...

When starting the app in the background, start it hidden.

CMakeLists.txt

@@ -51,6 +51,8 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx)
 
+add_compile_definitions(NDEBUG)
+
 set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)

README.md

@@ -405,6 +405,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Writeopia](https://github.com/Writeopia/Writeopia) (Text editor with integration with Ollama)
 - [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
+- [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
+- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
 
 ### Cloud
 
@@ -448,6 +450,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.
 - [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull, and download models from Ollama Registry in your terminal.
 - [GGUF-to-Ollama](https://github.com/jonathanhecl/gguf-to-ollama) - Importing GGUF to Ollama made easy (multiplatform)
+- [AWS-Strands-With-Ollama](https://github.com/rapidarchitect/ollama_strands) - AWS Strands Agents with Ollama Examples
 
 ### Apple Vision Pro
 
@@ -584,6 +587,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Simple-Discord-AI](https://github.com/zyphixor/simple-discord-ai)
 - [LLM Telegram Bot](https://github.com/innightwolfsleep/llm_telegram_bot) (telegram bot, primary for RP. Oobabooga-like buttons, [A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) API integration e.t.c)
 - [mcp-llm](https://github.com/sammcj/mcp-llm) (MCP Server to allow LLMs to call other LLMs)
+- [SimpleOllamaUnity](https://github.com/HardCodeDev777/SimpleOllamaUnity) (Unity Engine extension for communicating with Ollama in a few lines of code. Also works at runtime)
 - [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)
 
 ### Supported backends

api/client.go

@@ -24,7 +24,10 @@ import (
 	"net/http"
 	"net/url"
 	"runtime"
+	"strconv"
+	"time"
 
+	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/version"
@@ -76,6 +79,14 @@ func NewClient(base *url.URL, http *http.Client) *Client {
 	}
 }
 
+func getAuthorizationToken(ctx context.Context, challenge string) (string, error) {
+	token, err := auth.Sign(ctx, []byte(challenge))
+	if err != nil {
+		return "", err
+	}
+	return token, nil
+}
+
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
 	var reqBody io.Reader
 	var data []byte
@@ -97,6 +108,21 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	}
 
 	requestURL := c.base.JoinPath(path)
+
+	var token string
+	if envconfig.UseAuth() || c.base.Hostname() == "ollama.com" {
+		now := strconv.FormatInt(time.Now().Unix(), 10)
+		chal := fmt.Sprintf("%s,%s?ts=%s", method, path, now)
+		token, err = getAuthorizationToken(ctx, chal)
+		if err != nil {
+			return err
+		}
+
+		q := requestURL.Query()
+		q.Set("ts", now)
+		requestURL.RawQuery = q.Encode()
+	}
+
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), reqBody)
 	if err != nil {
 		return err
@@ -106,6 +132,10 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	request.Header.Set("Accept", "application/json")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
+	if token != "" {
+		request.Header.Set("Authorization", token)
+	}
+
 	respObj, err := c.http.Do(request)
 	if err != nil {
 		return err
@@ -143,6 +173,22 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	}
 
 	requestURL := c.base.JoinPath(path)
+
+	var token string
+	if envconfig.UseAuth() || c.base.Hostname() == "ollama.com" {
+		var err error
+		now := strconv.FormatInt(time.Now().Unix(), 10)
+		chal := fmt.Sprintf("%s,%s?ts=%s", method, path, now)
+		token, err = getAuthorizationToken(ctx, chal)
+		if err != nil {
+			return err
+		}
+
+		q := requestURL.Query()
+		q.Set("ts", now)
+		requestURL.RawQuery = q.Encode()
+	}
+
 	request, err := http.NewRequestWithContext(ctx, method, requestURL.String(), buf)
 	if err != nil {
 		return err
@@ -152,6 +198,10 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	request.Header.Set("Accept", "application/x-ndjson")
 	request.Header.Set("User-Agent", fmt.Sprintf("ollama/%s (%s %s) Go/%s", version.Version, runtime.GOARCH, runtime.GOOS, runtime.Version()))
 
+	if token != "" {
+		request.Header.Set("Authorization", token)
+	}
+
 	response, err := c.http.Do(request)
 	if err != nil {
 		return err

api/types.go

@@ -83,6 +83,12 @@ type GenerateRequest struct {
 	// Options lists model-specific options. For example, temperature can be
 	// set through this field, if the model supports it.
 	Options map[string]any `json:"options"`
+
+	// Think controls whether thinking/reasoning models will think before
+	// responding. Needs to be a pointer so we can distinguish between false
+	// (request that thinking _not_ be used) and unset (use the old behavior
+	// before this option was introduced)
+	Think *bool `json:"think,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -108,6 +114,10 @@ type ChatRequest struct {
 
 	// Options lists model-specific options.
 	Options map[string]any `json:"options"`
+
+	// Think controls whether thinking/reasoning models will think before
+	// responding
+	Think *bool `json:"think,omitempty"`
 }
 
 type Tools []Tool
@@ -126,8 +136,11 @@ func (t Tool) String() string {
 // role ("system", "user", or "assistant"), the content and an optional list
 // of images.
 type Message struct {
-	Role      string      `json:"role"`
-	Content   string      `json:"content"`
+	Role    string `json:"role"`
+	Content string `json:"content"`
+	// Thinking contains the text that was inside thinking tags in the
+	// original model output when ChatRequest.Think is enabled.
+	Thinking  string      `json:"thinking,omitempty"`
 	Images    []ImageData `json:"images,omitempty"`
 	ToolCalls []ToolCall  `json:"tool_calls,omitempty"`
 }
@@ -444,12 +457,13 @@ type ProcessResponse struct {
 
 // ListModelResponse is a single model description in [ListResponse].
 type ListModelResponse struct {
-	Name       string       `json:"name"`
-	Model      string       `json:"model"`
-	ModifiedAt time.Time    `json:"modified_at"`
-	Size       int64        `json:"size"`
-	Digest     string       `json:"digest"`
-	Details    ModelDetails `json:"details,omitempty"`
+	Name         string             `json:"name"`
+	Model        string             `json:"model"`
+	ModifiedAt   time.Time          `json:"modified_at"`
+	Size         int64              `json:"size"`
+	Digest       string             `json:"digest"`
+	Capabilities []model.Capability `json:"capabilities,omitempty"`
+	Details      ModelDetails       `json:"details,omitempty"`
 }
 
 // ProcessModelResponse is a single model description in [ProcessResponse].
@@ -478,6 +492,10 @@ type GenerateResponse struct {
 	// Response is the textual response itself.
 	Response string `json:"response"`
 
+	// Thinking contains the text that was inside thinking tags in the
+	// original model output when ChatRequest.Think is enabled.
+	Thinking string `json:"thinking,omitempty"`
+
 	// Done specifies if the response is complete.
 	Done bool `json:"done"`
 

api/types_test.go

@@ -372,3 +372,50 @@ func TestPropertyType_MarshalJSON(t *testing.T) {
 		})
 	}
 }
+
+func TestThinking_UnmarshalJSON(t *testing.T) {
+	trueVal := true
+	falseVal := false
+
+	tests := []struct {
+		name             string
+		input            string
+		expectedThinking *bool
+		expectedError    bool
+	}{
+		{
+			name:             "true",
+			input:            `{ "think": true }`,
+			expectedThinking: &trueVal,
+		},
+		{
+			name:             "false",
+			input:            `{ "think": false }`,
+			expectedThinking: &falseVal,
+		},
+		{
+			name:             "unset",
+			input:            `{ }`,
+			expectedThinking: nil,
+		},
+		{
+			name:             "invalid",
+			input:            `{ "think": "true" }`,
+			expectedThinking: nil,
+			expectedError:    true,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var req GenerateRequest
+			err := json.Unmarshal([]byte(test.input), &req)
+			if test.expectedError {
+				require.Error(t, err)
+			} else {
+				require.NoError(t, err)
+				assert.Equal(t, test.expectedThinking, req.Think)
+			}
+		})
+	}
+}

cmd/cmd.go

@@ -39,6 +39,7 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/runner"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
@@ -46,6 +47,23 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
+// ensureThinkingSupport emits a warning if the model does not advertise thinking support
+func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
+	if name == "" {
+		return
+	}
+	resp, err := client.Show(ctx, &api.ShowRequest{Model: name})
+	if err != nil {
+		return
+	}
+	for _, cap := range resp.Capabilities {
+		if cap == model.CapabilityThinking {
+			return
+		}
+	}
+	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
+}
+
 var errModelfileNotFound = errors.New("specified Modelfile wasn't found")
 
 func getModelfileName(cmd *cobra.Command) (string, error) {
@@ -265,6 +283,9 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 	req := &api.GenerateRequest{
 		Model:     opts.Model,
 		KeepAlive: opts.KeepAlive,
+
+		// pass Think here so we fail before getting to the chat prompt if the model doesn't support it
+		Think: opts.Think,
 	}
 
 	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
@@ -299,6 +320,22 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.Format = format
 
+	thinkFlag := cmd.Flags().Lookup("think")
+	if thinkFlag.Changed {
+		think, err := cmd.Flags().GetBool("think")
+		if err != nil {
+			return err
+		}
+		opts.Think = &think
+	} else {
+		opts.Think = nil
+	}
+	hidethinking, err := cmd.Flags().GetBool("hidethinking")
+	if err != nil {
+		return err
+	}
+	opts.HideThinking = hidethinking
+
 	keepAlive, err := cmd.Flags().GetString("keepalive")
 	if err != nil {
 		return err
@@ -362,6 +399,11 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	opts.Think, err = inferThinkingOption(&info.Capabilities, &opts, thinkFlag.Changed)
+	if err != nil {
+		return err
+	}
+
 	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision)
 
 	// TODO: remove the projector info and vision info checks below,
@@ -747,11 +789,38 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 				case float64:
 					v = fmt.Sprintf("%g", vData)
 				case []any:
-					n := 3
-					if len(vData) < n {
-						n = len(vData)
+					targetWidth := 10 // Small width where we are displaying the data in a column
+
+					var itemsToShow int
+					totalWidth := 1 // Start with 1 for opening bracket
+
+					// Find how many we can fit
+					for i := range vData {
+						itemStr := fmt.Sprintf("%v", vData[i])
+						width := runewidth.StringWidth(itemStr)
+
+						// Add separator width (", ") for all items except the first
+						if i > 0 {
+							width += 2
+						}
+
+						// Check if adding this item would exceed our width limit
+						if totalWidth+width > targetWidth && i > 0 {
+							break
+						}
+
+						totalWidth += width
+						itemsToShow++
+					}
+
+					// Format the output
+					if itemsToShow < len(vData) {
+						v = fmt.Sprintf("%v", vData[:itemsToShow])
+						v = strings.TrimSuffix(v, "]")
+						v += fmt.Sprintf(" ...+%d more]", len(vData)-itemsToShow)
+					} else {
+						v = fmt.Sprintf("%v", vData)
 					}
-					v = fmt.Sprintf("%v", vData[:n])
 				default:
 					v = fmt.Sprintf("%T", vData)
 				}
@@ -772,10 +841,19 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 
 	head := func(s string, n int) (rows [][]string) {
 		scanner := bufio.NewScanner(strings.NewReader(s))
-		for scanner.Scan() && (len(rows) < n || n < 0) {
-			if text := scanner.Text(); text != "" {
-				rows = append(rows, []string{"", strings.TrimSpace(text)})
+		count := 0
+		for scanner.Scan() {
+			text := strings.TrimSpace(scanner.Text())
+			if text == "" {
+				continue
 			}
+			count++
+			if n < 0 || count <= n {
+				rows = append(rows, []string{"", text})
+			}
+		}
+		if n >= 0 && count > n {
+			rows = append(rows, []string{"", "..."})
 		}
 		return
 	}
@@ -887,17 +965,19 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 type generateContextKey string
 
 type runOptions struct {
-	Model       string
-	ParentModel string
-	Prompt      string
-	Messages    []api.Message
-	WordWrap    bool
-	Format      string
-	System      string
-	Images      []api.ImageData
-	Options     map[string]any
-	MultiModal  bool
-	KeepAlive   *api.Duration
+	Model        string
+	ParentModel  string
+	Prompt       string
+	Messages     []api.Message
+	WordWrap     bool
+	Format       string
+	System       string
+	Images       []api.ImageData
+	Options      map[string]any
+	MultiModal   bool
+	KeepAlive    *api.Duration
+	Think        *bool
+	HideThinking bool
 }
 
 type displayResponseState struct {
@@ -953,6 +1033,26 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	}
 }
 
+func thinkingOutputOpeningText(plainText bool) string {
+	text := "Thinking...\n"
+
+	if plainText {
+		return text
+	}
+
+	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault + readline.ColorGrey
+}
+
+func thinkingOutputClosingText(plainText bool) string {
+	text := "...done thinking.\n\n"
+
+	if plainText {
+		return text
+	}
+
+	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault
+}
+
 func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -980,14 +1080,34 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	var latest api.ChatResponse
 	var fullResponse strings.Builder
 	var role string
+	var thinkTagOpened bool = false
+	var thinkTagClosed bool = false
 
 	fn := func(response api.ChatResponse) error {
-		p.StopAndClear()
+		if response.Message.Content != "" || !opts.HideThinking {
+			p.StopAndClear()
+		}
 
 		latest = response
 
 		role = response.Message.Role
+		if response.Message.Thinking != "" && !opts.HideThinking {
+			if !thinkTagOpened {
+				fmt.Print(thinkingOutputOpeningText(false))
+				thinkTagOpened = true
+			}
+			displayResponse(response.Message.Thinking, opts.WordWrap, state)
+		}
+
 		content := response.Message.Content
+		if thinkTagOpened && !thinkTagClosed && content != "" {
+			fmt.Print(thinkingOutputClosingText(false))
+			thinkTagClosed = true
+		}
+		// purposefully not putting thinking blocks in the response, which would
+		// only be needed if we later added tool calling to the cli (they get
+		// filtered out anyway since current models don't expect them unless you're
+		// about to finish some tool calls)
 		fullResponse.WriteString(content)
 
 		displayResponse(content, opts.WordWrap, state)
@@ -1004,6 +1124,7 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		Messages: opts.Messages,
 		Format:   json.RawMessage(opts.Format),
 		Options:  opts.Options,
+		Think:    opts.Think,
 	}
 
 	if opts.KeepAlive != nil {
@@ -1065,13 +1186,32 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	}()
 
 	var state *displayResponseState = &displayResponseState{}
+	var thinkTagOpened bool = false
+	var thinkTagClosed bool = false
+
+	plainText := !term.IsTerminal(int(os.Stdout.Fd()))
 
 	fn := func(response api.GenerateResponse) error {
-		p.StopAndClear()
-
 		latest = response
 		content := response.Response
 
+		if response.Response != "" || !opts.HideThinking {
+			p.StopAndClear()
+		}
+
+		if response.Thinking != "" && !opts.HideThinking {
+			if !thinkTagOpened {
+				fmt.Print(thinkingOutputOpeningText(plainText))
+				thinkTagOpened = true
+			}
+			displayResponse(response.Thinking, opts.WordWrap, state)
+		}
+
+		if thinkTagOpened && !thinkTagClosed && content != "" {
+			fmt.Print(thinkingOutputClosingText(plainText))
+			thinkTagClosed = true
+		}
+
 		displayResponse(content, opts.WordWrap, state)
 
 		return nil
@@ -1097,6 +1237,7 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		System:    opts.System,
 		Options:   opts.Options,
 		KeepAlive: opts.KeepAlive,
+		Think:     opts.Think,
 	}
 
 	if err := client.Generate(ctx, &request, fn); err != nil {
@@ -1200,11 +1341,11 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 		return err
 	}
 	if err := client.Heartbeat(cmd.Context()); err != nil {
-		if !strings.Contains(err.Error(), " refused") {
+		if !(strings.Contains(err.Error(), " refused") || strings.Contains(err.Error(), "could not connect")) {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return errors.New("could not connect to ollama app, is it running?")
+			return fmt.Errorf("ollama server not responding - %w", err)
 		}
 	}
 	return nil
@@ -1282,7 +1423,7 @@ func NewCLI() *cobra.Command {
 	}
 
 	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\"")
-	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")
+	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1312,6 +1453,8 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 	runCmd.Flags().String("format", "", "Response format (e.g. json)")
+	runCmd.Flags().Bool("think", false, "Whether to use thinking mode for supported models")
+	runCmd.Flags().Bool("hidethinking", false, "Hide thinking output (if provided)")
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",
@@ -1363,7 +1506,6 @@ func NewCLI() *cobra.Command {
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListRunningHandler,
 	}
-
 	copyCmd := &cobra.Command{
 		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
@@ -1452,3 +1594,45 @@ func NewCLI() *cobra.Command {
 
 	return rootCmd
 }
+
+// If the user has explicitly set thinking options, either through the CLI or
+// through the `/set think` or `set nothink` interactive options, then we
+// respect them. Otherwise, we check model capabilities to see if the model
+// supports thinking. If the model does support thinking, we enable it.
+// Otherwise, we unset the thinking option (which is different than setting it
+// to false).
+//
+// If capabilities are not provided, we fetch them from the server.
+func inferThinkingOption(caps *[]model.Capability, runOpts *runOptions, explicitlySetByUser bool) (*bool, error) {
+	if explicitlySetByUser {
+		return runOpts.Think, nil
+	}
+
+	if caps == nil {
+		client, err := api.ClientFromEnvironment()
+		if err != nil {
+			return nil, err
+		}
+		ret, err := client.Show(context.Background(), &api.ShowRequest{
+			Model: runOpts.Model,
+		})
+		if err != nil {
+			return nil, err
+		}
+		caps = &ret.Capabilities
+	}
+
+	thinkingSupported := false
+	for _, cap := range *caps {
+		if cap == model.CapabilityThinking {
+			thinkingSupported = true
+		}
+	}
+
+	if thinkingSupported {
+		thinking := true
+		return &thinking, nil
+	}
+
+	return nil, nil
+}

cmd/cmd_test.go

@@ -225,6 +225,7 @@ Weigh anchor!
   System
     You are a pirate!    
     Ahoy, matey!         
+    ...                  
 
 `
 		if diff := cmp.Diff(expect, b.String()); diff != "" {

cmd/interactive.go

@@ -62,6 +62,8 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /set noformat          Disable formatting")
 		fmt.Fprintln(os.Stderr, "  /set verbose           Show LLM stats")
 		fmt.Fprintln(os.Stderr, "  /set quiet             Disable LLM stats")
+		fmt.Fprintln(os.Stderr, "  /set think             Enable thinking")
+		fmt.Fprintln(os.Stderr, "  /set nothink           Disable thinking")
 		fmt.Fprintln(os.Stderr, "")
 	}
 
@@ -128,6 +130,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 	var sb strings.Builder
 	var multiline MultilineState
+	var thinkExplicitlySet bool = opts.Think != nil
 
 	for {
 		line, err := scanner.Readline()
@@ -195,11 +198,19 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Model = args[1]
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
+			opts.Think, err = inferThinkingOption(nil, &opts, thinkExplicitlySet)
+			if err != nil {
+				return err
+			}
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				if strings.Contains(err.Error(), "not found") {
 					fmt.Printf("error: %v\n", err)
 					continue
 				}
+				if strings.Contains(err.Error(), "does not support thinking") {
+					fmt.Printf("error: %v\n", err)
+					continue
+				}
 				return err
 			}
 			continue
@@ -260,6 +271,22 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						return err
 					}
 					fmt.Println("Set 'quiet' mode.")
+				case "think":
+					think := true
+					opts.Think = &think
+					thinkExplicitlySet = true
+					if client, err := api.ClientFromEnvironment(); err == nil {
+						ensureThinkingSupport(cmd.Context(), client, opts.Model)
+					}
+					fmt.Println("Set 'think' mode.")
+				case "nothink":
+					think := false
+					opts.Think = &think
+					thinkExplicitlySet = true
+					if client, err := api.ClientFromEnvironment(); err == nil {
+						ensureThinkingSupport(cmd.Context(), client, opts.Model)
+					}
+					fmt.Println("Set 'nothink' mode.")
 				case "format":
 					if len(args) < 3 || args[2] != "json" {
 						fmt.Println("Invalid or missing format. For 'json' mode use '/set format json'")
@@ -448,6 +475,11 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 			assistant, err := chat(cmd, opts)
 			if err != nil {
+				if strings.Contains(err.Error(), "does not support thinking") {
+					fmt.Printf("error: %v\n", err)
+					sb.Reset()
+					continue
+				}
 				return err
 			}
 			if assistant != nil {

cmd/start_darwin.go

@@ -23,7 +23,7 @@ func startApp(ctx context.Context, client *api.Client) error {
 		return errors.New("could not find ollama app")
 	}
 	path := strings.Split(link, "Ollama.app")
-	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
+	if err := exec.Command("/usr/bin/open", "-j", "-a", path[0]+"Ollama.app").Run(); err != nil {
 		return err
 	}
 	return waitForServer(ctx, client)

cmd/start_windows.go

@@ -4,17 +4,27 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"log/slog"
 	"os"
 	"os/exec"
+	"path"
 	"path/filepath"
 	"strings"
 	"syscall"
+	"unsafe"
 
 	"github.com/ollama/ollama/api"
+	"golang.org/x/sys/windows"
+)
+
+const (
+	Installer = "OllamaSetup.exe"
 )
 
 func startApp(ctx context.Context, client *api.Client) error {
-	// log.Printf("XXX Attempting to find and start ollama app")
+	if len(isProcRunning(Installer)) > 0 {
+		return fmt.Errorf("upgrade in progress...")
+	}
 	AppName := "ollama app.exe"
 	exe, err := os.Executable()
 	if err != nil {
@@ -35,14 +45,11 @@ func startApp(ctx context.Context, client *api.Client) error {
 			}
 		}
 	}
-	// log.Printf("XXX attempting to start app %s", appExe)
 
 	cmd_path := "c:\\Windows\\system32\\cmd.exe"
-	cmd := exec.Command(cmd_path, "/c", appExe)
-	// TODO - these hide flags aren't working - still pops up a command window for some reason
+	cmd := exec.Command(cmd_path, "/c", appExe, "hidden")
 	cmd.SysProcAttr = &syscall.SysProcAttr{CreationFlags: 0x08000000, HideWindow: true}
 
-	// TODO this didn't help either...
 	cmd.Stdin = strings.NewReader("")
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
@@ -56,3 +63,50 @@ func startApp(ctx context.Context, client *api.Client) error {
 	}
 	return waitForServer(ctx, client)
 }
+
+func isProcRunning(procName string) []uint32 {
+	pids := make([]uint32, 2048)
+	var ret uint32
+	if err := windows.EnumProcesses(pids, &ret); err != nil || ret == 0 {
+		slog.Debug("failed to check for running installers", "error", err)
+		return nil
+	}
+	if ret > uint32(len(pids)) {
+		pids = make([]uint32, ret+10)
+		if err := windows.EnumProcesses(pids, &ret); err != nil || ret == 0 {
+			slog.Debug("failed to check for running installers", "error", err)
+			return nil
+		}
+	}
+	if ret < uint32(len(pids)) {
+		pids = pids[:ret]
+	}
+	var matches []uint32
+	for _, pid := range pids {
+		if pid == 0 {
+			continue
+		}
+		hProcess, err := windows.OpenProcess(windows.PROCESS_QUERY_INFORMATION|windows.PROCESS_VM_READ, false, pid)
+		if err != nil {
+			continue
+		}
+		defer windows.CloseHandle(hProcess)
+		var module windows.Handle
+		var cbNeeded uint32
+		cb := (uint32)(unsafe.Sizeof(module))
+		if err := windows.EnumProcessModules(hProcess, &module, cb, &cbNeeded); err != nil {
+			continue
+		}
+		var sz uint32 = 1024 * 8
+		moduleName := make([]uint16, sz)
+		cb = uint32(len(moduleName)) * (uint32)(unsafe.Sizeof(uint16(0)))
+		if err := windows.GetModuleBaseName(hProcess, module, &moduleName[0], cb); err != nil && err != syscall.ERROR_INSUFFICIENT_BUFFER {
+			continue
+		}
+		exeFile := path.Base(strings.ToLower(syscall.UTF16ToString(moduleName)))
+		if strings.EqualFold(exeFile, procName) {
+			matches = append(matches, pid)
+		}
+	}
+	return matches
+}

cmd/warn_thinking_test.go

@@ -0,0 +1,63 @@
+package cmd
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
+)
+
+// Test that a warning is printed when thinking is requested but not supported.
+func TestWarnMissingThinking(t *testing.T) {
+	cases := []struct {
+		capabilities []model.Capability
+		expectWarn   bool
+	}{
+		{capabilities: []model.Capability{model.CapabilityThinking}, expectWarn: false},
+		{capabilities: []model.Capability{}, expectWarn: true},
+	}
+
+	for _, tc := range cases {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if r.URL.Path != "/api/show" || r.Method != http.MethodPost {
+				t.Fatalf("unexpected request to %s %s", r.URL.Path, r.Method)
+			}
+			var req api.ShowRequest
+			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+				t.Fatalf("decode request: %v", err)
+			}
+			resp := api.ShowResponse{Capabilities: tc.capabilities}
+			if err := json.NewEncoder(w).Encode(resp); err != nil {
+				t.Fatalf("encode response: %v", err)
+			}
+		}))
+		defer srv.Close()
+
+		t.Setenv("OLLAMA_HOST", srv.URL)
+		client, err := api.ClientFromEnvironment()
+		if err != nil {
+			t.Fatal(err)
+		}
+		oldStderr := os.Stderr
+		r, w, _ := os.Pipe()
+		os.Stderr = w
+		ensureThinkingSupport(t.Context(), client, "m")
+		w.Close()
+		os.Stderr = oldStderr
+		out, _ := io.ReadAll(r)
+
+		warned := strings.Contains(string(out), "warning:")
+		if tc.expectWarn && !warned {
+			t.Errorf("expected warning, got none")
+		}
+		if !tc.expectWarn && warned {
+			t.Errorf("did not expect warning, got: %s", string(out))
+		}
+	}
+}

convert/convert.go

@@ -53,8 +53,11 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	}
 
 	for _, sv := range t.SpecialVocabulary {
-		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
 		kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
+		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
+		if len(sv.IDs) > 0 {
+			kv[fmt.Sprintf("tokenizer.ggml.%s_token_ids", sv.Key())] = sv.IDs
+		}
 	}
 
 	return kv

convert/convert_llama.go

@@ -139,7 +139,8 @@ func (p *llamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	}
 
 	for _, t := range ts {
-		if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
+		if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") ||
+			strings.HasSuffix(t.Name(), "attn_q_proj.weight") || strings.HasSuffix(t.Name(), "attn_k_proj.weight") {
 			if !p.skipRepack {
 				t.SetRepacker(p.repack)
 			}
@@ -181,9 +182,9 @@ func (p *llamaModel) repack(name string, data []float32, shape []uint64) ([]floa
 	}
 
 	var heads uint32
-	if strings.HasSuffix(name, "attn_q.weight") {
+	if strings.HasSuffix(name, "attn_q.weight") || strings.HasSuffix(name, "attn_q_proj.weight") {
 		heads = p.NumAttentionHeads
-	} else if strings.HasSuffix(name, "attn_k.weight") {
+	} else if strings.HasSuffix(name, "attn_k.weight") || strings.HasSuffix(name, "attn_k_proj.weight") {
 		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
 	} else {
 		return nil, fmt.Errorf("unknown tensor for repack: %s", name)

convert/convert_mllama.go

@@ -94,7 +94,9 @@ func (m *mllamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	var out []*ggml.Tensor
 	var text []Tensor
 	for _, t := range ts {
-		if t.Name() == "v.position_embd.gate" {
+		if !strings.HasPrefix(t.Name(), "v.") && !strings.HasPrefix(t.Name(), "mm.") {
+			text = append(text, t)
+		} else if t.Name() == "v.position_embd.gate" {
 			for _, name := range []string{"v.position_embd.gate", "v.tile_position_embd.gate"} {
 				tt := t.Clone()
 				tt.SetRepacker(m.repack(name))
@@ -105,23 +107,21 @@ func (m *mllamaModel) Tensors(ts []Tensor) []*ggml.Tensor {
 					WriterTo: tt,
 				})
 			}
-		} else if t.Name() == "v.pre_tile_position_embd.gate" || t.Name() == "v.post_tile_position_embd.gate" {
-			t.SetRepacker(m.repack(t.Name()))
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
-		} else if strings.HasPrefix(t.Name(), "v.") || strings.HasPrefix(t.Name(), "mm.") {
-			out = append(out, &ggml.Tensor{
-				Name:     t.Name(),
-				Kind:     t.Kind(),
-				Shape:    t.Shape(),
-				WriterTo: t,
-			})
 		} else {
-			text = append(text, t)
+			if t.Name() == "v.pre_tile_position_embd.gate" || t.Name() == "v.post_tile_position_embd.gate" {
+				t.SetRepacker(m.repack(t.Name()))
+			} else if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
+				t.SetRepacker(m.repack(t.Name()))
+			} else if strings.HasSuffix(t.Name(), "attn_gate") || strings.HasSuffix(t.Name(), "ffn_gate") {
+				t.SetRepacker(m.repack(t.Name()))
+			}
+
+			out = append(out, &ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
 		}
 	}
 
@@ -137,16 +137,35 @@ func (m *mllamaModel) repack(name string) Repacker {
 
 		var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
 
-		t, err = tensor.Tanh(t)
-		if err != nil {
-			return nil, err
-		}
+		if strings.HasSuffix(name, "attn_q.weight") || strings.HasSuffix(name, "attn_k.weight") {
+			heads := m.VisionModel.AttentionHeads
+			if err := t.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+				return nil, err
+			}
 
-		if name == "v.position_embd.gate" {
-			t, err = tensor.Sub(float32(1), t)
+			if err := t.T(0, 2, 1, 3); err != nil {
+				return nil, err
+			}
+
+			if err := t.Reshape(dims...); err != nil {
+				return nil, err
+			}
+
+			if err := t.Transpose(); err != nil {
+				return nil, err
+			}
+		} else {
+			t, err = tensor.Tanh(t)
 			if err != nil {
 				return nil, err
 			}
+
+			if name == "v.position_embd.gate" {
+				t, err = tensor.Sub(float32(1), t)
+				if err != nil {
+					return nil, err
+				}
+			}
 		}
 
 		t = tensor.Materialize(t)

convert/convert_test.go

@@ -47,7 +47,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	}
 	t.Cleanup(func() { r.Close() })
 
-	m, _, err := ggml.Decode(r, -1)
+	m, err := ggml.Decode(r, -1)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -332,7 +332,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := ggml.Decode(r, -1)
+			m, err := ggml.Decode(r, -1)
 			if err != nil {
 				t.Fatal(err)
 			}

convert/tokenizer.go

@@ -110,6 +110,7 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 	}
 
 	if f, err := fsys.Open("tokenizer_config.json"); errors.Is(err, os.ErrNotExist) {
+		// noop
 	} else if err != nil {
 		return nil, err
 	} else {
@@ -171,6 +172,34 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 		}
 	}
 
+	if f, err := fsys.Open("generation_config.json"); errors.Is(err, os.ErrNotExist) {
+	} else if err != nil {
+		return nil, err
+	} else {
+		defer f.Close()
+
+		var p map[string]json.RawMessage
+		if err := json.NewDecoder(f).Decode(&p); err != nil {
+			return nil, err
+		}
+
+		for _, st := range specialTokenTypes {
+			if bts, ok := p[fmt.Sprintf("%s_token_id", st)]; ok {
+				var ids []int32
+				if err := json.Unmarshal(bts, &ids); err != nil {
+					// value is not a list so the existing ID is used
+					continue
+				}
+
+				if i := slices.IndexFunc(t.SpecialVocabulary, func(sv *SpecialVocabulary) bool {
+					return sv.Type == st
+				}); i >= 0 {
+					t.SpecialVocabulary[i].IDs = ids
+				}
+			}
+		}
+	}
+
 	return t, nil
 }
 
@@ -280,6 +309,9 @@ type SpecialVocabulary struct {
 	ID       int
 	Content  string
 	AddToken bool
+
+	// IDs is populated by generation_config.json
+	IDs []int32
 }
 
 func (sv SpecialVocabulary) Key() string {

convert/tokenizer_test.go

@@ -247,6 +247,67 @@ func TestParseTokenizer(t *testing.T) {
 				Pre: "default",
 			},
 		},
+		{
+			name: "generation config eos token ids",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 0,
+							"content": "<bos>",
+							"special": true
+						},
+						{
+							"id": 1,
+							"content": "<eos>",
+							"special": true
+						},
+						{
+							"id": 2,
+							"content": "<eot>",
+							"special": true
+						},
+						{
+							"id": 3,
+							"content": "<eom>",
+							"special": true
+						}
+					],
+					"model": {
+						"vocab": {
+							"<bos>": 0,
+							"<eos>": 1,
+							"<eot>": 2,
+							"<eom>": 3
+						}
+					}
+				}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"add_bos_token": true,
+					"add_eos_token": false,
+					"bos_token": "<bos>",
+					"eos_token": "<eos>"
+				}`),
+				"generation_config.json": strings.NewReader(`{
+					"bos_token_id": 0,
+					"eos_token_id": [1, 2, 3]
+				}`),
+			}),
+			specialTokenTypes: []string{"pad", "eos", "bos", "unk"},
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<bos>", "<eos>", "<eot>", "<eom>"},
+					Scores: []float32{0, 1, 2, 3},
+					Types:  []int32{3, 3, 3, 3},
+				},
+				SpecialVocabulary: []*SpecialVocabulary{
+					{Type: "eos", Content: "<eos>", ID: 1, IDs: []int32{1, 2, 3}, AddToken: false},
+					{Type: "bos", Content: "<bos>", ID: 0, AddToken: true},
+				},
+				Pre: "default",
+			},
+		},
 	}
 
 	for _, tt := range cases {

docs/api.md

@@ -43,6 +43,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 - `prompt`: the prompt to generate a response for
 - `suffix`: the text after the model response
 - `images`: (optional) a list of base64-encoded images (for multimodal models such as `llava`)
+- `think`: (for thinking models) should the model think before responding?
 
