parser: tidy up parameter/message parsing

This change addresses how parameters and messages are handled while parsing a Modelfile.
Currently a `MESSAGE` command creates a string that looks like "<role>: <message>" which
then has to be re-parsed, and `PARAMETER` ends up being the default data type which makes
it difficult to add other multi-part commands.

This change introduces a Message and a Parameter type which properly handle properties
such as the role and the name of the parameter.

.gitignore

@@ -6,7 +6,6 @@
 dist
 build
 .cache
-.gocache
 *.exe
 .idea
 test_data

CMakeLists.txt

@@ -98,17 +98,14 @@ check_language(HIP)
 if(CMAKE_HIP_COMPILER)
     set(HIP_PLATFORM "amd")
 
+    find_package(hip REQUIRED)
     if(NOT AMDGPU_TARGETS)
-        find_package(hip REQUIRED)
         list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
-    endif()
-
-    if(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
+    elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
         list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
     endif()
 
     if(AMDGPU_TARGETS)
-        find_package(hip REQUIRED)
         add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
 
         if (WIN32)

Dockerfile

@@ -1,7 +1,6 @@
 # vim: filetype=dockerfile
 
 ARG FLAVOR=${TARGETARCH}
-ARG PARALLEL=8
 
 ARG ROCMVERSION=6.3.3
 ARG JETPACK5VERSION=r35.4.1
@@ -35,51 +34,46 @@ ENV LDFLAGS=-s
 FROM base AS cpu
 RUN dnf install -y gcc-toolset-11-gcc gcc-toolset-11-gcc-c++
 ENV PATH=/opt/rh/gcc-toolset-11/root/usr/bin:$PATH
-ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'CPU' \
-        && cmake --build --parallel ${PARALLEL} --preset 'CPU' \
-        && cmake --install build --component CPU --strip --parallel ${PARALLEL}
+        && cmake --build --parallel --preset 'CPU' \
+        && cmake --install build --component CPU --strip --parallel 8
 
 FROM base AS cuda-11
 ARG CUDA11VERSION=11.8
 RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
 ENV PATH=/usr/local/cuda-11/bin:$PATH
-ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'CUDA 11' -DOLLAMA_RUNNER_DIR="cuda_v11" \
-        && cmake --build --parallel ${PARALLEL} --preset 'CUDA 11' \
-        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}
+        && cmake --build --parallel --preset 'CUDA 11' \
+        && cmake --install build --component CUDA --strip --parallel 8
 
 FROM base AS cuda-12
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
-ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'CUDA 12' -DOLLAMA_RUNNER_DIR="cuda_v12"\
-        && cmake --build --parallel ${PARALLEL} --preset 'CUDA 12' \
-        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}
+        && cmake --build --parallel --preset 'CUDA 12' \
+        && cmake --install build --component CUDA --strip --parallel 8
 
 
 FROM base AS cuda-13
 ARG CUDA13VERSION=13.0
 RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-}
 ENV PATH=/usr/local/cuda-13/bin:$PATH
-ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'CUDA 13' -DOLLAMA_RUNNER_DIR="cuda_v13" \
-        && cmake --build --parallel ${PARALLEL} --preset 'CUDA 13' \
-        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}
+        && cmake --build --parallel --preset 'CUDA 13' \
+        && cmake --install build --component CUDA --strip --parallel 8
 
 
 FROM base AS rocm-6
 ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
-ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'ROCm 6' \
-        && cmake --build --parallel ${PARALLEL} --preset 'ROCm 6' \
-        && cmake --install build --component HIP --strip --parallel ${PARALLEL}
+        && cmake --build --parallel --preset 'ROCm 6' \
+        && cmake --install build --component HIP --strip --parallel 8
 
 FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
 ARG CMAKEVERSION
@@ -87,11 +81,10 @@ RUN apt-get update && apt-get install -y curl ccache \
     && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
 COPY CMakeLists.txt CMakePresets.json .
 COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'JetPack 5' \
-        && cmake --build --parallel ${PARALLEL} --preset 'JetPack 5' \
-        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}
+        && cmake --build --parallel --preset 'JetPack 5' \
+        && cmake --install build --component CUDA --strip --parallel 8
 
 FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
 ARG CMAKEVERSION
@@ -99,11 +92,10 @@ RUN apt-get update && apt-get install -y curl ccache \
     && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
 COPY CMakeLists.txt CMakePresets.json .
 COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-ARG PARALLEL
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'JetPack 6' \
-        && cmake --build --parallel ${PARALLEL} --preset 'JetPack 6' \
-        && cmake --install build --component CUDA --strip --parallel ${PARALLEL}
+        && cmake --build --parallel --preset 'JetPack 6' \
+        && cmake --install build --component CUDA --strip --parallel 8
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama

api/client.go

@@ -45,12 +45,6 @@ func checkError(resp *http.Response, body []byte) error {
 		return nil
 	}
 
-	if resp.StatusCode == http.StatusUnauthorized {
-		authError := AuthorizationError{StatusCode: resp.StatusCode}
-		json.Unmarshal(body, &authError)
-		return authError
-	}
-
 	apiError := StatusError{StatusCode: resp.StatusCode}
 
 	err := json.Unmarshal(body, &apiError)
@@ -220,8 +214,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	scanner.Buffer(scanBuf, maxBufferSize)
 	for scanner.Scan() {
 		var errorResponse struct {
-			Error     string `json:"error,omitempty"`
-			SigninURL string `json:"signin_url,omitempty"`
+			Error string `json:"error,omitempty"`
 		}
 
 		bts := scanner.Bytes()
@@ -229,13 +222,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 			return fmt.Errorf("unmarshal: %w", err)
 		}
 
-		if response.StatusCode == http.StatusUnauthorized {
-			return AuthorizationError{
-				StatusCode: response.StatusCode,
-				Status:     response.Status,
-				SigninURL:  errorResponse.SigninURL,
-			}
-		} else if response.StatusCode >= http.StatusBadRequest {
+		if response.StatusCode >= http.StatusBadRequest {
 			return StatusError{
 				StatusCode:   response.StatusCode,
 				Status:       response.Status,
@@ -414,24 +401,6 @@ func (c *Client) Embed(ctx context.Context, req *EmbedRequest) (*EmbedResponse,
 	return &resp, nil
 }
 
-// WebSearch performs a web search via the Ollama server.
-func (c *Client) WebSearch(ctx context.Context, req *WebSearchRequest) (*WebSearchResponse, error) {
-	var resp WebSearchResponse
-	if err := c.do(ctx, http.MethodPost, "/api/web_search", req, &resp); err != nil {
-		return nil, err
-	}
-	return &resp, nil
-}
-
-// WebFetch fetches the contents of a web page via the Ollama server.
-func (c *Client) WebFetch(ctx context.Context, req *WebFetchRequest) (*WebFetchResponse, error) {
-	var resp WebFetchResponse
-	if err := c.do(ctx, http.MethodPost, "/api/web_fetch", req, &resp); err != nil {
-		return nil, err
-	}
-	return &resp, nil
-}
-
 // Embeddings generates an embedding from a model.
 func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*EmbeddingResponse, error) {
 	var resp EmbeddingResponse
@@ -459,21 +428,3 @@ func (c *Client) Version(ctx context.Context) (string, error) {
 
 	return version.Version, nil
 }
-
-// Signout will signout a client for a local ollama server.
-func (c *Client) Signout(ctx context.Context) error {
-	return c.do(ctx, http.MethodPost, "/api/signout", nil, nil)
-}
-
-// Disconnect will disconnect an ollama instance from ollama.com.
-func (c *Client) Disconnect(ctx context.Context, encodedKey string) error {
-	return c.do(ctx, http.MethodDelete, fmt.Sprintf("/api/user/keys/%s", encodedKey), nil, nil)
-}
-
-func (c *Client) Whoami(ctx context.Context) (*UserResponse, error) {
-	var resp UserResponse
-	if err := c.do(ctx, http.MethodPost, "/api/me", nil, &resp); err != nil {
-		return nil, err
-	}
-	return &resp, nil
-}

api/client_test.go

@@ -262,135 +262,3 @@ func TestClientDo(t *testing.T) {
 		})
 	}
 }
-
-func TestClientWebSearch(t *testing.T) {
-	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.Method != http.MethodPost {
-			t.Fatalf("expected POST, got %s", r.Method)
-		}
-		if !strings.HasSuffix(r.URL.Path, "/api/web_search") {
-			t.Fatalf("unexpected path: %s", r.URL.Path)
-		}
-
-		var req WebSearchRequest
-		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-			t.Fatalf("failed to decode request: %v", err)
-		}
-
-		if req.Query != "what is ollama" {
-			t.Fatalf("unexpected query: %s", req.Query)
-		}
-
-		if req.MaxResults != 3 {
-			t.Fatalf("unexpected max_results: %d", req.MaxResults)
-		}
-
-		resp := WebSearchResponse{
-			Results: []WebSearchResult{{
-				Title:   "Ollama",
-				URL:     "https://ollama.com",
-				Content: "Cloud models are now available...",
-			}},
-		}
-		if err := json.NewEncoder(w).Encode(resp); err != nil {
-			t.Fatalf("failed to encode response: %v", err)
-		}
-	}))
-	defer ts.Close()
-
-	u, err := url.Parse(ts.URL)
-	if err != nil {
-		t.Fatalf("parse server URL: %v", err)
-	}
-
-	client := NewClient(u, ts.Client())
-
-	resp, err := client.WebSearch(t.Context(), &WebSearchRequest{Query: "what is ollama", MaxResults: 3})
-	if err != nil {
-		t.Fatalf("WebSearch returned error: %v", err)
-	}
-
-	if len(resp.Results) != 1 {
-		t.Fatalf("expected 1 result, got %d", len(resp.Results))
-	}
-
-	if resp.Results[0].Title != "Ollama" {
-		t.Fatalf("unexpected title: %s", resp.Results[0].Title)
-	}
-}
-
-func TestClientWebSearchUnauthorized(t *testing.T) {
-	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		w.WriteHeader(http.StatusUnauthorized)
-		_ = json.NewEncoder(w).Encode(map[string]string{
-			"signin_url": "https://ollama.com/connect",
-		})
-	}))
-	defer ts.Close()
-
-	u, err := url.Parse(ts.URL)
-	if err != nil {
-		t.Fatalf("parse server URL: %v", err)
-	}
-
-	client := NewClient(u, ts.Client())
-
-	_, err = client.WebSearch(t.Context(), &WebSearchRequest{Query: "what is ollama"})
-	if err == nil {
-		t.Fatal("expected error, got nil")
-	}
-
-	if _, ok := err.(AuthorizationError); !ok {
-		t.Fatalf("expected AuthorizationError, got %T", err)
-	}
-}
-
-func TestClientWebFetch(t *testing.T) {
-	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.Method != http.MethodPost {
-			t.Fatalf("expected POST, got %s", r.Method)
-		}
-		if !strings.HasSuffix(r.URL.Path, "/api/web_fetch") {
-			t.Fatalf("unexpected path: %s", r.URL.Path)
-		}
-
-		var req WebFetchRequest
-		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-			t.Fatalf("failed to decode request: %v", err)
-		}
-
-		if req.URL != "https://ollama.com" {
-			t.Fatalf("unexpected url: %s", req.URL)
-		}
-
-		resp := WebFetchResponse{
-			Title:   "Ollama",
-			Content: "Cloud models are now available...",
-			Links:   []string{"https://ollama.com/models"},
-		}
-		if err := json.NewEncoder(w).Encode(resp); err != nil {
-			t.Fatalf("failed to encode response: %v", err)
-		}
-	}))
-	defer ts.Close()
-
-	u, err := url.Parse(ts.URL)
-	if err != nil {
-		t.Fatalf("parse server URL: %v", err)
-	}
-
-	client := NewClient(u, ts.Client())
-
-	resp, err := client.WebFetch(t.Context(), &WebFetchRequest{URL: "https://ollama.com"})
-	if err != nil {
-		t.Fatalf("WebFetch returned error: %v", err)
-	}
-
-	if resp.Title != "Ollama" {
-		t.Fatalf("unexpected title: %s", resp.Title)
-	}
-
-	if len(resp.Links) != 1 || resp.Links[0] != "https://ollama.com/models" {
-		t.Fatalf("unexpected links: %v", resp.Links)
-	}
-}

api/types.go

@@ -11,8 +11,6 @@ import (
 	"strings"
 	"time"
 
-	"github.com/google/uuid"
-
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/types/model"
 )
@@ -38,19 +36,6 @@ func (e StatusError) Error() string {
 	}
 }
 
-type AuthorizationError struct {
-	StatusCode int
-	Status     string
-	SigninURL  string `json:"signin_url"`
-}
-
-func (e AuthorizationError) Error() string {
-	if e.Status != "" {
-		return e.Status
-	}
-	return "something went wrong, please see the ollama server logs for details"
-}
-
 // ImageData represents the raw binary data of an image file.
 type ImageData []byte
 
@@ -328,29 +313,13 @@ func (t *ToolFunction) String() string {
 // ChatResponse is the response returned by [Client.Chat]. Its fields are
 // similar to [GenerateResponse].
 type ChatResponse struct {
-	// Model is the model name that generated the response.
-	Model string `json:"model"`
+	Model      string    `json:"model"`
+	CreatedAt  time.Time `json:"created_at"`
+	Message    Message   `json:"message"`
+	DoneReason string    `json:"done_reason,omitempty"`
 
-	// RemoteModel is the name of the upstream model that generated the response.
-	RemoteModel string `json:"remote_model,omitempty"`
-
-	// RemoteHost is the URL of the upstream Ollama host that generated the response.
-	RemoteHost string `json:"remote_host,omitempty"`
-
-	// CreatedAt is the timestamp of the response.
-	CreatedAt time.Time `json:"created_at"`
-
-	// Message contains the message or part of a message from the model.
-	Message Message `json:"message"`
-
-	// Done specifies if the response is complete.
 	Done bool `json:"done"`
 
-	// DoneReason is the reason the model stopped generating text.
-	DoneReason string `json:"done_reason,omitempty"`
-
-	DebugInfo *DebugInfo `json:"_debug_info,omitempty"`
-
 	Metrics
 }
 
@@ -360,6 +329,13 @@ type DebugInfo struct {
 	ImageCount       int    `json:"image_count,omitempty"`
 }
 
+// DebugTemplateResponse is returned when _debug_render_only is set to true
+type DebugTemplateResponse struct {
+	Model     string    `json:"model"`
+	CreatedAt time.Time `json:"created_at"`
+	DebugInfo DebugInfo `json:"_debug_info"`
+}
+
 type Metrics struct {
 	TotalDuration      time.Duration `json:"total_duration,omitempty"`
 	LoadDuration       time.Duration `json:"load_duration,omitempty"`
@@ -453,83 +429,20 @@ type EmbeddingResponse struct {
 	Embedding []float64 `json:"embedding"`
 }
 
-// WebSearchRequest is the request passed to [Client.WebSearch].
-type WebSearchRequest struct {
-	// Query is the search query string.
-	Query string `json:"query"`
-
-	// MaxResults is the optional maximum number of results to return (default 5, max 10).
-	MaxResults int `json:"max_results,omitempty"`
-}
-
-// WebSearchResult represents a single web search result.
-type WebSearchResult struct {
-	Title   string `json:"title"`
-	URL     string `json:"url"`
-	Content string `json:"content"`
-}
-
-// WebSearchResponse is the response from [Client.WebSearch].
-type WebSearchResponse struct {
-	Results []WebSearchResult `json:"results"`
-}
-
-// WebFetchRequest is the request passed to [Client.WebFetch].
-type WebFetchRequest struct {
-	// URL is the address of the page to fetch.
-	URL string `json:"url"`
-}
-
-// WebFetchResponse is the response from [Client.WebFetch].
-type WebFetchResponse struct {
-	Title   string   `json:"title"`
-	Content string   `json:"content"`
-	Links   []string `json:"links"`
-}
-
 // CreateRequest is the request passed to [Client.Create].
 type CreateRequest struct {
-	// Model is the model name to create.
-	Model string `json:"model"`
-
-	// Stream specifies whether the response is streaming; it is true by default.
-	Stream *bool `json:"stream,omitempty"`
-
-	// Quantize is the quantization format for the model; leave blank to not change the quantization level.
+	Model    string `json:"model"`
+	Stream   *bool  `json:"stream,omitempty"`
 	Quantize string `json:"quantize,omitempty"`
 
-	// From is the name of the model or file to use as the source.
-	From string `json:"from,omitempty"`
-
-	// RemoteHost is the URL of the upstream ollama API for the model (if any).
-	RemoteHost string `json:"remote_host,omitempty"`
-
-	// Files is a map of files include when creating the model.
-	Files map[string]string `json:"files,omitempty"`
-
-	// Adapters is a map of LoRA adapters to include when creating the model.
-	Adapters map[string]string `json:"adapters,omitempty"`
-
-	// Template is the template used when constructing a request to the model.
-	Template string `json:"template,omitempty"`
-
-	// License is a string or list of strings for licenses.
-	License any `json:"license,omitempty"`
-
-	// System is the system prompt for the model.
-	System string `json:"system,omitempty"`
-
-	// Parameters is a map of hyper-parameters which are applied to the model.
-	Parameters map[string]any `json:"parameters,omitempty"`
-
-	// Messages is a list of messages added to the model before chat and generation requests.
-	Messages []Message `json:"messages,omitempty"`
-
-	Renderer string `json:"renderer,omitempty"`
-	Parser   string `json:"parser,omitempty"`
-
-	// Info is a map of additional information for the model
-	Info map[string]any `json:"info,omitempty"`
+	From       string            `json:"from,omitempty"`
+	Files      map[string]string `json:"files,omitempty"`
+	Adapters   map[string]string `json:"adapters,omitempty"`
+	Template   string            `json:"template,omitempty"`
+	License    any               `json:"license,omitempty"`
+	System     string            `json:"system,omitempty"`
+	Parameters map[string]any    `json:"parameters,omitempty"`
+	Messages   []Message         `json:"messages,omitempty"`
 
 	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
@@ -567,12 +480,8 @@ type ShowResponse struct {
 	Parameters    string             `json:"parameters,omitempty"`
 	Template      string             `json:"template,omitempty"`
 	System        string             `json:"system,omitempty"`
-	Renderer      string             `json:"renderer,omitempty"`
-	Parser        string             `json:"parser,omitempty"`
 	Details       ModelDetails       `json:"details,omitempty"`
 	Messages      []Message          `json:"messages,omitempty"`
-	RemoteModel   string             `json:"remote_model,omitempty"`
-	RemoteHost    string             `json:"remote_host,omitempty"`
 	ModelInfo     map[string]any     `json:"model_info,omitempty"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
@@ -631,14 +540,12 @@ type ProcessResponse struct {
 
 // ListModelResponse is a single model description in [ListResponse].
 type ListModelResponse struct {
-	Name        string       `json:"name"`
-	Model       string       `json:"model"`
-	RemoteModel string       `json:"remote_model,omitempty"`
-	RemoteHost  string       `json:"remote_host,omitempty"`
-	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"`
+	Details    ModelDetails `json:"details,omitempty"`
 }
 
 // ProcessModelResponse is a single model description in [ProcessResponse].
@@ -662,12 +569,6 @@ type GenerateResponse struct {
 	// Model is the model name that generated the response.
 	Model string `json:"model"`
 
-	// RemoteModel is the name of the upstream model that generated the response.
-	RemoteModel string `json:"remote_model,omitempty"`
-
-	// RemoteHost is the URL of the upstream Ollama host that generated the response.
-	RemoteHost string `json:"remote_host,omitempty"`
-
 	// CreatedAt is the timestamp of the response.
 	CreatedAt time.Time `json:"created_at"`
 
@@ -691,8 +592,6 @@ type GenerateResponse struct {
 	Metrics
 
 	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
-
-	DebugInfo *DebugInfo `json:"_debug_info,omitempty"`
 }
 
 // ModelDetails provides details about a model.
@@ -705,18 +604,6 @@ type ModelDetails struct {
 	QuantizationLevel string   `json:"quantization_level"`
 }
 
-// UserResponse provides information about a user.
-type UserResponse struct {
-	ID        uuid.UUID `json:"id"`
-	Email     string    `json:"email"`
-	Name      string    `json:"name"`
-	Bio       string    `json:"bio,omitempty"`
-	AvatarURL string    `json:"avatarurl,omitempty"`
-	FirstName string    `json:"firstname,omitempty"`
-	LastName  string    `json:"lastname,omitempty"`
-	Plan      string    `json:"plan,omitempty"`
-}
-
 // Tensor describes the metadata for a given tensor.
 type Tensor struct {
 	Name  string   `json:"name"`

auth/auth.go

@@ -18,13 +18,21 @@ import (
 
 const defaultPrivateKey = "id_ed25519"
 
-func GetPublicKey() (string, error) {
+func keyPath() (string, error) {
 	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}
 
-	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
+	return filepath.Join(home, ".ollama", defaultPrivateKey), nil
+}
+
+func GetPublicKey() (string, error) {
+	keyPath, err := keyPath()
+	if err != nil {
+		return "", err
+	}
+
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
@@ -51,12 +59,11 @@ func NewNonce(r io.Reader, length int) (string, error) {
 }
 
 func Sign(ctx context.Context, bts []byte) (string, error) {
-	home, err := os.UserHomeDir()
+	keyPath, err := keyPath()
 	if err != nil {
 		return "", err
 	}
 
-	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))

cmd/cmd.go

@@ -47,8 +47,6 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
-const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
-
 // ensureThinkingSupport emits a warning if the model does not advertise thinking support
 func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
 	if name == "" {
@@ -288,17 +286,7 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 		Think: opts.Think,
 	}
 
-	return client.Generate(cmd.Context(), req, func(r api.GenerateResponse) error {
-		if r.RemoteModel != "" && opts.ShowConnect {
-			p.StopAndClear()
-			if strings.HasPrefix(r.RemoteHost, "https://ollama.com") {
-				fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", r.RemoteModel)
-			} else {
-				fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", r.RemoteModel, r.RemoteHost)
-			}
-		}
-		return nil
-	})
+	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
 }
 
 func StopHandler(cmd *cobra.Command, args []string) error {
@@ -319,10 +307,9 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	interactive := true
 
 	opts := runOptions{
-		Model:       args[0],
-		WordWrap:    os.Getenv("TERM") == "xterm-256color",
-		Options:     map[string]any{},
-		ShowConnect: true,
+		Model:    args[0],
+		WordWrap: os.Getenv("TERM") == "xterm-256color",
+		Options:  map[string]any{},
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -380,7 +367,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		}
 
 		prompts = append([]string{string(in)}, prompts...)
-		opts.ShowConnect = false
 		opts.WordWrap = false
 		interactive = false
 	}
@@ -447,15 +433,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
-			var sErr api.AuthorizationError
-			if errors.As(err, &sErr) && sErr.StatusCode == http.StatusUnauthorized {
-				fmt.Printf("You need to be signed in to Ollama to run Cloud models.\n\n")
-
-				if sErr.SigninURL != "" {
-					fmt.Printf(ConnectInstructions, sErr.SigninURL)
-				}
-				return nil
-			}
 			return err
 		}
 
@@ -476,59 +453,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	return generate(cmd, opts)
 }
 
-func SigninHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	user, err := client.Whoami(cmd.Context())
-	if err != nil {
-		var aErr api.AuthorizationError
-		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
-			fmt.Println("You need to be signed in to Ollama to run Cloud models.")
-			fmt.Println()
-
-			if aErr.SigninURL != "" {
-				fmt.Printf(ConnectInstructions, aErr.SigninURL)
-			}
-			return nil
-		}
-		return err
-	}
-
-	if user != nil && user.Name != "" {
-		fmt.Printf("You are already signed in as user '%s'\n", user.Name)
-		fmt.Println()
-		return nil
-	}
-
-	return nil
-}
-
-func SignoutHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	err = client.Signout(cmd.Context())
-	if err != nil {
-		var aErr api.AuthorizationError
-		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
-			fmt.Println("You are not signed in to ollama.com")
-			fmt.Println()
-			return nil
-		} else {
-			return err
-		}
-	}
-
-	fmt.Println("You have signed out of ollama.com")
-	fmt.Println()
-	return nil
-}
-
 func PushHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -540,25 +464,6 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	n := model.ParseName(args[0])
-	if strings.HasSuffix(n.Host, ".ollama.ai") || strings.HasSuffix(n.Host, ".ollama.com") {
-		_, err := client.Whoami(cmd.Context())
-		if err != nil {
-			var aErr api.AuthorizationError
-			if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
-				fmt.Println("You need to be signed in to push models to ollama.com.")
-				fmt.Println()
-
-				if aErr.SigninURL != "" {
-					fmt.Printf(ConnectInstructions, aErr.SigninURL)
-				}
-				return nil
-			}
-
-			return err
-		}
-	}
-
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
@@ -595,12 +500,12 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 	request := api.PushRequest{Name: args[0], Insecure: insecure}
 
+	n := model.ParseName(args[0])
 	if err := client.Push(cmd.Context(), &request, fn); err != nil {
 		if spinner != nil {
 			spinner.Stop()
 		}
-		errStr := strings.ToLower(err.Error())
-		if strings.Contains(errStr, "access denied") || strings.Contains(errStr, "unauthorized") {
+		if strings.Contains(err.Error(), "access denied") {
 			return errors.New("you are not authorized to push to this namespace, create the model under a namespace you own")
 		}
 		return err
@@ -634,14 +539,7 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 
 	for _, m := range models.Models {
 		if len(args) == 0 || strings.HasPrefix(strings.ToLower(m.Name), strings.ToLower(args[0])) {
-			var size string
-			if m.RemoteModel != "" {
-				size = "-"
-			} else {
-				size = format.HumanBytes(m.Size)
-			}
-
-			data = append(data, []string{m.Name, m.Digest[:12], size, format.HumanTime(m.ModifiedAt, "Never")})
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), format.HumanTime(m.ModifiedAt, "Never")})
 		}
 	}
 
@@ -726,8 +624,8 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 		KeepAlive: &api.Duration{Duration: 0},
 	}
 	if err := loadOrUnloadModel(cmd, opts); err != nil {
-		if !strings.Contains(strings.ToLower(err.Error()), "not found") {
-			fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
+		if !strings.Contains(err.Error(), "not found") {
+			return fmt.Errorf("unable to stop existing running model \"%s\": %s", args[0], err)
 		}
 	}
 
@@ -838,36 +736,12 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 	}
 
 	tableRender("Model", func() (rows [][]string) {
-		if resp.RemoteHost != "" {
-			rows = append(rows, []string{"", "Remote model", resp.RemoteModel})
-			rows = append(rows, []string{"", "Remote URL", resp.RemoteHost})
-		}
-
 		if resp.ModelInfo != nil {
 			arch := resp.ModelInfo["general.architecture"].(string)
 			rows = append(rows, []string{"", "architecture", arch})
-
-			var paramStr string
-			if resp.Details.ParameterSize != "" {
-				paramStr = resp.Details.ParameterSize
-			} else if v, ok := resp.ModelInfo["general.parameter_count"]; ok {
-				if f, ok := v.(float64); ok {
-					paramStr = format.HumanNumber(uint64(f))
-				}
-			}
-			rows = append(rows, []string{"", "parameters", paramStr})
-
-			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)]; ok {
-				if f, ok := v.(float64); ok {
-					rows = append(rows, []string{"", "context length", strconv.FormatFloat(f, 'f', -1, 64)})
-				}
-			}
-
-			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)]; ok {
-				if f, ok := v.(float64); ok {
-					rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(f, 'f', -1, 64)})
-				}
-			}
+			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ModelInfo["general.parameter_count"].(float64)))})
+			rows = append(rows, []string{"", "context length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64), 'f', -1, 64)})
+			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64), 'f', -1, 64)})
 		} else {
 			rows = append(rows, []string{"", "architecture", resp.Details.Family})
 			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
@@ -1115,52 +989,6 @@ type runOptions struct {
 	KeepAlive    *api.Duration
 	Think        *api.ThinkValue
 	HideThinking bool
-	ShowConnect  bool
-}
-
-func (r runOptions) Copy() runOptions {
-	var messages []api.Message
-	if r.Messages != nil {
-		messages = make([]api.Message, len(r.Messages))
-		copy(messages, r.Messages)
-	}
-
-	var images []api.ImageData
-	if r.Images != nil {
-		images = make([]api.ImageData, len(r.Images))
-		copy(images, r.Images)
-	}
-
-	var opts map[string]any
-	if r.Options != nil {
-		opts = make(map[string]any, len(r.Options))
-		for k, v := range r.Options {
-			opts[k] = v
-		}
-	}
-
-	var think *api.ThinkValue
-	if r.Think != nil {
-		cThink := *r.Think
-		think = &cThink
-	}
-
-	return runOptions{
-		Model:        r.Model,
-		ParentModel:  r.ParentModel,
-		Prompt:       r.Prompt,
-		Messages:     messages,
-		WordWrap:     r.WordWrap,
-		Format:       r.Format,
-		System:       r.System,
-		Images:       images,
-		Options:      opts,
-		MultiModal:   r.MultiModal,
-		KeepAlive:    r.KeepAlive,
-		Think:        think,
-		HideThinking: r.HideThinking,
-		ShowConnect:  r.ShowConnect,
-	}
 }
 
 type displayResponseState struct {
@@ -1716,22 +1544,6 @@ func NewCLI() *cobra.Command {
 
 	pushCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 
-	signinCmd := &cobra.Command{
-		Use:     "signin",
-		Short:   "Sign in to ollama.com",
-		Args:    cobra.ExactArgs(0),
-		PreRunE: checkServerHeartbeat,
-		RunE:    SigninHandler,
-	}
-
-	signoutCmd := &cobra.Command{
-		Use:     "signout",
-		Short:   "Sign out from ollama.com",
-		Args:    cobra.ExactArgs(0),
-		PreRunE: checkServerHeartbeat,
-		RunE:    SignoutHandler,
-	}
-
 	listCmd := &cobra.Command{
 		Use:     "list",
 		Aliases: []string{"ls"},
@@ -1826,8 +1638,6 @@ func NewCLI() *cobra.Command {
 		stopCmd,
 		pullCmd,
 		pushCmd,
-		signinCmd,
-		signoutCmd,
 		listCmd,
 		psCmd,
 		copyCmd,

cmd/cmd_test.go

@@ -3,12 +3,10 @@ package cmd
 import (
 	"bytes"
 	"encoding/json"
-	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
-	"reflect"
 	"strings"
 	"testing"
 	"time"
@@ -306,8 +304,6 @@ func TestDeleteHandler(t *testing.T) {
 				w.WriteHeader(http.StatusOK)
 			} else {
 				w.WriteHeader(http.StatusNotFound)
-				errPayload := `{"error":"model '%s' not found"}`
-				w.Write([]byte(fmt.Sprintf(errPayload, req.Name)))
 			}
 			return
 		}
@@ -350,7 +346,7 @@ func TestDeleteHandler(t *testing.T) {
 	}
 
 	err := DeleteHandler(cmd, []string{"test-model-not-found"})
-	if err == nil || !strings.Contains(err.Error(), "model 'test-model-not-found' not found") {
+	if err == nil || !strings.Contains(err.Error(), "unable to stop existing running model \"test-model-not-found\"") {
 		t.Fatalf("DeleteHandler failed: expected error about stopping non-existent model, got %v", err)
 	}
 }
@@ -492,35 +488,9 @@ func TestPushHandler(t *testing.T) {
 						w.(http.Flusher).Flush()
 					}
 				},
-				"/api/me": func(w http.ResponseWriter, r *http.Request) {
-					if r.Method != http.MethodPost {
-						t.Errorf("expected POST request, got %s", r.Method)
-					}
-				},
 			},
 			expectedOutput: "\nYou can find your model at:\n\n\thttps://ollama.com/test-model\n",
 		},
-		{
-			name:      "not signed in push",
-			modelName: "notsignedin-model",
-			serverResponse: map[string]func(w http.ResponseWriter, r *http.Request){
-				"/api/me": func(w http.ResponseWriter, r *http.Request) {
-					if r.Method != http.MethodPost {
-						t.Errorf("expected POST request, got %s", r.Method)
-					}
-					w.Header().Set("Content-Type", "application/json")
-					w.WriteHeader(http.StatusUnauthorized)
-					err := json.NewEncoder(w).Encode(map[string]string{
-						"error":      "unauthorized",
-						"signin_url": "https://somethingsomething",
-					})
-					if err != nil {
-						t.Fatal(err)
-					}
-				},
-			},
-			expectedOutput: "You need to be signed in to push",
-		},
 		{
 			name:      "unauthorized push",
 			modelName: "unauthorized-model",
@@ -529,17 +499,12 @@ func TestPushHandler(t *testing.T) {
 					w.Header().Set("Content-Type", "application/json")
 					w.WriteHeader(http.StatusUnauthorized)
 					err := json.NewEncoder(w).Encode(map[string]string{
-						"error": "403: {\"errors\":[{\"code\":\"ACCESS DENIED\", \"message\":\"access denied\"}]}",
+						"error": "access denied",
 					})
 					if err != nil {
 						t.Fatal(err)
 					}
 				},
-				"/api/me": func(w http.ResponseWriter, r *http.Request) {
-					if r.Method != http.MethodPost {
-						t.Errorf("expected POST request, got %s", r.Method)
-					}
-				},
 			},
 			expectedError: "you are not authorized to push to this namespace, create the model under a namespace you own",
 		},
@@ -557,10 +522,6 @@ func TestPushHandler(t *testing.T) {
 			defer mockServer.Close()
 
 			t.Setenv("OLLAMA_HOST", mockServer.URL)
-			tmpDir := t.TempDir()
-			t.Setenv("HOME", tmpDir)
-			t.Setenv("USERPROFILE", tmpDir)
-			initializeKeypair()
 
