llm: consider head_dim in llama arch (#5817)

Adjust windows ROCm discovery

.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"
@@ -304,11 +304,6 @@ jobs:
           write-host "Installing plugin"
           & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
           write-host "plugin installed"
-      - name: remove unwanted mingw dll.a files
-        run: |
-          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libpthread.dll.a" -File | Remove-Item -Force
-          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libwinpthread.dll.a" -File | Remove-Item -Force
-          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libstdc++.dll.a" -File | Remove-Item -Force
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod

.github/workflows/test.yaml

@@ -126,7 +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:
@@ -169,7 +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

README.md

@@ -293,6 +293,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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
@@ -173,6 +239,30 @@ type Runner struct {
 	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].
 type EmbeddingRequest struct {
 	// Model is the model name.
@@ -219,8 +309,10 @@ 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"`
 

api/types_test.go

@@ -208,3 +208,26 @@ func TestUseMmapFormatParams(t *testing.T) {
 		})
 	}
 }
+
+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/ollama.iss

@@ -127,6 +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

@@ -843,7 +843,6 @@ type runOptions struct {
 	WordWrap    bool
 	Format      string
 	System      string
-	Template    string
 	Images      []api.ImageData
 	Options     map[string]interface{}
 	MultiModal  bool
@@ -1037,7 +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,
 	}

cmd/interactive.go

@@ -27,7 +27,6 @@ const (
 	MultilineNone MultilineState = iota
 	MultilinePrompt
 	MultilineSystem
-	MultilineTemplate
 )
 
 func loadModel(cmd *cobra.Command, opts *runOptions) error {
@@ -94,7 +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")
@@ -204,10 +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
@@ -326,17 +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, `"""`)
@@ -356,23 +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
@@ -393,7 +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)
@@ -437,12 +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:
@@ -536,10 +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]

docs/faq.md

@@ -272,4 +272,4 @@ The following server settings may be used to adjust how Ollama handles concurren
 - `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 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.
+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

@@ -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/openai.md

@@ -103,10 +103,6 @@ curl http://localhost:11434/v1/chat/completions \
 - [ ] `user`
 - [ ] `n`
 
-#### Notes
-
-- `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached
-
 ## Models
 
 Before using a model, pull it locally `ollama pull`:

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")
@@ -132,10 +132,8 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 				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/gpu.go

@@ -274,6 +274,28 @@ 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)
 			}
@@ -338,14 +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
@@ -374,9 +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),

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_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_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
@@ -52,7 +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/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/server.cpp

@@ -1413,7 +1413,7 @@ struct llama_server_context
             return get_slot(-1);
         }
 
-        LOG_INFO("slot with common prefix found", {{
+        LOG_DEBUG("slot with common prefix found", {{
             "slot_id", slot->id,
             "characters", longest
         }});
@@ -1688,22 +1688,8 @@ struct llama_server_context
                     }
                     slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep);
 
-                    char buf[256];
-                    llama_model_meta_val_str(model, "general.architecture", buf, 256);
-                    bool gemma2 = strcmp(buf, "gemma2") == 0;
-
-                    int32_t truncate_at = slot.n_ctx;
-
-                    // truncate at 2/3 of the context length for gemma2 models
-                    // as they do not support context shifts (from the sliding window implementation).
-                    // this way, prompts that almost fit the context length can still generate a full
-                    // response without a sudden stop from hitting the context limit
-                    if (gemma2) {
-                        truncate_at = 2 * slot.n_ctx / 3;
-                    }
-
                     // if input prompt is too big, truncate it, if group attention self-extend is disabled
-                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= truncate_at)
+                    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_shift = n_left / 2;
@@ -1731,19 +1717,6 @@ struct llama_server_context
                         GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
                     }
 
-                    // Models with sliding window attention do not work with context shifts, so
-                    // limit their prediction to the context length
-                    if (gemma2) {
-                        int32_t limit = slot.n_ctx - slot.n_prompt_tokens;
-                        slot.n_predict = limit;
-                        slot.params.n_predict = limit;
-                        LOG_INFO("model does not support sliding window, limiting generation", {
-                            {"n_ctx", slot.n_ctx},
-                            {"n_prompt_tokens", slot.n_prompt_tokens},
-                            {"n_predict", slot.n_predict}
-                        });
-                    }
-
                     if (!slot.params.cache_prompt)
                     {
                         llama_sampling_reset(slot.ctx_sampling);
@@ -3215,26 +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_linux.sh

@@ -77,7 +77,7 @@ 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
@@ -93,7 +93,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
         # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
         # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
 
-        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=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)
@@ -178,7 +178,7 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
         CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
         echo "Building custom CUDA GPU"
     else
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DGGML_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} -DCMAKE_LIBRARY_PATH=/usr/local/cuda/compat"
+        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${CUDA_VARIANT}"
@@ -254,7 +254,7 @@ if [ -z "${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} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DCMAKE_C_COMPILER=$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

@@ -6,18 +6,9 @@ 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"
@@ -366,6 +357,7 @@ function build_rocm() {
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
             "-DGGML_HIPBLAS=on",
+            "-DLLAMA_CUDA_NO_PEER_COPY=on",
             "-DHIP_PLATFORM=amd",
             "-DGGML_AVX=on",
             "-DGGML_AVX2=off",
@@ -394,7 +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\"

llm/ggml.go

@@ -424,6 +424,32 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			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/gguf.go

@@ -537,6 +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

@@ -4,8 +4,8 @@ package llm
 // #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: -L${SRCDIR}/build/windows/amd64_static -L${SRCDIR}/build/windows/amd64_static/src -L${SRCDIR}/build/windows/amd64_static/ggml/src
-// #cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_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>
@@ -33,7 +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/patches/05-default-pretokenizer.diff

@@ -1,11 +1,11 @@
 diff --git a/src/llama.cpp b/src/llama.cpp
-index 73f52435..2b81b4bd 100644
+index 2b9ace28..172640e2 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -5092,16 +5092,7 @@ static void llm_load_vocab(
- 
-         // for now, only BPE models have pre-tokenizers
+@@ -5357,16 +5357,7 @@ 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__);
@@ -20,7 +20,7 @@ index 73f52435..2b81b4bd 100644
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -5164,7 +5155,8 @@ static void llm_load_vocab(
+@@ -5439,7 +5430,8 @@ static void llm_load_vocab(
                  tokenizer_pre == "jais") {
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
              } else {

llm/patches/10-tekken.diff

@@ -0,0 +1,43 @@
+diff --git a/include/llama.h b/include/llama.h
+index bb4b05ba..a92174e0 100644
+--- a/include/llama.h
++++ b/include/llama.h
+@@ -92,6 +92,7 @@ extern "C" {
+         LLAMA_VOCAB_PRE_TYPE_CHATGLM4       = 17,
+         LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,
+         LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,
++        LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
+     };
+ 
+     // note: these values should be synchronized with ggml_rope
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 18364976..435b6fe5 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -5429,6 +5429,12 @@ static void llm_load_vocab(
+             } else if (
+                 tokenizer_pre == "jais") {
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
++            } else if (
++                tokenizer_pre == "tekken") {
++                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_TEKKEN;
++                vocab.tokenizer_clean_spaces = false;
++                vocab.tokenizer_ignore_merges = true;
++                vocab.tokenizer_add_bos = true;
+             } else {
+                 LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+@@ -15448,6 +15454,13 @@ struct llm_tokenizer_bpe {
+                     " ?[^(\\s|.,!?…。，、।۔،)]+",
+                 };
+                 break;
++            case LLAMA_VOCAB_PRE_TYPE_TEKKEN:
++                    // original regex from tokenizer.json
++                    // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
++                regex_exprs = {
++                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
++                };
++                break;
+             default:
+                 // default regex for BPE tokenization pre-processing
+                 regex_exprs = {

llm/patches/11-embd_kv.diff

@@ -0,0 +1,19 @@
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 2b9ace28..e60d3d8d 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -6052,10 +6052,10 @@ static bool llm_load_tensors(
+ 
+                         layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+ 
+-                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+-                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+-                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+-                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
++                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd,  n_embd_head_k * n_head});
++                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
++                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
++                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+ 
+                         // optional bias tensors
+                         layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);

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
@@ -88,6 +88,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	var estimate MemoryEstimate
 	var systemTotalMemory uint64
 	var systemFreeMemory uint64
+	var systemSwapFreeMemory uint64
 
 	systemMemInfo, err := gpu.GetCPUMem()
 	if err != nil {
@@ -95,7 +96,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	} else {
 		systemTotalMemory = systemMemInfo.TotalMemory
 		systemFreeMemory = systemMemInfo.FreeMemory
-		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", systemFreeMemory)
+		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
@@ -122,6 +124,16 @@ 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
@@ -254,10 +266,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		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 == "" {
@@ -377,8 +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=") ||
@@ -679,7 +689,7 @@ type CompletionRequest struct {
 	Prompt  string
 	Format  string
 	Images  []ImageData
-	Options api.Options
+	Options *api.Options
 }
 
 type CompletionResponse struct {
@@ -699,10 +709,9 @@ 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{
@@ -860,15 +869,15 @@ 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
@@ -883,7 +892,7 @@ 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)
 	}
@@ -910,7 +919,7 @@ 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)
 	}

openai/openai.go

@@ -3,11 +3,14 @@ package openai
 
 import (
 	"bytes"
+	"encoding/base64"
 	"encoding/json"
 	"fmt"
 	"io"
+	"log/slog"
 	"math/rand"
 	"net/http"
+	"strings"
 	"time"
 
 	"github.com/gin-gonic/gin"
@@ -27,8 +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 {
@@ -59,6 +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"`
@@ -71,6 +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 {
@@ -104,6 +114,7 @@ type CompletionRequest struct {
 	Stream           bool     `json:"stream"`
 	Temperature      *float32 `json:"temperature"`
 	TopP             float32  `json:"top_p"`
+	Suffix           string   `json:"suffix"`
 }
 
 type Completion struct {
@@ -125,6 +136,15 @@ type CompletionChunk struct {
 	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"`
@@ -132,11 +152,23 @@ type Model struct {
 	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 {
@@ -151,7 +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",
@@ -160,7 +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
@@ -169,7 +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,
@@ -215,7 +270,6 @@ func toCompletion(id string, r api.GenerateResponse) Completion {
 			}(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,
@@ -260,6 +314,27 @@ func toListCompletion(r api.ListResponse) ListCompletion {
 	}
 }
 
+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,
@@ -269,10 +344,77 @@ func toModel(r api.ShowResponse, m string) Model {
 	}
 }
 
-func fromChatRequest(r ChatCompletionRequest) api.ChatRequest {
+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{})
@@ -323,13 +465,14 @@ func fromChatRequest(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
 }
 
 func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
@@ -338,12 +481,16 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 	switch stop := r.Stop.(type) {
 	case string:
 		options["stop"] = []string{stop}
-	case []string:
-		options["stop"] = stop
-	default:
-		if r.Stop != nil {
-			return api.GenerateRequest{}, fmt.Errorf("invalid type for 'stop' field: %T", r.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 {
@@ -375,6 +522,7 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 		Prompt:  r.Prompt,
 		Options: options,
 		Stream:  &r.Stream,
+		Suffix:  r.Suffix,
 	}, nil
 }
 