 Advanced parameters (optional):
 
@@ -490,11 +491,13 @@ Generate the next message in a chat with a provided model. This is a streaming e
 - `model`: (required) the [model name](#model-names)
 - `messages`: the messages of the chat, this can be used to keep a chat memory
 - `tools`: list of tools in JSON for the model to use if supported
+- `think`: (for thinking models) should the model think before responding?
 
 The `message` object has the following fields:
 
 - `role`: the role of the message, either `system`, `user`, `assistant`, or `tool`
 - `content`: the content of the message
+- `thinking`: (for thinking models) the model's thinking process
 - `images` (optional): a list of images to include in the message (for multimodal models such as `llava`)
 - `tool_calls` (optional): a list of tools in JSON that the model wants to use
 
@@ -1154,11 +1157,15 @@ A single JSON object will be returned.
 {
   "models": [
     {
-      "name": "deepseek-r1:latest",
-      "model": "deepseek-r1:latest",
-      "modified_at": "2025-05-10T08:06:48.639712648-07:00",
-      "size": 4683075271,
-      "digest": "0a8c266910232fd3291e71e5ba1e058cc5af9d411192cf88b6d30e92b6e73163",
+
+      "model": "codellama:13b",
+      "modified_at": "2023-11-04T14:56:49.277302595-07:00",
+      "size": 7365960935,
+      "digest": "9f438cb9cd581fc025612d27f7c1a6669ff83a8bb0ed86c94fcf4c5440555697",
+      "capabilities": [
+        "completion"
+      ],
+
       "details": {
         "parent_model": "",
         "format": "gguf",
@@ -1171,11 +1178,16 @@ A single JSON object will be returned.
       }
     },
     {
-      "name": "llama3.2:latest",
-      "model": "llama3.2:latest",
-      "modified_at": "2025-05-04T17:37:44.706015396-07:00",
-      "size": 2019393189,
-      "digest": "a80c4f17acd55265feec403c7aef86be0c25983ab279d83f3bcd3abbcb5b8b72",
+
+      "model": "llama4:latest",
+      "modified_at": "2023-12-07T09:32:18.757212583-08:00",
+      "size": 3825819519,
+      "digest": "fe938a131f40e6f6d40083c9f0f430a515233eb2edaa6d72eb85c50d64f2300e",
+      "capabilities": [
+        "completion",
+        "vision"
+      ],
+
       "details": {
         "parent_model": "",
         "format": "gguf",