 			cmd := &cobra.Command{}
 			cmd.Flags().Bool("insecure", false, "")
@@ -596,7 +557,7 @@ func TestPushHandler(t *testing.T) {
 					t.Errorf("expected no error, got %v", err)
 				}
 				if tt.expectedOutput != "" {
-					if got := string(stdout); !strings.Contains(got, tt.expectedOutput) {
+					if got := string(stdout); got != tt.expectedOutput {
 						t.Errorf("expected output %q, got %q", tt.expectedOutput, got)
 					}
 				}
@@ -954,286 +915,3 @@ func TestNewCreateRequest(t *testing.T) {
 		})
 	}
 }
-
-func TestRunOptions_Copy(t *testing.T) {
-	// Setup test data
-	originalKeepAlive := &api.Duration{Duration: 5 * time.Minute}
-	originalThink := &api.ThinkValue{Value: "test reasoning"}
-
-	original := runOptions{
-		Model:       "test-model",
-		ParentModel: "parent-model",
-		Prompt:      "test prompt",
-		Messages: []api.Message{
-			{Role: "user", Content: "hello"},
-			{Role: "assistant", Content: "hi there"},
-		},
-		WordWrap: true,
-		Format:   "json",
-		System:   "system prompt",
-		Images: []api.ImageData{
-			[]byte("image1"),
-			[]byte("image2"),
-		},
-		Options: map[string]any{
-			"temperature": 0.7,
-			"max_tokens":  1000,
-			"top_p":       0.9,
-		},
-		MultiModal:   true,
-		KeepAlive:    originalKeepAlive,
-		Think:        originalThink,
-		HideThinking: false,
-		ShowConnect:  true,
-	}
-
-	// Test the copy
-	copied := original.Copy()
-
-	// Test 1: Verify the copy is not the same instance
-	if &copied == &original {
-		t.Error("Copy should return a different instance")
-	}
-
-	// Test 2: Verify all fields are copied correctly
-	tests := []struct {
-		name string
-		got  interface{}
-		want interface{}
-	}{
-		{"Model", copied.Model, original.Model},
-		{"ParentModel", copied.ParentModel, original.ParentModel},
-		{"Prompt", copied.Prompt, original.Prompt},
-		{"WordWrap", copied.WordWrap, original.WordWrap},
-		{"Format", copied.Format, original.Format},
-		{"System", copied.System, original.System},
-		{"MultiModal", copied.MultiModal, original.MultiModal},
-		{"HideThinking", copied.HideThinking, original.HideThinking},
-		{"ShowConnect", copied.ShowConnect, original.ShowConnect},
-	}
-
-	for _, tt := range tests {
-		if !reflect.DeepEqual(tt.got, tt.want) {
-			t.Errorf("%s mismatch: got %v, want %v", tt.name, tt.got, tt.want)
-		}
-	}
-
-	// Test 3: Verify Messages slice is deeply copied
-	if len(copied.Messages) != len(original.Messages) {
-		t.Errorf("Messages length mismatch: got %d, want %d", len(copied.Messages), len(original.Messages))
-	}
-
-	if len(copied.Messages) > 0 && &copied.Messages[0] == &original.Messages[0] {
-		t.Error("Messages should be different instances")
-	}
-
-	// Modify original to verify independence
-	if len(original.Messages) > 0 {
-		originalContent := original.Messages[0].Content
-		original.Messages[0].Content = "modified"
-		if len(copied.Messages) > 0 && copied.Messages[0].Content == "modified" {
-			t.Error("Messages should be independent after copy")
-		}
-		// Restore for other tests
-		original.Messages[0].Content = originalContent
-	}
-
-	// Test 4: Verify Images slice is deeply copied
-	if len(copied.Images) != len(original.Images) {
-		t.Errorf("Images length mismatch: got %d, want %d", len(copied.Images), len(original.Images))
-	}
-
-	if len(copied.Images) > 0 && &copied.Images[0] == &original.Images[0] {
-		t.Error("Images should be different instances")
-	}
-
-	// Modify original to verify independence
-	if len(original.Images) > 0 {
-		originalImage := original.Images[0]
-		original.Images[0] = []byte("modified")
-		if len(copied.Images) > 0 && string(copied.Images[0]) == "modified" {
-			t.Error("Images should be independent after copy")
-		}
-		// Restore for other tests
-		original.Images[0] = originalImage
-	}
-
-	// Test 5: Verify Options map is deeply copied
-	if len(copied.Options) != len(original.Options) {
-		t.Errorf("Options length mismatch: got %d, want %d", len(copied.Options), len(original.Options))
-	}
-
-	if len(copied.Options) > 0 && &copied.Options == &original.Options {
-		t.Error("Options map should be different instances")
-	}
-
-	// Modify original to verify independence
-	if len(original.Options) > 0 {
-		originalTemp := original.Options["temperature"]
-		original.Options["temperature"] = 0.9
-		if copied.Options["temperature"] == 0.9 {
-			t.Error("Options should be independent after copy")
-		}
-		// Restore for other tests
-		original.Options["temperature"] = originalTemp
-	}
-
-	// Test 6: Verify KeepAlive pointer is copied (shallow copy)
-	if copied.KeepAlive != original.KeepAlive {
-		t.Error("KeepAlive pointer should be the same (shallow copy)")
-	}
-
-	// Test 7: Verify Think pointer creates a new instance
-	if original.Think != nil && copied.Think == original.Think {
-		t.Error("Think should be a different instance")
-	}
-
-	if original.Think != nil && copied.Think != nil {
-		if !reflect.DeepEqual(copied.Think.Value, original.Think.Value) {
-			t.Errorf("Think.Value mismatch: got %v, want %v", copied.Think.Value, original.Think.Value)
-		}
-	}
-
-	// Test 8: Test with zero values
-	zeroOriginal := runOptions{}
-	zeroCopy := zeroOriginal.Copy()
-
-	if !reflect.DeepEqual(zeroCopy, zeroOriginal) {
-		fmt.Printf("orig: %#v\ncopy: %#v\n", zeroOriginal, zeroCopy)
-		t.Error("Copy of zero value should equal original zero value")
-	}
-}
-
-func TestRunOptions_Copy_EmptySlicesAndMaps(t *testing.T) {
-	// Test with empty slices and maps
-	original := runOptions{
-		Messages: []api.Message{},
-		Images:   []api.ImageData{},
-		Options:  map[string]any{},
-	}
-
-	copied := original.Copy()
-
-	if copied.Messages == nil {
-		t.Error("Empty Messages slice should remain empty, not nil")
-	}
-
-	if copied.Images == nil {
-		t.Error("Empty Images slice should remain empty, not nil")
-	}
-
-	if copied.Options == nil {
-		t.Error("Empty Options map should remain empty, not nil")
-	}
-
-	if len(copied.Messages) != 0 {
-		t.Error("Empty Messages slice should remain empty")
-	}
-
-	if len(copied.Images) != 0 {
-		t.Error("Empty Images slice should remain empty")
-	}
-
-	if len(copied.Options) != 0 {
-		t.Error("Empty Options map should remain empty")
-	}
-}
-
-func TestRunOptions_Copy_NilPointers(t *testing.T) {
-	// Test with nil pointers
-	original := runOptions{
-		KeepAlive: nil,
-		Think:     nil,
-	}
-
-	copied := original.Copy()
-
-	if copied.KeepAlive != nil {
-		t.Error("Nil KeepAlive should remain nil")
-	}
-
-	if copied.Think != nil {
-		t.Error("Nil Think should remain nil")
-	}
-}
-
-func TestRunOptions_Copy_ThinkValueVariants(t *testing.T) {
-	tests := []struct {
-		name  string
-		think *api.ThinkValue
-	}{
-		{"nil Think", nil},
-		{"bool true", &api.ThinkValue{Value: true}},
-		{"bool false", &api.ThinkValue{Value: false}},
-		{"string value", &api.ThinkValue{Value: "reasoning text"}},
-		{"int value", &api.ThinkValue{Value: 42}},
-		{"nil value", &api.ThinkValue{Value: nil}},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			original := runOptions{Think: tt.think}
-			copied := original.Copy()
-
-			if tt.think == nil {
-				if copied.Think != nil {
-					t.Error("Nil Think should remain nil")
-				}
-				return
-			}
-
-			if copied.Think == nil {
-				t.Error("Non-nil Think should not become nil")
-				return
-			}
-
-			if copied.Think == original.Think {
-				t.Error("Think should be a different instance")
-			}
-
-			if !reflect.DeepEqual(copied.Think.Value, original.Think.Value) {
-				t.Errorf("Think.Value mismatch: got %v, want %v", copied.Think.Value, original.Think.Value)
-			}
-		})
-	}
-}
-
-func TestRunOptions_Copy_Independence(t *testing.T) {
-	// Test that modifications to original don't affect copy
-	originalThink := &api.ThinkValue{Value: "original"}
-	original := runOptions{
-		Model:    "original-model",
-		Messages: []api.Message{{Role: "user", Content: "original"}},
-		Options:  map[string]any{"key": "value"},
-		Think:    originalThink,
-	}
-
-	copied := original.Copy()
-
-	// Modify original
-	original.Model = "modified-model"
-	if len(original.Messages) > 0 {
-		original.Messages[0].Content = "modified"
-	}
-	original.Options["key"] = "modified"
-	if original.Think != nil {
-		original.Think.Value = "modified"
-	}
-
-	// Verify copy is unchanged
-	if copied.Model == "modified-model" {
-		t.Error("Copy Model should not be affected by original modification")
-	}
-
-	if len(copied.Messages) > 0 && copied.Messages[0].Content == "modified" {
-		t.Error("Copy Messages should not be affected by original modification")
-	}
-
-	if copied.Options["key"] == "modified" {
-		t.Error("Copy Options should not be affected by original modification")
-	}
-
-	if copied.Think != nil && copied.Think.Value == "modified" {
-		t.Error("Copy Think should not be affected by original modification")
-	}
-}

cmd/interactive.go

@@ -195,24 +195,16 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				fmt.Println("Usage:\n  /load <modelname>")
 				continue
 			}
-			origOpts := opts.Copy()
-
 			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 {
-				if strings.Contains(err.Error(), "not found") {
-					fmt.Printf("Couldn't find model '%s'\n", opts.Model)
-					opts = origOpts.Copy()
-					continue
-				}
 				return err
 			}
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				if strings.Contains(err.Error(), "not found") {
-					fmt.Printf("Couldn't find model '%s'\n", opts.Model)
-					opts = origOpts.Copy()
+					fmt.Printf("error: %v\n", err)
 					continue
 				}
 				if strings.Contains(err.Error(), "does not support thinking") {

convert/reader_safetensors.go

@@ -96,7 +96,7 @@ type safetensor struct {
 
 func (st safetensor) Kind() uint32 {
 	kind := st.tensorBase.Kind()
-	if !strings.HasPrefix(st.name, "v.") && st.dtype == "BF16" && kind != tensorKindFP32 {
+	if st.dtype == "BF16" && kind != tensorKindFP32 {
 		kind = tensorKindBF16
 	}
 

convert/reader_test.go

@@ -230,65 +230,3 @@ func TestSafetensors(t *testing.T) {
 		})
 	}
 }
-
-func TestSafetensorKind(t *testing.T) {
-	tests := []struct {
-		name     string
-		st       safetensor
-		expected uint32
-	}{
-		{
-			name: "BF16 dtype with non-v. prefix and non-FP32 base kind should return BF16",
-			st: safetensor{
-				tensorBase: &tensorBase{
-					name:  "weight.matrix",
-					shape: []uint64{10, 10}, // will default to FP16
-				},
-				dtype: "BF16",
-			},
-			expected: tensorKindBF16,
-		},
-		{
-			name: "BF16 dtype with v. prefix should return base kind",
-			st: safetensor{
-				tensorBase: &tensorBase{
-					name:  "v.weight.matrix",
-					shape: []uint64{10, 10}, // will default to FP16
-				},
-				dtype: "BF16",
-			},
-			expected: tensorKindFP16,
-		},
-		{
-			name: "BF16 dtype with FP32 base kind should return FP32",
-			st: safetensor{
-				tensorBase: &tensorBase{
-					name:  "weight.matrix",
-					shape: []uint64{10}, // will default to FP32
-				},
-				dtype: "BF16",
-			},
-			expected: tensorKindFP32,
-		},
-		{
-			name: "Non-BF16 dtype should return base kind",
-			st: safetensor{
-				tensorBase: &tensorBase{
-					name:  "weight.matrix",
-					shape: []uint64{10, 10}, // will default to FP16
-				},
-				dtype: "FP16",
-			},
-			expected: tensorKindFP16,
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			result := tt.st.Kind()
-			if result != tt.expected {
-				t.Errorf("Kind() = %d, expected %d", result, tt.expected)
-			}
-		})
-	}
-}

discover/cuda_common.go

@@ -16,7 +16,7 @@ import (
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
 
-func cudaVariant(gpuInfos []CudaGPUInfo) string {
+func cudaVariant(gpuInfo CudaGPUInfo) string {
 	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
 		if CudaTegra != "" {
 			ver := strings.Split(CudaTegra, ".")
@@ -45,19 +45,12 @@ func cudaVariant(gpuInfos []CudaGPUInfo) string {
 		}
 	}
 
-	// Check GPU compute capability FIRST, lowest common denominator if multi-gpu
-	for _, gpuInfo := range gpuInfos {
-		if gpuInfo.computeMajor < 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor < 5) {
-			// GPU is Pascal or older (CC <= 7.4) - use CUDA v12 (supports CC 6.1)
-			return "v12"
-		}
-	}
-
-	// GPU is Turing or newer (CC >= 7.5) - can use newer CUDA
-	if len(gpuInfos) > 0 && gpuInfos[0].DriverMajor < 13 {
+	if gpuInfo.DriverMajor < 13 {
 		// The detected driver is older than 580 (Aug 2025)
 		// Warn if their CC is compatible with v13 and they should upgrade their driver to get better performance
-		slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfos[0].DriverMajor, gpuInfos[0].DriverMinor))
+		if gpuInfo.computeMajor > 7 || (gpuInfo.computeMajor == 7 && gpuInfo.computeMinor >= 5) {
+			slog.Warn("old CUDA driver detected - please upgrade to a newer driver for best performance", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
+		}
 		return "v12"
 	}
 	return "v13"

discover/gpu.go

@@ -284,8 +284,18 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.MinimumMemory = cudaMinimumMemory
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
+				variant := cudaVariant(gpuInfo)
 
+				// Start with our bundled libraries
+				if variant != "" {
+					variantPath := filepath.Join(LibOllamaPath, "cuda_"+variant)
+					if _, err := os.Stat(variantPath); err == nil {
+						// Put the variant directory first in the search path to avoid runtime linking to the wrong library
+						gpuInfo.DependencyPath = append([]string{variantPath}, gpuInfo.DependencyPath...)
+					}
+				}
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+				gpuInfo.Variant = variant
 
 				if int(memInfo.major) < cudaComputeMajorMin || (int(memInfo.major) == cudaComputeMajorMin && int(memInfo.minor) < cudaComputeMinorMin) {
 					unsupportedGPUs = append(unsupportedGPUs,
@@ -323,24 +333,6 @@ func GetGPUInfo() GpuInfoList {
 				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
 				cudaGPUs = append(cudaGPUs, gpuInfo)
 			}
-			// Second pass on NVIDIA GPUs to set lowest common denominator variant and DependencyPaths
-			variant := cudaVariant(cudaGPUs)
-			var variantPath string
-			// Start with our bundled libraries
-			if variant != "" {
-				variantPath = filepath.Join(LibOllamaPath, "cuda_"+variant)
-				if _, err := os.Stat(variantPath); err != nil {
-					variantPath = ""
-				}
-			}
-
-			for i := range cudaGPUs {
-				cudaGPUs[i].Variant = variant
-				if variantPath != "" {
-					// Put the variant directory first in the search path to avoid runtime linking to the wrong library
-					cudaGPUs[i].DependencyPath = append([]string{variantPath}, cudaGPUs[i].DependencyPath...)
-				}
-			}
 		}
 
 		// Intel

docs/cloud.md

@@ -1,40 +0,0 @@
-# Cloud
-
-| Ollama's cloud is currently in preview. For full documentation, see [Ollama's documentation](https://docs.ollama.com/cloud).
-
-## Cloud Models
-
-[Cloud models](https://ollama.com/cloud) are a new kind of model in Ollama that can run without a powerful GPU. Instead, cloud models are automatically offloaded to Ollama's cloud while offering the same capabilities as local models, making it possible to keep using your local tools while running larger models that wouldn’t fit on a personal computer.
-
-Ollama currently supports the following cloud models, with more coming soon:
-
-- `gpt-oss:20b-cloud`
-- `gpt-oss:120b-cloud`
-- `deepseek-v3.1:671b-cloud`
-- `qwen3-coder:480b-cloud`
-
-### Get started
-
-To run a cloud model, open the terminal and run:
-
-```
-ollama run gpt-oss:120b-cloud
-```
-
-To run cloud models with integrations that work with Ollama, first download the cloud model:
-
-```
-ollama pull qwen3-coder:480b-cloud
-```
-
-Then sign in to Ollama:
-
-```
-ollama signin
-```
-
-Finally, access the model using the model name `qwen3-coder:480b-cloud` via Ollama's local API or tooling.
-
-## Cloud API access
-
-Cloud models can also be accessed directly on ollama.com's API. For more information, see the [docs](https://docs.ollama.com/cloud).

docs/turbo.md

@@ -0,0 +1,107 @@
+# Turbo
+
+> ⚠️ Turbo is preview
+
+Ollama’s [Turbo](https://ollama.com/turbo) is a new way to run open-source models with acceleration from datacenter-grade hardware.
+
+Currently, the following models are available in Turbo:
+
+- `gpt-oss:20b`
+- `gpt-oss:120b`
+
+## Get started
+
+### Ollama for macOS & Windows
+
+Download Ollama
+
+- Select a model such as `gpt-oss:20b` or `gpt-oss:120b`
+- Click on **Turbo**. You’ll be prompted to create an account or sign in
+
+### Ollama’s CLI
+
+- [Sign up](https://ollama.com/signup) for an Ollama account
+- Add your Ollama key [to ollama.com](https://ollama.com/settings/keys).
+
+  On macOS and Linux:
+
+  ```shell
+  cat ~/.ollama/id_ed25519.pub
+  ```
+
+  On Windows:
+
+  ```
+  type "%USERPROFILE%\.ollama\id_ed25519.pub"
+  ```
+
+- Then run a model setting `OLLAMA_HOST` to `ollama.com`:
+  ```shell
+  OLLAMA_HOST=ollama.com ollama run gpt-oss:120b
+  ```
+
+### Ollama’s Python library
+
+- Download Ollama's [Python library](https://github.com/ollama/ollama-python)
+- [Sign up](https://ollama.com/signup) for an Ollama account
+- Create an API key by visiting https://ollama.com/settings/keys
+
+```python
+from ollama import Client
+
+client = Client(
+    host="https://ollama.com",
+    headers={'Authorization': '<api key>'}
+)
+
+messages = [
+  {
+    'role': 'user',
+    'content': 'Why is the sky blue?',
+  },
+]
+
+for part in client.chat('gpt-oss:120b', messages=messages, stream=True):
+  print(part['message']['content'], end='', flush=True)
+```
+
+### Ollama’s JavaScript library
+
+- Download Ollama's [JavaScript library](https://github.com/ollama/ollama-js)
+- [Sign up](https://ollama.com/signup) for an Ollama account
+- Create an API key by visiting https://ollama.com/settings/keys
+
+```typescript
+import { Ollama } from 'ollama';
+
+const ollama = new Ollama({
+  host: 'https://ollama.com',
+  headers: {
+	  Authorization: "Bearer <api key>"
+  }
+});
+
+const response = await ollama.chat({
+  model: 'gpt-oss:120b',
+  messages: [{ role: 'user', content: 'Explain quantum computing' }],
+  stream: true
+});
+
+for await (const part of response) {
+    process.stdout.write(part.message.content)
+}
+```
+
+### Community integrations
+
+Turbo mode is also compatible with several community integrations.
+
+#### Open WebUI
+
+- Go to **settings** → **Admin settings** → **Connections**
+- Under **Ollama API,** click **+**
+- For the **URL** put `https://ollama.com`
+- For the **API key,** create an API key on https://ollama.com/settings/keys and add it.
+- Click **Save**
+
+Now, if you navigate to the model selector, Turbo models should be available under **External**.

envconfig/config.go

@@ -134,17 +134,6 @@ func LoadTimeout() (loadTimeout time.Duration) {
 	return loadTimeout
 }
 
-func Remotes() []string {
-	var r []string
-	raw := strings.TrimSpace(Var("OLLAMA_REMOTES"))
-	if raw == "" {
-		r = []string{"ollama.com"}
-	} else {
-		r = strings.Split(raw, ",")
-	}
-	return r
-}
-
 func Bool(k string) func() bool {
 	return func() bool {
 		if s := Var(k); s != "" {
@@ -281,7 +270,6 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
 		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
-		"OLLAMA_REMOTES":           {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},
 
 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},

fs/ggml/ggml.go

@@ -243,8 +243,6 @@ func (kv KV) OllamaEngineRequired() bool {
 		"gemma3",
 		"gemma3n",
 		"mistral3",
-		"qwen3",
-		"qwen3moe",
 		"llama4",
 		"mllama",
 		"qwen25vl",

harmony/harmonyparser.go

@@ -1,7 +1,6 @@
 package harmony
 
 import (
-	"encoding/json"
 	"fmt"
 	"log/slog"
 	"strings"
@@ -266,8 +265,6 @@ type HarmonyMessageHandler struct {
 	state           harmonyMessageState
 	HarmonyParser   *HarmonyParser
 	FunctionNameMap *FunctionNameMap
-	toolAccumulator *HarmonyToolCallAccumulator
-	convertedTools  map[string]struct{}
 }
 
 // NewHarmonyMessageHandler creates a new message handler
@@ -280,7 +277,6 @@ func NewHarmonyMessageHandler() *HarmonyMessageHandler {
 			HeaderEndTag:    "<|message|>",
 		},
 		FunctionNameMap: NewFunctionNameMap(),
-		convertedTools:  make(map[string]struct{}),
 	}
 }
 
@@ -388,85 +384,8 @@ func NewFunctionNameMap() *FunctionNameMap {
 	}
 }
 
-// Init initializes the handler with tools and optional last message
-// Implements the Parser interface
-func (h *HarmonyMessageHandler) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
-	// Initialize the harmony parser
-	if h.HarmonyParser == nil {
-		h.HarmonyParser = &HarmonyParser{
-			MessageStartTag: "<|start|>",
-			MessageEndTag:   "<|end|>",
-			HeaderEndTag:    "<|message|>",
-		}
-	}
-
-	// Handle prefill for chat mode
-	if lastMessage != nil {
-		h.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
-	} else {
-		h.HarmonyParser.AddImplicitStart()
-	}
-
-	// Initialize tool accumulator
-	h.toolAccumulator = h.CreateToolParser()
-
-	// Process tools and return renamed versions
-	if len(tools) == 0 {
-		return tools
-	}
-
-	processedTools := make([]api.Tool, len(tools))
-	copy(processedTools, tools)
-	for i, tool := range processedTools {
-		if tool.Function.Name != "" {
-			processedTools[i].Function.Name = h.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
-			h.convertedTools[tool.Function.Name] = struct{}{}
-		}
-	}
-	return processedTools
-}
-
-// Add implements the Parser interface - processes streamed content and extracts content, thinking, and tool calls
-func (h *HarmonyMessageHandler) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	content, thinking, toolContent := h.AddContent(s, h.toolAccumulator)
-	if toolContent != "" {
-		h.toolAccumulator.Add(toolContent)
-	}
-
-	// tool calls always happen one at a time, and always at the end of a message,
-	// so for simplicity we defer parsing them until we know we're done
-	if done {
-		toolName, raw := h.toolAccumulator.Drain()
-		if toolName != nil {
-			name := strings.TrimPrefix(*toolName, "functions.")
-			name = h.FunctionNameMap.OriginalFromConverted(name)
-			var args api.ToolCallFunctionArguments
-			if err := json.Unmarshal([]byte(raw), &args); err != nil {
-				return "", "", nil, fmt.Errorf("error parsing tool call: raw='%s', err=%w", raw, err)
-			}
-			calls = append(calls, api.ToolCall{Function: api.ToolCallFunction{Name: name, Arguments: args}})
-		}
-	}
-
-	return content, thinking, calls, nil
-}
-
-// HasToolSupport implements the Parser interface
-func (h *HarmonyMessageHandler) HasToolSupport() bool {
-	return true
-}
-
-// HasThinkingSupport implements the Parser interface
-func (h *HarmonyMessageHandler) HasThinkingSupport() bool {
-	return true
-}
-
 func (m *FunctionNameMap) ConvertAndAdd(userFunctionName string) string {
 	harmonyFunctionName := m.deriveName(userFunctionName)
-	// built-in functions should not be renamed
-	if userFunctionName == "browser.open" || userFunctionName == "browser.search" || userFunctionName == "browser.find" || userFunctionName == "python" {
-		harmonyFunctionName = userFunctionName
-	}
 	m.userToHarmony[userFunctionName] = harmonyFunctionName
 	m.harmonyToUser[harmonyFunctionName] = userFunctionName
 	return harmonyFunctionName

harmony/harmonyparser_test.go

@@ -513,7 +513,6 @@ func TestFunctionConvertAndAdd(t *testing.T) {
 		{name: "dupes from different user-specified names", in: []string{"get weather", "get_weather", "get-weather"}, want: []string{"get_weather", "get_weather_2", "get_weather_3"}},
 		{name: "non dupes after dupes", in: []string{"get weather", "get_weather", "get-weather", "something-different"}, want: []string{"get_weather", "get_weather_2", "get_weather_3", "something_different"}},
 		{name: "multiple sets of dupes", in: []string{"a", "a", "b", "a", "a", "b", "a"}, want: []string{"a", "a_2", "b", "a_3", "a_4", "b_2", "a_5"}},
-		{name: "built-in functions should not be renamed", in: []string{"browser.open", "python", "not.a.built-in.function", "browser.not_a_real_built_in"}, want: []string{"browser.open", "python", "not_a_built_in_function", "browser_not_a_real_built_in"}},
 	}
 
 	for i, tt := range tests {

integration/README.md

@@ -12,6 +12,3 @@ The integration tests have 2 modes of operating.
 
 > [!IMPORTANT]
 > Before running the tests locally without the "test existing" setting, compile ollama from the top of the source tree  `go build .` in addition to GPU support with cmake if applicable on your platform.  The integration tests expect to find an ollama binary at the top of the tree.
-
-
-Many tests use a default small model suitable to run on many systems.  You can override this default model by setting `OLLAMA_TEST_DEFAULT_MODEL`

integration/api_test.go

@@ -22,12 +22,13 @@ func TestAPIGenerate(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: blueSkyPrompt,
+		Prompt: "why is the sky blue? be brief",
 		Options: map[string]interface{}{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
+	anyResp := []string{"rayleigh", "scattering"}
 
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
@@ -119,14 +120,14 @@ func TestAPIGenerate(t *testing.T) {
 				// Verify the response contains the expected data
 				response := buf.String()
 				atLeastOne := false
-				for _, resp := range blueSkyExpected {
+				for _, resp := range anyResp {
 					if strings.Contains(strings.ToLower(response), resp) {
 						atLeastOne = true
 						break
 					}
 				}
 				if !atLeastOne {
-					t.Errorf("none of %v found in %s", blueSkyExpected, response)
+					t.Errorf("none of %v found in %s", anyResp, response)
 				}
 			case <-ctx.Done():
 				t.Error("outer test context done while waiting for generate")
@@ -180,7 +181,7 @@ func TestAPIChat(t *testing.T) {
 		Messages: []api.Message{
 			{
 				Role:    "user",
-				Content: blueSkyPrompt,
+				Content: "why is the sky blue?  be brief",
 			},
 		},
 		Options: map[string]interface{}{
@@ -188,6 +189,7 @@ func TestAPIChat(t *testing.T) {
 			"seed":        123,
 		},
 	}
+	anyResp := []string{"rayleigh", "scattering"}
 
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
@@ -277,14 +279,14 @@ func TestAPIChat(t *testing.T) {
 				// Verify the response contains the expected data
 				response := buf.String()
 				atLeastOne := false
-				for _, resp := range blueSkyExpected {
+				for _, resp := range anyResp {
 					if strings.Contains(strings.ToLower(response), resp) {
 						atLeastOne = true
 						break
 					}
 				}
 				if !atLeastOne {
-					t.Errorf("none of %v found in %s", blueSkyExpected, response)
+					t.Errorf("none of %v found in %s", anyResp, response)
 				}
 			case <-ctx.Done():
 				t.Error("outer test context done while waiting for chat")

integration/basic_test.go

@@ -19,14 +19,14 @@ func TestBlueSky(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: blueSkyPrompt,
+		Prompt: "why is the sky blue?",
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, blueSkyExpected)
+	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
 }
 
 func TestUnicode(t *testing.T) {
@@ -110,12 +110,12 @@ func TestUnicodeModelDir(t *testing.T) {
 
 	req := api.GenerateRequest{
 		Model:  smol,
-		Prompt: blueSkyPrompt,
+		Prompt: "why is the sky blue?",
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, blueSkyExpected)
+	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
 }

integration/context_test.go

@@ -63,11 +63,11 @@ func TestContextExhaustion(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived", "sunny", "cloudy", "clear", "water", "time", "travel", "world"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived", "sunny", "cloudy", "clear", "water"}, 120*time.Second, 10*time.Second)
 }
 
 // Send multiple generate requests with prior context and ensure the response is coherant and expected
-func TestParallelGenerateWithHistory(t *testing.T) {
+func TestGenerateWithHistory(t *testing.T) {
 	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
 	req, resp := GenerateRequests()
 	numParallel := 2
@@ -113,48 +113,8 @@ func TestParallelGenerateWithHistory(t *testing.T) {
 	wg.Wait()
 }
 
-// Send generate requests with prior context and ensure the response is coherant and expected
-func TestGenerateWithHistory(t *testing.T) {
-	req := api.GenerateRequest{
-		Model:     smol,
-		Prompt:    rainbowPrompt,
-		Stream:    &stream,
-		KeepAlive: &api.Duration{Duration: 10 * time.Second},
-		Options: map[string]any{
-			"num_ctx": 16384,
-		},
-	}
-
-	softTimeout, hardTimeout := getTimeouts(t)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	// Get the server running (if applicable) warm the model up with a single initial request
-	slog.Info("loading", "model", req.Model)
-	err := client.Generate(ctx,
-		&api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 10 * time.Second}, Options: req.Options},
-		func(response api.GenerateResponse) error { return nil },
-	)
-	if err != nil {
-		t.Fatalf("failed to load model %s: %s", req.Model, err)
-	}
-
-	req.Context = DoGenerate(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
-
-	for i := 0; i < len(rainbowFollowups); i++ {
-		req.Prompt = rainbowFollowups[i]
-		if time.Now().Sub(started) > softTimeout {
-			slog.Info("exceeded soft timeout, winding down test")
-			return
-		}
-		req.Context = DoGenerate(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
-	}
-}
-
 // Send multiple chat requests with prior context and ensure the response is coherant and expected
-func TestParallelChatWithHistory(t *testing.T) {
+func TestChatWithHistory(t *testing.T) {
 	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
 	req, resp := ChatRequests()
 	numParallel := 2
@@ -204,55 +164,3 @@ func TestParallelChatWithHistory(t *testing.T) {
 	}
 	wg.Wait()
 }
-
-// Send generate requests with prior context and ensure the response is coherant and expected
-func TestChatWithHistory(t *testing.T) {
-	req := api.ChatRequest{
-		Model:     smol,
-		Stream:    &stream,
-		KeepAlive: &api.Duration{Duration: 10 * time.Second},
-		Options: map[string]any{
-			"num_ctx": 16384,
-		},
-		Messages: []api.Message{
-			{
-				Role:    "user",
-				Content: rainbowPrompt,
-			},
-		},
-	}
-
-	softTimeout, hardTimeout := getTimeouts(t)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	// Get the server running (if applicable) warm the model up with a single initial request
-	slog.Info("loading", "model", req.Model)
-	err := client.Generate(ctx,
-		&api.GenerateRequest{Model: req.Model, KeepAlive: &api.Duration{Duration: 10 * time.Second}, Options: req.Options},
-		func(response api.GenerateResponse) error { return nil },
-	)
-	if err != nil {
-		t.Fatalf("failed to load model %s: %s", req.Model, err)
-	}
-
-	assistant := DoChat(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
-
-	for i := 0; i < len(rainbowFollowups); i++ {
-		if time.Now().Sub(started) > softTimeout {
-			slog.Info("exceeded soft timeout, winding down test")
-			return
-		}
-		req.Messages = append(req.Messages,
-			*assistant,
-			api.Message{Role: "user", Content: rainbowFollowups[i]},
-		)
-
-		assistant = DoChat(ctx, t, client, req, rainbowExpected, 30*time.Second, 20*time.Second)
-		if assistant == nil {
-			t.Fatalf("didn't get an assistant response for context")
-		}
-	}
-}

integration/embed_test.go

@@ -8,7 +8,6 @@ import (
 	"testing"
 	"time"
 
-	"github.com/google/go-cmp/cmp"
 	"github.com/ollama/ollama/api"
 )
 
@@ -45,8 +44,9 @@ func TestAllMiniLMEmbeddings(t *testing.T) {
 	}
 
 	res, err := embeddingTestHelper(ctx, client, t, req)
+
 	if err != nil {
-		t.Fatal(err)
+		t.Fatalf("error: %v", err)
 	}
 
 	if len(res.Embedding) != 384 {
@@ -74,8 +74,9 @@ func TestAllMiniLMEmbed(t *testing.T) {
 	}
 
 	res, err := embedTestHelper(ctx, client, t, req)
+
 	if err != nil {
-		t.Fatal(err)
+		t.Fatalf("error: %v", err)
 	}
 
 	if len(res.Embeddings) != 1 {
@@ -111,8 +112,9 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 	}
 
 	res, err := embedTestHelper(ctx, client, t, req)
+
 	if err != nil {
-		t.Fatal(err)
+		t.Fatalf("error: %v", err)
 	}
 
 	if len(res.Embeddings) != 2 {
@@ -154,135 +156,93 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 
 	truncTrue, truncFalse := true, false
 
-	want, err := embedTestHelper(ctx, client, t, api.EmbedRequest{
-		Model: "all-minilm",
-		Input: "why",
-	})
-	if err != nil {
-		t.Fatal(err)
+	type testReq struct {
+		Name    string
+		Request api.EmbedRequest
 	}
 
-	cases := []struct {
-		name    string
-		request api.EmbedRequest
-		check   func(*api.EmbedResponse, error)
-	}{
+	reqs := []testReq{
 		{
-			name: "target truncation",
-			request: api.EmbedRequest{
+			Name: "Target Truncation",
+			Request: api.EmbedRequest{
 				Model: "all-minilm",
 				Input: "why",
 			},
-			check: func(got *api.EmbedResponse, err error) {
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
-					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
-				}
-			},
 		},
 		{
-			name: "default truncate",
-			request: api.EmbedRequest{
+			Name: "Default Truncate",
+			Request: api.EmbedRequest{
 				Model:   "all-minilm",
 				Input:   "why is the sky blue?",
-				Options: map[string]any{"num_ctx": 3},
-			},
-			check: func(got *api.EmbedResponse, err error) {
-				if err != nil {
-					t.Fatal(err)
-				}
-				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
-					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
-				}
+				Options: map[string]any{"num_ctx": 1},
 			},
 		},
 		{
-			name: "explicit truncate",
-			request: api.EmbedRequest{
-				Model:    "all-minilm",
-				Input:    "why is the sky blue?",
-				Truncate: &truncTrue,
-				Options:  map[string]any{"num_ctx": 3},
-			},
-			check: func(got *api.EmbedResponse, err error) {
-				if err != nil {
-					t.Fatal(err)
-				}
-				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
-					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
-				}
-			},
-		},
-		{
-			name: "truncate error",
-			request: api.EmbedRequest{
-				Model:    "all-minilm",
-				Input:    "why is the sky blue?",
-				Truncate: &truncFalse,
-				Options:  map[string]any{"num_ctx": 3},
-			},
-			check: func(res *api.EmbedResponse, err error) {
-				if err.Error() != "input exceeds maximum context length" {
-					t.Fatalf("expected truncation error, got: %v", err)
-				}
-			},
-		},
-		{
-			name: "input after truncate error",
-			request: api.EmbedRequest{
+			Name: "Explicit Truncate",
+			Request: api.EmbedRequest{
 				Model:    "all-minilm",
 				Input:    "why is the sky blue?",
 				Truncate: &truncTrue,
 				Options:  map[string]any{"num_ctx": 1},
 			},
-			check: func(res *api.EmbedResponse, err error) {
-				if err.Error() != "input after truncation exceeds maximum context length" {
-					t.Fatalf("expected truncation error, got: %v", err)
-				}
-			},
-		},
-		{
-			name: "input after truncate error",
-			request: api.EmbedRequest{
-				Model:    "all-minilm",
-				Input:    "why is the sky blue?",
-				Truncate: &truncTrue,
-				Options:  map[string]any{"num_ctx": 0},
-			},
-			check: func(res *api.EmbedResponse, err error) {
-				if err.Error() != "input after truncation exceeds maximum context length" {
-					t.Fatalf("expected truncation error, got: %v", err)
-				}
-			},
 		},
 	}
 
-	for _, req := range cases {
-		t.Run(req.name, func(t *testing.T) {
-			req.check(embedTestHelper(ctx, client, t, req.request))
-		})
+	res := make(map[string]*api.EmbedResponse)
+
+	for _, req := range reqs {
+		response, err := embedTestHelper(ctx, client, t, req.Request)
+		if err != nil {
+			t.Fatalf("error: %v", err)
+		}
+		res[req.Name] = response
+	}
+
+	if res["Target Truncation"].Embeddings[0][0] != res["Default Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request to truncate correctly")
+	}
+
+	if res["Default Truncate"].Embeddings[0][0] != res["Explicit Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request and truncate true request to be the same")
+	}
+
+	// check that truncate set to false returns an error if context length is exceeded
+	_, err := embedTestHelper(ctx, client, t, api.EmbedRequest{
+		Model:    "all-minilm",
+		Input:    "why is the sky blue?",
+		Truncate: &truncFalse,
+		Options:  map[string]any{"num_ctx": 1},
+	})
+
+	if err == nil {
+		t.Fatal("expected error, got nil")
 	}
 }
 
 func embeddingTestHelper(ctx context.Context, client *api.Client, t *testing.T, req api.EmbeddingRequest) (*api.EmbeddingResponse, error) {
-	t.Helper()
-
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatal(err)
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
 	}
 
-	return client.Embeddings(ctx, &req)
+	response, err := client.Embeddings(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
 }
 
 func embedTestHelper(ctx context.Context, client *api.Client, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
-	t.Helper()
-
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatal(err)
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
 	}
 
-	return client.Embed(ctx, &req)
+	response, err := client.Embed(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
 }

integration/library_models_test.go

@@ -4,9 +4,7 @@ package integration
 
 import (
 	"context"
-	"fmt"
 	"log/slog"
-	"os"
 	"testing"
 	"time"
 
@@ -22,7 +20,6 @@ func TestLibraryModelsGenerate(t *testing.T) {
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	targetArch := os.Getenv("OLLAMA_TEST_ARCHITECTURE")
 
 	chatModels := libraryChatModels
 	for _, model := range chatModels {
@@ -33,26 +30,16 @@ func TestLibraryModelsGenerate(t *testing.T) {
 			if err := PullIfMissing(ctx, client, model); err != nil {
 				t.Fatalf("pull failed %s", err)
 			}
-			if targetArch != "" {
-				resp, err := client.Show(ctx, &api.ShowRequest{Name: model})
-				if err != nil {
-					t.Fatalf("unable to show model: %s", err)
-				}
-				arch := resp.ModelInfo["general.architecture"].(string)
-				if arch != targetArch {
-					t.Skip(fmt.Sprintf("Skipping %s architecture %s != %s", model, arch, targetArch))
-				}
-			}
 			req := api.GenerateRequest{
 				Model:     model,
-				Prompt:    blueSkyPrompt,
+				Prompt:    "why is the sky blue?",
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
 				Options: map[string]interface{}{
 					"temperature": 0.1,
 					"seed":        123,
 				},
 			}
-			anyResp := blueSkyExpected
+			anyResp := []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength"}
 			// Special cases
 			if model == "duckdb-nsql" {
 				anyResp = []string{"select", "from"}

integration/model_arch_test.go

@@ -68,13 +68,14 @@ func TestModelsGenerate(t *testing.T) {
 			// TODO - fiddle with context size
 			req := api.GenerateRequest{
 				Model:  model,
-				Prompt: blueSkyPrompt,
+				Prompt: "why is the sky blue?",
 				Options: map[string]interface{}{
 					"temperature": 0,
 					"seed":        123,
 				},
 			}
-			DoGenerate(ctx, t, client, req, blueSkyExpected, 120*time.Second, 30*time.Second)
+			anyResp := []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}
+			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
 		})
 	}
 }

integration/model_perf_test.go

@@ -40,18 +40,6 @@ var (
 // cat int.log | grep MODEL_PERF_HEADER | head -1| cut -f2- -d: > perf.csv
 // cat int.log | grep MODEL_PERF_DATA | cut -f2- -d: >> perf.csv
 func TestModelsPerf(t *testing.T) {
-	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
-		doModelPerfTest(t, ollamaEngineChatModels)
-	} else {
-		doModelPerfTest(t, append(ollamaEngineChatModels, llamaRunnerChatModels...))
-	}
-}
-
-func TestLibraryModelsPerf(t *testing.T) {
-	doModelPerfTest(t, libraryChatModels)
-}
-
-func doModelPerfTest(t *testing.T, chatModels []string) {
 	softTimeout, hardTimeout := getTimeouts(t)
 	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
 	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
@@ -77,12 +65,14 @@ func doModelPerfTest(t *testing.T, chatModels []string) {
 	}
 	longPrompt := "summarize the following: " + string(data)
 
