Debug("prompt now fits in context window", "required", required, "window", window)
+		if c > opts.NumCtx {
+			slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[i:]))
 			break
+		} else {
+			n = i
 		}
-
-		prompt := &prompts[0]
-
-		if len(prompt.images) > 1 {
-			img := prompt.images[0]
-			slog.Debug("prompt longer than context window, removing image", "id", img, "required", required, "window", window)
-			prompt.images = prompt.images[1:]
-			prompt.Prompt = strings.Replace(prompt.Prompt, fmt.Sprintf(" [img-%d]", img), "", 1)
-			prompt.tokens -= 768
-			continue
-		}
-
-		if len(prompts) > 1 {
-			slog.Debug("required tokens longer than context window, removing first prompt", "prompt", prompts[0].tokens, "required", required, "window", window)
-			system := prompt.System
-			prompts = prompts[1:]
-
-			if system != "" && prompts[0].System == "" {
-				prompts[0].System = system
-
-				tokens, err := countTokens(tmpl, prompts[0].System, prompts[0].Prompt, prompts[0].Response, encode)
-				if err != nil {
-					return "", err
-				}
-
-				prompts[0].tokens = tokens + len(prompts[0].images)*768
-			}
-
-			continue
-		}
-
-		// stop truncating if there's only one prompt left
-		break
 	}
 
-	var sb strings.Builder
-	for i, p := range prompts {
-		// last prompt should leave the response unrendered (for completion)
-		rendered, err := Prompt(tmpl, p.System, p.Prompt, p.Response, i == len(prompts)-1)
-		if err != nil {
-			return "", err
-		}
-		sb.WriteString(rendered)
+	// truncate any messages that do not fit into the context window
+	var b bytes.Buffer
+	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[n:]...), Tools: tools}); err != nil {
+		return "", nil, err
 	}
 
-	return sb.String(), nil
+	for _, m := range msgs[n:] {
+		for _, i := range m.Images {
+			images = append(images, llm.ImageData{
+				ID:   len(images),
+				Data: i,
+			})
+		}
+	}
+
+	return b.String(), images, nil
 }

server/prompt_test.go

@@ -1,204 +1,209 @@
 package server
 
 import (
-	"strings"
+	"bytes"
+	"context"
 	"testing"
 
+	"github.com/google/go-cmp/cmp"
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/template"
 )
 
-func TestPrompt(t *testing.T) {
-	tests := []struct {
-		name     string
-		template string
-		system   string
-		prompt   string
-		response string
-		generate bool
-		want     string
-	}{
-		{
-			name:     "simple prompt",
-			template: "[INST] {{ .System }} {{ .Prompt }} [/INST]",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			want:     "[INST] You are a Wizard. What are the potion ingredients? [/INST]",
-		},
-		{
-			name:     "implicit response",
-			template: "[INST] {{ .System }} {{ .Prompt }} [/INST]",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			want:     "[INST] You are a Wizard. What are the potion ingredients? [/INST]I don't know.",
-		},
-		{
-			name:     "response",
-			template: "[INST] {{ .System }} {{ .Prompt }} [/INST] {{ .Response }}",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			want:     "[INST] You are a Wizard. What are the potion ingredients? [/INST] I don't know.",
-		},
-		{
-			name:     "cut",
-			template: "<system>{{ .System }}</system><user>{{ .Prompt }}</user><assistant>{{ .Response }}</assistant>",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			generate: true,
-			want:     "<system>You are a Wizard.</system><user>What are the potion ingredients?</user><assistant>I don't know.",
-		},
-		{
-			name:     "nocut",
-			template: "<system>{{ .System }}</system><user>{{ .Prompt }}</user><assistant>{{ .Response }}</assistant>",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			want:     "<system>You are a Wizard.</system><user>What are the potion ingredients?</user><assistant>I don't know.</assistant>",
-		},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			got, err := Prompt(tc.template, tc.system, tc.prompt, tc.response, tc.generate)
-			if err != nil {
-				t.Errorf("error = %v", err)
-			}
-
-			if got != tc.want {
-				t.Errorf("got = %v, want %v", got, tc.want)
-			}
-		})
-	}
-}
-
 func TestChatPrompt(t *testing.T) {
-	tests := []struct {
-		name     string
-		template string
-		messages []api.Message
-		window   int
-		want     string
+	type expect struct {
+		prompt string
+		images [][]byte
+	}
+
+	cases := []struct {
+		name  string
+		limit int
+		msgs  []api.Message
+		expect
 	}{
 		{
-			name:     "simple prompt",
-			template: "[INST] {{ .Prompt }} [/INST]",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello"},
+			name:  "messages",
+			limit: 64,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
 			},
-			window: 1024,
-			want:   "[INST] Hello [/INST]",
 		},
 		{
-			name:     "with system message",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST]",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
+			name:  "truncate messages",
+			limit: 1,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "A test. And a thumping good one at that, I'd wager. ",
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> Hello [/INST]",
 		},
 		{
-			name:     "with response",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST] {{ .Response }}",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
-				{Role: "assistant", Content: "I am?"},
+			name:  "truncate messages with image",
+			limit: 64,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("something")}},
+			},
+			expect: expect{
+				prompt: "[img-0] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+				},
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> Hello [/INST] I am?",
 		},
 		{
-			name:     "with implicit response",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST]",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
-				{Role: "assistant", Content: "I am?"},
+			name:  "truncate messages with images",
+			limit: 64,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
+			},
+			expect: expect{
+				prompt: "[img-0] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> Hello [/INST]I am?",
 		},
 		{
-			name:     "with conversation",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST] {{ .Response }} ",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "What are the potion ingredients?"},
-				{Role: "assistant", Content: "sugar"},
-				{Role: "user", Content: "Anything else?"},
+			name:  "messages with images",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
+			},
+			expect: expect{
+				prompt: "[img-0] You're a test, Harry! I-I'm a what? [img-1] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> What are the potion ingredients? [/INST] sugar [INST] Anything else? [/INST] ",
 		},
 		{
-			name:     "with truncation",
-			template: "{{ .System }} {{ .Prompt }} {{ .Response }} ",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
-				{Role: "assistant", Content: "I am?"},
-				{Role: "user", Content: "Why is the sky blue?"},
-				{Role: "assistant", Content: "The sky is blue from rayleigh scattering"},
+			name:  "message with image tag",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry! [img]", Images: []api.ImageData{[]byte("something")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry! [img-0] I-I'm a what? [img-1] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+					[]byte("somethingelse"),
+				},
 			},
-			window: 10,
-			want:   "You are a Wizard. Why is the sky blue? The sky is blue from rayleigh scattering",
 		},
 		{
-			name:     "images",
-			template: "{{ .System }} {{ .Prompt }}",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello", Images: []api.ImageData{[]byte("base64")}},
+			name:  "messages with interleaved images",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "user", Images: []api.ImageData{[]byte("something")}},
+				{Role: "user", Images: []api.ImageData{[]byte("somethingelse")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry!\n\n[img-0]\n\n[img-1] I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "You are a Wizard. [img-0] Hello",
 		},
 		{
-			name:     "images truncated",
-			template: "{{ .System }} {{ .Prompt }}",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello", Images: []api.ImageData{[]byte("img1"), []byte("img2")}},
+			name:  "truncate message with interleaved images",
+			limit: 1024,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "user", Images: []api.ImageData{[]byte("something")}},
+				{Role: "user", Images: []api.ImageData{[]byte("somethingelse")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "[img-0] I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "You are a Wizard. [img-0] [img-1] Hello",
 		},
 		{
-			name:     "empty list",
-			template: "{{ .System }} {{ .Prompt }}",
-			messages: []api.Message{},
-			window:   1024,
-			want:     "",
+			name:  "message with system prompt",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "system", Content: "You are the Test Who Lived."},
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You are the Test Who Lived. You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
+			},
 		},
 		{
-			name:     "empty prompt",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST] {{ .Response }} ",
-			messages: []api.Message{
-				{Role: "user", Content: ""},
+			name:  "out of order system",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "system", Content: "You are the Test Who Lived."},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry! I-I'm a what? You are the Test Who Lived. A test. And a thumping good one at that, I'd wager. ",
 			},
-			window: 1024,
-			want:   "",
 		},
 	}
 
-	encode := func(s string) ([]int, error) {
-		words := strings.Fields(s)
-		return make([]int, len(words)), nil
+	tmpl, err := template.Parse(`
+{{- if .System }}{{ .System }} {{ end }}
+{{- if .Prompt }}{{ .Prompt }} {{ end }}
+{{- if .Response }}{{ .Response }} {{ end }}`)
+	if err != nil {
+		t.Fatal(err)
 	}
 
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			got, err := ChatPrompt(tc.template, tc.messages, tc.window, encode)
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			model := Model{Template: tmpl, ProjectorPaths: []string{"vision"}}
+			opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
+			prompt, images, err := chatPrompt(context.TODO(), &model, mockRunner{}.Tokenize, &opts, tt.msgs, nil)
 			if err != nil {
-				t.Errorf("error = %v", err)
+				t.Fatal(err)
 			}
 