docs/development.md

@@ -118,7 +118,7 @@ To run tests, use `go test`:
 go test ./...
 ```
 
-> NOTE: In rare cirumstances, you may nedd to change a package using the new
+> NOTE: In rare cirumstances, you may need to change a package using the new
 > "synctest" package in go1.24.
 >
 > If you do not have the "synctest" package enabled, you will not see build or

docs/import.md

@@ -132,22 +132,12 @@ success
 
 ### Supported Quantizations
 
-- `q4_0`
-- `q4_1`
-- `q5_0`
-- `q5_1`
 - `q8_0`
 
 #### K-means Quantizations
 
-- `q3_K_S`
-- `q3_K_M`
-- `q3_K_L`
 - `q4_K_S`
 - `q4_K_M`
-- `q5_K_S`
-- `q5_K_M`
-- `q6_K`
 
 
 ## Sharing your model on ollama.com

envconfig/config.go

@@ -183,6 +183,8 @@ var (
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
 	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
+	// Auth enables authentication between the Ollama client and server
+	UseAuth = Bool("OLLAMA_AUTH")
 )
 
 func String(s string) func() string {

fs/ggml/ggml.go

@@ -6,7 +6,6 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"math"
 	"slices"
 	"strings"
 
@@ -16,6 +15,7 @@ import (
 type GGML struct {
 	container
 	model
+	Length int64
 }
 
 type model interface {
@@ -387,12 +387,12 @@ func DetectContentType(b []byte) string {
 //
 // It collects array values for arrays with a size less than or equal to
 // maxArraySize. If the maxArraySize is negative, all arrays are collected.
-func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
+func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
 
 	var magic uint32
 	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
-		return nil, 0, err
+		return nil, err
 	}
 
 	var c container
@@ -402,24 +402,25 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	case FILE_MAGIC_GGUF_BE:
 		c = &containerGGUF{ByteOrder: binary.BigEndian, maxArraySize: maxArraySize}
 	default:
-		return nil, 0, errors.New("invalid file magic")
+		return nil, errors.New("invalid file magic")
 	}
 
 	model, err := c.Decode(rs)
 	if err != nil {
-		return nil, 0, err
+		return nil, err
 	}
 
 	offset, err := rs.Seek(0, io.SeekCurrent)
 	if err != nil {
-		return nil, 0, err
+		return nil, err
 	}
 
 	// final model type
 	return &GGML{
 		container: c,
 		model:     model,
-	}, offset, nil
+		Length:    offset,
+	}, nil
 }
 
 func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
@@ -653,24 +654,15 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
 			numPatches*numPatches*headCount)
 	case "qwen25vl":
 		maxPixels := uint64(llm.KV().Uint("vision.max_pixels", 28*28*1280))
-		mergeSize := uint64(llm.KV().Uint("vision.spatial_merge_size", 2))
-		temporalPatchSize := uint64(2)
 
-		// Calculate max possible patches based on max_pixels
-		maxHeight := uint64(math.Sqrt(float64(maxPixels)))
-		maxWidth := maxPixels / maxHeight
-		maxGridHeight := maxHeight / patchSize
-		maxGridWidth := maxWidth / patchSize
-		// Account for merged patches (2x2 grid)
-		numPatches := (maxGridHeight * maxGridWidth) / (mergeSize * mergeSize)
+		numPatches := maxPixels / (patchSize * patchSize)
 
-		// Calculate graph size based on typical operations in ProcessImage and createPatches
 		graphSize = 4 * (maxPixels*numChannels + // Original image storage
 			// Normalized pixels
 			maxPixels*numChannels +
-			// Patches storage (numPatches * channels * temporalPatchSize * patchSize^2)
-			numPatches*numChannels*temporalPatchSize*patchSize*patchSize +
-			// Self-attention calculations (similar to other architectures)
+			// Patches storage (numPatches * channels * patchSize^2)
+			numPatches*numChannels*patchSize*patchSize +
+			// Self-attention calculations
 			numPatches*numPatches*headCount +
 			// Additional buffer for processing
 			embeddingLength*numPatches)

fs/ggml/gguf_test.go

@@ -35,7 +35,7 @@ func TestWriteGGUF(t *testing.T) {
 	}
 	defer r.Close()
 
-	ff, _, err := Decode(r, 0)
+	ff, err := Decode(r, 0)
 	if err != nil {
 		t.Fatal(err)
 	}

integration/llm_image_test.go

@@ -19,7 +19,7 @@ func TestVisionModels(t *testing.T) {
 	}
 	testCases := []testCase{
 		{
-			model: "llava:7b",
+			model: "qwen2.5vl",
 		},
 		{
 			model: "llama3.2-vision",
@@ -60,6 +60,7 @@ func TestVisionModels(t *testing.T) {
 }
 
 func TestIntegrationSplitBatch(t *testing.T) {
+	skipUnderMinVRAM(t, 6)
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
 	require.NoError(t, err)
 	req := api.GenerateRequest{

kvcache/causal.go

@@ -30,6 +30,11 @@ type Causal struct {
 
 	// ** current forward pass **
 
+	// curReserve indicates that this forward pass is only for
+	// memory reservation and we should not update our metadata
+	// based on it.
+	curReserve bool
+
 	// the active layer for Get and Put
 	curLayer int
 
@@ -159,12 +164,13 @@ func (c *Causal) Close() {
 }
 
 func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
+	c.curReserve = reserve
 	c.curBatchSize = len(batch.Positions)
 	c.curSequences = batch.Sequences
 	c.curPositions = batch.Positions
 	c.opts.Except = nil
 
-	if !reserve {
+	if !c.curReserve {
 		c.updateSlidingWindow()
 
 		var err error
@@ -211,10 +217,9 @@ func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) e
 		c.curCellRange.max = len(c.cells) - 1
 	}
 
-	var err error
-	c.curMask, err = c.buildMask(ctx)
+	c.curMask = c.buildMask(ctx)
 
-	return err
+	return nil
 }
 
 func newRange() cellRange {
@@ -297,7 +302,7 @@ func roundUp(length, pad int) int {
 // Builds a mask of history x batch indicating whether for each token in the batch the
 // token in the history should apply. This is based on both the sequence and causality (the
 // position of the history is not ahead of the token in the batch).
-func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
+func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 	// Align and pad the two dimensions as required by the backend
 	batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)
 
@@ -305,6 +310,11 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 	c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1
 
 	length := c.curCellRange.max - c.curCellRange.min + 1
+
+	if c.curReserve {
+		return ctx.Input().Empty(c.config.MaskDType, length, batchSize)
+	}
+
 	mask := make([]float32, batchSize*length)
 
 	for i := range c.curBatchSize {
@@ -325,10 +335,7 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 		mask[i] = float32(math.Inf(-1))
 	}
 
-	maskTensor, err := ctx.Input().FromFloatSlice(mask, length, batchSize)
-	if err != nil {
-		return nil, err
-	}
+	maskTensor := ctx.Input().FromFloatSlice(mask, length, batchSize)
 
 	if c.config.MaskDType != ml.DTypeF32 {
 		out := ctx.Input().Empty(c.config.MaskDType, maskTensor.Shape()...)
@@ -336,7 +343,7 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 		maskTensor = out
 	}
 
-	return maskTensor, nil
+	return maskTensor
 }
 
 func (c *Causal) moveCells(ctx ml.Context, src, dst, length int) {
@@ -491,12 +498,7 @@ func (c *Causal) SetCausal(ctx ml.Context, opts CausalOptions) {
 	if !slices.Equal(c.opts.Except, opts.Except) {
 		c.opts = opts
 		if ctx != nil {
-			var err error
-			c.curMask, err = c.buildMask(ctx)
-			if err != nil {
-				// This error should never occur because we have previously built a mask with the same shape
-				panic(fmt.Errorf("SetCausal: %w", err))
-			}
+			c.curMask = c.buildMask(ctx)
 		}
 	}
 }
@@ -652,10 +654,7 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 		}
 	}
 
-	kShift, err := ctx.Input().FromIntSlice(offsets, len(offsets))
-	if err != nil {
-		return err
-	}
+	kShift := ctx.Input().FromIntSlice(offsets, len(offsets))
 
 	for i, key := range c.keys {
 		if key == nil {

kvcache/causal_test.go

@@ -344,7 +344,7 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 			}
 
 			cache.SetLayer(0)
-			tensor, _ := context.FromFloatSlice(test.in, test.inShape...)
+			tensor := context.FromFloatSlice(test.in, test.inShape...)
 			cache.Put(context, tensor, tensor)
 
 			out, _, mask := cache.Get(context)
@@ -386,7 +386,7 @@ func TestCanResume(t *testing.T) {
 	}
 
 	cache.SetLayer(0)
-	tensor, _ := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
+	tensor := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
 	cache.Put(context, tensor, tensor)
 
 	// with window size 4, nothing has slid out of the window yet
@@ -413,7 +413,7 @@ func TestCanResume(t *testing.T) {
 	}
 
 	cache.SetLayer(0)
-	tensor, _ = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
+	tensor = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
 	cache.Put(context, tensor, tensor)
 
 	// only the latest position has overlapping windows
@@ -470,24 +470,24 @@ func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
 	return c.Empty(dtype, shape...)
 }
 
-func (c *testContext) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
+func (c *testContext) FromFloatSlice(s []float32, shape ...int) ml.Tensor {
 	t := c.Empty(ml.DTypeF32, shape...).(*testTensor)
 
 	copy(t.data, s)
 
-	return t, nil
+	return t
 }
 
-func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
+func (c *testContext) FromIntSlice(s []int32, shape ...int) ml.Tensor {
 	f := make([]float32, len(s))
 	for i := range f {
 		f[i] = float32(s[i])
 	}
 
-	out, _ := c.FromFloatSlice(f, shape...)
+	out := c.FromFloatSlice(f, shape...)
 	out.(*testTensor).dtype = ml.DTypeI32
 
-	return out, nil
+	return out
 }
 
 func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
@@ -496,7 +496,7 @@ func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tenso
 		s = append(s, i)
 	}
 
-	out, _ := c.FromFloatSlice(s, len(s))
+	out := c.FromFloatSlice(s, len(s))
 	out.(*testTensor).dtype = dtype
 	return out
 }
@@ -508,7 +508,7 @@ func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }
 
 func (c *testContext) Compute(...ml.Tensor) {}
 
-func (c *testContext) Reserve() error { return nil }
+func (c *testContext) Reserve() {}
 
 func (c *testContext) MaxGraphNodes() int {
 	return 10

llama/llama.go

@@ -544,7 +544,7 @@ func NewSamplingContext(model *Model, params SamplingParams) (*SamplingContext,
 	cparams.penalty_last_n = C.int32_t(params.RepeatLastN)
 	cparams.penalty_repeat = C.float(params.PenaltyRepeat)
 	cparams.penalty_freq = C.float(params.PenaltyFreq)
-	cparams.penalty_present = C.float(params.PenaltyFreq)
+	cparams.penalty_present = C.float(params.PenaltyPresent)
 	cparams.seed = C.uint32_t(params.Seed)
 
 	grammar := C.CString(params.Grammar)
@@ -580,7 +580,7 @@ func SchemaToGrammar(schema []byte) []byte {
 	defer C.free(unsafe.Pointer(cStr))
 
 	// Allocate buffer for grammar based on schema length but with upper bound
-	maxLen := min(1024*1024, len(schema)*4)
+	maxLen := max(32768, min(1024*1024, len(schema)*4))
 	buf := make([]byte, maxLen)
 
 	// Call C function to convert schema to grammar
@@ -602,7 +602,7 @@ type Grammar struct {
 	mu sync.Mutex
 }
 
-func NewGrammar(grammar string, vocabIds []uint32, vocabValues []string, eogTokens []uint32) *Grammar {
+func NewGrammar(grammar string, vocabIds []uint32, vocabValues []string, eogTokens []int32) *Grammar {
 	cGrammar := C.CString(grammar)
 	defer C.free(unsafe.Pointer(cGrammar))
 
@@ -622,7 +622,7 @@ func NewGrammar(grammar string, vocabIds []uint32, vocabValues []string, eogToke
 		cEogTokens[i] = C.uint32_t(token)
 	}
 
-	g := C.grammar_init(cGrammar, (*C.uint32_t)(unsafe.Pointer(&cTokens[0])), C.size_t(len(cTokens)), (**C.char)(unsafe.Pointer(&cPieces[0])), (*C.uint32_t)(unsafe.Pointer(&cEogTokens[0])), C.size_t(len(cEogTokens)))
+	g := C.grammar_init(cGrammar, unsafe.SliceData(cTokens), C.size_t(len(cTokens)), unsafe.SliceData(cPieces), unsafe.SliceData(cEogTokens), C.size_t(len(cEogTokens)))
 	if g == nil {
 		return nil
 	}

llama/patches/0016-graph-memory-reporting-on-failure.patch

@@ -0,0 +1,156 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Jesse Gross <jesse@ollama.com>
+Date: Fri, 18 Apr 2025 15:58:19 -0700
+Subject: [PATCH] graph memory reporting on failure
+
+---
+ ggml/include/ggml-alloc.h   |  6 ++++++
+ ggml/include/ggml-backend.h |  6 ++++++
+ ggml/src/ggml-alloc.c       | 38 +++++++++++++++++++++++++++++++++----
+ ggml/src/ggml-backend.cpp   | 10 ++++++++++
+ 4 files changed, 56 insertions(+), 4 deletions(-)
+
+diff --git a/ggml/include/ggml-alloc.h b/ggml/include/ggml-alloc.h
+index 2cb150fd..781b1e10 100644
+--- a/ggml/include/ggml-alloc.h
++++ b/ggml/include/ggml-alloc.h
+@@ -66,6 +66,12 @@ GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph
+ 
+ GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
+ 
++struct ggml_allocr_buffer_status {
++    size_t size;
++    bool allocated;
++};
++GGML_API struct ggml_allocr_buffer_status ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id);
++
+ // Utils
+ // Create a buffer and allocate all the tensors in a ggml_context
+ GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
+index 778927f6..74e46716 100644
+--- a/ggml/include/ggml-backend.h
++++ b/ggml/include/ggml-backend.h
+@@ -304,6 +304,12 @@ extern "C" {
+ 
+     GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+ 
++    struct ggml_backend_buffer_status {
++        size_t size;
++        bool allocated;
++    };
++    GGML_API struct ggml_backend_buffer_status ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
++
+     GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+     GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+ 
+diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
+index 5fd379f6..04812990 100644
+--- a/ggml/src/ggml-alloc.c
++++ b/ggml/src/ggml-alloc.c
+@@ -364,6 +364,7 @@ struct node_alloc {
+ struct ggml_gallocr {
+     ggml_backend_buffer_type_t * bufts; // [n_buffers]
+     ggml_backend_buffer_t * buffers; // [n_buffers]
++    size_t *buffer_sizes; // [n_buffers]
+     struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
+     int n_buffers;
+ 
+@@ -387,6 +388,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
+     galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
+     GGML_ASSERT(galloc->buffers != NULL);
+ 
++    galloc->buffer_sizes = calloc(n_bufs, sizeof(size_t));
++    GGML_ASSERT(galloc->buffer_sizes != NULL);
++
+     galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
+     GGML_ASSERT(galloc->buf_tallocs != NULL);
+ 
+@@ -453,6 +457,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
+     ggml_hash_set_free(&galloc->hash_set);
+     free(galloc->hash_values);
+     free(galloc->bufts);
++    free(galloc->buffer_sizes);
+     free(galloc->buffers);
+     free(galloc->buf_tallocs);
+     free(galloc->node_allocs);
+@@ -748,6 +753,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+         }
+     }
+ 
++    bool success = true;
++
+     // reallocate buffers if needed
+     for (int i = 0; i < galloc->n_buffers; i++) {
+         // if the buffer type is used multiple times, we reuse the same buffer
+@@ -769,15 +776,20 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+ 
+             ggml_backend_buffer_free(galloc->buffers[i]);
+             galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
+-            if (galloc->buffers[i] == NULL) {
++            if (galloc->buffers[i]) {
++                galloc->buffer_sizes[i] = ggml_backend_buffer_get_size(galloc->buffers[i]);
++                ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
++            } else {
+                 GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+-                return false;
++                galloc->buffer_sizes[i] = new_size;
++                success = false;
+             }
+-            ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
++        } else {
++            galloc->buffer_sizes[i] = ggml_backend_buffer_get_size(galloc->buffers[i]);
+         }
+     }
+ 
+-    return true;
++    return success;
+ }
+ 
+ bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
+@@ -934,6 +946,24 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+     return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
+ }
+ 
++struct ggml_allocr_buffer_status ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
++    GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
++
++    for (int i = 0; i < buffer_id; i++) {
++        if (galloc->buf_tallocs[i] == galloc->buf_tallocs[buffer_id]) {
++            // This buffer is the same as a previous one due to the same buffer type being used multiple times
++            // (See above.) However, we need a different check because multiple buffers might be NULL in our
++            // case and we still want to know the attempted size.
++
++            struct ggml_allocr_buffer_status status = {0, true};
++            return status;
++        }
++    }
++
++    struct ggml_allocr_buffer_status status = {galloc->buffer_sizes[buffer_id], galloc->buffers[buffer_id] != NULL};
++    return status;
++}
++
+ // utils
+ 
+ static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) {
+diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
+index 0ce73a99..be335e8c 100644
+--- a/ggml/src/ggml-backend.cpp
++++ b/ggml/src/ggml-backend.cpp
+@@ -1629,6 +1629,16 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
+     return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
+ }
+ 
++struct ggml_backend_buffer_status ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
++    int backend_index = ggml_backend_sched_backend_id(sched, backend);
++    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
++
++    struct ggml_allocr_buffer_status allocr_status = ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
++    struct ggml_backend_buffer_status status = {allocr_status.size, allocr_status.allocated};
++
++    return status;
++}
++
+ void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+     int backend_index = ggml_backend_sched_backend_id(sched, backend);
+     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);

llama/patches/0017-ggml-Export-GPU-UUIDs.patch

@@ -0,0 +1,102 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Jesse Gross <jesse@ollama.com>
+Date: Thu, 24 Apr 2025 14:48:51 -0700
+Subject: [PATCH] ggml: Export GPU UUIDs
+
+This enables matching up devices and information reported by the backend
+with tools (e.g. nvidia-smi) and system management libraries (e.g. nvml).
+---
+ ggml/include/ggml-backend.h      |  1 +
+ ggml/src/ggml-cuda/ggml-cuda.cu  | 33 ++++++++++++++++++++++++++++++++
+ ggml/src/ggml-metal/ggml-metal.m |  1 +
+ 3 files changed, 35 insertions(+)
+
+diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
+index 74e46716..a880df33 100644
+--- a/ggml/include/ggml-backend.h
++++ b/ggml/include/ggml-backend.h
+@@ -152,6 +152,7 @@ extern "C" {
+     struct ggml_backend_dev_props {
+         const char * name;
+         const char * description;
++        const char * uuid;
+         size_t memory_free;
+         size_t memory_total;
+         enum ggml_backend_dev_type type;
+diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
+index cb0d8528..4c829153 100644
+--- a/ggml/src/ggml-cuda/ggml-cuda.cu
++++ b/ggml/src/ggml-cuda/ggml-cuda.cu
+@@ -2884,6 +2884,7 @@ struct ggml_backend_cuda_device_context {
+     int device;
+     std::string name;
+     std::string description;
++    std::string uuid;
+ };
+ 
+ static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
+@@ -2896,6 +2897,11 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
+     return ctx->description.c_str();
+ }
+ 
++static const char * ggml_backend_cuda_device_get_uuid(ggml_backend_dev_t dev) {
++    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
++    return ctx->uuid.c_str();
++}
++
+ static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+     ggml_cuda_set_device(ctx->device);
+@@ -2910,6 +2916,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
+ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+     props->name        = ggml_backend_cuda_device_get_name(dev);
+     props->description = ggml_backend_cuda_device_get_description(dev);
++    props->uuid        = ggml_backend_cuda_device_get_uuid(dev);
+     props->type        = ggml_backend_cuda_device_get_type(dev);
+     ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
+ 
+@@ -3458,6 +3465,32 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
+                 dev_ctx->description = prop.name;
+ 
++                #if !defined(GGML_USE_HIP)
++                char uuid[64];
++                snprintf(uuid, sizeof(uuid),
++                    "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
++                    (unsigned char)prop.uuid.bytes[0],
++                    (unsigned char)prop.uuid.bytes[1],
++                    (unsigned char)prop.uuid.bytes[2],
++                    (unsigned char)prop.uuid.bytes[3],
++                    (unsigned char)prop.uuid.bytes[4],
++                    (unsigned char)prop.uuid.bytes[5],
++                    (unsigned char)prop.uuid.bytes[6],
++                    (unsigned char)prop.uuid.bytes[7],
++                    (unsigned char)prop.uuid.bytes[8],
++                    (unsigned char)prop.uuid.bytes[9],
++                    (unsigned char)prop.uuid.bytes[10],
++                    (unsigned char)prop.uuid.bytes[11],
++                    (unsigned char)prop.uuid.bytes[12],
++                    (unsigned char)prop.uuid.bytes[13],
++                    (unsigned char)prop.uuid.bytes[14],
++                    (unsigned char)prop.uuid.bytes[15]
++                  );
++                dev_ctx->uuid = uuid;
++                #else
++                dev_ctx->uuid = "GPU-" + std::string(prop.uuid.bytes, 16);
++                #endif
++
+                 ggml_backend_dev_t dev = new ggml_backend_device {
+                     /* .iface   = */ ggml_backend_cuda_device_interface,
+                     /* .reg     = */ &reg,
+diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
+index 1b56f858..ee4f2dcb 100644
+--- a/ggml/src/ggml-metal/ggml-metal.m
++++ b/ggml/src/ggml-metal/ggml-metal.m
+@@ -5703,6 +5703,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
+ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+     props->name        = ggml_backend_metal_device_get_name(dev);
+     props->description = ggml_backend_metal_device_get_description(dev);
++    props->uuid        = "0";
+     props->type        = ggml_backend_metal_device_get_type(dev);
+     ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
+     props->caps = (struct ggml_backend_dev_caps) {

llm/memory.go

@@ -1,12 +1,9 @@
 package llm
 
 import (
-	"cmp"
 	"fmt"
 	"log/slog"
-	"maps"
 	"os"
-	"slices"
 	"strconv"
 	"strings"
 
@@ -85,8 +82,11 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	var graphOffload uint64
 
 	// Projectors loaded into GPU0 only
-	var projectorWeights uint64
-	var projectorGraph uint64
+	var llamaEngineProjectorWeights uint64
+
+	// Projectors loaded with output layer
+	var ollamaEngineProjectorWeights uint64
+	var ollamaEngineProjectorGraph uint64
 
 	// Conditional output size on GPU 0
 	var memoryLayerOutput uint64
@@ -111,21 +111,23 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", availableList)
 
 	for _, projector := range projectors {
-		weight := projectorMemoryRequirements(projector)
-		projectorWeights += weight
+		llamaEngineProjectorWeights += projectorMemoryRequirements(projector)
 
 		// multimodal models require at least 2048 context
 		opts.NumCtx = max(opts.NumCtx, 2048)
 	}
-	if projectorWeights == 0 && projectorGraph == 0 {
-		projectorWeights, projectorGraph = f.VisionGraphSize()
+	if llamaEngineProjectorWeights == 0 {
+		ollamaEngineProjectorWeights, ollamaEngineProjectorGraph = f.VisionGraphSize()
+		opts.NumCtx = max(opts.NumCtx, 2048)
 	}
 
 	layers := f.Tensors().GroupLayers()
-	// add one layer (chosing the max layer) worth of memory as a buffer
-	layerSize = slices.MaxFunc(slices.Collect(maps.Values(layers)), func(a, b ggml.Layer) int {
-		return cmp.Compare(a.Size(), b.Size())
-	}).Size()
+	// add one layer worth of memory as a buffer
+	if blk0, ok := layers["blk.0"]; ok {
+		layerSize = blk0.Size()
+	} else {
+		slog.Warn("model missing blk.0 layer size")
+	}
 
 	var kvct string
 	if envconfig.FlashAttention() &&
@@ -163,6 +165,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		graphFullOffload = graphPartialOffload
 	}
 
+	// Output layer handled at the end if we have space
 	if layer, ok := layers["output_norm"]; ok {
 		memoryLayerOutput += layer.Size()
 	}
@@ -172,8 +175,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		memoryLayerOutput += layer.Size()
 	}
 
-	// Output layer handled at the end if we have space
-	gpuZeroOverhead := projectorWeights + projectorGraph
+	gpuZeroOverhead := llamaEngineProjectorWeights
 
 	// Reduce set of GPUs to only those that have sufficient space to fit overhead and at least one layer
 	var layerCount int
@@ -216,6 +218,8 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	if len(gpusWithSpace) > 0 {
 		gpuZeroID = gpusWithSpace[0].i
 		gpuAllocations[gpuZeroID] += gpuZeroOverhead
+	} else {
+		overflow += gpuZeroOverhead
 	}
 
 	// For all the layers, find where they can fit on the GPU(s)
@@ -256,21 +260,24 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	}
 
 	// Determine if we need to consider output then find where it fits
-	if memoryLayerOutput > 0 && (opts.NumGPU < 0 || layerCount < opts.NumGPU) {
-		for j := len(gpusWithSpace); j > 0; j-- {
-			g := gpusWithSpace[layerCount%j]
-			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
-			if g.g.FreeMemory > overhead+used+memoryLayerOutput {
-				gpuAllocations[g.i] += memoryLayerOutput
-				layerCounts[g.i]++
-				layerCount++
-				break
+	memoryLastLayer := memoryLayerOutput + ollamaEngineProjectorWeights + ollamaEngineProjectorGraph
+	if memoryLastLayer > 0 {
+		if opts.NumGPU < 0 || layerCount < opts.NumGPU {
+			for j := len(gpusWithSpace); j > 0; j-- {
+				g := gpusWithSpace[layerCount%j]
+				used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
+				if g.g.FreeMemory > overhead+used+memoryLastLayer {
+					gpuAllocations[g.i] += memoryLastLayer
+					layerCounts[g.i]++
+					layerCount++
+					break
+				}
 			}
 		}
 
 		if layerCount < int(f.KV().BlockCount())+1 {
 			fullyLoaded = false
-			overflow += memoryLayerOutput
+			overflow += memoryLastLayer
 		}
 	}
 
@@ -328,8 +335,8 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		memoryLayerOutput:   memoryLayerOutput,
 		graphFullOffload:    graphFullOffload,
 		graphPartialOffload: graphPartialOffload,
-		projectorWeights:    projectorWeights,
-		projectorGraph:      projectorGraph,
+		projectorWeights:    llamaEngineProjectorWeights + ollamaEngineProjectorWeights,
+		projectorGraph:      ollamaEngineProjectorGraph,
 	}
 
 	if gpus[0].Library == "cpu" {
@@ -415,7 +422,7 @@ func projectorMemoryRequirements(filename string) (weights uint64) {
 	}
 	defer file.Close()
 
-	ggml, _, err := ggml.Decode(file, 1024)
+	ggml, err := ggml.Decode(file, 1024)
 	if err != nil {
 		return 0
 	}

llm/server.go

@@ -121,7 +121,7 @@ func LoadModel(model string, maxArraySize int) (*ggml.GGML, error) {
 	}
 	defer f.Close()
 
-	ggml, _, err := ggml.Decode(f, maxArraySize)
+	ggml, err := ggml.Decode(f, maxArraySize)
 	return ggml, err
 }
 
@@ -797,7 +797,8 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 
 	res, err := http.DefaultClient.Do(serverReq)
 	if err != nil {
-		return fmt.Errorf("POST predict: %v", err)
+		slog.Error("post predict", "error", err)
+		return errors.New("model runner has unexpectedly stopped, this may be due to resource limitations or an internal error, check ollama server logs for details")
 	}
 	defer res.Body.Close()
 

ml/backend.go

@@ -5,8 +5,8 @@ import (
 	"context"
 	"encoding/binary"
 	"fmt"
+	"log/slog"
 	"math"
-	"os"
 	"slices"
 	"strconv"
 	"strings"
@@ -15,6 +15,11 @@ import (
 )
 
 type Backend interface {
+	Load(ctx context.Context, progress func(float32)) error
+
+	// BackendMemory returns the memory allocations that were made for this model
+	BackendMemory() BackendMemory
+
 	Config() fs.Config
 	Get(name string) Tensor
 	NewContext() Context
@@ -52,10 +57,6 @@ type CacheConfig struct {
 
 // BackendParams controls how the backend loads and executes models
 type BackendParams struct {
-	// Progress is a callback function that allows reporting percentage completion
-	// of model loading
-	Progress func(float32)
-
 	// NumThreads sets the number of threads to use if running on the CPU
 	NumThreads int
 
@@ -72,9 +73,130 @@ type BackendParams struct {
 	FlashAttention bool
 }
 
-var backends = make(map[string]func(context.Context, *os.File, BackendParams) (Backend, error))
+// ErrNoMem is returned when panicing due to insufficient memory. It includes
+// the attempted memory allocation.
+type ErrNoMem struct {
+	BackendMemory
+}
 
-func RegisterBackend(name string, f func(context.Context, *os.File, BackendParams) (Backend, error)) {
+func (e ErrNoMem) Error() string {
+	return fmt.Sprintf("insufficient memory - required allocations: %+v", e.BackendMemory)
+}
+
+type AllocationStatus int
+
+const (
+	// Unallocated memory - have not yet attempted to allocate
+	Unallocated AllocationStatus = iota
+
+	// Failed memory - tried to allocate the memory and did not succeed
+	Failed
+
+	// Allocated memory = tried and succeeded to allocate memory
+	Allocated
+)
+
+// Memory is the size of an allocation and whether it was successful.
+type Memory struct {
+	Size   uint64
+	Status AllocationStatus
+}
+
+func (m Memory) String() string {
+	s := fmt.Sprint(m.Size)
+
+	switch m.Status {
+	case Unallocated:
+		s += "U"
+	case Failed:
+		s += "F"
+	case Allocated:
+		s += "A"
+	}
+
+	return s
+}
+
+// DeviceMemory provides a breakdown of the memory needed
+// per device, such as a CPU or GPU.
+type DeviceMemory struct {
+	// Name is the name of the device as labeled by the backend. It
+	// may not be persistent across instances of the runner.
+	Name string
+
+	// UUID is a unique persistent identifier for the device for matching
+	// with system management libraries
+	UUID string
+
+	// Weights is the per-layer memory needed for the model weights.
+	Weights []Memory
+
+	// Cache is the per-layer memory needed for the KV cache.
+	Cache []Memory
+
+	// Graph is the size of the compute graph. It is not per-layer.
+	Graph Memory
+}
+
+func memoryPresent(mem []Memory) bool {
+	return slices.ContainsFunc(mem, func(m Memory) bool { return m.Size != 0 })
+}
+
+func (m DeviceMemory) LogValue() slog.Value {
+	var attrs []slog.Attr
+	if memoryPresent(m.Weights) {
+		attrs = append(attrs, slog.Any("Weights", m.Weights))
+	}
+
+	if memoryPresent(m.Cache) {
+		attrs = append(attrs, slog.Any("Cache", m.Cache))
+	}
+
+	if m.Graph.Size != 0 {
+		attrs = append(attrs, slog.Any("Graph", m.Graph))
+	}
+
+	if len(attrs) > 0 && m.UUID != "" {
+		attrs = append([]slog.Attr{slog.String("UUID", m.UUID)}, attrs...)
+	}
+
+	return slog.GroupValue(attrs...)
+}
+
+// BackendMemory provides the amount of memory required to load the model
+// per device based on the BackendParams. In some cases, not all required
+// allocations will be known at this point. However, the size of the most recent
+// allocation is guaranteed to be provided so that if it failed, the caller can
+// accommodate that to make forward progress.
+type BackendMemory struct {
+	// InputsWeights are always located on the CPU and cannot be moved
+	InputWeights Memory
+
+	// CPU model components are located in system memory. This does not
+	// include unified memory allocated through the GPU.
+	CPU DeviceMemory
+
+	// GPU model components are located on one or more GPUs.
+	GPUs []DeviceMemory
+}
+
+func (m BackendMemory) LogValue() slog.Value {
+	var attrs []slog.Attr
+	if m.InputWeights.Size != 0 {
+		attrs = append(attrs, slog.Any("InputWeights", m.InputWeights))
+	}
+
+	attrs = append(attrs, slog.Any(m.CPU.Name, m.CPU))
+	for _, g := range m.GPUs {
+		attrs = append(attrs, slog.Any(g.Name, g))
+	}
+
+	return slog.GroupValue(attrs...)
+}
+
+var backends = make(map[string]func(string, BackendParams) (Backend, error))
+
+func RegisterBackend(name string, f func(string, BackendParams) (Backend, error)) {
 	if _, ok := backends[name]; ok {
 		panic("backend: backend already registered")
 	}
@@ -82,9 +204,9 @@ func RegisterBackend(name string, f func(context.Context, *os.File, BackendParam
 	backends[name] = f
 }
 
-func NewBackend(ctx context.Context, f *os.File, params BackendParams) (Backend, error) {
+func NewBackend(modelPath string, params BackendParams) (Backend, error) {
 	if backend, ok := backends["ggml"]; ok {
-		return backend(ctx, f, params)
+		return backend(modelPath, params)
 	}
 
 	return nil, fmt.Errorf("unsupported backend")
@@ -93,8 +215,8 @@ func NewBackend(ctx context.Context, f *os.File, params BackendParams) (Backend,
 type Context interface {
 	Empty(dtype DType, shape ...int) Tensor
 	Zeros(dtype DType, shape ...int) Tensor
-	FromFloatSlice(s []float32, shape ...int) (Tensor, error)
-	FromIntSlice(s []int32, shape ...int) (Tensor, error)
+	FromFloatSlice(s []float32, shape ...int) Tensor
+	FromIntSlice(s []int32, shape ...int) Tensor
 
 	// Arange creates a 1D tensor with values within an interval (start, stop] increased by step.
 	Arange(start, stop, step float32, dtype DType) Tensor
@@ -106,7 +228,7 @@ type Context interface {
 	// graph, simply preallocates memory. Typically called with a
 	// worst case graph to ensure all resources are available for
 	// for future inference.
-	Reserve() error
+	Reserve()
 
 	MaxGraphNodes() int
 	Close()
@@ -119,21 +241,6 @@ type Context interface {
 	Layer(int) Context
 }
 
-// RopeOptions contains optional parameters for RoPE function
-type RopeOptions struct {
-	OriginalContextLen uint32
-}
-
-// RopeOption defines a function that modifies RopeOpts
-type RopeOption func(*RopeOptions)
-
-// WithContextLen sets a custom context length
-func WithContextLen(len uint32) RopeOption {
-	return func(opts *RopeOptions) {
-		opts.OriginalContextLen = len
-	}
-}
-
 type Tensor interface {
 	Dim(n int) int
 	Stride(n int) int
@@ -147,6 +254,8 @@ type Tensor interface {
 	Neg(ctx Context) Tensor
 	Add(ctx Context, t2 Tensor) Tensor
 	Mul(ctx Context, t2 Tensor) Tensor
+	Div(ctx Context, t2 Tensor) Tensor
+
 	Mulmat(ctx Context, t2 Tensor) Tensor
 	MulmatFullPrec(ctx Context, t2 Tensor) Tensor
 	MulmatID(ctx Context, t2, ids Tensor) Tensor
@@ -155,11 +264,11 @@ type Tensor interface {
 	LayerNorm(ctx Context, weight, bias Tensor, eps float32) Tensor
 	RMSNorm(ctx Context, weight Tensor, eps float32) Tensor
 	Scale(ctx Context, s float64) Tensor
+	SumRows(ctx Context) Tensor
 
 	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 
-	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32, options ...RopeOption) Tensor
 	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
 
 	Sin(ctx Context) Tensor

ml/backend/ggml/ggml.go

@@ -10,7 +10,6 @@ import "C"
 
 import (
 	"context"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -30,6 +29,7 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