@@ -403,6 +551,11 @@ type RetrieveWriter struct {
 	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)
@@ -568,6 +721,33 @@ func (w *RetrieveWriter) Write(data []byte) (int, error) {
 	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{
@@ -631,6 +811,47 @@ func CompletionsMiddleware() gin.HandlerFunc {
 			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()
@@ -652,7 +873,14 @@ func ChatMiddleware() gin.HandlerFunc {
 		}
 
 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(fromChatRequest(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
 		}

openai/openai_test.go

@@ -2,8 +2,8 @@ package openai
 
 import (
 	"bytes"
+	"encoding/base64"
 	"encoding/json"
-	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -16,7 +16,387 @@ import (
 	"github.com/stretchr/testify/assert"
 )
 
-func TestMiddleware(t *testing.T) {
+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
@@ -29,188 +409,6 @@ func TestMiddleware(t *testing.T) {
 	}
 
 	testCases := []testCase{
-		{
-			Name:     "chat handler",
-			Method:   http.MethodPost,
-			Path:     "/api/chat",
-			TestPath: "/api/chat",
-			Handler:  ChatMiddleware,
-			Endpoint: func(c *gin.Context) {
-				var chatReq api.ChatRequest
-				if err := c.ShouldBindJSON(&chatReq); err != nil {
-					c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"})
-					return
-				}
-
-				userMessage := chatReq.Messages[0].Content
-				var assistantMessage string
-
-				switch userMessage {
-				case "Hello":
-					assistantMessage = "Hello!"
-				default:
-					assistantMessage = "I'm not sure how to respond to that."
-				}
-
-				c.JSON(http.StatusOK, api.ChatResponse{
-					Message: api.Message{
-						Role:    "assistant",
-						Content: assistantMessage,
-					},
-				})
-			},
-			Setup: func(t *testing.T, req *http.Request) {
-				body := ChatCompletionRequest{
-					Model:    "test-model",
-					Messages: []Message{{Role: "user", Content: "Hello"}},
-				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
-				assert.Equal(t, http.StatusOK, resp.Code)
-
-				var chatResp ChatCompletion
-				if err := json.NewDecoder(resp.Body).Decode(&chatResp); err != nil {
-					t.Fatal(err)
-				}
-
-				if chatResp.Object != "chat.completion" {
-					t.Fatalf("expected chat.completion, got %s", chatResp.Object)
-				}
-
-				if chatResp.Choices[0].Message.Content != "Hello!" {
-					t.Fatalf("expected Hello!, got %s", chatResp.Choices[0].Message.Content)
-				}
-			},
-		},
-		{
-			Name:     "completions handler",
-			Method:   http.MethodPost,
-			Path:     "/api/generate",
-			TestPath: "/api/generate",
-			Handler:  CompletionsMiddleware,
-			Endpoint: func(c *gin.Context) {
-				c.JSON(http.StatusOK, api.GenerateResponse{
-					Response: "Hello!",
-				})
-			},
-			Setup: func(t *testing.T, req *http.Request) {
-				body := CompletionRequest{
-					Model:  "test-model",
-					Prompt: "Hello",
-				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
-				assert.Equal(t, http.StatusOK, resp.Code)
-				var completionResp Completion
-				if err := json.NewDecoder(resp.Body).Decode(&completionResp); err != nil {
-					t.Fatal(err)
-				}
-
-				if completionResp.Object != "text_completion" {
-					t.Fatalf("expected text_completion, got %s", completionResp.Object)
-				}
-
-				if completionResp.Choices[0].Text != "Hello!" {
-					t.Fatalf("expected Hello!, got %s", completionResp.Choices[0].Text)
-				}
-			},
-		},
-		{
-			Name:     "completions handler with params",
-			Method:   http.MethodPost,
-			Path:     "/api/generate",
-			TestPath: "/api/generate",
-			Handler:  CompletionsMiddleware,
-			Endpoint: func(c *gin.Context) {
-				var generateReq api.GenerateRequest
-				if err := c.ShouldBindJSON(&generateReq); err != nil {
-					c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"})
-					return
-				}
-
-				temperature := generateReq.Options["temperature"].(float64)
-				var assistantMessage string
-
-				switch temperature {
-				case 1.6:
-					assistantMessage = "Received temperature of 1.6"
-				default:
-					assistantMessage = fmt.Sprintf("Received temperature of %f", temperature)
-				}
-
-				c.JSON(http.StatusOK, api.GenerateResponse{
-					Response: assistantMessage,
-				})
-			},
-			Setup: func(t *testing.T, req *http.Request) {
-				temp := float32(0.8)
-				body := CompletionRequest{
-					Model:       "test-model",
-					Prompt:      "Hello",
-					Temperature: &temp,
-				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
-				assert.Equal(t, http.StatusOK, resp.Code)
-				var completionResp Completion
-				if err := json.NewDecoder(resp.Body).Decode(&completionResp); err != nil {
-					t.Fatal(err)
-				}
-
-				if completionResp.Object != "text_completion" {
-					t.Fatalf("expected text_completion, got %s", completionResp.Object)
-				}
-
-				if completionResp.Choices[0].Text != "Received temperature of 1.6" {
-					t.Fatalf("expected Received temperature of 1.6, got %s", completionResp.Choices[0].Text)
-				}
-			},
-		},
-		{
-			Name:     "completions handler with error",
-			Method:   http.MethodPost,
-			Path:     "/api/generate",
-			TestPath: "/api/generate",
-			Handler:  CompletionsMiddleware,
-			Endpoint: func(c *gin.Context) {
-				c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"})
-			},
-			Setup: func(t *testing.T, req *http.Request) {
-				body := CompletionRequest{
-					Model:  "test-model",
-					Prompt: "Hello",
-				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
-				if resp.Code != http.StatusBadRequest {
-					t.Fatalf("expected 400, got %d", resp.Code)
-				}
-
-				if !strings.Contains(resp.Body.String(), `"invalid request"`) {
-					t.Fatalf("error was not forwarded")
-				}
-			},
-		},
 		{
 			Name:     "list handler",
 			Method:   http.MethodGet,
@@ -227,8 +425,6 @@ func TestMiddleware(t *testing.T) {
 				})
 			},
 			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
-				assert.Equal(t, http.StatusOK, resp.Code)
-
 				var listResp ListCompletion
 				if err := json.NewDecoder(resp.Body).Decode(&listResp); err != nil {
 					t.Fatal(err)
@@ -292,6 +488,8 @@ func TestMiddleware(t *testing.T) {
 			resp := httptest.NewRecorder()
 			router.ServeHTTP(resp, req)
 
+			assert.Equal(t, http.StatusOK, resp.Code)
+
 			tc.Expected(t, resp)
 		})
 	}

scripts/build_windows.ps1

@@ -107,9 +107,12 @@ function gatherDependencies() {
 
     # TODO - this varies based on host build system and MSVC version - drive from dumpbin output
     # currently works for Win11 + MSVC 2019 + Cuda V11
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140.dll" "${script:DEPS_DIR}\ollama_runners\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140*.dll" "${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\"

server/images.go

@@ -34,9 +34,20 @@ import (
 	"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")
+const (
+	CapabilityCompletion = Capability("completion")
+	CapabilityTools      = Capability("tools")
+	CapabilityInsert     = Capability("insert")
+)
 
 type registryOptions struct {
 	Insecure bool
@@ -62,7 +73,10 @@ type Model struct {
 	Template *template.Template
 }
 
-func (m *Model) Has(caps ...Capability) bool {
+// 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:
@@ -81,15 +95,28 @@ func (m *Model) Has(caps ...Capability) bool {
 			}
 
 			if _, ok := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]; ok {
-				return false
+				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 false
+			return fmt.Errorf("unknown capability: %s", cap)
 		}
 	}
 
-	return true
+	if err := errors.Join(errs...); err != nil {
+		return fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
+	}
+
+	return nil
 }
 
 func (m *Model) String() string {
@@ -465,6 +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 {

server/model.go

@@ -4,6 +4,7 @@ import (
 	"archive/zip"
 	"bytes"
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -11,6 +12,9 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
+	"slices"
+	"strings"
+	"text/template/parse"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/convert"
@@ -289,3 +293,92 @@ 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
+		}
+	}
+
+	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())
+			objs = append(objs, obj)
+		}
+	}
+
+	var toolCalls []api.ToolCall
+	for _, kv := range objs {
+		var call api.ToolCall
+		for k, v := range kv {
+			switch k {
+			case name:
+				call.Function.Name = v.(string)
+			case arguments:
+				call.Function.Arguments = v.(map[string]any)
+			}
+		}
+
+		toolCalls = append(toolCalls, call)
+	}
+
+	return toolCalls, len(toolCalls) > 0
+}

server/model_test.go

@@ -3,7 +3,9 @@ package server
 import (
 	"archive/zip"
 	"bytes"
+	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"os"
 	"path/filepath"
@@ -11,7 +13,9 @@ import (
 	"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 {
@@ -110,3 +114,122 @@ func TestExtractFromZipFile(t *testing.T) {
 		})
 	}
 }
+
+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},
+	}
+
+	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/prompt.go

@@ -1,217 +1,74 @@
 package server
 
 import (
-	"fmt"
+	"bytes"
+	"context"
 	"log/slog"
-	"strings"
-
-	"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 *template.Template, system, prompt, response string, generate bool) (string, error) {
-	formatTemplateForResponse(tmpl, generate)
-
-	vars := map[string]any{
-		"System":   system,
-		"Prompt":   prompt,
-		"Response": response,
-	}
-
-	var sb strings.Builder
-	if err := tmpl.Execute(&sb, vars); err != nil {
-		return "", err
-	}
-
-	return sb.String(), nil
-}
-
-func countTokens(tmpl *template.Template, 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 *template.Template, 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

@@ -1,215 +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) {
-			tmpl, err := template.Parse(tc.template)
-			if err != nil {
-				t.Fatal(err)
-			}
-
-			got, err := Prompt(tmpl, 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) {
-			tmpl, err := template.Parse(tc.template)
+	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.Fatal(err)
 			}
 
-			got, err := ChatPrompt(tmpl, tc.messages, tc.window, encode)
-			if err != nil {
-				t.Errorf("error = %v", err)
+			if diff := cmp.Diff(prompt, tt.prompt); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
 			}
 
-			if got != tc.want {
-				t.Errorf("got: %q, want: %q", got, tc.want)
+			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.go

@@ -1,14 +1,15 @@
 package server
 
 import (
+	"bytes"
 	"cmp"
 	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
-	"io/fs"
 	"log/slog"
+	"math"
 	"net"
 	"net/http"
 	"net/netip"
@@ -54,6 +55,9 @@ func init() {
 	gin.SetMode(mode)
 }
 