-			if got != tc.want {
-				t.Errorf("got: %q, want: %q", got, tc.want)
+			if diff := cmp.Diff(prompt, tt.prompt); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+
+			if len(images) != len(tt.images) {
+				t.Fatalf("expected %d images, got %d", len(tt.images), len(images))
+			}
+
+			for i := range images {
+				if images[i].ID != i {
+					t.Errorf("expected ID %d, got %d", i, images[i].ID)
+				}
+
+				if !bytes.Equal(images[i].Data, tt.images[i]) {
+					t.Errorf("expected %q, got %q", tt.images[i], images[i])
+				}
 			}
 		})
 	}

server/routes_create_test.go

@@ -85,6 +85,8 @@ func checkFileExists(t *testing.T, p string, expect []string) {
 }
 
 func TestCreateFromBin(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -111,6 +113,8 @@ func TestCreateFromBin(t *testing.T) {
 }
 
 func TestCreateFromModel(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -152,6 +156,8 @@ func TestCreateFromModel(t *testing.T) {
 }
 
 func TestCreateRemovesLayers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -199,6 +205,8 @@ func TestCreateRemovesLayers(t *testing.T) {
 }
 
 func TestCreateUnsetsSystem(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -255,6 +263,8 @@ func TestCreateUnsetsSystem(t *testing.T) {
 }
 
 func TestCreateMergeParameters(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -358,6 +368,8 @@ func TestCreateMergeParameters(t *testing.T) {
 }
 
 func TestCreateReplacesMessages(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -434,6 +446,8 @@ func TestCreateReplacesMessages(t *testing.T) {
 }
 
 func TestCreateTemplateSystem(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -477,9 +491,47 @@ func TestCreateTemplateSystem(t *testing.T) {
 	if string(system) != "Say bye!" {
 		t.Errorf("expected \"Say bye!\", actual %s", system)
 	}
+
+	t.Run("incomplete template", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ .Prompt", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+	
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
+
+	t.Run("template with unclosed if", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ if .Prompt }}", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+	
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
+
+	t.Run("template with undefined function", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{  Prompt }}", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+	
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
 }
 
 func TestCreateLicenses(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -526,6 +578,8 @@ func TestCreateLicenses(t *testing.T) {
 }
 
 func TestCreateDetectTemplate(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -545,9 +599,9 @@ func TestCreateDetectTemplate(t *testing.T) {
 		}
 
 		checkFileExists(t, filepath.Join(p, "blobs", "*"), []string{
-			filepath.Join(p, "blobs", "sha256-2f8e594e6f34b1b4d36a246628eeb3365ce442303d656f1fcc69e821722acea0"),
-			filepath.Join(p, "blobs", "sha256-542b217f179c7825eeb5bca3c77d2b75ed05bafbd3451d9188891a60a85337c6"),
 			filepath.Join(p, "blobs", "sha256-553c4a3f747b3d22a4946875f1cc8ed011c2930d83f864a0c7265f9ec0a20413"),
+			filepath.Join(p, "blobs", "sha256-c608dc615584cd20d9d830363dabf8a4783ae5d34245c3d8c115edb3bc7b28e4"),
+			filepath.Join(p, "blobs", "sha256-f836ee110db21567f826332e4cedd746c06d10664fd5a9ea3659e3683a944510"),
 		})
 	})
 

server/routes_delete_test.go

@@ -8,12 +8,15 @@ import (
 	"path/filepath"
 	"testing"
 
+	"github.com/gin-gonic/gin"
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/types/model"
 )
 
 func TestDelete(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	envconfig.LoadConfig()
@@ -77,6 +80,8 @@ func TestDelete(t *testing.T) {
 }
 
 func TestDeleteDuplicateLayers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	var s Server