 	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
+	"github.com/ollama/ollama/ml/nn/rope"
 	"golang.org/x/sync/errgroup"
 )
 
@@ -44,8 +44,15 @@ func devices() []*C.struct_ggml_backend_device {
 }
 
 type Backend struct {
+	// modelPath is the location of the model data
+	modelPath string
+
 	meta *fsggml.GGML
 
+	// tensorLoadTargets maps from the name of the tensor in the file
+	// to the name that is used by the model definition
+	tensorLoadTargets map[string][]string
+
 	sched         *C.struct_ggml_backend_sched
 	schedBackends []*C.struct_ggml_backend
 	schedBufts    []*C.struct_ggml_backend_buffer_type
@@ -58,14 +65,26 @@ type Backend struct {
 	// layers is the backend used for repeating layers
 	layers map[int]*C.struct_ggml_backend_buffer_type
 
+	// requiredMemory is the cumulative memory allocations needed by the backend
+	requiredMemory *ml.BackendMemory
+
+	// btDeviceMemory maps from a buffer type to the memory allocations associated with that device
+	btDeviceMemory map[*C.struct_ggml_backend_buffer_type]*ml.DeviceMemory
+
 	flashAttention bool
 
 	// maxGraphNodes is the maximum allowed number of graph nodes in this scheduler
 	maxGraphNodes int
 }
 
-func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend, error) {
-	meta, n, err := fsggml.Decode(r, -1)
+func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
+	r, err := os.Open(modelPath)
+	if err != nil {
+		return nil, err
+	}
+	defer r.Close()
+
+	meta, err := fsggml.Decode(r, -1)
 	if err != nil {
 		return nil, err
 	}
@@ -80,6 +99,9 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 		"num_key_values", len(meta.KV()),
 	)
 
+	var requiredMemory ml.BackendMemory
+	btDeviceMemory := make(map[*C.struct_ggml_backend_buffer_type]*ml.DeviceMemory)
+
 	type deviceBufferType struct {
 		d   *C.struct_ggml_backend_device
 		bts []*C.struct_ggml_backend_buffer_type
@@ -100,6 +122,8 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 		}
 	}
 
+	blocks := int(meta.KV().BlockCount())
+
 	// create list of buffer types for the cpu
 	cpuDeviceBufferType := deviceBufferType{d: C.ggml_backend_dev_by_type(C.GGML_BACKEND_DEVICE_TYPE_CPU)}
 	for _, d := range append(accels, append(gpus, cpus...)...) {
@@ -107,17 +131,33 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 		case C.GGML_BACKEND_DEVICE_TYPE_CPU,
 			C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
 			cpuDeviceBufferType.bts = append(cpuDeviceBufferType.bts, C.ggml_backend_dev_buffer_type(d))
+			btDeviceMemory[C.ggml_backend_dev_buffer_type(d)] = &requiredMemory.CPU
 		}
 	}
 
+	requiredMemory.CPU.Name = C.GoString(C.ggml_backend_dev_name(cpuDeviceBufferType.d))
+	var props C.struct_ggml_backend_dev_props
+	C.ggml_backend_dev_get_props(cpuDeviceBufferType.d, &props)
+	requiredMemory.CPU.UUID = C.GoString(props.uuid)
+	requiredMemory.CPU.Weights = make([]ml.Memory, blocks+1)
+	requiredMemory.CPU.Cache = make([]ml.Memory, blocks+1)
+
 	// create list of buffer types for each gpu
 	var gpuDeviceBufferTypes []deviceBufferType
-	for _, d := range gpus {
+	requiredMemory.GPUs = make([]ml.DeviceMemory, len(gpus))
+	for i, d := range gpus {
 		bt := C.ggml_backend_dev_buffer_type(d)
 		gpuDeviceBufferTypes = append(gpuDeviceBufferTypes, deviceBufferType{
 			d:   d,
 			bts: append([]*C.struct_ggml_backend_buffer_type{bt}, cpuDeviceBufferType.bts...),
 		})
+		btDeviceMemory[bt] = &requiredMemory.GPUs[i]
+		requiredMemory.GPUs[i].Name = C.GoString(C.ggml_backend_dev_name(d))
+		var props C.struct_ggml_backend_dev_props
+		C.ggml_backend_dev_get_props(d, &props)
+		requiredMemory.GPUs[i].UUID = C.GoString(props.uuid)
+		requiredMemory.GPUs[i].Weights = make([]ml.Memory, blocks+1)
+		requiredMemory.GPUs[i].Cache = make([]ml.Memory, blocks+1)
 	}
 
 	useDefaultSplit := true
@@ -156,8 +196,6 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 	// inputs always use cpu
 	input := cpuDeviceBufferType
 
-	blocks := int(meta.KV().BlockCount())
-
 	// define a range of gpu layers. anything outside of this range is assigned to the cpu
 	gpuRangeStart := max(0, blocks-params.NumGPULayers)
 	gpuRangeStop := min(gpuRangeStart+params.NumGPULayers, blocks+1)
@@ -198,7 +236,7 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 
 	// contexts are shared by tensors of the same buffer type
 	ctxs := make(map[*C.struct_ggml_backend_buffer_type]*C.struct_ggml_context)
-	createTensor := func(t tensor, bts []*C.struct_ggml_backend_buffer_type) *C.struct_ggml_tensor {
+	createTensor := func(t tensor, bts []*C.struct_ggml_backend_buffer_type, layer int) *C.struct_ggml_tensor {
 		for _, bt := range bts {
 			if _, ok := ctxs[bt]; !ok {
 				ctxs[bt] = C.ggml_init(C.struct_ggml_init_params{
@@ -224,6 +262,16 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 			C.ggml_set_name(tt, cname)
 
 			slog.Log(context.TODO(), logutil.LevelTrace, "created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
+
+			size := pad(C.ggml_backend_buft_get_alloc_size(bt, tt), C.ggml_backend_buft_get_alignment(bt))
+			if layer == -1 {
+				// Assume that InputWeights can be allocated - they're always in system memory and can't be moved in any case
+				requiredMemory.InputWeights.Status = ml.Allocated
+				requiredMemory.InputWeights.Size += uint64(size)
+			} else {
+				btDeviceMemory[bt].Weights[layer].Size += uint64(size)
+			}
+
 			//nolint:staticcheck // TODO: check if buffer type supports this tensor
 			return tt
 		}
@@ -245,22 +293,22 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 	for _, t := range meta.Tensors().Items() {
 		switch {
 		case contains(t.Name, "position_embd", "token_embd", "token_norm_embd", "token_types"):
-			createTensor(tensor{source: t}, input.bts)
+			createTensor(tensor{source: t}, input.bts, -1)
 			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
-				createTensor(tensor{source: t, target: "output.weight"}, output.bts)
+				createTensor(tensor{source: t, target: "output.weight"}, output.bts, blocks)
 			}
 		case contains(t.Name, "cls", "output", "output_norm"):
-			createTensor(tensor{source: t}, output.bts)
+			createTensor(tensor{source: t}, output.bts, blocks)
 		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
 			// TODO: assign vision tensors to the gpu if possible
-			createTensor(tensor{source: t}, output.bts)
+			createTensor(tensor{source: t}, output.bts, blocks)
 		case contains(t.Name, "rope_freqs", "rope_factors_long", "rope_factors_short"):
 			// these tensors should be repeated per layer
 			for i, layer := range layers {
 				createTensor(tensor{
 					source: t,
 					target: "blk." + strconv.Itoa(i) + "." + t.Name,
-				}, layer.bts)
+				}, layer.bts, i)
 			}
 		default:
 			layerIndex := -1
@@ -271,10 +319,10 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 			}
 
 			if layerIndex >= 0 {
-				createTensor(tensor{source: t}, layers[layerIndex].bts)
+				createTensor(tensor{source: t}, layers[layerIndex].bts, layerIndex)
 			} else {
 				// load all other tensors on the cpu
-				createTensor(tensor{source: t}, input.bts)
+				createTensor(tensor{source: t}, input.bts, -1)
 			}
 		}
 	}
@@ -287,8 +335,18 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 		}
 
 		b := C.ggml_backend_alloc_ctx_tensors_from_buft(c, bt)
+		for i := range btDeviceMemory[bt].Weights {
+			if btDeviceMemory[bt].Weights[i].Size != 0 {
+				if b != nil {
+					btDeviceMemory[bt].Weights[i].Status = ml.Allocated
+				} else {
+					btDeviceMemory[bt].Weights[i].Status = ml.Failed
+				}
+			}
+		}
+
 		if b == nil {
-			return nil, fmt.Errorf("unable to allocate memory from device %v for model weights", C.GoString(C.ggml_backend_buft_name(bt)))
+			panic(ml.ErrNoMem{BackendMemory: requiredMemory})
 		}
 
 		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
@@ -307,73 +365,6 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 		}
 	}
 
-	var doneBytes atomic.Uint64
-	totalBytes := uint64(n) - meta.Tensors().Offset
-
-	g, ctx := errgroup.WithContext(ctx)
-	g.SetLimit(runtime.GOMAXPROCS(0))
-	for _, t := range meta.Tensors().Items() {
-		t := t
-		g.Go(func() error {
-			tts := make([]*C.struct_ggml_tensor, max(1, len(targets[t.Name])))
-			for i := range tts {
-				target := targets[t.Name][i]
-				if target == "" {
-					target = t.Name
-				}
-
-				tt, ok := tensors[target]
-				if !ok {
-					return fmt.Errorf("unassigned tensor: %s", t.Name)
-				}
-
-				tts[i] = tt
-			}
-
-			// Create a new FD for each goroutine so that each FD is read sequentially, rather than
-			// seeking around within an FD shared between all goroutines.
-			file, err := os.Open(r.Name())
-			if err != nil {
-				slog.Warn("file open error", "file", r.Name(), "error", err)
-				return err
-			}
-			defer file.Close()
-			sr := io.NewSectionReader(file, int64(meta.Tensors().Offset+t.Offset), int64(t.Size()))
-			bts := make([]byte, 128*format.KibiByte)
-
-			var s uint64
-			for s < t.Size() {
-				// Stop if either the parent context has been canceled or if any of the other tensors returned an error
-				if err := ctx.Err(); err != nil {
-					return err
-				}
-
-				n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Size()-s))])
-				if err != nil {
-					slog.Warn("file read error", "file", r.Name(), "error", err)
-					return err
-				}
-
-				for _, tt := range tts {
-					C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), C.size_t(s), C.size_t(n))
-				}
-
-				s += uint64(n)
-
-				if params.Progress != nil {
-					done := doneBytes.Add(uint64(n))
-					params.Progress(float32(done) / float32(totalBytes))
-				}
-			}
-
-			return nil
-		})
-	}
-
-	if err := g.Wait(); err != nil {
-		return nil, err
-	}
-
 	// map devices to backend buffer types so new tensors can be assigned to the correct device
 	deviceBufferTypes := make(map[*C.struct_ggml_backend_device]*C.struct_ggml_backend_buffer_type)
 
@@ -397,9 +388,11 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 
 	maxGraphNodes := max(8192, len(meta.Tensors().Items())*5)
 	return &Backend{
-		flashAttention: params.FlashAttention,
-		meta:           meta,
-		tensors:        tensors,
+		modelPath:         modelPath,
+		flashAttention:    params.FlashAttention,
+		meta:              meta,
+		tensorLoadTargets: targets,
+		tensors:           tensors,
 		sched: C.ggml_backend_sched_new(
 			(*C.ggml_backend_t)(unsafe.Pointer(&schedBackends[0])),
 			(*C.ggml_backend_buffer_type_t)(unsafe.Pointer(&schedBufts[0])),
@@ -418,7 +411,9 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 			}
 			return m
 		}(),
-		maxGraphNodes: maxGraphNodes,
+		requiredMemory: &requiredMemory,
+		btDeviceMemory: btDeviceMemory,
+		maxGraphNodes:  maxGraphNodes,
 	}, nil
 }
 
@@ -426,6 +421,81 @@ func init() {
 	ml.RegisterBackend("ggml", New)
 }
 
+func (b *Backend) Load(ctx context.Context, progress func(float32)) error {
+	var doneBytes atomic.Uint64
+	totalBytes := uint64(b.meta.Length) - b.meta.Tensors().Offset
+
+	g, ctx := errgroup.WithContext(ctx)
+	g.SetLimit(runtime.GOMAXPROCS(0))
+	for _, t := range b.meta.Tensors().Items() {
+		t := t
+		g.Go(func() error {
+			tts := make([]*C.struct_ggml_tensor, max(1, len(b.tensorLoadTargets[t.Name])))
+			for i := range tts {
+				target := b.tensorLoadTargets[t.Name][i]
+				if target == "" {
+					target = t.Name
+				}
+
+				tt, ok := b.tensors[target]
+				if !ok {
+					return fmt.Errorf("unassigned tensor: %s", t.Name)
+				}
+
+				tts[i] = tt
+			}
+
+			// Create a new FD for each goroutine so that each FD is read sequentially, rather than
+			// seeking around within an FD shared between all goroutines.
+			file, err := os.Open(b.modelPath)
+			if err != nil {
+				slog.Warn("file open error", "file", b.modelPath, "error", err)
+				return err
+			}
+			defer file.Close()
+			sr := io.NewSectionReader(file, int64(b.meta.Tensors().Offset+t.Offset), int64(t.Size()))
+			bts := make([]byte, 128*format.KibiByte)
+
+			var s uint64
+			for s < t.Size() {
+				// Stop if either the parent context has been canceled or if any of the other tensors returned an error
+				if err := ctx.Err(); err != nil {
+					return err
+				}
+
+				n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Size()-s))])
+				if err != nil {
+					slog.Warn("file read error", "file", b.modelPath, "error", err)
+					return err
+				}
+
+				for _, tt := range tts {
+					C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), C.size_t(s), C.size_t(n))
+				}
+
+				s += uint64(n)
+
+				if progress != nil {
+					done := doneBytes.Add(uint64(n))
+					progress(float32(done) / float32(totalBytes))
+				}
+			}
+
+			return nil
+		})
+	}
+
+	if err := g.Wait(); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (b *Backend) BackendMemory() ml.BackendMemory {
+	return *b.requiredMemory
+}
+
 func (b *Backend) Config() fs.Config {
 	return b.meta.KV()
 }
@@ -457,6 +527,7 @@ func (b *Backend) NewContextSize(n int) ml.Context {
 			no_alloc: true,
 		}),
 		allocatedBuffers: &allocatedBuffers,
+		layer:            -1,
 	}
 }
 
@@ -483,6 +554,9 @@ type Context struct {
 
 	// maxGraphNodes is the maximum allowed number of graph nodes in this context
 	maxGraphNodes int
+
+	// layer is the graph layer that this context is allocating for - assumed to be cache
+	layer int
 }
 
 func (c *Context) Input() ml.Context {
@@ -493,6 +567,7 @@ func (c *Context) Input() ml.Context {
 			buft:             c.b.input,
 			allocatedBuffers: c.allocatedBuffers,
 			maxGraphNodes:    c.maxGraphNodes,
+			layer:            -1,
 		}
 	}
 
@@ -507,6 +582,7 @@ func (c *Context) Layer(i int) ml.Context {
 			buft:             buft,
 			allocatedBuffers: c.allocatedBuffers,
 			maxGraphNodes:    c.maxGraphNodes,
+			layer:            i,
 		}
 	}
 
@@ -544,22 +620,34 @@ func (c *Context) Compute(tensors ...ml.Tensor) {
 	}
 }
 
-func (c *Context) Reserve() error {
-	if !C.ggml_backend_sched_reserve(c.b.sched, c.graph) {
-		C.ggml_backend_sched_reset(c.b.sched)
-		return errors.New("failed to reserve graph")
-	}
+func (c *Context) Reserve() {
+	reserved := C.ggml_backend_sched_reserve(c.b.sched, c.graph)
 
 	slog.Debug("compute graph", "nodes", C.ggml_graph_n_nodes(c.graph), "splits", C.ggml_backend_sched_get_n_splits(c.b.sched))
-	for i := range c.b.schedBackends {
-		size := C.ggml_backend_sched_get_buffer_size(c.b.sched, c.b.schedBackends[i])
-		slog.Info("compute graph", "backend", C.GoString(C.ggml_backend_name(c.b.schedBackends[i])), "buffer_type", C.GoString(C.ggml_backend_buft_name(c.b.schedBufts[i])),
-			"size", format.HumanBytes2(uint64(size)))
+
+	// Reserve may get called multiple times for different graphs - we just want the last run, which will contain the max allocations
+	for _, bt := range c.b.schedBufts {
+		c.b.btDeviceMemory[bt].Graph = ml.Memory{}
 	}
 
-	C.ggml_backend_sched_reset(c.b.sched)
+	for i := range c.b.schedBackends {
+		bufferStatus := C.ggml_backend_sched_get_attempted_buffer_size(c.b.sched, c.b.schedBackends[i])
 
-	return nil
+		graph := &c.b.btDeviceMemory[c.b.schedBufts[i]].Graph
+		graph.Size += uint64(bufferStatus.size)
+		if bufferStatus.allocated && graph.Status != ml.Failed {
+			graph.Status = ml.Allocated
+		} else {
+			graph.Status = ml.Failed
+		}
+
+		slog.Info("compute graph", "backend", C.GoString(C.ggml_backend_name(c.b.schedBackends[i])), "buffer_type", C.GoString(C.ggml_backend_buft_name(c.b.schedBufts[i])),
+			"size", format.HumanBytes2(uint64(bufferStatus.size)))
+	}
+
+	if !reserved {
+		panic(ml.ErrNoMem{BackendMemory: *c.b.requiredMemory})
+	}
 }
 
 func (c *Context) MaxGraphNodes() int {
@@ -579,7 +667,7 @@ func pad(length, pad C.size_t) C.size_t {
 	return ((length + pad - 1) / pad) * pad
 }
 
-func (c *Context) newTensor(dtype ml.DType, shape []int) (ml.Tensor, error) {
+func (c *Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
 	if c.buft == nil {
 		panic("set Input or Layer before creating tensors")
 	}
@@ -602,7 +690,7 @@ func (c *Context) newTensor(dtype ml.DType, shape []int) (ml.Tensor, error) {
 
 	if len(shape) < 1 || shape[0] == 0 {
 		var shape C.int64_t = 0
-		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}, nil
+		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}
 	} else if len(shape) > 4 {
 		panic("unsupported number of dimensions")
 	}
@@ -615,40 +703,43 @@ func (c *Context) newTensor(dtype ml.DType, shape []int) (ml.Tensor, error) {
 
 	t := C.ggml_new_tensor(c.ctx, cdtype, C.int(len(shape)), shapeToGGML(shape))
 	size := pad(C.ggml_backend_buft_get_alloc_size(c.buft, t), C.ggml_backend_buft_get_alignment(c.buft))
-	b := C.ggml_backend_buft_alloc_buffer(c.buft, size)
-	if b == nil {
-		return nil, fmt.Errorf("unable to allocate %v from device %v for new tensor", format.HumanBytes2(uint64(size)), C.GoString(C.ggml_backend_buft_name(c.buft)))
-	}
-	*c.allocatedBuffers = append(*c.allocatedBuffers, b)
 
+	b := C.ggml_backend_buft_alloc_buffer(c.buft, size)
+	if c.layer >= 0 {
+		cache := &c.b.btDeviceMemory[c.buft].Cache[c.layer]
+
+		cache.Size += uint64(size)
+		if b != nil {
+			cache.Status = ml.Allocated
+		} else {
+			cache.Status = ml.Failed
+		}
+	}
+
+	if b == nil {
+		panic(ml.ErrNoMem{BackendMemory: *c.b.requiredMemory})
+	}
+
+	*c.allocatedBuffers = append(*c.allocatedBuffers, b)
 	C.ggml_backend_tensor_alloc(b, t, C.ggml_backend_buffer_get_base(b))
-	return &Tensor{b: c.b, t: t}, nil
+	return &Tensor{b: c.b, t: t}
 }
 
 func (c *Context) Empty(dtype ml.DType, shape ...int) ml.Tensor {
-	t, err := c.newTensor(dtype, shape)
-	if err != nil {
-		panic(err)
-	}
-
-	return t
+	return c.newTensor(dtype, shape)
 }
 
 func (c *Context) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
-	t, err := c.newTensor(dtype, shape)
-	if err != nil {
-		panic(err)
-	}
-
+	t := c.newTensor(dtype, shape)
 	C.ggml_set_zero(t.(*Tensor).t)
 	return t
 }
 
-func checkShape[S ~[]E, E any](s S, shape ...int) error {
+func checkShape[S ~[]E, E any](s S, shape ...int) {
 	n := len(s)
 
 	if n == 0 {
-		return nil
+		return
 	}
 
 	for _, v := range shape {
@@ -656,44 +747,32 @@ func checkShape[S ~[]E, E any](s S, shape ...int) error {
 	}
 
 	if n != 1 {
-		return fmt.Errorf("invalid shape: %v", shape)
+		panic(fmt.Errorf("invalid shape: %v", shape))
 	}
-
-	return nil
 }
 
-func (c *Context) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
-	if err := checkShape(s, shape...); err != nil {
-		return nil, err
-	}
+func (c *Context) FromFloatSlice(s []float32, shape ...int) ml.Tensor {
+	checkShape(s, shape...)
 
-	t, err := c.newTensor(ml.DTypeF32, shape)
-	if err != nil {
-		return nil, err
-	}
+	t := c.newTensor(ml.DTypeF32, shape)
 
 	if len(s) > 0 {
 		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
 	}
 
-	return t, nil
+	return t
 }
 
-func (c *Context) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
-	if err := checkShape(s, shape...); err != nil {
-		return nil, err
-	}
+func (c *Context) FromIntSlice(s []int32, shape ...int) ml.Tensor {
+	checkShape(s, shape...)
 
-	t, err := c.newTensor(ml.DTypeI32, shape)
-	if err != nil {
-		return nil, err
-	}
+	t := c.newTensor(ml.DTypeI32, shape)
 
 	if len(s) > 0 {
 		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
 	}
 
-	return t, nil
+	return t
 }
 
 func (c Context) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
@@ -711,12 +790,7 @@ func (c Context) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
 			arange = append(arange, int32(i))
 		}
 
-		t, err := c.Input().FromIntSlice(arange, len(arange))
-		if err != nil {
-			panic(err)
-		}
-
-		return t
+		return c.Input().FromIntSlice(arange, len(arange))
 	default:
 		panic("unsupported dtype for arange")
 	}
@@ -867,6 +941,13 @@ func (t *Tensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	}
 }
 
+func (t *Tensor) Div(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_div(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	return &Tensor{
 		b: t.b,
@@ -915,6 +996,8 @@ func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
 func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
 	if len(shape) != 4 {
 		panic("expected 4 dimensions")
+	} else if shape[3] != 0 {
+		panic("cuda does not support 4d tensors")
 	}
 
 	return &Tensor{
@@ -982,6 +1065,13 @@ func (t *Tensor) Scale(ctx ml.Context, s float64) ml.Tensor {
 	}
 }
 
+func (t *Tensor) SumRows(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sum_rows(ctx.(*Context).ctx, t.t),
+	}
+}
+
 func (t *Tensor) Softmax(ctx ml.Context) ml.Tensor {
 	return &Tensor{
 		b: t.b,
@@ -1053,28 +1143,15 @@ func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	}
 }
 
-const (
-	ropeTypeNorm   C.int = 0
-	ropeTypeNeox   C.int = 2
-	ropeTypeMrope  C.int = 8
-	ropeTypeVision C.int = 24
-)
-
-func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim, ropeType uint32, ropeBase, ropeScale float32, options ...ml.RopeOption) ml.Tensor {
+func (t *Tensor) RoPE(ctx ml.Context, positions ml.Tensor, ropeDim int, ropeBase, ropeScale float32, options ...func(*rope.Options)) ml.Tensor {
 	// Default options
-	opts := &ml.RopeOptions{
-		OriginalContextLen: 131072,
-	}
+	opts := &rope.Options{OriginalContextLength: 131072, Factors: &Tensor{}}
 
 	// Apply any provided options
 	for _, option := range options {
 		option(opts)
 	}
 
-	if ropeFactors == nil {
-		ropeFactors = &Tensor{b: t.b}
-	}
-
 	dequant := t.t
 	if C.ggml_is_quantized(t.t._type) {
 		dequant = C.ggml_cast(ctx.(*Context).ctx, t.t, C.GGML_TYPE_F32)
@@ -1085,11 +1162,11 @@ func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDi
 		t: C.ggml_rope_ext(
 			ctx.(*Context).ctx,
 			dequant,
-			positionIDs.(*Tensor).t,
-			ropeFactors.(*Tensor).t,
+			positions.(*Tensor).t,
+			opts.Factors.(*Tensor).t,
 			C.int(ropeDim),
-			C.int(ropeType),
-			C.int(opts.OriginalContextLen),
+			C.int(opts.Type),
+			C.int(opts.OriginalContextLength),
 			C.float(ropeBase),
 			C.float(ropeScale),
 			C.float(0.0),

ml/backend/ggml/ggml/include/ggml-alloc.h

@@ -66,6 +66,12 @@ GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph
 
 GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
 
+struct ggml_allocr_buffer_status {
+    size_t size;
+    bool allocated;
+};
+GGML_API struct ggml_allocr_buffer_status ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id);
+
 // Utils
 // Create a buffer and allocate all the tensors in a ggml_context
 GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);

ml/backend/ggml/ggml/include/ggml-backend.h

@@ -152,6 +152,7 @@ extern "C" {
     struct ggml_backend_dev_props {
         const char * name;
         const char * description;
+        const char * uuid;
         size_t memory_free;
         size_t memory_total;
         enum ggml_backend_dev_type type;
@@ -304,6 +305,12 @@ extern "C" {
 
     GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
 
+    struct ggml_backend_buffer_status {
+        size_t size;
+        bool allocated;
+    };
+    GGML_API struct ggml_backend_buffer_status ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+
     GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
     GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 

ml/backend/ggml/ggml/src/ggml-alloc.c

@@ -364,6 +364,7 @@ struct node_alloc {
 struct ggml_gallocr {
     ggml_backend_buffer_type_t * bufts; // [n_buffers]
     ggml_backend_buffer_t * buffers; // [n_buffers]
+    size_t *buffer_sizes; // [n_buffers]
     struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
     int n_buffers;
 
@@ -387,6 +388,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
     galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
     GGML_ASSERT(galloc->buffers != NULL);
 
+    galloc->buffer_sizes = calloc(n_bufs, sizeof(size_t));
+    GGML_ASSERT(galloc->buffer_sizes != NULL);
+
     galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
     GGML_ASSERT(galloc->buf_tallocs != NULL);
 
@@ -453,6 +457,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
     ggml_hash_set_free(&galloc->hash_set);
     free(galloc->hash_values);
     free(galloc->bufts);
+    free(galloc->buffer_sizes);
     free(galloc->buffers);
     free(galloc->buf_tallocs);
     free(galloc->node_allocs);
@@ -748,6 +753,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         }
     }
 
+    bool success = true;
+
     // reallocate buffers if needed
     for (int i = 0; i < galloc->n_buffers; i++) {
         // if the buffer type is used multiple times, we reuse the same buffer
@@ -769,15 +776,20 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
 
             ggml_backend_buffer_free(galloc->buffers[i]);
             galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
-            if (galloc->buffers[i] == NULL) {
+            if (galloc->buffers[i]) {
+                galloc->buffer_sizes[i] = ggml_backend_buffer_get_size(galloc->buffers[i]);
+                ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
+            } else {
                 GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
-                return false;
+                galloc->buffer_sizes[i] = new_size;
+                success = false;
             }
-            ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
+        } else {
+            galloc->buffer_sizes[i] = ggml_backend_buffer_get_size(galloc->buffers[i]);
         }
     }
 
-    return true;
+    return success;
 }
 
 bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
@@ -934,6 +946,24 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
     return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
 }
 
+struct ggml_allocr_buffer_status ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
+
+    for (int i = 0; i < buffer_id; i++) {
+        if (galloc->buf_tallocs[i] == galloc->buf_tallocs[buffer_id]) {
+            // This buffer is the same as a previous one due to the same buffer type being used multiple times
+            // (See above.) However, we need a different check because multiple buffers might be NULL in our
+            // case and we still want to know the attempted size.
+
+            struct ggml_allocr_buffer_status status = {0, true};
+            return status;
+        }
+    }
+
+    struct ggml_allocr_buffer_status status = {galloc->buffer_sizes[buffer_id], galloc->buffers[buffer_id] != NULL};
+    return status;
+}
+
 // utils
 
 static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) {

ml/backend/ggml/ggml/src/ggml-backend.cpp

@@ -1629,6 +1629,16 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
     return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
 }
 
+struct ggml_backend_buffer_status ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+
+    struct ggml_allocr_buffer_status allocr_status = ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
+    struct ggml_backend_buffer_status status = {allocr_status.size, allocr_status.allocated};
+
+    return status;
+}
+
 void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);

ml/backend/ggml/ggml/src/ggml-cpu/cpu.go

@@ -3,7 +3,7 @@ package cpu
 // #cgo CFLAGS: -O3 -Wno-implicit-function-declaration
 // #cgo CXXFLAGS: -std=c++17
 // #cgo CPPFLAGS: -I${SRCDIR}/amx -I${SRCDIR}/llamafile -I${SRCDIR}/.. -I${SRCDIR}/../../include
-// #cgo CPPFLAGS: -DGGML_USE_LLAMAFILE
+// #cgo CPPFLAGS: -DNDEBUG -DGGML_USE_LLAMAFILE
 // #cgo linux CPPFLAGS: -D_GNU_SOURCE
 // #cgo darwin,arm64 CPPFLAGS: -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
 // #cgo darwin,arm64 LDFLAGS: -framework Accelerate

ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2884,6 +2884,7 @@ struct ggml_backend_cuda_device_context {
     int device;
     std::string name;
     std::string description;
+    std::string uuid;
 };
 
 static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -2896,6 +2897,11 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
     return ctx->description.c_str();
 }
 
+static const char * ggml_backend_cuda_device_get_uuid(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ctx->uuid.c_str();
+}
+
 static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     ggml_cuda_set_device(ctx->device);
@@ -2910,6 +2916,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
 static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
     props->name        = ggml_backend_cuda_device_get_name(dev);
     props->description = ggml_backend_cuda_device_get_description(dev);
+    props->uuid        = ggml_backend_cuda_device_get_uuid(dev);
     props->type        = ggml_backend_cuda_device_get_type(dev);
     ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
 
@@ -3458,6 +3465,32 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;
 
+                #if !defined(GGML_USE_HIP)
+                char uuid[64];
+                snprintf(uuid, sizeof(uuid),
+                    "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
+                    (unsigned char)prop.uuid.bytes[0],
+                    (unsigned char)prop.uuid.bytes[1],
+                    (unsigned char)prop.uuid.bytes[2],
+                    (unsigned char)prop.uuid.bytes[3],
+                    (unsigned char)prop.uuid.bytes[4],
+                    (unsigned char)prop.uuid.bytes[5],
+                    (unsigned char)prop.uuid.bytes[6],
+                    (unsigned char)prop.uuid.bytes[7],
+                    (unsigned char)prop.uuid.bytes[8],
+                    (unsigned char)prop.uuid.bytes[9],
+                    (unsigned char)prop.uuid.bytes[10],
+                    (unsigned char)prop.uuid.bytes[11],
+                    (unsigned char)prop.uuid.bytes[12],
+                    (unsigned char)prop.uuid.bytes[13],
+                    (unsigned char)prop.uuid.bytes[14],
+                    (unsigned char)prop.uuid.bytes[15]
+                  );
+                dev_ctx->uuid = uuid;
+                #else
+                dev_ctx->uuid = "GPU-" + std::string(prop.uuid.bytes, 16);
+                #endif
+
                 ggml_backend_dev_t dev = new ggml_backend_device {
                     /* .iface   = */ ggml_backend_cuda_device_interface,
                     /* .reg     = */ &reg,

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.m

@@ -5703,6 +5703,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
 static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
     props->name        = ggml_backend_metal_device_get_name(dev);
     props->description = ggml_backend_metal_device_get_description(dev);
+    props->uuid        = "0";
     props->type        = ggml_backend_metal_device_get_type(dev);
     ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
     props->caps = (struct ggml_backend_dev_caps) {

ml/backend/ggml/ggml/src/ggml-metal/metal.go

@@ -4,6 +4,6 @@ package metal
 
 //go:generate sh -c "{ echo // Code generated by 'go generate'. DO NOT EDIT.; sed -e '/__embed_ggml-common.h__/r ../ggml-common.h' -e '/__embed_ggml-common.h__/d' -e '/#include \"ggml-metal-impl.h\"/r ggml-metal-impl.h' -e '/#include \"ggml-metal-impl.h\"/d' ggml-metal.metal; } >ggml-metal-embed.metal"
 
-// #cgo CPPFLAGS: -DGGML_METAL_EMBED_LIBRARY -I.. -I../../include
+// #cgo CPPFLAGS: -DGGML_METAL_NDEBUG -DGGML_METAL_EMBED_LIBRARY -I.. -I../../include
 // #cgo LDFLAGS: -framework Metal -framework MetalKit
 import "C"

ml/nn/fast/rope.go

@@ -0,0 +1,21 @@
+// fast provides implementations of fast (fused) operations for increased performance.
+package fast
+
+import (
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn/rope"
+)
+
+// fastRoPE is an interface for tensors that support fast rotary positional embedding.
+type fastRoPE interface {
+	RoPE(ctx ml.Context, positionIDs ml.Tensor, dim int, base, scale float32, options ...func(*rope.Options)) ml.Tensor
+}
+
+// RoPE applies rotary positional embedding to tensor `t`.
+func RoPE(ctx ml.Context, t, positions ml.Tensor, dim int, base, scale float32, options ...func(*rope.Options)) ml.Tensor {
+	if t, ok := t.(fastRoPE); ok {
+		return t.RoPE(ctx, positions, dim, base, scale, options...)
+	}
+
+	panic("RoPE not implemented for this tensor type")
+}

ml/nn/rope/rope.go

@@ -0,0 +1,33 @@
+package rope
+
+import "github.com/ollama/ollama/ml"
+
+// Options contains optional parameters for RoPE function
+type Options struct {
+	OriginalContextLength int
+	Type                  int
+	Factors               ml.Tensor
+}
+
+// WithOriginalContextLength sets a custom context length
+func WithOriginalContextLength(n int) func(*Options) {
+	return func(opts *Options) {
+		opts.OriginalContextLength = n
+	}
+}
+
+// WithType sets RoPE type to NeoX
+func WithTypeNeoX() func(*Options) {
+	return func(opts *Options) {
+		opts.Type = 2
+	}
+}
+
+// WithFactors sets custom rope factors
+func WithFactors(factors ml.Tensor) func(*Options) {
+	return func(opts *Options) {
+		if factors != nil {
+			opts.Factors = factors
+		}
+	}
+}

model/bytepairencoding.go

@@ -3,118 +3,16 @@ package model
 import (
 	"cmp"
 	"context"
+	"fmt"
 	"iter"
 	"log/slog"
-	"slices"
 	"strings"
-	"sync"
 
 	"github.com/dlclark/regexp2"
 	heap "github.com/emirpasic/gods/v2/trees/binaryheap"
 	"github.com/ollama/ollama/logutil"
 )
 
-type Special int32
-
-const (
-	SpecialBOS Special = iota
-	SpecialEOS
-)
-
-const (
-	TOKEN_TYPE_NORMAL = iota + 1
-	TOKEN_TYPE_UNKNOWN
-	TOKEN_TYPE_CONTROL
-	TOKEN_TYPE_USER_DEFINED
-	TOKEN_TYPE_UNUSED
-	TOKEN_TYPE_BYTE
-)
-
-type TextProcessor interface {
-	Encode(s string, addSpecial bool) ([]int32, error)
-	Decode([]int32) (string, error)
-	Is(int32, Special) bool
-	Vocabulary() *Vocabulary
-}
-
-type Vocabulary struct {
-	Values []string
-	Types  []int32
-	Scores []float32
-	Merges []string
-
-	BOS, EOS, EOT          int32
-	AddBOS, AddEOS, AddEOT bool
-
-	specialOnce sync.Once
-	special     []string
-
-	valuesOnce sync.Once
-	values     map[string]int32
-
-	mergeOnce sync.Once
-	merge     map[string]int32
-}
-
-func (v *Vocabulary) Is(id int32, special Special) bool {
-	switch special {
-	case SpecialBOS:
-		return id == v.BOS
-	case SpecialEOS:
-		return id == v.EOS || id == v.EOT
-	default:
-		return false
-	}
-}
-
-func (v *Vocabulary) Encode(s string) int32 {
-	v.valuesOnce.Do(func() {
-		v.values = make(map[string]int32, len(v.Values))
-		for i, value := range v.Values {
-			v.values[value] = int32(i)
-		}
-	})
-
-	if id, ok := v.values[s]; ok {
-		return id
-	}
-
-	return -1
-}
-
-func (v *Vocabulary) Decode(id int32) string {
-	return v.Values[id]
-}
-
-func (v *Vocabulary) SpecialVocabulary() []string {
-	v.specialOnce.Do(func() {
-		for i := range v.Values {
-			if slices.Contains([]int{105, 106}, i) {
-				v.special = append(v.special, v.Values[i])
-			} else if v.Types[i] == TOKEN_TYPE_CONTROL {
-				v.special = append(v.special, v.Values[i])
-			}
-		}
-	})
-
-	return v.special
-}
-
-func (v *Vocabulary) Merge(left, right string) int {
-	v.mergeOnce.Do(func() {
-		v.merge = make(map[string]int32, len(v.Merges))
-		for i, merge := range v.Merges {
-			v.merge[merge] = int32(i)
-		}
-	})
-
-	if id, ok := v.merge[left+" "+right]; ok {
-		return int(id)
-	}
-
-	return -1
-}
-