+var errRequired = errors.New("is required")
+var errBadTemplate = errors.New("template error")
+
 func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options, error) {
 	opts := api.DefaultOptions()
 	if err := opts.FromMap(model.Options); err != nil {
@@ -67,250 +71,220 @@ func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options
 	return opts, nil
 }
 
-func isSupportedImageType(image []byte) bool {
-	contentType := http.DetectContentType(image)
-	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png"}
-	return slices.Contains(allowedTypes, contentType)
+// scheduleRunner schedules a runner after validating inputs such as capabilities and model options.
+// It returns the allocated runner, model instance, and consolidated options if successful and error otherwise.
+func (s *Server) scheduleRunner(ctx context.Context, name string, caps []Capability, requestOpts map[string]any, keepAlive *api.Duration) (llm.LlamaServer, *Model, *api.Options, error) {
+	if name == "" {
+		return nil, nil, nil, fmt.Errorf("model %w", errRequired)
+	}
+
+	model, err := GetModel(name)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+
+	if err := model.CheckCapabilities(caps...); err != nil {
+		return nil, nil, nil, fmt.Errorf("%s %w", name, err)
+	}
+
+	opts, err := modelOptions(model, requestOpts)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+
+	runnerCh, errCh := s.sched.GetRunner(ctx, model, opts, keepAlive)
+	var runner *runnerRef
+	select {
+	case runner = <-runnerCh:
+	case err = <-errCh:
+		return nil, nil, nil, err
+	}
+
+	return runner.llama, model, &opts, nil
 }
 
 func (s *Server) GenerateHandler(c *gin.Context) {
 	checkpointStart := time.Now()
 	var req api.GenerateRequest
-	err := c.ShouldBindJSON(&req)
-
-	switch {
-	case errors.Is(err, io.EOF):
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
 		return
-	case err != nil:
+	} else if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
-	// validate the request
-	switch {
-	case req.Model == "":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
+	if req.Format != "" && req.Format != "json" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be empty or \"json\""})
 		return
-	case len(req.Format) > 0 && req.Format != "json":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be json"})
-		return
-	case req.Raw && (req.Template != "" || req.System != "" || len(req.Context) > 0):
+	} else if req.Raw && (req.Template != "" || req.System != "" || len(req.Context) > 0) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "raw mode does not support template, system, or context"})
 		return
 	}
 
-	for _, img := range req.Images {
-		if !isSupportedImageType(img) {
-			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "unsupported image format"})
-			return
-		}
+	caps := []Capability{CapabilityCompletion}
+	if req.Suffix != "" {
+		caps = append(caps, CapabilityInsert)
 	}
 
-	model, err := GetModel(req.Model)
-	if err != nil {
-		var pErr *fs.PathError
-		if errors.As(err, &pErr) {
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
-			return
-		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+	r, m, opts, err := s.scheduleRunner(c.Request.Context(), req.Model, caps, req.Options, req.KeepAlive)
+	if errors.Is(err, errCapabilityCompletion) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%q does not support generate", req.Model)})
+		return
+	} else if err != nil {
+		handleScheduleError(c, req.Model, err)
 		return
 	}
 
-	if !model.Has(CapabilityCompletion) {
-		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%s does not support generate", req.Model)})
-		return
-	}
+	checkpointLoaded := time.Now()
 
-	opts, err := modelOptions(model, req.Options)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
-	var runner *runnerRef
-	select {
-	case runner = <-rCh:
-	case err = <-eCh:
-		handleErrorResponse(c, err)
-		return
-	}
-
-	// an empty request loads the model
-	// note: for a short while template was used in lieu
-	// of `raw` mode so we need to check for it too
-	if req.Prompt == "" && req.Template == "" && req.System == "" {
+	if req.Prompt == "" {
 		c.JSON(http.StatusOK, api.GenerateResponse{
-			CreatedAt:  time.Now().UTC(),
 			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
 			Done:       true,
 			DoneReason: "load",
 		})
 		return
 	}
 
-	tmpl, err := template.Parse(req.Template)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
+	images := make([]llm.ImageData, len(req.Images))
+	for i := range req.Images {
+		images[i] = llm.ImageData{ID: i, Data: req.Images[i]}
 	}
 
-	checkpointLoaded := time.Now()
-
-	var prompt string
-	switch {
-	case req.Raw:
-		prompt = req.Prompt
-	case req.Prompt != "":
-		if req.Template == "" {
-			tmpl = model.Template
+	prompt := req.Prompt
+	if !req.Raw {
+		tmpl := m.Template
+		if req.Template != "" {
+			tmpl, err = template.Parse(req.Template)
+			if err != nil {
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+				return
+			}
 		}
 
-		if req.System == "" {
-			req.System = model.System
-		}
-
-		slog.Debug("generate handler", "prompt", req.Prompt)
-		slog.Debug("generate handler", "template", req.Template)
-		slog.Debug("generate handler", "system", req.System)
-
-		var sb strings.Builder
-		for i := range req.Images {
-			fmt.Fprintf(&sb, "[img-%d] ", i)
-		}
-
-		sb.WriteString(req.Prompt)
-
-		p, err := Prompt(tmpl, req.System, sb.String(), "", true)
-		if err != nil {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-
-		sb.Reset()
+		var b bytes.Buffer
 		if req.Context != nil {
-			prev, err := runner.llama.Detokenize(c.Request.Context(), req.Context)
+			s, err := r.Detokenize(c.Request.Context(), req.Context)
 			if err != nil {
 				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 				return
 			}
 
-			sb.WriteString(prev)
+			b.WriteString(s)
 		}
 
-		sb.WriteString(p)
-
-		prompt = sb.String()
-	}
-
-	slog.Debug("generate handler", "prompt", prompt)
-
-	ch := make(chan any)
-	var generated strings.Builder
-	go func() {
-		defer close(ch)
-
-		fn := func(r llm.CompletionResponse) {
-			// Build up the full response
-			if _, err := generated.WriteString(r.Content); err != nil {
-				ch <- gin.H{"error": err.Error()}
-				return
+		var values template.Values
+		if req.Suffix != "" {
+			values.Prompt = prompt
+			values.Suffix = req.Suffix
+		} else {
+			var msgs []api.Message
+			if req.System != "" {
+				msgs = append(msgs, api.Message{Role: "system", Content: req.System})
+			} else if m.System != "" {
+				msgs = append(msgs, api.Message{Role: "system", Content: m.System})
 			}
 
-			resp := api.GenerateResponse{
+			for _, i := range images {
+				msgs = append(msgs, api.Message{Role: "user", Content: fmt.Sprintf("[img-%d]", i.ID)})
+			}
+
+			values.Messages = append(msgs, api.Message{Role: "user", Content: req.Prompt})
+		}
+
+		if err := tmpl.Execute(&b, values); err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+
+		prompt = b.String()
+	}
+
+	slog.Debug("generate request", "prompt", prompt, "images", images)
+
+	ch := make(chan any)
+	go func() {
+		// TODO (jmorganca): avoid building the response twice both here and below
+		var sb strings.Builder
+		defer close(ch)
+		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
+		}, func(cr llm.CompletionResponse) {
+			res := api.GenerateResponse{
 				Model:      req.Model,
 				CreatedAt:  time.Now().UTC(),
-				Done:       r.Done,
-				Response:   r.Content,
-				DoneReason: r.DoneReason,
+				Response:   cr.Content,
+				Done:       cr.Done,
+				DoneReason: cr.DoneReason,
 				Metrics: api.Metrics{
-					PromptEvalCount:    r.PromptEvalCount,
-					PromptEvalDuration: r.PromptEvalDuration,
-					EvalCount:          r.EvalCount,
-					EvalDuration:       r.EvalDuration,
+					PromptEvalCount:    cr.PromptEvalCount,
+					PromptEvalDuration: cr.PromptEvalDuration,
+					EvalCount:          cr.EvalCount,
+					EvalDuration:       cr.EvalDuration,
 				},
 			}
 
-			if r.Done {
-				resp.TotalDuration = time.Since(checkpointStart)
-				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+			if _, err := sb.WriteString(cr.Content); err != nil {
+				ch <- gin.H{"error": err.Error()}
+			}
+
+			if cr.Done {
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
 
 				if !req.Raw {
-					p, err := Prompt(tmpl, req.System, req.Prompt, generated.String(), false)
-					if err != nil {
-						c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-						return
-					}
-
-					// TODO (jmorganca): encode() should not strip special tokens
-					tokens, err := runner.llama.Tokenize(c.Request.Context(), p)
+					tokens, err := r.Tokenize(c.Request.Context(), prompt+sb.String())
 					if err != nil {
 						ch <- gin.H{"error": err.Error()}
 						return
 					}
-
-					resp.Context = append(req.Context, tokens...)
+					res.Context = append(req.Context, tokens...)
 				}
 			}
 
-			ch <- resp
-		}
-
-		var images []llm.ImageData
-		for i := range req.Images {
-			images = append(images, llm.ImageData{
-				ID:   i,
-				Data: req.Images[i],
-			})
-		}
-
-		// Start prediction
-		req := llm.CompletionRequest{
-			Prompt:  prompt,
-			Format:  req.Format,
-			Images:  images,
-			Options: opts,
-		}
-		if err := runner.llama.Completion(c.Request.Context(), req, fn); err != nil {
+			ch <- res
+		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
 
 	if req.Stream != nil && !*req.Stream {
-		// Accumulate responses into the final response
-		var final api.GenerateResponse
+		var r api.GenerateResponse
 		var sb strings.Builder
-		for resp := range ch {
-			switch r := resp.(type) {
+		for rr := range ch {
+			switch t := rr.(type) {
 			case api.GenerateResponse:
-				sb.WriteString(r.Response)
-				final = r
+				sb.WriteString(t.Response)
+				r = t
 			case gin.H:
-				if errorMsg, ok := r["error"].(string); ok {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
-					return
-				} else {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
-					return
+				msg, ok := t["error"].(string)
+				if !ok {
+					msg = "unexpected error format in response"
 				}
+
+				c.JSON(http.StatusInternalServerError, gin.H{"error": msg})
+				return
 			default:
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected response"})
 				return
 			}
 		}
 
-		final.Response = sb.String()
-		c.JSON(http.StatusOK, final)
+		r.Response = sb.String()
+		c.JSON(http.StatusOK, r)
 		return
 	}
 
 	streamResponse(c, ch)
 }
 
-func (s *Server) EmbeddingsHandler(c *gin.Context) {
-	var req api.EmbeddingRequest
+func (s *Server) EmbedHandler(c *gin.Context) {
+	var req api.EmbedRequest
 	err := c.ShouldBindJSON(&req)
 	switch {
 	case errors.Is(err, io.EOF):
@@ -321,34 +295,122 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	if req.Model == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
+	truncate := true
+
+	if req.Truncate != nil && !*req.Truncate {
+		truncate = false
+	}
+
+	var input []string
+
+	switch i := req.Input.(type) {
+	case string:
+		if len(i) > 0 {
+			input = append(input, i)
+		}
+	case []any:
+		for _, v := range i {
+			if _, ok := v.(string); !ok {
+				c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "invalid input type"})
+				return
+			}
+			input = append(input, v.(string))
+		}
+	default:
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "invalid input type"})
 		return
 	}
 