server/routes_generate_test.go

@@ -0,0 +1,714 @@
+package server
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/gpu"
+	"github.com/ollama/ollama/llm"
+)
+
+type mockRunner struct {
+	llm.LlamaServer
+
+	// CompletionRequest is only valid until the next call to Completion
+	llm.CompletionRequest
+	llm.CompletionResponse
+}
+
+func (m *mockRunner) Completion(_ context.Context, r llm.CompletionRequest, fn func(r llm.CompletionResponse)) error {
+	m.CompletionRequest = r
+	fn(m.CompletionResponse)
+	return nil
+}
+
+func (mockRunner) Tokenize(_ context.Context, s string) (tokens []int, err error) {
+	for range strings.Fields(s) {
+		tokens = append(tokens, len(tokens))
+	}
+
+	return
+}
+
+func newMockServer(mock *mockRunner) func(gpu.GpuInfoList, string, *llm.GGML, []string, []string, api.Options, int) (llm.LlamaServer, error) {
+	return func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, projectors, system []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
+		return mock, nil
+	}
+}
+
+func TestGenerateChat(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	mock := mockRunner{
+		CompletionResponse: llm.CompletionResponse{
+			Done:               true,
+			DoneReason:         "stop",
+			PromptEvalCount:    1,
+			PromptEvalDuration: 1,
+			EvalCount:          1,
+			EvalDuration:       1,
+		},
+	}
+
+	s := Server{
+		sched: &Scheduler{
+			pendingReqCh:  make(chan *LlmRequest, 1),
+			finishedReqCh: make(chan *LlmRequest, 1),
+			expiredCh:     make(chan *runnerRef, 1),
+			unloadedCh:    make(chan any, 1),
+			loaded:        make(map[string]*runnerRef),
+			newServerFn:   newMockServer(&mock),
+			getGpuFn:      gpu.GetGPUInfo,
+			getCpuFn:      gpu.GetCPUInfo,
+			reschedDelay:  250 * time.Millisecond,
+			loadFn: func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+				// add small delay to simulate loading
+				time.Sleep(time.Millisecond)
+				req.successCh <- &runnerRef{
+					llama: &mock,
+				}
+			},
+		},
+	}
+
+	go s.sched.Run(context.TODO())
+
+	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		Model: "test",
+		Modelfile: fmt.Sprintf(`FROM %s
+		TEMPLATE """
+{{- if .System }}System: {{ .System }} {{ end }}
+{{- if .Prompt }}User: {{ .Prompt }} {{ end }}
+{{- if .Response }}Assistant: {{ .Response }} {{ end }}"""
+`, createBinFile(t, llm.KV{
+			"general.architecture":          "llama",
+			"llama.block_count":             uint32(1),
+			"llama.context_length":          uint32(8192),
+			"llama.embedding_length":        uint32(4096),
+			"llama.attention.head_count":    uint32(32),
+			"llama.attention.head_count_kv": uint32(8),
+			"tokenizer.ggml.tokens":         []string{""},
+			"tokenizer.ggml.scores":         []float32{0},
+			"tokenizer.ggml.token_type":     []int32{0},
+		}, []llm.Tensor{
+			{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		})),
+		Stream: &stream,
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("missing body", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, nil)
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing model", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{})
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing capabilities chat", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Model: "bert",
+			Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, llm.KV{
+				"general.architecture": "bert",
+				"bert.pooling_type":    uint32(0),
+			}, []llm.Tensor{})),
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d", w.Code)
+		}
+
+		w = createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "bert",
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"\"bert\" does not support chat"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("load model", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		var actual api.ChatResponse
+		if err := json.NewDecoder(w.Body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != "test" {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done true, got false")
+		}
+
+		if actual.DoneReason != "load" {
+			t.Errorf("expected done reason load, got %s", actual.DoneReason)
+		}
+	})
+
+	checkChatResponse := func(t *testing.T, body io.Reader, model, content string) {
+		t.Helper()
+
+		var actual api.ChatResponse
+		if err := json.NewDecoder(body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != model {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done false, got true")
+		}
+
+		if actual.DoneReason != "stop" {
+			t.Errorf("expected done reason stop, got %s", actual.DoneReason)
+		}
+
+		if diff := cmp.Diff(actual.Message, api.Message{
+			Role:    "assistant",
+			Content: content,
+		}); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		if actual.PromptEvalCount == 0 {
+			t.Errorf("expected prompt eval count > 0, got 0")
+		}
+
+		if actual.PromptEvalDuration == 0 {
+			t.Errorf("expected prompt eval duration > 0, got 0")
+		}
+
+		if actual.EvalCount == 0 {
+			t.Errorf("expected eval count > 0, got 0")
+		}
+
+		if actual.EvalDuration == 0 {
+			t.Errorf("expected eval duration > 0, got 0")
+		}
+
+		if actual.LoadDuration == 0 {
+			t.Errorf("expected load duration > 0, got 0")
+		}
+
+		if actual.TotalDuration == 0 {
+			t.Errorf("expected total duration > 0, got 0")
+		}
+	}
+
+	mock.CompletionResponse.Content = "Hi!"
+	t.Run("messages", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test",
+			Messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test", "Hi!")
+	})
+
+	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		Model:     "test-system",
+		Modelfile: "FROM test\nSYSTEM You are a helpful assistant.",
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("messages with model system", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test-system",
+			Messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You are a helpful assistant. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test-system", "Hi!")
+	})
+
+	mock.CompletionResponse.Content = "Abra kadabra!"
+	t.Run("messages with system", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test-system",
+			Messages: []api.Message{
+				{Role: "system", Content: "You can perform magic tricks."},
+				{Role: "user", Content: "Hello!"},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You can perform magic tricks. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+
+	t.Run("messages with interleaved system", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test-system",
+			Messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+				{Role: "assistant", Content: "I can help you with that."},
+				{Role: "system", Content: "You can perform magic tricks."},
+				{Role: "user", Content: "Help me write tests."},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You are a helpful assistant. User: Hello! Assistant: I can help you with that. System: You can perform magic tricks. User: Help me write tests. "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+}
+
+func TestGenerate(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	mock := mockRunner{
+		CompletionResponse: llm.CompletionResponse{
+			Done:               true,
+			DoneReason:         "stop",
+			PromptEvalCount:    1,
+			PromptEvalDuration: 1,
+			EvalCount:          1,
+			EvalDuration:       1,
+		},
+	}
+
+	s := Server{
+		sched: &Scheduler{
+			pendingReqCh:  make(chan *LlmRequest, 1),
+			finishedReqCh: make(chan *LlmRequest, 1),
+			expiredCh:     make(chan *runnerRef, 1),
+			unloadedCh:    make(chan any, 1),
+			loaded:        make(map[string]*runnerRef),
+			newServerFn:   newMockServer(&mock),
+			getGpuFn:      gpu.GetGPUInfo,
+			getCpuFn:      gpu.GetCPUInfo,
+			reschedDelay:  250 * time.Millisecond,
+			loadFn: func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+				// add small delay to simulate loading
+				time.Sleep(time.Millisecond)
+				req.successCh <- &runnerRef{
+					llama: &mock,
+				}
+			},
+		},
+	}
+
+	go s.sched.Run(context.TODO())
+
+	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		Model: "test",
+		Modelfile: fmt.Sprintf(`FROM %s
+		TEMPLATE """
+{{- if .System }}System: {{ .System }} {{ end }}
+{{- if .Prompt }}User: {{ .Prompt }} {{ end }}
+{{- if .Response }}Assistant: {{ .Response }} {{ end }}"""
+`, createBinFile(t, llm.KV{
+			"general.architecture":          "llama",
+			"llama.block_count":             uint32(1),
+			"llama.context_length":          uint32(8192),
+			"llama.embedding_length":        uint32(4096),
+			"llama.attention.head_count":    uint32(32),
+			"llama.attention.head_count_kv": uint32(8),
+			"tokenizer.ggml.tokens":         []string{""},
+			"tokenizer.ggml.scores":         []float32{0},
+			"tokenizer.ggml.token_type":     []int32{0},
+		}, []llm.Tensor{
+			{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		})),
+		Stream: &stream,
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("missing body", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, nil)
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing model", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{})
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing capabilities generate", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Model: "bert",
+			Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, llm.KV{
+				"general.architecture": "bert",
+				"bert.pooling_type":    uint32(0),
+			}, []llm.Tensor{})),
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d", w.Code)
+		}
+
+		w = createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model: "bert",
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"\"bert\" does not support generate"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing capabilities suffix", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test",
+			Prompt: "def add(",
+			Suffix: "    return c",
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"test does not support insert"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("load model", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model: "test",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		var actual api.GenerateResponse
+		if err := json.NewDecoder(w.Body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != "test" {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done true, got false")
+		}
+
+		if actual.DoneReason != "load" {
+			t.Errorf("expected done reason load, got %s", actual.DoneReason)
+		}
+	})
+
+	checkGenerateResponse := func(t *testing.T, body io.Reader, model, content string) {
+		t.Helper()
+
+		var actual api.GenerateResponse
+		if err := json.NewDecoder(body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != model {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done false, got true")
+		}
+
+		if actual.DoneReason != "stop" {
+			t.Errorf("expected done reason stop, got %s", actual.DoneReason)
+		}
+
+		if actual.Response != content {
+			t.Errorf("expected response %s, got %s", content, actual.Response)
+		}
+
+		if actual.Context == nil {
+			t.Errorf("expected context not nil")
+		}
+
+		if actual.PromptEvalCount == 0 {
+			t.Errorf("expected prompt eval count > 0, got 0")
+		}
+
+		if actual.PromptEvalDuration == 0 {
+			t.Errorf("expected prompt eval duration > 0, got 0")
+		}
+
+		if actual.EvalCount == 0 {
+			t.Errorf("expected eval count > 0, got 0")
+		}
+
+		if actual.EvalDuration == 0 {
+			t.Errorf("expected eval duration > 0, got 0")
+		}
+
+		if actual.LoadDuration == 0 {
+			t.Errorf("expected load duration > 0, got 0")
+		}
+
+		if actual.TotalDuration == 0 {
+			t.Errorf("expected total duration > 0, got 0")
+		}
+	}
+
+	mock.CompletionResponse.Content = "Hi!"
+	t.Run("prompt", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test",
+			Prompt: "Hello!",
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test", "Hi!")
+	})
+
+	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		Model:     "test-system",
+		Modelfile: "FROM test\nSYSTEM You are a helpful assistant.",
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("prompt with model system", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Hello!",
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You are a helpful assistant. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test-system", "Hi!")
+	})
+
+	mock.CompletionResponse.Content = "Abra kadabra!"
+	t.Run("prompt with system", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Hello!",
+			System: "You can perform magic tricks.",
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You can perform magic tricks. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+
+	t.Run("prompt with template", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Help me write tests.",
+			System: "You can perform magic tricks.",
+			Template: `{{- if .System }}{{ .System }} {{ end }}
+{{- if .Prompt }}### USER {{ .Prompt }} {{ end }}
+{{- if .Response }}### ASSISTANT {{ .Response }} {{ end }}`,
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "You can perform magic tricks. ### USER Help me write tests. "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+
+	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		Model: "test-suffix",
+		Modelfile: `FROM test
+TEMPLATE """{{- if .Suffix }}<PRE> {{ .Prompt }} <SUF>{{ .Suffix }} <MID>
+{{- else }}{{ .Prompt }}
+{{- end }}"""`,
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("prompt with suffix", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-suffix",
+			Prompt: "def add(",
+			Suffix: "    return c",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "<PRE> def add( <SUF>    return c <MID>"); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("prompt without suffix", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-suffix",
+			Prompt: "def add(",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "def add("); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("raw", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Help me write tests.",
+			Raw:    true,
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "Help me write tests."); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+}

server/routes_list_test.go

@@ -7,11 +7,14 @@ import (
 	"slices"
 	"testing"
 
+	"github.com/gin-gonic/gin"
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 )
 
 func TestList(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	t.Setenv("OLLAMA_MODELS", t.TempDir())
 	envconfig.LoadConfig()
 