-	targetArch := os.Getenv("OLLAMA_TEST_ARCHITECTURE")
+	var chatModels []string
+	if s := os.Getenv("OLLAMA_NEW_ENGINE"); s != "" {
+		chatModels = ollamaEngineChatModels
+	} else {
+		chatModels = append(ollamaEngineChatModels, llamaRunnerChatModels...)
+	}
 
 	for _, model := range chatModels {
-		if !strings.Contains(model, ":") {
-			model = model + ":latest"
-		}
 		t.Run(model, func(t *testing.T) {
 			if time.Now().Sub(started) > softTimeout {
 				t.Skip("skipping remaining tests to avoid excessive runtime")
@@ -98,9 +88,6 @@ func doModelPerfTest(t *testing.T, chatModels []string) {
 			}
 			arch := resp.ModelInfo["general.architecture"].(string)
 			maxContext = int(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))
-			if targetArch != "" && arch != targetArch {
-				t.Skip(fmt.Sprintf("Skipping %s architecture %s != %s", model, arch, targetArch))
-			}
 
 			if maxVram > 0 {
 				resp, err := client.List(ctx)
@@ -164,8 +151,8 @@ func doModelPerfTest(t *testing.T, chatModels []string) {
 					prompt  string
 					anyResp []string
 				}{
-					{blueSkyPrompt, blueSkyExpected},
-					{maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy", "love", "sorrow", "beauty"}},
+					{"why is the sky blue?", []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}},
+					{maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy"}},
 				}
 				var gpuPercent int
 				for _, tc := range testCases {
@@ -254,12 +241,11 @@ func doModelPerfTest(t *testing.T, chatModels []string) {
 							}
 						}
 					}
-					// Round the logged prompt count for comparisons across versions/configurations which can vary slightly
 					fmt.Fprintf(os.Stderr, "MODEL_PERF_HEADER:%s,%s,%s,%s,%s,%s,%s\n",
 						"MODEL",
 						"CONTEXT",
 						"GPU PERCENT",
-						"APPROX PROMPT COUNT",
+						"PROMPT COUNT",
 						"LOAD TIME",
 						"PROMPT EVAL TPS",
 						"EVAL TPS",
@@ -268,7 +254,7 @@ func doModelPerfTest(t *testing.T, chatModels []string) {
 						model,
 						numCtx,
 						gpuPercent,
-						(resp.PromptEvalCount/10)*10,
+						resp.PromptEvalCount,
 						float64(resp.LoadDuration)/1000000000.0,
 						float64(resp.PromptEvalCount)/(float64(resp.PromptEvalDuration)/1000000000.0),
 						float64(resp.EvalCount)/(float64(resp.EvalDuration)/1000000000.0),

integration/quantization_test.go

@@ -76,7 +76,7 @@ func TestQuantization(t *testing.T) {
 				stream := true
 				genReq := api.GenerateRequest{
 					Model:     newName,
-					Prompt:    blueSkyPrompt,
+					Prompt:    "why is the sky blue?",
 					KeepAlive: &api.Duration{Duration: 3 * time.Second},
 					Options: map[string]any{
 						"seed":        42,
@@ -88,13 +88,14 @@ func TestQuantization(t *testing.T) {
 
 				// Some smaller quantizations can cause models to have poor quality
 				// or get stuck in repetition loops, so we stop as soon as we have any matches
+				anyResp := []string{"rayleigh", "scattering", "day", "sun", "moon", "color", "nitrogen", "oxygen"}
 				reqCtx, reqCancel := context.WithCancel(ctx)
 				atLeastOne := false
 				var buf bytes.Buffer
 				genfn := func(response api.GenerateResponse) error {
 					buf.Write([]byte(response.Response))
 					fullResp := strings.ToLower(buf.String())
-					for _, resp := range blueSkyExpected {
+					for _, resp := range anyResp {
 						if strings.Contains(fullResp, resp) {
 							atLeastOne = true
 							t.Log(fullResp)

integration/utils_test.go

@@ -256,29 +256,13 @@ var (
 		"snowflake-arctic-embed",
 		"snowflake-arctic-embed2",
 	}
-
-	blueSkyPrompt   = "why is the sky blue? Be brief but factual in your reply"
-	blueSkyExpected = []string{"rayleigh", "scatter", "atmosphere", "nitrogen", "oxygen", "wavelength", "interact"}
-
-	rainbowPrompt    = "how do rainbows form? Be brief but factual in your reply"
-	rainbowFollowups = []string{
-		"Explain the physics involved in them.  Be breif in your reply",
-		"Explain the chemistry involved in them.  Be breif in your reply",
-		"Explain the quantum mechanics involved in them. Be breif in your reply",
-		"What are common myths related to them? Be brief in your reply",
-		"What are common fairytales related to them? Be brief in your reply",
-		"Can they form if there is no rain?  Be breif in your reply",
-		"Can they form if there are no clouds?  Be breif in your reply",
-		"Do they happen on other planets? Be brief in your reply",
-	}
-	rainbowExpected = []string{"water", "droplet", "mist", "glow", "refracted", "reflect", "color", "spectrum", "frequency", "end", "gold", "fortune", "blessing", "prosperity"}
 )
 
 func init() {
 	lifecycle.InitLogging()
-	custom := os.Getenv("OLLAMA_TEST_DEFAULT_MODEL")
+	custom := os.Getenv("OLLAMA_TEST_SMOL_MODEL")
 	if custom != "" {
-		slog.Info("setting default test model to " + custom)
+		slog.Info("setting smol test model to " + custom)
 		smol = custom
 	}
 }
@@ -593,11 +577,11 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			},
 		},
 		[][]string{
-			{"sunlight", "scatter", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorb", "wavelength", "water", "molecule"},
-			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigment", "particle", "iron oxide", "rust", "air", "water", "wet", "mixture", "mixing", "mineral", "element", "decomposed", "matter", "wavelength"},
-			{"water", "droplet", "refract", "reflect", "color", "spectrum", "raindrop"},
+			{"sunlight", "scattering", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorbs", "wavelength"},
+			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigments", "particles", "iron oxide", "rust", "air", "water", "mixture", "mixing"},
+			{"water", "droplet", "refracted", "reflect", "color", "spectrum"},
 			{"fourth", "july", "declaration", "independence"},
-			{"nitrogen", "oxygen", "carbon", "dioxide", "water", "vapor", "fluid", "particles", "gas"},
+			{"nitrogen", "oxygen", "carbon", "dioxide", "water", "vapor"},
 		}
 }
 

llama/patches/0014-graph-memory-reporting-on-failure.patch

@@ -4,38 +4,48 @@ Date: Fri, 18 Apr 2025 15:58:19 -0700
 Subject: [PATCH] graph memory reporting on failure
 
 ---
- ggml/include/ggml-alloc.h   |  1 +
- ggml/include/ggml-backend.h |  1 +
- ggml/src/ggml-alloc.c       | 36 ++++++++++++++++++++++++++++++++----
- ggml/src/ggml-backend.cpp   |  7 +++++++
- 4 files changed, 41 insertions(+), 4 deletions(-)
+ 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 2cb150fd2..7ab3f0192 100644
+index 2cb150fd..781b1e10 100644
 --- a/ggml/include/ggml-alloc.h
 +++ b/ggml/include/ggml-alloc.h
-@@ -65,6 +65,7 @@ GGML_API bool ggml_gallocr_reserve_n(
- GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+@@ -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);
-+GGML_API size_t ggml_gallocr_get_attempted_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 a2977ea2e..e8cf30841 100644
+index a2977ea2..8a91b381 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
-@@ -303,6 +303,7 @@ extern "C" {
-     GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
+@@ -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);
-+    GGML_API size_t               ggml_backend_sched_get_attempted_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 8b6e60283..b58bd671d 100644
+index 8b6e6028..41c8c4a2 100644
 --- a/ggml/src/ggml-alloc.c
 +++ b/ggml/src/ggml-alloc.c
 @@ -350,6 +350,7 @@ struct node_alloc {
@@ -98,11 +108,11 @@ index 8b6e60283..b58bd671d 100644
  }
  
  bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-@@ -920,6 +932,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+@@ -920,6 +932,24 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
      return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
  }
  
-+size_t ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int 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++) {
@@ -111,29 +121,34 @@ index 8b6e60283..b58bd671d 100644
 +            // (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.
 +
-+            return 0;
++            struct ggml_allocr_buffer_status status = {0, true};
++            return status;
 +        }
 +    }
 +
-+    return galloc->buffer_sizes[buffer_id];
++    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 97f47abd2..d02a40e60 100644
+index 97f47abd..eded0291 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -1631,6 +1631,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
+@@ -1631,6 +1631,16 @@ 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);
  }
  
-+size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
++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);
 +
-+    return ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
++    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) {

llama/patches/0022-ggml-No-alloc-mode.patch

@@ -3,45 +3,35 @@ From: Jesse Gross <jesse@ollama.com>
 Date: Wed, 23 Jul 2025 11:58:49 -0700
 Subject: [PATCH] ggml: No-alloc mode
 
-Callers can set a scheduler to be no-alloc, meaning that
+Callers can set a backend buffer type to be no-alloc, meaning that
 it does not allocate memory for tensors or operations. This can
 be used for calculating memory requirements. Tensors and graphs
 must be recreated with no-alloc set to false before loading data.
+
+Defaults to false for newly created backend buffer types.
 ---
- ggml/include/ggml-backend.h     |   1 +
- ggml/src/ggml-backend-impl.h    |  16 +++
- ggml/src/ggml-backend.cpp       |  72 ++++++++++-
- ggml/src/ggml-cuda/common.cuh   |  48 ++++++-
- ggml/src/ggml-cuda/ggml-cuda.cu | 217 ++++++++++++++++++++++++++------
- 5 files changed, 310 insertions(+), 44 deletions(-)
+ ggml/include/ggml-backend.h  |  1 +
+ ggml/src/ggml-backend-impl.h |  2 ++
+ ggml/src/ggml-backend.cpp    | 19 ++++++++++++++++++-
+ 3 files changed, 21 insertions(+), 1 deletion(-)
 
 diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index 2773cc310..ae94887dd 100644
+index 9424394e..b602a7c7 100644
 --- a/ggml/include/ggml-backend.h
 +++ b/ggml/include/ggml-backend.h
-@@ -291,6 +291,7 @@ extern "C" {
+@@ -35,6 +35,7 @@ extern "C" {
+     //
  
-     // Initialize a backend scheduler, backends with low index are given priority over backends with high index
-     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
-+    GGML_API ggml_backend_sched_t ggml_backend_sched_new_ext(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload, bool alloc_buffers);
-     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
- 
-     // Initialize backend buffers from a measure graph
+     GGML_API const char *          ggml_backend_buft_name          (ggml_backend_buffer_type_t buft);
++    GGML_API void                  ggml_backend_buft_set_alloc     (ggml_backend_buffer_type_t buft, bool alloc);
+     GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer  (ggml_backend_buffer_type_t buft, size_t size);
+     GGML_API size_t                ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
+     GGML_API size_t                ggml_backend_buft_get_max_size  (ggml_backend_buffer_type_t buft);
 diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
-index c36c12d65..369e9e25a 100644
+index c36c12d6..81749a5a 100644
 --- a/ggml/src/ggml-backend-impl.h
 +++ b/ggml/src/ggml-backend-impl.h
-@@ -26,12 +26,17 @@ extern "C" {
-         size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
-         // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
-         bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
-+
-+        // (optional) returns a dummy buffer that is equivalent to one created by alloc_buffer but without actually being backed
-+        // by memory
-+        ggml_backend_buffer_t (*noalloc_buffer)(ggml_backend_buffer_type_t buft, size_t size);
-     };
- 
-     struct ggml_backend_buffer_type {
+@@ -32,6 +32,7 @@ extern "C" {
          struct ggml_backend_buffer_type_i  iface;
          ggml_backend_dev_t device;
          void * context;
@@ -49,7 +39,7 @@ index c36c12d65..369e9e25a 100644
      };
  
      //
-@@ -63,6 +68,7 @@ extern "C" {
+@@ -63,6 +64,7 @@ extern "C" {
          void * context;
          size_t size;
          enum ggml_backend_buffer_usage usage;
@@ -57,40 +47,26 @@ index c36c12d65..369e9e25a 100644
      };
  
      GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
-@@ -114,6 +120,16 @@ extern "C" {
-         void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
-         // wait for an event on on a different stream
-         void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
-+
-+        // (optional) reserves intermediate buffers needed for the compution
-+        // if alloc is true, memory is actually allocated, otherwise the required amount is just returned by buffer_size
-+        enum ggml_status          (*graph_reserve)     (ggml_backend_t backend, struct ggml_cgraph * cgraph, bool alloc);
-+
-+        // (optional) returns the memory needed after calling graph_reserve
-+        size_t                    (*buffer_size)       (ggml_backend_t backend);
-+
-+        // (optional) frees memory from intermediate buffers that was allocated either by graph_compute or graph_reserve
-+        void                      (*reset)             (ggml_backend_t backend);
-     };
- 
-     struct ggml_backend {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index d02a40e60..6b4dee4c7 100644
+index eded0291..05a842ed 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -41,6 +41,19 @@ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t
+@@ -35,12 +35,22 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
+     return buft->iface.get_name(buft);
+ }
+ 
++void ggml_backend_buft_set_alloc(ggml_backend_buffer_type_t buft, bool alloc) {
++    buft->no_alloc = !alloc;
++}
++
+ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+     if (size == 0) {
+         // return a dummy buffer for zero-sized allocations
          return ggml_backend_buffer_init(buft, {}, NULL, 0);
      }
  
 +    if (buft->no_alloc) {
-+        ggml_backend_buffer_t buf;
-+
-+        if (buft->iface.noalloc_buffer != NULL) {
-+            buf = buft->iface.noalloc_buffer(buft, size);
-+        } else {
-+            buf = ggml_backend_buffer_init(buft, {}, NULL, size);
-+        }
-+
++        ggml_backend_buffer_t buf = ggml_backend_buffer_init(buft, {}, NULL, size);
 +        buf->no_alloc = true;
 +        return buf;
 +    }
@@ -98,7 +74,7 @@ index d02a40e60..6b4dee4c7 100644
      return buft->iface.alloc_buffer(buft, size);
  }
  
-@@ -89,7 +102,8 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
+@@ -89,7 +99,8 @@ ggml_backend_buffer_t ggml_backend_buffer_init(
          /* .buft      = */ buft,
          /* .context   = */ context,
          /* .size      = */ size,
@@ -108,7 +84,7 @@ index d02a40e60..6b4dee4c7 100644
      };
  
      return buffer;
-@@ -119,6 +133,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+@@ -119,6 +130,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
          return NULL;
      }
  
@@ -121,532 +97,3 @@ index d02a40e60..6b4dee4c7 100644
      void * base = buffer->iface.get_base(buffer);
  
      GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
-@@ -663,6 +683,12 @@ struct ggml_backend_sched {
-     bool op_offload;
- 
-     int debug;
-+
-+    // allocate buffers on attached ggml_backend_buffer_type_t's and during reservation
-+    // if false, dummy buffers are used for faster memory sizing calculations
-+    // the scheduler needs to be recreated with allocated buffers before it can be used
-+    // for computation
-+    bool alloc_buffers;
- };
- 
- #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
-@@ -1449,6 +1475,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
-         size_t graph_size,
-         bool parallel,
-         bool op_offload) {
-+            return ggml_backend_sched_new_ext(backends, bufts, n_backends, graph_size, parallel, op_offload, true);
-+        }
-+
-+ggml_backend_sched_t ggml_backend_sched_new_ext(
-+        ggml_backend_t * backends,
-+        ggml_backend_buffer_type_t * bufts,
-+        int n_backends,
-+        size_t graph_size,
-+        bool parallel,
-+        bool op_offload,
-+        bool alloc_buffers) {
-     GGML_ASSERT(n_backends > 0);
-     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
-     GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
-@@ -1490,10 +1527,13 @@ ggml_backend_sched_t ggml_backend_sched_new(
-                 sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
-             }
-         }
-+
-+        sched->bufts[b]->no_alloc = !alloc_buffers;
-     }
- 
-     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
-     sched->op_offload = op_offload;
-+    sched->alloc_buffers = alloc_buffers;
- 
-     ggml_backend_sched_reset(sched);
- 
-@@ -1508,6 +1548,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
-         for (int c = 0; c < sched->n_copies; c++) {
-             ggml_backend_event_free(sched->events[b][c]);
-         }
-+
-+        if (sched->backends[b]->iface.reset != NULL) {
-+            sched->backends[b]->iface.reset(sched->backends[b]);
-+        }
-     }
-     ggml_gallocr_free(sched->galloc);
-     ggml_free(sched->ctx);
-@@ -1547,6 +1591,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
-         return false;
-     }
- 
-+    if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
-+        return false;
-+    }
-+
-+    struct ggml_backend_sched_split * splits = sched->splits;
-+    for (int i = 0; i < sched->n_splits; i++) {
-+        struct ggml_backend_sched_split * split = &splits[i];
-+        int split_backend_id = split->backend_id;
-+        ggml_backend_t split_backend = sched->backends[split_backend_id];
-+
-+        if (split_backend->iface.graph_reserve != NULL) {
-+            enum ggml_status ec = split_backend->iface.graph_reserve(split_backend, &split->graph, sched->alloc_buffers);
-+            if (ec != GGML_STATUS_SUCCESS) {
-+                return false;
-+            }
-+        }
-+    }
-+
-     ggml_backend_sched_reset(sched);
- 
-     return true;
-@@ -1635,7 +1697,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
-     int backend_index = ggml_backend_sched_backend_id(sched, backend);
-     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
- 
--    return ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
-+    size_t size = ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
-+
-+    if (backend->iface.buffer_size != NULL) {
-+        size += backend->iface.buffer_size(backend);
-+    }
-+
-+    return size;
- }
- 
- void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
-diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
-index 2e5d48797..b915ee1b8 100644
---- a/ggml/src/ggml-cuda/common.cuh
-+++ b/ggml/src/ggml-cuda/common.cuh
-@@ -35,6 +35,31 @@
- #include "vendors/cuda.h"
- #endif // defined(GGML_USE_HIP)
- 
-+extern bool reserving_graph;
-+
-+// If we are reserving the graph, pointers might be invalid and will fail if cudaMemcpyAsync tries to validate them.
-+// However, since we don't actually expect a result, we don't need to actually do the memcpy.
-+static cudaError_t cudaMemcpyAsyncReserve ( void* dst, const void* src, size_t count, cudaMemcpyKind kind, cudaStream_t stream = 0 ) {
-+    if (!reserving_graph) {
-+        return cudaMemcpyAsync(dst, src, count, kind, stream);
-+    } else {
-+        return cudaSuccess;
-+    }
-+}
-+
-+static cudaError_t cudaMemcpy2DAsyncReserve ( void* dst, size_t dpitch, const void* src, size_t spitch, size_t width, size_t height, cudaMemcpyKind kind, cudaStream_t stream = 0 ) {
-+    if (!reserving_graph) {
-+        return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, kind, stream);
-+    } else {
-+        return cudaSuccess;
-+    }
-+}
-+
-+#undef cudaMemcpyAsync
-+#define cudaMemcpyAsync cudaMemcpyAsyncReserve
-+#undef cudaMemcpy2DAsync
-+#define cudaMemcpy2DAsync cudaMemcpy2DAsyncReserve
-+
- #define STRINGIZE_IMPL(...) #__VA_ARGS__
- #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
- 
-@@ -771,6 +796,9 @@ struct ggml_cuda_pool {
- 
-     virtual void * alloc(size_t size, size_t * actual_size) = 0;
-     virtual void free(void * ptr, size_t size) = 0;
-+
-+    virtual bool alloc_memory() = 0;
-+    virtual size_t alloc_size() = 0;
- };
- 
- template<typename T>
-@@ -914,11 +942,11 @@ struct ggml_backend_cuda_context {
-     // pool
-     std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
- 
--    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
-+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, bool alloc);
- 
-     ggml_cuda_pool & pool(int device) {
-         if (pools[device] == nullptr) {
--            pools[device] = new_pool_for_device(device);
-+            pools[device] = new_pool_for_device(device, true);
-         }
-         return *pools[device];
-     }
-@@ -926,4 +954,20 @@ struct ggml_backend_cuda_context {
-     ggml_cuda_pool & pool() {
-         return pool(device);
-     }
-+
-+    void pool_set_alloc(bool alloc) {
-+        GGML_ASSERT(pools[device] == nullptr || pools[device]->alloc_memory() == alloc);
-+
-+        if (pools[device] == nullptr) {
-+            pools[device] = new_pool_for_device(device, alloc);
-+        }
-+    }
-+
-+    size_t pool_get_alloc_size() {
-+        if (pools[device] == nullptr) {
-+            return 0;
-+        }
-+
-+        return pools[device]->alloc_size();
-+    }
- };
-diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index c7f9dc3a5..d5abe09e0 100644
---- a/ggml/src/ggml-cuda/ggml-cuda.cu
-+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -350,6 +350,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {
- 
- // #define DEBUG_CUDA_MALLOC
- 
-+#define CUDA_ALIGNMENT 128
-+
- // buffer pool for cuda (legacy)
- struct ggml_cuda_pool_leg : public ggml_cuda_pool {
-     static const int MAX_BUFFERS = 256;
-@@ -362,9 +364,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
- 
-     ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
-     size_t pool_size = 0;
-+    bool allocate = true;
-+    size_t last_alloc = 0;
- 
--    explicit ggml_cuda_pool_leg(int device) :
--        device(device) {
-+    explicit ggml_cuda_pool_leg(int device, bool alloc) :
-+        device(device),
-+        allocate(alloc) {
-     }
- 
-     ~ggml_cuda_pool_leg() {
-@@ -372,7 +377,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
-         for (int i = 0; i < MAX_BUFFERS; ++i) {
-             ggml_cuda_buffer & b = buffer_pool[i];
-             if (b.ptr != nullptr) {
--                CUDA_CHECK(cudaFree(b.ptr));
-+                if (allocate) {
-+                    CUDA_CHECK(cudaFree(b.ptr));
-+                }
-                 pool_size -= b.size;
-             }
-         }
-@@ -420,8 +427,15 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
-         void * ptr;
-         size_t look_ahead_size = (size_t) (1.05 * size);
-         look_ahead_size = 256 * ((look_ahead_size + 255)/256);
--        ggml_cuda_set_device(device);
--        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
-+        if (allocate) {
-+            ggml_cuda_set_device(device);
-+            if (ggml_cuda_device_malloc(&ptr, look_ahead_size, device) != cudaSuccess) {
-+                    last_alloc = look_ahead_size;
-+                    throw std::bad_alloc();
-+            }
-+        } else {
-+            ptr = (void *)CUDA_ALIGNMENT;
-+        }
-         *actual_size = look_ahead_size;
-         pool_size += look_ahead_size;
- #ifdef DEBUG_CUDA_MALLOC
-@@ -441,10 +455,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
-             }
-         }
-         GGML_LOG_DEBUG(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
--        ggml_cuda_set_device(device);
--        CUDA_CHECK(cudaFree(ptr));
-+        if (allocate) {
-+            ggml_cuda_set_device(device);
-+            CUDA_CHECK(cudaFree(ptr));
-+        }
-         pool_size -= size;
-     }
-+
-+    bool alloc_memory() override {
-+        return allocate;
-+    }
-+
-+    size_t alloc_size() override {
-+        return pool_size + last_alloc;
-+    }
- };
- 
- // pool with virtual memory
-@@ -456,18 +480,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
-     CUdeviceptr pool_addr = 0;
-     size_t pool_used = 0;
-     size_t pool_size = 0;
-+    bool allocate = true;
-+    size_t last_alloc = 0;
-     size_t granularity;
- #if defined(GGML_USE_HIP)
-     std::vector<std::pair<CUdeviceptr, size_t>> mappings;
- #endif
- 
--    explicit ggml_cuda_pool_vmm(int device) :
-+    explicit ggml_cuda_pool_vmm(int device, bool alloc) :
-         device(device),
--        granularity(ggml_cuda_info().devices[device].vmm_granularity) {
-+        granularity(ggml_cuda_info().devices[device].vmm_granularity),
-+        allocate(alloc) {
-+        if (!allocate) {
-+            pool_addr = (CUdeviceptr)CUDA_ALIGNMENT;
-+        }
-     }
- 
-     ~ggml_cuda_pool_vmm() {
--        if (pool_addr != 0) {
-+        if (pool_addr != 0 && allocate) {
- #if defined(GGML_USE_HIP)
-             // Workaround for https://github.com/ROCm/ROCR-Runtime/issues/285
-             for (std::pair<CUdeviceptr, size_t> & mapping : mappings) {
-@@ -494,35 +524,49 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
- 
-             GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
- 
--            // allocate more physical memory
--            CUmemAllocationProp prop = {};
--            prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
--            prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
--            prop.location.id = device;
--            CUmemGenericAllocationHandle handle;
--            CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));
--
--            // reserve virtual address space (if not already reserved)
--            if (pool_addr == 0) {
--                CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
--            }
-+            if (allocate) {
-+                // allocate more physical memory
-+                CUmemAllocationProp prop = {};
-+                prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
-+                prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-+                prop.location.id = device;
-+                CUmemGenericAllocationHandle handle;
-+                if (cuMemCreate(&handle, reserve_size, &prop, 0) != CUDA_SUCCESS) {
-+                    last_alloc = reserve_size;
-+                    throw std::bad_alloc();
-+                }
- 
--            // map at the end of the pool
--            CUdeviceptr start_ptr = (CUdeviceptr)((char *)(pool_addr) + pool_size);
--            CU_CHECK(cuMemMap(start_ptr, reserve_size, 0, handle, 0));
--#if defined(GGML_USE_HIP)
--            mappings.push_back({start_ptr, reserve_size});
--#endif
-+                // reserve virtual address space (if not already reserved)
-+                if (pool_addr == 0) {
-+                    CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
-+                }
- 
--            // the memory allocation handle is no longer needed after mapping
--            CU_CHECK(cuMemRelease(handle));
-+                // map at the end of the pool
-+                CUdeviceptr start_ptr = (CUdeviceptr)((char *)(pool_addr) + pool_size);
-+                if (cuMemMap(start_ptr, reserve_size, 0, handle, 0) != CUDA_SUCCESS) {
-+                    last_alloc = reserve_size;
-+                    CU_CHECK(cuMemRelease(handle));
-+                    throw std::bad_alloc();
-+                }
-+
-+                // the memory allocation handle is no longer needed after mapping
-+                CU_CHECK(cuMemRelease(handle));
-+
-+                // set access
-+                CUmemAccessDesc access = {};
-+                access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-+                access.location.id = device;
-+                access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
-+                if (cuMemSetAccess((CUdeviceptr)((char *)(pool_addr) + pool_size), reserve_size, &access, 1) != CUDA_SUCCESS) {
-+                    CU_CHECK(cuMemUnmap(start_ptr, reserve_size));
-+                    last_alloc = reserve_size;
-+                    throw std::bad_alloc();
-+                }
- 
--            // set access
--            CUmemAccessDesc access = {};
--            access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
--            access.location.id = device;
--            access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
--            CU_CHECK(cuMemSetAccess((CUdeviceptr)((char *)(pool_addr) + pool_size), reserve_size, &access, 1));
-+    #if defined(GGML_USE_HIP)
-+                mappings.push_back({start_ptr, reserve_size});
-+    #endif
-+            }
- 
-             // add to the pool
-             pool_size += reserve_size;
-@@ -555,16 +599,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
-         // all deallocations must be in reverse order of the allocations
-         GGML_ASSERT(ptr == (void *) ((char *)(pool_addr) + pool_used));
-     }
-+
-+    bool alloc_memory() override {
-+        return allocate;
-+    }
-+
-+    size_t alloc_size() override {
-+        return pool_size + last_alloc;
-+    }
- };
- #endif // defined(GGML_USE_VMM)
- 
--std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device, bool alloc) {
- #if defined(GGML_USE_VMM)
-     if (ggml_cuda_info().devices[device].vmm) {
--        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
-+        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device, alloc));
-     }
- #endif // defined(GGML_USE_VMM)
--    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
-+    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device, alloc));
- }
- 
- // destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
-@@ -748,11 +800,20 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
- }
- 
- static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
--    return 128;
-+    return CUDA_ALIGNMENT;
- 
-     GGML_UNUSED(buft);
- }
- 
-+static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_noalloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-+    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
-+
-+    void * dev_ptr = (void *)ggml_backend_cuda_buffer_type_get_alignment(buft);
-+    ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr);
-+
-+    return ggml_backend_buffer_init(buft, {}, ctx, size);
-+}
-+
- static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
-     size_t size = ggml_nbytes(tensor);
-     int64_t ne0 = tensor->ne[0];
-@@ -776,6 +837,7 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
-     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-     /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
-     /* .is_host          = */ NULL,
-+    /* .noalloc_buffer   = */ ggml_backend_cuda_buffer_type_noalloc_buffer,
- };
- 
- ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
-@@ -2936,6 +2998,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
- 
- static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
-     bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
-+
-     // flag used to determine whether it is an integrated_gpu
-     const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
- 
-@@ -2951,6 +3014,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
-                     continue;
-                 }
- 
-+                // When reserving, we are forcing CUDA graphs but this operation is not graph-safe so we need to skip it
-+                if (reserving_graph && node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
-+                    continue;
-+                }
-+
-                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
-                 if (!disable_fusion) {
-                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL }, {})) {
-@@ -3022,6 +3090,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
- 
- static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
-+    cuda_ctx->pool_set_alloc(true);
- 
-     ggml_cuda_set_device(cuda_ctx->device);
- 
-@@ -3101,6 +3170,71 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
-     return GGML_STATUS_SUCCESS;
- }
- 
-+// This is used to skip operations that are not graph safe during the reservation process.
-+bool reserving_graph = false;
-+
-+static enum ggml_status ggml_backend_cuda_graph_reserve(ggml_backend_t backend, ggml_cgraph * cgraph, bool alloc) {
-+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
-+    cuda_ctx->pool_set_alloc(alloc);
-+
-+    #ifdef USE_CUDA_GRAPH
-+    if (cuda_ctx->cuda_graph == nullptr) {
-+        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
-+    }
-+    #endif
-+
-+    ggml_cuda_set_device(cuda_ctx->device);
-+
-+    {
-+        std::lock_guard<std::mutex> lock(ggml_cuda_lock);
-+        ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed);
-+    }
-+
-+    reserving_graph = true;
-+
-+    // Create CuBLAS handles early to avoid synchronous allocations during graph capture.
-+    cuda_ctx->cublas_handle();
-+
-+    CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
-+
-+    enum ggml_status result = GGML_STATUS_SUCCESS;
-+
-+    try {
-+        bool use_cuda_graph = false;
-+        bool cuda_graph_update_required = false;
-+        bool graph_evaluated_or_captured = false;
-+
-+        evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
-+    } catch (const std::exception &e) {
-+        result = GGML_STATUS_FAILED;
-+    }
-+
-+    cudaGraph_t graph;
-+    CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &graph));
-+    CUDA_CHECK(cudaGraphDestroy(graph));
-+
-+    reserving_graph = false;
-+
-+    {
-+        std::lock_guard<std::mutex> lock(ggml_cuda_lock);
-+        if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) {
-+            ggml_cuda_lock_cv.notify_all();
-+        }
-+    }
-+
-+    return result;
-+}
-+
-+static size_t ggml_backend_cuda_buffer_size(ggml_backend_t backend) {
-+    ggml_backend_cuda_context * ctx = (ggml_backend_cuda_context *)backend->context;
-+    return ctx->pool_get_alloc_size();
-+}
-+
-+static void ggml_backend_cuda_reset(ggml_backend_t backend) {
-+    ggml_backend_cuda_context * ctx = (ggml_backend_cuda_context *)backend->context;
-+    ctx->pools[ctx->device] = NULL;
-+}
-+
- static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
-     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
- 
-@@ -3140,6 +3274,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
-     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
-     /* .event_record            = */ ggml_backend_cuda_event_record,
-     /* .event_wait              = */ ggml_backend_cuda_event_wait,
-+    /* .graph_reserve           = */ ggml_backend_cuda_graph_reserve,
-+    /* .buffer_size             = */ ggml_backend_cuda_buffer_size,
-+    /* .reset                   = */ ggml_backend_cuda_reset,
- };
- 
- static ggml_guid_t ggml_backend_cuda_guid() {

llama/patches/0026-ggml-Backport-scale-kernel-fixes.patch

@@ -1,57 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Jesse Gross <jesse@ollama.com>
-Date: Tue, 23 Sep 2025 15:41:58 -0700
-Subject: [PATCH] ggml: Backport scale kernel fixes
-
-The GGML scale kernel uses signed 32-bit ints to represent
-the number of elements in the tensor. For large images,
-mistral-small3.2 overflows this, triggering CUDA errors due
-to negative arguments.
-
-Currently, this can happen when the user passes a large image
-to mistral-small3.2. However, with upcoming changes to reserve
-CUDA memory, it happens every time mistral-small is loaded as
-we reserve using a worst case batch.
-
-This patch is part of an upstream GGML commit and should be removed
-after GGML is updated past 0a1b398 "ggml: add ops for WAN video model
-(cuda && cpu) (#15669)".
-
-Fixes #10388
----
- ggml/src/ggml-cuda/scale.cu | 19 ++++++++++---------
- 1 file changed, 10 insertions(+), 9 deletions(-)
-
-diff --git a/ggml/src/ggml-cuda/scale.cu b/ggml/src/ggml-cuda/scale.cu
-index 2ee9e5889..0ddeff6a1 100644
---- a/ggml/src/ggml-cuda/scale.cu
-+++ b/ggml/src/ggml-cuda/scale.cu
-@@ -1,18 +1,19 @@
- #include "scale.cuh"
- 
--static __global__ void scale_f32(const float * x, float * dst, const float scale, const float bias, const int k) {
--    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-+#define MAX_GRIDDIM_X 0x7FFFFFFF
- 
--    if (i >= k) {
--        return;
--    }
-+static __global__ void scale_f32(const float * x, float * dst, const float scale, const float bias, const int64_t nelements) {
-+    int64_t tid = (int64_t)blockIdx.x * (int64_t)blockDim.x + (int64_t)threadIdx.x;
-+    int64_t stride = (int64_t)blockDim.x * (int64_t)gridDim.x;
- 
--    dst[i] = scale * x[i] + bias;
-+    for (int64_t i = tid; i < nelements; i += stride) {
-+        dst[i] = scale * x[i] + bias;
-+    }
- }
- 
--static void scale_f32_cuda(const float * x, float * dst, const float scale, const float bias, const int k, cudaStream_t stream) {
--    const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
--    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, bias, k);
-+static void scale_f32_cuda(const float * x, float * dst, const float scale, const float bias, const int64_t nelements, cudaStream_t stream) {
-+    const int64_t num_blocks = (nelements + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
-+    scale_f32<<<MIN(MAX_GRIDDIM_X, num_blocks), CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, bias, nelements);
- }
- 
- void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

llm/server.go

@@ -853,19 +853,19 @@ func (s *ollamaServer) createLayout(systemInfo discover.SystemInfo, systemGPUs d
 
 	if memory == nil {
 		memory = &ml.BackendMemory{CPU: ml.DeviceMemory{
-			Weights: make([]uint64, s.totalLayers),
-			Cache:   make([]uint64, s.totalLayers),
+			Weights: make([]ml.Memory, s.totalLayers),
+			Cache:   make([]ml.Memory, s.totalLayers),
 		}}
 	}
 
 	layers := make([]uint64, len(memory.CPU.Weights))
 	for i := range layers {
 		for j := range memory.GPUs {
-			layers[i] += memory.GPUs[j].Weights[i]
-			layers[i] += memory.GPUs[j].Cache[i]
+			layers[i] += memory.GPUs[j].Weights[i].Size
+			layers[i] += memory.GPUs[j].Cache[i].Size
 		}
-		layers[i] += memory.CPU.Weights[i]
-		layers[i] += memory.CPU.Cache[i]
+		layers[i] += memory.CPU.Weights[i].Size
+		layers[i] += memory.CPU.Cache[i].Size
 		logutil.Trace("layer to assign", "layer", i, "size", format.HumanBytes2(layers[i]))
 	}
 
@@ -880,11 +880,11 @@ func (s *ollamaServer) createLayout(systemInfo discover.SystemInfo, systemGPUs d
 			found := false
 			for j := range memory.GPUs {
 				if gl[i].ID == memory.GPUs[j].ID {
-					if memory.GPUs[j].Graph != 0 {
+					if memory.GPUs[j].Graph.Size != 0 {
 						lastUsedGPU = i
 					}
 
-					reserved := uint64(float32(gl[i].FreeMemory)*backoff) + gl[i].MinimumMemory + envconfig.GpuOverhead() + memory.GPUs[j].Graph
+					reserved := uint64(float32(gl[i].FreeMemory)*backoff) + gl[i].MinimumMemory + envconfig.GpuOverhead() + memory.GPUs[j].Graph.Size
 					if gl[i].FreeMemory > reserved {
 						gl[i].FreeMemory -= reserved
 					} else {
@@ -895,7 +895,7 @@ func (s *ollamaServer) createLayout(systemInfo discover.SystemInfo, systemGPUs d
 						"available layer vram", format.HumanBytes2(gl[i].FreeMemory),
 						"backoff", fmt.Sprintf("%.2f", backoff), "minimum", format.HumanBytes2(gl[i].MinimumMemory),
 						"overhead", format.HumanBytes2(envconfig.GpuOverhead()),
-						"graph", format.HumanBytes2(memory.GPUs[j].Graph))
+						"graph", format.HumanBytes2(memory.GPUs[j].Graph.Size))
 
 					found = true
 					break
@@ -914,12 +914,12 @@ func (s *ollamaServer) createLayout(systemInfo discover.SystemInfo, systemGPUs d
 	}
 