 type BytePairEncoding struct {
 	pre   *regexp2.Regexp
 	vocab *Vocabulary
@@ -304,30 +202,23 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
 		}
 	}
 
+	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
+
 	if addSpecial && len(ids) > 0 {
-		if bpe.vocab.AddBOS {
-			if ids[0] == bpe.vocab.BOS {
-				slog.Warn("adding bos token to prompt which already has it", "id", bpe.vocab.BOS)
-			}
-
-			slog.Debug("adding bos token to prompt", "id", bpe.vocab.BOS)
-			ids = append([]int32{bpe.vocab.BOS}, ids...)
-		}
-
-		if bpe.vocab.AddEOS {
-			if ids[len(ids)-1] == bpe.vocab.EOS {
-				slog.Warn("adding eos token to prompt which already has it", "id", bpe.vocab.EOS)
-			}
-
-			slog.Debug("adding eos token to prompt", "id", bpe.vocab.EOS)
-			ids = append(ids, bpe.vocab.EOS)
-		}
+		ids = bpe.vocab.addSpecials(ids)
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "ids", ids)
 	return ids, nil
 }
 
+type lazyIdsString struct {
+	ids []int32
+}
+
+func (l lazyIdsString) LogValue() slog.Value {
+	return slog.AnyValue(fmt.Sprint(l.ids))
+}
+
 func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 	var sb strings.Builder
 	for _, id := range ids {
@@ -352,6 +243,6 @@ func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String())
+	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
 	return sb.String(), nil
 }

model/input/input.go

@@ -2,16 +2,30 @@ package input
 
 import "github.com/ollama/ollama/ml"
 
+// Multimodal is a multimodal embedding or a component of one.
+// For example, it could be a row of an image that can be processed
+// independently.
+type Multimodal struct {
+	// Tensor is the embedding data. Implementations may chose what to
+	// store here or it may be nil if not needed. However, any ml.Tensor
+	// objects must be stored here and not in Data.
+	Tensor ml.Tensor
+
+	// Data is implementation-specific opaque data, such as metadata on how
+	// to layout Tensor. It may be nil if not needed. It may also store larger
+	// objects such as complete images if they are to be processed later.
+	Data any
+}
+
 // Input represents one token in the input stream
 type Input struct {
 	// Token is a single element of text.
 	Token int32
 
-	// Multimodal is opaque data representing a non-text
-	// element such as an image (or part of one if the image
-	// can be processed in pieces). It may be either together
-	// with Token or on its own.
-	Multimodal any
+	// Multimodal is represents a non-text element such as an
+	// image (or part of one if the image can be processed in pieces).
+	// It may be used either together with Token or on its own.
+	Multimodal []Multimodal
 
 	// MultimodalHash is a unique representation of the data
 	// stored in Multimodal, used for caching and comparing
@@ -32,7 +46,7 @@ type Input struct {
 // Positions slice.
 type MultimodalIndex struct {
 	Index      int
-	Multimodal any
+	Multimodal []Multimodal
 }
 
 // Batch contains the inputs for a model forward pass

model/model.go

@@ -40,12 +40,13 @@ type MultimodalProcessor interface {
 	// EncodeMultimodal processes a single input (such as an image) and
 	// generates an output (typically an embedding) that can be used by the model.
 	//
-	// The return value is most typically an ml.Tensor, however, different
-	// type are possible, such as an object containing a tensor plus
-	// additional metadata, a slice of tensors or even just the original input.
+	// The return value is one or more tensors, each with optional model-specific
+	// opaque metadata. Typically, the tensors might be views into an embedding
+	// with each view representing a chunk of data that can be processed independently
+	// in different batches.
 	//
 	// The result may be cached by the runner.
-	EncodeMultimodal(ml.Context, []byte) (any, error)
+	EncodeMultimodal(ml.Context, []byte) ([]input.Multimodal, error)
 
 	// PostTokenize is called after tokenization to allow the model to edit the
 	// input stream to correctly arrange multimodal elements.
@@ -97,14 +98,8 @@ func Register(name string, f func(fs.Config) (Model, error)) {
 }
 
 // New initializes a new model instance with the provided configuration based on the metadata in the model file
-func New(ctx context.Context, modelPath string, params ml.BackendParams) (Model, error) {
-	r, err := os.Open(modelPath)
-	if err != nil {
-		return nil, err
-	}
-	defer r.Close()
-
-	b, err := ml.NewBackend(ctx, r, params)
+func New(modelPath string, params ml.BackendParams) (Model, error) {
+	b, err := ml.NewBackend(modelPath, params)
 	if err != nil {
 		return nil, err
 	}
@@ -133,7 +128,7 @@ func NewTextProcessor(s string) (TextProcessor, error) {
 		return nil, err
 	}
 	defer r.Close()
-	meta, _, err := fsggml.Decode(r, -1)
+	meta, err := fsggml.Decode(r, -1)
 	if err != nil {
 		return nil, err
 	}
@@ -292,11 +287,7 @@ func Forward(ctx ml.Context, m Model, inputs []int32, batch input.Batch) (ml.Ten
 		return nil, errors.New("batch size cannot be less than 1")
 	}
 
-	var err error
-	batch.Inputs, err = ctx.Input().FromIntSlice(inputs, len(inputs))
-	if err != nil {
-		return nil, err
-	}
+	batch.Inputs = ctx.Input().FromIntSlice(inputs, len(inputs))
 
 	cache := m.Config().Cache
 	if cache != nil {

model/models/gemma2/model.go

@@ -7,6 +7,8 @@ import (
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
 )
@@ -43,10 +45,13 @@ func New(c fs.Config) (model.Model, error) {
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
-				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
-				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
-				// TODO: set EOT to EOS otherwise 0 will stop generation
-				EOT: int32(c.Uint("tokenizer.ggml.eos_token_id")),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
 			},
 		),
 		Layers: make([]Layer, c.Uint("block_count")),
@@ -80,11 +85,10 @@ type SelfAttention struct {
 
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
-	ropeType := uint32(2)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
+	q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -94,7 +98,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
+	k = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -124,7 +128,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, nil, uint32(m.Options.attnKeyLen), uint32(2), m.Options.ropeBase, m.Options.ropeScale), nil
+	return fast.RoPE(ctx, key, shift, m.Options.attnKeyLen, m.Options.ropeBase, m.Options.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 type MLP struct {
@@ -171,15 +175,8 @@ func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Ten
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
-	if err != nil {
-		return nil, err
-	}
-
-	outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
-	if err != nil {
-		return nil, err
-	}
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
 
 	hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
 	hiddenState = hiddenState.Scale(ctx, math.Sqrt(float64(m.Options.hiddenSize)))

model/models/gemma3/model.go

@@ -60,12 +60,16 @@ func New(c fs.Config) (model.Model, error) {
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
-				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
 				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
-				EOS:    int32(1),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
 				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-				EOT:    int32(106),
-				AddEOT: c.Bool("tokenizer.ggml.add_eot_token", false),
+				EOS: append(
+					[]int32{
+						int32(c.Uint("tokenizer.ggml.eos_token_id")),
+						int32(c.Uint("tokenizer.ggml.eot_token_id", 106)),
+					},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
 			},
 		),
 		ImageProcessor: newImageProcessor(c),
@@ -82,7 +86,7 @@ func New(c fs.Config) (model.Model, error) {
 	return &m, nil
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Layers) == 0 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -97,33 +101,30 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 		return nil, err
 	}
 
-	pixelValues, err := ctx.Input().FromFloatSlice(f32s,
+	pixelValues := ctx.Input().FromFloatSlice(f32s,
 		m.ImageProcessor.imageSize,
 		m.ImageProcessor.imageSize,
 		m.ImageProcessor.numChannels,
 	)
-	if err != nil {
-		return nil, err
-	}
 
 	visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
 	visionOutputs = m.MultiModalProjector.Forward(ctx, visionOutputs, m.imageSize, m.patchSize, m.VisionModel.eps)
-	return visionOutputs, nil
+	return []input.Multimodal{{Tensor: visionOutputs}}, nil
 }
 
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	var result []input.Input
 
 	for _, inp := range inputs {
-		if inp.Multimodal == nil {
+		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 		} else {
-			inputMultimodal := inp.Multimodal.(ml.Tensor)
+			inputMultimodal := inp.Multimodal[0].Tensor
 
 			result = append(result,
-				input.Input{Token: 108, SameBatch: inputMultimodal.Dim(1) + 3},               // "\n\n"
-				input.Input{Token: 255999},                                                   // "<start_of_image>""
-				input.Input{Multimodal: inputMultimodal, MultimodalHash: inp.MultimodalHash}, // image data is on the first placeholder
+				input.Input{Token: 108, SameBatch: inputMultimodal.Dim(1) + 3}, // "\n\n"
+				input.Input{Token: 255999},                                     // "<start_of_image>""
+				input.Input{Multimodal: []input.Multimodal{{Tensor: inputMultimodal}}, MultimodalHash: inp.MultimodalHash}, // image data is on the first placeholder
 			)
 
 			// add image token placeholders
@@ -140,15 +141,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
-	if err != nil {
-		return nil, err
-	}
-
-	outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
-	if err != nil {
-		return nil, err
-	}
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
 
 	return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
 }

model/models/gemma3/model_text.go

@@ -7,6 +7,8 @@ import (
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model/input"
 )
 
@@ -73,7 +75,6 @@ type TextSelfAttention struct {
 
 func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *TextConfig) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
-	ropeType := uint32(2)
 
 	ropeBase := opts.ropeLocalBase
 	if (layer+1)%gemmaGlobalCacheCount == 0 {
@@ -83,7 +84,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
 	q = sa.QueryNorm.Forward(ctx, q, opts.eps)
-	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
+	q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, ropeBase, opts.ropeScale, rope.WithTypeNeoX())
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -94,7 +95,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
 	k = sa.KeyNorm.Forward(ctx, k, opts.eps)
-	k = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
+	k = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, ropeBase, opts.ropeScale, rope.WithTypeNeoX())
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -112,7 +113,7 @@ func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.T
 		ropeBase = m.TextConfig.ropeGlobalBase
 	}
 
-	return key.RoPE(ctx, shift, nil, uint32(m.TextConfig.attnKeyLen), uint32(2), ropeBase, m.TextConfig.ropeScale), nil
+	return fast.RoPE(ctx, key, shift, m.TextConfig.attnKeyLen, ropeBase, m.TextConfig.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 type TextMLP struct {
@@ -165,7 +166,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	// set image embeddings
 	var except []int
 	for _, image := range batch.Multimodal {
-		visionOutputs := image.Multimodal.(ml.Tensor)
+		visionOutputs := image.Multimodal[0].Tensor
 		ctx.Forward(visionOutputs.Copy(ctx, hiddenState.View(ctx, image.Index*hiddenState.Stride(1), visionOutputs.Dim(0)*visionOutputs.Dim(1))))
 
 		for i := range visionOutputs.Dim(1) {

model/models/llama/model.go

@@ -1,22 +1,23 @@
 package llama
 
 import (
-	"fmt"
+	"cmp"
 	"math"
-	"strings"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
 )
 
 type Options struct {
 	hiddenSize, numHeads, numKVHeads int
+	headDim, ropeDim                 int
 	eps, ropeBase, ropeScale         float32
-	ropeDim                          uint32
 }
 
 type Model struct {
@@ -32,10 +33,6 @@ type Model struct {
 }
 
 func New(c fs.Config) (model.Model, error) {
-	if !strings.EqualFold(c.String("tokenizer.ggml.model"), "gpt2") {
-		return nil, fmt.Errorf("tokenizer %s not yet supported", c.String("tokenizer.ggml.model"))
-	}
-
 	m := Model{
 		BytePairEncoding: model.NewBytePairEncoding(
 			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
@@ -43,13 +40,13 @@ func New(c fs.Config) (model.Model, error) {
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),
-				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
 				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
-				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
 				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-				// TODO: set EOT to EOS otherwise 0 will stop generation
-				EOT:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
-				AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
 			},
 		),
 		Layers: make([]Layer, c.Uint("block_count")),
@@ -57,10 +54,11 @@ func New(c fs.Config) (model.Model, error) {
 			hiddenSize: int(c.Uint("embedding_length")),
 			numHeads:   int(c.Uint("attention.head_count")),
 			numKVHeads: int(c.Uint("attention.head_count_kv")),
+			headDim:    int(c.Uint("attention.key_length")),
+			ropeDim:    int(c.Uint("rope.dimension_count")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:   c.Float("rope.freq_base"),
 			ropeScale:  c.Float("rope.freq_scale", 1),
-			ropeDim:    c.Uint("rope.dimension_count"),
 		},
 	}
 
@@ -77,31 +75,31 @@ type SelfAttention struct {
 	RopeFactors ml.Tensor  `gguf:"rope_freqs.weight"`
 }
 
-func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
-	headDim := opts.hiddenSize / opts.numHeads
-	ropeType := uint32(0)
+	headDim := cmp.Or(opts.headDim, opts.hiddenSize/opts.numHeads)
+	ropeDim := cmp.Or(opts.ropeDim, headDim)
 
-	q := sa.Query.Forward(ctx, hiddenState)
-	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	query := sa.Query.Forward(ctx, hiddenState)
+	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
 
-	k := sa.Key.Forward(ctx, hiddenState)
-	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	key := sa.Key.Forward(ctx, hiddenState)
+	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
-	v := sa.Value.Forward(ctx, hiddenState)
-	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
+	value := sa.Value.Forward(ctx, hiddenState)
+	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
-	scaleFactor := 1.0 / math.Sqrt(float64(headDim))
-	kqv := nn.Attention(ctx, q, k, v, scaleFactor, cache)
-	kqv = kqv.Reshape(ctx, opts.hiddenSize, batchSize)
+	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 
-	return sa.Output.Forward(ctx, kqv)
+	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
+	attention = attention.Reshape(ctx, headDim*opts.numHeads, batchSize)
+	return sa.Output.Forward(ctx, attention)
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, uint32(0), m.ropeDim, m.ropeBase, m.ropeScale), nil
+	ropeDim := cmp.Or(m.ropeDim, m.hiddenSize/m.numHeads)
+	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(m.Layers[layer].SelfAttention.RopeFactors)), nil
 }
 
 type MLP struct {
@@ -122,11 +120,11 @@ type Layer struct {
 	MLP           *MLP
 }
 
-func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+func (l *Layer) Forward(ctx ml.Context, hiddenState, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	residual := hiddenState
 
 	hiddenState = l.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
-	hiddenState = l.SelfAttention.Forward(ctx, hiddenState, positionIDs, cache, opts)
+	hiddenState = l.SelfAttention.Forward(ctx, hiddenState, positions, cache, opts)
 
 	// In the final layer (outputs != nil), optimize by pruning to just the token positions
 	// we need logits for.
@@ -144,27 +142,19 @@ func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Ten
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
-	if err != nil {
-		return nil, err
-	}
-
-	outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
-	if err != nil {
-		return nil, err
-	}
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
 
 	hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
 
 	for i, layer := range m.Layers {
 		m.Cache.SetLayer(i)
 
-		var lastLayerOutputs ml.Tensor
+		var outputs ml.Tensor
 		if i == len(m.Layers)-1 {
-			lastLayerOutputs = outputs
+			outputs = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
 		}
 
-		hiddenState = layer.Forward(ctx, hiddenState, positions, lastLayerOutputs, m.Cache, m.Options)
+		hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, m.Cache, m.Options)
 	}
 
 	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)

model/models/llama4/model.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"image"
 	"slices"
-	"sync"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -41,13 +40,13 @@ func New(c fs.Config) (model.Model, error) {
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),
-				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
 				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
-				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
 				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-				// TODO: set EOT to EOS otherwise 0 will stop generation
-				EOT:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
-				AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
 			},
 		),
 		ImageProcessor: newImageProcessor(c),
@@ -63,7 +62,7 @@ func New(c fs.Config) (model.Model, error) {
 	return &m, nil
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Layers) < 1 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -78,10 +77,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 		return nil, err
 	}
 
-	tilesLocal, err := ctx.Input().FromFloatSlice(pixelsLocal, size.X, size.Y, m.numChannels)
-	if err != nil {
-		return nil, err
-	}
+	tilesLocal := ctx.Input().FromFloatSlice(pixelsLocal, size.X, size.Y, m.numChannels)
 
 	ratioW, ratioH := size.X/m.imageSize, size.Y/m.imageSize
 
@@ -92,81 +88,86 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 	pixelValues := tilesLocal
 
 	if len(pixelsGlobal) > 0 {
-		tilesGlobal, err := ctx.Input().FromFloatSlice(pixelsGlobal, m.imageSize, m.imageSize, m.numChannels)
-		if err != nil {
-			return nil, err
-		}
-
+		tilesGlobal := ctx.Input().FromFloatSlice(pixelsGlobal, m.imageSize, m.imageSize, m.numChannels)
 		pixelValues = pixelValues.Concat(ctx, tilesGlobal, 3)
 	}
 
 	visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
 	visionOutputs = visionOutputs.Reshape(ctx, visionOutputs.Dim(0), visionOutputs.Dim(1)*visionOutputs.Dim(2)*visionOutputs.Dim(3))
 	projectedOutputs := m.Projector.Forward(ctx, visionOutputs)
-	return &chunks{Model: m, Tensor: projectedOutputs, aspectRatio: image.Point{ratioW, ratioH}}, nil
+
+	var multimodal []input.Multimodal
+	aspectRatio := image.Point{ratioW, ratioH}
+
+	var offset int
+	patchesPerChunk := projectedOutputs.Dim(1)
+	if aspectRatio.Y*aspectRatio.X > 1 {
+		patchesPerChunk = projectedOutputs.Dim(1) / (aspectRatio.X*aspectRatio.Y + 1)
+
+		for range aspectRatio.Y {
+			for x := range aspectRatio.X {
+				view := projectedOutputs.View(ctx, projectedOutputs.Stride(1)*offset,
+					projectedOutputs.Dim(0), projectedOutputs.Stride(1),
+					patchesPerChunk)
+				var separator separator
+				if x < aspectRatio.X-1 {
+					separator.x = true // <|tile_x_separator|>
+				} else {
+					separator.y = true // <|tile_y_separator|>
+				}
+				multimodal = append(multimodal, input.Multimodal{Tensor: view, Data: &separator})
+				offset += patchesPerChunk
+			}
+		}
+	}
+
+	view := projectedOutputs.View(ctx, projectedOutputs.Stride(1)*offset,
+		projectedOutputs.Dim(0), projectedOutputs.Stride(1),
+		patchesPerChunk)
+	multimodal = append(multimodal, input.Multimodal{Tensor: view, Data: &separator{}})
+
+	return multimodal, nil
 }
 
-type chunks struct {
-	*Model
-	ml.Tensor
-	aspectRatio image.Point
-
-	dataOnce sync.Once
-	data     []float32
-}
-
-type chunk struct {
-	*chunks
-	s, n int
-}
-
-func (r *chunk) floats() []float32 {
-	r.dataOnce.Do(func() {
-		temp := r.Backend().NewContext()
-		defer temp.Close()
-		temp.Forward(r.Tensor).Compute(r.Tensor)
-		r.data = r.Floats()
-	})
-
-	return r.data[r.s*r.Dim(0) : (r.s+r.n)*r.Dim(0)]
+type separator struct {
+	x bool
+	y bool
 }
 
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	var result []input.Input
 	for _, inp := range inputs {
-		if inp.Multimodal == nil {
+		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 			continue
 		}
 
-		t := inp.Multimodal.(*chunks)
 		var imageInputs []input.Input
 		imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_start|>
 
-		var offset int
-		patchesPerChunk := t.Dim(1)
-		if t.aspectRatio.Y*t.aspectRatio.X > 1 {
-			patchesPerChunk = t.Dim(1) / (t.aspectRatio.X*t.aspectRatio.Y + 1)
+		for i, mm := range inp.Multimodal {
+			patchesPerChunk := mm.Tensor.Dim(1)
 
-			for range t.aspectRatio.Y {
-				for x := range t.aspectRatio.X {
-					imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: &chunk{t, offset, patchesPerChunk}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
-					imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
-					if x < t.aspectRatio.X-1 {
-						imageInputs = append(imageInputs, input.Input{Token: 200084}) // <|tile_x_separator|>
-					}
-					offset += patchesPerChunk
+			if i < len(inp.Multimodal)-1 {
+				separator := mm.Data.(*separator)
+
+				imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
+				imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
+
+				if separator.x {
+					imageInputs = append(imageInputs, input.Input{Token: 200084}) // <|tile_x_separator|>
 				}
-
-				imageInputs = append(imageInputs, input.Input{Token: 200085}) // <|tile_y_separator|>
+				if separator.y {
+					imageInputs = append(imageInputs, input.Input{Token: 200085}) // <|tile_y_separator|>
+				}
+			} else {
+				imageInputs = append(imageInputs, input.Input{Token: 200090})                                                                                                                      // <|image|>
+				imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
+				imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
+				imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_end|>
 			}
 		}
 
-		imageInputs = append(imageInputs, input.Input{Token: 200090})                                                                                                                 // <|image|>
-		imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: &chunk{t, offset, patchesPerChunk}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
-		imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
-		imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_end|>
-
 		result = append(result, imageInputs...)
 	}
 
@@ -174,15 +175,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
-	if err != nil {
-		return nil, err
-	}
-
-	outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
-	if err != nil {
-		return nil, err
-	}
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
 
 	return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
 }

model/models/llama4/model_text.go

@@ -8,6 +8,8 @@ import (
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model/input"
 )
 
@@ -31,8 +33,8 @@ func (sa *TextAttention) Forward(ctx ml.Context, hiddenStates, positions, attent
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
 	if useRope {
-		query = query.RoPE(ctx, positions, sa.RopeFactors, uint32(opts.ropeDim), uint32(0), opts.ropeBase, opts.ropeScale)
-		key = key.RoPE(ctx, positions, sa.RopeFactors, uint32(opts.ropeDim), uint32(0), opts.ropeBase, opts.ropeScale)
+		query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+		key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 	}
 
 	if opts.useQKNorm {
@@ -80,7 +82,7 @@ func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tens
 
 	nextStates := downStates.View(ctx, 0, hiddenStates.Dim(0), downStates.Stride(2), hiddenStates.Dim(2))
 	for i := 1; i < opts.numExpertsUsed; i++ {
-		nextStates.Add(ctx, downStates.View(ctx, i*downStates.Stride(1), hiddenStates.Dim(0), downStates.Stride(2), hiddenStates.Dim(2)))
+		nextStates = nextStates.Add(ctx, downStates.View(ctx, i*downStates.Stride(1), hiddenStates.Dim(0), downStates.Stride(2), hiddenStates.Dim(2)))
 	}
 
 	return nextStates
@@ -210,12 +212,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	hiddenStates := m.TokenEmbedding.Forward(ctx, inputs).Duplicate(ctx)
 
 	for _, mi := range batch.Multimodal {
-		f32s := mi.Multimodal.(*chunk).floats()
-		img, err := ctx.Input().FromFloatSlice(f32s, len(f32s)/m.hiddenSize, m.hiddenSize)
-		if err != nil {
-			panic(err)
-		}
-
+		img := mi.Multimodal[0].Tensor
 		ctx.Forward(img.Copy(ctx, hiddenStates.View(ctx, mi.Index*hiddenStates.Stride(1), img.Dim(0)*img.Dim(1))))
 	}
 
@@ -226,11 +223,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 			scales[i] = float32(math.Log(math.Floor(((float64(p)+1.0)/float64(m.attentionFloorScale))+1.0))*m.attentionScale + 1.0)
 		}
 
-		var err error
-		attentionScales, err = ctx.Input().FromFloatSlice(scales, 1, 1, len(scales))
-		if err != nil {
-			panic(err)
-		}
+		attentionScales = ctx.Input().FromFloatSlice(scales, 1, 1, len(scales))
 	}
 
 	for i, layer := range m.Layers {
@@ -255,5 +248,5 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, m.Layers[layer].Attention.RopeFactors, uint32(0), uint32(m.ropeDim), m.ropeBase, m.ropeScale), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(m.Layers[layer].Attention.RopeFactors)), nil
 }

model/models/llama4/model_vision.go

@@ -245,10 +245,7 @@ func (m *VisionModel) rotaryEmbedding(ctx ml.Context) (ml.Tensor, ml.Tensor) {
 		}
 	}
 
-	ropeFreqs, err := ctx.Input().FromFloatSlice(freqs, freqDim/2, numPatches, 2)
-	if err != nil {
-		panic(err)
-	}
+	ropeFreqs := ctx.Input().FromFloatSlice(freqs, freqDim/2, numPatches, 2)
 
 	ropeFreqs = ropeFreqs.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
 	ropeFreqs = ropeFreqs.Reshape(ctx, freqDim, 1, numPatches)

model/models/mistral3/model.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"image"
 	"slices"
-	"sync"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -32,31 +31,26 @@ var _ model.MultimodalProcessor = (*Model)(nil)
 var _ model.TextProcessor = (*Model)(nil)
 
 func New(c fs.Config) (model.Model, error) {
-	textModel, err := NewTextModel(c)
-	if err != nil {
-		return nil, err
-	}
-
 	m := &Model{
-		TextModel:           textModel,
-		VisionModel:         newVisionModel(c),
-		ImageProcessor:      newImageProcessor(c),
-		MultiModalProjector: newMultiModalProjector(c),
 		BytePairEncoding: model.NewBytePairEncoding(
 			c.String("tokenizer.ggml.pretokenizer", `[^\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+`),
 			&model.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),
-				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id", 1)),
 				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
-				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id", 2)),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
 				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-				// TODO: set EOT to EOS otherwise 0 will stop generation
-				EOT:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
-				AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
 			},
 		),
+		TextModel:           newTextModel(c),
+		VisionModel:         newVisionModel(c),
+		ImageProcessor:      newImageProcessor(c),
+		MultiModalProjector: newMultiModalProjector(c),
 	}
 
 	m.Cache = kvcache.NewCausalCache(m.TextModel.Shift)
@@ -105,7 +99,7 @@ func newMultiModalProjector(c fs.Config) *MultiModalProjector {
 	}
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Layers) == 0 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -120,46 +114,20 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 		return nil, err
 	}
 
-	pixelValues, err := ctx.Input().FromFloatSlice(f32s, size.X, size.Y, m.ImageProcessor.numChannels)
-	if err != nil {
-		return nil, err
-	}
+	pixelValues := ctx.Input().FromFloatSlice(f32s, size.X, size.Y, m.ImageProcessor.numChannels)
 
 	visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
 	features, size := m.MultiModalProjector.Forward(ctx, visionOutputs, size)
 
 	// split into patches to be sent to the text transformer
-	parent := imageFeatures{tensor: features}
-	rows := make([]*imageRow, size.Y)
+	rows := make([]input.Multimodal, size.Y)
 	for i := range rows {
-		rows[i] = &imageRow{parent: &parent, s: i, shape: []int{features.Dim(0), size.X}}
+		rows[i].Tensor = features.View(ctx, features.Stride(1)*size.X*i, features.Dim(0), features.Stride(1), size.X)
 	}
 
 	return rows, nil
 }
 
-type imageFeatures struct {
-	tensor ml.Tensor
-
-	dataOnce sync.Once
-	data     []float32
-}
-
-type imageRow struct {
-	parent *imageFeatures
-	s      int
-	shape  []int
-}
-
-func (r *imageRow) data() []float32 {
-	n := 1
-	for _, s := range r.shape {
-		n *= s
-	}
-
-	return r.parent.data[r.s*n : (r.s+1)*n]
-}
-
 // PostTokenize arranges Mistral 3's inputs for the forward pass
 // In Mistral 3 and Pixtral, the input patches are arranged as follows:
 // [IMG]...[IMG][IMG_BREAK][IMG]...[IMG][IMG_BREAK][IMG]...[IMG][IMG_END]
@@ -168,15 +136,14 @@ func (r *imageRow) data() []float32 {
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	var result []input.Input
 	for _, inp := range inputs {
-		if inp.Multimodal == nil {
+		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 		} else {
-			inputMultimodal := inp.Multimodal.([]*imageRow)
-			for i, row := range inputMultimodal {
+			for i, row := range inp.Multimodal {
 				// [IMG]
-				result = append(result, input.Input{Token: 10, Multimodal: row, MultimodalHash: inp.MultimodalHash, SameBatch: row.shape[1]})
-				result = append(result, slices.Repeat([]input.Input{{Token: 10}}, row.shape[1]-1)...)
-				if i == len(inputMultimodal)-1 {
+				result = append(result, input.Input{Token: 10, Multimodal: []input.Multimodal{{Tensor: row.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: row.Tensor.Dim(1)})
+				result = append(result, slices.Repeat([]input.Input{{Token: 10}}, row.Tensor.Dim(1)-1)...)
+				if i == len(inp.Multimodal)-1 {
 					// [IMG_END]
 					result = append(result, input.Input{Token: 13})
 				} else {
@@ -191,15 +158,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
-	if err != nil {
-		return nil, err
-	}
-
-	outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
-	if err != nil {
-		return nil, err
-	}
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
 
 	return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache), nil
 }

model/models/mistral3/model_text.go

@@ -1,27 +1,24 @@
 package mistral3
 
 import (
-	"fmt"
+	"cmp"
 	"math"
-	"strings"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
-	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/ml/nn/fast"
 	"github.com/ollama/ollama/model/input"
 )
 
 type TextOptions struct {
-	hiddenSize, numHeads, numKVHeads, headDim int
-	eps, ropeBase, ropeScale                  float32
-	ropeDim                                   uint32
+	hiddenSize, numHeads, numKVHeads int
+	headDim, ropeDim                 int
+	eps, ropeBase, ropeScale         float32
 }
 
 type TextModel struct {
-	model.Base
-
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
 	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
@@ -39,19 +36,15 @@ type SelfAttention struct {
 
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *TextOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
-	ropeType := uint32(0)
-	headDim := opts.headDim
-	if headDim == 0 {
-		headDim = opts.hiddenSize / opts.numHeads
-	}
+	headDim := cmp.Or(opts.headDim, opts.hiddenSize/opts.numHeads)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, nil, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	q = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale)
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, nil, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	k = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -62,7 +55,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, nil, uint32(0), m.ropeDim, m.ropeBase, m.ropeScale), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale), nil
 }
 
 type MLP struct {
@@ -109,20 +102,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 
 	// image embeddings
 	for _, image := range batch.Multimodal {
-		row := image.Multimodal.(*imageRow)
-		row.parent.dataOnce.Do(func() {
-			// use a new, throwaway context so the image tensor is not added to the graph
-			temp := m.Backend().NewContext()
-			temp.Forward(row.parent.tensor).Compute(row.parent.tensor)
-			row.parent.data = row.parent.tensor.Floats()
-			temp.Close()
-		})
-
-		imageFeature, err := ctx.Input().FromFloatSlice(row.data(), row.shape...)
-		if err != nil {
-			panic(err)
-		}
-
+		imageFeature := image.Multimodal[0].Tensor
 		ctx.Forward(imageFeature.Copy(ctx, hiddenState.View(ctx, image.Index*hiddenState.Stride(1), imageFeature.Dim(0)*imageFeature.Dim(1))))
 	}
 
@@ -141,24 +121,18 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	return m.Output.Forward(ctx, hiddenState)
 }
 
-func NewTextModel(c fs.Config) (*TextModel, error) {
-	if !strings.EqualFold(c.String("tokenizer.ggml.model"), "gpt2") {
-		return nil, fmt.Errorf("tokenizer %s not yet supported", c.String("tokenizer.ggml.model"))
-	}
-
-	textModel := &TextModel{
+func newTextModel(c fs.Config) *TextModel {
+	return &TextModel{
 		Layers: make([]Layer, c.Uint("block_count")),
 		TextOptions: &TextOptions{
 			hiddenSize: int(c.Uint("embedding_length")),
 			numHeads:   int(c.Uint("attention.head_count")),
 			numKVHeads: int(c.Uint("attention.head_count_kv")),
 			headDim:    int(c.Uint("attention.key_length")),
+			ropeDim:    int(c.Uint("rope.dimension_count")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:   c.Float("rope.freq_base"),
 			ropeScale:  c.Float("rope.freq_scale", 1),
-			ropeDim:    c.Uint("rope.dimension_count"),
 		},
 	}
-
-	return textModel, nil
 }

model/models/mistral3/model_vision.go

@@ -110,15 +110,8 @@ func (m *VisionModel) positionalEmbedding(ctx ml.Context, positionIDs ml.Tensor)
 		}
 	}
 
-	h, err := ctx.Input().FromFloatSlice(frequenciesHeight, maxPatchesPerSide, frequencies/2)
-	if err != nil {
-		panic(err)
-	}
-
-	w, err := ctx.Input().FromFloatSlice(frequenciesWidth, maxPatchesPerSide, frequencies/2)
-	if err != nil {
-		panic(err)
-	}
+	h := ctx.Input().FromFloatSlice(frequenciesHeight, maxPatchesPerSide, frequencies/2)
+	w := ctx.Input().FromFloatSlice(frequenciesWidth, maxPatchesPerSide, frequencies/2)
 
 	h = h.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
 	w = w.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
@@ -151,10 +144,7 @@ func (m *VisionModel) Forward(ctx ml.Context, pixelValues ml.Tensor) ml.Tensor {