server/routes_test.go

@@ -7,6 +7,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
+	"math"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -19,6 +20,8 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -104,6 +107,24 @@ func Test_Routes(t *testing.T) {
 				assert.Empty(t, len(modelList.Models))
 			},
 		},
+		{
+			Name:   "openai empty list",
+			Method: http.MethodGet,
+			Path:   "/v1/models",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var modelList openai.ListCompletion
+				err = json.Unmarshal(body, &modelList)
+				require.NoError(t, err)
+
+				assert.Equal(t, "list", modelList.Object)
+				assert.Empty(t, modelList.Data)
+			},
+		},
 		{
 			Name:   "Tags Handler (yes tags)",
 			Method: http.MethodGet,
@@ -127,6 +148,25 @@ func Test_Routes(t *testing.T) {
 				assert.Equal(t, "test-model:latest", modelList.Models[0].Name)
 			},
 		},
+		{
+			Name:   "openai list models with tags",
+			Method: http.MethodGet,
+			Path:   "/v1/models",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var modelList openai.ListCompletion
+				err = json.Unmarshal(body, &modelList)
+				require.NoError(t, err)
+
+				assert.Len(t, modelList.Data, 1)
+				assert.Equal(t, "test-model:latest", modelList.Data[0].Id)
+				assert.Equal(t, "library", modelList.Data[0].OwnedBy)
+			},
+		},
 		{
 			Name:   "Create Model Handler",
 			Method: http.MethodPost,
@@ -212,6 +252,96 @@ func Test_Routes(t *testing.T) {
 					"top_p 0.9",
 				}
 				assert.Equal(t, expectedParams, params)
+				assert.InDelta(t, 0, showResp.ModelInfo["general.parameter_count"], 1e-9, "Parameter count should be 0")
+			},
+		},
+		{
+			Name:   "openai retrieve model handler",
+			Method: http.MethodGet,
+			Path:   "/v1/models/show-model",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var retrieveResp api.RetrieveModelResponse
+				err = json.Unmarshal(body, &retrieveResp)
+				require.NoError(t, err)
+
+				assert.Equal(t, "show-model", retrieveResp.Id)
+				assert.Equal(t, "library", retrieveResp.OwnedBy)
+			},
+		},
+		{
+			Name:   "Embed Handler Empty Input",
+			Method: http.MethodPost,
+			Path:   "/api/embed",
+			Setup: func(t *testing.T, req *http.Request) {
+				embedReq := api.EmbedRequest{
+					Model: "t-bone",
+					Input: "",
+				}
+				jsonData, err := json.Marshal(embedReq)
+				require.NoError(t, err)
+				req.Body = io.NopCloser(bytes.NewReader(jsonData))
+			},
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				if contentType != "application/json; charset=utf-8" {
+					t.Fatalf("expected content type application/json; charset=utf-8, got %s", contentType)
+				}
+				body, err := io.ReadAll(resp.Body)
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				var embedResp api.EmbedResponse
+				err = json.Unmarshal(body, &embedResp)
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if embedResp.Model != "t-bone" {
+					t.Fatalf("expected model t-bone, got %s", embedResp.Model)
+				}
+
+				if embedResp.Embeddings == nil {
+					t.Fatalf("expected embeddings to not be nil, got %v", embedResp.Embeddings)
+				}
+
+				if len(embedResp.Embeddings) != 0 {
+					t.Fatalf("expected embeddings to be empty, got %v", embedResp.Embeddings)
+				}
+			},
+		},
+		{
+			Name:   "Embed Handler Invalid Input",
+			Method: http.MethodPost,
+			Path:   "/api/embed",
+			Setup: func(t *testing.T, req *http.Request) {
+				embedReq := api.EmbedRequest{
+					Model: "t-bone",
+					Input: 2,
+				}
+				jsonData, err := json.Marshal(embedReq)
+				require.NoError(t, err)
+				req.Body = io.NopCloser(bytes.NewReader(jsonData))
+			},
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				if contentType != "application/json; charset=utf-8" {
+					t.Fatalf("expected content type application/json; charset=utf-8, got %s", contentType)
+				}
+				_, err := io.ReadAll(resp.Body)
+
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if resp.StatusCode != http.StatusBadRequest {
+					t.Fatalf("expected status code 400, got %d", resp.StatusCode)
+				}
 			},
 		},
 	}
@@ -325,3 +455,75 @@ func TestCase(t *testing.T) {
 		})
 	}
 }
+
+func TestShow(t *testing.T) {
+	t.Setenv("OLLAMA_MODELS", t.TempDir())
+	envconfig.LoadConfig()
+
+	var s Server
+
+	createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		Name: "show-model",
+		Modelfile: fmt.Sprintf(
+			"FROM %s\nFROM %s",
+			createBinFile(t, llm.KV{"general.architecture": "test"}, nil),
+			createBinFile(t, llm.KV{"general.architecture": "clip"}, nil),
+		),
+	})
+
+	w := createRequest(t, s.ShowModelHandler, api.ShowRequest{
+		Name: "show-model",
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status code 200, actual %d", w.Code)
+	}
+
+	var resp api.ShowResponse
+	if err := json.NewDecoder(w.Body).Decode(&resp); err != nil {
+		t.Fatal(err)
+	}
+
+	if resp.ModelInfo["general.architecture"] != "test" {
+		t.Fatal("Expected model architecture to be 'test', but got", resp.ModelInfo["general.architecture"])
+	}
+
+	if resp.ProjectorInfo["general.architecture"] != "clip" {
+		t.Fatal("Expected projector architecture to be 'clip', but got", resp.ProjectorInfo["general.architecture"])
+	}
+}
+
+func TestNormalize(t *testing.T) {
+	type testCase struct {
+		input []float32
+	}
+
+	testCases := []testCase{
+		{input: []float32{1}},
+		{input: []float32{0, 1, 2, 3}},
+		{input: []float32{0.1, 0.2, 0.3}},
+		{input: []float32{-0.1, 0.2, 0.3, -0.4}},
+		{input: []float32{0, 0, 0}},
+	}
+
+	isNormalized := func(vec []float32) (res bool) {
+		sum := 0.0
+		for _, v := range vec {
+			sum += float64(v * v)
+		}
+		if math.Abs(sum-1) > 1e-6 {
+			return sum == 0
+		} else {
+			return true
+		}
+	}
+
+	for _, tc := range testCases {
+		t.Run("", func(t *testing.T) {
+			normalized := normalize(tc.input)
+			if !isNormalized(normalized) {
+				t.Errorf("Vector %v is not normalized", tc.input)
+			}
+		})
+	}
+}

server/sched.go

@@ -23,7 +23,8 @@ type LlmRequest struct {
 	ctx             context.Context //nolint:containedctx
 	model           *Model
 	opts            api.Options
-	sessionDuration time.Duration
+	origNumCtx      int // Track the initial ctx request
+	sessionDuration *api.Duration
 	successCh       chan *runnerRef
 	errCh           chan error
 	schedAttempts   uint
@@ -38,13 +39,23 @@ type Scheduler struct {
 	loaded   map[string]*runnerRef
 	loadedMu sync.Mutex
 
-	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
-	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error)
+	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int)
+	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error)
 	getGpuFn     func() gpu.GpuInfoList
 	getCpuFn     func() gpu.GpuInfoList
 	reschedDelay time.Duration
 }
 
+// Default automatic value for number of models we allow per GPU
+// Model will still need to fit in VRAM, but loading many small models
+// on a large GPU can cause stalling
+var defaultModelsPerGPU = 3
+
+// Default automatic value for parallel setting
+// Model will still need to fit in VRAM.  If this setting wont fit
+// we'll back off down to 1 to try to get it to fit
+var defaultParallel = 4
+
 var ErrMaxQueue = fmt.Errorf("server busy, please try again.  maximum pending requests exceeded")
 
 func InitScheduler(ctx context.Context) *Scheduler {
@@ -64,14 +75,11 @@ func InitScheduler(ctx context.Context) *Scheduler {
 }
 
 // context must be canceled to decrement ref count and release the runner
-func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) {
-	// allocate a large enough kv cache for all parallel requests
+func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration *api.Duration) (chan *runnerRef, chan error) {
 	if opts.NumCtx < 4 {
 		opts.NumCtx = 4
 	}
 
-	opts.NumCtx *= envconfig.NumParallel
-
 	req := &LlmRequest{
 		ctx:             c,
 		model:           model,
@@ -110,11 +118,21 @@ func (s *Scheduler) processPending(ctx context.Context) {
 		case pending := <-s.pendingReqCh:
 			// Block other requests until we get this pending request running
 			pending.schedAttempts++
+			if pending.origNumCtx == 0 {
+				pending.origNumCtx = pending.opts.NumCtx
+			}
 
 			if pending.ctx.Err() != nil {
 				slog.Debug("pending request cancelled or timed out, skipping scheduling")
 				continue
 			}
+			numParallel := envconfig.NumParallel
+			// TODO (jmorganca): multimodal models don't support parallel yet
+			// see https://github.com/ollama/ollama/issues/4165
+			if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 {
+				numParallel = 1
+				slog.Warn("multimodal models don't support parallel requests yet")
+			}
 
 			for {
 				var runnerToExpire *runnerRef
@@ -143,8 +161,28 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						gpus = s.getGpuFn()
 					}
 
+					if envconfig.MaxRunners <= 0 {
+						// No user specified MaxRunners, so figure out what automatic setting to use
+						// If all GPUs have reliable free memory reporting, defaultModelsPerGPU * the number of GPUs
+						// if any GPU has unreliable free memory reporting, 1x the number of GPUs
+						allReliable := true
+						for _, gpu := range gpus {
+							if gpu.UnreliableFreeMemory {
+								allReliable = false
+								break
+							}
+						}
+						if allReliable {
+							envconfig.MaxRunners = defaultModelsPerGPU * len(gpus)
+							slog.Debug("updating default concurrency", "OLLAMA_MAX_LOADED_MODELS", envconfig.MaxRunners, "gpu_count", len(gpus))
+						} else {
+							slog.Info("one or more GPUs detected that are unable to accurately report free memory - disabling default concurrency")
+							envconfig.MaxRunners = len(gpus)
+						}
+					}
+
 					// Load model for fitting
-					ggml, err := llm.LoadModel(pending.model.ModelPath)
+					ggml, err := llm.LoadModel(pending.model.ModelPath, 0)
 					if err != nil {
 						pending.errCh <- err
 						break
@@ -152,26 +190,33 @@ func (s *Scheduler) processPending(ctx context.Context) {
 