-	model, err := GetModel(req.Model)
+	if len(input) == 0 {
+		c.JSON(http.StatusOK, api.EmbedResponse{Model: req.Model, Embeddings: [][]float32{}})
+		return
+	}
+
+	r, m, opts, err := s.scheduleRunner(c.Request.Context(), req.Model, []Capability{}, req.Options, req.KeepAlive)
 	if err != nil {
-		var pErr *fs.PathError
-		if errors.As(err, &pErr) {
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
+		handleScheduleError(c, req.Model, err)
+		return
+	}
+
+	kvData, err := getKVData(m.ModelPath, false)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	for i, s := range input {
+		tokens, err := r.Tokenize(c.Request.Context(), s)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 			return
 		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
 
-	opts, err := modelOptions(model, req.Options)
+		ctxLen := min(opts.NumCtx, int(kvData.ContextLength()))
+		if len(tokens) > ctxLen {
+			if !truncate {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "input length exceeds maximum context length"})
+				return
+			}
+
+			tokens = tokens[:ctxLen]
+			s, err = r.Detokenize(c.Request.Context(), tokens)
+			if err != nil {
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+				return
+			}
+		}
+
+		input[i] = s
+	}
+	embeddings, err := r.Embed(c.Request.Context(), input)
+
 	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		slog.Error("embedding generation failed", "error", err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to generate embedding"})
 		return
 	}
 
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
-	var runner *runnerRef
-	select {
-	case runner = <-rCh:
-	case err = <-eCh:
-		handleErrorResponse(c, err)
+	for i, e := range embeddings {
+		embeddings[i] = normalize(e)
+	}
+
+	resp := api.EmbedResponse{
+		Model:      req.Model,
+		Embeddings: embeddings,
+	}
+	c.JSON(http.StatusOK, resp)
+}
+
+func normalize(vec []float32) []float32 {
+	var sum float32
+	for _, v := range vec {
+		sum += v * v
+	}
+
+	norm := float32(0.0)
+	if sum > 0 {
+		norm = float32(1.0 / math.Sqrt(float64(sum)))
+	}
+
+	for i := range vec {
+		vec[i] *= norm
+	}
+	return vec
+}
+
+func (s *Server) EmbeddingsHandler(c *gin.Context) {
+	var req api.EmbeddingRequest
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	} else if err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	r, _, _, err := s.scheduleRunner(c.Request.Context(), req.Model, []Capability{}, req.Options, req.KeepAlive)
+	if err != nil {
+		handleScheduleError(c, req.Model, err)
 		return
 	}
 
@@ -358,13 +420,20 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	embedding, err := runner.llama.Embedding(c.Request.Context(), req.Prompt)
+	embeddings, err := r.Embed(c.Request.Context(), []string{req.Prompt})
+
 	if err != nil {
 		slog.Info(fmt.Sprintf("embedding generation failed: %v", err))
 		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to generate embedding"})
 		return
 	}
 
+	embedding := make([]float64, len(embeddings[0]))
+
+	for i, v := range embeddings[0] {
+		embedding[i] = float64(v)
+	}
+
 	resp := api.EmbeddingResponse{
 		Embedding: embedding,
 	}
@@ -541,8 +610,11 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 
 		quantization := cmp.Or(r.Quantize, r.Quantization)
 		if err := CreateModel(ctx, name, filepath.Dir(r.Path), strings.ToUpper(quantization), f, fn); err != nil {
+			if errors.Is(err, errBadTemplate) {
+			  ch <- gin.H{"error": err.Error(), "status": http.StatusBadRequest}
+			}
 			ch <- gin.H{"error": err.Error()}
-		}
+		  }
 	}()
 
 	if r.Stream != nil && !*r.Stream {
@@ -642,16 +714,9 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		m.System = req.System
 	}
 
-	if req.Template != "" {
-		m.Template, err = template.Parse(req.Template)
-		if err != nil {
-			return nil, err
-		}
-	}
-
-	msgs := make([]api.Message, 0)
-	for _, msg := range m.Messages {
-		msgs = append(msgs, api.Message{Role: msg.Role, Content: msg.Content})
+	msgs := make([]api.Message, len(m.Messages))
+	for i, msg := range m.Messages {
+		msgs[i] = api.Message{Role: msg.Role, Content: msg.Content}
 	}
 
 	n := model.ParseName(req.Model)
@@ -994,6 +1059,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 	r.POST("/api/pull", s.PullModelHandler)
 	r.POST("/api/generate", s.GenerateHandler)
 	r.POST("/api/chat", s.ChatHandler)
+	r.POST("/api/embed", s.EmbedHandler)
 	r.POST("/api/embeddings", s.EmbeddingsHandler)
 	r.POST("/api/create", s.CreateModelHandler)
 	r.POST("/api/push", s.PushModelHandler)
@@ -1007,6 +1073,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 	// Compatibility endpoints
 	r.POST("/v1/chat/completions", openai.ChatMiddleware(), s.ChatHandler)
 	r.POST("/v1/completions", openai.CompletionsMiddleware(), s.GenerateHandler)
+	r.POST("/v1/embeddings", openai.EmbeddingsMiddleware(), s.EmbedHandler)
 	r.GET("/v1/models", openai.ListMiddleware(), s.ListModelsHandler)
 	r.GET("/v1/models/:model", openai.RetrieveMiddleware(), s.ShowModelHandler)
 
@@ -1133,11 +1200,15 @@ func waitForStream(c *gin.Context, ch chan interface{}) {
 				return
 			}
 		case gin.H:
+			status, ok := r["status"].(int)
+			if !ok {
+				status = http.StatusInternalServerError
+			}
 			if errorMsg, ok := r["error"].(string); ok {
-				c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
+				c.JSON(status, gin.H{"error": errorMsg})
 				return
 			} else {
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in progress response"})
+				c.JSON(status, gin.H{"error": "unexpected error format in progress response"})
 				return
 			}
 		default:
@@ -1214,132 +1285,67 @@ func (s *Server) ProcessHandler(c *gin.Context) {
 	c.JSON(http.StatusOK, api.ProcessResponse{Models: models})
 }
 
-// ChatPrompt builds up a prompt from a series of messages for the currently `loaded` model
-func chatPrompt(ctx context.Context, runner *runnerRef, template *template.Template, messages []api.Message, numCtx int) (string, error) {
-	encode := func(s string) ([]int, error) {
-		return runner.llama.Tokenize(ctx, s)
-	}
-
-	prompt, err := ChatPrompt(template, messages, numCtx, encode)
-	if err != nil {
-		return "", err
-	}
-
-	return prompt, nil
-}
-
 func (s *Server) ChatHandler(c *gin.Context) {
 	checkpointStart := time.Now()
 
 	var req api.ChatRequest
-	err := c.ShouldBindJSON(&req)
-	switch {
-	case errors.Is(err, io.EOF):
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
 		return
-	case err != nil:
+	} else if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
-	// validate the request
-	switch {
-	case req.Model == "":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
-		return
-	case len(req.Format) > 0 && req.Format != "json":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be json"})
-		return
+	caps := []Capability{CapabilityCompletion}
+	if len(req.Tools) > 0 {
+		caps = append(caps, CapabilityTools)
 	}
 
-	model, err := GetModel(req.Model)
-	if err != nil {
-		var pErr *fs.PathError
-		if errors.As(err, &pErr) {
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
-			return
-		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+	r, m, opts, err := s.scheduleRunner(c.Request.Context(), req.Model, caps, req.Options, req.KeepAlive)
+	if errors.Is(err, errCapabilityCompletion) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%q does not support chat", req.Model)})
 		return
-	}
-
-	if !model.Has(CapabilityCompletion) {
-		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%s does not support chat", req.Model)})
-		return
-	}
-
-	opts, err := modelOptions(model, req.Options)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
-	var runner *runnerRef
-	select {
-	case runner = <-rCh:
-	case err = <-eCh:
-		handleErrorResponse(c, err)
+	} else if err != nil {
+		handleScheduleError(c, req.Model, err)
 		return
 	}
 
 	checkpointLoaded := time.Now()
 
-	// if the first message is not a system message, then add the model's default system message
-	if len(req.Messages) > 0 && req.Messages[0].Role != "system" {
-		req.Messages = append([]api.Message{
-			{
-				Role:    "system",
-				Content: model.System,
-			},
-		}, req.Messages...)
-	}
-
-	prompt, err := chatPrompt(c.Request.Context(), runner, model.Template, req.Messages, opts.NumCtx)
-	if err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	// an empty request loads the model
-	if len(req.Messages) == 0 || prompt == "" {
-		resp := api.ChatResponse{
-			CreatedAt:  time.Now().UTC(),
+	if len(req.Messages) == 0 {
+		c.JSON(http.StatusOK, api.ChatResponse{
 			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
+			Message:    api.Message{Role: "assistant"},
 			Done:       true,
 			DoneReason: "load",
-			Message:    api.Message{Role: "assistant"},
-		}
-		c.JSON(http.StatusOK, resp)
+		})
 		return
 	}
 
-	// only send images that are in the prompt
-	var i int
-	var images []llm.ImageData
-	for _, m := range req.Messages {
-		for _, img := range m.Images {
-			if !isSupportedImageType(img) {
-				c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "unsupported image format"})
-				return
-			}
-
-			if strings.Contains(prompt, fmt.Sprintf("[img-%d]", i)) {
-				images = append(images, llm.ImageData{Data: img, ID: i})
-			}
-			i += 1
-		}
+	if req.Messages[0].Role != "system" && m.System != "" {
+		req.Messages = append([]api.Message{{Role: "system", Content: m.System}}, req.Messages...)
 	}
 
-	slog.Debug("chat handler", "prompt", prompt, "images", len(images))
+	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, req.Messages, req.Tools)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	slog.Debug("chat request", "images", len(images), "prompt", prompt)
 
 	ch := make(chan any)
-
 	go func() {
 		defer close(ch)
-
-		fn := func(r llm.CompletionResponse) {
-			resp := api.ChatResponse{
+		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
+		}, func(r llm.CompletionResponse) {
+			res := api.ChatResponse{
 				Model:      req.Model,
 				CreatedAt:  time.Now().UTC(),
 				Message:    api.Message{Role: "assistant", Content: r.Content},
@@ -1354,62 +1360,65 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 
 			if r.Done {
-				resp.TotalDuration = time.Since(checkpointStart)
-				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
 			}
 