 	// These sizes will only increase as we go through additional iterations and get additional information.
-	cpuSize := memory.InputWeights + memory.CPU.Graph
+	cpuSize := memory.InputWeights.Size + memory.CPU.Graph.Size
 	var vramSize uint64
 	for _, gl := range gpuLayers {
 		for _, gpu := range memory.GPUs {
 			if gl.ID == gpu.ID {
-				vramSize += gpu.Graph
+				vramSize += gpu.Graph.Size
 				break
 			}
 		}
@@ -1723,21 +1723,21 @@ func (s *ollamaServer) VRAMSize() uint64 {
 	var mem uint64
 
 	for _, g := range s.mem.GPUs {
-		mem += g.Size()
+		mem += g.Allocated()
 	}
 
 	// Some elements are always on CPU. However, if we have allocated all layers
 	// on the GPU then include the CPU components as well, to represent complete offloading.
 	noCPULayers := true
 	for i := range s.mem.CPU.Weights {
-		if s.mem.CPU.Weights[i] != 0 || s.mem.CPU.Cache[i] != 0 {
+		if s.mem.CPU.Weights[i].Size != 0 || s.mem.CPU.Cache[i].Size != 0 {
 			noCPULayers = false
 			break
 		}
 	}
 	if noCPULayers {
-		mem += s.mem.InputWeights
-		mem += s.mem.CPU.Graph
+		mem += s.mem.InputWeights.Size
+		mem += s.mem.CPU.Graph.Size
 	}
 
 	return mem
@@ -1748,10 +1748,10 @@ func (s *ollamaServer) TotalSize() uint64 {
 		return 0
 	}
 
-	mem := s.mem.InputWeights
-	mem += s.mem.CPU.Size()
+	mem := s.mem.InputWeights.Size
+	mem += s.mem.CPU.Allocated()
 	for _, g := range s.mem.GPUs {
-		mem += g.Size()
+		mem += g.Allocated()
 	}
 
 	return mem
@@ -1764,7 +1764,7 @@ func (s *ollamaServer) VRAMByGPU(gpuID string) uint64 {
 
 	for _, g := range s.mem.GPUs {
 		if g.ID == gpuID {
-			return g.Size()
+			return g.Allocated()
 		}
 	}
 

llm/server_test.go

@@ -155,18 +155,18 @@ func TestLLMServerFitGPU(t *testing.T) {
 			}
 
 			s.mem = &ml.BackendMemory{CPU: ml.DeviceMemory{
-				Weights: make([]uint64, s.totalLayers),
-				Cache:   make([]uint64, s.totalLayers),
+				Weights: make([]ml.Memory, s.totalLayers),
+				Cache:   make([]ml.Memory, s.totalLayers),
 			}, GPUs: make([]ml.DeviceMemory, len(gpus))}
 
 			for i := range tt.layers {
-				s.mem.CPU.Weights[i] = uint64(tt.layers[i])
+				s.mem.CPU.Weights[i].Size = uint64(tt.layers[i])
 			}
 
 			for i := range s.mem.GPUs {
 				s.mem.GPUs[i].ID = fmt.Sprintf("gpu%d", i)
-				s.mem.GPUs[i].Weights = make([]uint64, s.totalLayers)
-				s.mem.GPUs[i].Cache = make([]uint64, s.totalLayers)
+				s.mem.GPUs[i].Weights = make([]ml.Memory, s.totalLayers)
+				s.mem.GPUs[i].Cache = make([]ml.Memory, s.totalLayers)
 			}
 
 			gpuLayers, err := s.createLayout(systemInfo, gpus, s.mem, tt.requireFull, 0)

logutil/logutil.go

@@ -5,8 +5,6 @@ import (
 	"io"
 	"log/slog"
 	"path/filepath"
-	"runtime"
-	"time"
 )
 
 const LevelTrace slog.Level = -8
@@ -31,18 +29,10 @@ func NewLogger(w io.Writer, level slog.Level) *slog.Logger {
 	}))
 }
 
-type key string
-
 func Trace(msg string, args ...any) {
-	TraceContext(context.WithValue(context.TODO(), key("skip"), 1), msg, args...)
+	slog.Log(context.TODO(), LevelTrace, msg, args...)
 }
 
 func TraceContext(ctx context.Context, msg string, args ...any) {
-	if logger := slog.Default(); logger.Enabled(ctx, LevelTrace) {
-		skip, _ := ctx.Value(key("skip")).(int)
-		pc, _, _, _ := runtime.Caller(1 + skip)
-		record := slog.NewRecord(time.Now(), LevelTrace, msg, pc)
-		record.Add(args...)
-		logger.Handler().Handle(ctx, record)
-	}
+	slog.Log(ctx, LevelTrace, msg, args...)
 }

ml/backend.go

@@ -158,6 +158,40 @@ 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 {
@@ -170,32 +204,39 @@ type DeviceMemory struct {
 	ID string
 
 	// Weights is the per-layer memory needed for the model weights.
-	Weights []uint64
+	Weights []Memory
 
 	// Cache is the per-layer memory needed for the KV cache.
-	Cache []uint64
+	Cache []Memory
 
 	// Graph is the size of the compute graph. It is not per-layer.
-	Graph uint64
+	Graph Memory
 }
 
-func sumMemory(mem []uint64) uint64 {
-	var sum uint64
+// Allocated returns the total size of the memory that has been successfully
+// allocated on this device
+func (m DeviceMemory) Allocated() uint64 {
+	var mem uint64
 
-	for _, m := range mem {
-		sum += m
+	for _, w := range m.Weights {
+		if w.Status == Allocated {
+			mem += w.Size
+		}
+	}
+	for _, c := range m.Cache {
+		if c.Status == Allocated {
+			mem += c.Size
+		}
+	}
+	if m.Graph.Status == Allocated {
+		mem += m.Graph.Size
 	}
 
-	return sum
+	return mem
 }
 
-// Size returns the total size of the memory required by this device
-func (m DeviceMemory) Size() uint64 {
-	return sumMemory(m.Weights) + sumMemory(m.Cache) + m.Graph
-}
-
-func memoryPresent(mem []uint64) bool {
-	return slices.ContainsFunc(mem, func(m uint64) bool { return m != 0 })
+func memoryPresent(mem []Memory) bool {
+	return slices.ContainsFunc(mem, func(m Memory) bool { return m.Size != 0 })
 }
 
 func (m DeviceMemory) LogValue() slog.Value {
@@ -208,7 +249,7 @@ func (m DeviceMemory) LogValue() slog.Value {
 		attrs = append(attrs, slog.Any("Cache", m.Cache))
 	}
 
-	if m.Graph != 0 {
+	if m.Graph.Size != 0 {
 		attrs = append(attrs, slog.Any("Graph", m.Graph))
 	}
 
@@ -226,7 +267,7 @@ func (m DeviceMemory) LogValue() slog.Value {
 // accommodate that to make forward progress.
 type BackendMemory struct {
 	// InputWeights are always located on the CPU and cannot be moved
-	InputWeights uint64
+	InputWeights Memory
 
 	// CPU model components are located in system memory. This does not
 	// include unified memory allocated through the GPU.
@@ -238,7 +279,7 @@ type BackendMemory struct {
 
 func (m BackendMemory) LogValue() slog.Value {
 	var attrs []slog.Attr
-	if m.InputWeights != 0 {
+	if m.InputWeights.Size != 0 {
 		attrs = append(attrs, slog.Any("InputWeights", m.InputWeights))
 	}
 
@@ -250,7 +291,17 @@ func (m BackendMemory) LogValue() slog.Value {
 	return slog.GroupValue(attrs...)
 }
 
-// Log prints a high level summary of the memory
+func sumMemory(mem []Memory) uint64 {
+	var sum uint64
+
+	for _, m := range mem {
+		sum += m.Size
+	}
+
+	return sum
+}
+
+// Log prints a high level summary of the memory (allocated or not)
 func (m BackendMemory) Log(level slog.Level) {
 	var total uint64
 
@@ -260,7 +311,7 @@ func (m BackendMemory) Log(level slog.Level) {
 			total += sum
 		}
 	}
-	if sum := m.InputWeights + sumMemory(m.CPU.Weights); sum > 0 {
+	if sum := m.InputWeights.Size + sumMemory(m.CPU.Weights); sum > 0 {
 		slog.Log(context.TODO(), level, "model weights", "device", m.CPU.Name, "size", format.HumanBytes2(sum))
 		total += sum
 	}
@@ -277,12 +328,12 @@ func (m BackendMemory) Log(level slog.Level) {
 	}
 
 	for _, gpu := range m.GPUs {
-		if sum := gpu.Graph; sum > 0 {
+		if sum := gpu.Graph.Size; sum > 0 {
 			slog.Log(context.TODO(), level, "compute graph", "device", gpu.Name, "size", format.HumanBytes2(sum))
 			total += sum
 		}
 	}
-	if sum := m.CPU.Graph; sum > 0 {
+	if sum := m.CPU.Graph.Size; sum > 0 {
 		slog.Log(context.TODO(), level, "compute graph", "device", m.CPU.Name, "size", format.HumanBytes2(sum))
 		total += sum
 	}
@@ -379,13 +430,12 @@ type Tensor interface {
 	Sin(ctx Context) Tensor
 	Cos(ctx Context) Tensor
 	Tanh(ctx Context) Tensor
-	GELU(ctx Context, up ...Tensor) Tensor
-	SILU(ctx Context, up ...Tensor) Tensor
-	RELU(ctx Context, up ...Tensor) Tensor
+	GELU(ctx Context) Tensor
+	QuickGELU(ctx Context) Tensor
+	SILU(ctx Context) Tensor
+	RELU(ctx Context) Tensor
 	Sigmoid(ctx Context) Tensor
-
-	// AlphaLimitSILU is a variant of SILU that clamps the input to the range [-limit, limit]
-	SILUAlphaLimit(ctx Context, up Tensor, alpha, limit float32) Tensor
+	SwiGLU(ctx Context, up Tensor, alpha, limit float32) Tensor
 
 	Reshape(ctx Context, shape ...int) Tensor
 	View(ctx Context, offset int, shape ...int) Tensor

ml/backend/ggml/ggml.go

@@ -159,6 +159,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
 			bt := C.ggml_backend_dev_buffer_type(d)
 			cpuDeviceBufferType.bts = append(cpuDeviceBufferType.bts, bt)
+			C.ggml_backend_buft_set_alloc(bt, C.bool(params.AllocMemory))
 
 			btDeviceMemory[C.ggml_backend_dev_buffer_type(d)] = &requiredMemory.CPU
 		}
@@ -168,8 +169,8 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	var props C.struct_ggml_backend_dev_props
 	C.ggml_backend_dev_get_props(cpuDeviceBufferType.d, &props)
 	requiredMemory.CPU.ID = C.GoString(props.id)
-	requiredMemory.CPU.Weights = make([]uint64, blocks+1)
-	requiredMemory.CPU.Cache = make([]uint64, blocks+1)
+	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
@@ -180,14 +181,15 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			d:   d,
 			bts: append([]C.ggml_backend_buffer_type_t{bt}, cpuDeviceBufferType.bts...),
 		})
+		C.ggml_backend_buft_set_alloc(bt, C.bool(params.AllocMemory))
 
 		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].ID = C.GoString(props.id)
-		requiredMemory.GPUs[i].Weights = make([]uint64, blocks+1)
-		requiredMemory.GPUs[i].Cache = make([]uint64, blocks+1)
+		requiredMemory.GPUs[i].Weights = make([]ml.Memory, blocks+1)
+		requiredMemory.GPUs[i].Cache = make([]ml.Memory, blocks+1)
 	}
 
 	// inputs always use cpu
@@ -273,9 +275,13 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 
 			size := pad(C.ggml_backend_buft_get_alloc_size(bt, tt), C.ggml_backend_buft_get_alignment(bt))
 			if layer == -1 {
-				requiredMemory.InputWeights += uint64(size)
+				// Assume that InputWeights can be allocated - they're always in system memory and can't be moved in any case
+				if params.AllocMemory {
+					requiredMemory.InputWeights.Status = ml.Allocated
+				}
+				requiredMemory.InputWeights.Size += uint64(size)
 			} else {
-				btDeviceMemory[bt].Weights[layer] += uint64(size)
+				btDeviceMemory[bt].Weights[layer].Size += uint64(size)
 			}
 
 			//nolint:staticcheck // TODO: check if buffer type supports this tensor
@@ -335,6 +341,47 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 		}
 	}
 
+	// allocate buffers for each context
+	bbs := make(map[*C.struct_ggml_context]C.ggml_backend_buffer_t, len(ctxs))
+	for bt, c := range ctxs {
+		if C.ggml_get_first_tensor(c) == nil {
+			continue
+		}
+
+		b := C.ggml_backend_alloc_ctx_tensors_from_buft(c, bt)
+		if params.AllocMemory {
+			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 {
+			for _, b := range bbs {
+				C.ggml_backend_buffer_free(b)
+			}
+
+			for _, ctx := range ctxs {
+				C.ggml_free(ctx)
+			}
+
+			panic(ml.ErrNoMem{BackendMemory: requiredMemory})
+		}
+
+		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
+		bbs[c] = b
+	}
+
+	for bs := range maps.Values(bbs) {
+		logutil.Trace("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)),
+			"size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
+	}
+
 	// map tensor names to tensors for easy lookup later
 	tensors := make(map[string]*C.struct_ggml_tensor)
 	for _, c := range ctxs {
@@ -372,46 +419,6 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	}
 
 	maxGraphNodes := max(8192, len(meta.Tensors().Items())*5)
-
-	sched := C.ggml_backend_sched_new_ext(
-		(*C.ggml_backend_t)(unsafe.Pointer(&schedBackends[0])),
-		(*C.ggml_backend_buffer_type_t)(unsafe.Pointer(&schedBufts[0])),
-		C.int(len(schedBackends)),
-		C.size_t(maxGraphNodes),
-		C._Bool(false),
-		C._Bool(false),
-		C._Bool(params.AllocMemory),
-	)
-
-	// allocate buffers for each context
-	bbs := make(map[*C.struct_ggml_context]C.ggml_backend_buffer_t, len(ctxs))
-	for bt, c := range ctxs {
-		if C.ggml_get_first_tensor(c) == nil {
-			continue
-		}
-
-		b := C.ggml_backend_alloc_ctx_tensors_from_buft(c, bt)
-		if b == nil {
-			for _, b := range bbs {
-				C.ggml_backend_buffer_free(b)
-			}
-
-			for _, ctx := range ctxs {
-				C.ggml_free(ctx)
-			}
-
-			panic(ml.ErrNoMem{BackendMemory: requiredMemory})
-		}
-
-		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
-		bbs[c] = b
-	}
-
-	for bs := range maps.Values(bbs) {
-		logutil.Trace("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)),
-			"size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
-	}
-
 	return &Backend{
 		modelPath:         modelPath,
 		allocMemory:       params.AllocMemory,
@@ -419,11 +426,18 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 		meta:              meta,
 		tensorLoadTargets: targets,
 		tensors:           tensors,
-		sched:             sched,
-		schedBackends:     schedBackends,
-		schedBufts:        schedBufts,
-		input:             deviceBufferTypes[input.d],
-		output:            output.d,
+		sched: C.ggml_backend_sched_new(
+			(*C.ggml_backend_t)(unsafe.Pointer(&schedBackends[0])),
+			(*C.ggml_backend_buffer_type_t)(unsafe.Pointer(&schedBufts[0])),
+			C.int(len(schedBackends)),
+			C.size_t(maxGraphNodes),
+			C._Bool(false),
+			C._Bool(false),
+		),
+		schedBackends: schedBackends,
+		schedBufts:    schedBufts,
+		input:         deviceBufferTypes[input.d],
+		output:        output.d,
 		layers: func() map[int]layerDevice {
 			m := make(map[int]layerDevice)
 			for i, layer := range layers {
@@ -781,15 +795,24 @@ func (c *Context) Reserve() {
 
 	// 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 = 0
+		c.b.btDeviceMemory[bt].Graph = ml.Memory{}
 	}
 
 	for i := range c.b.schedBackends {
-		bufferSize := C.ggml_backend_sched_get_attempted_buffer_size(c.b.sched, c.b.schedBackends[i])
-		c.b.btDeviceMemory[c.b.schedBufts[i]].Graph += uint64(bufferSize)
+		bufferStatus := C.ggml_backend_sched_get_attempted_buffer_size(c.b.sched, c.b.schedBackends[i])
+
+		graph := &c.b.btDeviceMemory[c.b.schedBufts[i]].Graph
+		graph.Size += uint64(bufferStatus.size)
+		if c.b.allocMemory {
+			if bufferStatus.allocated && graph.Status != ml.Failed {
+				graph.Status = ml.Allocated
+			} else {
+				graph.Status = ml.Failed
+			}
+		}
 
 		logutil.Trace("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(bufferSize)))
+			"buffer_type", C.GoString(C.ggml_backend_buft_name(c.b.schedBufts[i])), "size", format.HumanBytes2(uint64(bufferStatus.size)))
 	}
 
 	if !reserved {
@@ -839,7 +862,16 @@ func (c *Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
 
 	b := C.ggml_backend_buft_alloc_buffer(c.buft, size)
 	if c.layer >= 0 {
-		c.b.btDeviceMemory[c.buft].Cache[c.layer] += uint64(size)
+		cache := &c.b.btDeviceMemory[c.buft].Cache[c.layer]
+
+		cache.Size += uint64(size)
+		if c.b.allocMemory {
+			if b != nil {
+				cache.Status = ml.Allocated
+			} else {
+				cache.Status = ml.Failed
+			}
+		}
 	}
 
 	if b == nil {
@@ -1399,46 +1431,35 @@ func (t *Tensor) IM2Col(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int
 	}
 }
 
-func (t *Tensor) GELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
-	if len(t2) > 0 {
-		return &Tensor{
-			b: t.b,
-			t: C.ggml_geglu_split(ctx.(*Context).ctx, t.t, t2[0].(*Tensor).t),
-		}
-	}
+func (t *Tensor) GELU(ctx ml.Context) ml.Tensor {
 	return &Tensor{
 		b: t.b,
 		t: C.ggml_gelu_inplace(ctx.(*Context).ctx, t.t),
 	}
 }
 
-func (t *Tensor) SILU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
-	if len(t2) > 0 {
-		return &Tensor{
-			b: t.b,
-			t: C.ggml_swiglu_split(ctx.(*Context).ctx, t.t, t2[0].(*Tensor).t),
-		}
+func (t *Tensor) QuickGELU(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_gelu_quick_inplace(ctx.(*Context).ctx, t.t),
 	}
+}
+
+func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
 	return &Tensor{
 		b: t.b,
 		t: C.ggml_silu_inplace(ctx.(*Context).ctx, t.t),
 	}
 }
 
-func (t *Tensor) RELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
-	if len(t2) > 0 {
-		return &Tensor{
-			b: t.b,
-			t: C.ggml_reglu_split(ctx.(*Context).ctx, t.t, t2[0].(*Tensor).t),
-		}
-	}
+func (t *Tensor) RELU(ctx ml.Context) ml.Tensor {
 	return &Tensor{
 		b: t.b,
 		t: C.ggml_relu_inplace(ctx.(*Context).ctx, t.t),
 	}
 }
 
-func (t *Tensor) SILUAlphaLimit(ctx ml.Context, up ml.Tensor, alpha, limit float32) ml.Tensor {
+func (t *Tensor) SwiGLU(ctx ml.Context, up ml.Tensor, alpha, limit float32) ml.Tensor {
 	return &Tensor{
 		b: t.b,
 		t: C.ggml_swiglu_oai(ctx.(*Context).ctx, t.t, up.(*Tensor).t, C.float(alpha), C.float(limit)),

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

@@ -65,7 +65,12 @@ GGML_API bool ggml_gallocr_reserve_n(
 GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
 
 GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
-GGML_API size_t ggml_gallocr_get_attempted_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

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

@@ -35,6 +35,7 @@ extern "C" {
     //
 
     GGML_API const char *          ggml_backend_buft_name          (ggml_backend_buffer_type_t buft);
+    GGML_API void                  ggml_backend_buft_set_alloc     (ggml_backend_buffer_type_t buft, bool alloc);
     GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer  (ggml_backend_buffer_type_t buft, size_t size);
     GGML_API size_t                ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
     GGML_API size_t                ggml_backend_buft_get_max_size  (ggml_backend_buffer_type_t buft);
@@ -292,7 +293,6 @@ extern "C" {
 
     // Initialize a backend scheduler, backends with low index are given priority over backends with high index
     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new_ext(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload, bool alloc_buffers);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph
@@ -306,7 +306,12 @@ extern "C" {
     GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
 
     GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-    GGML_API size_t               ggml_backend_sched_get_attempted_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

@@ -932,7 +932,7 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
     return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
 }
 
-size_t ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int 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++) {
@@ -941,11 +941,13 @@ size_t ggml_gallocr_get_attempted_buffer_size(ggml_gallocr_t galloc, int buffer_
             // (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.
 
-            return 0;
+            struct ggml_allocr_buffer_status status = {0, true};
+            return status;
         }
     }
 
-    return galloc->buffer_sizes[buffer_id];
+    struct ggml_allocr_buffer_status status = {galloc->buffer_sizes[buffer_id], galloc->buffers[buffer_id] != NULL};
+    return status;
 }
 
 // utils

ml/backend/ggml/ggml/src/ggml-backend-impl.h

@@ -26,10 +26,6 @@ extern "C" {
         size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
         // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
         bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
-
-        // (optional) returns a dummy buffer that is equivalent to one created by alloc_buffer but without actually being backed
-        // by memory
-        ggml_backend_buffer_t (*noalloc_buffer)(ggml_backend_buffer_type_t buft, size_t size);
     };
 
     struct ggml_backend_buffer_type {
@@ -120,16 +116,6 @@ extern "C" {
         void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
         // wait for an event on on a different stream
         void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
-
-        // (optional) reserves intermediate buffers needed for the compution
-        // if alloc is true, memory is actually allocated, otherwise the required amount is just returned by buffer_size
-        enum ggml_status          (*graph_reserve)     (ggml_backend_t backend, struct ggml_cgraph * cgraph, bool alloc);
-
-        // (optional) returns the memory needed after calling graph_reserve
-        size_t                    (*buffer_size)       (ggml_backend_t backend);
-
-        // (optional) frees memory from intermediate buffers that was allocated either by graph_compute or graph_reserve
-        void                      (*reset)             (ggml_backend_t backend);
     };
 
     struct ggml_backend {

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

@@ -35,6 +35,10 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_name(buft);
 }
 
+void ggml_backend_buft_set_alloc(ggml_backend_buffer_type_t buft, bool alloc) {
+    buft->no_alloc = !alloc;
+}
+
 ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     if (size == 0) {
         // return a dummy buffer for zero-sized allocations
@@ -42,14 +46,7 @@ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t
     }
 
     if (buft->no_alloc) {
-        ggml_backend_buffer_t buf;
-
-        if (buft->iface.noalloc_buffer != NULL) {
-            buf = buft->iface.noalloc_buffer(buft, size);
-        } else {
-            buf = ggml_backend_buffer_init(buft, {}, NULL, size);
-        }
-
+        ggml_backend_buffer_t buf = ggml_backend_buffer_init(buft, {}, NULL, size);
         buf->no_alloc = true;
         return buf;
     }
@@ -691,12 +688,6 @@ struct ggml_backend_sched {
     bool op_offload;
 
     int debug;
-
-    // allocate buffers on attached ggml_backend_buffer_type_t's and during reservation
-    // if false, dummy buffers are used for faster memory sizing calculations
-    // the scheduler needs to be recreated with allocated buffers before it can be used
-    // for computation
-    bool alloc_buffers;
 };
 
 #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
@@ -1483,17 +1474,6 @@ ggml_backend_sched_t ggml_backend_sched_new(
         size_t graph_size,
         bool parallel,
         bool op_offload) {
-            return ggml_backend_sched_new_ext(backends, bufts, n_backends, graph_size, parallel, op_offload, true);
-        }
-
-ggml_backend_sched_t ggml_backend_sched_new_ext(
-        ggml_backend_t * backends,
-        ggml_backend_buffer_type_t * bufts,
-        int n_backends,
-        size_t graph_size,
-        bool parallel,
-        bool op_offload,
-        bool alloc_buffers) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
@@ -1535,13 +1515,10 @@ ggml_backend_sched_t ggml_backend_sched_new_ext(
                 sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
             }
         }
-
-        sched->bufts[b]->no_alloc = !alloc_buffers;
     }
 
     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
     sched->op_offload = op_offload;
-    sched->alloc_buffers = alloc_buffers;
 
     ggml_backend_sched_reset(sched);
 
@@ -1556,10 +1533,6 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
         for (int c = 0; c < sched->n_copies; c++) {
             ggml_backend_event_free(sched->events[b][c]);
         }
-
-        if (sched->backends[b]->iface.reset != NULL) {
-            sched->backends[b]->iface.reset(sched->backends[b]);
-        }
     }
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
@@ -1599,24 +1572,6 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
         return false;
     }
 
-    if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
-        return false;
-    }
-
-    struct ggml_backend_sched_split * splits = sched->splits;
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
-        int split_backend_id = split->backend_id;
-        ggml_backend_t split_backend = sched->backends[split_backend_id];
-
-        if (split_backend->iface.graph_reserve != NULL) {
-            enum ggml_status ec = split_backend->iface.graph_reserve(split_backend, &split->graph, sched->alloc_buffers);
-            if (ec != GGML_STATUS_SUCCESS) {
-                return false;
-            }
-        }
-    }
-
     ggml_backend_sched_reset(sched);
 
     return true;
@@ -1701,17 +1656,14 @@ 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);
 }
 
-size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
+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);
 
-    size_t size = ggml_gallocr_get_attempted_buffer_size(sched->galloc, backend_index);
+    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};
 
-    if (backend->iface.buffer_size != NULL) {
-        size += backend->iface.buffer_size(backend);
-    }
-
-    return size;
+    return status;
 }
 
 void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {

ml/backend/ggml/ggml/src/ggml-cuda/common.cuh

@@ -35,31 +35,6 @@
 #include "vendors/cuda.h"
 #endif // defined(GGML_USE_HIP)
 
-extern bool reserving_graph;
-
-// If we are reserving the graph, pointers might be invalid and will fail if cudaMemcpyAsync tries to validate them.
-// However, since we don't actually expect a result, we don't need to actually do the memcpy.
-static cudaError_t cudaMemcpyAsyncReserve ( void* dst, const void* src, size_t count, cudaMemcpyKind kind, cudaStream_t stream = 0 ) {
-    if (!reserving_graph) {
-        return cudaMemcpyAsync(dst, src, count, kind, stream);
-    } else {
-        return cudaSuccess;
-    }
-}
-
-static cudaError_t cudaMemcpy2DAsyncReserve ( void* dst, size_t dpitch, const void* src, size_t spitch, size_t width, size_t height, cudaMemcpyKind kind, cudaStream_t stream = 0 ) {
-    if (!reserving_graph) {
-        return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, kind, stream);
-    } else {
-        return cudaSuccess;
-    }
-}
-
-#undef cudaMemcpyAsync
-#define cudaMemcpyAsync cudaMemcpyAsyncReserve
-#undef cudaMemcpy2DAsync
-#define cudaMemcpy2DAsync cudaMemcpy2DAsyncReserve
-
 #define STRINGIZE_IMPL(...) #__VA_ARGS__
 #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
 
@@ -796,9 +771,6 @@ struct ggml_cuda_pool {
 
     virtual void * alloc(size_t size, size_t * actual_size) = 0;
     virtual void free(void * ptr, size_t size) = 0;
-
-    virtual bool alloc_memory() = 0;
-    virtual size_t alloc_size() = 0;
 };
 
 template<typename T>
@@ -942,11 +914,11 @@ struct ggml_backend_cuda_context {
     // pool
     std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
 
-    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, bool alloc);
+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
 
     ggml_cuda_pool & pool(int device) {
         if (pools[device] == nullptr) {
-            pools[device] = new_pool_for_device(device, true);
+            pools[device] = new_pool_for_device(device);
         }
         return *pools[device];
     }
@@ -954,20 +926,4 @@ struct ggml_backend_cuda_context {
     ggml_cuda_pool & pool() {
         return pool(device);
     }
-
-    void pool_set_alloc(bool alloc) {
-        GGML_ASSERT(pools[device] == nullptr || pools[device]->alloc_memory() == alloc);
-
-        if (pools[device] == nullptr) {
-            pools[device] = new_pool_for_device(device, alloc);
-        }
-    }
-
-    size_t pool_get_alloc_size() {
-        if (pools[device] == nullptr) {
-            return 0;
-        }
-
-        return pools[device]->alloc_size();
-    }
 };

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

@@ -355,8 +355,6 @@ const ggml_cuda_device_info & ggml_cuda_info() {
 
 // #define DEBUG_CUDA_MALLOC
 
-#define CUDA_ALIGNMENT 128
-
 // buffer pool for cuda (legacy)
 struct ggml_cuda_pool_leg : public ggml_cuda_pool {
     static const int MAX_BUFFERS = 256;
@@ -369,12 +367,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 
     ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
     size_t pool_size = 0;
-    bool allocate = true;
-    size_t last_alloc = 0;
 
-    explicit ggml_cuda_pool_leg(int device, bool alloc) :
-        device(device),
-        allocate(alloc) {
+    explicit ggml_cuda_pool_leg(int device) :
+        device(device) {
     }
 
     ~ggml_cuda_pool_leg() {
@@ -382,9 +377,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cuda_buffer & b = buffer_pool[i];
             if (b.ptr != nullptr) {
-                if (allocate) {
-                    CUDA_CHECK(cudaFree(b.ptr));
-                }
+                CUDA_CHECK(cudaFree(b.ptr));
                 pool_size -= b.size;
             }
         }
@@ -432,15 +425,8 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
         void * ptr;
         size_t look_ahead_size = (size_t) (1.05 * size);
         look_ahead_size = 256 * ((look_ahead_size + 255)/256);
-        if (allocate) {
-            ggml_cuda_set_device(device);
-            if (ggml_cuda_device_malloc(&ptr, look_ahead_size, device) != cudaSuccess) {
-                    last_alloc = look_ahead_size;
-                    throw std::bad_alloc();
-            }
-        } else {
-            ptr = (void *)CUDA_ALIGNMENT;
-        }
+        ggml_cuda_set_device(device);
+        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
@@ -460,20 +446,10 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
             }
         }
         GGML_LOG_DEBUG(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
-        if (allocate) {
-            ggml_cuda_set_device(device);
-            CUDA_CHECK(cudaFree(ptr));
-        }
+        ggml_cuda_set_device(device);
+        CUDA_CHECK(cudaFree(ptr));
         pool_size -= size;
     }
-
-    bool alloc_memory() override {
-        return allocate;
-    }
-
-    size_t alloc_size() override {
-        return pool_size + last_alloc;
-    }
 };
 
 // pool with virtual memory
@@ -485,24 +461,18 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     CUdeviceptr pool_addr = 0;
     size_t pool_used = 0;
     size_t pool_size = 0;
-    bool allocate = true;
-    size_t last_alloc = 0;
     size_t granularity;
 #if defined(GGML_USE_HIP)
     std::vector<std::pair<CUdeviceptr, size_t>> mappings;
 #endif
 
-    explicit ggml_cuda_pool_vmm(int device, bool alloc) :
+    explicit ggml_cuda_pool_vmm(int device) :
         device(device),
-        granularity(ggml_cuda_info().devices[device].vmm_granularity),
-        allocate(alloc) {
-        if (!allocate) {
-            pool_addr = (CUdeviceptr)CUDA_ALIGNMENT;
-        }
+        granularity(ggml_cuda_info().devices[device].vmm_granularity) {
     }
 
     ~ggml_cuda_pool_vmm() {
-        if (pool_addr != 0 && allocate) {
+        if (pool_addr != 0) {
 #if defined(GGML_USE_HIP)
             // Workaround for https://github.com/ROCm/ROCR-Runtime/issues/285
             for (std::pair<CUdeviceptr, size_t> & mapping : mappings) {
@@ -529,50 +499,36 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
 
             GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
 
-            if (allocate) {
-                // allocate more physical memory
-                CUmemAllocationProp prop = {};
-                prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
-                prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-                prop.location.id = device;
-                CUmemGenericAllocationHandle handle;
-                if (cuMemCreate(&handle, reserve_size, &prop, 0) != CUDA_SUCCESS) {
-                    last_alloc = reserve_size;
-                    throw std::bad_alloc();
-                }
+            // allocate more physical memory
+            CUmemAllocationProp prop = {};
+            prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
+            prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+            prop.location.id = device;
+            CUmemGenericAllocationHandle handle;
+            CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));
 
-                // reserve virtual address space (if not already reserved)
-                if (pool_addr == 0) {
-                    CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
-                }
-
-                // map at the end of the pool
-                CUdeviceptr start_ptr = (CUdeviceptr)((char *)(pool_addr) + pool_size);
-                if (cuMemMap(start_ptr, reserve_size, 0, handle, 0) != CUDA_SUCCESS) {
-                    last_alloc = reserve_size;
-                    CU_CHECK(cuMemRelease(handle));
-                    throw std::bad_alloc();
-                }
-
-                // the memory allocation handle is no longer needed after mapping
-                CU_CHECK(cuMemRelease(handle));
-
-                // set access
-                CUmemAccessDesc access = {};
-                access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-                access.location.id = device;
-                access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
-                if (cuMemSetAccess((CUdeviceptr)((char *)(pool_addr) + pool_size), reserve_size, &access, 1) != CUDA_SUCCESS) {
-                    CU_CHECK(cuMemUnmap(start_ptr, reserve_size));
-                    last_alloc = reserve_size;
-                    throw std::bad_alloc();
-                }
-
-    #if defined(GGML_USE_HIP)
-                mappings.push_back({start_ptr, reserve_size});
-    #endif
+            // reserve virtual address space (if not already reserved)
+            if (pool_addr == 0) {
+                CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
             }
 
+            // map at the end of the pool
+            CUdeviceptr start_ptr = (CUdeviceptr)((char *)(pool_addr) + pool_size);
+            CU_CHECK(cuMemMap(start_ptr, reserve_size, 0, handle, 0));
+#if defined(GGML_USE_HIP)
+            mappings.push_back({start_ptr, reserve_size});
+#endif
+
+            // the memory allocation handle is no longer needed after mapping
+            CU_CHECK(cuMemRelease(handle));
+
+            // set access
+            CUmemAccessDesc access = {};
+            access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+            access.location.id = device;
+            access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+            CU_CHECK(cuMemSetAccess((CUdeviceptr)((char *)(pool_addr) + pool_size), reserve_size, &access, 1));
+
             // add to the pool
             pool_size += reserve_size;
 
@@ -604,24 +560,16 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
         // all deallocations must be in reverse order of the allocations
         GGML_ASSERT(ptr == (void *) ((char *)(pool_addr) + pool_used));
     }
-
-    bool alloc_memory() override {
-        return allocate;
-    }
-
-    size_t alloc_size() override {
-        return pool_size + last_alloc;
-    }
 };
 #endif // defined(GGML_USE_VMM)
 
-std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device, bool alloc) {
+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
 #if defined(GGML_USE_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
-        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device, alloc));
+        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
 #endif // defined(GGML_USE_VMM)
-    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device, alloc));
+    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
 // destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
@@ -805,20 +753,11 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
 }
 
 static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return CUDA_ALIGNMENT;
+    return 128;
 
     GGML_UNUSED(buft);
 }
 
-static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_noalloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
-
-    void * dev_ptr = (void *)ggml_backend_cuda_buffer_type_get_alignment(buft);
-    ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr);
-
-    return ggml_backend_buffer_init(buft, {}, ctx, size);
-}
-
 static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
@@ -842,7 +781,6 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
     /* .is_host          = */ NULL,
-    /* .noalloc_buffer   = */ ggml_backend_cuda_buffer_type_noalloc_buffer,
 };
 
 ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
@@ -3003,7 +2941,6 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
 
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
     bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
-
     // flag used to determine whether it is an integrated_gpu
     const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
 
@@ -3019,11 +2956,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     continue;
                 }
 
-                // When reserving, we are forcing CUDA graphs but this operation is not graph-safe so we need to skip it
-                if (reserving_graph && node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
-                    continue;
-                }
-
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                 if (!disable_fusion) {
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL }, {})) {
@@ -3095,7 +3027,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
 
 static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
-    cuda_ctx->pool_set_alloc(true);
 
     ggml_cuda_set_device(cuda_ctx->device);
 
@@ -3175,71 +3106,6 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
     return GGML_STATUS_SUCCESS;
 }
 
-// This is used to skip operations that are not graph safe during the reservation process.
-bool reserving_graph = false;
-
-static enum ggml_status ggml_backend_cuda_graph_reserve(ggml_backend_t backend, ggml_cgraph * cgraph, bool alloc) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
-    cuda_ctx->pool_set_alloc(alloc);
-
-    #ifdef USE_CUDA_GRAPH
-    if (cuda_ctx->cuda_graph == nullptr) {
-        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
-    }
-    #endif
-
-    ggml_cuda_set_device(cuda_ctx->device);
-
-    {
-        std::lock_guard<std::mutex> lock(ggml_cuda_lock);
-        ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed);
-    }
-
-    reserving_graph = true;
-
-    // Create CuBLAS handles early to avoid synchronous allocations during graph capture.
-    cuda_ctx->cublas_handle();
-
-    CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
-
-    enum ggml_status result = GGML_STATUS_SUCCESS;
-
-    try {
-        bool use_cuda_graph = false;
-        bool cuda_graph_update_required = false;
-        bool graph_evaluated_or_captured = false;
-
-        evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required);
-    } catch (const std::exception &e) {
-        result = GGML_STATUS_FAILED;
-    }
-
-    cudaGraph_t graph;
-    CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &graph));
-    CUDA_CHECK(cudaGraphDestroy(graph));
-
-    reserving_graph = false;
-
-    {
-        std::lock_guard<std::mutex> lock(ggml_cuda_lock);
-        if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) {
-            ggml_cuda_lock_cv.notify_all();
-        }
-    }
-
-    return result;
-}
-
-static size_t ggml_backend_cuda_buffer_size(ggml_backend_t backend) {
-    ggml_backend_cuda_context * ctx = (ggml_backend_cuda_context *)backend->context;
-    return ctx->pool_get_alloc_size();
-}
-
-static void ggml_backend_cuda_reset(ggml_backend_t backend) {
-    ggml_backend_cuda_context * ctx = (ggml_backend_cuda_context *)backend->context;
-    ctx->pools[ctx->device] = NULL;
-}
-
 static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
 