 		}
 	}
 
-	positionIDs, err := ctx.Input().FromIntSlice(positions, len(positions))
-	if err != nil {
-		panic(err)
-	}
+	positionIDs := ctx.Input().FromIntSlice(positions, len(positions))
 
 	positionEmbedding := m.positionalEmbedding(ctx, positionIDs)
 	cos, sin := positionEmbedding.Cos(ctx), positionEmbedding.Sin(ctx)
@@ -170,7 +160,7 @@ func (m *VisionModel) Forward(ctx ml.Context, pixelValues ml.Tensor) ml.Tensor {
 
 func newVisionModel(c fs.Config) *VisionModel {
 	return &VisionModel{
-		Layers: make([]VisionEncoderLayer, c.Uint("vision.block_count", 24)),
+		Layers: make([]VisionEncoderLayer, c.Uint("vision.block_count")),
 		VisionModelOptions: &VisionModelOptions{
 			hiddenSize:       int(c.Uint("vision.embedding_length", 1024)),
 			numHeads:         int(c.Uint("vision.attention.head_count", 16)),

model/models/mllama/model.go

@@ -3,6 +3,7 @@ package mllama
 import (
 	"bytes"
 	"image"
+	"slices"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -37,13 +38,13 @@ func New(c fs.Config) (model.Model, error) {
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),
-				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
 				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
-				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
 				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-				// TODO: set EOT to EOS otherwise 0 will stop generation
-				EOT:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
-				AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
 			},
 		),
 		ImageProcessor: newImageProcessor(c),
@@ -58,7 +59,7 @@ func New(c fs.Config) (model.Model, error) {
 	return &m, nil
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Transformer.Layers) == 0 || len(m.GlobalTransformer.Layers) == 0 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -73,21 +74,20 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 		return nil, err
 	}
 
-	pixelValues, err := ctx.Input().FromFloatSlice(f32s, m.imageSize, m.imageSize, m.numChannels, ratio.numTiles())
-	if err != nil {
-		return nil, err
+	if ratio.numTiles() < m.maxNumTiles {
+		// Pad tiles to maxNumTiles
+		f32s = slices.Grow(f32s, m.imageSize*m.imageSize*m.numChannels*m.maxNumTiles)
+		f32s = f32s[:m.imageSize*m.imageSize*m.numChannels*m.maxNumTiles]
 	}
 
-	pixelValues = pixelValues.Pad(ctx, 0, 0, 0, m.ImageProcessor.maxNumTiles-ratio.numTiles())
-
-	aspectRatio, err := ctx.Input().FromIntSlice([]int32{int32(ratio.rank)}, 1)
-	if err != nil {
-		return nil, err
-	}
+	pixelValues := ctx.Input().FromFloatSlice(f32s, m.imageSize, m.imageSize, m.numChannels, m.maxNumTiles)
+	aspectRatio := ctx.Input().FromIntSlice([]int32{int32(ratio.rank)}, 1)
 
 	positionIDs := ctx.Arange(0, 1601, 1, ml.DTypeI32)
 	crossAttentionStates := m.VisionModel.Forward(ctx, pixelValues, positionIDs, aspectRatio)
-	return m.Projector.Forward(ctx, crossAttentionStates), nil
+	projectedOutputs := m.Projector.Forward(ctx, crossAttentionStates)
+
+	return []input.Multimodal{{Tensor: projectedOutputs}}, nil
 }
 
 func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
@@ -103,18 +103,11 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 	var crossAttentionStates ml.Tensor
 	if len(batch.Multimodal) > 0 {
-		crossAttentionStates = batch.Multimodal[len(batch.Multimodal)-1].Multimodal.(ml.Tensor)
+		crossAttentionStates = batch.Multimodal[len(batch.Multimodal)-1].Multimodal[0].Tensor
 	}
 
-	positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
-	if err != nil {
-		return nil, err
-	}
-
-	outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
-	if err != nil {
-		return nil, err
-	}
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
 
 	// TODO: attention mask, cross attention mask
 	return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, crossAttentionStates, nil, m.Cache.(*kvcache.WrapperCache)), nil

model/models/mllama/model_text.go

@@ -8,6 +8,8 @@ import (
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
 )
 
 type TextSelfAttention struct {
@@ -21,15 +23,14 @@ type TextSelfAttention struct {
 func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
-	ropeType := uint32(0)
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	query = query.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 
 	key := sa.Key.Forward(ctx, hiddenState)
 	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	key = key.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
+	key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithFactors(sa.RopeFactors))
 
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -44,7 +45,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.T
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
 	// This will only get called for layers in the cache, which are just the self attention layers
 	if sa, ok := m.Transformer.Layers[layer].(*TextSelfAttentionDecoderLayer); ok {
-		return key.RoPE(ctx, shift, sa.SelfAttention.RopeFactors, m.ropeDim, uint32(0), m.ropeBase, m.ropeScale), nil
+		return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(sa.SelfAttention.RopeFactors)), nil
 	}
 
 	return key, nil
@@ -199,8 +200,8 @@ func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, outputs,
 
 type TextModelOptions struct {
 	hiddenSize, numHeads, numKVHeads int
+	ropeDim                          int
 	eps, ropeBase, ropeScale         float32
-	ropeDim                          uint32
 
 	crossAttentionLayers []int32
 }
@@ -240,10 +241,10 @@ func newTextModel(c fs.Config) *TextModel {
 			hiddenSize:           int(c.Uint("embedding_length")),
 			numHeads:             int(c.Uint("attention.head_count")),
 			numKVHeads:           int(c.Uint("attention.head_count_kv")),
+			ropeDim:              int(c.Uint("rope.dimension_count")),
 			eps:                  c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:             c.Float("rope.freq_base"),
 			ropeScale:            c.Float("rope.freq_scale", 1),
-			ropeDim:              c.Uint("rope.dimension_count"),
 			crossAttentionLayers: c.Ints("attention.cross_attention_layers"),
 		},
 	}

model/models/mllama/model_vision.go

@@ -16,8 +16,6 @@ type VisionSelfAttention struct {
 	Key    *nn.Linear `gguf:"attn_k"`
 	Value  *nn.Linear `gguf:"attn_v"`
 	Output *nn.Linear `gguf:"attn_output"`
-
-	Gate ml.Tensor `gguf:"attn_gate"`
 }
 
 func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *VisionModelOptions) ml.Tensor {
@@ -25,27 +23,16 @@ func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenState ml.Tensor, op
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, query.Dim(1), batchSize)
-	query = query.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
 
 	key := sa.Key.Forward(ctx, hiddenState)
 	key = key.Reshape(ctx, headDim, opts.numHeads, key.Dim(1), batchSize)
-	key = key.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
 
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numHeads, value.Dim(1), batchSize)
-	value = value.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)
 
-	scores := key.Mulmat(ctx, query)
-	scores = scores.Scale(ctx, 1.0/math.Sqrt(float64(headDim)))
-	scores = scores.Softmax(ctx)
-
-	attention := value.Mulmat(ctx, scores)
-	attention = attention.Reshape(ctx, headDim, attention.Dim(1), opts.numHeads, batchSize)
-	attention = attention.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
+	attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(headDim)), nil)
 	attention = attention.Reshape(ctx, opts.hiddenSize, attention.Dim(2), batchSize)
-
-	hiddenState = sa.Output.Forward(ctx, attention)
-	return hiddenState
+	return sa.Output.Forward(ctx, attention)
 }
 
 type VisionMLP struct {
@@ -76,21 +63,18 @@ func (e *VisionEncoderLayer) Forward(ctx ml.Context, hiddenState ml.Tensor, opts
 	// self attention
 	hiddenState = e.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
 	hiddenState = e.SelfAttention.Forward(ctx, hiddenState, opts)
-
 	if e.AttentionGate != nil {
 		hiddenState = hiddenState.Mul(ctx, e.AttentionGate)
 	}
 	hiddenState = hiddenState.Add(ctx, residual)
 	residual = hiddenState
 
-	// feed forward
 	hiddenState = e.MLPNorm.Forward(ctx, hiddenState, opts.eps)
 	hiddenState = e.MLP.Forward(ctx, hiddenState, opts)
-	hiddenState = hiddenState.Add(ctx, residual)
 	if e.MLPGate != nil {
 		hiddenState = hiddenState.Mul(ctx, e.MLPGate)
 	}
-
+	hiddenState = hiddenState.Add(ctx, residual)
 	return hiddenState
 }
 

model/models/models.go

@@ -7,5 +7,7 @@ import (
 	_ "github.com/ollama/ollama/model/models/llama4"
 	_ "github.com/ollama/ollama/model/models/mistral3"
 	_ "github.com/ollama/ollama/model/models/mllama"
+	_ "github.com/ollama/ollama/model/models/qwen2"
 	_ "github.com/ollama/ollama/model/models/qwen25vl"
+	_ "github.com/ollama/ollama/model/models/qwen3"
 )

model/models/qwen2/model.go

@@ -0,0 +1,164 @@
+package qwen2
+
+import (
+	"cmp"
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	hiddenSize, numHeads, numKVHeads int
+	headDim, ropeDim                 int
+	eps, ropeBase, ropeScale         float32
+}
+
+type Attention struct {
+	Query  *nn.Linear `gguf:"attn_q"`
+	Key    *nn.Linear `gguf:"attn_k"`
+	Value  *nn.Linear `gguf:"attn_v"`
+	Output *nn.Linear `gguf:"attn_output"`
+}
+
+func (attn Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	batchSize := hiddenStates.Dim(1)
+	headDim := cmp.Or(opts.headDim, opts.hiddenSize/opts.numHeads)
+	ropeDim := cmp.Or(opts.ropeDim, headDim)
+
+	query := attn.Query.Forward(ctx, hiddenStates)
+	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
+
+	key := attn.Key.Forward(ctx, hiddenStates)
+	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
+
+	value := attn.Value.Forward(ctx, hiddenStates)
+	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
+
+	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+
+	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
+	attention = attention.Reshape(ctx, headDim*opts.numHeads, batchSize)
+
+	return attn.Output.Forward(ctx, attention)
+}
+
+type MLP struct {
+	Gate *nn.Linear `gguf:"ffn_gate"`
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+}
+
+func (mlp MLP) Forward(ctx ml.Context, hiddenStates ml.Tensor) ml.Tensor {
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	return mlp.Down.Forward(ctx, hiddenStates)
+}
+
+type DecoderLayer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	Attention     *Attention
+	MLPNorm       *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP           *MLP
+}
+
+func (d DecoderLayer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	residual := hiddenStates
+
+	hiddenStates = d.AttentionNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = d.Attention.Forward(ctx, hiddenStates, positions, cache, opts)
+	if outputs != nil {
+		hiddenStates = hiddenStates.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenStates = hiddenStates.Add(ctx, residual)
+	residual = hiddenStates
+
+	hiddenStates = d.MLPNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = d.MLP.Forward(ctx, hiddenStates)
+	return hiddenStates.Add(ctx, residual)
+}
+
+type Model struct {
+	model.Base
+	model.BytePairEncoding
+
+	TokenEmbedding *nn.Embedding  `gguf:"token_embd"`
+	Layers         []DecoderLayer `gguf:"blk"`
+	OutputNorm     *nn.RMSNorm    `gguf:"output_norm"`
+	Output         *nn.Linear     `gguf:"output,alt:token_embd"`
+
+	Options
+}
+
+// Forward implements model.Model.
+func (m Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+
+	hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
+		}
+
+		hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, &m.Options)
+	}
+
+	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
+	hiddenStates = m.Output.Forward(ctx, hiddenStates)
+	return hiddenStates, nil
+}
+
+func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	ropeDim := cmp.Or(m.ropeDim, m.hiddenSize/m.numHeads)
+	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
+}
+
+func New(c fs.Config) (model.Model, error) {
+	m := Model{
+		Layers: make([]DecoderLayer, c.Uint("block_count")),
+		BytePairEncoding: model.NewBytePairEncoding(
+			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+		),
+		Options: Options{
+			hiddenSize: int(c.Uint("embedding_length")),
+			numHeads:   int(c.Uint("attention.head_count")),
+			numKVHeads: int(c.Uint("attention.head_count_kv")),
+			headDim:    int(c.Uint("attention.key_length")),
+			ropeDim:    int(c.Uint("rope.dimension_count")),
+			ropeBase:   c.Float("rope.freq_base"),
+			ropeScale:  c.Float("rope.freq_scale", 1),
+			eps:        c.Float("attention.layer_norm_rms_epsilon"),
+		},
+	}
+
+	m.Cache = kvcache.NewCausalCache(m.Shift)
+	return &m, nil
+}
+
+func init() {
+	model.Register("qwen2", New)
+}

model/models/qwen25vl/model.go

@@ -5,7 +5,6 @@ import (
 	"fmt"
 	"image"
 	"slices"
-	"sync"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -35,12 +34,13 @@ func New(c fs.Config) (model.Model, error) {
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),
-				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
-				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),
-				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
 				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
-				EOT:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
-				AddEOT: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
 			},
 		),
 		TextModel:      NewTextModel(c),
@@ -69,15 +69,12 @@ func (m *Model) PixelValues(ctx ml.Context, multimodalData []byte) (ml.Tensor, *
 		m.ImageProcessor.patchSize * m.ImageProcessor.patchSize
 	numPatches := grid.Temporal * grid.Height * grid.Width
 
-	pixelValues, err := ctx.Input().FromFloatSlice(f32s, patchDim, numPatches)
-	if err != nil {
-		return nil, nil, fmt.Errorf("failed to create tensor from image: %w", err)
-	}
+	pixelValues := ctx.Input().FromFloatSlice(f32s, patchDim, numPatches)
 
 	return pixelValues, grid, nil
 }
 
-func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input.Multimodal, error) {
 	if len(m.VisionModel.Layers) == 0 {
 		return nil, model.ErrNoVisionModel
 	}
@@ -88,31 +85,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, er
 	}
 
 	visionOutputs := m.VisionModel.Forward(ctx, pixels, grid)
-	return &chunks{Model: m, Tensor: visionOutputs}, nil
-}
-
-type chunks struct {
-	*Model
-	ml.Tensor
-
-	dataOnce sync.Once
-	data     []float32
-}
-
-type chunk struct {
-	*chunks
-	s, n int
-}
-
-func (r *chunk) floats() []float32 {
-	r.dataOnce.Do(func() {
-		temp := r.Backend().NewContext()
-		defer temp.Close()
-		temp.Forward(r.Tensor).Compute(r.Tensor)
-		r.data = r.Floats()
-	})
-
-	return r.data[r.s*r.Dim(0) : (r.s+r.n)*r.Dim(0)]
+	return []input.Multimodal{{Tensor: visionOutputs}}, nil
 }
 
 // PostTokenize arranges Qwen-2.5-VL's inputs for the forward pass
@@ -142,18 +115,15 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 				result = append(result, input.Input{Token: pre[i]})
 			}
 
-			// This is an image token with multimodal data
-			chunksData := inp.Multimodal.(*chunks)
-			patchesPerChunk := chunksData.Dim(1)
+			patchesPerChunk := inp.Multimodal[0].Tensor.Dim(1)
 
 			// First add the vision start token
-			result = append(result, input.Input{Token: visionStartToken, SameBatch: patchesPerChunk + 2})
+			result = append(result, input.Input{Token: visionStartToken})
 
 			// Add the image token with the multimodal tensor data at the first position
-			// Create a chunk with proper s and n values
 			result = append(result, input.Input{
 				Token:          imageToken,
-				Multimodal:     &chunk{chunks: chunksData, s: 0, n: patchesPerChunk},
+				Multimodal:     inp.Multimodal,
 				MultimodalHash: inp.MultimodalHash,
 				SameBatch:      patchesPerChunk,
 			})
@@ -169,15 +139,8 @@ func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions, err := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
-	if err != nil {
-		return nil, err
-	}
-
-	outputs, err := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
-	if err != nil {
-		return nil, err
-	}
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	outputs := ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
 
 	return m.TextModel.Forward(ctx, batch.Inputs, positions, outputs, batch, m.Cache)
 }

model/models/qwen25vl/model_text.go

@@ -7,13 +7,15 @@ import (
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model/input"
 )
 
 type TextOptions struct {
-	ctxLen, hiddenSize, numHeads, numKVHeads int
-	eps, ropeBase, ropeScale                 float32
-	ropeDim, defaultContextLen               uint32
+	hiddenSize, numHeads, numKVHeads int
+	ropeDim, originalContextLength   int
+	eps, ropeBase, ropeScale         float32
 }
 
 type TextModel struct {
@@ -29,15 +31,14 @@ func NewTextModel(c fs.Config) *TextModel {
 	m := TextModel{
 		Layers: make([]Layer, c.Uint("block_count")),
 		TextOptions: &TextOptions{
-			ctxLen:            int(c.Uint("context_length")),
-			hiddenSize:        int(c.Uint("embedding_length")),
-			numHeads:          int(c.Uint("attention.head_count")),
-			numKVHeads:        int(c.Uint("attention.head_count_kv")),
-			eps:               c.Float("attention.layer_norm_rms_epsilon"),
-			ropeBase:          c.Float("rope.freq_base"),
-			ropeScale:         c.Float("rope.freq_scale", 1),
-			ropeDim:           c.Uint("rope.dimension_count", 128),
-			defaultContextLen: c.Uint("context_length", 128000),
+			hiddenSize:            int(c.Uint("embedding_length")),
+			numHeads:              int(c.Uint("attention.head_count")),
+			numKVHeads:            int(c.Uint("attention.head_count_kv")),
+			ropeDim:               int(c.Uint("rope.dimension_count", 128)),
+			originalContextLength: int(c.Uint("context_length", 128000)),
+			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:              c.Float("rope.freq_base"),
+			ropeScale:             c.Float("rope.freq_scale", 1),
 		},
 	}
 
@@ -59,11 +60,11 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, nil, opts.ropeDim, 2, opts.ropeBase, opts.ropeScale, ml.WithContextLen(opts.defaultContextLen))
+	q = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, nil, opts.ropeDim, 2, opts.ropeBase, opts.ropeScale, ml.WithContextLen(opts.defaultContextLen))
+	k = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -77,7 +78,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 // Shift applies rotary position embeddings to the key tensor for causal attention caching
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, nil, m.ropeDim, 2, m.ropeBase, m.ropeScale, ml.WithContextLen(m.defaultContextLen)), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithOriginalContextLength(m.originalContextLength), rope.WithTypeNeoX()), nil
 }
 
 // MLP implements the feed-forward network component with SwiGLU activation
@@ -129,12 +130,7 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	hiddenStates := m.TokenEmbedding.Forward(ctx, inputs).Duplicate(ctx)
 
 	for _, mi := range batch.Multimodal {
-		f32s := mi.Multimodal.(*chunk).floats()
-		img, err := ctx.Input().FromFloatSlice(f32s, len(f32s)/m.hiddenSize, m.hiddenSize)
-		if err != nil {
-			panic(err)
-		}
-
+		img := mi.Multimodal[0].Tensor
 		ctx.Forward(img.Copy(ctx, hiddenStates.View(ctx, mi.Index*hiddenStates.Stride(1), img.Dim(0)*img.Dim(1))))
 	}
 

model/models/qwen25vl/model_vision.go

@@ -1,7 +1,6 @@
 package qwen25vl
 
 import (
-	"fmt"
 	"math"
 	"slices"
 
@@ -44,10 +43,8 @@ func blockDiagonalMask(ctx ml.Context, seqLength int, bounds []int, numHeads int
 		}
 	}
 
-	mask, err := ctx.Input().FromFloatSlice(flat, seqLength, seqLength)
-	if err != nil {
-		panic(err)
-	}
+	mask := ctx.Input().FromFloatSlice(flat, seqLength, seqLength)
+
 	// Reshape to match [seqLength, seqLength, 1] for broadcasting
 	mask = mask.Reshape(ctx, seqLength, seqLength, 1)
 
@@ -303,10 +300,7 @@ func (m *VisionModel) WindowIndex(ctx ml.Context, grid *Grid) (ml.Tensor, []int)
 		}
 	}
 
-	t, err := ctx.Input().FromIntSlice(index, len(index))
-	if err != nil {
-		panic(err)
-	}
+	t := ctx.Input().FromIntSlice(index, len(index))
 
 	return t, bounds
 }
@@ -326,10 +320,7 @@ func (m *VisionModel) PositionalEmbedding(ctx ml.Context, grid *Grid) ml.Tensor
 			freqVals[i*freq+j] = float32(i) / float32(math.Pow(theta, float64(j*2)/float64(dim)))
 		}
 	}
-	freqs, err := ctx.Input().FromFloatSlice(freqVals, freq, maxGridSize)
-	if err != nil {
-		panic(fmt.Errorf("failed to create tensor from frequencies: %w", err))
-	}
+	freqs := ctx.Input().FromFloatSlice(freqVals, freq, maxGridSize)
 
 	// Create position coordinates (y,x pairs) for the grid
 	// In PyTorch: Equivalent to generating position ids with torch.arange()
@@ -339,10 +330,7 @@ func (m *VisionModel) PositionalEmbedding(ctx ml.Context, grid *Grid) ml.Tensor
 			coords = append(coords, int32(y), int32(x))
 		}
 	}
-	pos, err := ctx.Input().FromIntSlice(coords, 2, grid.Width, grid.Height)
-	if err != nil {
-		panic(fmt.Errorf("failed to create tensor from positions: %w", err))
-	}
+	pos := ctx.Input().FromIntSlice(coords, 2, grid.Width, grid.Height)
 
 	// Reshape and permute positions to match spatial merging pattern
 	pos = pos.Reshape(ctx, 2, grid.Width, merge, grid.Height/merge)

model/models/qwen3/model.go

@@ -0,0 +1,233 @@
+package qwen3
+
+import (
+	"cmp"
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	hiddenSize, numHeads, numKVHeads int
+	eps                              float32
+	ropeBase, ropeScale              float32
+
+	keyLength, valueLength int
+
+	numExperts, numExpertsUsed int
+	normTopKProb               bool
+}
+
+func (o Options) headDim() int {
+	return cmp.Or(o.keyLength, o.valueLength, o.hiddenSize/o.numHeads)
+}
+
+type Attention struct {
+	QueryNorm *nn.RMSNorm `gguf:"attn_q_norm"`
+	Query     *nn.Linear  `gguf:"attn_q"`
+	KeyNorm   *nn.RMSNorm `gguf:"attn_k_norm"`
+	Key       *nn.Linear  `gguf:"attn_k"`
+	Value     *nn.Linear  `gguf:"attn_v"`
+	Output    *nn.Linear  `gguf:"attn_output"`
+}
+
+func (sa *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	batchSize := hiddenStates.Dim(1)
+
+	query := sa.Query.Forward(ctx, hiddenStates)
+	key := sa.Key.Forward(ctx, hiddenStates)
+	value := sa.Value.Forward(ctx, hiddenStates)
+
+	query = query.Reshape(ctx, opts.headDim(), opts.numHeads, batchSize)
+	key = key.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
+	value = value.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
+
+	query = sa.QueryNorm.Forward(ctx, query, opts.eps)
+	key = sa.KeyNorm.Forward(ctx, key, opts.eps)
+
+	query = fast.RoPE(ctx, query, positions, opts.headDim(), opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+	key = fast.RoPE(ctx, key, positions, opts.headDim(), opts.ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+
+	attention := nn.Attention(ctx, query, key, value, 1./math.Sqrt(float64(opts.headDim())), cache)
+	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)
+	return sa.Output.Forward(ctx, attention)
+}
+
+type MLP interface {
+	Forward(ml.Context, ml.Tensor, *Options) ml.Tensor
+}
+
+type sparse struct {
+	Router *nn.Linear `gguf:"ffn_gate_inp"`
+	Gate   ml.Tensor  `gguf:"ffn_gate_exps.weight"`
+	Up     ml.Tensor  `gguf:"ffn_up_exps.weight"`
+	Down   ml.Tensor  `gguf:"ffn_down_exps.weight"`
+}
+
+func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
+	hiddenDim, sequenceLength, batchSize := hiddenStates.Dim(0), hiddenStates.Dim(1), hiddenStates.Dim(2)
+	hiddenStates = hiddenStates.Reshape(ctx, hiddenDim, sequenceLength*batchSize)
+	routerLogits := mlp.Router.Forward(ctx, hiddenStates)
+
+	routingWeights := routerLogits.Softmax(ctx)
+	selectedExperts := routingWeights.TopK(ctx, opts.numExpertsUsed)
+	routingWeights = routingWeights.Reshape(ctx, 1, opts.numExperts, hiddenStates.Dim(1)).Rows(ctx, selectedExperts)
+	if opts.normTopKProb {
+		routingWeights = routingWeights.Reshape(ctx, opts.numExpertsUsed, hiddenStates.Dim(1))
+		routingWeights = routingWeights.Div(ctx, routingWeights.SumRows(ctx))
+		routingWeights = routingWeights.Reshape(ctx, 1, opts.numExpertsUsed, hiddenStates.Dim(1))
+	}
+
+	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0), 1, hiddenStates.Dim(1))
+
+	upStates := mlp.Up.MulmatID(ctx, hiddenStates, selectedExperts)
+
+	hiddenStates = mlp.Gate.MulmatID(ctx, hiddenStates, selectedExperts)
+	hiddenStates = hiddenStates.SILU(ctx)
+	hiddenStates = hiddenStates.Mul(ctx, upStates)
+
+	experts := mlp.Down.MulmatID(ctx, hiddenStates, selectedExperts)
+	experts = experts.Mul(ctx, routingWeights)
+
+	nextStates := experts.View(ctx, 0, experts.Dim(0), experts.Stride(2), experts.Dim(2))
+	for i := 1; i < opts.numExpertsUsed; i++ {
+		nextStates = nextStates.Add(ctx, experts.View(ctx, i*experts.Stride(1), experts.Dim(0), experts.Stride(2), experts.Dim(2)))
+	}
+
+	return nextStates
+}
+
+type dense struct {
+	Gate *nn.Linear `gguf:"ffn_gate"`
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+}
+
+func (mlp *dense) Forward(ctx ml.Context, hiddenStates ml.Tensor, _ *Options) ml.Tensor {
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	return mlp.Down.Forward(ctx, hiddenStates)
+}
+
+type Layer struct {
+	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
+	*Attention
+
+	MLPNorm *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP
+}
+
+func (d *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	residual := hiddenStates
+	hiddenStates = d.AttentionNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = d.Attention.Forward(ctx, hiddenStates, positions, cache, opts)
+
+	if outputs != nil {
+		hiddenStates = hiddenStates.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenStates = hiddenStates.Add(ctx, residual)
+
+	residual = hiddenStates
+	hiddenStates = d.MLPNorm.Forward(ctx, hiddenStates, opts.eps)
+	hiddenStates = d.MLP.Forward(ctx, hiddenStates, opts)
+	return hiddenStates.Add(ctx, residual)
+}
+
+type Model struct {
+	model.Base
+	model.BytePairEncoding
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
+	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
+
+	Layers []Layer `gguf:"blk"`
+
+	*Options
+}
+
+// Forward implements model.Model.
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+
+	hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs))
+		}
+
+		hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, m.Options)
+	}
+
+	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
+	return m.Output.Forward(ctx, hiddenStates), nil
+}
+
+func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	return fast.RoPE(ctx, key, shift, m.headDim(), m.ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
+}
+
+var _ model.Model = (*Model)(nil)
+
+func New(c fs.Config) (model.Model, error) {
+	layers := make([]Layer, c.Uint("block_count"))
+	for i := range layers {
+		if c.String("general.architecture") == "qwen3moe" {
+			layers[i].MLP = &sparse{}
+		} else {
+			layers[i].MLP = &dense{}
+		}
+	}
+
+	m := Model{
+		BytePairEncoding: model.NewBytePairEncoding(
+			`(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+		),
+		Layers: layers,
+		Options: &Options{
+			hiddenSize:     int(c.Uint("embedding_length")),
+			numHeads:       int(c.Uint("attention.head_count")),
+			numKVHeads:     int(c.Uint("attention.head_count_kv")),
+			keyLength:      int(c.Uint("attention.key_length")),
+			valueLength:    int(c.Uint("attention.value_length")),
+			eps:            c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:       c.Float("rope.freq_base"),
+			ropeScale:      c.Float("rope.freq_scale", 1),
+			numExperts:     int(c.Uint("expert_count")),
+			numExpertsUsed: int(c.Uint("expert_used_count")),
+			normTopKProb:   c.Bool("norm_top_k_prob", true),
+		},
+	}
+
+	m.Cache = kvcache.NewCausalCache(m.Shift)
+	return &m, nil
+}
+
+func init() {
+	model.Register("qwen3", New)
+	model.Register("qwen3moe", New)
+}

model/sentencepiece.go

@@ -182,27 +182,12 @@ func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error)
 		}
 	}
 
+	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
+
 	if addSpecial && len(ids) > 0 {
-		if spm.vocab.AddBOS {
-			if ids[0] == spm.vocab.BOS {
-				slog.Warn("adding bos token to prompt which already has it", "id", spm.vocab.BOS)
-			}
-
-			slog.Debug("adding bos token to prompt", "id", spm.vocab.BOS)
-			ids = append([]int32{spm.vocab.BOS}, ids...)
-		}
-
-		if spm.vocab.AddEOS {
-			if ids[len(ids)-1] == spm.vocab.EOS {
-				slog.Warn("adding eos token to prompt which already has it", "id", spm.vocab.EOS)
-			}
-
-			slog.Debug("adding eos token to prompt", "id", spm.vocab.EOS)
-			ids = append(ids, spm.vocab.EOS)
-		}
+		ids = spm.vocab.addSpecials(ids)
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "ids", ids)
 	return ids, nil
 }
 
@@ -261,6 +246,6 @@ func (spm SentencePieceModel) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String())
+	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "ids", ids, "string", sb.String())
 	return sb.String(), nil
 }

model/textprocessor.go

@@ -0,0 +1,17 @@
+package model
+
+const (
+	TOKEN_TYPE_NORMAL = iota + 1
+	TOKEN_TYPE_UNKNOWN
+	TOKEN_TYPE_CONTROL
+	TOKEN_TYPE_USER_DEFINED
+	TOKEN_TYPE_UNUSED
+	TOKEN_TYPE_BYTE
+)
+
+type TextProcessor interface {
+	Encode(s string, addSpecial bool) ([]int32, error)
+	Decode([]int32) (string, error)
+	Is(int32, Special) bool
+	Vocabulary() *Vocabulary
+}

model/vocabulary.go

@@ -0,0 +1,112 @@
+package model
+
+import (
+	"log/slog"
+	"slices"
+	"sync"
+)
+
+type Special int32
+
+const (
+	SpecialBOS Special = iota
+	SpecialEOS
+)
+
+type Vocabulary struct {
+	Values []string
+	Types  []int32
+	Scores []float32
+	Merges []string
+
+	BOS, EOS       []int32
+	AddBOS, AddEOS bool
+
+	specialOnce sync.Once
+	special     []string
+
+	valuesOnce sync.Once
+	values     map[string]int32
+
+	mergeOnce sync.Once
+	merge     map[string]int32
+}
+
+func (v *Vocabulary) Is(id int32, special Special) bool {
+	switch special {
+	case SpecialBOS:
+		return slices.Contains(v.BOS, id)
+	case SpecialEOS:
+		return slices.Contains(v.EOS, id)
+	default:
+		return false
+	}
+}
+
+func (v *Vocabulary) addSpecials(ids []int32) []int32 {
+	if v.AddBOS && len(v.BOS) > 0 {
+		if slices.Contains(v.BOS, ids[0]) {
+			slog.Warn("adding bos token to prompt which already has it", "id", v.BOS)
+		}
+
+		slog.Debug("adding bos token to prompt", "id", v.BOS)
+		ids = append([]int32{v.BOS[0]}, ids...)
+	}
+
+	if v.AddEOS && len(v.EOS) > 0 {
+		if slices.Contains(v.BOS, ids[len(ids)-1]) {
+			slog.Warn("adding eos token to prompt which already has it", "id", v.EOS)
+		}
+
+		slog.Debug("adding eos token to prompt", "id", v.EOS)
+		ids = append(ids, v.EOS[0])
+	}
+
+	return ids
+}
+
+func (v *Vocabulary) Encode(s string) int32 {
+	v.valuesOnce.Do(func() {
+		v.values = make(map[string]int32, len(v.Values))
+		for i, value := range v.Values {
+			v.values[value] = int32(i)
+		}
+	})
+
+	if id, ok := v.values[s]; ok {
+		return id
+	}
+
+	return -1
+}
+
+func (v *Vocabulary) Decode(id int32) string {
+	return v.Values[id]
+}
+
+func (v *Vocabulary) SpecialVocabulary() []string {
+	v.specialOnce.Do(func() {
+		for i := range v.Values {
+			if v.Types[i] == TOKEN_TYPE_CONTROL {
+				v.special = append(v.special, v.Values[i])
+			}
+		}
+	})
+
+	return v.special
+}
+
+func (v *Vocabulary) Merge(left, right string) int {
+	v.mergeOnce.Do(func() {
+		v.merge = make(map[string]int32, len(v.Merges))
+		for i, merge := range v.Merges {
+			v.merge[merge] = int32(i)
+		}
+	})
+
+	if id, ok := v.merge[left+" "+right]; ok {
+		return int(id)
+	}
+
+	return -1
+}

readline/types.go

@@ -61,6 +61,8 @@ const (
 	ColorGrey    = Esc + "[38;5;245m"
 	ColorDefault = Esc + "[0m"
 
+	ColorBold = Esc + "[1m"
+
 	StartBracketedPaste = Esc + "[?2004h"
 	EndBracketedPaste   = Esc + "[?2004l"
 )

runner/llamarunner/cache.go

@@ -104,8 +104,8 @@ func (c *InputCache) LoadCacheSlot(prompt []input, cachePrompt bool) (*InputCach
 	slog.Debug("loading cache slot", "id", slot.Id, "cache", len(slot.Inputs), "prompt", len(prompt),
 		"used", numPast, "remaining", len(prompt)-numPast)
 
+	slot.Inputs = prompt[:numPast]
 	prompt = prompt[numPast:]
-	slot.Inputs = slot.Inputs[:numPast]
 
 	return slot, prompt, nil
 }

runner/ollamarunner/cache.go

@@ -136,8 +136,8 @@ func (c *InputCache) LoadCacheSlot(prompt []input.Input) (*InputCacheSlot, []inp
 	slog.Debug("loading cache slot", "id", slot.Id, "cache", len(slot.Inputs), "prompt", len(prompt),
 		"used", numPast, "remaining", int32(len(prompt))-numPast)
 
+	slot.Inputs = prompt[:numPast]
 	prompt = prompt[numPast:]
-	slot.Inputs = slot.Inputs[:numPast]
 
 	return slot, prompt, nil
 }

runner/ollamarunner/cache_test.go

@@ -3,7 +3,6 @@ package ollamarunner
 import (
 	"errors"
 	"fmt"
-	"image"
 	"testing"
 	"time"
 
@@ -12,10 +11,6 @@ import (
 )
 
 func TestCountCommon(t *testing.T) {
-	imgA := image.NewRGBA(image.Rect(0, 0, 100, 100))
-	imgB := image.NewRGBA(image.Rect(0, 0, 50, 50))
-	imgC := image.NewRGBA(image.Rect(50, 50, 100, 100))
-
 	tests := []struct {
 		name     string
 		t1       []input.Input
@@ -36,20 +31,20 @@ func TestCountCommon(t *testing.T) {
 		},
 		{
 			name:     "Image Prefix",
-			t1:       []input.Input{{Multimodal: imgA, MultimodalHash: 1}},
-			t2:       []input.Input{{Multimodal: imgA, MultimodalHash: 1}, {Multimodal: imgB, MultimodalHash: 2}, {Multimodal: imgC, MultimodalHash: 3}},
+			t1:       []input.Input{{MultimodalHash: 1}},
+			t2:       []input.Input{{MultimodalHash: 1}, {MultimodalHash: 2}, {MultimodalHash: 3}},
 			expected: 1,
 		},
 		{
 			name:     "Mixed",
-			t1:       []input.Input{{Token: 1}, {Multimodal: imgA, MultimodalHash: 1}},
-			t2:       []input.Input{{Token: 1}, {Multimodal: imgA, MultimodalHash: 1}, {Token: 5}},
+			t1:       []input.Input{{Token: 1}, {MultimodalHash: 1}},
+			t2:       []input.Input{{Token: 1}, {MultimodalHash: 1}, {Token: 5}},
 			expected: 2,
 		},
 		{
 			name:     "Mixed, Same Length",
-			t1:       []input.Input{{Token: 1}, {Multimodal: imgA, MultimodalHash: 1}},
-			t2:       []input.Input{{Token: 1}, {Multimodal: imgB, MultimodalHash: 2}},
+			t1:       []input.Input{{Token: 1}, {MultimodalHash: 1}},
+			t2:       []input.Input{{Token: 1}, {MultimodalHash: 2}},
 			expected: 1,
 		},
 		{

runner/ollamarunner/multimodal.go

@@ -0,0 +1,113 @@
+package ollamarunner
+
+import (
+	"errors"
+
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
+)
+
+// Tensors can't be used across multiple compute graphs. This is a problem
+// if a single embedding is split across batches using views since all of
+// the views will have the same source tensor. We also don't want to
+// recompute the entire embedding for each batch.
+//
+// To avoid this, we compute all of the tensors for the embedding on the
+// first use and then store the result in system memory. When we need
+// additional tensors, we recreate them from the stored data.
+
+// multimodalEntry represents the embeddings of a single object (such
+// as an image).
+type multimodalEntry struct {
+	// mm is the original set of tensors created by EncodeMultimodal
+	mm []input.Multimodal
+