-			ch <- resp
-		}
-
-		if err := runner.llama.Completion(c.Request.Context(), llm.CompletionRequest{
-			Prompt:  prompt,
-			Format:  req.Format,
-			Images:  images,
-			Options: opts,
-		}, fn); err != nil {
+			ch <- res
+		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
 
 	if req.Stream != nil && !*req.Stream {
-		// Accumulate responses into the final response
-		var final api.ChatResponse
+		var resp api.ChatResponse
 		var sb strings.Builder
-		for resp := range ch {
-			switch r := resp.(type) {
+		for rr := range ch {
+			switch t := rr.(type) {
 			case api.ChatResponse:
-				sb.WriteString(r.Message.Content)
-				final = r
+				sb.WriteString(t.Message.Content)
+				resp = t
 			case gin.H:
-				if errorMsg, ok := r["error"].(string); ok {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
-					return
-				} else {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
-					return
+				msg, ok := t["error"].(string)
+				if !ok {
+					msg = "unexpected error format in response"
 				}
+
+				c.JSON(http.StatusInternalServerError, gin.H{"error": msg})
+				return
 			default:
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected response"})
 				return
 			}
 		}
 
-		final.Message = api.Message{Role: "assistant", Content: sb.String()}
-		c.JSON(http.StatusOK, final)
+		resp.Message.Content = sb.String()
+
+		if len(req.Tools) > 0 {
+			if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
+				resp.Message.ToolCalls = toolCalls
+				resp.Message.Content = ""
+			}
+		}
+
+		c.JSON(http.StatusOK, resp)
 		return
 	}
 
 	streamResponse(c, ch)
 }
 
-func handleErrorResponse(c *gin.Context, err error) {
-	if errors.Is(err, context.Canceled) {
+func handleScheduleError(c *gin.Context, name string, err error) {
+	switch {
+	case errors.Is(err, errCapabilities), errors.Is(err, errRequired):
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	case errors.Is(err, context.Canceled):
 		c.JSON(499, gin.H{"error": "request canceled"})
-		return
-	}
-	if errors.Is(err, ErrMaxQueue) {
+	case errors.Is(err, ErrMaxQueue):
 		c.JSON(http.StatusServiceUnavailable, gin.H{"error": err.Error()})
-		return
+	case errors.Is(err, os.ErrNotExist):
+		c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model %q not found, try pulling it first", name)})
+	default:
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 	}
-	c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 }

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"
@@ -272,6 +273,77 @@ func Test_Routes(t *testing.T) {
 				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)
+				}
+			},
+		},
 	}
 
 	t.Setenv("OLLAMA_MODELS", t.TempDir())
@@ -420,3 +492,38 @@ func TestShow(t *testing.T) {
 		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

@@ -133,17 +133,8 @@ func (s *Scheduler) processPending(ctx context.Context) {
 				numParallel = 1
 				slog.Warn("multimodal models don't support parallel requests yet")
 			}
-			// Keep NumCtx and numParallel in sync
-			if numParallel > 1 {
-				pending.opts.NumCtx = pending.origNumCtx * numParallel
-			}
 
 			for {
-				cpus := s.getCpuFn()
-				var systemMem gpu.GpuInfo
-				if len(cpus) > 0 {
-					systemMem = cpus[0]
-				}
 				var runnerToExpire *runnerRef
 				s.loadedMu.Lock()
 				runner := s.loaded[pending.model.ModelPath]
@@ -197,35 +188,15 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						break
 					}
 
-					// Block attempting to load a model larger than system memory + GPU memory
-					estimate := llm.EstimateGPULayers(gpus, ggml, pending.model.ProjectorPaths, pending.opts)
-					maxSize := systemMem.FreeMemory
-					for _, gpu := range gpus {
-						if gpu.Library == "cpu" {
-							continue
-						}
-						if loadedCount == 0 {
-							// If no other models are loaded, set the limit based on what's available
-							maxSize += gpu.FreeMemory
-						} else {
-							// Other models could be unloaded, favor total memory for limit
-							maxSize += gpu.TotalMemory
-						}
-					}
-					if estimate.TotalSize > maxSize {
-						slog.Warn("model request too large for system", "requested", format.HumanBytes2(estimate.TotalSize), "system", format.HumanBytes2(maxSize))
-						pending.errCh <- fmt.Errorf("requested model (%s) is too large for this system (%s)", format.HumanBytes2(estimate.TotalSize), format.HumanBytes2(maxSize))
-						break
-					}
-
 					// 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
 						}
 
+						pending.opts.NumCtx = pending.origNumCtx * numParallel
+
 						if loadedCount == 0 {
 							slog.Debug("cpu mode with first model, loading")
 							s.loadFn(pending, ggml, gpus, numParallel)

server/sched_test.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"log/slog"
 	"os"
+	"runtime"
 	"testing"
 	"time"
 
@@ -94,7 +95,7 @@ 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 +103,13 @@ 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, numParallel int) (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)
@@ -135,124 +136,154 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 
 	fname := f.Name()
 	model := &Model{Name: modelName, ModelPath: fname}
-	scenario.ggml, err = llm.LoadModel(model.ModelPath, 0)
+	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: &api.Duration{Duration: 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 = &api.Duration{Duration: 5 * time.Millisecond}
-	scenario1b := newScenario(t, ctx, "ollama-model-1", 11)
-	scenario1b.req.model = scenario1a.req.model
-	scenario1b.ggml = scenario1a.ggml
-	scenario1b.req.sessionDuration = &api.Duration{Duration: 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 = &api.Duration{Duration: 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)
-	case err := <-scenario1a.req.errCh:
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		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)
-	case err := <-scenario1b.req.errCh:
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
+	case <-ctx.Done():
+		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)
-	case err := <-scenario2a.req.errCh:
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		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)
-	case err := <-scenario3a.req.errCh:
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
@@ -262,15 +293,15 @@ func TestRequests(t *testing.T) {
 	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)
-	case err := <-scenario3b.req.errCh:
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
@@ -280,15 +311,15 @@ func TestRequests(t *testing.T) {
 	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)
-	case err := <-scenario3c.req.errCh:
+		require.Empty(t, c.req.errCh)
+	case err := <-c.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
@@ -298,25 +329,25 @@ func TestRequests(t *testing.T) {
 	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.Fatal("timeout")
 	}
@@ -325,30 +356,59 @@ func TestRequests(t *testing.T) {
 	s.loadedMu.Unlock()
 }
 
+func TestRequestsModelTooBigForSystem(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	s.getGpuFn = func() gpu.GpuInfoList {
+		g := gpu.GpuInfo{Library: "metal"}
+		g.TotalMemory = 4 * format.MebiByte
+		g.FreeMemory = 3 * format.MebiByte
+		return []gpu.GpuInfo{g}
+	}
+
+	s.getCpuFn = func() gpu.GpuInfoList {
+		g := gpu.GpuInfo{Library: "cpu"}
+		g.TotalMemory = 4 * format.MebiByte
+		g.FreeMemory = 2 * format.MebiByte
+		return []gpu.GpuInfo{g}
+	}
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	select {
+	case <-a.req.successCh:
+		if runtime.GOOS == "linux" {
+			t.Fatal("request should have been rejected with out of space")
+		}
+		// else - Darwin and Windows don't reject right now
+	case err := <-a.req.errCh:
+		require.Contains(t, err.Error(), "too large")
+	case <-ctx.Done():
+		t.Fatal("timeout")
+	}
+}
 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 = &api.Duration{Duration: 0}
-	scenario1b := newScenario(t, ctx, "ollama-model-1b", 10)
-	scenario1b.req.sessionDuration = &api.Duration{Duration: 0}
-	scenario1c := newScenario(t, ctx, "ollama-model-1c", 10)
-	scenario1c.req.sessionDuration = &api.Duration{Duration: 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)
@@ -357,22 +417,24 @@ 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.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)
@@ -380,7 +442,7 @@ 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
@@ -389,7 +451,7 @@ 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"}
@@ -411,6 +473,8 @@ 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.Fatal("timeout")
 	}
@@ -446,6 +510,8 @@ func TestUseLoadedRunner(t *testing.T) {
 	select {
 	case success := <-req.successCh:
 		require.Equal(t, r1, success)
+	case err := <-req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -625,8 +691,7 @@ func TestAlreadyCanceled(t *testing.T) {
 	defer done()
 	dctx, done2 := context.WithCancel(ctx)
 	done2()
-	scenario1a := newScenario(t, dctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = &api.Duration{Duration: 0}
+	scenario1a := newScenarioRequest(t, dctx, "ollama-model-1", 10, &api.Duration{Duration: 0})
 	s := InitScheduler(ctx)
 	slog.Info("scenario1a")
 	s.pendingReqCh <- scenario1a.req
@@ -642,8 +707,8 @@ type mockLlm struct {
 	pingResp           error
 	waitResp           error
 	completionResp     error
-	embeddingResp      []float64
-	embeddingRespErr   error
+	embedResp          [][]float32
+	embedRespErr       error
 	tokenizeResp       []int
 	tokenizeRespErr    error
 	detokenizeResp     string
@@ -660,8 +725,8 @@ 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|>
+

server/testdata/tools/messages.json

@@ -0,0 +1,39 @@
+[
+  {
+    "role": "system",
+    "content": "You are a knowledgable assistant. You can answer questions and perform tasks."
+  },
+  {
+    "role": "user",
+    "content": "What's the weather like today in Paris?"
+  },
+  {
+    "role": "assistant",
+    "tool_calls": [
+      {
+        "id": "89a1e453-0bce-4de3-a456-c54bed09c520",
+        "type": "function",
+        "function": {
+          "name": "get_current_weather",
+          "arguments": {
+            "location": "Paris, France",
+            "format": "celsius"
+          }
+        }
+      }
+    ]
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "89a1e453-0bce-4de3-a456-c54bed09c520",
+    "content": "22"
+  },
+  {
+    "role": "assistant",
+    "content": "The current temperature in Paris, France is 22 degrees Celsius."
+  },
+  {
+    "role": "user",
+    "content": "What's the weather like today in San Francisco and Toronto?"
+  }
+]

server/testdata/tools/mistral.gotmpl

@@ -0,0 +1,15 @@
+{{- range $index, $_ := .Messages }}
+{{- if eq .Role "user" }}
+{{- if and (eq (len (slice $.Messages $index)) 1) $.Tools }}[AVAILABLE_TOOLS] {{ $.Tools }}[/AVAILABLE_TOOLS]
+{{- end }}[INST] {{ if and (eq (len (slice $.Messages $index)) 1) $.System }}{{ $.System }}
+
+{{ end }}{{ .Content }}[/INST]
+{{- else if eq .Role "assistant" }}
+{{- if .Content }} {{ .Content }}</s>
+{{- else if .ToolCalls }}[TOOL_CALLS] [
+{{- range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
+{{- end }}]</s>
+{{- end }}
+{{- else if eq .Role "tool" }}[TOOL_RESULTS] {"content": {{ .Content }}}[/TOOL_RESULTS]
+{{- end }}
+{{- end }}