@@ -3279,9 +3145,6 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .event_record            = */ ggml_backend_cuda_event_record,
     /* .event_wait              = */ ggml_backend_cuda_event_wait,
-    /* .graph_reserve           = */ ggml_backend_cuda_graph_reserve,
-    /* .buffer_size             = */ ggml_backend_cuda_buffer_size,
-    /* .reset                   = */ ggml_backend_cuda_reset,
 };
 
 static ggml_guid_t ggml_backend_cuda_guid() {

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

@@ -1,19 +1,18 @@
 #include "scale.cuh"
 
-#define MAX_GRIDDIM_X 0x7FFFFFFF
+static __global__ void scale_f32(const float * x, float * dst, const float scale, const float bias, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
-static __global__ void scale_f32(const float * x, float * dst, const float scale, const float bias, const int64_t nelements) {
-    int64_t tid = (int64_t)blockIdx.x * (int64_t)blockDim.x + (int64_t)threadIdx.x;
-    int64_t stride = (int64_t)blockDim.x * (int64_t)gridDim.x;
-
-    for (int64_t i = tid; i < nelements; i += stride) {
-        dst[i] = scale * x[i] + bias;
+    if (i >= k) {
+        return;
     }
+
+    dst[i] = scale * x[i] + bias;
 }
 
-static void scale_f32_cuda(const float * x, float * dst, const float scale, const float bias, const int64_t nelements, cudaStream_t stream) {
-    const int64_t num_blocks = (nelements + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
-    scale_f32<<<MIN(MAX_GRIDDIM_X, num_blocks), CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, bias, nelements);
+static void scale_f32_cuda(const float * x, float * dst, const float scale, const float bias, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
+    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, bias, k);
 }
 
 void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ml/nn/attention.go

@@ -26,7 +26,6 @@ func Attention(ctx ml.Context, query, key, value ml.Tensor, scale float64, cache
 }
 
 func AttentionWithSinks(ctx ml.Context, query, key, value, sinks ml.Tensor, scale float64, cache kvcache.Cache) ml.Tensor {
-	ctx.Forward(query)
 	if key != nil && value != nil {
 		if query.Dim(0) != key.Dim(0) {
 			panic(fmt.Errorf("d_k in attention operation does not match between query(%v) and key(%v)", query.Dim(0), key.Dim(0)))
@@ -40,7 +39,6 @@ func AttentionWithSinks(ctx ml.Context, query, key, value, sinks ml.Tensor, scal
 			panic(fmt.Errorf("seq_len_k in attention operation does not match between key(%v) and value(%v)", key.Dim(2), value.Dim(2)))
 		}
 
-		ctx.Forward(key, value)
 		if cache != nil {
 			cache.Put(ctx, key, value)
 		}

ml/nn/pooling/pooling.go

@@ -11,32 +11,26 @@ const (
 	TypeMean
 	TypeCLS
 	TypeLast
+	TypeRank
+
+	TypeUnknown     = 0xFFFFFFFE
+	TypeUnspecified = 0xFFFFFFFF
 )
 
-func (t Type) String() string {
-	switch t {
-	case TypeMean:
-		return "Mean"
-	case TypeCLS:
-		return "CLS"
-	case TypeLast:
-		return "Last"
-	default:
-		return "Unknown"
-	}
-}
-
-func (t Type) Forward(ctx ml.Context, hiddenStates ml.Tensor) ml.Tensor {
-	switch t {
+func Pooling(ctx ml.Context, hiddenStates ml.Tensor, poolingType Type) ml.Tensor {
+	switch poolingType {
+	case TypeNone:
+		return hiddenStates
 	case TypeMean:
 		hiddenStates = hiddenStates.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx).Mean(ctx)
 		return hiddenStates.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
 	case TypeCLS:
 		return hiddenStates.View(ctx, 0, hiddenStates.Dim(0))
 	case TypeLast:
-		hiddenStates = hiddenStates.View(ctx, (hiddenStates.Dim(1)-1)*hiddenStates.Stride(1), hiddenStates.Dim(0))
-		return hiddenStates
+		panic("not implemented")
+	case TypeRank:
+		panic("not implemented")
 	default:
-		panic("unknown pooling type")
+		panic("not implemented")
 	}
 }

ml/nn/pooling/pooling_test.go

@@ -1,79 +0,0 @@
-package pooling_test
-
-import (
-	"bytes"
-	"os"
-	"slices"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/discover"
-	fsggml "github.com/ollama/ollama/fs/ggml"
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/ml/backend/ggml"
-	"github.com/ollama/ollama/ml/nn/pooling"
-)
-
-func setup(tb testing.TB, n int) ml.Backend {
-	tb.Helper()
-
-	f, err := os.CreateTemp(tb.TempDir(), "*.bin")
-	if err != nil {
-		tb.Fatal(err)
-	}
-	defer f.Close()
-
-	if err := fsggml.WriteGGUF(f, fsggml.KV{
-		"general.architecture": "test",
-		"test.block_count":     uint32(1),
-	}, []*fsggml.Tensor{
-		{Name: "blk.0.weight", Shape: []uint64{1}, WriterTo: bytes.NewBuffer(make([]byte, 4))},
-	}); err != nil {
-		tb.Fatal(err)
-	}
-
-	var gpuLayers ml.GPULayersList
-	if gpus := discover.GetGPUInfo(); len(gpus) > 0 {
-		gpuLayers = append(gpuLayers, ml.GPULayers{
-			ID: gpus[0].ID,
-			Layers: slices.Collect(func(yield func(int) bool) {
-				for i := range n {
-					if !yield(i) {
-						return
-					}
-				}
-			}),
-		})
-	}
-	b, err := ggml.New(f.Name(), ml.BackendParams{AllocMemory: true, GPULayers: gpuLayers})
-	if err != nil {
-		tb.Fatal(err)
-	}
-
-	return b
-}
-
-func TestForward(t *testing.T) {
-	cases := map[pooling.Type][]float32{
-		pooling.TypeMean: {4, 5, 6, 7, 8, 9, 10, 11},
-		pooling.TypeCLS:  {0, 1, 2, 3, 4, 5, 6, 7},
-		pooling.TypeLast: {8, 9, 10, 11, 12, 13, 14, 15},
-	}
-	for typ, want := range cases {
-		t.Run(typ.String(), func(t *testing.T) {
-			b := setup(t, 99)
-			defer b.Close()
-
-			ctx := b.NewContext()
-			defer ctx.Close()
-
-			tt := ctx.Input().Arange(0, 16, 1, ml.DTypeF32).Reshape(ctx, 8, 2)
-			tt = typ.Forward(ctx, tt)
-
-			ctx.Forward(tt).Compute(tt)
-			if diff := cmp.Diff(want, tt.Floats()); diff != "" {
-				t.Error(diff)
-			}
-		})
-	}
-}

model/bytepairencoding.go

@@ -5,7 +5,6 @@ import (
 	"fmt"
 	"iter"
 	"log/slog"
-	"slices"
 	"strings"
 
 	"github.com/dlclark/regexp2"
@@ -14,28 +13,16 @@ import (
 )
 
 type BytePairEncoding struct {
-	vocab   *Vocabulary
-	regexps []*regexp2.Regexp
+	pre   *regexp2.Regexp
+	vocab *Vocabulary
 }
 
 var _ TextProcessor = (*BytePairEncoding)(nil)
 
-func NewBytePairEncoding(vocab *Vocabulary, pretokenizers ...string) BytePairEncoding {
-	if len(pretokenizers) == 0 {
-		// set default byte-level pretokenizer if none provided, e.g.
-		// https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/byte_level.rs#L44
-		pretokenizers = []string{`'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+`}
-	}
-
+func NewBytePairEncoding(pre string, vocab *Vocabulary) BytePairEncoding {
 	return BytePairEncoding{
+		pre:   regexp2.MustCompile(pre, regexp2.None),
 		vocab: vocab,
-		regexps: slices.Collect(func(yield func(*regexp2.Regexp) bool) {
-			for _, p := range pretokenizers {
-				if !yield(regexp2.MustCompile(p, regexp2.RE2)) {
-					return
-				}
-			}
-		}),
 	}
 }
 
@@ -48,36 +35,13 @@ func (bpe BytePairEncoding) Is(id int32, special Special) bool {
 }
 
 func (bpe *BytePairEncoding) split(s string) iter.Seq[string] {
-	parts := []string{s}
-	for _, re := range bpe.regexps {
-		parts = slices.Collect(func(yield func(string) bool) {
-			for _, part := range parts {
-				r := []rune(part)
-				var offset int
-				for m, _ := re.FindRunesMatch(r); m != nil; m, _ = re.FindNextMatch(m) {
-					if offset-m.Index != 0 {
-						if !yield(string(r[:m.Index])) {
-							return
-						}
-					}
-
-					if !yield(m.String()) {
-						return
-					}
-
-					offset = m.Index + m.Length
-				}
-
-				if offset < len(r) {
-					if !yield(string(r[offset:])) {
-						return
-					}
-				}
+	return func(yield func(string) bool) {
+		for m, _ := bpe.pre.FindStringMatch(s); m != nil; m, _ = bpe.pre.FindNextMatch(m) {
+			if !yield(m.String()) {
+				break
 			}
-		})
+		}
 	}
-
-	return slices.Values(parts)
 }
 
 // fragment is a string fragment and their corresponding token IDs

model/bytepairencoding_test.go

@@ -59,12 +59,12 @@ func llama(t testing.TB) BytePairEncoding {
 	}
 
 	return NewBytePairEncoding(
+		`(?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+`,
 		&Vocabulary{
 			Values: tokens,
 			Types:  types,
 			Merges: merges,
 		},
-		"(?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+",
 	)
 }
 
@@ -282,41 +282,3 @@ func BenchmarkBytePairEncoding(b *testing.B) {
 		})
 	}
 }
-
-func TestSplit(t *testing.T) {
-	cases := []struct {
-		name string
-		patterns,
-		want []string
-	}{
-		{
-			name: "default",
-			want: []string{"Hello", ",", " WORLD", "!!", " How", "'s", " it", " going", "?", " 123", " 一二三"},
-		},
-		{
-			name: "unicode",
-			patterns: []string{
-				"\\p{N}{1,3}",
-				`[一-龥぀-ゟ゠-ヿ]+`,
-				"[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
-			},
-			want: []string{"Hello", ",", " WORLD", "!!", " How", "'s", " it", " going", "?", " ", "123", " ", "一二三"},
-		},
-		{
-			name: "individual digits",
-			patterns: []string{
-				"(?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+",
-			},
-			want: []string{"Hello", ",", " WORLD", "!!", " How", "'s", " it", " going", "?", " ", "1", "2", "3", " 一二三"},
-		},
-	}
-
-	for _, tt := range cases {
-		t.Run(tt.name, func(t *testing.T) {
-			tokenizer := NewBytePairEncoding(nil, tt.patterns...)
-			if diff := cmp.Diff(tt.want, slices.Collect(tokenizer.split("Hello, WORLD!! How's it going? 123 一二三"))); diff != "" {
-				t.Errorf("no match (-theirs +ours):\n%s", diff)
-			}
-		})
-	}
-}

model/model.go

@@ -5,7 +5,7 @@ import (
 	"fmt"
 	_ "image/jpeg"
 	_ "image/png"
-	"log/slog"
+	"math"
 	"os"
 	"reflect"
 	"strconv"
@@ -21,7 +21,6 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
 	_ "github.com/ollama/ollama/ml/backend"
-	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model/input"
 )
 
@@ -108,12 +107,23 @@ func New(modelPath string, params ml.BackendParams) (Model, error) {
 		return nil, err
 	}
 
-	m, err := modelForArch(b.Config())
+	arch := b.Config().Architecture()
+	if b.Config().Uint("pooling_type", math.MaxUint32) != math.MaxUint32 {
+		arch = arch + "_embed"
+	}
+
+	f, ok := models[arch]
+	if !ok {
+		return nil, fmt.Errorf("unsupported model architecture %q", arch)
+	}
+
+	m, err := f(b.Config())
 	if err != nil {
 		return nil, err
 	}
 
 	base := Base{b: b, config: m.Config()}
+
 	v := reflect.ValueOf(m)
 	v.Elem().Set(populateFields(base, v.Elem()))
 	return m, nil
@@ -125,38 +135,30 @@ func NewTextProcessor(s string) (TextProcessor, error) {
 		return nil, err
 	}
 	defer r.Close()
-
 	meta, err := fsggml.Decode(r, -1)
 	if err != nil {
 		return nil, err
 	}
+	return getTextProcessor(meta.KV())
+}
 
-	m, err := modelForArch(meta.KV())
+func getTextProcessor(kv fsggml.KV) (TextProcessor, error) {
+	arch := kv.Architecture()
+	f, ok := models[arch]
+	if !ok {
+		return nil, fmt.Errorf("unsupported model architecture %q", arch)
+	}
+	m, err := f(kv)
 	if err != nil {
 		return nil, err
 	}
-
 	tp, ok := m.(TextProcessor)
 	if !ok {
-		return nil, ErrUnsupportedTokenizer
+		return nil, fmt.Errorf("%v is not a TextProcessor", m)
 	}
 	return tp, nil
 }
 
-func modelForArch(c fs.Config) (Model, error) {
-	arch := c.Architecture()
-	if pooling.Type(c.Uint("pooling_type")) != pooling.TypeNone {
-		arch = arch + "_embed"
-	}
-
-	f, ok := models[arch]
-	if !ok {
-		return nil, ErrUnsupportedModel
-	}
-
-	return f(c)
-}
-
 func populateFields(base Base, v reflect.Value, tags ...Tag) reflect.Value {
 	t := v.Type()
 
@@ -172,44 +174,35 @@ func populateFields(base Base, v reflect.Value, tags ...Tag) reflect.Value {
 			// make a copy
 			tagsCopy := tags
 			if tag := t.Field(i).Tag.Get("gguf"); tag != "" {
-				tagsCopy = append(tagsCopy, parseTag(tag))
+				tagsCopy = append(tagsCopy, ParseTags(tag))
 			}
 
 			if tt == reflect.TypeOf((*Base)(nil)).Elem() {
 				vv.Set(reflect.ValueOf(base))
 			} else if tt == reflect.TypeOf((*ml.Tensor)(nil)).Elem() {
-				var fn func([]Tag, string, string) [][]string
-				fn = func(tags []Tag, prefix, suffix string) (fullNames [][]string) {
+				var fn func([]Tag) [][]string
+				fn = func(tags []Tag) (names [][]string) {
 					if len(tags) > 0 {
-						var names []string
-						if tags[0].name != "" {
-							for _, n := range append([]string{tags[0].name}, tags[0].alternatives...) {
-								names = append(names, prefix+n+suffix)
-							}
-						}
-						childNames := fn(tags[1:], tags[0].prefix, tags[0].suffix)
-						if len(names) == 0 {
-							// current tag has no name, use child names only
-							fullNames = append(fullNames, childNames...)
-						} else if len(childNames) == 0 {
-							// current tag has names but no children, create branches for each name
-							for _, name := range names {
-								fullNames = append(fullNames, []string{name})
-							}
-						} else {
-							// merge each name with each child
-							for _, name := range names {
-								for _, childName := range childNames {
-									fullNames = append(fullNames, append([]string{name}, childName...))
+						localNames := []string{tags[0].Name}
+						localNames = append(localNames, tags[0].Alternate...)
+
+						for _, localName := range localNames {
+							fullName := []string{localName}
+							nested := fn(tags[1:])
+							if len(nested) > 0 {
+								for _, rest := range nested {
+									names = append(names, append(fullName, rest...))
 								}
+							} else {
+								names = append(names, fullName)
 							}
 						}
 					}
 
-					return fullNames
+					return names
 				}
 
-				names := fn(tagsCopy, "", "")
+				names := fn(tagsCopy)
 				for _, name := range names {
 					if tensor := base.Backend().Get(strings.Join(name, ".")); tensor != nil {
 						logutil.Trace("found tensor", "", tensor)
@@ -223,9 +216,9 @@ func populateFields(base Base, v reflect.Value, tags ...Tag) reflect.Value {
 				for i := range vv.Len() {
 					vvv := vv.Index(i)
 					if vvv.Kind() == reflect.Pointer || vvv.Kind() == reflect.Interface {
-						setPointer(base, vvv, append(tagsCopy, Tag{name: strconv.Itoa(i)}))
+						setPointer(base, vvv, append(tagsCopy, Tag{Name: strconv.Itoa(i)}))
 					} else {
-						vvv.Set(populateFields(base, vvv, append(tagsCopy, Tag{name: strconv.Itoa(i)})...))
+						vvv.Set(populateFields(base, vvv, append(tagsCopy, Tag{Name: strconv.Itoa(i)})...))
 					}
 				}
 			}
@@ -264,31 +257,18 @@ func setPointer(base Base, v reflect.Value, tags []Tag) {
 }
 
 type Tag struct {
-	name,
-	// prefix and suffix are applied to child tags
-	prefix,
-	suffix string
-	alternatives []string
+	Name      string
+	Alternate []string
 }
 
-func parseTag(s string) (tag Tag) {
+func ParseTags(s string) (tag Tag) {
 	parts := strings.Split(s, ",")
 	if len(parts) > 0 {
-		tag.name = parts[0]
+		tag.Name = parts[0]
 
 		for _, part := range parts[1:] {
-			if value, ok := strings.CutPrefix(part, "alt:"); ok && tag.name == "" {
-				// elevate alternative to primary if no primary given
-				tag.name = value
-				slog.Warn("gguf tag has alt: but no primary name", "tag", s)
-			} else if ok {
-				tag.alternatives = append(tag.alternatives, value)
-			}
-			if value, ok := strings.CutPrefix(part, "pre:"); ok {
-				tag.prefix = value
-			}
-			if value, ok := strings.CutPrefix(part, "suf:"); ok {
-				tag.suffix = value
+			if value, ok := strings.CutPrefix(part, "alt:"); ok {
+				tag.Alternate = append(tag.Alternate, value)
 			}
 		}
 	}

model/model_test.go

@@ -1,9 +1,9 @@
 package model
 
 import (
-	"errors"
 	"reflect"
 	"slices"
+	"strings"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
@@ -12,6 +12,7 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/backend/ggml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model/input"
 )
 
 func TestParseTags(t *testing.T) {
@@ -22,14 +23,14 @@ func TestParseTags(t *testing.T) {
 		{
 			value: "output",
 			want: Tag{
-				name: "output",
+				Name: "output",
 			},
 		},
 		{
 			value: "output,alt:token_embd",
 			want: Tag{
-				name: "output",
-				alternatives: []string{
+				Name: "output",
+				Alternate: []string{
 					"token_embd",
 				},
 			},
@@ -38,8 +39,8 @@ func TestParseTags(t *testing.T) {
 
 	for _, tt := range cases {
 		t.Run(tt.value, func(t *testing.T) {
-			got := parseTag(tt.value)
-			if diff := cmp.Diff(tt.want, got, cmp.AllowUnexported((Tag{}))); diff != "" {
+			got := ParseTags(tt.value)
+			if diff := cmp.Diff(tt.want, got); diff != "" {
 				t.Errorf("ParseTags() returned unexpected values (-want +got):\n%s", diff)
 			}
 		})
@@ -125,7 +126,6 @@ func TestPopulateFieldsAlternateName(t *testing.T) {
 		Input  *nn.Embedding `gguf:"input"`
 		Output *nn.Linear    `gguf:"output,alt:input"`
 		Nested *nested       `gguf:"nested"`
-		Tensor ml.Tensor     `gguf:"leaf,alt:tensor"`
 	}
 
 	var m fakeModel
@@ -134,7 +134,6 @@ func TestPopulateFieldsAlternateName(t *testing.T) {
 		names: []string{
 			"input.weight",
 			"nested.b.weight",
-			"leaf",
 		},
 	}}, v.Elem()))
 
@@ -144,115 +143,44 @@ func TestPopulateFieldsAlternateName(t *testing.T) {
 		Nested: &nested{
 			Weight: &nn.Linear{Weight: &fakeTensor{Name: "nested.b.weight"}},
 		},
-		Tensor: &fakeTensor{Name: "leaf"},
 	}, m); diff != "" {
 		t.Errorf("populateFields() set incorrect values (-want +got):\n%s", diff)
 	}
 }
 
-func TestPopulateFieldsPrefixSuffixName(t *testing.T) {
-	type fakeBlock struct {
-		A  *nn.Linear `gguf:"a"`
-		B  *nn.Linear `gguf:",pre:b_"`
-		C  *nn.Linear `gguf:",suf:_c"`
-		XY *nn.Linear `gguf:",pre:x_,suf:_y"`
+func TestGetTextProcessor(t *testing.T) {
+	tp, err := getTextProcessor(fsggml.KV{})
+	if err == nil {
+		t.Error("expected error")
+	} else if !strings.Contains(err.Error(), "unsupported model architecture") {
+		t.Errorf("unexpected error: %v", err)
+	} else if tp != nil {
+		t.Error("expected nil tp")
 	}
 
-	type fakeModel struct {
-		Blocks []fakeBlock `gguf:"blk"`
+	models["dummy"] = func(fs.Config) (Model, error) {
+		return notTextProcessorModel{}, nil
 	}
-
-	m := fakeModel{
-		Blocks: make([]fakeBlock, 2),
-	}
-	v := reflect.ValueOf(&m)
-	v.Elem().Set(populateFields(Base{b: &fakeBackend{
-		names: []string{
-			"blk.0.a.weight",
-			"blk.0.b_weight",
-			"blk.0.b_bias",
-			"blk.0.weight_c",
-			"blk.0.x_weight_y",
-			"blk.1.a.weight",
-			"blk.1.b_weight",
-			"blk.1.b_bias",
-			"blk.1.weight_c",
-			"blk.1.x_weight_y",
-		},
-	}}, v.Elem()))
-
-	if diff := cmp.Diff(fakeModel{
-		Blocks: []fakeBlock{
-			{
-				A:  &nn.Linear{Weight: &fakeTensor{Name: "blk.0.a.weight"}},
-				B:  &nn.Linear{Weight: &fakeTensor{Name: "blk.0.b_weight"}, Bias: &fakeTensor{Name: "blk.0.b_bias"}},
-				C:  &nn.Linear{Weight: &fakeTensor{Name: "blk.0.weight_c"}},
-				XY: &nn.Linear{Weight: &fakeTensor{Name: "blk.0.x_weight_y"}},
-			},
-			{
-				A:  &nn.Linear{Weight: &fakeTensor{Name: "blk.1.a.weight"}},
-				B:  &nn.Linear{Weight: &fakeTensor{Name: "blk.1.b_weight"}, Bias: &fakeTensor{Name: "blk.1.b_bias"}},
-				C:  &nn.Linear{Weight: &fakeTensor{Name: "blk.1.weight_c"}},
-				XY: &nn.Linear{Weight: &fakeTensor{Name: "blk.1.x_weight_y"}},
-			},
-		},
-	}, m); diff != "" {
-		t.Errorf("populateFields() set incorrect values (-want +got):\n%s", diff)
+	tp, err = getTextProcessor(fsggml.KV{"general.architecture": "dummy"})
+	if err == nil {
+		t.Error("expected error")
+	} else if !strings.Contains(err.Error(), "not a TextProcessor") {
+		t.Errorf("unexpected error: %v", err)
+	} else if tp != nil {
+		t.Error("expected nil tp")
 	}
 }
 
-func TestModelForArch(t *testing.T) {
-	type fakeModel struct {
-		Model
-	}
+type notTextProcessorModel struct{}
 
-	type fakeEmbeddingModel struct {
-		Model
-	}
-
-	models["model"] = func(c fs.Config) (Model, error) { return fakeModel{}, nil }
-	models["model_embed"] = func(c fs.Config) (Model, error) { return fakeEmbeddingModel{}, nil }
-
-	cases := []struct {
-		name   string
-		config fs.Config
-		want   any
-		err    error
-	}{
-		{
-			name: "model",
-			config: fsggml.KV{
-				"general.architecture": "model",
-			},
-			want: fakeModel{},
-		},
-		{
-			name: "embedding",
-			config: fsggml.KV{
-				"general.architecture": "model",
-				"model.pooling_type":   uint32(1),
-			},
-			want: fakeEmbeddingModel{},
-		},
-		{
-			name: "unsupported",
-			config: fsggml.KV{
-				"general.architecture": "unsupported",
-			},
-			err: ErrUnsupportedModel,
-		},
-	}
-
-	for _, tt := range cases {
-		t.Run(tt.name, func(t *testing.T) {
-			got, err := modelForArch(tt.config)
-			if !errors.Is(err, tt.err) {
-				t.Fatal(err)
-			}
-
-			if diff := cmp.Diff(tt.want, got); diff != "" {
-				t.Errorf("modelForArch() returned unexpected values (-want +got):\n%s", diff)
-			}
-		})
-	}
+func (notTextProcessorModel) Forward(ml.Context, input.Batch) (ml.Tensor, error) {
+	panic("unimplemented")
+}
+
+func (notTextProcessorModel) Backend() ml.Backend {
+	panic("unimplemented")
+}
+
+func (notTextProcessorModel) Config() config {
+	panic("unimplemented")
 }

model/models/bert/model.go

@@ -37,7 +37,7 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 		hiddenStates = layer.Forward(ctx, hiddenStates, &m.Options)
 	}
 
-	hiddenStates = m.poolingType.Forward(ctx, hiddenStates)
+	hiddenStates = pooling.Pooling(ctx, hiddenStates, m.poolingType)
 	if m.normalize {
 		hiddenStates = hiddenStates.L2Norm(ctx, 1e-12)
 	}

model/models/deepseek2/model.go

@@ -1,324 +0,0 @@
-package deepseek2
-
-// uses deepseek 2 architecture but written based on deepseek 3 model
-
-import (
-	"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 {
-	numExpertsUsed      int
-	numExperts          int
-	normTopKProb        bool
-	routedScalingFactor float32
-
-	kvLoraRank,
-	qkNopeHeadDim,
-	qkRopeHeadDim,
-	kqNopeHeadDim,
-	qkHeadDim int
-	qLoraRank int
-	vHeadDim  int
-
-	hiddenSize,
-	numHeads,
-	numKVHeads,
-	keyLength,
-	valueLength,
-	originalContextLength int
-
-	eps,
-	ropeBase,
-	ropeScale float32
-	kqScale float64
-}
-
-func (o Options) RoPEOptions() []func(*rope.Options) {
-	attnFactor := float32(1.0 / (1.0 + 0.1*math.Log(float64(o.ropeScale))))
-	return []func(*rope.Options){
-		rope.WithOriginalContextLength(o.originalContextLength),
-		rope.WithExtrapolationFactor(1.),
-		rope.WithAttentionFactor(attnFactor),
-	}
-}
-
-type Attention struct {
-	Q *nn.Linear `gguf:"attn_q"`
-
-	QA     *nn.Linear  `gguf:"attn_q_a"`
-	QANorm *nn.RMSNorm `gguf:"attn_q_a_norm"`
-	QB     *nn.Linear  `gguf:"attn_q_b"`
-
-	KVA     *nn.Linear  `gguf:"attn_kv_a_mqa"`
-	KVANorm *nn.RMSNorm `gguf:"attn_kv_a_norm"`
-	KVB     *nn.Linear  `gguf:"attn_kv_b"`
-
-	Output *nn.Linear `gguf:"attn_out,alt:attn_output"`
-}
-
-func (attn *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
-	seqLength := hiddenStates.Dim(1)
-
-	var query ml.Tensor
-	if opts.qLoraRank == 0 { // nil {
-		query = attn.Q.Forward(ctx, hiddenStates)
-	} else {
-		query = attn.QA.Forward(ctx, hiddenStates)
-		query = attn.QANorm.Forward(ctx, query, opts.eps)
-		query = attn.QB.Forward(ctx, query)
-	}
-
-	query = query.Reshape(ctx, query.Dim(0)/opts.numHeads, opts.numHeads, seqLength)
-
-	qPass := query.View(ctx, 0,
-		opts.qkNopeHeadDim, query.Stride(1),
-		query.Dim(1), query.Stride(2),
-		query.Dim(2))
-
-	qRot := query.View(ctx, opts.qkNopeHeadDim*query.Stride(0),
-		opts.qkRopeHeadDim, query.Stride(1),
-		query.Dim(1), query.Stride(2),
-		query.Dim(2))
-
-	compressedKV := attn.KVA.Forward(ctx, hiddenStates)
-
-	kPass := compressedKV.View(ctx, 0, opts.kvLoraRank, compressedKV.Stride(1), compressedKV.Dim(1))
-	kRot := compressedKV.View(ctx, opts.kvLoraRank*compressedKV.Stride(0),
-		opts.qkRopeHeadDim, compressedKV.Stride(1),
-		1, compressedKV.Stride(1),
-		compressedKV.Dim(1))
-
-	kPass = attn.KVANorm.Forward(ctx, kPass, opts.eps)
-	kPass = attn.KVB.Forward(ctx, kPass)
-
-	kv := kPass.Reshape(ctx, kPass.Dim(0)/opts.numKVHeads, opts.numKVHeads, seqLength)
-	kPass = kv.View(ctx, 0, opts.kqNopeHeadDim, kv.Stride(1), kv.Dim(1), kv.Stride(2), kv.Dim(2))
-	value := kv.View(ctx, opts.kqNopeHeadDim*kv.Stride(0),
-		opts.vHeadDim, kv.Stride(1),
-		kv.Dim(1), kv.Stride(2),
-		kv.Dim(2)).Contiguous(ctx)
-
-	qRot = fast.RoPE(ctx, qRot, positions, opts.qkRopeHeadDim, opts.ropeBase, 1./opts.ropeScale, opts.RoPEOptions()...)
-	kRot = fast.RoPE(ctx, kRot, positions, opts.qkRopeHeadDim, opts.ropeBase, 1./opts.ropeScale, opts.RoPEOptions()...)
-
-	kRot = kRot.Repeat(ctx, 1, qPass.Dim(1))
-
-	query = qRot.Concat(ctx, qPass, 0)
-	key := kRot.Concat(ctx, kPass, 0)
-
-	attention := nn.Attention(ctx, query, key, value, opts.kqScale, cache)
-	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), seqLength)
-	return attn.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         *nn.Linear `gguf:"ffn_gate_exps"`
-	Up           *nn.Linear `gguf:"ffn_up_exps"`
-	Down         *nn.Linear `gguf:"ffn_down_exps"`
-	SharedExpert *dense     `gguf:",suf:_shexp"`
-	ExpProbsBias ml.Tensor  `gguf:"exp_probs_b.bias,alt:exp_probs_b"`
-}
-
-func (moe *sparse) Moe(ctx ml.Context, hiddenStates, topKIndices, topKWeights ml.Tensor, opts *Options) ml.Tensor {
-	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0), 1, hiddenStates.Dim(1))
-
-	upStates := moe.Up.Weight.MulmatID(ctx, hiddenStates, topKIndices)
-	hiddenStates = moe.Gate.Weight.MulmatID(ctx, hiddenStates, topKIndices)
-	hiddenStates = hiddenStates.SILU(ctx, upStates)
-
-	experts := moe.Down.Weight.MulmatID(ctx, hiddenStates, topKIndices)
-	experts = experts.Mul(ctx, topKWeights)
-	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
-}
-
-func (moe *sparse) topKIndices(ctx ml.Context, scores ml.Tensor, opts *Options) ml.Tensor {
-	scores = scores.Add(ctx, moe.ExpProbsBias)
-	topKIndices := scores.TopK(ctx, opts.numExpertsUsed)
-	return topKIndices
-}
-
-func (moe *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
-	residuals := hiddenStates
-
-	routerLogits := moe.Router.Forward(ctx, hiddenStates)
-	scores := routerLogits.Sigmoid(ctx)
-	topKIndices := moe.topKIndices(ctx, scores, opts)
-	topKWeights := scores.Reshape(ctx, 1, opts.numExperts, hiddenStates.Dim(1)).Rows(ctx, topKIndices)
-
-	if opts.normTopKProb {
-		topKWeights = topKWeights.Reshape(ctx, opts.numExpertsUsed, hiddenStates.Dim(1))
-		topKWeights = topKWeights.Div(ctx, topKWeights.SumRows(ctx))
-		topKWeights = topKWeights.Reshape(ctx, 1, opts.numExpertsUsed, hiddenStates.Dim(1))
-	}
-
-	topKWeights = topKWeights.Scale(ctx, float64(opts.routedScalingFactor))
-	hiddenStates = moe.Moe(ctx, hiddenStates, topKIndices, topKWeights, opts)
-	sharedExpertResult := moe.SharedExpert.Forward(ctx, residuals, opts)
-
-	hiddenStates = hiddenStates.Add(ctx, sharedExpertResult)
-	return hiddenStates
-}
-
-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, opts *Options) ml.Tensor {
-	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
-	return mlp.Down.Forward(ctx, hiddenStates)
-}
-
-type Layer struct {
-	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
-	Attention     *Attention
-
-	MLPNorm *nn.RMSNorm `gguf:"ffn_norm"`
-	MLP     MLP
-}
-
-func (t *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
-	residual := hiddenStates
-	hiddenStates = t.AttentionNorm.Forward(ctx, hiddenStates, opts.eps)
-	hiddenStates = t.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 = t.MLPNorm.Forward(ctx, hiddenStates, opts.eps)
-	hiddenStates = t.MLP.Forward(ctx, hiddenStates, opts)
-	hiddenStates = hiddenStates.Add(ctx, residual)
-	return hiddenStates
-}
-
-type Model struct {
-	model.Base
-	model.BytePairEncoding
-
-	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
-	Layers         []Layer       `gguf:"blk"`
-
-	OutputNorm *nn.RMSNorm `gguf:"output_norm"`
-	Output     *nn.Linear  `gguf:"output,alt:token_embd"`
-
-	*Options
-}
-
-func New(c fs.Config) (model.Model, error) {
-	layers := make([]Layer, c.Uint("block_count"))
-
-	firstDenseLayerIndex := int(c.Uint("leading_dense_block_count"))
-	for i := range layers {
-		if i < firstDenseLayerIndex {
-			layers[i].MLP = &dense{}
-		} else {
-			layers[i].MLP = &sparse{}
-		}
-	}
-
-	mScale := float32(1.0 + float64(c.Float("rope.scaling.yarn_log_multiplier"))*math.Log(float64(c.Float("rope.scaling.factor"))))
-	kqScale := float64(mScale) * float64(mScale) / math.Sqrt(float64(c.Uint("attention.key_length")))
-
-	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&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")...,
-				),
-			},
-			// Split regex into multiple parts (according to DeepSeek3's regex)
-			"\\p{N}{1,3}",
-			`[一-龥぀-ゟ゠-ヿ]+`,
-			"[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
-		),
-		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.scaling.factor", 1),
-			numExperts:     int(c.Uint("expert_count")),
-			numExpertsUsed: int(c.Uint("expert_used_count")),
-			normTopKProb:   c.Bool("expert_weights_norm", true),
-
-			qLoraRank:     int(c.Uint("attention.q_lora_rank")), //&qLoraRankVal,
-			kvLoraRank:    int(c.Uint("attention.kv_lora_rank")),
-			qkHeadDim:     int(c.Uint("attention.key_length")),
-			vHeadDim:      int(c.Uint("attention.value_length")),
-			qkRopeHeadDim: int(c.Uint("rope.dimension_count")),
-			qkNopeHeadDim: int(c.Uint("attention.key_length")) - int(c.Uint("rope.dimension_count")),
-			kqNopeHeadDim: int(c.Uint("attention.key_length")) - int(c.Uint("rope.dimension_count")),
-
-			routedScalingFactor:   c.Float("expert_weights_scale"),
-			originalContextLength: int(c.Uint("rope.scaling.original_context_length")),
-
-			kqScale: kqScale,
-		},
-	}
-
-	m.Cache = kvcache.NewCausalCache(m.Shift)
-	return &m, nil
-}
-
-func (m Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return fast.RoPE(ctx, key, shift, m.qkRopeHeadDim, m.ropeBase, 1./m.ropeScale, m.RoPEOptions()...), nil
-}
-
-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 = 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 init() {
-	model.Register("deepseek2", New)
-}

model/models/gemma2/model.go

@@ -63,7 +63,7 @@ func New(c fs.Config) (model.Model, error) {
 			attnValLen:        int(c.Uint("attention.value_length")),
 			eps:               c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:          c.Float("rope.freq_base", 10000.0),
-			ropeScale:         c.Float("rope.scaling.factor", 1.0),
+			ropeScale:         c.Float("rope.freq_scale", 1.0),
 			attnLogitSoftcap:  c.Float("attn_logit_softcapping"),
 			finalLogitSoftcap: c.Float("final_logit_softcapping"),
 		},
@@ -88,7 +88,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
-	q = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+	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)))
@@ -98,7 +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 = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+	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)
@@ -128,7 +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 fast.RoPE(ctx, key, shift, m.Options.attnKeyLen, m.Options.ropeBase, 1/m.Options.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, m.Options.attnKeyLen, m.Options.ropeBase, m.Options.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 type MLP struct {
@@ -138,7 +138,7 @@ type MLP struct {
 }
 
 func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor {
-	hiddenState = mlp.Gate.Forward(ctx, hiddenState).GELU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).GELU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
 	return mlp.Down.Forward(ctx, hiddenState)
 }
 

model/models/gemma3/embed.go

@@ -22,7 +22,7 @@ type embedModel struct {
 
 func (m *embedModel) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 	hiddenStates := m.TextModel.Forward(ctx, batch, m.Cache)
-	hiddenStates = m.poolingType.Forward(ctx, hiddenStates)
+	hiddenStates = pooling.Pooling(ctx, hiddenStates, m.poolingType)
 	for _, dense := range m.Dense {
 		hiddenStates = dense.Forward(ctx, hiddenStates)
 	}

model/models/gemma3/model_text.go

@@ -53,10 +53,7 @@ func newTextModel(c fs.Config) *TextModel {
 			eps:            c.Float("attention.layer_norm_rms_epsilon", 1e-06),
 			ropeLocalBase:  c.Float("rope.local.freq_base", 10000.0),
 			ropeGlobalBase: c.Float("rope.global.freq_base", 1000000.0),
-			ropeScale:      1,
-			// NOTE: the rope.scaling.factor is set incorrectly in the official QAT weights
-			//       (8 instead of 1)
-			// ropeScale:      c.Float("rope.scaling.factor", 1.0),
+			ropeScale:      c.Float("rope.freq_scale", 1.0),
 		},
 	}
 