 					// 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
+						if numParallel <= 0 {
+							numParallel = defaultParallel
+						}
+
+						pending.opts.NumCtx = pending.origNumCtx * numParallel
+
 						if loadedCount == 0 {
 							slog.Debug("cpu mode with first model, loading")
-							s.loadFn(pending, ggml, gpus)
+							s.loadFn(pending, ggml, gpus, numParallel)
 							break
 						}
 						runnerToExpire = s.maybeFindCPURunnerToUnload(pending, ggml, gpus)
 						if runnerToExpire == nil {
 							slog.Debug("cpu mode with available system memory or first model, loading")
-							s.loadFn(pending, ggml, gpus)
+							s.loadFn(pending, ggml, gpus, numParallel)
 							break
 						}
 						// else we need to expire a runner
 					} else if loadedCount == 0 {
 						// No models loaded. Load the model but prefer the best fit.
 						slog.Debug("loading first model", "model", pending.model.ModelPath)
-						g := pickBestFitGPUs(pending, ggml, gpus)
+						g := pickBestFitGPUs(pending, ggml, gpus, &numParallel)
 						if g != nil {
 							gpus = g
 						}
-						s.loadFn(pending, ggml, gpus)
+						s.loadFn(pending, ggml, gpus, numParallel)
 						break
 					}
 
@@ -186,10 +231,10 @@ func (s *Scheduler) processPending(ctx context.Context) {
 
 						// Update free memory from currently loaded models
 						s.updateFreeSpace(availGpus)
-						fitGpus := pickBestFitGPUs(pending, ggml, availGpus)
+						fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel)
 						if fitGpus != nil {
 							slog.Debug("new model fits with existing models, loading")
-							s.loadFn(pending, ggml, fitGpus)
+							s.loadFn(pending, ggml, fitGpus, numParallel)
 							break
 						}
 
@@ -341,7 +386,9 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 		runner.expireTimer.Stop()
 		runner.expireTimer = nil
 	}
-	runner.sessionDuration = pending.sessionDuration
+	if pending.sessionDuration != nil {
+		runner.sessionDuration = pending.sessionDuration.Duration
+	}
 	pending.successCh <- runner
 	go func() {
 		<-pending.ctx.Done()
@@ -350,8 +397,15 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 	}()
 }
 
-func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) {
-	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+	if numParallel < 1 {
+		numParallel = 1
+	}
+	sessionDuration := envconfig.KeepAlive
+	if req.sessionDuration != nil {
+		sessionDuration = req.sessionDuration.Duration
+	}
+	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel)
 	if err != nil {
 		// some older models are not compatible with newer versions of llama.cpp
 		// show a generalized compatibility error until there is a better way to
@@ -368,13 +422,14 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 		modelPath:       req.model.ModelPath,
 		llama:           llama,
 		Options:         &req.opts,
-		sessionDuration: req.sessionDuration,
+		sessionDuration: sessionDuration,
 		gpus:            gpus,
 		estimatedVRAM:   llama.EstimatedVRAM(),
 		estimatedTotal:  llama.EstimatedTotal(),
 		loading:         true,
 		refCount:        1,
 	}
+	runner.numParallel = numParallel
 	runner.refMu.Lock()
 
 	s.loadedMu.Lock()
@@ -483,8 +538,9 @@ type runnerRef struct {
 	expireTimer     *time.Timer
 	expiresAt       time.Time
 
-	model     *Model
-	modelPath string
+	model       *Model
+	modelPath   string
+	numParallel int
 	*api.Options
 }
 
@@ -525,6 +581,9 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 		optsNew.NumGPU = -1
 	}
 
+	// Normalize the NumCtx for parallelism
+	optsExisting.NumCtx = optsExisting.NumCtx / runner.numParallel
+
 	ctx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
 	if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed?
@@ -611,22 +670,38 @@ func (a ByDuration) Less(i, j int) bool {
 
 // pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
 // If the model can not be fit fully within the available GPU(s) nil is returned
-func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.GpuInfoList {
+// If numParallel is <= 0, this will attempt try to optimize parallism based on available VRAM, and adjust
+// opts.NumCtx accordingly
+func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
 	var estimatedVRAM uint64
+
+	var numParallelToTry []int
+	if *numParallel <= 0 {
+		// If no specific parallel setting was provided, try larger then smaller, always end with 1
+		numParallelToTry = append(numParallelToTry, defaultParallel, 1)
+	} else {
+		numParallelToTry = []int{*numParallel}
+	}
+
 	for _, gl := range gpus.ByLibrary() {
 		var ok bool
 		sgl := append(make(gpu.GpuInfoList, 0, len(gl)), gl...)
 
 		// TODO - potentially sort by performance capability, existing models loaded, etc.
+		// TODO - Eliminate any GPUs that already have envconfig.MaxRunners loaded on them
 		// Note: at present, this will favor more VRAM over faster GPU speed in mixed setups
 		sort.Sort(sort.Reverse(gpu.ByFreeMemory(sgl)))
 
 		// First attempt to fit the model into a single GPU
-		if !envconfig.SchedSpread {
-			for _, g := range sgl {
-				if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
-					slog.Debug("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
-					return []gpu.GpuInfo{g}
+		for _, p := range numParallelToTry {
+			req.opts.NumCtx = req.origNumCtx * p
+			if !envconfig.SchedSpread {
+				for _, g := range sgl {
+					if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
+						*numParallel = p
+						return []gpu.GpuInfo{g}
+					}
 				}
 			}
 		}
@@ -636,9 +711,13 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.
 		// - try subsets of GPUs instead of just falling back to 1 or all in a family
 
 		// Now try all the GPUs
-		if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
-			slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
-			return sgl
+		for _, p := range numParallelToTry {
+			req.opts.NumCtx = req.origNumCtx * p
+			if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+				slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM))
+				*numParallel = p
+				return sgl
+			}
 		}
 	}
 	return nil

server/sched_test.go

@@ -44,14 +44,14 @@ func TestLoad(t *testing.T) {
 		opts:            api.DefaultOptions(),
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
-		sessionDuration: 2,
+		sessionDuration: &api.Duration{Duration: 2 * time.Second},
 	}
 	// Fail to load model first
-	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
 		return nil, fmt.Errorf("something failed to load model blah")
 	}
 	gpus := gpu.GpuInfoList{}
-	s.load(req, ggml, gpus)
+	s.load(req, ggml, gpus, 0)
 	require.Empty(t, req.successCh)
 	require.Len(t, req.errCh, 1)
 	s.loadedMu.Lock()
@@ -61,10 +61,10 @@ func TestLoad(t *testing.T) {
 	require.Contains(t, err.Error(), "this model may be incompatible")
 
 	server := &mockLlm{estimatedVRAM: 10, estimatedVRAMByGPU: map[string]uint64{}}
-	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
 		return server, nil
 	}
-	s.load(req, ggml, gpus)
+	s.load(req, ggml, gpus, 0)
 	select {
 	case err := <-req.errCh:
 		require.NoError(t, err)
@@ -78,12 +78,12 @@ func TestLoad(t *testing.T) {
 
 	req.model.ModelPath = "dummy_model_path"
 	server.waitResp = fmt.Errorf("wait failure")
-	s.load(req, ggml, gpus)
+	s.load(req, ggml, gpus, 0)
 	select {
 	case err := <-req.errCh:
 		require.Contains(t, err.Error(), "wait failure")
 	case resp := <-req.successCh:
-		t.Errorf("unexpected success %v", resp)
+		t.Fatalf("unexpected success %v", resp)
 	}
 	s.loadedMu.Lock()
 	runner := s.loaded["dummy_model_path"]
@@ -94,7 +94,7 @@ func TestLoad(t *testing.T) {
 	require.Len(t, s.expiredCh, 1)
 }
 
-type bundle struct {
+type reqBundle struct {
 	ctx     context.Context //nolint:containedctx
 	ctxDone func()
 	srv     *mockLlm
@@ -102,13 +102,13 @@ type bundle struct {
 	ggml    *llm.GGML
 }
 
-func (scenario *bundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
+func (scenario *reqBundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
 	return scenario.srv, nil
 }
 
-func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedVRAM uint64) *bundle {
-	scenario := &bundle{}
-	scenario.ctx, scenario.ctxDone = context.WithCancel(ctx)
+func newScenarioRequest(t *testing.T, ctx context.Context, modelName string, estimatedVRAM uint64, duration *api.Duration) *reqBundle {
+	b := &reqBundle{}
+	b.ctx, b.ctxDone = context.WithCancel(ctx)
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), modelName)
@@ -128,185 +128,227 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 		"tokenizer.ggml.scores":         []float32{0},
 		"tokenizer.ggml.token_type":     []int32{0},
 	}, []llm.Tensor{
-		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
-		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
+		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
+		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
 	})
 	require.NoError(t, err)
 
 	fname := f.Name()
 	model := &Model{Name: modelName, ModelPath: fname}
-	scenario.ggml, err = llm.LoadModel(model.ModelPath)
+	b.ggml, err = llm.LoadModel(model.ModelPath, 0)
 	require.NoError(t, err)
 