server/testdata/tools/mistral.out

@@ -0,0 +1,3 @@
+[INST] What's the weather like today in Paris?[/INST][TOOL_CALLS] [{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]</s>[TOOL_RESULTS] {"content": 22}[/TOOL_RESULTS] The current temperature in Paris, France is 22 degrees Celsius.</s>[AVAILABLE_TOOLS] [{"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"}}}}}][/AVAILABLE_TOOLS][INST] You are a knowledgable assistant. You can answer questions and perform tasks.
+
+What's the weather like today in San Francisco and Toronto?[/INST]

server/testdata/tools/tools.json

@@ -0,0 +1,30 @@
+[
+  {
+    "type": "function",
+    "function": {
+      "name": "get_current_weather",
+      "description": "Get the current weather",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "location": {
+            "type": "string",
+            "description": "The city and state, e.g. San Francisco, CA"
+          },
+          "format": {
+            "type": "string",
+            "enum": [
+              "celsius",
+              "fahrenheit"
+            ],
+            "description": "The temperature unit to use. Infer this from the users location."
+          }
+        },
+        "required": [
+          "location",
+          "format"
+        ]
+      }
+    }
+  }
+]

template/alpaca.gotmpl

@@ -4,4 +4,5 @@
 {{ .Prompt }}
 
 {{ end }}### Response:
-{{ .Response }}
+{{ .Response }}
+

template/chatml.gotmpl

@@ -3,4 +3,4 @@
 {{ end }}{{ if .Prompt }}<|im_start|>user
 {{ .Prompt }}<|im_end|>
 {{ end }}<|im_start|>assistant
-{{ .Response }}<|im_end|>
+{{ .Response }}<|im_end|>

template/chatqa.gotmpl

@@ -2,4 +2,5 @@
 
 {{ end }}{{ if .Prompt }}User: {{ .Prompt }}
 
-{{ end }}Assistant: <|begin_of_text|>{{ .Response }}
+{{ end }}Assistant: {{ .Response }}
+

template/codellama-70b-instruct.gotmpl

@@ -1,8 +1,10 @@
-{{ if .System }} Source: system
+{{ if .System }}Source: system
 
- {{ .System }} <step>{{ end }} Source: user
+ {{ .System }} <step> {{ end }}Source: user
 
  {{ .Prompt }} <step> Source: assistant
+{{- if not .Response }}
 Destination: user
+{{- end }}
 
- {{ .Response }}<step>
+ {{ .Response }} <step> 

template/falcon-instruct.gotmpl

@@ -1,3 +1,5 @@
-{{ if .System }}{{ .System }}
-{{ end }}{{ if .Prompt }}User: {{ .Prompt }}
-{{ end }}Assistant: {{ .Response }}
+{{ if .System }}System: {{ .System }}
+{{ end }}{{ if .Prompt }}User:
+{{ .Prompt }}
+{{ end }}Falcon:
+{{ .Response }}

template/gemma-instruct.gotmpl

@@ -1,4 +1,5 @@
 <start_of_turn>user
-{{ if .System }}{{ .System }} {{ end }}{{ .Prompt }}<end_of_turn>
+{{ if .System }}{{ .System }}
+{{ end }}{{ .Prompt }}<end_of_turn>
 <start_of_turn>model
-{{ .Response }}<end_of_turn>
+{{ .Response }}<end_of_turn>

template/granite-instruct.gotmpl

@@ -1,9 +1,9 @@
-{{ if .System }}
-System:
+{{ if .System }}System:
 {{ .System }}
 
 {{ end }}{{ if .Prompt }}Question:
 {{ .Prompt }}
 
 {{ end }}Answer:
-{{ .Response }}
+{{ .Response }}
+

template/llama2-chat.gotmpl

@@ -1,3 +1,6 @@
-[INST] <<SYS>>{{ .System }}<</SYS>>
+[INST] <<SYS>>
+{{- if .System }}
+{{ .System }}
+{{ end }}<</SYS>>
 
-{{ .Prompt }} [/INST] {{ .Response }}
+{{ .Prompt }} [/INST] {{ .Response }}</s><s>

template/magicoder.gotmpl

@@ -4,4 +4,5 @@
 {{ .Prompt }}
 
 {{ end }}@@ Response
-{{ .Response }}
+{{ .Response }}
+

template/mistral-instruct.gotmpl

@@ -1,6 +1,3 @@
-{{ if .System }}<|im_start|>system
-{{ .System }}<|im_end|>
-{{ end }}{{ if .Prompt }}<|im_start|>user
-{{ .Prompt }}<|im_end|>
-{{ end }}<|im_start|>assistant
-{{ .Response }}<|im_end|>
+[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}</s>

template/openchat.gotmpl

@@ -1 +1 @@
-{{ .System }}<|end_of_turn|>GPT4 Correct User: {{ .Prompt }}<|end_of_turn|>GPT4 Correct Assistant: {{ .Response }}<|end_of_turn|>
+{{ if .System }}GPT4 Correct System: {{ .System }}<|end_of_turn|>{{ end }}GPT4 Correct User: {{ .Prompt }}<|end_of_turn|>GPT4 Correct Assistant: {{ .Response }}<|end_of_turn|>

template/phi-3.gotmpl

@@ -3,4 +3,4 @@
 {{ end }}{{ if .Prompt }}<|user|>
 {{ .Prompt }}<|end|>
 {{ end }}<|assistant|>
-{{ .Response }}<|end|>
+{{ .Response }}<|end|>

template/solar-instruct.gotmpl

@@ -5,4 +5,5 @@
 {{ .Prompt }}
 
 {{ end }}### Assistant:
-{{ .Response }}
+{{ .Response }}</s>
+

template/starcoder2-instruct.gotmpl

@@ -3,7 +3,6 @@
 {{ end }}{{ if .Prompt }}### Instruction
 {{ .Prompt }}
 
-
 {{ end }}### Response
 {{ .Response }}<|endoftext|>
 

template/template.go

@@ -5,6 +5,7 @@ import (
 	"embed"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"math"
 	"slices"
@@ -14,6 +15,7 @@ import (
 	"text/template/parse"
 
 	"github.com/agnivade/levenshtein"
+	"github.com/ollama/ollama/api"
 	"golang.org/x/exp/maps"
 )
 
@@ -74,30 +76,66 @@ func Named(s string) (*named, error) {
 	return nil, errors.New("no matching template found")
 }
 
+var DefaultTemplate, _ = Parse("{{ .Prompt }}")
+
 type Template struct {
 	*template.Template
 	raw string
 }
 
+// response is a template node that can be added to templates that don't already have one
+var response = parse.ActionNode{
+	NodeType: parse.NodeAction,
+	Pipe: &parse.PipeNode{
+		NodeType: parse.NodePipe,
+		Cmds: []*parse.CommandNode{
+			{
+				NodeType: parse.NodeCommand,
+				Args: []parse.Node{
+					&parse.FieldNode{
+						NodeType: parse.NodeField,
+						Ident:    []string{"Response"},
+					},
+				},
+			},
+		},
+	},
+}
+
+var funcs = template.FuncMap{
+	"json": func(v any) string {
+		b, _ := json.Marshal(v)
+		return string(b)
+	},
+}
+
+func Parse(s string) (*Template, error) {
+	tmpl := template.New("").Option("missingkey=zero").Funcs(funcs)
+
+	tmpl, err := tmpl.Parse(s)
+	if err != nil {
+		return nil, err
+	}
+
+	t := Template{Template: tmpl, raw: s}
+	if vars := t.Vars(); !slices.Contains(vars, "messages") && !slices.Contains(vars, "response") {
+		// touch up the template and append {{ .Response }}
+		tmpl.Tree.Root.Nodes = append(tmpl.Tree.Root.Nodes, &response)
+	}
+
+	return &t, nil
+}
+
 func (t *Template) String() string {
 	return t.raw
 }
 
-var DefaultTemplate, _ = Parse("{{ .Prompt }}")
-
-func Parse(s string) (*Template, error) {
-	t, err := template.New("").Option("missingkey=zero").Parse(s)
-	if err != nil {
-		return nil, err
-	}
-
-	return &Template{Template: t, raw: s}, nil
-}
-
 func (t *Template) Vars() []string {
 	var vars []string
-	for _, n := range t.Tree.Root.Nodes {
-		vars = append(vars, parseNode(n)...)
+	for _, tt := range t.Templates() {
+		for _, n := range tt.Root.Nodes {
+			vars = append(vars, Identifiers(n)...)
+		}
 	}
 
 	set := make(map[string]struct{})
@@ -110,49 +148,285 @@ func (t *Template) Vars() []string {
 	return vars
 }
 