@@ -87,7 +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 = fast.RoPE(ctx, q, positionIDs, opts.attnKeyLen, ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+	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)))
@@ -98,7 +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 = fast.RoPE(ctx, k, positionIDs, opts.attnKeyLen, ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+	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)
@@ -116,7 +113,7 @@ func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.T
 		ropeBase = m.TextConfig.ropeGlobalBase
 	}
 
-	return fast.RoPE(ctx, key, shift, m.TextConfig.attnKeyLen, ropeBase, 1/m.TextConfig.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, m.TextConfig.attnKeyLen, ropeBase, m.TextConfig.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 type TextMLP struct {
@@ -126,7 +123,7 @@ type TextMLP struct {
 }
 
 func (mlp *TextMLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *TextConfig) ml.Tensor {
-	hiddenState = mlp.Gate.Forward(ctx, hiddenState).GELU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).GELU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
 	return mlp.Down.Forward(ctx, hiddenState)
 }
 

model/models/gemma3n/model_text.go

@@ -95,7 +95,7 @@ func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.T
 		ropeBase = m.ropeBaseLocal
 	}
 
-	return fast.RoPE(ctx, key, shift, m.headDim(), ropeBase, 1./m.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, m.headDim(), ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 type TextScaledWordEmbedding struct {
@@ -170,7 +170,8 @@ func (d TextLayer) Forward(ctx ml.Context, hiddenStates, perLayerInput, position
 	}
 
 	active = d.PerLayerInputGate.Forward(ctx, active)
-	active = active.GELU(ctx, perLayerInput)
+	active = active.GELU(ctx)
+	active = active.Mul(ctx, perLayerInput)
 
 	active = d.PerLayerProjection.Forward(ctx, active)
 	active = d.PostPerLayerNorm.Forward(ctx, active, opts.eps)
@@ -256,14 +257,14 @@ func (attn TextAttention) Forward(ctx ml.Context, hiddenStates, positions ml.Ten
 	query := attn.Query.Forward(ctx, hiddenStates)
 	query = query.Reshape(ctx, opts.headDim(), opts.numHeads, batchSize)
 	query = attn.QueryNorm.Forward(ctx, query, opts.eps)
-	query = fast.RoPE(ctx, query, positions, opts.headDim(), ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+	query = fast.RoPE(ctx, query, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
 
 	var key, value ml.Tensor
 	if !sharedKV {
 		key = attn.Key.Forward(ctx, hiddenStates)
 		key = key.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
 		key = attn.KeyNorm.Forward(ctx, key, opts.eps)
-		key = fast.RoPE(ctx, key, positions, opts.headDim(), ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+		key = fast.RoPE(ctx, key, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
 
 		value = attn.Value.Forward(ctx, hiddenStates)
 		value = value.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
@@ -291,7 +292,7 @@ func (mlp TextMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, activationSpa
 		hiddenStates = hiddenStates.Sub(ctx, cutoff).RELU(ctx)
 	}
 
-	hiddenStates = hiddenStates.GELU(ctx, upStates)
+	hiddenStates = hiddenStates.GELU(ctx).Mul(ctx, upStates)
 	hiddenStates = mlp.Down.Forward(ctx, hiddenStates)
 	return hiddenStates
 }
@@ -349,7 +350,7 @@ func newTextModel(c fs.Config) *TextModel {
 			eps:           c.Float("attention.layer_norm_rms_epsilon", 1e-06),
 			ropeBase:      c.Float("rope.freq_base", 1_000_000),
 			ropeBaseLocal: c.Float("rope.freq_base_local", 10_000),
-			ropeScale:     c.Float("rope.scaling.factor", 1.0),
+			ropeScale:     c.Float("rope.freq_scale", 1.0),
 
 			slidingWindowPattern:    c.Bools("attention.sliding_window_pattern"),
 			activationSparsityScale: c.Floats("activation_sparsity_scale"),

model/models/gptoss/model.go

@@ -210,7 +210,7 @@ func (mlp *MLPBlock) Forward(ctx ml.Context, hiddenStates, one ml.Tensor, opts *
 		up = mlp.Up.Forward(ctx, hiddenStates, selectedExperts)
 	}
 
-	hiddenStates = gate.SILUAlphaLimit(ctx, up, 1.702, 7)
+	hiddenStates = gate.SwiGLU(ctx, up, 1.702, 7)
 
 	experts := mlp.Down.Forward(ctx, hiddenStates, selectedExperts)
 	experts = experts.Mul(ctx, routingWeights)
@@ -227,6 +227,17 @@ func New(c fs.Config) (model.Model, error) {
 	m := Transformer{
 		TransformerBlocks: make([]TransformerBlock, c.Uint("block_count")),
 		BytePairEncoding: model.NewBytePairEncoding(
+			c.String("tokenizer.ggml.pretokenizer",
+				strings.Join([]string{
+					`[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]*[\p{Ll}\p{Lm}\p{Lo}\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?`,
+					`[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?`,
+					`\p{N}{1,3}`,
+					` ?[^\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"),
@@ -239,15 +250,6 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			strings.Join([]string{
-				`[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]*[\p{Ll}\p{Lm}\p{Lo}\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?`,
-				`[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?`,
-				`\p{N}{1,3}`,
-				` ?[^\s\p{L}\p{N}]+[\r\n/]*`,
-				`\s*[\r\n]+`,
-				`\s+(?!\S)`,
-				`\s+`,
-			}, "|"),
 		),
 		Options: Options{
 			hiddenSize:            int(c.Uint("embedding_length")),

model/models/llama/model.go

@@ -2,6 +2,7 @@ package llama
 
 import (
 	"cmp"
+	"fmt"
 	"math"
 
 	"github.com/ollama/ollama/fs"
@@ -22,80 +23,51 @@ type Options struct {
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	model.BytePairEncoding
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
 	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
 	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
 
-	Options
+	*Options
 }
 
 func New(c fs.Config) (model.Model, error) {
-	if c.Uint("expert_count") > 0 {
-		// TODO: support mixtures of experts
-		return nil, model.ErrUnsupportedModel
+	// This model currently only supports the gpt2 tokenizer
+	if c.String("tokenizer.ggml.model") == "llama" {
+		return nil, fmt.Errorf("unsupported tokenizer: llama")
 	}
-
-	var processor model.TextProcessor
-	vocabulary := model.Vocabulary{
-		Values: c.Strings("tokenizer.ggml.tokens"),
-		Scores: c.Floats("tokenizer.ggml.scores"),
-		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")...,
-		),
+	// Best effort detection of library/deepseek-coder model(s) which are incompatible
+	if c.String("general.name") == "deepseek-ai" {
+		return nil, fmt.Errorf("unsupported model: %s", c.String("general.name"))
 	}
-	switch c.String("tokenizer.ggml.model") {
-	case "gpt2":
-		var pretokenizers []string
-		switch c.String("tokenizer.ggml.pre") {
-		case "default":
-			// no-op use the default bpe pretokenizer
-		case "qwen2":
-			pretokenizers = []string{
-				"(?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+",
-			}
-		case "refact":
-			pretokenizers = []string{
-				`\p{N}`,
-				`'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+`,
-			}
-		case "tekken":
-			pretokenizers = []string{
-				"[^\\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+",
-			}
-		default:
-			// use a llama-style pretokenizer
-			pretokenizers = []string{
-				"(?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+",
-			}
-		}
-		processor = model.NewBytePairEncoding(&vocabulary, pretokenizers...)
-	case "llama":
-		processor = model.NewSentencePiece(&vocabulary)
-	default:
-		return nil, model.ErrUnsupportedTokenizer
-	}
-
 	m := Model{
-		TextProcessor: processor,
-		Layers:        make([]Layer, c.Uint("block_count")),
-		Options: Options{
+		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+`),
+			&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: make([]Layer, c.Uint("block_count")),
+		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")),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
-			ropeBase:   c.Float("rope.freq_base", 1e5),
-			ropeScale:  c.Float("rope.scaling.factor", 1),
+			ropeBase:   c.Float("rope.freq_base"),
+			ropeScale:  c.Float("rope.freq_scale", 1),
 		},
 	}
 
@@ -126,8 +98,8 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tenso
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
-	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
-	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	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))
 
 	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
 	attention = attention.Reshape(ctx, headDim*opts.numHeads, batchSize)
@@ -136,7 +108,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tenso
 
 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, 1./m.ropeScale, rope.WithFactors(m.Layers[layer].SelfAttention.RopeFactors)), nil
+	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(m.Layers[layer].SelfAttention.RopeFactors)), nil
 }
 
 type MLP struct {
@@ -146,7 +118,7 @@ type MLP struct {
 }
 
 func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor {
-	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
 	return mlp.Down.Forward(ctx, hiddenState)
 }
 
@@ -191,7 +163,7 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 			outputs = batch.Outputs
 		}
 
-		hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, 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

@@ -34,6 +34,8 @@ func (p *Projector) Forward(ctx ml.Context, visionOutputs ml.Tensor) ml.Tensor {
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
 		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}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\p{N}{1,3}| ?[^\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"),
@@ -46,7 +48,6 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			`[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]*[\p{Ll}\p{Lm}\p{Lo}\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}]+[\p{Ll}\p{Lm}\p{Lo}\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n/]*|\s*[\r\n]+|\s+(?!\S)|\s+`,
 		),
 		ImageProcessor: newImageProcessor(c),
 		VisionModel:    newVisionModel(c),

model/models/llama4/model_text.go

@@ -33,8 +33,8 @@ func (sa *TextAttention) Forward(ctx ml.Context, hiddenStates, positions, attent
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
 	if useRope {
-		query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
-		key = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+		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 {
@@ -58,14 +58,14 @@ type TextMLP struct {
 }
 
 func (mlp *TextMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
-	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
 	return mlp.Down.Forward(ctx, hiddenStates)
 }
 
 type TextExperts struct {
-	Gate *nn.LinearBatch `gguf:"ffn_gate_exps"`
-	Up   *nn.LinearBatch `gguf:"ffn_up_exps"`
-	Down *nn.LinearBatch `gguf:"ffn_down_exps"`
+	Gate *nn.Linear `gguf:"ffn_gate_exps"`
+	Up   *nn.Linear `gguf:"ffn_up_exps"`
+	Down *nn.Linear `gguf:"ffn_down_exps"`
 }
 
 func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tensor, opts *TextOptions) ml.Tensor {
@@ -76,9 +76,9 @@ func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tens
 	hiddenStates = hiddenStates.Repeat(ctx, 1, opts.numExpertsUsed)
 	hiddenStates = hiddenStates.Mul(ctx, scores)
 
-	upStates := e.Up.Forward(ctx, hiddenStates, experts)
-	gateStates := e.Gate.Forward(ctx, hiddenStates, experts)
-	downStates := e.Down.Forward(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
+	upStates := e.Up.Weight.MulmatID(ctx, hiddenStates, experts)
+	gateStates := e.Gate.Weight.MulmatID(ctx, hiddenStates, experts)
+	downStates := e.Down.Weight.MulmatID(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
 
 	nextStates := downStates.View(ctx, 0, hiddenStates.Dim(0), downStates.Stride(2), hiddenStates.Dim(2))
 	for i := 1; i < opts.numExpertsUsed; i++ {
@@ -88,10 +88,22 @@ func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tens
 	return nextStates
 }
 
+// TextSharedExpert is TextMLP with different tensor names
+type TextSharedExpert struct {
+	Gate *nn.Linear `gguf:"ffn_gate_shexp"`
+	Up   *nn.Linear `gguf:"ffn_up_shexp"`
+	Down *nn.Linear `gguf:"ffn_down_shexp"`
+}
+
+func (mlp *TextSharedExpert) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	return mlp.Down.Forward(ctx, hiddenStates)
+}
+
 type TextMOE struct {
 	Router       *nn.Linear `gguf:"ffn_gate_inp"`
 	Experts      *TextExperts
-	SharedExpert *TextMLP `gguf:",suf:_shexp"`
+	SharedExpert *TextSharedExpert
 }
 
 func (moe *TextMOE) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
@@ -184,7 +196,7 @@ func newTextModel(c fs.Config) *TextModel {
 			numExpertsUsed:             int(c.Uint("expert_used_count")),
 			ropeDim:                    int(c.Uint("rope.dimension_count")),
 			ropeBase:                   c.Float("rope.freq_base"),
-			ropeScale:                  c.Float("rope.scaling.factor", 1),
+			ropeScale:                  c.Float("rope.freq_scale", 1),
 			eps:                        c.Float("attention.layer_norm_rms_epsilon"),
 			interleaveLayerStep:        int(c.Uint("interleave_moe_layer_step", 1)),
 			noRopeInterval:             int(c.Uint("no_rope_interval", 4)),
@@ -236,5 +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 fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithFactors(m.Layers[layer].Attention.RopeFactors)), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(m.Layers[layer].Attention.RopeFactors)), nil
 }

model/models/mistral3/model.go

@@ -33,6 +33,7 @@ var _ model.TextProcessor = (*Model)(nil)
 func New(c fs.Config) (model.Model, error) {
 	m := &Model{
 		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"),
@@ -45,7 +46,6 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			`[^\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+`,
 		),
 		TextModel:           newTextModel(c),
 		VisionModel:         newVisionModel(c),

model/models/mistral3/model_text.go

@@ -40,11 +40,11 @@ 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 = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, 1./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 = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, 1./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)
@@ -55,7 +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 fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale), nil
+	return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale), nil
 }
 
 type MLP struct {
@@ -65,7 +65,7 @@ type MLP struct {
 }
 
 func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *TextOptions) ml.Tensor {
-	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
 	return mlp.Down.Forward(ctx, hiddenState)
 }
 
@@ -132,7 +132,7 @@ func newTextModel(c fs.Config) *TextModel {
 			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.scaling.factor", 1),
+			ropeScale:  c.Float("rope.freq_scale", 1),
 		},
 	}
 }

model/models/mistral3/model_vision.go

@@ -51,7 +51,7 @@ type VisionMLP struct {
 }
 
 func (mlp *VisionMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *VisionModelOptions) ml.Tensor {
-	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
 	return mlp.Down.Forward(ctx, hiddenStates)
 }
 

model/models/mllama/model.go

@@ -33,6 +33,7 @@ const (
 func New(c fs.Config) (model.Model, error) {
 	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+`),
 			&model.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -45,7 +46,6 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			`(?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+`,
 		),
 		ImageProcessor: newImageProcessor(c),
 		VisionModel:    newVisionModel(c),

model/models/mllama/model_text.go

@@ -26,11 +26,11 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.T
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	query = fast.RoPE(ctx, query, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	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 = fast.RoPE(ctx, key, positions, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithFactors(sa.RopeFactors))
+	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)
@@ -45,7 +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 fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithFactors(sa.SelfAttention.RopeFactors)), nil
+		return fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, m.ropeScale, rope.WithFactors(sa.SelfAttention.RopeFactors)), nil
 	}
 
 	return key, nil
@@ -58,7 +58,7 @@ type TextMLP struct {
 }
 
 func (mlp *TextMLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *TextModelOptions) ml.Tensor {
-	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
 	return mlp.Down.Forward(ctx, hiddenState)
 }
 
@@ -244,7 +244,7 @@ func newTextModel(c fs.Config) *TextModel {
 			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.scaling.factor", 1),
+			ropeScale:            c.Float("rope.freq_scale", 1),
 			crossAttentionLayers: c.Ints("attention.cross_attention_layers"),
 		},
 	}

model/models/models.go

@@ -2,7 +2,6 @@ package models
 
 import (
 	_ "github.com/ollama/ollama/model/models/bert"
-	_ "github.com/ollama/ollama/model/models/deepseek2"
 	_ "github.com/ollama/ollama/model/models/gemma2"
 	_ "github.com/ollama/ollama/model/models/gemma3"
 	_ "github.com/ollama/ollama/model/models/gemma3n"

model/models/qwen2/model.go

@@ -43,8 +43,8 @@ func (attn Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor,
 	value := attn.Value.Forward(ctx, hiddenStates)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
 
-	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
-	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+	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)
@@ -59,7 +59,7 @@ type MLP struct {
 }
 
 func (mlp MLP) Forward(ctx ml.Context, hiddenStates ml.Tensor) ml.Tensor {
-	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
 	return mlp.Down.Forward(ctx, hiddenStates)
 }
 
@@ -124,7 +124,7 @@ func (m Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 
 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, 1./m.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 func New(c fs.Config) (model.Model, error) {
@@ -139,6 +139,7 @@ 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"),
@@ -151,7 +152,6 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			`(?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+`,
 		),
 		Options: Options{
 			hiddenSize: int(c.Uint("embedding_length")),
@@ -160,7 +160,7 @@ func New(c fs.Config) (model.Model, error) {
 			headDim:    int(c.Uint("attention.key_length")),
 			ropeDim:    int(c.Uint("rope.dimension_count")),
 			ropeBase:   c.Float("rope.freq_base"),
-			ropeScale:  c.Float("rope.scaling.factor", 1),
+			ropeScale:  c.Float("rope.freq_scale", 1),
 			eps:        c.Float("attention.layer_norm_rms_epsilon"),
 		},
 	}

model/models/qwen25vl/model.go

@@ -29,6 +29,7 @@ var _ model.MultimodalProcessor = (*Model)(nil)
 func New(c fs.Config) (model.Model, error) {
 	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}| ?[^\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"),
@@ -41,7 +42,6 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			`(?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+`,
 		),
 		TextModel:      NewTextModel(c),
 		VisionModel:    newVisionModel(c),

model/models/qwen25vl/model_text.go

@@ -38,7 +38,7 @@ func NewTextModel(c fs.Config) *TextModel {
 			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.scaling.factor", 1),
+			ropeScale:             c.Float("rope.freq_scale", 1),
 		},
 	}
 
@@ -60,11 +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 = fast.RoPE(ctx, q, positionIDs, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
+	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 = fast.RoPE(ctx, k, positionIDs, opts.ropeDim, opts.ropeBase, 1./opts.ropeScale, rope.WithOriginalContextLength(opts.originalContextLength), rope.WithTypeNeoX())
+	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)
@@ -78,7 +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 fast.RoPE(ctx, key, shift, m.ropeDim, m.ropeBase, 1./m.ropeScale, rope.WithOriginalContextLength(m.originalContextLength), rope.WithTypeNeoX()), 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
@@ -90,7 +90,7 @@ type MLP struct {
 
 func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *TextOptions) ml.Tensor {
 	// Apply SwiGLU activation gating
-	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
 	// Project back to hidden dimension
 	return mlp.Down.Forward(ctx, hiddenState)
 }

model/models/qwen25vl/model_vision.go

@@ -100,7 +100,8 @@ type VisionMLP struct {
 func (mlp *VisionMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *VisionModelOptions) ml.Tensor {
 	// Using activation as specified in config (likely GELU or SiLU/Swish)
 	gateOutput := mlp.Gate.Forward(ctx, hiddenStates)
-	hiddenStates = gateOutput.SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	upOutput := mlp.Up.Forward(ctx, hiddenStates)
+	hiddenStates = gateOutput.SILU(ctx).Mul(ctx, upOutput)
 
 	return mlp.Down.Forward(ctx, hiddenStates)
 }

model/models/qwen3/embed.go

@@ -1,73 +0,0 @@
-package qwen3
-
-import (
-	"github.com/ollama/ollama/fs"
-	"github.com/ollama/ollama/kvcache"
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/ml/nn/pooling"
-	"github.com/ollama/ollama/model"
-	"github.com/ollama/ollama/model/input"
-)
-
-type embedModel struct {
-	model.Base
-	model.BytePairEncoding
-
-	*Model
-	poolingType pooling.Type
-}
-
-func (m *embedModel) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	hiddenStates, err := m.forward(ctx, batch)
-	if err != nil {
-		return nil, err
-	}
-
-	hiddenStates = m.poolingType.Forward(ctx, hiddenStates)
-	hiddenStates = hiddenStates.L2Norm(ctx, 1e-12)
-	return hiddenStates, nil
-}
-
-func newEmbed(c fs.Config) (model.Model, error) {
-	layers := make([]Layer, c.Uint("block_count"))
-	for i := range layers {
-		layers[i].MLP = &dense{}
-	}
-	m := embedModel{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&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")...,
-				),
-			},
-			`(?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: &Model{
-			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),
-			},
-		},
-		poolingType: pooling.Type(c.Uint("pooling_type")),
-	}
-
-	m.Cache = kvcache.NewCausalCache(m.Shift)
-	return &m, nil
-}

model/models/qwen3/model.go

@@ -30,10 +30,10 @@ func (o Options) headDim() int {
 }
 
 type Attention struct {
-	Query     *nn.Linear  `gguf:"attn_q"`
 	QueryNorm *nn.RMSNorm `gguf:"attn_q_norm"`
-	Key       *nn.Linear  `gguf:"attn_k"`
+	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"`
 }
@@ -52,8 +52,8 @@ func (sa *Attention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor,
 	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, 1./opts.ropeScale, rope.WithTypeNeoX())
-	key = fast.RoPE(ctx, key, positions, opts.headDim(), opts.ropeBase, 1./opts.ropeScale, rope.WithTypeNeoX())
+	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)
@@ -65,10 +65,10 @@ type MLP interface {
 }
 
 type sparse struct {
-	Router *nn.Linear      `gguf:"ffn_gate_inp"`
-	Gate   *nn.LinearBatch `gguf:"ffn_gate_exps"`
-	Up     *nn.LinearBatch `gguf:"ffn_up_exps"`
-	Down   *nn.LinearBatch `gguf:"ffn_down_exps"`
+	Router *nn.Linear `gguf:"ffn_gate_inp"`
+	Gate   *nn.Linear `gguf:"ffn_gate_exps"`
+	Up     *nn.Linear `gguf:"ffn_up_exps"`
+	Down   *nn.Linear `gguf:"ffn_down_exps"`
 }
 
 func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
@@ -87,9 +87,13 @@ func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options
 
 	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0), 1, hiddenStates.Dim(1))
 
-	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates, selectedExperts).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates, selectedExperts))
+	upStates := mlp.Up.Weight.MulmatID(ctx, hiddenStates, selectedExperts)
 
-	experts := mlp.Down.Forward(ctx, hiddenStates, selectedExperts)
+	hiddenStates = mlp.Gate.Weight.MulmatID(ctx, hiddenStates, selectedExperts)
+	hiddenStates = hiddenStates.SILU(ctx)
+	hiddenStates = hiddenStates.Mul(ctx, upStates)
+
+	experts := mlp.Down.Weight.MulmatID(ctx, hiddenStates, selectedExperts)
 	experts = experts.Mul(ctx, routingWeights)
 
 	nextStates := experts.View(ctx, 0, experts.Dim(0), experts.Stride(2), experts.Dim(2))
@@ -107,8 +111,7 @@ type dense struct {
 }
 
 func (mlp *dense) Forward(ctx ml.Context, hiddenStates ml.Tensor, _ *Options) ml.Tensor {
-	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).
-		SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenStates))
 	return mlp.Down.Forward(ctx, hiddenStates)
 }
 
@@ -151,25 +154,14 @@ type Model struct {
 	*Options
 }
 
-func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	hiddenStates, err := m.forward(ctx, batch)
-	if err != nil {
-		return nil, err
-	}
-
-	return m.Output.Forward(ctx, hiddenStates), nil
-}
-
 // Forward implements model.Model.
-func (m *Model) forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+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 {
-		if m.Cache != nil {
-			m.Cache.SetLayer(i)
-		}
+		m.Cache.SetLayer(i)
 
 		var outputs ml.Tensor
 		if i == len(m.Layers)-1 {
@@ -179,11 +171,12 @@ func (m *Model) forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 		hiddenStates = layer.Forward(ctx, hiddenStates, positions, outputs, m.Cache, m.Options)
 	}
 
-	return m.OutputNorm.Forward(ctx, hiddenStates, m.eps), nil
+	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, 1./m.ropeScale, rope.WithTypeNeoX()), nil
+	return fast.RoPE(ctx, key, shift, m.headDim(), m.ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
 }
 
 var _ model.Model = (*Model)(nil)
@@ -200,6 +193,7 @@ func New(c fs.Config) (model.Model, error) {
 
 	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"),
@@ -212,7 +206,6 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			`(?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+`,
 		),
 		Layers: layers,
 		Options: &Options{
@@ -223,7 +216,7 @@ func New(c fs.Config) (model.Model, error) {
 			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.scaling.factor", 1),
+			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),
@@ -237,5 +230,4 @@ func New(c fs.Config) (model.Model, error) {
 func init() {
 	model.Register("qwen3", New)
 	model.Register("qwen3moe", New)
-	model.Register("qwen3_embed", newEmbed)
 }

model/parsers/parsers.go

@@ -1,49 +0,0 @@
-package parsers
-
-import (
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/harmony"
-)
-
-type Parser interface {
-	// Init initializes the parser with tools and optional last message for chat prefill
-	// Returns processed tools if the parser needs to modify them (e.g., harmony renames them)
-	Init(tools []api.Tool, lastMessage *api.Message) []api.Tool
-	// Add processes streamed content and returns parsed content, thinking, and tool calls
-	// The done flag indicates if this is the last chunk (used for draining accumulators)
-	Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error)
-	HasToolSupport() bool
-	HasThinkingSupport() bool
-}
-
-func ParserForName(name string) Parser {
-	switch name {
-	case "qwen3-coder":
-		parser := &Qwen3CoderParser{}
-		return parser
-	case "passthrough":
-		return &PassthroughParser{}
-	case "harmony":
-		return harmony.NewHarmonyMessageHandler()
-	default:
-		return nil
-	}
-}
-
-type PassthroughParser struct{}
-
-func (p *PassthroughParser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
-	return tools // passthrough doesn't modify tools
-}
-
-func (p *PassthroughParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	return s, "", nil, nil
-}
-
-func (p *PassthroughParser) HasToolSupport() bool {
-	return false
-}
-
-func (p *PassthroughParser) HasThinkingSupport() bool {
-	return false
-}

model/parsers/qwen3coder.go

@@ -1,463 +0,0 @@
-package parsers
-
-import (
-	"context"
-	"encoding/json"
-	"encoding/xml"
-	"fmt"
-	"log/slog"
-	"math"
-	"regexp"
-	"strconv"
-	"strings"
-	"unicode"
-	"unicode/utf8"
-
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/logutil"
-)
-
-type qwenParserState int
-
-const (
-	toolOpenTag  = "<tool_call>"
-	toolCloseTag = "</tool_call>"
-)
-
-const (
-	qwenParserState_LookingForToolStart qwenParserState = iota
-	qwenParserState_CollectingToolContent
-)
-
-type Qwen3CoderParser struct {
-	state qwenParserState
-	acc   strings.Builder
-	tools []api.Tool
-}
-
-func (p *Qwen3CoderParser) HasToolSupport() bool {
-	return true
-}
-
-func (p *Qwen3CoderParser) HasThinkingSupport() bool {
-	return false
-}
-
-func (p *Qwen3CoderParser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
-	p.tools = tools
-	return tools // Qwen doesn't modify tools
-}
-
-func (p *Qwen3CoderParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	p.acc.WriteString(s)
-
-	events := p.parseEvents()
-
-	var toolCalls []api.ToolCall
-	var sb strings.Builder
-	for _, event := range events {
-		switch event := event.(type) {
-		case qwenEventRawToolCall:
-			toolCall, err := parseToolCall(event, p.tools)
-			if err != nil {
-				slog.Warn("qwen tool call parsing failed", "error", err)
-				return "", "", nil, err
-			}
-			toolCalls = append(toolCalls, toolCall)
-		case qwenEventContent:
-			// TODO(drifkin): if the same turn contains multiple interleaved content
-			// events, we naively append them together here. See the note below about
-			// `qwenEvent`s for more details
-			sb.WriteString(event.content)
-		}
-	}
-
-	return sb.String(), "", toolCalls, nil
-}
-
-func (p *Qwen3CoderParser) parseEvents() []qwenEvent {
-	var all []qwenEvent
-
-	keepLooping := true
-	for keepLooping {
-		var events []qwenEvent
-		events, keepLooping = eat(p)
-		if len(events) > 0 {
-			all = append(all, events...)
-		}
-	}
-
-	if len(all) > 0 {
-		slog.Log(context.TODO(), logutil.LevelTrace, "qwen events parsed", "events", all, "state", p.state, "acc", p.acc.String())
-	}
-
-	return all
-}
-
-// we use some internal event types in order to communicate between `Add` and
-// `eat`. We do this to support interleaving content and parallel tool calls in
-// the parser, even though qwen3-coder isn't supposed to do this. Our API
-// doesn't currently support models outputting multiple messages in a turn, so
-// we wouldn't be able to represent it yet, but there's no reason to prevent the
-// parser from supporting it, especially for future models if they end up using
-// a similar format.
-type qwenEvent interface {
-	isQwenEvent()
-}
-
-type qwenEventRawToolCall struct {
-	raw string
-}
-
-type qwenEventContent struct {
-	content string
-}
-
-func (qwenEventContent) isQwenEvent()     {}
-func (qwenEventRawToolCall) isQwenEvent() {}
-
-// eat consumes the parser's buffer, and returns a list of any unambiguous
-// events from the current parser state. If the parser transitions to another
-// state, it may have additional events to emit on the next call, which is what
-// the second return value indicates
-func eat(p *Qwen3CoderParser) ([]qwenEvent, bool) {
-	var events []qwenEvent
-
-	switch p.state {
-	case qwenParserState_LookingForToolStart:
-		if strings.Contains(p.acc.String(), toolOpenTag) {
-			// we found a full tool open tag, so we can emit the content before the
-			// tag, being sure to trim any trailing whitespace
-			split := strings.SplitN(p.acc.String(), toolOpenTag, 2)
-			before := split[0]
-			before = strings.TrimRightFunc(before, unicode.IsSpace)
-			if len(before) > 0 {
-				events = append(events, qwenEventContent{content: before})
-			}
-			after := split[1]
-			p.acc.Reset()
-			p.acc.WriteString(after)
-			p.state = qwenParserState_CollectingToolContent
-			return events, true
-		} else if overlap := overlap(p.acc.String(), toolOpenTag); overlap > 0 {
-			// we found a partial tool open tag, so we can emit the unambiguous part,
-			// which is the (trailing-whitespace trimmed) content before the partial
-			// tool open tag
-			beforePartialTag := p.acc.String()[:len(p.acc.String())-overlap]
-			trailingWhitespaceLen := trailingWhitespaceLen(beforePartialTag)
-			ambiguousStart := len(beforePartialTag) - trailingWhitespaceLen
-			unambiguous := p.acc.String()[:ambiguousStart]
-			ambiguous := p.acc.String()[ambiguousStart:]
-			p.acc.Reset()
-			p.acc.WriteString(ambiguous)
-			events = append(events, qwenEventContent{content: unambiguous})
-			return events, false
-		} else {
-			// we found content that is entirely not a tool call. We should withhold
-			// any trailing whitespace in case this is the end of the content
-			whitespaceLen := trailingWhitespaceLen(p.acc.String())
-			ambiguousStart := len(p.acc.String()) - whitespaceLen
-			unambiguous := p.acc.String()[:ambiguousStart]
-			ambiguous := p.acc.String()[ambiguousStart:]
-			p.acc.Reset()
-			p.acc.WriteString(ambiguous)
-			if len(unambiguous) > 0 {
-				events = append(events, qwenEventContent{content: unambiguous})
-			}
-			return events, false
-		}
-	case qwenParserState_CollectingToolContent:
-		if strings.Contains(p.acc.String(), toolCloseTag) {
-			split := strings.SplitN(p.acc.String(), toolCloseTag, 2)
-			before := split[0]
-			if len(before) == 0 {
-				slog.Warn("qwen tool call closing tag found but no content before it")
-			}
-			// remove any whitespace between the tool call and any content after it
-			after := strings.TrimLeftFunc(split[1], unicode.IsSpace)
-			p.acc.Reset()
-			p.acc.WriteString(after)
-			events = append(events, qwenEventRawToolCall{raw: before})
-			p.state = qwenParserState_LookingForToolStart
-			return events, true
-		} else {
-			// note that we don't need to check the overlap here because we only plan
-			// on parsing the tool call once we see the full closing tag. We don't
-			// stream back the unparsed tool content, so there's no need to be eager
-			// here
-			return events, false
-		}
-	default:
-		panic("unreachable")
-	}
-}
-
-// TODO(drifkin): move this to a shared location
-// 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
-}
-
-func trailingWhitespaceLen(s string) int {
-	remaining := s
-	total := 0
-	for len(remaining) > 0 {
-		r, size := utf8.DecodeLastRuneInString(remaining)
-		// if it's an invalid utf8 rune, assume it isn't whitespace
-		if r == utf8.RuneError && size == 1 {
-			break
-		}
-		if !unicode.IsSpace(r) {
-			break
-		}
-		total += size
-		remaining = remaining[:len(remaining)-size]
-	}
-	return total
-}
-
-type XMLFunctionCall struct {
-	XMLName    xml.Name       `xml:"function"`
-	Name       string         `xml:"name,attr"`
-	Parameters []XMLParameter `xml:"parameter"`
-}
-
-type XMLParameter struct {
-	Name  string `xml:"name,attr"`
-	Value string `xml:",chardata"`
-}
-
-// parseToolCall parses a raw tool call string into an api.ToolCall.
-// The raw string follows an xml-like format, here's an example:
-//
-// <function=get_current_temperature>
-// <parameter=location>
-// San Francisco
-// </parameter>
-// <parameter=unit>
-// celsius
-// </parameter>
-// </function>
-func parseToolCall(raw qwenEventRawToolCall, tools []api.Tool) (api.ToolCall, error) {
-	toolCall := api.ToolCall{}
-
-	xmlString := transformToXML(raw.raw)
-
-	var functionCall XMLFunctionCall
-	err := xml.Unmarshal([]byte(xmlString), &functionCall)
-	if err != nil {
-		return api.ToolCall{}, err
-	}
-
-	toolCall.Function = api.ToolCallFunction{
-		Name: functionCall.Name,
-	}
-
-	// Find the matching tool to get parameter types
-	var matchedTool *api.Tool
-	for i := range tools {
-		if tools[i].Function.Name == functionCall.Name {
-			matchedTool = &tools[i]
-			break
-		}
-	}
-
-	toolCall.Function.Arguments = make(api.ToolCallFunctionArguments)
-	for _, parameter := range functionCall.Parameters {
-		// Look up the parameter type if we found the tool
-		var paramType api.PropertyType
-		if matchedTool != nil && matchedTool.Function.Parameters.Properties != nil {
-			if prop, ok := matchedTool.Function.Parameters.Properties[parameter.Name]; ok {
-				paramType = prop.Type
-			}
-		}
-
-		toolCall.Function.Arguments[parameter.Name] = parseValue(parameter.Value, paramType)
-	}
-
-	return toolCall, nil
-}
-
-// parseValue converts a raw string value to the appropriate type based on the parameter type specification.
-//
-// For union types (multiple types in PropertyType, which we support but doesn't
-// seem as though the reference parser does type coercion with those types in
-// mind) we use a type precedence approach:
-// 1. null - checked first regardless of declared types (matches reference implementation)
-// 2. boolean - only "true"/"false" are valid booleans
-// 3. integer - must parse as a whole number
-// 4. number - must parse as numeric (returns int if no decimal part)
-// 5. array - must parse as valid JSON array
-// 6. object - must parse as valid JSON object
-// 7. string - always succeeds (least specific type)
-//
-// This precedence ensures we return the most specific type that successfully parses,
-// following the principle of least surprise. For example, with PropertyType{"string", "number"},
-// "123" becomes 123 (number), while "hello" becomes "hello" (string).
-func parseValue(raw string, paramType api.PropertyType) any {
-	// first remove a single leading newlines, and a single trailing newline (if
-	// they exist). This follows the reference implementation
-	raw = strings.TrimPrefix(raw, "\n")
-	raw = strings.TrimSuffix(raw, "\n")
-
-	// Check for null first (case-insensitive) - this takes precedence over any type
-	if strings.ToLower(raw) == "null" {
-		return nil
-	}
-
-	// If no type is specified, default to string
-	if len(paramType) == 0 {
-		return raw
-	}
-
-	// Check if any of the specified types match, using type precedence
-	// Order: boolean -> integer -> number -> array -> object -> string
-	typeSet := make(map[string]bool)
-	for _, t := range paramType {
-		typeSet[t] = true
-	}
-
-	// Try boolean first (most restrictive)
-	if typeSet["boolean"] {
-		lower := strings.ToLower(raw)
-		switch lower {
-		case "true":
-			return true
-		case "false":
-			return false
-		}
-		// If not a valid boolean but boolean is the only type, return false (matching reference)
-		if len(paramType) == 1 {
-			return false
-		}
-		// Otherwise try other types
-	}
-
-	// Try integer
-	if typeSet["integer"] {
-		if i, err := strconv.ParseInt(raw, 10, 64); err == nil {
-			// Return as int if it fits in int32, otherwise int64
-			if i >= math.MinInt32 && i <= math.MaxInt32 {
-				return int(i)
-			}
-			return i
-		}
-		// If integer is the only type and parsing failed, fall back to string
-		if len(paramType) == 1 {
-			return raw
-		}
-	}
-
-	// Try number (float)
-	if typeSet["number"] {
-		if f, err := strconv.ParseFloat(raw, 64); err == nil {
-			// If the number has no decimal part, return as int (matching reference)
-			if f == math.Trunc(f) {
-				i := int64(f)
-				if i >= math.MinInt32 && i <= math.MaxInt32 {
-					return int(i)
-				}
-				return i
-			}
-			return f
-		}
-		// If number is the only type and parsing failed, fall back to string
-		if len(paramType) == 1 {
-			return raw
-		}
-	}
-
-	// Try array
-	if typeSet["array"] {
-		var arr []any
-		if err := json.Unmarshal([]byte(raw), &arr); err == nil {
-			return arr
-		}
-		// If array is the only type and parsing failed, fall back to string
-		if len(paramType) == 1 {
-			return raw
-		}
-	}
-
-	// Try object
-	if typeSet["object"] {
-		var obj map[string]any
-		if err := json.Unmarshal([]byte(raw), &obj); err == nil {
-			return obj
-		}
-		// If object is the only type and parsing failed, fall back to string
-		if len(paramType) == 1 {
-			return raw
-		}
-	}
-
-	// String always succeeds (or if "string" is in the type set)
-	if typeSet["string"] {
-		return raw
-	}
-
-	// If we get here, none of the types matched and string wasn't an option
-	// We return string as a fallback. The reference implementation will attempt
-	// to parse the value as a python literal, but we purposefully don't support
-	// that
-	return raw
-}
-
-var (
-	qwenTagRegex    = regexp.MustCompile(`<(\w+)=([^>]+)>`)
-	qwenXMLTagRegex = regexp.MustCompile(`</?(?:function|parameter)(?:\s+name="[^"]*")?>`)
-)
-
-// transformToXML transforms a raw qwen tool call with xml-like tags into valid
-// xml so that it can be parsed by any xml parser
-func transformToXML(raw string) string {
-	// take the form `<tag=abc>` and transform it to `<tag name="abc">`, taking
-	// care to properly escape the string that becomes the attribute value
-	transformed := qwenTagRegex.ReplaceAllStringFunc(raw, func(match string) string {
-		groups := qwenTagRegex.FindStringSubmatch(match)
-		tag := groups[1]
-		var escapedValue strings.Builder
-		xml.EscapeText(&escapedValue, []byte(groups[2]))
-		return fmt.Sprintf(`<%s name="%s">`, tag, escapedValue.String())
-	})
-
-	// Walk the resulting string, escaping any character data that sits between the
-	// xml tags we just emitted
-	var out strings.Builder
-	lastIdx := 0
-	for _, loc := range qwenXMLTagRegex.FindAllStringIndex(transformed, -1) {
-		if loc[0] > lastIdx {
-			escapeTextNode(&out, transformed[lastIdx:loc[0]])
-		}
-		out.WriteString(transformed[loc[0]:loc[1]])
-		lastIdx = loc[1]
-	}
-	if lastIdx < len(transformed) {
-		escapeTextNode(&out, transformed[lastIdx:])
-	}
-
-	return out.String()
-}
-
-// escapeTextNode escapes XML character data without altering other characters
-// like newlines or tabs (which is why we don't use xml.EscapeText for this)
-func escapeTextNode(sb *strings.Builder, s string) {
-	for _, r := range s {
-		switch r {
-		case '&':
-			sb.WriteString("&amp;")
-		case '<':
-			sb.WriteString("&lt;")
-		case '>':
-			sb.WriteString("&gt;")
-		default:
-			sb.WriteRune(r)
-		}
-	}
-}