-	scenario.req = &LlmRequest{
-		ctx:             scenario.ctx,
+	if duration == nil {
+		duration = &api.Duration{Duration: 5 * time.Millisecond}
+	}
+	b.req = &LlmRequest{
+		ctx:             b.ctx,
 		model:           model,
 		opts:            api.DefaultOptions(),
-		sessionDuration: 5 * time.Millisecond,
+		sessionDuration: duration,
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
 	}
-	scenario.srv = &mockLlm{estimatedVRAM: estimatedVRAM, estimatedVRAMByGPU: map[string]uint64{"": estimatedVRAM}}
-	return scenario
+	b.srv = &mockLlm{estimatedVRAM: estimatedVRAM, estimatedVRAMByGPU: map[string]uint64{"": estimatedVRAM}}
+	return b
 }
 
-func TestRequests(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 10*time.Second)
+func getGpuFn() gpu.GpuInfoList {
+	g := gpu.GpuInfo{Library: "metal"}
+	g.TotalMemory = 24 * format.GigaByte
+	g.FreeMemory = 12 * format.GigaByte
+	return []gpu.GpuInfo{g}
+}
+
+func getCpuFn() gpu.GpuInfoList {
+	g := gpu.GpuInfo{Library: "cpu"}
+	g.TotalMemory = 32 * format.GigaByte
+	g.FreeMemory = 26 * format.GigaByte
+	return []gpu.GpuInfo{g}
+}
+
+func TestRequestsSameModelSameRequest(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
 	defer done()
-
-	// Same model, same request
-	scenario1a := newScenario(t, ctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = 5 * time.Millisecond
-	scenario1b := newScenario(t, ctx, "ollama-model-1", 11)
-	scenario1b.req.model = scenario1a.req.model
-	scenario1b.ggml = scenario1a.ggml
-	scenario1b.req.sessionDuration = 0
-
-	// simple reload of same model
-	scenario2a := newScenario(t, ctx, "ollama-model-1", 20)
-	tmpModel := *scenario1a.req.model
-	scenario2a.req.model = &tmpModel
-	scenario2a.ggml = scenario1a.ggml
-	scenario2a.req.sessionDuration = 5 * time.Millisecond
-
-	// Multiple loaded models
-	scenario3a := newScenario(t, ctx, "ollama-model-3a", 1*format.GigaByte)
-	scenario3b := newScenario(t, ctx, "ollama-model-3b", 24*format.GigaByte)
-	scenario3c := newScenario(t, ctx, "ollama-model-4a", 30)
-	scenario3c.req.opts.NumGPU = 0                           // CPU load, will be allowed
-	scenario3d := newScenario(t, ctx, "ollama-model-3c", 30) // Needs prior unloaded
-
 	s := InitScheduler(ctx)
-	s.getGpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "metal"}
-		g.TotalMemory = 24 * format.GigaByte
-		g.FreeMemory = 12 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.getCpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "cpu"}
-		g.TotalMemory = 32 * format.GigaByte
-		g.FreeMemory = 26 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.newServerFn = scenario1a.newServer
-	slog.Info("scenario1a")
-	s.pendingReqCh <- scenario1a.req
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1", 11, &api.Duration{Duration: 0})
+	b.req.model = a.req.model
+	b.ggml = a.ggml
+
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
 	require.Len(t, s.pendingReqCh, 1)
 	s.Run(ctx)
 	select {
-	case resp := <-scenario1a.req.successCh:
-		require.Equal(t, resp.llama, scenario1a.srv)
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario1a.req.errCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 
 	// Same runner as first request due to not needing a reload
-	s.newServerFn = scenario1b.newServer
-	slog.Info("scenario1b")
-	s.pendingReqCh <- scenario1b.req
+	s.newServerFn = b.newServer
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	select {
-	case resp := <-scenario1b.req.successCh:
-		require.Equal(t, resp.llama, scenario1a.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario1b.req.errCh)
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
+	}
+}
+
+func TestRequestsSimpleReloadSameModel(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1", 20, &api.Duration{Duration: 5 * time.Millisecond})
+	tmpModel := *a.req.model
+	b.req.model = &tmpModel
+	b.ggml = a.ggml
+
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	select {
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
+		require.Empty(t, s.pendingReqCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
+	case <-ctx.Done():
+		t.Fatal("timeout")
 	}
 
 	// Trigger a reload
-	s.newServerFn = scenario2a.newServer
-	scenario2a.req.model.AdapterPaths = []string{"new"}
-	slog.Info("scenario2a")
-	s.pendingReqCh <- scenario2a.req
+	s.newServerFn = b.newServer
+	b.req.model.AdapterPaths = []string{"new"}
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	// finish first two requests, so model can reload
 	time.Sleep(1 * time.Millisecond)
-	scenario1a.ctxDone()
-	scenario1b.ctxDone()
+	a.ctxDone()
 	select {
-	case resp := <-scenario2a.req.successCh:
-		require.Equal(t, resp.llama, scenario2a.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, b.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario2a.req.errCh)
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
+}
+
+func TestRequestsMultipleLoadedModels(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+
+	// Multiple loaded models
+	a := newScenarioRequest(t, ctx, "ollama-model-3a", 1*format.GigaByte, nil)
+	b := newScenarioRequest(t, ctx, "ollama-model-3b", 24*format.GigaByte, nil)
+	c := newScenarioRequest(t, ctx, "ollama-model-4a", 30, nil)
+	c.req.opts.NumGPU = 0                                       // CPU load, will be allowed
+	d := newScenarioRequest(t, ctx, "ollama-model-3c", 30, nil) // Needs prior unloaded
 
 	envconfig.MaxRunners = 1
-	s.newServerFn = scenario3a.newServer
-	slog.Info("scenario3a")
-	s.pendingReqCh <- scenario3a.req
-	// finish prior request, so new model can load
-	time.Sleep(1 * time.Millisecond)
-	scenario2a.ctxDone()
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	s.Run(ctx)
 	select {
-	case resp := <-scenario3a.req.successCh:
-		require.Equal(t, resp.llama, scenario3a.srv)
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3a.req.errCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 1)
 	s.loadedMu.Unlock()
 
 	envconfig.MaxRunners = 0
-	s.newServerFn = scenario3b.newServer
-	slog.Info("scenario3b")
-	s.pendingReqCh <- scenario3b.req
+	s.newServerFn = b.newServer
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	select {
-	case resp := <-scenario3b.req.successCh:
-		require.Equal(t, resp.llama, scenario3b.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, b.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3b.req.errCh)
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 2)
 	s.loadedMu.Unlock()
 
 	// This is a CPU load with NumGPU = 0 so it should load
-	s.newServerFn = scenario3c.newServer
-	slog.Info("scenario3c")
-	s.pendingReqCh <- scenario3c.req
+	s.newServerFn = c.newServer
+	slog.Info("c")
+	s.pendingReqCh <- c.req
 	select {
-	case resp := <-scenario3c.req.successCh:
-		require.Equal(t, resp.llama, scenario3c.srv)
+	case resp := <-c.req.successCh:
+		require.Equal(t, resp.llama, c.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3c.req.errCh)
+		require.Empty(t, c.req.errCh)
+	case err := <-c.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 3)
 	s.loadedMu.Unlock()
 
 	// Try to load a model that wont fit
-	s.newServerFn = scenario3d.newServer
-	slog.Info("scenario3d")
+	s.newServerFn = d.newServer
+	slog.Info("d")
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 3)
 	s.loadedMu.Unlock()
-	scenario3a.ctxDone() // Won't help since this one isn't big enough to make room
+	a.ctxDone() // Won't help since this one isn't big enough to make room
 	time.Sleep(2 * time.Millisecond)
-	s.pendingReqCh <- scenario3d.req
+	s.pendingReqCh <- d.req
 	// finish prior request, so new model can load
 	time.Sleep(6 * time.Millisecond)
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 2)
 	s.loadedMu.Unlock()
-	scenario3b.ctxDone()
+	b.ctxDone()
 	select {
-	case resp := <-scenario3d.req.successCh:
-		require.Equal(t, resp.llama, scenario3d.srv)
+	case resp := <-d.req.successCh:
+		require.Equal(t, resp.llama, d.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3d.req.errCh)
+		require.Empty(t, d.req.errCh)
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 2)
@@ -317,26 +359,19 @@ func TestGetRunner(t *testing.T) {
 	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
 	defer done()
 