-func parseNode(n parse.Node) []string {
+type Values struct {
+	Messages []api.Message
+	api.Tools
+	Prompt string
+	Suffix string
+
+	// forceLegacy is a flag used to test compatibility with legacy templates
+	forceLegacy bool
+}
+
+func (t *Template) Subtree(fn func(parse.Node) bool) *template.Template {
+	var walk func(parse.Node) parse.Node
+	walk = func(n parse.Node) parse.Node {
+		if fn(n) {
+			return n
+		}
+
+		switch t := n.(type) {
+		case *parse.ListNode:
+			for _, c := range t.Nodes {
+				if n := walk(c); n != nil {
+					return n
+				}
+			}
+		case *parse.BranchNode:
+			for _, n := range []*parse.ListNode{t.List, t.ElseList} {
+				if n != nil {
+					if n := walk(n); n != nil {
+						return n
+					}
+				}
+			}
+		case *parse.IfNode:
+			return walk(&t.BranchNode)
+		case *parse.WithNode:
+			return walk(&t.BranchNode)
+		case *parse.RangeNode:
+			return walk(&t.BranchNode)
+		}
+
+		return nil
+	}
+
+	if n := walk(t.Tree.Root); n != nil {
+		return (&template.Template{
+			Tree: &parse.Tree{
+				Root: &parse.ListNode{
+					Nodes: []parse.Node{n},
+				},
+			},
+		}).Funcs(funcs)
+	}
+
+	return nil
+}
+
+func (t *Template) Execute(w io.Writer, v Values) error {
+	system, messages := collate(v.Messages)
+	if v.Prompt != "" && v.Suffix != "" {
+		return t.Template.Execute(w, map[string]any{
+			"Prompt":   v.Prompt,
+			"Suffix":   v.Suffix,
+			"Response": "",
+		})
+	} else if !v.forceLegacy && slices.Contains(t.Vars(), "messages") {
+		return t.Template.Execute(w, map[string]any{
+			"System":   system,
+			"Messages": messages,
+			"Tools":    v.Tools,
+			"Response": "",
+		})
+	}
+
+	system = ""
+	var b bytes.Buffer
+	var prompt, response string
+	for _, m := range messages {
+		execute := func() error {
+			if err := t.Template.Execute(&b, map[string]any{
+				"System":   system,
+				"Prompt":   prompt,
+				"Response": response,
+			}); err != nil {
+				return err
+			}
+
+			system = ""
+			prompt = ""
+			response = ""
+			return nil
+		}
+
+		switch m.Role {
+		case "system":
+			if prompt != "" || response != "" {
+				if err := execute(); err != nil {
+					return err
+				}
+			}
+			system = m.Content
+		case "user":
+			if response != "" {
+				if err := execute(); err != nil {
+					return err
+				}
+			}
+			prompt = m.Content
+		case "assistant":
+			response = m.Content
+		}
+	}
+
+	var cut bool
+	nodes := deleteNode(t.Template.Root.Copy(), func(n parse.Node) bool {
+		if field, ok := n.(*parse.FieldNode); ok && slices.Contains(field.Ident, "Response") {
+			cut = true
+			return false
+		}
+
+		return cut
+	})
+
+	tree := parse.Tree{Root: nodes.(*parse.ListNode)}
+	if err := template.Must(template.New("").AddParseTree("", &tree)).Execute(&b, map[string]any{
+		"System":   system,
+		"Prompt":   prompt,
+		"Response": response,
+	}); err != nil {
+		return err
+	}
+
+	_, err := io.Copy(w, &b)
+	return err
+}
+
+// collate messages based on role. consecutive messages of the same role are merged
+// into a single message. collate also collects and returns all system messages.
+// collate mutates message content adding image tags ([img-%d]) as needed
+func collate(msgs []api.Message) (string, []*api.Message) {
+	var n int
+
+	var system []string
+	var collated []*api.Message
+	for i := range msgs {
+		msg := msgs[i]
+		for range msg.Images {
+			imageTag := fmt.Sprintf("[img-%d]", n)
+			if !strings.Contains(msg.Content, "[img]") {
+				msg.Content = strings.TrimSpace("[img] " + msg.Content)
+			}
+
+			msg.Content = strings.Replace(msg.Content, "[img]", imageTag, 1)
+			n++
+		}
+
+		if msg.Role == "system" {
+			system = append(system, msg.Content)
+		}
+
+		if len(collated) > 0 && collated[len(collated)-1].Role == msg.Role {
+			collated[len(collated)-1].Content += "\n\n" + msg.Content
+		} else {
+			collated = append(collated, &msg)
+		}
+	}
+
+	return strings.Join(system, "\n\n"), collated
+}
+
+// Identifiers walks the node tree returning any identifiers it finds along the way
+func Identifiers(n parse.Node) []string {
 	switch n := n.(type) {
+	case *parse.ListNode:
+		var names []string
+		for _, n := range n.Nodes {
+			names = append(names, Identifiers(n)...)
+		}
+
+		return names
+	case *parse.TemplateNode:
+		return Identifiers(n.Pipe)
 	case *parse.ActionNode:
-		return parseNode(n.Pipe)
+		return Identifiers(n.Pipe)
+	case *parse.BranchNode:
+		names := Identifiers(n.Pipe)
+		for _, n := range []*parse.ListNode{n.List, n.ElseList} {
+			if n != nil {
+				names = append(names, Identifiers(n)...)
+			}
+		}
+		return names
 	case *parse.IfNode:
-		names := parseNode(n.Pipe)
-		names = append(names, parseNode(n.List)...)
-		if n.ElseList != nil {
-			names = append(names, parseNode(n.ElseList)...)
-		}
-		return names
+		return Identifiers(&n.BranchNode)
 	case *parse.RangeNode:
-		names := parseNode(n.Pipe)
-		names = append(names, parseNode(n.List)...)
-		if n.ElseList != nil {
-			names = append(names, parseNode(n.ElseList)...)
-		}
-		return names
+		return Identifiers(&n.BranchNode)
 	case *parse.WithNode:
-		names := parseNode(n.Pipe)
-		names = append(names, parseNode(n.List)...)
-		if n.ElseList != nil {
-			names = append(names, parseNode(n.ElseList)...)
-		}
-		return names
+		return Identifiers(&n.BranchNode)
 	case *parse.PipeNode:
 		var names []string
 		for _, c := range n.Cmds {
 			for _, a := range c.Args {
-				names = append(names, parseNode(a)...)
+				names = append(names, Identifiers(a)...)
 			}
 		}
-		return names
-	case *parse.ListNode:
-		var names []string
-		for _, n := range n.Nodes {
-			names = append(names, parseNode(n)...)
-		}
-
 		return names
 	case *parse.FieldNode:
 		return n.Ident
+	case *parse.VariableNode:
+		return n.Ident
 	}
 
 	return nil
 }
+
+// deleteNode walks the node list and deletes nodes that match the predicate
+// this is currently to remove the {{ .Response }} node from templates
+func deleteNode(n parse.Node, fn func(parse.Node) bool) parse.Node {
+	var walk func(n parse.Node) parse.Node
+	walk = func(n parse.Node) parse.Node {
+		if fn(n) {
+			return nil
+		}
+
+		switch t := n.(type) {
+		case *parse.ListNode:
+			var nodes []parse.Node
+			for _, c := range t.Nodes {
+				if n := walk(c); n != nil {
+					nodes = append(nodes, n)
+				}
+			}
+
+			t.Nodes = nodes
+			return t
+		case *parse.IfNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.WithNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.RangeNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.BranchNode:
+			t.List = walk(t.List).(*parse.ListNode)
+			if t.ElseList != nil {
+				t.ElseList = walk(t.ElseList).(*parse.ListNode)
+			}
+		case *parse.ActionNode:
+			n := walk(t.Pipe)
+			if n == nil {
+				return nil
+			}
+
+			t.Pipe = n.(*parse.PipeNode)
+		case *parse.PipeNode:
+			var commands []*parse.CommandNode
+			for _, c := range t.Cmds {
+				var args []parse.Node
+				for _, a := range c.Args {
+					if n := walk(a); n != nil {
+						args = append(args, n)
+					}
+				}
+
+				if len(args) == 0 {
+					return nil
+				}
+
+				c.Args = args
+				commands = append(commands, c)
+			}
+
+			if len(commands) == 0 {
+				return nil
+			}
+
+			t.Cmds = commands
+		}
+
+		return n
+	}
+
+	return walk(n)
+}

template/template_test.go

@@ -8,9 +8,11 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
+	"strings"
 	"testing"
-	"text/template"
 
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/llm"
 )
 
@@ -46,7 +48,7 @@ func TestNamed(t *testing.T) {
 					t.Fatal(err)
 				}
 
-				tmpl, err := template.New(s).Parse(b.String())
+				tmpl, err := Parse(b.String())
 				if err != nil {
 					t.Fatal(err)
 				}
@@ -59,18 +61,125 @@ func TestNamed(t *testing.T) {
 	}
 }
 
+func TestTemplate(t *testing.T) {
+	cases := make(map[string][]api.Message)
+	for _, mm := range [][]api.Message{
+		{
+			{Role: "user", Content: "Hello, how are you?"},
+		},
+		{
+			{Role: "user", Content: "Hello, how are you?"},
+			{Role: "assistant", Content: "I'm doing great. How can I help you today?"},
+			{Role: "user", Content: "I'd like to show off how chat templating works!"},
+		},
+		{
+			{Role: "system", Content: "You are a helpful assistant."},
+			{Role: "user", Content: "Hello, how are you?"},
+			{Role: "assistant", Content: "I'm doing great. How can I help you today?"},
+			{Role: "user", Content: "I'd like to show off how chat templating works!"},
+		},
+	} {
+		var roles []string
+		for _, m := range mm {
+			roles = append(roles, m.Role)
+		}
+
+		cases[strings.Join(roles, "-")] = mm
+	}
+
+	matches, err := filepath.Glob("*.gotmpl")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	for _, match := range matches {
+		t.Run(match, func(t *testing.T) {
+			bts, err := os.ReadFile(match)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			tmpl, err := Parse(string(bts))
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			for n, tt := range cases {
+				var actual bytes.Buffer
+				t.Run(n, func(t *testing.T) {
+					if err := tmpl.Execute(&actual, Values{Messages: tt}); err != nil {
+						t.Fatal(err)
+					}
+
+					expect, err := os.ReadFile(filepath.Join("testdata", match, n))
+					if err != nil {
+						t.Fatal(err)
+					}
+
+					bts := actual.Bytes()
+
+					if slices.Contains([]string{"chatqa.gotmpl", "llama2-chat.gotmpl", "mistral-instruct.gotmpl", "openchat.gotmpl", "vicuna.gotmpl"}, match) && bts[len(bts)-1] == ' ' {
+						t.Log("removing trailing space from output")
+						bts = bts[:len(bts)-1]
+					}
+
+					if diff := cmp.Diff(bts, expect); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				})
+
+				t.Run("legacy", func(t *testing.T) {
+					t.Skip("legacy outputs are currently default outputs")
+					var legacy bytes.Buffer
+					if err := tmpl.Execute(&legacy, Values{Messages: tt, forceLegacy: true}); err != nil {
+						t.Fatal(err)
+					}
+
+					legacyBytes := legacy.Bytes()
+					if slices.Contains([]string{"chatqa.gotmpl", "openchat.gotmpl", "vicuna.gotmpl"}, match) && legacyBytes[len(legacyBytes)-1] == ' ' {
+						t.Log("removing trailing space from legacy output")
+						legacyBytes = legacyBytes[:len(legacyBytes)-1]
+					} else if slices.Contains([]string{"codellama-70b-instruct.gotmpl", "llama2-chat.gotmpl", "mistral-instruct.gotmpl"}, match) {
+						t.Skip("legacy outputs cannot be compared to messages outputs")
+					}
+
+					if diff := cmp.Diff(legacyBytes, actual.Bytes()); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				})
+			}
+		})
+	}
+}
+
 func TestParse(t *testing.T) {
 	cases := []struct {
 		template string
 		vars     []string
 	}{
-		{"{{ .Prompt }}", []string{"prompt"}},
-		{"{{ .System }} {{ .Prompt }}", []string{"prompt", "system"}},
+		{"{{ .Prompt }}", []string{"prompt", "response"}},
+		{"{{ .System }} {{ .Prompt }}", []string{"prompt", "response", "system"}},
 		{"{{ .System }} {{ .Prompt }} {{ .Response }}", []string{"prompt", "response", "system"}},
-		{"{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}", []string{"prompt", "system", "tools"}},
+		{"{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}", []string{"prompt", "response", "system", "tools"}},
 		{"{{ range .Messages }}{{ .Role }} {{ .Content }}{{ end }}", []string{"content", "messages", "role"}},
-		{"{{ range .Messages }}{{ if eq .Role \"system\" }}SYSTEM: {{ .Content }}{{ else if eq .Role \"user\" }}USER: {{ .Content }}{{ else if eq .Role \"assistant\" }}ASSISTANT: {{ .Content }}{{ end }}{{ end }}", []string{"content", "messages", "role"}},
-		{"{{ .Prompt }} {{ .Suffix }}", []string{"prompt", "suffix"}},
+		{`{{- range .Messages }}
+{{- if eq .Role "system" }}SYSTEM:
+{{- else if eq .Role "user" }}USER:
+{{- else if eq .Role "assistant" }}ASSISTANT:
+{{- end }} {{ .Content }}
+{{- end }}`, []string{"content", "messages", "role"}},
+		{`{{- if .Messages }}
+{{- range .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ else -}}
+{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}{{ if .Prompt }}<|im_start|>user
+{{ .Prompt }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ .Response }}<|im_end|>
+{{- end -}}`, []string{"content", "messages", "prompt", "response", "role", "system"}},
 	}
 
 	for _, tt := range cases {
@@ -80,9 +189,227 @@ func TestParse(t *testing.T) {
 				t.Fatal(err)
 			}
 