+	// data is the computed result of mm. Nil if not yet computed
+	data [][]float32
+}
+
+// multimodalStore maps from an individual tensor (of which there
+// may be many in a single multimodal object) to its parent embedding
+type multimodalStore map[ml.Tensor]*multimodalEntry
+
+func newMultimodalStore() multimodalStore {
+	return make(multimodalStore)
+}
+
+// addMultimodal stores an embedding for later use in a compute graph
+func (m multimodalStore) addMultimodal(embedding []input.Multimodal) {
+	entry := &multimodalEntry{mm: embedding}
+
+	for _, e := range embedding {
+		if e.Tensor != nil {
+			m[e.Tensor] = entry
+		}
+	}
+}
+
+// getMultimodal takes a source set of tensors (which may contain a whole or
+// parts of one or more images) and returns the equivalent that can be used in
+// the current context
+func (m multimodalStore) getMultimodal(backend ml.Backend, ctx ml.Context, in []input.Multimodal, reserve bool) ([]input.Multimodal, error) {
+	out := make([]input.Multimodal, len(in))
+	for i := range out {
+		if in[i].Tensor != nil {
+			var err error
+			out[i].Tensor, err = m.getTensor(backend, ctx, in[i].Tensor, reserve)
+			if err != nil {
+				return nil, err
+			}
+		}
+
+		out[i].Data = in[i].Data
+	}
+
+	return out, nil
+}
+
+func (m multimodalStore) getTensor(backend ml.Backend, ctx ml.Context, in ml.Tensor, reserve bool) (ml.Tensor, error) {
+	entry := m[in]
+
+	if entry.data == nil {
+		computeCtx := backend.NewContext()
+		defer computeCtx.Close()
+
+		var tensors []ml.Tensor
+		for _, t := range entry.mm {
+			if t.Tensor != nil {
+				tensors = append(tensors, t.Tensor)
+			}
+		}
+
+		if len(tensors) == 0 {
+			return nil, nil
+		}
+
+		computeCtx.Forward(tensors...)
+		entry.data = make([][]float32, len(entry.mm))
+
+		if !reserve {
+			computeCtx.Compute(tensors...)
+
+			for i, t := range entry.mm {
+				if t.Tensor != nil {
+					entry.data[i] = t.Tensor.Floats()
+				}
+			}
+		} else {
+			computeCtx.Reserve()
+		}
+	}
+
+	for i, t := range entry.mm {
+		if in == t.Tensor {
+			if !reserve {
+				return ctx.Input().FromFloatSlice(entry.data[i], t.Tensor.Shape()...), nil
+			} else {
+				return ctx.Input().Empty(t.Tensor.DType(), t.Tensor.Shape()...), nil
+			}
+		}
+	}
+
+	return nil, errors.New("multimodal tensor not found")
+}

runner/ollamarunner/runner.go

@@ -1,12 +1,14 @@
 package ollamarunner
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
 	"flag"
 	"fmt"
 	"hash/maphash"
+	"image"
 	"log"
 	"log/slog"
 	"net"
@@ -20,6 +22,7 @@ import (
 	"time"
 	"unicode/utf8"
 
+	"golang.org/x/image/bmp"
 	"golang.org/x/sync/semaphore"
 
 	"github.com/ollama/ollama/api"
@@ -40,6 +43,9 @@ type Sequence struct {
 	// multimodal embeddings
 	ctxs []ml.Context
 
+	// mmStore holds multimodal embeddings to mange memory and enable splitting across batches
+	mmStore multimodalStore
+
 	// batch index
 	iBatch int
 
@@ -101,7 +107,7 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 
 	startTime := time.Now()
 
-	inputs, ctxs, err := s.inputs(prompt, images)
+	inputs, ctxs, mmStore, err := s.inputs(prompt, images)
 	if err != nil {
 		return nil, fmt.Errorf("failed to process inputs: %w", err)
 	} else if len(inputs) == 0 {
@@ -156,6 +162,7 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 
 	return &Sequence{
 		ctxs:                ctxs,
+		mmStore:             mmStore,
 		inputs:              inputs,
 		numPromptInputs:     len(inputs),
 		startProcessingTime: startTime,
@@ -174,9 +181,10 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 // inputs processes the prompt and images into a list of inputs
 // by splitting the prompt on [img-<n>] tags, tokenizing text and
 // decoding images
-func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, []ml.Context, error) {
+func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, []ml.Context, multimodalStore, error) {
 	var inputs []input.Input
 	var ctxs []ml.Context
+	var mmStore multimodalStore
 
 	var parts []string
 	var matches [][]string
@@ -187,6 +195,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 		re := regexp.MustCompile(`\[img-(\d+)\]`)
 		parts = re.Split(prompt, -1)
 		matches = re.FindAllStringSubmatch(prompt, -1)
+		mmStore = newMultimodalStore()
 	} else {
 		parts = []string{prompt}
 	}
@@ -196,7 +205,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 		// text - tokenize
 		tokens, err := s.model.(model.TextProcessor).Encode(part, i == 0)
 		if err != nil {
-			return nil, nil, err
+			return nil, nil, nil, err
 		}
 
 		for _, t := range tokens {
@@ -216,7 +225,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 			}
 
 			if imageIndex < 0 {
-				return nil, nil, fmt.Errorf("invalid image index: %d", n)
+				return nil, nil, nil, fmt.Errorf("invalid image index: %d", n)
 			}
 
 			ctx := s.model.Backend().NewContext()
@@ -224,13 +233,15 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 			ctxs = append(ctxs, ctx)
 			imageEmbeddings, err := multimodalProcessor.EncodeMultimodal(ctx, images[imageIndex].Data)
 			if err != nil {
-				return nil, nil, err
+				return nil, nil, nil, err
 			}
 
 			s.multimodalHash.Reset()
 			_, _ = s.multimodalHash.Write(images[imageIndex].Data)
 			imageHash := s.multimodalHash.Sum64()
 
+			mmStore.addMultimodal(imageEmbeddings)
+
 			inputs = append(inputs, input.Input{Multimodal: imageEmbeddings, MultimodalHash: imageHash})
 			postTokenize = true
 		}
@@ -240,11 +251,11 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, [
 		var err error
 		inputs, err = multimodalProcessor.PostTokenize(inputs)
 		if err != nil {
-			return nil, nil, err
+			return nil, nil, nil, err
 		}
 	}
 
-	return inputs, ctxs, nil
+	return inputs, ctxs, mmStore, nil
 }
 
 type Server struct {
@@ -363,6 +374,9 @@ func (s *Server) processBatch() error {
 	}
 	defer s.mu.Unlock()
 
+	ctx := s.model.Backend().NewContext()
+	defer ctx.Close()
+
 	var batchInputs []int32
 	var batch input.Batch
 
@@ -433,7 +447,11 @@ func (s *Server) processBatch() error {
 
 			batchInputs = append(batchInputs, inp.Token)
 			if inp.Multimodal != nil {
-				batch.Multimodal = append(batch.Multimodal, input.MultimodalIndex{Index: len(batchInputs) - 1, Multimodal: inp.Multimodal})
+				mm, err := seq.mmStore.getMultimodal(s.model.Backend(), ctx, inp.Multimodal, false)
+				if err != nil {
+					return err
+				}
+				batch.Multimodal = append(batch.Multimodal, input.MultimodalIndex{Index: len(batchInputs) - 1, Multimodal: mm})
 			}
 
 			batch.Positions = append(batch.Positions, int32(len(seq.cache.Inputs)+len(seq.pendingInputs)))
@@ -459,9 +477,6 @@ func (s *Server) processBatch() error {
 		return nil
 	}
 
-	ctx := s.model.Backend().NewContext()
-	defer ctx.Close()
-
 	modelOutput, err := model.Forward(ctx, s.model, batchInputs, batch)
 	if err != nil {
 		return fmt.Errorf("failed to decode batch: %w", err)
@@ -720,12 +735,71 @@ func (s *Server) reserveWorstCaseGraph() error {
 	ctx := s.model.Backend().NewContext()
 	defer ctx.Close()
 
+	var err error
+	inputs := make([]input.Input, s.batchSize)
+	mmStore := newMultimodalStore()
+
+	// Multimodal strategy:
+	// - Encode a 2048x2048 image. This assumes that a single image of this
+	//   size is sufficient to trigger the worst case. This is currently true
+	//   because for existing models, only a single image fits in a batch.
+	// - Add the embedding to a full batch of tokens - this is necessary because
+	//   the model may be looking for non-image data, such as <image> tags.
+	// - Run PostTokenize to execute any transformations between generated
+	//   embeddings and what the forward pass expects.
+	// - The result may now be larger than a batch (images may not fit in a
+	//   single batch), so trim based on what will fit and must be grouped together.
+	// - Fill out the rest of the space with text tokens.
+	if multimodalProcessor, ok := s.model.(model.MultimodalProcessor); ok {
+		mmCtx := s.model.Backend().NewContext()
+		defer mmCtx.Close()
+
+		img := image.NewGray(image.Rect(0, 0, 2048, 2048))
+		var buf bytes.Buffer
+		bmp.Encode(&buf, img)
+
+		if inputs[0].Multimodal, err = multimodalProcessor.EncodeMultimodal(mmCtx, buf.Bytes()); err == nil {
+			mmStore.addMultimodal(inputs[0].Multimodal)
+
+			inputs, err = multimodalProcessor.PostTokenize(inputs)
+			if err != nil {
+				return err
+			}
+
+			for i, inp := range inputs {
+				minBatch := 1 + inp.SameBatch
+				if minBatch > s.batchSize {
+					inputs = inputs[i:min(i+minBatch, len(inputs))]
+					break
+				} else if i+minBatch > s.batchSize {
+					inputs = inputs[:i]
+					break
+				}
+			}
+
+			if len(inputs) < s.batchSize {
+				newInputs := make([]input.Input, s.batchSize)
+				copy(newInputs, inputs)
+				inputs = newInputs
+			}
+		}
+	}
+
 	var batch input.Batch
 
-	inputs := make([]int32, s.batchSize)
+	batchInputs := make([]int32, len(inputs))
 	batch.Positions = make([]int32, len(inputs))
 	batch.Sequences = make([]int, len(inputs))
-	for i := range inputs {
+	for i, inp := range inputs {
+		batchInputs[i] = inp.Token
+		if inp.Multimodal != nil {
+			mm, err := mmStore.getMultimodal(s.model.Backend(), ctx, inp.Multimodal, true)
+			if err != nil {
+				return err
+			}
+			batch.Multimodal = append(batch.Multimodal, input.MultimodalIndex{Index: i, Multimodal: mm})
+		}
+
 		batch.Positions[i] = int32(i)
 	}
 
@@ -734,11 +808,7 @@ func (s *Server) reserveWorstCaseGraph() error {
 		batch.Outputs[i] = int32(i)
 	}
 
-	var err error
-	batch.Inputs, err = ctx.Input().FromIntSlice(inputs, len(inputs))
-	if err != nil {
-		return err
-	}
+	batch.Inputs = ctx.Input().FromIntSlice(batchInputs, len(batchInputs))
 
 	cache := s.model.Config().Cache
 	if cache != nil {
@@ -753,16 +823,12 @@ func (s *Server) reserveWorstCaseGraph() error {
 		return err
 	}
 
-	err = ctx.Forward(t).Reserve()
-	if err != nil {
-		return err
-	}
+	ctx.Forward(t).Reserve()
 
 	return nil
 }
 
-func (s *Server) loadModel(
-	ctx context.Context,
+func (s *Server) initModel(
 	mpath string,
 	params ml.BackendParams,
 	lpath multiLPath,
@@ -770,21 +836,21 @@ func (s *Server) loadModel(
 	kvCacheType string,
 	kvSize int,
 	multiUserCache bool,
-) {
+) error {
 	var err error
-	s.model, err = model.New(ctx, mpath, params)
+	s.model, err = model.New(mpath, params)
 	if err != nil {
-		panic(err)
+		return err
 	}
 
 	// TODO(jessegross): LoRA loading
 	if lpath.String() != "" {
-		panic("loras are not yet implemented")
+		return errors.New("loras are not yet implemented")
 	}
 
 	s.cache, err = NewInputCache(s.model, kvCacheType, int32(kvSize), parallel, s.batchSize, multiUserCache)
 	if err != nil {
-		panic(err)
+		return err
 	}
 
 	if !s.cache.enabled && parallel > 1 {
@@ -796,7 +862,30 @@ func (s *Server) loadModel(
 	s.seqs = make([]*Sequence, s.parallel)
 	s.seqsSem = semaphore.NewWeighted(int64(s.parallel))
 
-	err = s.reserveWorstCaseGraph()
+	return s.reserveWorstCaseGraph()
+}
+
+func (s *Server) load(
+	ctx context.Context,
+	mpath string,
+	params ml.BackendParams,
+	lpath multiLPath,
+	parallel int,
+	kvCacheType string,
+	kvSize int,
+	multiUserCache bool,
+) {
+	err := s.initModel(mpath, params, lpath, parallel, kvCacheType, kvSize, multiUserCache)
+	if err != nil {
+		panic(err)
+	}
+
+	slog.Debug("memory", "allocated", s.model.Backend().BackendMemory())
+
+	err = s.model.Backend().Load(ctx,
+		func(progress float32) {
+			s.progress = progress
+		})
 	if err != nil {
 		panic(err)
 	}
@@ -840,9 +929,14 @@ func Execute(args []string) error {
 		status:    llm.ServerStatusLoadingModel,
 	}
 
+	server.cond = sync.NewCond(&server.mu)
+	server.ready.Add(1)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
 	// TODO(jessegross): Parameters that need to be implemented:
 	//	no-mmap
-	//	mlock
 
 	var tensorSplitFloats []float32
 	if *tensorSplit != "" {
@@ -855,9 +949,6 @@ func Execute(args []string) error {
 	}
 
 	params := ml.BackendParams{
-		Progress: func(progress float32) {
-			server.progress = progress
-		},
 		NumThreads:     *threads,
 		NumGPULayers:   *numGPULayers,
 		MainGPU:        *mainGPU,
@@ -865,14 +956,7 @@ func Execute(args []string) error {
 		FlashAttention: *flashAttention,
 	}
 
-	server.ready.Add(1)
-	ctx, cancel := context.WithCancel(context.Background())
-	defer cancel()
-
-	go server.loadModel(ctx, *mpath, params, lpaths, *parallel, *kvCacheType, *kvSize, *multiUserCache)
-
-	server.cond = sync.NewCond(&server.mu)
-
+	go server.load(ctx, *mpath, params, lpaths, *parallel, *kvCacheType, *kvSize, *multiUserCache)
 	go server.run(ctx)
 
 	addr := "127.0.0.1:" + strconv.Itoa(*port)

runner/ollamarunner/runner_test.go

@@ -1,218 +0,0 @@
-package ollamarunner
-
-import (
-	"context"
-	"sync"
-	"testing"
-
-	"github.com/ollama/ollama/fs"
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/model"
-	"github.com/ollama/ollama/model/input"
-	"github.com/ollama/ollama/sample"
-	"golang.org/x/sync/semaphore"
-)
-
-// testBackend implements ml.Backend with minimal functionality required for tests.
-type testBackend struct{}
-
-func (b *testBackend) Config() fs.Config             { return testConfig{} }
-func (b *testBackend) Get(string) ml.Tensor          { return nil }
-func (b *testBackend) NewContext() ml.Context        { return &testContext{} }
-func (b *testBackend) NewContextSize(int) ml.Context { return &testContext{} }
-
-// testConfig is a stub implementation of fs.Config used by testBackend.
-type testConfig struct{}
-
-func (testConfig) Architecture() string                  { return "" }
-func (testConfig) String(string, ...string) string       { return "" }
-func (testConfig) Uint(string, ...uint32) uint32         { return 0 }
-func (testConfig) Float(string, ...float32) float32      { return 0 }
-func (testConfig) Bool(string, ...bool) bool             { return false }
-func (testConfig) Strings(string, ...[]string) []string  { return nil }
-func (testConfig) Ints(string, ...[]int32) []int32       { return nil }
-func (testConfig) Floats(string, ...[]float32) []float32 { return nil }
-
-type testContext struct{}
-
-func (c *testContext) Empty(dtype ml.DType, shape ...int) ml.Tensor {
-	sz := 1
-	for _, s := range shape {
-		sz *= s
-	}
-	return &testTensor{dtype: dtype, data: make([]float32, sz), shape: shape}
-}
-func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor { return c.Empty(dtype, shape...) }
-func (c *testContext) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
-	t := c.Empty(ml.DTypeF32, shape...).(*testTensor)
-	copy(t.data, s)
-	return t, nil
-}
-func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
-	f := make([]float32, len(s))
-	for i, v := range s {
-		f[i] = float32(v)
-	}
-	out, _ := c.FromFloatSlice(f, shape...)
-	out.(*testTensor).dtype = ml.DTypeI32
-	return out, nil
-}
-func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
-	return c.Empty(dtype, int((stop-start)/step))
-}
-func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }
-func (c *testContext) Compute(...ml.Tensor)            {}
-func (c *testContext) Reserve() error                  { return nil }
-func (c *testContext) MaxGraphNodes() int              { return 0 }
-func (c *testContext) Close()                          {}
-func (c *testContext) Input() ml.Context               { return c }
-func (c *testContext) Layer(int) ml.Context            { return c }
-
-type testTensor struct {
-	ml.Tensor
-	dtype ml.DType
-	data  []float32
-	shape []int
-}
-
-func (t *testTensor) Dim(n int) int    { return t.shape[n] }
-func (t *testTensor) Stride(n int) int { return 0 }
-func (t *testTensor) Shape() []int     { return t.shape }
-func (t *testTensor) DType() ml.DType  { return t.dtype }
-func (t *testTensor) Bytes() []byte    { return nil }
-func (t *testTensor) Floats() []float32 {
-	out := make([]float32, len(t.data))
-	copy(out, t.data)
-	return out
-}
-func (t *testTensor) Neg(ctx ml.Context) ml.Tensor { return nil }
-func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	out, _ := ctx.(*testContext).FromFloatSlice(nil, len(t.data))
-	return out
-}
-func (t *testTensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor            { return nil }
-func (t *testTensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor         { return nil }
-func (t *testTensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor { return nil }
-func (t *testTensor) MulmatID(ctx ml.Context, t2, ids ml.Tensor) ml.Tensor  { return nil }
-func (t *testTensor) Softmax(ctx ml.Context) ml.Tensor                      { return nil }
-func (t *testTensor) LayerNorm(ctx ml.Context, w, b ml.Tensor, e float32) ml.Tensor {
-	return nil
-}
-func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
-	return ctx.(*testContext).Empty(t.dtype, shape...)
-}
-func (t *testTensor) Copy(ctx ml.Context, dest ml.Tensor) ml.Tensor {
-	copy(dest.(*testTensor).data, t.data)
-	return nil
-}
-
-// fakeModel implements model.Model and model.TextProcessor.
-type fakeModel struct {
-	model.Base
-	decode  map[int32]string
-	logits  [][]float32
-	call    int
-	backend ml.Backend
-}
-
-func (f *fakeModel) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	idx := f.call
-	if idx >= len(f.logits) {
-		idx = len(f.logits) - 1
-	}
-	f.call++
-	return ctx.FromFloatSlice(f.logits[idx], len(f.logits[idx]))
-}
-
-func (f *fakeModel) Backend() ml.Backend {
-	if f.backend == nil {
-		f.backend = &testBackend{}
-	}
-	return f.backend
-}
-
-func (f *fakeModel) Encode(string, bool) ([]int32, error) { return nil, nil }
-func (f *fakeModel) Decode(ids []int32) (string, error) {
-	var s string
-	for _, id := range ids {
-		s += f.decode[id]
-	}
-	return s, nil
-}
-func (f *fakeModel) Is(id int32, sp model.Special) bool { return false }
-func (f *fakeModel) Vocabulary() *model.Vocabulary      { return &model.Vocabulary{} }
-
-var _ model.Model = (*fakeModel)(nil)
-var _ model.TextProcessor = (*fakeModel)(nil)
-
-func TestProcessBatchUnicode(t *testing.T) {
-	tests := []struct {
-		name   string
-		decode map[int32]string
-		logits [][]float32
-		want   string
-	}{
-		{
-			name:   "emoji",
-			decode: map[int32]string{0: "A", 1: "😀", 2: "👍", 3: "!"},
-			logits: [][]float32{{10, 0, 0, 0}, {0, 10, 0, 0}, {0, 0, 10, 0}, {0, 0, 0, 10}},
-			want:   "A😀👍!",
-		},
-		{
-			name:   "ascii",
-			decode: map[int32]string{0: "H", 1: "e", 2: "y"},
-			logits: [][]float32{{10, 0, 0}, {0, 10, 0}, {0, 0, 10}},
-			want:   "Hey",
-		},
-		{
-			name:   "multibyte",
-			decode: map[int32]string{0: "世", 1: "界", 2: "😊"},
-			logits: [][]float32{{10, 0, 0}, {0, 10, 0}, {0, 0, 10}},
-			want:   "世界😊",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			m := &fakeModel{decode: tt.decode, logits: tt.logits}
-
-			s := &Server{model: m, batchSize: 1, parallel: 1}
-			s.cache = &InputCache{enabled: true, slots: []InputCacheSlot{{Id: 0}}, numCtx: 10}
-			s.seqs = make([]*Sequence, 1)
-			s.seqsSem = semaphore.NewWeighted(1)
-			if err := s.seqsSem.Acquire(context.Background(), 1); err != nil {
-				t.Fatal(err)
-			}
-			s.cond = sync.NewCond(&s.mu)
-
-			seq := &Sequence{
-				inputs:     []input.Input{{Token: 0}},
-				cache:      &s.cache.slots[0],
-				responses:  make(chan string, 10),
-				quit:       make(chan bool, 1),
-				numPredict: len(tt.logits),
-				sampler:    sample.NewSampler(0, 0, 0, 0, 0, nil),
-				embedding:  make(chan []float32, 1),
-			}
-			s.seqs[0] = seq
-
-			for {
-				if err := s.processBatch(); err != nil {
-					t.Fatal(err)
-				}
-				if s.seqs[0] == nil {
-					break
-				}
-			}
-
-			var result string
-			for r := range seq.responses {
-				result += r
-			}
-
-			if result != tt.want {
-				t.Fatalf("got %q want %q", result, tt.want)
-			}
-		})
-	}
-}

sample/samplers.go

@@ -176,7 +176,7 @@ func NewGrammarSampler(model model.TextProcessor, grammarStr string) (*GrammarSa
 		vocabIds[i] = uint32(i)
 	}
 
-	grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, []uint32{uint32(model.Vocabulary().EOS), uint32(model.Vocabulary().EOT)})
+	grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, model.Vocabulary().EOS)
 	if grammar == nil {
 		return nil, errors.New("sample: failed to initialize grammar")
 	}

server/create.go

@@ -295,7 +295,7 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 	}
 	defer bin.Close()
 
-	f, _, err := ggml.Decode(bin, -1)
+	f, err := ggml.Decode(bin, -1)
 	if err != nil {
 		return nil, err
 	}
@@ -430,7 +430,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 	fnWrap := func(n uint64) {
 		done := doneBytes.Add(n)
 		progress := float32(done) / float32(totalBytes)
-		fn(api.ProgressResponse{Status: fmt.Sprintf("quantizing %s model to %s", ft, quantizeType), Digest: "0", Total: layer.Size, Completed: int64(progress * float32(layer.Size))})
+		fn(api.ProgressResponse{Status: fmt.Sprintf("quantizing %s model to %s", ft, quantizeType), Digest: "0000000000000000000", Total: layer.Size, Completed: int64(progress * float32(layer.Size))})
 	}
 	ftype, err := ggml.ParseFileType(quantizeType)
 	if err != nil {
@@ -467,7 +467,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 		return nil, err
 	}
 
-	f, _, err := ggml.Decode(temp, 1024)
+	f, err := ggml.Decode(temp, 1024)
 	if err != nil {
 		slog.Error(fmt.Sprintf("error decoding ggml: %s\n", err))
 		return nil, err
@@ -501,48 +501,27 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 		return nil, errOnlyGGUFSupported
 	}
 
-	stat, err := blob.Stat()
+	f, err := ggml.Decode(blob, -1)
 	if err != nil {
 		return nil, err
 	}
 
-	var offset int64
-	for offset < stat.Size() {
-		f, n, err := ggml.Decode(blob, 1024)
-		if errors.Is(err, io.EOF) {
-			break
-		} else if err != nil {
-			return nil, err
-		}
-
-		mediatype := "application/vnd.ollama.image.model"
-		if f.KV().Kind() == "adapter" {
-			mediatype = "application/vnd.ollama.image.adapter"
-		} else if _, ok := f.KV()[fmt.Sprintf("%s.vision.block_count", f.KV().Architecture())]; ok || f.KV().Kind() == "projector" {
-			mediatype = "application/vnd.ollama.image.projector"
-		}
-
-		var layer Layer
-		if digest != "" && n == stat.Size() && offset == 0 {
-			layer, err = NewLayerFromLayer(digest, mediatype, blob.Name())
-			if err != nil {
-				slog.Debug("could not create new layer from layer", "error", err)
-				return nil, err
-			}
-		}
-
-		// Fallback to creating layer from file copy (either NewLayerFromLayer failed, or digest empty/n != stat.Size())
-		if layer.Digest == "" {
-			layer, err = NewLayer(io.NewSectionReader(blob, offset, n), mediatype)
-			if err != nil {
-				return nil, err
-			}
-		}
-
-		layers = append(layers, &layerGGML{layer, f})
-		offset = n
+	mediatype := "application/vnd.ollama.image.model"
+	if f.KV().Kind() == "adapter" {
+		mediatype = "application/vnd.ollama.image.adapter"
+	} else if (f.KV().Uint("block_count") == 0 && f.KV().Uint("vision.block_count") > 0) || f.KV().Kind() == "projector" {
+		// if a model has vision.block_count but not block_count, it is a standalone vision model
+		mediatype = "application/vnd.ollama.image.projector"
 	}
 
+	layer, err := NewLayerFromLayer(digest, mediatype, blob.Name())
+	if err != nil {
+		slog.Debug("could not create new layer from layer", "error", err)
+		return nil, err
+	}
+
+	layers = append(layers, &layerGGML{layer, f})
+
 	return detectChatTemplate(layers)
 }
 

server/download.go

@@ -464,6 +464,10 @@ type downloadOpts struct {
 
 // downloadBlob downloads a blob from the registry and stores it in the blobs directory
 func downloadBlob(ctx context.Context, opts downloadOpts) (cacheHit bool, _ error) {
+	if opts.digest == "" {
+		return false, fmt.Errorf(("%s: %s"), opts.mp.GetNamespaceRepository(), "digest is is empty")
+	}
+
 	fp, err := GetBlobsPath(opts.digest)
 	if err != nil {
 		return false, err

server/images.go

@@ -26,6 +26,7 @@ import (
 	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/template"
+	"github.com/ollama/ollama/thinking"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
@@ -37,6 +38,7 @@ var (
 	errCapabilityInsert     = errors.New("insert")
 	errCapabilityVision     = errors.New("vision")
 	errCapabilityEmbedding  = errors.New("embedding")
+	errCapabilityThinking   = errors.New("thinking")
 	errInsecureProtocol     = errors.New("insecure protocol http")
 )
 
@@ -75,7 +77,7 @@ func (m *Model) Capabilities() []model.Capability {
 	if err == nil {
 		defer r.Close()
 
-		f, _, err := ggml.Decode(r, 1024)
+		f, err := ggml.Decode(r, 1024)
 		if err == nil {
 			if _, ok := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]; ok {
 				capabilities = append(capabilities, model.CapabilityEmbedding)
@@ -111,6 +113,12 @@ func (m *Model) Capabilities() []model.Capability {
 		capabilities = append(capabilities, model.CapabilityVision)
 	}
 
+	// Check for thinking capability
+	openingTag, closingTag := thinking.InferTags(m.Template.Template)
+	if openingTag != "" && closingTag != "" {
+		capabilities = append(capabilities, model.CapabilityThinking)
+	}
+
 	return capabilities
 }
 
@@ -127,6 +135,7 @@ func (m *Model) CheckCapabilities(want ...model.Capability) error {
 		model.CapabilityInsert:     errCapabilityInsert,
 		model.CapabilityVision:     errCapabilityVision,
 		model.CapabilityEmbedding:  errCapabilityEmbedding,
+		model.CapabilityThinking:   errCapabilityThinking,
 	}
 
 	for _, cap := range want {
@@ -141,11 +150,19 @@ func (m *Model) CheckCapabilities(want ...model.Capability) error {
 		}
 	}
 
+	var err error
 	if len(errs) > 0 {
-		return fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
+		err = fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
 	}
 
-	return nil
+	if slices.Contains(errs, errCapabilityThinking) {
+		if m.Config.ModelFamily == "qwen3" || model.ParseName(m.Name).Model == "deepseek-r1" {
+			// append a message to the existing error
+			return fmt.Errorf("%w. Pull the model again to get the latest version with full thinking support", err)
+		}
+	}
+
+	return err
 }
 
 func (m *Model) String() string {

server/model.go

@@ -10,9 +10,6 @@ import (
 	"log/slog"
 	"net/http"
 	"os"
-	"slices"
-	"strings"
-	"text/template/parse"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/fs/ggml"
@@ -64,7 +61,7 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 			}
 			defer blob.Close()
 
-			f, _, err := ggml.Decode(blob, -1)
+			f, err := ggml.Decode(blob, -1)
 			if err != nil {
 				return nil, err
 			}
@@ -128,124 +125,3 @@ func detectContentType(r io.Reader) (string, error) {
 
 	return "unknown", nil
 }
-
-func parseObjects(s string) []map[string]any {
-	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 {
-			return nil
-		} else {
-			offset += int(decoder.InputOffset())
-			objs = append(objs, obj)
-		}
-	}
-
-	return objs
-}
-
-// 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
-	}
-
-	templateObjects := parseObjects(b.String())
-	if len(templateObjects) == 0 {
-		return nil, false
-	}
-
-	// find the keys that correspond to the name and arguments fields
-	var name, arguments string
-	for k, v := range templateObjects[0] {
-		switch v.(type) {
-		case string:
-			name = k
-		case map[string]any:
-			arguments = k
-		}
-	}
-
-	if name == "" || arguments == "" {
-		return nil, false
-	}
-
-	responseObjects := parseObjects(s)
-	if len(responseObjects) == 0 {
-		return nil, false
-	}
-
-	// collect all nested objects
-	var collect func(any) []map[string]any
-	collect = func(obj any) (all []map[string]any) {
-		switch o := obj.(type) {
-		case map[string]any:
-			all = append(all, o)
-			for _, v := range o {
-				all = append(all, collect(v)...)
-			}
-		case []any:
-			for _, v := range o {
-				all = append(all, collect(v)...)
-			}
-		}
-
-		return all
-	}
-
-	var objs []map[string]any
-	for _, p := range responseObjects {
-		objs = append(objs, collect(p)...)
-	}
-
-	var toolCalls []api.ToolCall
-	for _, kv := range objs {
-		n, nok := kv[name].(string)
-		a, aok := kv[arguments].(map[string]any)
-		if nok && aok {
-			toolCalls = append(toolCalls, api.ToolCall{
-				Function: api.ToolCallFunction{
-					Name:      n,
-					Arguments: a,
-				},
-			})
-		}
-	}
-
-	return toolCalls, len(toolCalls) > 0
-}

server/model_test.go

@@ -1,179 +0,0 @@
-package server
-
-import (
-	"bytes"
-	"encoding/json"
-	"fmt"
-	"os"
-	"path/filepath"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/template"
-)
-
-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", `[TOOL_CALLS]  [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"To }]`, false},
-		{"mistral", `I'm not aware of that information. However, I can suggest searching for the weather using the "get_current_weather" function:
-
-		[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
-		{"mistral", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
-		{"command-r-plus", "Action: ```json" + `
-[
-    {
-        "tool_name": "get_current_weather",
-        "parameters": {
-            "format": "fahrenheit",
-            "location": "San Francisco, CA"
-        }
-    },
-    {
-        "tool_name": "get_current_weather",
-        "parameters": {
-            "format": "celsius",
-            "location": "Toronto, Canada"
-        }
-    }
-]
-` + "```", true},
-		{"command-r-plus", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
-		{"firefunction", ` functools[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
-		{"firefunction", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
-		{"llama3-groq-tool-use", `<tool_call>
-{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}}
-{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}
-</tool_call>`, true},
-		{"xlam", `{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]}`, true},
-		{"nemotron", `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]} </toolcall>`, 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)
-					}
-				}
-			})
-		})
-	}
-}
-
-func TestParseObjects(t *testing.T) {
-	tests := []struct {
-		input string
-		want  []map[string]any
-	}{
-		{
-			input: `[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`,
-			want: []map[string]any{
-				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
-				{"name": "get_current_weather", "arguments": map[string]any{"format": "celsius", "location": "Toronto, Canada"}},
-			},
-		},
-		{
-			input: `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}} </toolcall>`,
-			want: []map[string]any{
-				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
-			},
-		},
-		{
-			input: `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}} </toolcall> <toolcall>{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, ON"}} </toolcall>`,
-			want: []map[string]any{
-				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
-				{"name": "get_current_weather", "arguments": map[string]any{"format": "celsius", "location": "Toronto, ON"}},
-			},
-		},
-		{
-			input: `{"name": "get_current_weather", "arguments": `,
-			want:  nil,
-		},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.input, func(t *testing.T) {
-			got := parseObjects(tc.input)
-
-			if diff := cmp.Diff(got, tc.want); diff != "" {
-				t.Errorf("mismatch (-got +want):\n%s", diff)
-			}
-		})
-	}
-}

server/modelpath.go

@@ -116,7 +116,7 @@ func (mp ModelPath) BaseURL() *url.URL {
 func GetManifestPath() (string, error) {
 	path := filepath.Join(envconfig.Models(), "manifests")
 	if err := os.MkdirAll(path, 0o755); err != nil {
-		return "", err
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
 	}
 
 	return path, nil
@@ -139,7 +139,7 @@ func GetBlobsPath(digest string) (string, error) {
 	}
 
 	if err := os.MkdirAll(dirPath, 0o755); err != nil {
-		return "", err
+		return "", fmt.Errorf("%w: ensure path elements are traversable", err)
 	}
 
 	return path, nil

server/prompt.go

@@ -19,7 +19,7 @@ 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) {
+func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool, think *bool) (prompt string, images []llm.ImageData, _ error) {
 	var system []api.Message
 
 	// TODO: Ideally we would compute this from the projector metadata but some pieces are implementation dependent
@@ -41,8 +41,12 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 			}
 		}
 
+		thinkVal := false
+		if think != nil {
+			thinkVal = *think
+		}
 		var b bytes.Buffer
-		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools}); err != nil {
+		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools, Think: thinkVal, IsThinkSet: think != nil}); err != nil {
 			return "", nil, err
 		}
 
@@ -96,7 +100,11 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 
 	// 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[currMsgIdx:]...), Tools: tools}); err != nil {
+	thinkVal := false
+	if think != nil {
+		thinkVal = *think
+	}
+	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[currMsgIdx:]...), Tools: tools, Think: thinkVal, IsThinkSet: think != nil}); err != nil {
 		return "", nil, err
 	}
 

server/prompt_test.go

@@ -208,7 +208,8 @@ func TestChatPrompt(t *testing.T) {
 		t.Run(tt.name, func(t *testing.T) {
 			model := tt.model
 			opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
-			prompt, images, err := chatPrompt(t.Context(), &model, mockRunner{}.Tokenize, &opts, tt.msgs, nil)
+			think := false
+			prompt, images, err := chatPrompt(t.Context(), &model, mockRunner{}.Tokenize, &opts, tt.msgs, nil, &think)
 			if tt.error == nil && err != nil {
 				t.Fatal(err)
 			} else if tt.error != nil && err != tt.error {

server/quantization.go

@@ -120,14 +120,30 @@ func getTensorNewType(kv fsggml.KV, qs *quantizeState, newType fsggml.TensorType
 
 	if newType.IsQuantized() {
 		nx := shape[0]
-		ny := uint64(1)
-		if len(shape) > 1 {
-			ny = shape[1]
-		}
 		qk_k := newType.BlockSize()
+
+		// Check if first dimension is divisible by block size
 		if nx%qk_k != 0 {
-			slog.Warn(fmt.Sprintf("tensor cols %d x %d are not divisible by %d, required for %s.  Falling back to quantization %s", nx, ny, qk_k, newType.String(), fsggml.TensorTypeF16.String()))
-			newType = fsggml.TensorTypeF16
+			// Store the original type for logging
+			originalType := newType
+
+			// Select appropriate fallback based on original type
+			switch newType {
+			case fsggml.TensorTypeQ4_K:
+				newType = fsggml.TensorTypeQ5_0
+			case fsggml.TensorTypeQ5_K:
+				newType = fsggml.TensorTypeQ5_1
+			case fsggml.TensorTypeQ6_K:
+				newType = fsggml.TensorTypeQ8_0
+			}
+
+			// Final check - if still incompatible, fall back to F16
+			if nx%newType.BlockSize() != 0 {
+				newType = fsggml.TensorTypeF16
+			}
+
+			slog.Warn(fmt.Sprintf("tensor cols %d are not divisible by %d, required for %s - using fallback quantization %s",
+				nx, qk_k, originalType.String(), newType.String()))
 		}
 	}
 	return newType

server/quantization_test.go

@@ -271,7 +271,7 @@ func TestQuantizeModel(t *testing.T) {
 				t.Fatal(err.Error())
 			}
 			defer fp.Close()
-			meta, _, err := fsggml.Decode(fp, -1)
+			meta, err := fsggml.Decode(fp, -1)
 			if err != nil {
 				t.Fatal(err.Error())
 			}
@@ -303,7 +303,7 @@ func TestQuantizeModel(t *testing.T) {