-	scenario1a := newScenario(t, ctx, "ollama-model-1a", 10)
-	scenario1a.req.sessionDuration = 0
-	scenario1b := newScenario(t, ctx, "ollama-model-1b", 10)
-	scenario1b.req.sessionDuration = 0
-	scenario1c := newScenario(t, ctx, "ollama-model-1c", 10)
-	scenario1c.req.sessionDuration = 0
+	a := newScenarioRequest(t, ctx, "ollama-model-1a", 10, &api.Duration{Duration: 2 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1b", 10, &api.Duration{Duration: 2 * time.Millisecond})
+	c := newScenarioRequest(t, ctx, "ollama-model-1c", 10, &api.Duration{Duration: 2 * time.Millisecond})
 	envconfig.MaxQueuedRequests = 1
 	s := InitScheduler(ctx)
-	s.getGpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "metal"}
-		g.TotalMemory = 24 * format.GigaByte
-		g.FreeMemory = 12 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.newServerFn = scenario1a.newServer
-	slog.Info("scenario1a")
-	successCh1a, errCh1a := s.GetRunner(scenario1a.ctx, scenario1a.req.model, scenario1a.req.opts, scenario1a.req.sessionDuration)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	successCh1a, errCh1a := s.GetRunner(a.ctx, a.req.model, a.req.opts, a.req.sessionDuration)
 	require.Len(t, s.pendingReqCh, 1)
-	slog.Info("scenario1b")
-	successCh1b, errCh1b := s.GetRunner(scenario1b.ctx, scenario1b.req.model, scenario1b.req.opts, scenario1b.req.sessionDuration)
+	slog.Info("b")
+	successCh1b, errCh1b := s.GetRunner(b.ctx, b.req.model, b.req.opts, b.req.sessionDuration)
 	require.Len(t, s.pendingReqCh, 1)
 	require.Empty(t, successCh1b)
 	require.Len(t, errCh1b, 1)
@@ -345,22 +380,24 @@ func TestGetRunner(t *testing.T) {
 	s.Run(ctx)
 	select {
 	case resp := <-successCh1a:
-		require.Equal(t, resp.llama, scenario1a.srv)
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, errCh1a)
+	case err := <-errCh1a:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
-	scenario1a.ctxDone()
+	a.ctxDone() // Set "a" model to idle so it can unload
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 1)
 	s.loadedMu.Unlock()
 
-	scenario1c.req.model.ModelPath = "bad path"
-	slog.Info("scenario1c")
-	successCh1c, errCh1c := s.GetRunner(scenario1c.ctx, scenario1c.req.model, scenario1c.req.opts, scenario1c.req.sessionDuration)
+	c.req.model.ModelPath = "bad path"
+	slog.Info("c")
+	successCh1c, errCh1c := s.GetRunner(c.ctx, c.req.model, c.req.opts, c.req.sessionDuration)
 	// Starts in pending channel, then should be quickly processsed to return an error
-	time.Sleep(5 * time.Millisecond)
+	time.Sleep(20 * time.Millisecond) // Long enough for the "a" model to expire and unload
 	require.Empty(t, successCh1c)
 	s.loadedMu.Lock()
 	require.Empty(t, s.loaded)
@@ -368,7 +405,7 @@ func TestGetRunner(t *testing.T) {
 	require.Len(t, errCh1c, 1)
 	err = <-errCh1c
 	require.Contains(t, err.Error(), "bad path")
-	scenario1b.ctxDone()
+	b.ctxDone()
 }
 
 // TODO - add one scenario that triggers the bogus finished event with positive ref count
@@ -377,7 +414,7 @@ func TestPrematureExpired(t *testing.T) {
 	defer done()
 
 	// Same model, same request
-	scenario1a := newScenario(t, ctx, "ollama-model-1a", 10)
+	scenario1a := newScenarioRequest(t, ctx, "ollama-model-1a", 10, nil)
 	s := InitScheduler(ctx)
 	s.getGpuFn = func() gpu.GpuInfoList {
 		g := gpu.GpuInfo{Library: "metal"}
@@ -399,10 +436,12 @@ func TestPrematureExpired(t *testing.T) {
 		s.loadedMu.Unlock()
 		slog.Info("sending premature expired event now")
 		s.expiredCh <- resp // Shouldn't happen in real life, but make sure its safe
+	case err := <-errCh1a:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
-	time.Sleep(scenario1a.req.sessionDuration)
+	time.Sleep(scenario1a.req.sessionDuration.Duration)
 	scenario1a.ctxDone()
 	time.Sleep(20 * time.Millisecond)
 	require.LessOrEqual(t, len(s.finishedReqCh), 1)
@@ -423,19 +462,21 @@ func TestUseLoadedRunner(t *testing.T) {
 		ctx:             ctx,
 		opts:            api.DefaultOptions(),
 		successCh:       make(chan *runnerRef, 1),
-		sessionDuration: 2,
+		sessionDuration: &api.Duration{Duration: 2},
 	}
 	finished := make(chan *LlmRequest)
 	llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}}
-	r1 := &runnerRef{llama: llm1, sessionDuration: 1}
+	r1 := &runnerRef{llama: llm1, sessionDuration: 1, numParallel: 1}
 	req.useLoadedRunner(r1, finished)
 	require.Equal(t, uint(1), r1.refCount)
 	require.Equal(t, time.Duration(2), r1.sessionDuration)
 	select {
 	case success := <-req.successCh:
 		require.Equal(t, r1, success)
+	case err := <-req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	done()
 	fin := <-finished
@@ -461,8 +502,8 @@ func TestUpdateFreeSpace(t *testing.T) {
 	gpus[1].FreeMemory = 1900
 	llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{"1": 50, "2": 50}}
 	llm2 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{"1": 125, "2": 75}}
-	r1 := &runnerRef{llama: llm1, gpus: gpus}
-	r2 := &runnerRef{llama: llm2, gpus: gpus}
+	r1 := &runnerRef{llama: llm1, gpus: gpus, numParallel: 1}
+	r2 := &runnerRef{llama: llm2, gpus: gpus, numParallel: 1}
 
 	s := InitScheduler(ctx)
 	s.loadedMu.Lock()
@@ -513,8 +554,8 @@ func TestFindRunnerToUnload(t *testing.T) {
 	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
 	defer done()
 
-	r1 := &runnerRef{refCount: 1, sessionDuration: 1}
-	r2 := &runnerRef{sessionDuration: 2}
+	r1 := &runnerRef{refCount: 1, sessionDuration: 1, numParallel: 1}
+	r2 := &runnerRef{sessionDuration: 2, numParallel: 1}
 
 	s := InitScheduler(ctx)
 	s.loadedMu.Lock()
@@ -536,9 +577,13 @@ func TestNeedsReload(t *testing.T) {
 	llm := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}}
 	do := api.DefaultOptions()
 	runner := &runnerRef{
-		model:   &Model{AdapterPaths: []string{"adapter1"}, ProjectorPaths: []string{"projector1"}},
-		Options: &do,
-		llama:   llm,
+		model: &Model{
+			AdapterPaths:   []string{"adapter1"},
+			ProjectorPaths: []string{"projector1"},
+		},
+		Options:     &do,
+		llama:       llm,
+		numParallel: 1,
 	}
 	req := &LlmRequest{
 		model: &Model{
@@ -581,8 +626,8 @@ func TestUnloadAllRunners(t *testing.T) {
 	s := InitScheduler(ctx)
 	s.unloadAllRunners()
 
-	r1 := &runnerRef{llama: llm1}
-	r2 := &runnerRef{llama: llm2}
+	r1 := &runnerRef{llama: llm1, numParallel: 1}
+	r2 := &runnerRef{llama: llm2, numParallel: 1}
 
 	s.loadedMu.Lock()
 	s.loaded["a"] = r1
@@ -596,20 +641,37 @@ func TestUnloadAllRunners(t *testing.T) {
 
 func TestUnload(t *testing.T) {
 	llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}}
-	r1 := &runnerRef{llama: llm1}
-	r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}}
+	r1 := &runnerRef{llama: llm1, numParallel: 1}
+	r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}, numParallel: 1}
 	r1.unload()
 	require.True(t, llm1.closeCalled)
 	r2.unload()
 	require.Nil(t, r2.model)
 }
 