model/renderers/qwen3coder.go

@@ -1,236 +0,0 @@
-package renderers
-
-import (
-	"encoding/json"
-	"fmt"
-	"reflect"
-	"strings"
-
-	"github.com/ollama/ollama/api"
-)
-
-var (
-	imStartTag = "<|im_start|>"
-	imEndTag   = "<|im_end|>"
-)
-
-// renderAdditionalKeys renders all JSON fields except the ones in handledKeys
-// This follows the same approach from the reference implementation, which gives
-// a particular key ordering
-func renderAdditionalKeys(obj any, handledKeys map[string]bool) string {
-	data, err := json.Marshal(obj)
-	if err != nil {
-		return ""
-	}
-
-	var m map[string]any
-	if err := json.Unmarshal(data, &m); err != nil {
-		return ""
-	}
-
-	var sb strings.Builder
-	for key, value := range m {
-		if handledKeys[key] {
-			continue
-		}
-
-		// Check if value is a map or array (needs JSON serialization)
-		switch v := value.(type) {
-		case map[string]any, []any:
-			jsonBytes, _ := json.Marshal(v)
-			// TODO(drifkin): it would be nice to format the JSON here similarly to
-			// python's default json.dumps behavior (spaces after commas and colons).
-			// This would let us be byte-for-byte compatible with the reference
-			// implementation for most common inputs
-			jsonStr := string(jsonBytes)
-			sb.WriteString("\n<" + key + ">" + jsonStr + "</" + key + ">")
-		case nil:
-			continue
-		default:
-			// Simple types, convert to string
-			sb.WriteString("\n<" + key + ">" + fmt.Sprintf("%v", value) + "</" + key + ">")
-		}
-	}
-
-	return sb.String()
-}
-
-func Qwen3CoderRenderer(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error) {
-	var sb strings.Builder
-
-	// filter out system messages and choose the first (if any) to win
-	var systemMessage string
-	var filteredMessages []api.Message
-	for _, message := range messages {
-		if message.Role != "system" {
-			filteredMessages = append(filteredMessages, message)
-			continue
-		}
-
-		if systemMessage == "" {
-			systemMessage = message.Content
-		}
-	}
-
-	if systemMessage != "" || len(tools) > 0 {
-		sb.WriteString(imStartTag + "system\n")
-
-		// if we have tools but no system message, match the reference implementation by providing a default system message
-		if systemMessage == "" {
-			systemMessage = "You are Qwen, a helpful AI assistant that can interact with a computer to solve tasks."
-		}
-
-		sb.WriteString(systemMessage)
-
-		if len(tools) > 0 {
-			sb.WriteString("\n\n# Tools\n\nYou have access to the following functions:\n\n")
-			sb.WriteString("<tools>")
-			for _, tool := range tools {
-				sb.WriteString("\n")
-				sb.WriteString("<function>\n")
-				sb.WriteString("<name>" + tool.Function.Name + "</name>")
-				if tool.Function.Description != "" {
-					sb.WriteString("\n<description>" + tool.Function.Description + "</description>")
-				}
-				sb.WriteString("\n<parameters>")
-
-				for name, prop := range tool.Function.Parameters.Properties {
-					sb.WriteString("\n<parameter>")
-					sb.WriteString("\n<name>" + name + "</name>")
-
-					if len(prop.Type) > 0 {
-						sb.WriteString("\n<type>" + formatToolDefinitionType(prop.Type) + "</type>")
-					}
-
-					if prop.Description != "" {
-						sb.WriteString("\n<description>" + prop.Description + "</description>")
-					}
-
-					// Render any additional keys not already handled
-					handledKeys := map[string]bool{
-						"type":        true,
-						"description": true,
-					}
-					sb.WriteString(renderAdditionalKeys(prop, handledKeys))
-
-					sb.WriteString("\n</parameter>")
-				}
-
-				// Render extra keys for parameters (everything except 'type' and 'properties')
-				paramHandledKeys := map[string]bool{
-					"type":       true,
-					"properties": true,
-				}
-				sb.WriteString(renderAdditionalKeys(tool.Function.Parameters, paramHandledKeys))
-
-				sb.WriteString("\n</parameters>")
-				sb.WriteString("\n</function>")
-			}
-			sb.WriteString("\n</tools>")
-			sb.WriteString("\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>")
-		}
-
-		sb.WriteString(imEndTag + "\n")
-	}
-
-	for i, message := range filteredMessages {
-		lastMessage := i == len(filteredMessages)-1
-		prefill := lastMessage && message.Role == "assistant"
-		switch message.Role {
-		case "assistant":
-			if len(message.ToolCalls) > 0 {
-				sb.WriteString(imStartTag + "assistant\n")
-				if message.Content != "" {
-					sb.WriteString(message.Content + "\n")
-				}
-				for _, toolCall := range message.ToolCalls {
-					sb.WriteString("\n<tool_call>\n<function=" + toolCall.Function.Name + ">")
-					for name, value := range toolCall.Function.Arguments {
-						valueStr := formatToolCallArgument(value)
-						sb.WriteString("\n<parameter=" + name + ">\n" + valueStr + "\n</parameter>")
-					}
-					sb.WriteString("\n</function>\n</tool_call>")
-				}
-				sb.WriteString("<|im_end|>\n")
-			} else {
-				sb.WriteString(imStartTag + "assistant\n")
-				sb.WriteString(message.Content)
-				if !prefill {
-					sb.WriteString(imEndTag + "\n")
-				}
-			}
-		case "tool":
-			// consecutive tool responses should share a single `<im_start>user`, but
-			// have their own <tool_response> tags
-
-			// only start a new user block if this is the first tool response
-			if i == 0 || filteredMessages[i-1].Role != "tool" {
-				sb.WriteString(imStartTag + "user\n")
-			}
-
-			sb.WriteString("<tool_response>\n")
-			sb.WriteString(message.Content)
-			sb.WriteString("\n</tool_response>\n")
-
-			// close the user block only if this is the last tool response
-			if i == len(filteredMessages)-1 || filteredMessages[i+1].Role != "tool" {
-				sb.WriteString(imEndTag + "\n")
-			}
-		default:
-			sb.WriteString(imStartTag + message.Role + "\n")
-			sb.WriteString(message.Content)
-			sb.WriteString(imEndTag + "\n")
-		}
-
-		if lastMessage && !prefill {
-			sb.WriteString(imStartTag + "assistant\n")
-		}
-	}
-
-	return sb.String(), nil
-}
-
-func formatToolCallArgument(value any) string {
-	if value == nil {
-		return "null"
-	}
-
-	switch v := value.(type) {
-	case string:
-		return v
-	case []byte:
-		return string(v)
-	}
-
-	if reflect.TypeOf(value) != nil {
-		kind := reflect.TypeOf(value).Kind()
-		if kind == reflect.Map || kind == reflect.Slice || kind == reflect.Array {
-			if marshalled, err := json.Marshal(value); err == nil {
-				return string(marshalled)
-			}
-		}
-	}
-
-	return fmt.Sprintf("%v", value)
-}
-
-func formatToolDefinitionType(tp api.PropertyType) string {
-	if len(tp) == 0 {
-		return "[]"
-	}
-
-	if len(tp) == 1 {
-		return tp[0]
-	}
-
-	// TODO(drifkin): it would be nice to format the JSON here similarly to
-	// python's default json.dumps behavior (spaces after commas and colons).
-	// This would let us be byte-for-byte compatible with the reference
-	// implementation for most common inputs
-	jsonBytes, err := json.Marshal(tp)
-	if err != nil {
-		return "[]"
-	}
-
-	return string(jsonBytes)
-}

model/renderers/qwen3coder_test.go

@@ -1,370 +0,0 @@
-package renderers
-
-import (
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-	"github.com/ollama/ollama/api"
-)
-
-func TestQwen3CoderRenderer(t *testing.T) {
-	tests := []struct {
-		name     string
-		msgs     []api.Message
-		tools    []api.Tool
-		expected string
-	}{
-		{
-			name: "basic",
-			msgs: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant."},
-				{Role: "user", Content: "Hello, how are you?"},
-			},
-			expected: `<|im_start|>system
-You are a helpful assistant.<|im_end|>
-<|im_start|>user
-Hello, how are you?<|im_end|>
-<|im_start|>assistant
-`,
-		},
-		{
-			name: "with tools and response",
-			msgs: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant with access to tools."},
-				{Role: "user", Content: "What is the weather like in San Francisco?"},
-				{
-					Role:    "assistant",
-					Content: "I'll check the weather in San Francisco for you.",
-					ToolCalls: []api.ToolCall{
-						{
-							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: map[string]any{
-									"unit": "fahrenheit",
-								},
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: "{\"location\": \"San Francisco, CA\", \"temperature\": 68, \"condition\": \"partly cloudy\", \"humidity\": 65, \"wind_speed\": 12}", ToolName: "get_weather"},
-				{Role: "user", Content: "That sounds nice! What about New York?"},
-			},
-			tools: []api.Tool{
-				{Function: api.ToolFunction{
-					Name:        "get_weather",
-					Description: "Get the current weather in a given location",
-					Parameters: api.ToolFunctionParameters{
-						Required: []string{"unit"},
-						Properties: map[string]api.ToolProperty{
-							"unit": {Type: api.PropertyType{"string"}, Enum: []any{"celsius", "fahrenheit"}, Description: "The unit of temperature"},
-							// TODO(drifkin): add multiple params back once we have predictable
-							// order via some sort of ordered map type (see
-							// <https://github.com/ollama/ollama/issues/12244>)
-							/*
-								"location": {Type: api.PropertyType{"string"}, Description: "The city and state, e.g. San Francisco, CA"},
-							*/
-						},
-					},
-				}},
-			},
-			expected: `<|im_start|>system
-You are a helpful assistant with access to tools.
-
-# Tools
-
-You have access to the following functions:
-
-<tools>
-<function>
-<name>get_weather</name>
-<description>Get the current weather in a given location</description>
-<parameters>
-<parameter>
-<name>unit</name>
-<type>string</type>
-<description>The unit of temperature</description>
-<enum>["celsius","fahrenheit"]</enum>
-</parameter>
-<required>["unit"]</required>
-</parameters>
-</function>
-</tools>
-
-If you choose to call a function ONLY reply in the following format with NO suffix:
-
-<tool_call>
-<function=example_function_name>
-<parameter=example_parameter_1>
-value_1
-</parameter>
-<parameter=example_parameter_2>
-This is the value for the second parameter
-that can span
-multiple lines
-</parameter>
-</function>
-</tool_call>
-
-<IMPORTANT>
-Reminder:
-- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags
-- Required parameters MUST be specified
-- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after
-- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls
-</IMPORTANT><|im_end|>
-<|im_start|>user
-What is the weather like in San Francisco?<|im_end|>
-<|im_start|>assistant
-I'll check the weather in San Francisco for you.
-
-<tool_call>
-<function=get_weather>
-<parameter=unit>
-fahrenheit
-</parameter>
-</function>
-</tool_call><|im_end|>
-<|im_start|>user
-<tool_response>
-{"location": "San Francisco, CA", "temperature": 68, "condition": "partly cloudy", "humidity": 65, "wind_speed": 12}
-</tool_response>
-<|im_end|>
-<|im_start|>user
-That sounds nice! What about New York?<|im_end|>
-<|im_start|>assistant
-`,
-		},
-		{
-			name: "parallel tool calls",
-			msgs: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant with access to tools."},
-				{Role: "user", Content: "call double(1) and triple(2)"},
-				{Role: "assistant", Content: "I'll call double(1) and triple(2) for you.", ToolCalls: []api.ToolCall{
-					{Function: api.ToolCallFunction{Name: "double", Arguments: map[string]any{"number": "1"}}},
-					{Function: api.ToolCallFunction{Name: "triple", Arguments: map[string]any{"number": "2"}}},
-				}},
-				{Role: "tool", Content: "{\"number\": 2}", ToolName: "double"},
-				{Role: "tool", Content: "{\"number\": 6}", ToolName: "triple"},
-			},
-			tools: []api.Tool{
-				{Function: api.ToolFunction{Name: "double", Description: "Double a number", Parameters: api.ToolFunctionParameters{Properties: map[string]api.ToolProperty{
-					"number": {Type: api.PropertyType{"string"}, Description: "The number to double"},
-				}}}},
-				{Function: api.ToolFunction{Name: "triple", Description: "Triple a number", Parameters: api.ToolFunctionParameters{Properties: map[string]api.ToolProperty{
-					"number": {Type: api.PropertyType{"string"}, Description: "The number to triple"},
-				}}}},
-			},
-			expected: `<|im_start|>system
-You are a helpful assistant with access to tools.
-
-# Tools
-
-You have access to the following functions:
-
-<tools>
-<function>
-<name>double</name>
-<description>Double a number</description>
-<parameters>
-<parameter>
-<name>number</name>
-<type>string</type>
-<description>The number to double</description>
-</parameter>
-</parameters>
-</function>
-<function>
-<name>triple</name>
-<description>Triple a number</description>
-<parameters>
-<parameter>
-<name>number</name>
-<type>string</type>
-<description>The number to triple</description>
-</parameter>
-</parameters>
-</function>
-</tools>
-
-If you choose to call a function ONLY reply in the following format with NO suffix:
-
-<tool_call>
-<function=example_function_name>
-<parameter=example_parameter_1>
-value_1
-</parameter>
-<parameter=example_parameter_2>
-This is the value for the second parameter
-that can span
-multiple lines
-</parameter>
-</function>
-</tool_call>
-
-<IMPORTANT>
-Reminder:
-- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags
-- Required parameters MUST be specified
-- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after
-- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls
-</IMPORTANT><|im_end|>
-<|im_start|>user
-call double(1) and triple(2)<|im_end|>
-<|im_start|>assistant
-I'll call double(1) and triple(2) for you.
-
-<tool_call>
-<function=double>
-<parameter=number>
-1
-</parameter>
-</function>
-</tool_call>
-<tool_call>
-<function=triple>
-<parameter=number>
-2
-</parameter>
-</function>
-</tool_call><|im_end|>
-<|im_start|>user
-<tool_response>
-{"number": 2}
-</tool_response>
-<tool_response>
-{"number": 6}
-</tool_response>
-<|im_end|>
-<|im_start|>assistant
-`,
-		},
-		{
-			name: "prefill",
-			msgs: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant."},
-				{Role: "user", Content: "Tell me something interesting."},
-				{Role: "assistant", Content: "I'll tell you something interesting about cats"},
-			},
-			expected: `<|im_start|>system
-You are a helpful assistant.<|im_end|>
-<|im_start|>user
-Tell me something interesting.<|im_end|>
-<|im_start|>assistant
-I'll tell you something interesting about cats`,
-		},
-		{
-			name: "complex tool call arguments should remain json encoded",
-			msgs: []api.Message{
-				{Role: "user", Content: "call tool"},
-				{Role: "assistant", ToolCalls: []api.ToolCall{
-					{Function: api.ToolCallFunction{
-						Name: "echo",
-						Arguments: map[string]any{
-							"payload": map[string]any{"foo": "bar"},
-						},
-					}},
-				}},
-				{Role: "tool", Content: "{\"payload\": {\"foo\": \"bar\"}}", ToolName: "echo"},
-			},
-			expected: `<|im_start|>user
-call tool<|im_end|>
-<|im_start|>assistant
-
-<tool_call>
-<function=echo>
-<parameter=payload>
-{"foo":"bar"}
-</parameter>
-</function>
-</tool_call><|im_end|>
-<|im_start|>user
-<tool_response>
-{"payload": {"foo": "bar"}}
-</tool_response>
-<|im_end|>
-<|im_start|>assistant
-`,
-		},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			rendered, err := Qwen3CoderRenderer(tt.msgs, tt.tools, nil)
-			if err != nil {
-				t.Fatal(err)
-			}
-			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
-				t.Errorf("mismatch (-got +want):\n%s", diff)
-			}
-		})
-	}
-}
-
-func TestFormatToolCallArgument(t *testing.T) {
-	tests := []struct {
-		name     string
-		arg      any
-		expected string
-	}{
-		{
-			name: "string",
-			arg:  "foo",
-			// notice no quotes around the string
-			expected: "foo",
-		},
-		{
-			name:     "map",
-			arg:      map[string]any{"foo": "bar"},
-			expected: "{\"foo\":\"bar\"}",
-		},
-		{
-			name:     "number",
-			arg:      1,
-			expected: "1",
-		},
-		{
-			name:     "boolean",
-			arg:      true,
-			expected: "true",
-		},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			got := formatToolCallArgument(tt.arg)
-			if got != tt.expected {
-				t.Errorf("formatToolCallArgument(%v) = %v, want %v", tt.arg, got, tt.expected)
-			}
-		})
-	}
-}
-
-func TestQwen3ToolDefinitionTypes(t *testing.T) {
-	tests := []struct {
-		name         string
-		propertyType api.PropertyType
-		expected     string
-	}{
-		{
-			name:         "simple",
-			propertyType: api.PropertyType{"string"},
-			expected:     "string",
-		},
-		{
-			name:         "multiple",
-			propertyType: api.PropertyType{"string", "number"},
-			expected:     "[\"string\",\"number\"]",
-		},
-		{
-			name:         "empty",
-			propertyType: api.PropertyType{},
-			expected:     "[]",
-		},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			got := formatToolDefinitionType(tt.propertyType)
-			if got != tt.expected {
-				t.Errorf("formatToolDefinitionType() = %v, want %v", got, tt.expected)
-			}
-		})
-	}
-}

model/renderers/renderer.go

@@ -1,26 +0,0 @@
-package renderers
-
-import (
-	"fmt"
-
-	"github.com/ollama/ollama/api"
-)
-
-type rendererFunc func([]api.Message, []api.Tool, *api.ThinkValue) (string, error)
-
-func RenderWithRenderer(name string, msgs []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error) {
-	renderer := rendererForName(name)
-	if renderer == nil {
-		return "", fmt.Errorf("unknown renderer %q", name)
-	}
-	return renderer(msgs, tools, think)
-}
-
-func rendererForName(name string) rendererFunc {
-	switch name {
-	case "qwen3-coder":
-		return Qwen3CoderRenderer
-	default:
-		return nil
-	}
-}

openai/openai.go

@@ -105,18 +105,16 @@ type ChatCompletionRequest struct {
 	Tools            []api.Tool      `json:"tools"`
 	Reasoning        *Reasoning      `json:"reasoning,omitempty"`
 	ReasoningEffort  *string         `json:"reasoning_effort,omitempty"`
-	DebugRenderOnly  bool            `json:"_debug_render_only"`
 }
 
 type ChatCompletion struct {
-	Id                string         `json:"id"`
-	Object            string         `json:"object"`
-	Created           int64          `json:"created"`
-	Model             string         `json:"model"`
-	SystemFingerprint string         `json:"system_fingerprint"`
-	Choices           []Choice       `json:"choices"`
-	Usage             Usage          `json:"usage,omitempty"`
-	DebugInfo         *api.DebugInfo `json:"_debug_info,omitempty"`
+	Id                string   `json:"id"`
+	Object            string   `json:"object"`
+	Created           int64    `json:"created"`
+	Model             string   `json:"model"`
+	SystemFingerprint string   `json:"system_fingerprint"`
+	Choices           []Choice `json:"choices"`
+	Usage             Usage    `json:"usage,omitempty"`
 }
 
 type ChatCompletionChunk struct {
@@ -143,7 +141,6 @@ type CompletionRequest struct {
 	Temperature      *float32       `json:"temperature"`
 	TopP             float32        `json:"top_p"`
 	Suffix           string         `json:"suffix"`
-	DebugRenderOnly  bool           `json:"_debug_render_only"`
 }
 
 type Completion struct {
@@ -276,8 +273,8 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 				}
 				return nil
 			}(r.DoneReason),
-		}}, Usage: toUsage(r),
-		DebugInfo: r.DebugInfo,
+		}},
+		Usage: toUsage(r),
 	}
 }
 
@@ -571,14 +568,13 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 	}
 
 	return &api.ChatRequest{
-		Model:           r.Model,
-		Messages:        messages,
-		Format:          format,
-		Options:         options,
-		Stream:          &r.Stream,
-		Tools:           r.Tools,
-		Think:           think,
-		DebugRenderOnly: r.DebugRenderOnly,
+		Model:    r.Model,
+		Messages: messages,
+		Format:   format,
+		Options:  options,
+		Stream:   &r.Stream,
+		Tools:    r.Tools,
+		Think:    think,
 	}, nil
 }
 
@@ -652,12 +648,11 @@ func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
 	}
 
 	return api.GenerateRequest{
-		Model:           r.Model,
-		Prompt:          r.Prompt,
-		Options:         options,
-		Stream:          &r.Stream,
-		Suffix:          r.Suffix,
-		DebugRenderOnly: r.DebugRenderOnly,
+		Model:   r.Model,
+		Prompt:  r.Prompt,
+		Options: options,
+		Stream:  &r.Stream,
+		Suffix:  r.Suffix,
 	}, nil
 }
 

parser/parser.go

@@ -62,14 +62,15 @@ func (f Modelfile) CreateRequest(relativeDir string) (*api.CreateRequest, error)
 	for _, c := range f.Commands {
 		switch c.Name {
 		case "model":
-			path, err := expandPath(c.Args, relativeDir)
+			name := c.Args.(string)
+			path, err := expandPath(name, relativeDir)
 			if err != nil {
 				return nil, err
 			}
 
 			digestMap, err := fileDigestMap(path)
 			if errors.Is(err, os.ErrNotExist) {
-				req.From = c.Args
+				req.From = name
 				continue
 			} else if err != nil {
 				return nil, err
@@ -83,7 +84,8 @@ func (f Modelfile) CreateRequest(relativeDir string) (*api.CreateRequest, error)
 				}
 			}
 		case "adapter":
-			path, err := expandPath(c.Args, relativeDir)
+			adapter := c.Args.(string)
+			path, err := expandPath(adapter, relativeDir)
 			if err != nil {
 				return nil, err
 			}
@@ -95,25 +97,25 @@ func (f Modelfile) CreateRequest(relativeDir string) (*api.CreateRequest, error)
 
 			req.Adapters = digestMap
 		case "template":
-			req.Template = c.Args
+			template := c.Args.(string)
+			req.Template = template
 		case "system":
-			req.System = c.Args
+			system := c.Args.(string)
+			req.System = system
 		case "license":
-			licenses = append(licenses, c.Args)
-		case "renderer":
-			req.Renderer = c.Args
-		case "parser":
-			req.Parser = c.Args
+			license := c.Args.(string)
+			licenses = append(licenses, license)
 		case "message":
-			role, msg, _ := strings.Cut(c.Args, ": ")
-			messages = append(messages, api.Message{Role: role, Content: msg})
-		default:
+			msg := c.Args.(*Message)
+			messages = append(messages, api.Message{Role: msg.Role, Content: msg.Content})
+		case "parameter":
 			if slices.Contains(deprecatedParameters, c.Name) {
-				fmt.Printf("warning: parameter %s is deprecated\n", c.Name)
+				fmt.Printf("warning: parameter '%s' is deprecated\n", c.Name)
 				break
 			}
 
-			ps, err := api.FormatParams(map[string][]string{c.Name: {c.Args}})
+			param := c.Args.(*Parameter)
+			ps, err := api.FormatParams(map[string][]string{param.Name: {param.Value}})
 			if err != nil {
 				return nil, err
 			}
@@ -127,6 +129,8 @@ func (f Modelfile) CreateRequest(relativeDir string) (*api.CreateRequest, error)
 					params[k] = v
 				}
 			}
+		default:
+			return nil, fmt.Errorf("warning: unknown command '%s'", c.Name)
 		}
 	}
 
@@ -316,7 +320,17 @@ func filesForModel(path string) ([]string, error) {
 
 type Command struct {
 	Name string
-	Args string
+	Args any
+}
+
+type Parameter struct {
+	Name  string
+	Value string
+}
+
+type Message struct {
+	Role    string
+	Content string
 }
 
 func (c Command) String() string {
@@ -324,13 +338,17 @@ func (c Command) String() string {
 	switch c.Name {
 	case "model":
 		fmt.Fprintf(&sb, "FROM %s", c.Args)
-	case "license", "template", "system", "adapter", "renderer", "parser":
-		fmt.Fprintf(&sb, "%s %s", strings.ToUpper(c.Name), quote(c.Args))
+	case "license", "template", "system", "adapter":
+		data := c.Args.(string)
+		fmt.Fprintf(&sb, "%s %s", strings.ToUpper(c.Name), quote(data))
 	case "message":
-		role, message, _ := strings.Cut(c.Args, ": ")
-		fmt.Fprintf(&sb, "MESSAGE %s %s", role, quote(message))
+		data := c.Args.(*Message)
+		fmt.Fprintf(&sb, "MESSAGE %s %s", data.Role, quote(data.Content))
+	case "parameter":
+		data := c.Args.(*Parameter)
+		fmt.Fprintf(&sb, "PARAMETER %s %s", data.Name, quote(data.Value))
 	default:
-		fmt.Fprintf(&sb, "PARAMETER %s %s", c.Name, quote(c.Args))
+		fmt.Printf("unknown command '%s'\n", c.Name)
 	}
 
 	return sb.String()
@@ -350,7 +368,7 @@ const (
 var (
 	errMissingFrom        = errors.New("no FROM line")
 	errInvalidMessageRole = errors.New("message role must be one of \"system\", \"user\", or \"assistant\"")
-	errInvalidCommand     = errors.New("command must be one of \"from\", \"license\", \"template\", \"system\", \"adapter\", \"renderer\", \"parser\", \"parameter\", or \"message\"")
+	errInvalidCommand     = errors.New("command must be one of \"from\", \"license\", \"template\", \"system\", \"adapter\", \"parameter\", or \"message\"")
 )
 
 type ParserError struct {
@@ -370,7 +388,6 @@ func ParseFile(r io.Reader) (*Modelfile, error) {
 	var curr state
 	var currLine int = 1
 	var b bytes.Buffer
-	var role string
 
 	var f Modelfile
 
@@ -417,6 +434,7 @@ func ParseFile(r io.Reader) (*Modelfile, error) {
 				case "parameter":
 					// transition to stateParameter which sets command name
 					next = stateParameter
+					cmd.Name = s
 				case "message":
 					// transition to stateMessage which validates the message role
 					next = stateMessage
@@ -425,16 +443,37 @@ func ParseFile(r io.Reader) (*Modelfile, error) {
 					cmd.Name = s
 				}
 			case stateParameter:
-				cmd.Name = b.String()
+				s, ok := unquote(strings.TrimSpace(b.String()))
+				if !ok || isSpace(r) {
+					if _, err := b.WriteRune(r); err != nil {
+						return nil, err
+					}
+
+					continue
+				}
+				cmd.Args = &Parameter{
+					Name: s,
+				}
 			case stateMessage:
-				if !isValidMessageRole(b.String()) {
+				s, ok := unquote(strings.TrimSpace(b.String()))
+				if !ok || isSpace(r) {
+					if _, err := b.WriteRune(r); err != nil {
+						return nil, err
+					}
+
+					continue
+				}
+
+				if !isValidMessageRole(s) {
 					return nil, &ParserError{
 						LineNumber: currLine,
 						Msg:        errInvalidMessageRole.Error(),
 					}
 				}
 
-				role = b.String()
+				cmd.Args = &Message{
+					Role: s,
+				}
 			case stateComment, stateNil:
 				// pass
 			case stateValue:
@@ -447,12 +486,16 @@ func ParseFile(r io.Reader) (*Modelfile, error) {
 					continue
 				}
 
-				if role != "" {
-					s = role + ": " + s
-					role = ""
+				switch cmd.Name {
+				case "parameter":
+					p := cmd.Args.(*Parameter)
+					p.Value = s
+				case "message":
+					m := cmd.Args.(*Message)
+					m.Content = s
+				default:
+					cmd.Args = s
 				}
-
-				cmd.Args = s
 				f.Commands = append(f.Commands, cmd)
 			}
 
@@ -477,11 +520,16 @@ func ParseFile(r io.Reader) (*Modelfile, error) {
 			return nil, io.ErrUnexpectedEOF
 		}
 
-		if role != "" {
-			s = role + ": " + s
+		switch cmd.Name {
+		case "parameter":
+			c := cmd.Args.(*Parameter)
+			c.Value = s
+		case "message":
+			c := cmd.Args.(*Message)
+			c.Content = s
+		default:
+			cmd.Args = s
 		}
-
-		cmd.Args = s
 		f.Commands = append(f.Commands, cmd)
 	default:
 		return nil, io.ErrUnexpectedEOF
@@ -610,7 +658,7 @@ func isValidMessageRole(role string) bool {
 
 func isValidCommand(cmd string) bool {
 	switch strings.ToLower(cmd) {
-	case "from", "license", "template", "system", "adapter", "renderer", "parser", "parameter", "message":
+	case "from", "license", "template", "system", "adapter", "parameter", "message":
 		return true
 	default:
 		return false

parser/parser_test.go

@@ -47,8 +47,8 @@ TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
 		{Name: "model", Args: "model1"},
 		{Name: "adapter", Args: "adapter1"},
 		{Name: "license", Args: "MIT"},
-		{Name: "param1", Args: "value1"},
-		{Name: "param2", Args: "value2"},
+		{Name: "parameter", Args: &Parameter{"param1", "value1"}},
+		{Name: "parameter", Args: &Parameter{"param2", "value2"}},
 		{Name: "template", Args: "{{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>"},
 	}
 
@@ -80,8 +80,8 @@ TEMPLATE """   {{ if .System }}<|start_header_id|>system<|end_header_id|>
 		{Name: "model", Args: "     model 1"},
 		{Name: "adapter", Args: "adapter3"},
 		{Name: "license", Args: "MIT       "},
-		{Name: "param1", Args: "value1"},
-		{Name: "param2", Args: "value2"},
+		{Name: "parameter", Args: &Parameter{"param1", "value1"}},
+		{Name: "parameter", Args: &Parameter{"param2", "value2"}},
 		{Name: "template", Args: "   {{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>   "},
 	}
 
@@ -101,7 +101,7 @@ func TestParseFileFrom(t *testing.T) {
 		},
 		{
 			"FROM \"FOO BAR\"\nPARAMETER param1 value1",
-			[]Command{{Name: "model", Args: "FOO BAR"}, {Name: "param1", Args: "value1"}},
+			[]Command{{Name: "model", Args: "FOO BAR"}, {Name: "parameter", Args: &Parameter{"param1", "value1"}}},
 			nil,
 		},
 		{
@@ -149,12 +149,12 @@ func TestParseFileFrom(t *testing.T) {
 		},
 		{
 			"PARAMETER param1 value1\nFROM foo",
-			[]Command{{Name: "param1", Args: "value1"}, {Name: "model", Args: "foo"}},
+			[]Command{{Name: "parameter", Args: &Parameter{"param1", "value1"}}, {Name: "model", Args: "foo"}},
 			nil,
 		},
 		{
 			"PARAMETER what the \nFROM lemons make lemonade ",
-			[]Command{{Name: "what", Args: "the"}, {Name: "model", Args: "lemons make lemonade"}},
+			[]Command{{Name: "parameter", Args: &Parameter{"what", "the"}}, {Name: "model", Args: "lemons make lemonade"}},
 			nil,
 		},
 	}
@@ -198,34 +198,6 @@ BADCOMMAND param1 value1
 	}
 }
 
-func TestParseFileRenderer(t *testing.T) {
-	input := `
-FROM foo
-RENDERER renderer1
-`
-
-	reader := strings.NewReader(input)
-
-	modelfile, err := ParseFile(reader)
-	require.NoError(t, err)
-
-	assert.Equal(t, []Command{{Name: "model", Args: "foo"}, {Name: "renderer", Args: "renderer1"}}, modelfile.Commands)
-}
-
-func TestParseFileParser(t *testing.T) {
-	input := `
-FROM foo
-PARSER parser1
-`
-
-	reader := strings.NewReader(input)
-
-	modelfile, err := ParseFile(reader)
-	require.NoError(t, err)
-
-	assert.Equal(t, []Command{{Name: "model", Args: "foo"}, {Name: "parser", Args: "parser1"}}, modelfile.Commands)
-}
-
 func TestParseFileMessages(t *testing.T) {
 	cases := []struct {
 		input    string
@@ -239,7 +211,7 @@ MESSAGE system You are a file parser. Always parse things.
 `,
 			[]Command{
 				{Name: "model", Args: "foo"},
-				{Name: "message", Args: "system: You are a file parser. Always parse things."},
+				{Name: "message", Args: &Message{"system", "You are a file parser. Always parse things."}},
 			},
 			nil,
 		},
@@ -249,7 +221,7 @@ FROM foo
 MESSAGE system You are a file parser. Always parse things.`,
 			[]Command{
 				{Name: "model", Args: "foo"},
-				{Name: "message", Args: "system: You are a file parser. Always parse things."},
+				{Name: "message", Args: &Message{"system", "You are a file parser. Always parse things."}},
 			},
 			nil,
 		},
@@ -262,9 +234,9 @@ MESSAGE assistant Hello, I want to parse all the things!
 `,
 			[]Command{
 				{Name: "model", Args: "foo"},
-				{Name: "message", Args: "system: You are a file parser. Always parse things."},
-				{Name: "message", Args: "user: Hey there!"},
-				{Name: "message", Args: "assistant: Hello, I want to parse all the things!"},
+				{Name: "message", Args: &Message{"system", "You are a file parser. Always parse things."}},
+				{Name: "message", Args: &Message{"user", "Hey there!"}},
+				{Name: "message", Args: &Message{"assistant", "Hello, I want to parse all the things!"}},
 			},
 			nil,
 		},
@@ -272,12 +244,12 @@ MESSAGE assistant Hello, I want to parse all the things!
 			`
 FROM foo
 MESSAGE system """
-You are a multiline file parser. Always parse things.
+You are a multiline file "parser". Always parse things.
 """
 			`,
 			[]Command{
 				{Name: "model", Args: "foo"},
-				{Name: "message", Args: "system: \nYou are a multiline file parser. Always parse things.\n"},
+				{Name: "message", Args: &Message{"system", "\nYou are a multiline file \"parser\". Always parse things.\n"}},
 			},
 			nil,
 		},
@@ -542,7 +514,7 @@ func TestParseFileParameters(t *testing.T) {
 
 			assert.Equal(t, []Command{
 				{Name: "model", Args: "foo"},
-				{Name: v.name, Args: v.value},
+				{Name: "parameter", Args: &Parameter{v.name, v.value}},
 			}, modelfile.Commands)
 		})
 	}
@@ -645,8 +617,8 @@ SYSTEM You are a utf16 file.
 