-			vars := tmpl.Vars()
-			if !slices.Equal(tt.vars, vars) {
-				t.Errorf("expected %v, got %v", tt.vars, vars)
+			if diff := cmp.Diff(tmpl.Vars(), tt.vars); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestExecuteWithMessages(t *testing.T) {
+	type template struct {
+		name     string
+		template string
+	}
+	cases := []struct {
+		name      string
+		templates []template
+		values    Values
+		expected  string
+	}{
+		{
+			"mistral",
+			[]template{
+				{"no response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}
+{{- range .Messages }}
+{{- if eq .Role "user" }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}[INST] {{ end }}
+{{- end }}`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+				},
+			},
+			`[INST] Hello friend![/INST] Hello human![INST] What is your name?[/INST] `,
+		},
+		{
+			"mistral system",
+			[]template{
+				{"no response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}
+{{- range .Messages }}
+{{- if eq .Role "user" }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}[INST] {{ end }}
+{{- end }}`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "system", Content: "You are a helpful assistant!"},
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+				},
+			},
+			`[INST] You are a helpful assistant!
+
+Hello friend![/INST] Hello human![INST] What is your name?[/INST] `,
+		},
+		{
+			"mistral assistant",
+			[]template{
+				{"no response", `[INST] {{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `
+{{- range $i, $m := .Messages }}
+{{- if eq .Role "user" }}[INST] {{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}{{ end }}
+{{- end }}`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+					{Role: "assistant", Content: "My name is Ollama and I"},
+				},
+			},
+			`[INST] Hello friend![/INST] Hello human![INST] What is your name?[/INST] My name is Ollama and I`,
+		},
+		{
+			"chatml",
+			[]template{
+				// this does not have a "no response" test because it's impossible to render the same output
+				{"response", `{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}{{ if .Prompt }}<|im_start|>user
+{{ .Prompt }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ .Response }}<|im_end|>
+`},
+				{"messages", `
+{{- range $index, $_ := .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
+`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "system", Content: "You are a helpful assistant!"},
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+				},
+			},
+			`<|im_start|>system
+You are a helpful assistant!<|im_end|>
+<|im_start|>user
+Hello friend!<|im_end|>
+<|im_start|>assistant
+Hello human!<|im_end|>
+<|im_start|>user
+What is your name?<|im_end|>
+<|im_start|>assistant
+`,
+		},
+		{
+			"moondream",
+			[]template{
+				// this does not have a "no response" test because it's impossible to render the same output
+				{"response", `{{ if .Prompt }}Question: {{ .Prompt }}
+
+{{ end }}Answer: {{ .Response }}
+
+`},
+				{"messages", `
+{{- range .Messages }}
+{{- if eq .Role "user" }}Question: {{ .Content }}
+
+{{ else if eq .Role "assistant" }}Answer: {{ .Content }}
+
+{{ end }}
+{{- end }}Answer: `},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "user", Content: "What's in this image?", Images: []api.ImageData{[]byte("")}},
+					{Role: "assistant", Content: "It's a hot dog."},
+					{Role: "user", Content: "What's in _this_ image?"},
+					{Role: "user", Images: []api.ImageData{[]byte("")}},
+					{Role: "user", Content: "Is it a hot dog?"},
+				},
+			},
+			`Question: [img-0] What's in this image?
+
+Answer: It's a hot dog.
+
+Question: What's in _this_ image?
+
+[img-1]
+
+Is it a hot dog?
+
+Answer: `,
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			for _, ttt := range tt.templates {
+				t.Run(ttt.name, func(t *testing.T) {
+					tmpl, err := Parse(ttt.template)
+					if err != nil {
+						t.Fatal(err)
+					}
+
+					var b bytes.Buffer
+					if err := tmpl.Execute(&b, tt.values); err != nil {
+						t.Fatal(err)
+					}
+
+					if diff := cmp.Diff(b.String(), tt.expected); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				})
+			}
+		})
+	}
+}
+
+func TestExecuteWithSuffix(t *testing.T) {
+	tmpl, err := Parse(`{{- if .Suffix }}<PRE> {{ .Prompt }} <SUF>{{ .Suffix }} <MID>
+{{- else }}{{ .Prompt }}
+{{- end }}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	cases := []struct {
+		name   string
+		values Values
+		expect string
+	}{
+		{
+			"message", Values{Messages: []api.Message{{Role: "user", Content: "hello"}}}, "hello",
+		},
+		{
+			"prompt suffix", Values{Prompt: "def add(", Suffix: "return x"}, "<PRE> def add( <SUF>return x <MID>",
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			var b bytes.Buffer
+			if err := tmpl.Execute(&b, tt.values); err != nil {
+				t.Fatal(err)
+			}
+
+			if diff := cmp.Diff(b.String(), tt.expect); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
 			}
 		})
 	}

template/testdata/alfred.gotmpl/system-user-assistant-user

@@ -0,0 +1 @@
+<start_system>You are a helpful assistant.<end_message><start_user>Hello, how are you?<end_message><start_assistant>I'm doing great. How can I help you today?<end_message><start_user>I'd like to show off how chat templating works!<end_message><start_assistant>

template/testdata/alfred.gotmpl/user

@@ -0,0 +1 @@
+<start_user>Hello, how are you?<end_message><start_assistant>

template/testdata/alfred.gotmpl/user-assistant-user

@@ -0,0 +1 @@
+<start_user>Hello, how are you?<end_message><start_assistant>I'm doing great. How can I help you today?<end_message><start_user>I'd like to show off how chat templating works!<end_message><start_assistant>

template/testdata/alpaca.gotmpl/system-user-assistant-user

@@ -0,0 +1,12 @@
+You are a helpful assistant.
+
+### Instruction:
+Hello, how are you?
+
+### Response:
+I'm doing great. How can I help you today?
+
+### Instruction:
+I'd like to show off how chat templating works!
+
+### Response:

template/testdata/alpaca.gotmpl/user

@@ -0,0 +1,4 @@
+### Instruction:
+Hello, how are you?
+
+### Response:

template/testdata/alpaca.gotmpl/user-assistant-user

@@ -0,0 +1,10 @@
+### Instruction:
+Hello, how are you?
+
+### Response:
+I'm doing great. How can I help you today?
+
+### Instruction:
+I'd like to show off how chat templating works!
+
+### Response:

template/testdata/chatml.gotmpl/system-user-assistant-user

@@ -0,0 +1,9 @@
+<|im_start|>system
+You are a helpful assistant.<|im_end|>
+<|im_start|>user
+Hello, how are you?<|im_end|>
+<|im_start|>assistant
+I'm doing great. How can I help you today?<|im_end|>
+<|im_start|>user
+I'd like to show off how chat templating works!<|im_end|>
+<|im_start|>assistant

template/testdata/chatml.gotmpl/user

@@ -0,0 +1,3 @@
+<|im_start|>user
+Hello, how are you?<|im_end|>
+<|im_start|>assistant

template/testdata/chatml.gotmpl/user-assistant-user

@@ -0,0 +1,7 @@
+<|im_start|>user
+Hello, how are you?<|im_end|>
+<|im_start|>assistant
+I'm doing great. How can I help you today?<|im_end|>
+<|im_start|>user
+I'd like to show off how chat templating works!<|im_end|>
+<|im_start|>assistant

template/testdata/chatqa.gotmpl/system-user-assistant-user

@@ -0,0 +1,9 @@
+System: You are a helpful assistant.
+
+User: Hello, how are you?
+
+Assistant: I'm doing great. How can I help you today?
+
+User: I'd like to show off how chat templating works!
+
+Assistant:

template/testdata/chatqa.gotmpl/user

@@ -0,0 +1,3 @@
+User: Hello, how are you?
+
+Assistant:

template/testdata/chatqa.gotmpl/user-assistant-user

@@ -0,0 +1,7 @@
+User: Hello, how are you?
+
+Assistant: I'm doing great. How can I help you today?
+
+User: I'd like to show off how chat templating works!
+
+Assistant:

template/testdata/codellama-70b-instruct.gotmpl/system-user-assistant-user

@@ -0,0 +1,12 @@
+Source: system
+
+ You are a helpful assistant. <step> Source: user
+
+ Hello, how are you? <step> Source: assistant
+
+ I'm doing great. How can I help you today? <step> Source: user
+
+ I'd like to show off how chat templating works! <step> Source: assistant
+Destination: user
+
+ 

template/testdata/codellama-70b-instruct.gotmpl/user

@@ -0,0 +1,6 @@
+Source: user
+
+ Hello, how are you? <step> Source: assistant
+Destination: user
+
+ 

template/testdata/codellama-70b-instruct.gotmpl/user-assistant-user

@@ -0,0 +1,10 @@
+Source: user
+
+ Hello, how are you? <step> Source: assistant
+
+ I'm doing great. How can I help you today? <step> Source: user
+
+ I'd like to show off how chat templating works! <step> Source: assistant
+Destination: user
+
+ 

template/testdata/falcon-instruct.gotmpl/system-user-assistant-user

@@ -0,0 +1,8 @@
+System: You are a helpful assistant.
+User:
+Hello, how are you?
+Falcon:
+I'm doing great. How can I help you today?
+User:
+I'd like to show off how chat templating works!
+Falcon:

template/testdata/falcon-instruct.gotmpl/user

@@ -0,0 +1,3 @@
+User:
+Hello, how are you?
+Falcon:

template/testdata/falcon-instruct.gotmpl/user-assistant-user

@@ -0,0 +1,7 @@
+User:
+Hello, how are you?
+Falcon:
+I'm doing great. How can I help you today?
+User:
+I'd like to show off how chat templating works!
+Falcon:

template/testdata/gemma-instruct.gotmpl/system-user-assistant-user

@@ -0,0 +1,8 @@
+<start_of_turn>user
+You are a helpful assistant.
+Hello, how are you?<end_of_turn>
+<start_of_turn>model
+I'm doing great. How can I help you today?<end_of_turn>
+<start_of_turn>user
+I'd like to show off how chat templating works!<end_of_turn>
+<start_of_turn>model

template/testdata/gemma-instruct.gotmpl/user

@@ -0,0 +1,3 @@
+<start_of_turn>user
+Hello, how are you?<end_of_turn>
+<start_of_turn>model