+func TestAlreadyCanceled(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	dctx, done2 := context.WithCancel(ctx)
+	done2()
+	scenario1a := newScenarioRequest(t, dctx, "ollama-model-1", 10, &api.Duration{Duration: 0})
+	s := InitScheduler(ctx)
+	slog.Info("scenario1a")
+	s.pendingReqCh <- scenario1a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	time.Sleep(5 * time.Millisecond)
+	require.Empty(t, s.pendingReqCh)
+	require.Empty(t, scenario1a.req.errCh)
+	require.Empty(t, scenario1a.req.successCh)
+}
+
 type mockLlm struct {
 	pingResp           error
 	waitResp           error
 	completionResp     error
-	embeddingResp      []float64
-	embeddingRespErr   error
+	embedResp          [][]float32
+	embedRespErr       error
 	tokenizeResp       []int
 	tokenizeRespErr    error
 	detokenizeResp     string
@@ -626,8 +688,8 @@ func (s *mockLlm) WaitUntilRunning(ctx context.Context) error { return s.waitRes
 func (s *mockLlm) Completion(ctx context.Context, req llm.CompletionRequest, fn func(llm.CompletionResponse)) error {
 	return s.completionResp
 }
-func (s *mockLlm) Embedding(ctx context.Context, prompt string) ([]float64, error) {
-	return s.embeddingResp, s.embeddingRespErr
+func (s *mockLlm) Embed(ctx context.Context, input []string) ([][]float32, error) {
+	return s.embedResp, s.embedRespErr
 }
 func (s *mockLlm) Tokenize(ctx context.Context, content string) ([]int, error) {
 	return s.tokenizeResp, s.tokenizeRespErr

server/testdata/tools/command-r-plus.gotmpl

@@ -0,0 +1,67 @@
+{{- if or .Tools .System }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>
+{{- if .Tools }}# Safety Preamble
+The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
+
+# System Preamble
+## Basic Rules
+You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
+
+{{ if .System }}# User Preamble
+{{ .System }}
+{{- end }}
+
+## Available Tools
+Here is a list of tools that you have available to you:
+{{- range .Tools }}
+
+```python
+def {{ .Function.Name }}(
+{{- range $name, $property := .Function.Parameters.Properties }}{{ $name }}: {{ $property.Type }}, {{ end }}) -> List[Dict]:
+    """{{ .Function.Description }}
+
+{{- if .Function.Parameters.Properties }}
+
+    Args:
+{{- range $name, $property := .Function.Parameters.Properties }}
+        {{ $name }} ({{ $property.Type }}): {{ $property.Description }}
+{{- end }}
+{{- end }}
+    """
+    pass
+```
+{{- end }}
+{{- else if .System }}{{ .System }}
+{{- end }}<|END_OF_TURN_TOKEN|>
+{{- end }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- continue }}
+{{- end }}<|START_OF_TURN_TOKEN|>
+{{- if eq .Role "user" }}<|USER_TOKEN|>{{ .Content }}
+{{- else if eq .Role "assistant" }}<|CHATBOT_TOKEN|>
+{{- if .Content }}{{ .Content }}
+{{- else if .ToolCalls }}
+Action: ```json
+[
+{{- range .ToolCalls }}
+    {
+        "tool_name": "{{ .Function.Name }}",
+        "parameters": {{ .Function.Arguments }}
+    }
+{{- end }}
+]```
+{{ continue }}
+{{ end }}
+{{- else if eq .Role "tool" }}<|SYSTEM_TOKEN|><results>
+{{ .Content }}</results>
+{{- end }}<|END_OF_TURN_TOKEN|>
+{{- end }}
+{{- if .Tools }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
+```json
+[
+    {
+        "tool_name": title of the tool in the specification,
+        "parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
+    }
+]```
+{{- end }}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

server/testdata/tools/command-r-plus.out

@@ -0,0 +1,39 @@
+<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|># Safety Preamble
+The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
+
+# System Preamble
+## Basic Rules
+You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
+
+# User Preamble
+You are a knowledgable assistant. You can answer questions and perform tasks.
+
+## Available Tools
+Here is a list of tools that you have available to you:
+
+```python
+def get_current_weather(format: string, location: string, ) -> List[Dict]:
+    """Get the current weather
+
+    Args:
+        format (string): The temperature unit to use. Infer this from the users location.
+        location (string): The city and state, e.g. San Francisco, CA
+    """
+    pass
+```<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>What's the weather like today in Paris?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
+Action: ```json
+[
+    {
+        "tool_name": "get_current_weather",
+        "parameters": {"format":"celsius","location":"Paris, France"}
+    }
+]```
+<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|><results>
+22</results><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>The current temperature in Paris, France is 22 degrees Celsius.<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>What's the weather like today in San Francisco and Toronto?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
+```json
+[
+    {
+        "tool_name": title of the tool in the specification,
+        "parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
+    }
+]```<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

server/testdata/tools/firefunction.gotmpl

@@ -0,0 +1,31 @@
+{{- if or .System .Tools }}<|start_header_id|>system<|end_header_id|>
+{{- if .System }}
+{{ .System }}
+{{- end }}
+In addition to plain text responses, you can chose to call one or more of the provided functions.
+
+Use the following rule to decide when to call a function:
+  * if the response can be generated from your internal knowledge (e.g., as in the case of queries like "What is the capital of Poland?"), do so
+  * if you need external information that can be obtained by calling one or more of the provided functions, generate a function calls
+
+If you decide to call functions:
+  * prefix function calls with functools marker (no closing marker required)
+  * all function calls should be generated in a single JSON list formatted as functools[{"name": [function name], "arguments": [function arguments as JSON]}, ...]
+  * follow the provided JSON schema. Do not hallucinate arguments or values. Do to blindly copy values from the provided samples
+  * respect the argument type formatting. E.g., if the type if number and format is float, write value 7 as 7.0
+  * make sure you pick the right functions that match the user intent
+
+Available functions as JSON spec:
+{{- if .Tools }}
+{{ .Tools }}
+{{- end }}<|eot_id|>
+{{- end }}
+{{- range .Messages }}<|start_header_id|>
+{{- if or (eq .Role "user") (eq .Role "assistant") (eq .Role "tool") }}{{ .Role }}
+{{- end }}<|end_header_id|>
+{{- if .Content }}{{ .Content }}
+{{- else if .ToolCalls }} functools[
+{{- range .ToolCalls }}{{ "{" }}"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}{{ "}" }}
+{{- end }}]
+{{- end }}<|eot_id|>
+{{- end }}<|start_header_id|>assistant<|end_header_id|>

server/testdata/tools/firefunction.out

@@ -0,0 +1,17 @@
+<|start_header_id|>system<|end_header_id|>
+You are a knowledgable assistant. You can answer questions and perform tasks.
+In addition to plain text responses, you can chose to call one or more of the provided functions.
+
+Use the following rule to decide when to call a function:
+  * if the response can be generated from your internal knowledge (e.g., as in the case of queries like "What is the capital of Poland?"), do so
+  * if you need external information that can be obtained by calling one or more of the provided functions, generate a function calls
+
+If you decide to call functions:
+  * prefix function calls with functools marker (no closing marker required)
+  * all function calls should be generated in a single JSON list formatted as functools[{"name": [function name], "arguments": [function arguments as JSON]}, ...]
+  * follow the provided JSON schema. Do not hallucinate arguments or values. Do to blindly copy values from the provided samples
+  * respect the argument type formatting. E.g., if the type if number and format is float, write value 7 as 7.0
+  * make sure you pick the right functions that match the user intent
+
+Available functions as JSON spec:
+[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}}]<|eot_id|><|start_header_id|><|end_header_id|>You are a knowledgable assistant. You can answer questions and perform tasks.<|eot_id|><|start_header_id|>user<|end_header_id|>What's the weather like today in Paris?<|eot_id|><|start_header_id|>assistant<|end_header_id|> functools[{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]<|eot_id|><|start_header_id|>tool<|end_header_id|>22<|eot_id|><|start_header_id|>assistant<|end_header_id|>The current temperature in Paris, France is 22 degrees Celsius.<|eot_id|><|start_header_id|>user<|end_header_id|>What's the weather like today in San Francisco and Toronto?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

server/testdata/tools/llama3-groq-tool-use.gotmpl

@@ -0,0 +1,43 @@
+{{- if .Messages }}
+{{- if or .System .Tools }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}
+{{- if .Tools }} You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
+<tool_call>
+{"name": <function-name>,"arguments": <args-dict>}
+</tool_call>
+
+Here are the available tools:
+<tools>
+{{- range .Tools }} {{ .Function }}
+{{- end }} </tools>
+{{- end }}
+{{- end }}<|eot_id|>
+{{- range .Messages }}
+{{- if ne .Role "system" }}<|start_header_id|>{{ .Role }}<|end_header_id|>
+
+{{ if eq .Role "user" }}{{ .Content }}
+{{- else if eq .Role "assistant" }}
+{{- if .Content }}{{ .Content }}
+{{- else if .ToolCalls }}<tool_call>
+{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
+{{- end }}
+</tool_call>
+{{- end }}
+{{- else if eq .Role "tool" }}<tool_response>
+{{ .Content }}
+</tool_response>
+{{- end }}<|eot_id|>
+{{- end }}
+{{- end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ else }}
+{{ if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
+
+{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ end }}{{ .Response }}
+{{- if .Response }}<|eot_id|>
+{{- end }}

server/testdata/tools/llama3-groq-tool-use.out

@@ -0,0 +1,24 @@
+<|start_header_id|>system<|end_header_id|>
+
+You are a knowledgable assistant. You can answer questions and perform tasks. You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
+<tool_call>
+{"name": <function-name>,"arguments": <args-dict>}
+</tool_call>
+
+Here are the available tools:
+<tools> {"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}} </tools><|eot_id|><|start_header_id|>user<|end_header_id|>
+
+What's the weather like today in Paris?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+<tool_call>
+{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}
+</tool_call><|eot_id|><|start_header_id|>tool<|end_header_id|>
+
+<tool_response>
+22
+</tool_response><|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+The current temperature in Paris, France is 22 degrees Celsius.<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+What's the weather like today in San Francisco and Toronto?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+