 				t.Fatalf("failed to load the quantized model %s: %s", tmp.Name(), err)
 			}
 			defer fpNew.Close()
-			newMeta, _, err := fsggml.Decode(fpNew, -1)
+			newMeta, err := fsggml.Decode(fpNew, -1)
 			if err != nil {
 				t.Fatalf("failed to load the quantized model %s: %s", tmp.Name(), err)
 			}

server/routes.go

@@ -17,7 +17,6 @@ import (
 	"net/netip"
 	"os"
 	"os/signal"
-	"regexp"
 	"slices"
 	"strings"
 	"syscall"
@@ -38,6 +37,8 @@ import (
 	"github.com/ollama/ollama/server/internal/client/ollama"
 	"github.com/ollama/ollama/server/internal/registry"
 	"github.com/ollama/ollama/template"
+	"github.com/ollama/ollama/thinking"
+	"github.com/ollama/ollama/tools"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -185,6 +186,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 	if req.Suffix != "" {
 		caps = append(caps, model.CapabilityInsert)
 	}
+	if req.Think != nil && *req.Think {
+		caps = append(caps, model.CapabilityThinking)
+		// TODO(drifkin): consider adding a warning if it's false and the model
+		// doesn't support thinking. It's not strictly required, but it can be a
+		// hint that the user is on an older qwen3/r1 model that doesn't have an
+		// updated template supporting thinking
+	}
 
 	r, m, opts, err := s.scheduleRunner(c.Request.Context(), name.String(), caps, req.Options, req.KeepAlive)
 	if errors.Is(err, errCapabilityCompletion) {
@@ -253,6 +261,9 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			values.Messages = append(msgs, api.Message{Role: "user", Content: req.Prompt})
 		}
 
+		values.Think = req.Think != nil && *req.Think
+		values.IsThinkSet = req.Think != nil
+
 		var b bytes.Buffer
 		if req.Context != nil {
 			slog.Warn("the context field is deprecated and will be removed in a future version of Ollama")
@@ -272,6 +283,15 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		prompt = b.String()
 	}
 
+	var thinkingState *thinking.Parser
+	openingTag, closingTag := thinking.InferTags(m.Template.Template)
+	if req.Think != nil && *req.Think && openingTag != "" && closingTag != "" {
+		thinkingState = &thinking.Parser{
+			OpeningTag: openingTag,
+			ClosingTag: closingTag,
+		}
+	}
+
 	ch := make(chan any)
 	go func() {
 		// TODO (jmorganca): avoid building the response twice both here and below
@@ -296,6 +316,12 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 				},
 			}
 
+			if thinkingState != nil {
+				thinking, content := thinkingState.AddContent(cr.Content)
+				res.Thinking = thinking
+				res.Response = content
+			}
+
 			if _, err := sb.WriteString(cr.Content); err != nil {
 				ch <- gin.H{"error": err.Error()}
 			}
@@ -323,11 +349,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	if req.Stream != nil && !*req.Stream {
 		var r api.GenerateResponse
-		var sb strings.Builder
+		var sbThinking strings.Builder
+		var sbContent strings.Builder
 		for rr := range ch {
 			switch t := rr.(type) {
 			case api.GenerateResponse:
-				sb.WriteString(t.Response)
+				sbThinking.WriteString(t.Thinking)
+				sbContent.WriteString(t.Response)
 				r = t
 			case gin.H:
 				msg, ok := t["error"].(string)
@@ -343,7 +371,9 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			}
 		}
 
-		r.Response = sb.String()
+		r.Thinking = sbThinking.String()
+		r.Response = sbContent.String()
+
 		c.JSON(http.StatusOK, r)
 		return
 	}
@@ -899,8 +929,7 @@ func (s *Server) ListHandler(c *gin.Context) {
 			}
 		}
 
-		// tag should never be masked
-		models = append(models, api.ListModelResponse{
+		r := api.ListModelResponse{
 			Model:      n.DisplayShortest(),
 			Name:       n.DisplayShortest(),
 			Size:       m.Size(),
@@ -913,7 +942,16 @@ func (s *Server) ListHandler(c *gin.Context) {
 				ParameterSize:     cf.ModelType,
 				QuantizationLevel: cf.FileType,
 			},
-		})
+		}
+
+		model, err := GetModel(n.String())
+		if err != nil {
+			slog.Warn("bad model details", "name", n, "error", err)
+		} else {
+			r.Capabilities = model.Capabilities()
+		}
+
+		models = append(models, r)
 	}
 
 	slices.SortStableFunc(models, func(i, j api.ListModelResponse) int {
@@ -1435,6 +1473,9 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	if len(req.Tools) > 0 {
 		caps = append(caps, model.CapabilityTools)
 	}
+	if req.Think != nil && *req.Think {
+		caps = append(caps, model.CapabilityThinking)
+	}
 
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
@@ -1475,18 +1516,36 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 	msgs = filterThinkTags(msgs, m)
 
-	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, req.Tools)
+	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, req.Tools, req.Think)
 	if err != nil {
 		slog.Error("chat prompt error", "error", err)
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
 
+	var thinkingState *thinking.Parser
+	openingTag, closingTag := thinking.InferTags(m.Template.Template)
+	if req.Think != nil && *req.Think && openingTag != "" && closingTag != "" {
+		thinkingState = &thinking.Parser{
+			OpeningTag: openingTag,
+			ClosingTag: closingTag,
+		}
+	}
+
+	var toolParser *tools.Parser
+	if len(req.Tools) > 0 {
+		toolParser, err = tools.NewParser(m.Template.Template)
+		if err != nil {
+			slog.Error("failed to create tool parser", "error", err)
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+	}
+
 	ch := make(chan any)
 	go func() {
 		defer close(ch)
-		var sb strings.Builder
-		var toolCallIndex int = 0
+
 		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
 			Prompt:  prompt,
 			Images:  images,
@@ -1506,43 +1565,40 @@ func (s *Server) ChatHandler(c *gin.Context) {
 				},
 			}
 
+			if thinkingState != nil {
+				thinkingContent, remainingContent := thinkingState.AddContent(res.Message.Content)
+				if thinkingContent == "" && remainingContent == "" && !r.Done {
+					// need to accumulate more to decide what to send
+					return
+				}
+				res.Message.Content = remainingContent
+				res.Message.Thinking = thinkingContent
+			}
+
 			if r.Done {
 				res.DoneReason = r.DoneReason.String()
 				res.TotalDuration = time.Since(checkpointStart)
 				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
 			}
 
-			// TODO: tool call checking and filtering should be moved outside of this callback once streaming
-			// however this was a simple change for now without reworking streaming logic of this (and other)
-			// handlers
-			if req.Stream != nil && !*req.Stream || len(req.Tools) == 0 {
-				ch <- res
-				return
+			if len(req.Tools) > 0 {
+				toolCalls, content := toolParser.Add(res.Message.Content)
+				if len(content) > 0 {
+					res.Message.Content = content
+				} else if len(toolCalls) > 0 {
+					res.Message.ToolCalls = toolCalls
+					res.Message.Content = ""
+				} else if res.Message.Thinking != "" {
+					// don't return
+				} else {
+					if r.Done {
+						ch <- res
+					}
+					return
+				}
 			}
 
-			// Streaming tool calls:
-			// If tools are recognized, use a flag to track the sending of a tool downstream
-			// This ensures that content is cleared from the message on the last chunk sent
-			sb.WriteString(r.Content)
-			if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
-				res.Message.ToolCalls = toolCalls
-				for i := range toolCalls {
-					toolCalls[i].Function.Index = toolCallIndex
-					toolCallIndex++
-				}
-				res.Message.Content = ""
-				sb.Reset()
-				ch <- res
-				return
-			}
-
-			if r.Done {
-				// Send any remaining content if no tool calls were detected
-				if toolCallIndex == 0 {
-					res.Message.Content = sb.String()
-				}
-				ch <- res
-			}
+			ch <- res
 		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
@@ -1550,12 +1606,18 @@ func (s *Server) ChatHandler(c *gin.Context) {
 
 	if req.Stream != nil && !*req.Stream {
 		var resp api.ChatResponse
-		var sb strings.Builder
+		var toolCalls []api.ToolCall
+		var sbThinking strings.Builder
+		var sbContent strings.Builder
 		for rr := range ch {
 			switch t := rr.(type) {
 			case api.ChatResponse:
-				sb.WriteString(t.Message.Content)
+				sbThinking.WriteString(t.Message.Thinking)
+				sbContent.WriteString(t.Message.Content)
 				resp = t
+				if len(req.Tools) > 0 {
+					toolCalls = append(toolCalls, t.Message.ToolCalls...)
+				}
 			case gin.H:
 				msg, ok := t["error"].(string)
 				if !ok {
@@ -1570,13 +1632,11 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 		}
 
-		resp.Message.Content = sb.String()
+		resp.Message.Content = sbContent.String()
+		resp.Message.Thinking = sbThinking.String()
 
-		if len(req.Tools) > 0 {
-			if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
-				resp.Message.ToolCalls = toolCalls
-				resp.Message.Content = ""
-			}
+		if len(toolCalls) > 0 {
+			resp.Message.ToolCalls = toolCalls
 		}
 
 		c.JSON(http.StatusOK, resp)
@@ -1601,8 +1661,6 @@ func handleScheduleError(c *gin.Context, name string, err error) {
 	}
 }
 
-var thinkTagRegexp = regexp.MustCompile(`<think>(?s).*?</think>(\n)*`)
-
 func filterThinkTags(msgs []api.Message, m *Model) []api.Message {
 	if m.Config.ModelFamily == "qwen3" || model.ParseName(m.Name).Model == "deepseek-r1" {
 		finalUserIndex := -1
@@ -1614,7 +1672,17 @@ func filterThinkTags(msgs []api.Message, m *Model) []api.Message {
 
 		for i, msg := range msgs {
 			if msg.Role == "assistant" && i < finalUserIndex {
-				msgs[i].Content = thinkTagRegexp.ReplaceAllString(msg.Content, "")
+				// TODO(drifkin): this is from before we added proper thinking support.
+				// However, even if thinking is not enabled (and therefore we shouldn't
+				// change the user output), we should probably perform this filtering
+				// for all thinking models (not just qwen3 & deepseek-r1) since it tends
+				// to save tokens and improve quality.
+				thinkingState := &thinking.Parser{
+					OpeningTag: "<think>",
+					ClosingTag: "</think>",
+				}
+				_, content := thinkingState.AddContent(msg.Content)
+				msgs[i].Content = content
 			}
 		}
 	}

server/routes_create_test.go

@@ -24,7 +24,7 @@ import (
 
 var stream bool = false
 
-func createBinFile(t *testing.T, kv map[string]any, ti []*ggml.Tensor) (string, string) {
+func createBinFile(t testing.TB, kv map[string]any, ti []*ggml.Tensor) (string, string) {
 	t.Helper()
 	t.Setenv("OLLAMA_MODELS", cmp.Or(os.Getenv("OLLAMA_MODELS"), t.TempDir()))
 
@@ -71,7 +71,7 @@ func (t *responseRecorder) CloseNotify() <-chan bool {
 	return make(chan bool)
 }
 
-func createRequest(t *testing.T, fn func(*gin.Context), body any) *httptest.ResponseRecorder {
+func createRequest(t testing.TB, fn func(*gin.Context), body any) *httptest.ResponseRecorder {
 	t.Helper()
 	// if OLLAMA_MODELS is not set, set it to the temp directory
 	t.Setenv("OLLAMA_MODELS", cmp.Or(os.Getenv("OLLAMA_MODELS"), t.TempDir()))

server/routes_generate_test.go

@@ -143,6 +143,25 @@ func TestGenerateChat(t *testing.T) {
 		}
 	})
 
+	t.Run("missing thinking capability", func(t *testing.T) {
+		think := true
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test",
+			Messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			Think: &think,
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"registry.ollama.ai/library/test:latest does not support thinking"}`); 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 {

server/routes_list_test.go

@@ -2,6 +2,7 @@ package server
 
 import (
 	"encoding/json"
+	"fmt"
 	"net/http"
 	"slices"
 	"testing"
@@ -64,3 +65,65 @@ func TestList(t *testing.T) {
 		t.Fatalf("expected slices to be equal %v", actualNames)
 	}
 }
+
+func BenchmarkListHandler(b *testing.B) {
+	gin.SetMode(gin.TestMode)
+
+	// Test with higher model counts to simulate real-world scenarios
+	modelCounts := []int{50, 100, 250, 500, 1000, 2000}
+
+	for _, count := range modelCounts {
+		b.Run(fmt.Sprintf("models_%d", count), func(b *testing.B) {
+			benchmarkListWithModelCount(b, count)
+		})
+	}
+}
+
+func benchmarkListWithModelCount(b *testing.B, modelCount int) {
+	// Setup
+	tempDir := b.TempDir()
+	b.Setenv("OLLAMA_MODELS", tempDir)
+
+	var s Server
+
+	// Create the specified number of models
+	b.Logf("Creating %d models for benchmark...", modelCount)
+	for i := range modelCount {
+		modelName := fmt.Sprintf("testmodel%d:latest", i)
+		_, digest := createBinFile(b, nil, nil)
+
+		createRequest(b, s.CreateHandler, api.CreateRequest{
+			Name:  modelName,
+			Files: map[string]string{"test.gguf": digest},
+		})
+
+		// Log progress for large numbers
+		if modelCount >= 500 && i%100 == 0 {
+			b.Logf("Created %d/%d models", i, modelCount)
+		}
+	}
+
+	b.Logf("Setup complete, starting benchmark with %d models", modelCount)
+
+	// Reset timer to exclude setup time
+	b.ResetTimer()
+
+	// Run the actual benchmark
+	for i := 0; i < b.N; i++ {
+		w := createRequest(b, s.ListHandler, nil)
+		if w.Code != http.StatusOK {
+			b.Fatalf("expected status code 200, actual %d", w.Code)
+		}
+
+		// Optional: Verify we got the expected number of models
+		if i == 0 { // Only check on first iteration to avoid overhead
+			var resp api.ListResponse
+			if err := json.NewDecoder(w.Body).Decode(&resp); err != nil {
+				b.Fatal(err)
+			}
+			if len(resp.Models) != modelCount {
+				b.Fatalf("expected %d models, got %d", modelCount, len(resp.Models))
+			}
+		}
+	}
+}

server/sched.go

@@ -387,6 +387,17 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 				s.loadedMu.Unlock()
 				runner.refMu.Unlock()
 				slog.Debug("duplicate expired event, ignoring", "runner", runner)
+			} else if runner.pid != runnerToUnload.pid {
+				// If the pids do not match, we likely had multiple load
+				// failures for the same model in quick succession due to
+				// request context canceled and are draining the queue of
+				// events. Ensure the orphaned runner is properly shut down, but
+				// do not delete the mismatched loaded runner, or wait for VRAM
+				// convergence.
+				slog.Debug("orphaned runner shutting down", "orphan", runner, "loaded", runnerToUnload)
+				runner.unload()
+				s.loadedMu.Unlock()
+				runner.refMu.Unlock()
 			} else {
 				slog.Debug("starting background wait for VRAM recovery", "runner", runner)
 				finished := runner.waitForVRAMRecovery()

template/template.go

@@ -167,6 +167,10 @@ type Values struct {
 	api.Tools
 	Prompt string
 	Suffix string
+	Think  bool
+	// whether or not the user explicitly set the thinking flag (vs. it being
+	// implicitly false). Templates can't see whether `Think` is nil
+	IsThinkSet bool
 
 	// forceLegacy is a flag used to test compatibility with legacy templates
 	forceLegacy bool
@@ -222,16 +226,20 @@ 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": "",
+			"Prompt":     v.Prompt,
+			"Suffix":     v.Suffix,
+			"Response":   "",
+			"Think":      v.Think,
+			"IsThinkSet": v.IsThinkSet,
 		})
 	} 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":     system,
+			"Messages":   messages,
+			"Tools":      v.Tools,
+			"Response":   "",
+			"Think":      v.Think,
+			"IsThinkSet": v.IsThinkSet,
 		})
 	}
 
@@ -241,9 +249,11 @@ func (t *Template) Execute(w io.Writer, v Values) error {
 	for _, m := range messages {
 		execute := func() error {
 			if err := t.Template.Execute(&b, map[string]any{
-				"System":   system,
-				"Prompt":   prompt,
-				"Response": response,
+				"System":     system,
+				"Prompt":     prompt,
+				"Response":   response,
+				"Think":      v.Think,
+				"IsThinkSet": v.IsThinkSet,
 			}); err != nil {
 				return err
 			}
@@ -286,9 +296,11 @@ func (t *Template) Execute(w io.Writer, v Values) error {
 
 	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,
+		"System":     system,
+		"Prompt":     prompt,
+		"Response":   response,
+		"Think":      v.Think,
+		"IsThinkSet": v.IsThinkSet,
 	}); err != nil {
 		return err
 	}

thinking/parser.go

@@ -0,0 +1,171 @@
+package thinking
+
+import (
+	"strings"
+	"unicode"
+)
+
+type thinkingState int
+
+const (
+	// We're looking for the opening tag, but we haven't seen any non-whitespace
+	// characters yet
+	thinkingState_LookingForOpening thinkingState = iota
+	// We've seen the opening tag, but we haven't seen any non-whitespace
+	// characters yet (we want to eat any whitespace between the opening tag and
+	// the thinking content)
+	thinkingState_ThinkingStartedEatingWhitespace
+	// We've seen non-whitespace characters after the opening tag, but we haven't
+	// seen the closing tag yet
+	thinkingState_Thinking
+	// We've seen the closing tag, but we haven't seen any non-whitespace
+	// characters after the closing tag yet (we want to eat any whitespace between
+	// the closing tag and the content)
+	thinkingState_ThinkingDoneEatingWhitespace
+	// We've seen the closing tag and seen at least one non-whitespace character
+	// after it
+	thinkingState_ThinkingDone
+)
+
+func (s thinkingState) String() string {
+	switch s {
+	case thinkingState_LookingForOpening:
+		return "LookingForOpening"
+	case thinkingState_ThinkingStartedEatingWhitespace:
+		return "ThinkingStartedEatingWhitespace"
+	case thinkingState_Thinking:
+		return "Thinking"
+	case thinkingState_ThinkingDoneEatingWhitespace:
+		return "ThinkingDoneEatingWhitespace"
+	case thinkingState_ThinkingDone:
+		return "ThinkingDone"
+	default:
+		return "Unknown"
+	}
+}
+
+type Parser struct {
+	state      thinkingState
+	OpeningTag string
+	ClosingTag string
+	acc        strings.Builder
+}
+
+// AddContent returns the thinking content and the non-thinking content that
+// should be immediately sent to the user. It will internally buffer if it needs
+// to see more raw content to disambiguate
+func (s *Parser) AddContent(content string) (string, string) {
+	s.acc.WriteString(content)
+
+	var thinkingSb, remainingSb strings.Builder
+
+	var thinking, remaining string
+	keepLooping := true
+	// we loop because we might pass through multiple parsing states in a single
+	// call to addContent, and we want to make sure callers don't have to wait for
+	// data that's already unambiguous
+	for keepLooping {
+		thinking, remaining, keepLooping = eat(s)
+		thinkingSb.WriteString(thinking)
+		remainingSb.WriteString(remaining)
+	}
+
+	return thinkingSb.String(), remainingSb.String()
+}
+
+// the additional bool return is true iff we should continue eating
+func eat(s *Parser) (string, string, bool) {
+	switch s.state {
+	case thinkingState_LookingForOpening:
+		trimmed := strings.TrimLeftFunc(s.acc.String(), unicode.IsSpace)
+		if strings.HasPrefix(trimmed, s.OpeningTag) {
+			after := strings.Join(strings.Split(trimmed, s.OpeningTag)[1:], s.OpeningTag)
+			after = strings.TrimLeftFunc(after, unicode.IsSpace)
+			// after might contain more than just thinking tokens, so we continue
+			// parsing instead of returning it as thinking tokens here
+			s.acc.Reset()
+			s.acc.WriteString(after)
+			if after == "" {
+				s.state = thinkingState_ThinkingStartedEatingWhitespace
+			} else {
+				s.state = thinkingState_Thinking
+			}
+			return "", "", true
+		} else if strings.HasPrefix(s.OpeningTag, trimmed) {
+			// partial opening seen, so let's keep accumulating
+			return "", "", false
+		} else if trimmed == "" {
+			// saw whitespace only, so let's keep accumulating
+			return "", "", false
+		} else {
+			// didn't see an opening tag, but we have content, so thinking was skipped
+			s.state = thinkingState_ThinkingDone
+			// note that we use the original content, not the trimmed one because we
+			// don't want to eat any whitespace in the real content if there were no
+			// thinking tags
+			return "", s.acc.String(), false
+		}
+	case thinkingState_ThinkingStartedEatingWhitespace:
+		trimmed := strings.TrimLeftFunc(s.acc.String(), unicode.IsSpace)
+		s.acc.Reset()
+		if trimmed == "" {
+			return "", "", false
+		} else {
+			s.state = thinkingState_Thinking
+			s.acc.WriteString(trimmed)
+			return "", "", true
+		}
+	case thinkingState_Thinking:
+		acc := s.acc.String()
+		if strings.Contains(acc, s.ClosingTag) {
+			split := strings.Split(acc, s.ClosingTag)
+			thinking := split[0]
+			remaining := strings.Join(split[1:], s.ClosingTag)
+			remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+			s.acc.Reset()
+			if remaining == "" {
+				s.state = thinkingState_ThinkingDoneEatingWhitespace
+			} else {
+				s.state = thinkingState_ThinkingDone
+			}
+			return thinking, remaining, false
+		} else if overlapLen := overlap(acc, s.ClosingTag); overlapLen > 0 {
+			thinking := acc[:len(acc)-overlapLen]
+			remaining := acc[len(acc)-overlapLen:]
+			s.acc.Reset()
+			// keep track of the candidate closing tag. We have to buffer it until it
+			// becomes disambiguated
+			s.acc.WriteString(remaining)
+			return thinking, "", false
+		} else {
+			// purely just thinking tokens, so we can return them
+			s.acc.Reset()
+			return acc, "", false
+		}
+	case thinkingState_ThinkingDoneEatingWhitespace:
+		trimmed := strings.TrimLeftFunc(s.acc.String(), unicode.IsSpace)
+		s.acc.Reset()
+		// if we see non-whitespace, we're done eating the leading whitespace of the content
+		if trimmed != "" {
+			s.state = thinkingState_ThinkingDone
+		}
+		return "", trimmed, false
+	case thinkingState_ThinkingDone:
+		acc := s.acc.String()
+		s.acc.Reset()
+		return "", acc, false
+	default:
+		panic("unknown state")
+	}
+}
+
+// longest overlap between suffix of s and prefix of delim
+func overlap(s, delim string) int {
+	max := min(len(delim), len(s))
+	for i := max; i > 0; i-- {
+		if strings.HasSuffix(s, delim[:i]) {
+			return i
+		}
+	}
+	return 0
+}

thinking/parser_test.go

@@ -0,0 +1,278 @@
+package thinking
+
+import (
+	"testing"
+)
+
+func TestExtractThinking(t *testing.T) {
+	tests := []struct {
+		in, wantContent, wantThink string
+	}{
+		{
+			in:          "<think> internal </think> world",
+			wantThink:   "internal ",
+			wantContent: "world",
+		},
+		{
+			in:          "<think>a</think><think>b</think>c",
+			wantThink:   "a",
+			wantContent: "<think>b</think>c",
+		},
+		{
+			in:          "no think",
+			wantThink:   "",
+			wantContent: "no think",
+		},
+	}
+	for i, tt := range tests {
+		parser := Parser{
+			OpeningTag: "<think>",
+			ClosingTag: "</think>",
+		}
+		gotThinking, gotContent := parser.AddContent(tt.in)
+		if gotContent != tt.wantContent || gotThinking != tt.wantThink {
+			t.Errorf("case %d: got (%q,%q), want (%q,%q)", i, gotThinking, gotContent, tt.wantThink, tt.wantContent)
+		}
+	}
+}
+
+func TestThinkingStreaming(t *testing.T) {
+	type step struct {
+		input          string
+		wantThinking   string
+		wantContent    string
+		wantStateAfter thinkingState
+	}
+
+	cases := []struct {
+		desc  string
+		skip  bool
+		steps []step
+	}{
+		{
+			desc: "content without a thinking tag",
+			steps: []step{
+				{
+					input:          "  abc",
+					wantThinking:   "",
+					wantContent:    "  abc",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "content before a thinking tag nerfs the thinking tag",
+			steps: []step{
+				{
+					input:          "  abc <think>def</think> ghi",
+					wantThinking:   "",
+					wantContent:    "  abc <think>def</think> ghi",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "building up a thinking tag partially",
+			steps: []step{
+				{
+					input:          "  <th",
+					wantThinking:   "",
+					wantContent:    "",
+					wantStateAfter: thinkingState_LookingForOpening,
+				},
+				{
+					input:          "in",
+					wantThinking:   "",
+					wantContent:    "",
+					wantStateAfter: thinkingState_LookingForOpening,
+				},
+				{
+					input:          "k>a",
+					wantThinking:   "a",
+					wantContent:    "",
+					wantStateAfter: thinkingState_Thinking,
+				},
+			},
+		},
+		{
+			desc: "partial closing tag",
+			steps: []step{
+				{
+					input:          "<think>abc</th",
+					wantThinking:   "abc",
+					wantContent:    "",
+					wantStateAfter: thinkingState_Thinking,
+				},
+				{
+					input:          "ink>def",
+					wantThinking:   "",
+					wantContent:    "def",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "partial closing tag fakeout",
+			steps: []step{
+				{
+					input:          "<think>abc</th",
+					wantThinking:   "abc",
+					wantContent:    "",
+					wantStateAfter: thinkingState_Thinking,
+				},
+				{
+					input:          "ing>def",
+					wantThinking:   "</thing>def",
+					wantContent:    "",
+					wantStateAfter: thinkingState_Thinking,
+				},
+				{
+					input:          "ghi</thi",
+					wantThinking:   "ghi",
+					wantContent:    "",
+					wantStateAfter: thinkingState_Thinking,
+				},
+				{
+					input:          "nk>jkl",
+					wantThinking:   "",
+					wantContent:    "jkl",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "whitespace after thinking tag",
+			steps: []step{
+				{
+					input:          "  <think>abc</think>\n\ndef",
+					wantThinking:   "abc",
+					wantContent:    "def",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "whitespace after thinking tag (incremental)",
+			steps: []step{
+				{
+					input:          "  <think>abc</think>",
+					wantThinking:   "abc",
+					wantContent:    "",
+					wantStateAfter: thinkingState_ThinkingDoneEatingWhitespace,
+				},
+				{
+					input:          "\n\ndef",
+					wantThinking:   "",
+					wantContent:    "def",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "whitespace after thinking tag with content and more whitespace",
+			steps: []step{
+				{
+					input:          "  <think>abc</think>\n\ndef ",
+					wantThinking:   "abc",
+					wantContent:    "def ",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+				{
+					input:          " ghi",
+					wantThinking:   "",
+					wantContent:    " ghi",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "token by token",
+			steps: []step{
+				{
+					input:          "<think>",
+					wantThinking:   "",
+					wantContent:    "",
+					wantStateAfter: thinkingState_ThinkingStartedEatingWhitespace,
+				},
+				{
+					input:          "\n",
+					wantThinking:   "",
+					wantContent:    "",
+					wantStateAfter: thinkingState_ThinkingStartedEatingWhitespace,
+				},
+				{
+					input:          "</think>",
+					wantThinking:   "",
+					wantContent:    "",
+					wantStateAfter: thinkingState_ThinkingDoneEatingWhitespace,
+				},
+				{
+					input:          "\n\n",
+					wantThinking:   "",
+					wantContent:    "",
+					wantStateAfter: thinkingState_ThinkingDoneEatingWhitespace,
+				},
+				{
+					input:          "Hi",
+					wantThinking:   "",
+					wantContent:    "Hi",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+				{
+					input:          " there",
+					wantThinking:   "",
+					wantContent:    " there",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+		{
+			desc: "leading thinking whitespace",
+			steps: []step{
+				{
+					input:          "  <think>   \t ",
+					wantThinking:   "",
+					wantContent:    "",
+					wantStateAfter: thinkingState_ThinkingStartedEatingWhitespace,
+				},
+				{
+					input:          "  these are some ",
+					wantThinking:   "these are some ",
+					wantContent:    "",
+					wantStateAfter: thinkingState_Thinking,
+				},
+				{
+					input:          "thoughts </think>  ",
+					wantThinking:   "thoughts ",
+					wantContent:    "",
+					wantStateAfter: thinkingState_ThinkingDoneEatingWhitespace,
+				},
+				{
+					input:          "  more content",
+					wantThinking:   "",
+					wantContent:    "more content",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
+			},
+		},
+	}
+
+	for _, c := range cases {
+		parser := Parser{
+			OpeningTag: "<think>",
+			ClosingTag: "</think>",
+		}
+		if c.skip {
+			continue
+		}
+		for i, step := range c.steps {
+			thinking, content := parser.AddContent(step.input)
+			if content != step.wantContent || thinking != step.wantThinking {
+				t.Errorf("case %q (step %d): got (%q,%q), want (%q,%q)", c.desc, i, content, thinking, step.wantContent, step.wantThinking)
+			}
+			if parser.state != step.wantStateAfter {
+				t.Errorf("case %q (step %d): got state %s, want %s", c.desc, i, parser.state, step.wantStateAfter)
+			}
+		}
+	}
+}

thinking/template.go

@@ -0,0 +1,134 @@
+package thinking
+
+import (
+	"strings"
+	"text/template"
+	"text/template/parse"
+)
+
+func templateVisit(n parse.Node, enterFn func(parse.Node) bool, exitFn func(parse.Node)) {
+	if n == nil {
+		return
+	}
+	shouldContinue := enterFn(n)
+	if !shouldContinue {
+		return
+	}
+	switch x := n.(type) {
+	case *parse.ListNode:
+		for _, c := range x.Nodes {
+			templateVisit(c, enterFn, exitFn)
+		}
+	case *parse.BranchNode:
+		if x.Pipe != nil {
+			templateVisit(x.Pipe, enterFn, exitFn)
+		}
+		if x.List != nil {
+			templateVisit(x.List, enterFn, exitFn)
+		}
+		if x.ElseList != nil {
+			templateVisit(x.ElseList, enterFn, exitFn)
+		}
+	case *parse.ActionNode:
+		templateVisit(x.Pipe, enterFn, exitFn)
+	case *parse.WithNode:
+		templateVisit(&x.BranchNode, enterFn, exitFn)
+	case *parse.RangeNode:
+		templateVisit(&x.BranchNode, enterFn, exitFn)
+	case *parse.IfNode:
+		templateVisit(&x.BranchNode, enterFn, exitFn)
+	case *parse.TemplateNode:
+		templateVisit(x.Pipe, enterFn, exitFn)
+	case *parse.PipeNode:
+		for _, c := range x.Cmds {
+			templateVisit(c, enterFn, exitFn)
+		}
+	case *parse.CommandNode:
+		for _, a := range x.Args {
+			templateVisit(a, enterFn, exitFn)
+		}
+		// text, field, number, etc. are leaves – nothing to recurse into
+	}
+	if exitFn != nil {
+		exitFn(n)
+	}
+}
+
+// InferTags uses a heuristic to infer the tags that surround thinking traces:
+// We look for a range node that iterates over "Messages" and then look for a
+// reference to "Thinking" like `{{.Thinking}}`. We then go up to the nearest
+// ListNode and take the first and last TextNodes as the opening and closing
+// tags.
+func InferTags(t *template.Template) (string, string) {
+	ancestors := []parse.Node{}
+
+	openingTag := ""
+	closingTag := ""
+
+	enterFn := func(n parse.Node) bool {
+		ancestors = append(ancestors, n)
+
+		switch x := n.(type) {
+		case *parse.FieldNode:
+			if len(x.Ident) > 0 && x.Ident[0] == "Thinking" {
+				var mostRecentRange *parse.RangeNode
+				for i := len(ancestors) - 1; i >= 0; i-- {
+					if r, ok := ancestors[i].(*parse.RangeNode); ok {
+						mostRecentRange = r
+						break
+					}
+				}
+				if mostRecentRange == nil || !rangeUsesField(mostRecentRange, "Messages") {
+					return true
+				}
+
+				// TODO(drifkin): to be more robust, check that it's in the action
+				// part, not the `if`'s pipeline part. We do match on the nearest list
+				// that starts and ends with text nodes, which makes this not strictly
+				// necessary for our heuristic
+
+				// go up to the nearest ancestor that is a *parse.ListNode
+				for i := len(ancestors) - 1; i >= 0; i-- {
+					if l, ok := ancestors[i].(*parse.ListNode); ok {
+						firstNode := l.Nodes[0]
+						if t, ok := firstNode.(*parse.TextNode); ok {
+							openingTag = strings.TrimSpace(t.String())
+						}
+						lastNode := l.Nodes[len(l.Nodes)-1]
+						if t, ok := lastNode.(*parse.TextNode); ok {
+							closingTag = strings.TrimSpace(t.String())
+						}
+
+						break
+					}
+				}
+			}
+		}
+
+		return true
+	}
+
+	exitFn := func(n parse.Node) {
+		ancestors = ancestors[:len(ancestors)-1]
+	}
+
+	templateVisit(t.Root, enterFn, exitFn)
+
+	return openingTag, closingTag
+}
+
+// checks to see if the given field name is present in the pipeline of the given range node
+func rangeUsesField(rangeNode *parse.RangeNode, field string) bool {
+	found := false
+	enterFn := func(n parse.Node) bool {
+		switch x := n.(type) {
+		case *parse.FieldNode:
+			if x.Ident[0] == field {
+				found = true
+			}
+		}
+		return true
+	}
+	templateVisit(rangeNode.BranchNode.Pipe, enterFn, nil)
+	return found
+}

thinking/template_test.go

@@ -0,0 +1,130 @@
+package thinking
+
+import (
+	"testing"
+	"text/template"
+)
+
+func TestInferThinkingTags(t *testing.T) {
+	cases := []struct {
+		desc           string
+		tmplString     string
+		wantOpeningTag string
+		wantClosingTag string
+	}{
+		{
+			desc: "basic",
+			tmplString: `
+			{{ if .Thinking}}
+				/think
+			{{ end }}
+			{{- range $i, $_ := .Messages }}
+				{{- $last := eq (len (slice $.Messages $i)) 1 -}}
+				{{ if and $last .Thinking }}
+					<think>{{ .Thinking }}</think>
+				{{ end }}
+			{{ end }}
+		`,
+			wantOpeningTag: "<think>",
+			wantClosingTag: "</think>",
+		},
+		{
+			desc: "doubly nested range",
+			tmplString: `
+			{{ if .Thinking}}
+				/think
+			{{ end }}
+			{{- range $i, $_ := .Messages }}
+				{{- range $j, $_ := .NotMessages }}
+					{{- $last := eq (len (slice $.Messages $i)) 1 -}}
+					{{ if and $last .Thinking }}
+						<think>{{ .Thinking }}</think>
+					{{ end }}
+				{{ end }}
+			{{ end }}
+		`,
+			wantOpeningTag: "",
+			wantClosingTag: "",
+		},
+		{
+			desc: "whitespace is trimmed",
+			tmplString: `
+			{{ if .Thinking}}
+				/think
+			{{ end }}
+			{{- range $i, $_ := .Messages }}
+				{{- $last := eq (len (slice $.Messages $i)) 1 -}}
+				{{ if and $last .Thinking }}
+					Some text before   {{ .Thinking }}    Some text after
+				{{ end }}
+			{{ end }}
+		`,
+			wantOpeningTag: "Some text before",
+			wantClosingTag: "Some text after",
+		},
+		{
+			desc: "qwen3",
+			tmplString: `
+{{- if or .System .Tools .Thinking }}<|im_start|>system
+{{- if .System }}
+{{ .System }}
+{{- end }}
+{{- if .Tools }}
+
+# Tools
+
+You may call one or more functions to assist with the user query.
+
+You are provided with function signatures within <tools></tools> XML tags:
+<tools>
+{{- range .Tools }}
+{"type": "function", "function": {{ .Function }}}
+{{- end }}
+</tools>
+
+For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
+<tool_call>
+{"name": <function-name>, "arguments": <args-json-object>}
+</tool_call>
+{{- end }}
+{{- if .Thinking }}
+/think
+{{- else }}
+/no_think
+{{- end }}<|im_end|>
+{{ end }}
+{{- range $i, $_ := .Messages }}
+{{- $last := eq (len (slice $.Messages $i)) 1 -}}
+{{- if eq .Role "user" }}<|im_start|>user
+{{ .Content }}<|im_end|>
+{{ else if eq .Role "assistant" }}<|im_start|>assistant
+{{ if and $last .Thinking }}
+<think>{{ .Thinking }}</think>
+{{ end }}
+{{ if .Content }}{{ .Content }}
+{{- else if .ToolCalls }}<tool_call>
+{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
+{{ end }}</tool_call>
+{{- end }}{{ if not $last }}<|im_end|>
+{{ end }}
+{{- else if eq .Role "tool" }}<|im_start|>user
+<tool_response>
+{{ .Content }}
+</tool_response><|im_end|>
+{{ end }}
+{{- if and (ne .Role "assistant") $last }}<|im_start|>assistant
+{{ end }}
+{{- end }}
+			`,
+			wantOpeningTag: "<think>",
+			wantClosingTag: "</think>",
+		},
+	}
+	for _, c := range cases {
+		tmpl := template.Must(template.New("test").Parse(c.tmplString))
+		openingTag, closingTag := InferTags(tmpl)
+		if openingTag != c.wantOpeningTag || closingTag != c.wantClosingTag {
+			t.Errorf("case %q: got (%q,%q), want (%q,%q)", c.desc, openingTag, closingTag, c.wantOpeningTag, c.wantClosingTag)
+		}
+	}
+}