 	expected := []Command{
 		{Name: "model", Args: "bob"},
-		{Name: "param1", Args: "1"},
-		{Name: "param2", Args: "4096"},
+		{Name: "parameter", Args: &Parameter{"param1", "1"}},
+		{Name: "parameter", Args: &Parameter{"param2", "4096"}},
 		{Name: "system", Args: "You are a utf16 file."},
 	}
 

runner/llamarunner/cache.go

@@ -204,8 +204,13 @@ func (c *InputCache) ShiftDiscard(inputLen int, numKeep int) int {
 	targetFree = max(targetFree, 1)
 
 	currentFree := c.numCtx - inputLen
+	discard := targetFree - currentFree
 
-	return max(targetFree-currentFree, 0)
+	if discard < 0 {
+		discard = 0
+	}
+
+	return discard
 }
 
 type ErrReprocessInputs struct {

runner/ollamarunner/cache.go

@@ -242,8 +242,13 @@ func (c *InputCache) ShiftDiscard(inputLen int32, numKeep int32) int32 {
 	targetFree = max(targetFree, 1)
 
 	currentFree := c.numCtx - inputLen
+	discard := targetFree - currentFree
 
-	return max(targetFree-currentFree, 0)
+	if discard < 0 {
+		discard = 0
+	}
+
+	return discard
 }
 
 type ErrReprocessInputs struct {

runner/ollamarunner/runner.go

@@ -11,6 +11,7 @@ import (
 	"image"
 	"log"
 	"log/slog"
+	"math"
 	"net"
 	"net/http"
 	"os"
@@ -31,7 +32,6 @@ import (
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
 	"github.com/ollama/ollama/runner/common"
@@ -405,7 +405,7 @@ func (s *Server) removeSequence(seqIndex int, reason llm.DoneReason) {
 func (s *Server) run(ctx context.Context) {
 	s.ready.Wait()
 
-	supportsAsync := pooling.Type(s.model.Backend().Config().Uint("pooling_type")) == pooling.TypeNone
+	supportsAsync := s.model.Backend().Config().Uint("pooling_type", math.MaxUint32) == math.MaxUint32
 
 	var activeBatch batchState
 	for {
@@ -900,7 +900,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 }
 
 func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
-	if pooling.Type(s.model.Backend().Config().Uint("pooling_type")) == pooling.TypeNone {
+	if s.model.Backend().Config().Uint("pooling_type", math.MaxUint32) == math.MaxUint32 {
 		http.Error(w, "this model does not support embeddings", http.StatusNotImplemented)
 		return
 	}

sample/samplers_test.go

@@ -82,6 +82,7 @@ func modelHelper(t testing.TB) model.BytePairEncoding {
 	merges := make([]string, 0, 1)
 	// Only need vocab for Grammar Test
 	return model.NewBytePairEncoding(
+		``,
 		&model.Vocabulary{
 			Values: tokens,
 			Types:  make([]int32, len(vocab)),

scripts/env.sh

@@ -16,7 +16,6 @@ OLLAMA_COMMON_BUILD_ARGS="--build-arg=VERSION \
     --build-arg=OLLAMA_FAST_BUILD \
     --build-arg=CUSTOM_CPU_FLAGS \
     --build-arg=GPU_RUNNER_CPU_FLAGS \
-    --build-arg=PARALLEL \
     --build-arg=AMDGPU_TARGETS"
 
 echo "Building Ollama"

server/create.go

@@ -10,11 +10,8 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
-	"net"
 	"net/http"
-	"net/url"
 	"os"
-	"path"
 	"path/filepath"
 	"slices"
 	"strings"
@@ -42,14 +39,6 @@ var (
 )
 
 func (s *Server) CreateHandler(c *gin.Context) {
-	config := &ConfigV2{
-		OS:           "linux",
-		Architecture: "amd64",
-		RootFS: RootFS{
-			Type: "layers",
-		},
-	}
-
 	var r api.CreateRequest
 	if err := c.ShouldBindJSON(&r); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
@@ -59,9 +48,6 @@ func (s *Server) CreateHandler(c *gin.Context) {
 		return
 	}
 
-	config.Renderer = r.Renderer
-	config.Parser = r.Parser
-
 	for v := range r.Files {
 		if !fs.ValidPath(v) {
 			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": errFilePath.Error()})
@@ -91,34 +77,20 @@ func (s *Server) CreateHandler(c *gin.Context) {
 		oldManifest, _ := ParseNamedManifest(name)
 
 		var baseLayers []*layerGGML
-		var err error
-		var remote bool
-
 		if r.From != "" {
-			slog.Debug("create model from model name", "from", r.From)
+			slog.Debug("create model from model name")
 			fromName := model.ParseName(r.From)
 			if !fromName.IsValid() {
 				ch <- gin.H{"error": errtypes.InvalidModelNameErrMsg, "status": http.StatusBadRequest}
 				return
 			}
-			if r.RemoteHost != "" {
-				ru, err := remoteURL(r.RemoteHost)
-				if err != nil {
-					ch <- gin.H{"error": "bad remote", "status": http.StatusBadRequest}
-					return
-				}
 
-				config.RemoteModel = r.From
-				config.RemoteHost = ru
-				remote = true
-			} else {
-				ctx, cancel := context.WithCancel(c.Request.Context())
-				defer cancel()
+			ctx, cancel := context.WithCancel(c.Request.Context())
+			defer cancel()
 
-				baseLayers, err = parseFromModel(ctx, fromName, fn)
-				if err != nil {
-					ch <- gin.H{"error": err.Error()}
-				}
+			baseLayers, err = parseFromModel(ctx, fromName, fn)
+			if err != nil {
+				ch <- gin.H{"error": err.Error()}
 			}
 		} else if r.Files != nil {
 			baseLayers, err = convertModelFromFiles(r.Files, baseLayers, false, fn)
@@ -138,7 +110,7 @@ func (s *Server) CreateHandler(c *gin.Context) {
 		}
 
 		var adapterLayers []*layerGGML
-		if !remote && r.Adapters != nil {
+		if r.Adapters != nil {
 			adapterLayers, err = convertModelFromFiles(r.Adapters, baseLayers, true, fn)
 			if err != nil {
 				for _, badReq := range []error{errNoFilesProvided, errOnlyOneAdapterSupported, errOnlyGGUFSupported, errUnknownType, errFilePath} {
@@ -156,56 +128,7 @@ func (s *Server) CreateHandler(c *gin.Context) {
 			baseLayers = append(baseLayers, adapterLayers...)
 		}
 
-		// Info is not currently exposed by Modelfiles, but allows overriding various
-		// config values
-		if r.Info != nil {
-			caps, ok := r.Info["capabilities"]
-			if ok {
-				switch tcaps := caps.(type) {
-				case []any:
-					caps := make([]string, len(tcaps))
-					for i, c := range tcaps {
-						str, ok := c.(string)
-						if !ok {
-							continue
-						}
-						caps[i] = str
-					}
-					config.Capabilities = append(config.Capabilities, caps...)
-				}
-			}
-
-			strFromInfo := func(k string) string {
-				v, ok := r.Info[k]
-				if ok {
-					val := v.(string)
-					return val
-				}
-				return ""
-			}
-
-			vFromInfo := func(k string) float64 {
-				v, ok := r.Info[k]
-				if ok {
-					val := v.(float64)
-					return val
-				}
-				return 0
-			}
-
-			config.ModelFamily = strFromInfo("model_family")
-			if config.ModelFamily != "" {
-				config.ModelFamilies = []string{config.ModelFamily}
-			}
-
-			config.BaseName = strFromInfo("base_name")
-			config.FileType = strFromInfo("quantization_level")
-			config.ModelType = strFromInfo("parameter_size")
-			config.ContextLen = int(vFromInfo("context_length"))
-			config.EmbedLen = int(vFromInfo("embedding_length"))
-		}
-
-		if err := createModel(r, name, baseLayers, config, fn); err != nil {
+		if err := createModel(r, name, baseLayers, fn); err != nil {
 			if errors.Is(err, errBadTemplate) {
 				ch <- gin.H{"error": err.Error(), "status": http.StatusBadRequest}
 				return
@@ -231,51 +154,6 @@ func (s *Server) CreateHandler(c *gin.Context) {
 	streamResponse(c, ch)
 }
 
-func remoteURL(raw string) (string, error) {
-	// Special‑case: user supplied only a path ("/foo/bar").
-	if strings.HasPrefix(raw, "/") {
-		return (&url.URL{
-			Scheme: "http",
-			Host:   net.JoinHostPort("localhost", "11434"),
-			Path:   path.Clean(raw),
-		}).String(), nil
-	}
-
-	if !strings.Contains(raw, "://") {
-		raw = "http://" + raw
-	}
-
-	if raw == "ollama.com" || raw == "http://ollama.com" {
-		raw = "https://ollama.com:443"
-	}
-
-	u, err := url.Parse(raw)
-	if err != nil {
-		return "", fmt.Errorf("parse error: %w", err)
-	}
-
-	if u.Host == "" {
-		u.Host = "localhost"
-	}
-
-	hostPart, portPart, err := net.SplitHostPort(u.Host)
-	if err == nil {
-		u.Host = net.JoinHostPort(hostPart, portPart)
-	} else {
-		u.Host = net.JoinHostPort(u.Host, "11434")
-	}
-
-	if u.Path != "" {
-		u.Path = path.Clean(u.Path)
-	}
-
-	if u.Path == "/" {
-		u.Path = ""
-	}
-
-	return u.String(), nil
-}
-
 func convertModelFromFiles(files map[string]string, baseLayers []*layerGGML, isAdapter bool, fn func(resp api.ProgressResponse)) ([]*layerGGML, error) {
 	switch detectModelTypeFromFiles(files) {
 	case "safetensors":
@@ -438,7 +316,15 @@ func kvFromLayers(baseLayers []*layerGGML) (ggml.KV, error) {
 	return ggml.KV{}, fmt.Errorf("no base model was found")
 }
 
-func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML, config *ConfigV2, fn func(resp api.ProgressResponse)) (err error) {
+func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML, fn func(resp api.ProgressResponse)) (err error) {
+	config := ConfigV2{
+		OS:           "linux",
+		Architecture: "amd64",
+		RootFS: RootFS{
+			Type: "layers",
+		},
+	}
+
 	var layers []Layer
 	for _, layer := range baseLayers {
 		if layer.GGML != nil {
@@ -518,7 +404,7 @@ func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML,
 		return err
 	}
 
-	configLayer, err := createConfigLayer(layers, *config)
+	configLayer, err := createConfigLayer(layers, config)
 	if err != nil {
 		return err
 	}

server/create_test.go

@@ -104,154 +104,3 @@ func TestConvertFromSafetensors(t *testing.T) {
 		})
 	}
 }
-
-func TestRemoteURL(t *testing.T) {
-	tests := []struct {
-		name     string
-		input    string
-		expected string
-		hasError bool
-	}{
-		{
-			name:     "absolute path",
-			input:    "/foo/bar",
-			expected: "http://localhost:11434/foo/bar",
-			hasError: false,
-		},
-		{
-			name:     "absolute path with cleanup",
-			input:    "/foo/../bar",
-			expected: "http://localhost:11434/bar",
-			hasError: false,
-		},
-		{
-			name:     "root path",
-			input:    "/",
-			expected: "http://localhost:11434/",
-			hasError: false,
-		},
-		{
-			name:     "host without scheme",
-			input:    "example.com",
-			expected: "http://example.com:11434",
-			hasError: false,
-		},
-		{
-			name:     "host with port",
-			input:    "example.com:8080",
-			expected: "http://example.com:8080",
-			hasError: false,
-		},
-		{
-			name:     "full URL",
-			input:    "https://example.com:8080/path",
-			expected: "https://example.com:8080/path",
-			hasError: false,
-		},
-		{
-			name:     "full URL with path cleanup",
-			input:    "https://example.com:8080/path/../other",
-			expected: "https://example.com:8080/other",
-			hasError: false,
-		},
-		{
-			name:     "ollama.com special case",
-			input:    "ollama.com",
-			expected: "https://ollama.com:443",
-			hasError: false,
-		},
-		{
-			name:     "http ollama.com special case",
-			input:    "http://ollama.com",
-			expected: "https://ollama.com:443",
-			hasError: false,
-		},
-		{
-			name:     "URL with only host",
-			input:    "http://example.com",
-			expected: "http://example.com:11434",
-			hasError: false,
-		},
-		{
-			name:     "URL with root path cleaned",
-			input:    "http://example.com/",
-			expected: "http://example.com:11434",
-			hasError: false,
-		},
-		{
-			name:     "invalid URL",
-			input:    "http://[::1]:namedport", // invalid port
-			expected: "",
-			hasError: true,
-		},
-		{
-			name:     "empty string",
-			input:    "",
-			expected: "http://localhost:11434",
-			hasError: false,
-		},
-		{
-			name:     "host with scheme but no port",
-			input:    "http://localhost",
-			expected: "http://localhost:11434",
-			hasError: false,
-		},
-		{
-			name:     "complex path cleanup",
-			input:    "/a/b/../../c/./d",
-			expected: "http://localhost:11434/c/d",
-			hasError: false,
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			result, err := remoteURL(tt.input)
-
-			if tt.hasError {
-				if err == nil {
-					t.Errorf("expected error but got none")
-				}
-				return
-			}
-
-			if err != nil {
-				t.Errorf("unexpected error: %v", err)
-				return
-			}
-
-			if result != tt.expected {
-				t.Errorf("expected %q, got %q", tt.expected, result)
-			}
-		})
-	}
-}
-
-func TestRemoteURL_Idempotent(t *testing.T) {
-	// Test that applying remoteURL twice gives the same result as applying it once
-	testInputs := []string{
-		"/foo/bar",
-		"example.com",
-		"https://example.com:8080/path",
-		"ollama.com",
-		"http://localhost:11434",
-	}
-
-	for _, input := range testInputs {
-		t.Run(input, func(t *testing.T) {
-			firstResult, err := remoteURL(input)
-			if err != nil {
-				t.Fatalf("first call failed: %v", err)
-			}
-
-			secondResult, err := remoteURL(firstResult)
-			if err != nil {
-				t.Fatalf("second call failed: %v", err)
-			}
-
-			if firstResult != secondResult {
-				t.Errorf("function is not idempotent: first=%q, second=%q", firstResult, secondResult)
-			}
-		})
-	}
-}

server/images.go

@@ -24,7 +24,6 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/gguf"
-	"github.com/ollama/ollama/model/parsers"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/thinking"
@@ -74,38 +73,29 @@ func (m *Model) Capabilities() []model.Capability {
 	capabilities := []model.Capability{}
 
 	// Check for completion capability
-	if m.ModelPath != "" {
-		f, err := gguf.Open(m.ModelPath)
-		if err == nil {
-			defer f.Close()
+	f, err := gguf.Open(m.ModelPath)
+	if err == nil {
+		defer f.Close()
 
-			if f.KeyValue("pooling_type").Valid() {
-				capabilities = append(capabilities, model.CapabilityEmbedding)
-			} else {
-				// If no embedding is specified, we assume the model supports completion
-				capabilities = append(capabilities, model.CapabilityCompletion)
-			}
-			if f.KeyValue("vision.block_count").Valid() {
-				capabilities = append(capabilities, model.CapabilityVision)
-			}
+		if f.KeyValue("pooling_type").Valid() {
+			capabilities = append(capabilities, model.CapabilityEmbedding)
 		} else {
-			slog.Error("couldn't open model file", "error", err)
+			// If no embedding is specified, we assume the model supports completion
+			capabilities = append(capabilities, model.CapabilityCompletion)
 		}
-	} else if len(m.Config.Capabilities) > 0 {
-		for _, c := range m.Config.Capabilities {
-			capabilities = append(capabilities, model.Capability(c))
+		if f.KeyValue("vision.block_count").Valid() {
+			capabilities = append(capabilities, model.CapabilityVision)
 		}
 	} else {
-		slog.Warn("unknown capabilities for model", "model", m.Name)
+		slog.Error("couldn't open model file", "error", err)
 	}
 
 	if m.Template == nil {
 		return capabilities
 	}
 
-	builtinParser := parsers.ParserForName(m.Config.Parser)
 	// Check for tools capability
-	if slices.Contains(m.Template.Vars(), "tools") || (builtinParser != nil && builtinParser.HasToolSupport()) {
+	if slices.Contains(m.Template.Vars(), "tools") {
 		capabilities = append(capabilities, model.CapabilityTools)
 	}
 
@@ -119,16 +109,10 @@ func (m *Model) Capabilities() []model.Capability {
 		capabilities = append(capabilities, model.CapabilityVision)
 	}
 
-	// Skip the thinking check if it's already set
-	if slices.Contains(capabilities, "thinking") {
-		return capabilities
-	}
-
 	// Check for thinking capability
 	openingTag, closingTag := thinking.InferTags(m.Template.Template)
 	hasTags := openingTag != "" && closingTag != ""
-	isGptoss := slices.Contains([]string{"gptoss", "gpt-oss"}, m.Config.ModelFamily)
-	if hasTags || isGptoss || (builtinParser != nil && builtinParser.HasThinkingSupport()) {
+	if hasTags || slices.Contains([]string{"gptoss", "gpt-oss"}, m.Config.ModelFamily) {
 		capabilities = append(capabilities, model.CapabilityThinking)
 	}
 
@@ -214,20 +198,6 @@ func (m *Model) String() string {
 		})
 	}
 
-	if m.Config.Renderer != "" {
-		modelfile.Commands = append(modelfile.Commands, parser.Command{
-			Name: "renderer",
-			Args: m.Config.Renderer,
-		})
-	}
-
-	if m.Config.Parser != "" {
-		modelfile.Commands = append(modelfile.Commands, parser.Command{
-			Name: "parser",
-			Args: m.Config.Parser,
-		})
-	}
-
 	for k, v := range m.Options {
 		switch v := v.(type) {
 		case []any:
@@ -266,19 +236,8 @@ type ConfigV2 struct {
 	ModelFormat   string   `json:"model_format"`
 	ModelFamily   string   `json:"model_family"`
 	ModelFamilies []string `json:"model_families"`
-	ModelType     string   `json:"model_type"` // shown as Parameter Size
-	FileType      string   `json:"file_type"`  // shown as Quantization Level
-	Renderer      string   `json:"renderer,omitempty"`
-	Parser        string   `json:"parser,omitempty"`
-
-	RemoteHost  string `json:"remote_host,omitempty"`
-	RemoteModel string `json:"remote_model,omitempty"`
-
-	// used for remotes
-	Capabilities []string `json:"capabilities,omitempty"`
-	ContextLen   int      `json:"context_length,omitempty"`
-	EmbedLen     int      `json:"embedding_length,omitempty"`
-	BaseName     string   `json:"base_name,omitempty"`
+	ModelType     string   `json:"model_type"`
+	FileType      string   `json:"file_type"`
 
 	// required by spec
 	Architecture string `json:"architecture"`

server/internal/internal/backoff/backoff.go

@@ -25,7 +25,10 @@ func Loop(ctx context.Context, maxBackoff time.Duration) iter.Seq2[int, error] {
 
 			// n^2 backoff timer is a little smoother than the
 			// common choice of 2^n.
-			d := min(time.Duration(n*n)*10*time.Millisecond, maxBackoff)
+			d := time.Duration(n*n) * 10 * time.Millisecond
+			if d > maxBackoff {
+				d = maxBackoff
+			}
 			// Randomize the delay between 0.5-1.5 x msec, in order
 			// to prevent accidental "thundering herd" problems.
 			d = time.Duration(float64(d) * (rand.Float64() + 0.5))

server/prompt.go

@@ -11,7 +11,6 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/llm"
-	"github.com/ollama/ollama/model/renderers"
 	"github.com/ollama/ollama/template"
 )
 
@@ -42,12 +41,18 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 			}
 		}
 
-		p, err := renderPrompt(m, append(system, msgs[i:]...), tools, think)
-		if err != nil {
+		thinkVal := false
+		thinkLevel := ""
+		if think != nil {
+			thinkVal = think.Bool()
+			thinkLevel = think.String()
+		}
+		var b bytes.Buffer
+		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools, Think: thinkVal, ThinkLevel: thinkLevel, IsThinkSet: think != nil}); err != nil {
 			return "", nil, err
 		}
 
-		s, err := tokenize(ctx, p)
+		s, err := tokenize(ctx, b.String())
 		if err != nil {
 			return "", nil, err
 		}
@@ -96,23 +101,6 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 	}
 
 	// truncate any messages that do not fit into the context window
-	p, err := renderPrompt(m, append(system, msgs[currMsgIdx:]...), tools, think)
-	if err != nil {
-		return "", nil, err
-	}
-
-	return p, images, nil
-}
-
-func renderPrompt(m *Model, msgs []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error) {
-	if m.Config.Renderer != "" {
-		rendered, err := renderers.RenderWithRenderer(m.Config.Renderer, msgs, tools, think)
-		if err != nil {
-			return "", err
-		}
-		return rendered, nil
-	}
-
 	var b bytes.Buffer
 	thinkVal := false
 	thinkLevel := ""
@@ -120,8 +108,9 @@ func renderPrompt(m *Model, msgs []api.Message, tools []api.Tool, think *api.Thi
 		thinkVal = think.Bool()
 		thinkLevel = think.String()
 	}
-	if err := m.Template.Execute(&b, template.Values{Messages: msgs, Tools: tools, Think: thinkVal, ThinkLevel: thinkLevel, IsThinkSet: think != nil}); err != nil {
-		return "", err
+	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[currMsgIdx:]...), Tools: tools, Think: thinkVal, ThinkLevel: thinkLevel, IsThinkSet: think != nil}); err != nil {
+		return "", nil, err
 	}
-	return b.String(), nil
+
+	return b.String(), images, nil
 }

server/routes.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"cmp"
 	"context"
-	"encoding/base64"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -16,7 +15,6 @@ import (
 	"net"
 	"net/http"
 	"net/netip"
-	"net/url"
 	"os"
 	"os/signal"
 	"slices"
@@ -30,14 +28,13 @@ import (
 	"golang.org/x/sync/errgroup"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/discover"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/harmony"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
-	"github.com/ollama/ollama/model/parsers"
 	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/server/internal/client/ollama"
 	"github.com/ollama/ollama/server/internal/registry"
@@ -49,8 +46,6 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
-const signinURLStr = "https://ollama.com/connect?name=%s&key=%s"
-
 func shouldUseHarmony(model *Model) bool {
 	if slices.Contains([]string{"gptoss", "gpt-oss"}, model.Config.ModelFamily) {
 		// heuristic to check whether the template expects to be parsed via harmony:
@@ -76,20 +71,9 @@ var lowVRAMThreshold uint64 = 20 * format.GibiByte
 var mode string = gin.DebugMode
 
 type Server struct {
-	addr         net.Addr
-	sched        *Scheduler
-	lowVRAM      bool
-	cloudBaseURL *url.URL
-}
-
-func (s *Server) webServiceBase() *url.URL {
-	defaultWebServiceURL := url.URL{Scheme: "https", Host: "ollama.com"}
-	if s != nil && s.cloudBaseURL != nil {
-		u := *s.cloudBaseURL
-		return &u
-	}
-	u := defaultWebServiceURL
-	return &u
+	addr    net.Addr
+	sched   *Scheduler
+	lowVRAM bool
 }
 
 func init() {
@@ -164,17 +148,6 @@ func (s *Server) scheduleRunner(ctx context.Context, name string, caps []model.C
 	return runner.llama, model, &opts, nil
 }
 
-func signinURL() (string, error) {
-	pubKey, err := auth.GetPublicKey()
-	if err != nil {
-		return "", err
-	}
-
-	encKey := base64.RawURLEncoding.EncodeToString([]byte(pubKey))
-	h, _ := os.Hostname()
-	return fmt.Sprintf(signinURLStr, url.PathEscape(h), encKey), nil
-}
-
 func (s *Server) GenerateHandler(c *gin.Context) {
 	checkpointStart := time.Now()
 	var req api.GenerateRequest
@@ -215,90 +188,6 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	if m.Config.RemoteHost != "" && m.Config.RemoteModel != "" {
-		origModel := req.Model
-
-		remoteURL, err := url.Parse(m.Config.RemoteHost)
-		if err != nil {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-
-		if !slices.Contains(envconfig.Remotes(), remoteURL.Hostname()) {
-			slog.Info("remote model", "remotes", envconfig.Remotes(), "remoteURL", m.Config.RemoteHost, "hostname", remoteURL.Hostname())
-			c.JSON(http.StatusBadRequest, gin.H{"error": "this server cannot run this remote model"})
-			return
-		}
-
-		req.Model = m.Config.RemoteModel
-
-		if req.Template == "" && m.Template.String() != "" {
-			req.Template = m.Template.String()
-		}
-
-		if req.Options == nil {
-			req.Options = map[string]any{}
-		}
-
-		for k, v := range m.Options {
-			if _, ok := req.Options[k]; !ok {
-				req.Options[k] = v
-			}
-		}
-
-		// update the system prompt from the model if one isn't already specified
-		if req.System == "" && m.System != "" {
-			req.System = m.System
-		}
-
-		if len(m.Messages) > 0 {
-			slog.Warn("embedded messages in the model not supported with '/api/generate'; try '/api/chat' instead")
-		}
-
-		fn := func(resp api.GenerateResponse) error {
-			resp.Model = origModel
-			resp.RemoteModel = m.Config.RemoteModel
-			resp.RemoteHost = m.Config.RemoteHost
-
-			data, err := json.Marshal(resp)
-			if err != nil {
-				return err
-			}
-
-			if _, err = c.Writer.Write(append(data, '\n')); err != nil {
-				return err
-			}
-			c.Writer.Flush()
-			return nil
-		}
-
-		client := api.NewClient(remoteURL, http.DefaultClient)
-		err = client.Generate(c, &req, fn)
-		if err != nil {
-			var authError api.AuthorizationError
-			if errors.As(err, &authError) {
-				sURL, sErr := signinURL()
-				if sErr != nil {
-					slog.Error(sErr.Error())
-					c.JSON(http.StatusInternalServerError, gin.H{"error": "error getting authorization details"})
-					return
-				}
-
-				c.JSON(authError.StatusCode, gin.H{"error": "unauthorized", "signin_url": sURL})
-				return
-			}
-			var apiError api.StatusError
-			if errors.As(err, &apiError) {
-				c.JSON(apiError.StatusCode, apiError)
-				return
-			}
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-
-		return
-	}
-
 	// expire the runner
 	if req.Prompt == "" && req.KeepAlive != nil && req.KeepAlive.Duration == 0 {
 		s.sched.expireRunner(m)
@@ -318,21 +207,17 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	var builtinParser parsers.Parser
-	if shouldUseHarmony(m) && m.Config.Parser == "" {
-		m.Config.Parser = "harmony"
+	useHarmony := shouldUseHarmony(m) && !req.Raw
+	var harmonyMessageHandler *harmony.HarmonyMessageHandler
+	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
+	if useHarmony {
+		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
+		harmonyMessageHandler.HarmonyParser.AddImplicitStart()
+		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
 	}
 
-	if !req.Raw && m.Config.Parser != "" {
-		builtinParser = parsers.ParserForName(m.Config.Parser)
-		if builtinParser != nil {
-			// no tools or last message for generate endpoint
-			builtinParser.Init(nil, nil)
-		}
-	}
-
-	// Validate Think value: string values currently only allowed for harmony/gptoss models
-	if req.Think != nil && req.Think.IsString() && m.Config.Parser != "harmony" {
+	// Validate Think value: string values currently only allowed for gptoss models
+	if req.Think != nil && req.Think.IsString() && !useHarmony {
 		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("think value %q is not supported for this model", req.Think.String())})
 		return
 	}
@@ -444,10 +329,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	// If debug mode is enabled, return the rendered template instead of calling the model
 	if req.DebugRenderOnly {
-		c.JSON(http.StatusOK, api.GenerateResponse{
+		c.JSON(http.StatusOK, api.DebugTemplateResponse{
 			Model:     req.Model,
 			CreatedAt: time.Now().UTC(),
-			DebugInfo: &api.DebugInfo{
+			DebugInfo: api.DebugInfo{
 				RenderedTemplate: prompt,
 				ImageCount:       len(images),
 			},
@@ -456,16 +341,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 	}
 
 	var thinkingState *thinking.Parser
-	if builtinParser == nil {
+	if !useHarmony {
 		openingTag, closingTag := thinking.InferTags(m.Template.Template)
 		if req.Think != nil && req.Think.Bool() && openingTag != "" && closingTag != "" {
 			thinkingState = &thinking.Parser{
 				OpeningTag: openingTag,
 				ClosingTag: closingTag,
 			}
-			if strings.HasSuffix(strings.TrimSpace(prompt), openingTag) {
-				thinkingState.AddContent(openingTag)
-			}
 		}
 	}
 
@@ -493,17 +375,11 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 				},
 			}
 
-			if builtinParser != nil {
-				content, thinking, toolCalls, err := builtinParser.Add(cr.Content, cr.Done)
-				if err != nil {
-					ch <- gin.H{"error": err.Error()}
-					return
-				}
+			if useHarmony {
+				content, thinking, toolContent := harmonyMessageHandler.AddContent(cr.Content, harmonyToolParser)
 				res.Response = content
 				res.Thinking = thinking
-				if cr.Done && len(toolCalls) > 0 {
-					res.ToolCalls = toolCalls
-				}
+				harmonyToolParser.Add(toolContent)
 			} else if thinkingState != nil {
 				thinking, content := thinkingState.AddContent(cr.Content)
 				res.Thinking = thinking
@@ -515,6 +391,26 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			}
 
 			if cr.Done {
+				if useHarmony {
+					toolName, toolContent := harmonyToolParser.Drain()
+					if toolName != nil {
+						*toolName = strings.TrimPrefix(*toolName, "functions.")
+						var args api.ToolCallFunctionArguments
+						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
+							errStr := fmt.Sprintf("error parsing tool call: raw='%s', err=%s", toolContent, err.Error())
+							ch <- gin.H{"error": errStr}
+							return
+						}
+
+						res.ToolCalls = append(res.ToolCalls, api.ToolCall{
+							Function: api.ToolCallFunction{
+								Name:      *toolName,
+								Arguments: args,
+							},
+						})
+					}
+				}
+
 				res.DoneReason = cr.DoneReason.String()
 				res.TotalDuration = time.Since(checkpointStart)
 				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
@@ -529,7 +425,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 				}
 			}
 
-			if builtinParser != nil {
+			if useHarmony {
 				// only send messages with meaningful content (empty messages confuse clients)
 				if res.Response != "" || res.Thinking != "" || res.Done || len(res.ToolCalls) > 0 {
 					ch <- res
@@ -654,7 +550,7 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 		ctxLen := min(opts.NumCtx, int(kvData.ContextLength()))
 		if len(tokens) > ctxLen {
 			if !truncate {
-				c.JSON(http.StatusBadRequest, gin.H{"error": "input exceeds maximum context length"})
+				c.JSON(http.StatusBadRequest, gin.H{"error": "input length exceeds maximum context length"})
 				return
 			}
 
@@ -666,13 +562,6 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 				ctxLen--
 			}
 
-			slog.Info("", "ctxLen", ctxLen, "tokenCount", len(tokens))
-			if ctxLen <= 0 {
-				// return error if the truncated input would be empty or just special tokens
-				c.JSON(http.StatusBadRequest, gin.H{"error": "input after truncation exceeds maximum context length"})
-				return
-			}
-
 			tokens = tokens[:ctxLen]
 
 			s, err = r.Detokenize(c.Request.Context(), tokens)
@@ -778,105 +667,6 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 	c.JSON(http.StatusOK, resp)
 }
 
-func (s *Server) WebSearchHandler(c *gin.Context) {
-	var req api.WebSearchRequest
-	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
-		return
-	} else if err != nil {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	req.Query = strings.TrimSpace(req.Query)
-	if req.Query == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "query is required"})
-		return
-	}
-
-	if req.MaxResults != 0 && (req.MaxResults < 1 || req.MaxResults > 10) {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "max_results must be between 1 and 10"})
-		return
-	}
-
-	webServiceClient := api.NewClient(s.webServiceBase(), http.DefaultClient)
-	resp, err := webServiceClient.WebSearch(c.Request.Context(), &req)
-	if err != nil {
-		var authError api.AuthorizationError
-		if errors.As(err, &authError) {
-			sURL, sErr := signinURL()
-			if sErr != nil {
-				slog.Error(sErr.Error())
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "error getting authorization details"})
-				return
-			}
-
-			c.JSON(authError.StatusCode, gin.H{"error": "unauthorized", "signin_url": sURL})
-			return
-		}
-		var apiError api.StatusError
-		if errors.As(err, &apiError) {
-			c.JSON(apiError.StatusCode, apiError)
-			return
-		}
-		c.JSON(http.StatusBadGateway, gin.H{"error": err.Error()})
-		return
-	}
-
-	if resp == nil {
-		resp = &api.WebSearchResponse{}
-	}
-
-	c.JSON(http.StatusOK, resp)
-}
-
-func (s *Server) WebFetchHandler(c *gin.Context) {
-	var req api.WebFetchRequest
-	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
-		return
-	} else if err != nil {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	req.URL = strings.TrimSpace(req.URL)
-	if req.URL == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "url is required"})
-		return
-	}
-
-	webServiceClient := api.NewClient(s.webServiceBase(), http.DefaultClient)
-	resp, err := webServiceClient.WebFetch(c.Request.Context(), &req)
-	if err != nil {
-		var authError api.AuthorizationError
-		if errors.As(err, &authError) {
-			sURL, sErr := signinURL()
-			if sErr != nil {
-				slog.Error(sErr.Error())
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "error getting authorization details"})
-				return
-			}
-
-			c.JSON(authError.StatusCode, gin.H{"error": "unauthorized", "signin_url": sURL})
-			return
-		}
-		var apiError api.StatusError
-		if errors.As(err, &apiError) {
-			c.JSON(apiError.StatusCode, apiError)
-			return
-		}
-		c.JSON(http.StatusBadGateway, gin.H{"error": err.Error()})
-		return
-	}
-
-	if resp == nil {
-		resp = &api.WebFetchResponse{}
-	}
-
-	c.JSON(http.StatusOK, resp)
-}
-
 func (s *Server) PullHandler(c *gin.Context) {
 	var req api.PullRequest
 	err := c.ShouldBindJSON(&req)
@@ -1140,28 +930,6 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		ModifiedAt:   manifest.fi.ModTime(),
 	}
 
-	if m.Config.RemoteHost != "" {
-		resp.RemoteHost = m.Config.RemoteHost
-		resp.RemoteModel = m.Config.RemoteModel
-
-		if m.Config.ModelFamily != "" {
-			resp.ModelInfo = make(map[string]any)
-			resp.ModelInfo["general.architecture"] = m.Config.ModelFamily
-
-			if m.Config.BaseName != "" {
-				resp.ModelInfo["general.basename"] = m.Config.BaseName
-			}
-
-			if m.Config.ContextLen > 0 {
-				resp.ModelInfo[fmt.Sprintf("%s.context_length", m.Config.ModelFamily)] = m.Config.ContextLen
-			}
-
-			if m.Config.EmbedLen > 0 {
-				resp.ModelInfo[fmt.Sprintf("%s.embedding_length", m.Config.ModelFamily)] = m.Config.EmbedLen
-			}
-		}
-	}
-
 	var params []string
 	cs := 30
 	for k, v := range m.Options {
@@ -1192,11 +960,6 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	fmt.Fprint(&sb, m.String())
 	resp.Modelfile = sb.String()
 
-	// skip loading tensor information if this is a remote model
-	if m.Config.RemoteHost != "" && m.Config.RemoteModel != "" {
-		return resp, nil
-	}
-
 	kvData, tensors, err := getModelData(m.ModelPath, req.Verbose)
 	if err != nil {
 		return nil, err
@@ -1273,13 +1036,11 @@ func (s *Server) ListHandler(c *gin.Context) {
 
 		// tag should never be masked
 		models = append(models, api.ListModelResponse{
-			Model:       n.DisplayShortest(),
-			Name:        n.DisplayShortest(),
-			RemoteModel: cf.RemoteModel,
-			RemoteHost:  cf.RemoteHost,
-			Size:        m.Size(),
-			Digest:      m.digest,
-			ModifiedAt:  m.fi.ModTime(),
+			Model:      n.DisplayShortest(),
+			Name:       n.DisplayShortest(),
+			Size:       m.Size(),
+			Digest:     m.digest,
+			ModifiedAt: m.fi.ModTime(),
 			Details: api.ModelDetails{
 				Format:            cf.ModelFormat,
 				Family:            cf.ModelFamily,
@@ -1539,12 +1300,6 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.POST("/api/show", s.ShowHandler)
 	r.DELETE("/api/delete", s.DeleteHandler)
 
-	r.POST("/api/me", s.WhoamiHandler)
-
-	r.POST("/api/signout", s.SignoutHandler)
-	// deprecated
-	r.DELETE("/api/user/keys/:encodedKey", s.SignoutHandler)
-
 	// Create
 	r.POST("/api/create", s.CreateHandler)
 	r.POST("/api/blobs/:digest", s.CreateBlobHandler)
@@ -1557,8 +1312,6 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.POST("/api/chat", s.ChatHandler)
 	r.POST("/api/embed", s.EmbedHandler)
 	r.POST("/api/embeddings", s.EmbeddingsHandler)
-	r.POST("/api/web_search", s.WebSearchHandler)
-	r.POST("/api/web_fetch", s.WebFetchHandler)
 
 	// Inference (OpenAI compatibility)
 	r.POST("/v1/chat/completions", openai.ChatMiddleware(), s.ChatHandler)
@@ -1743,70 +1496,6 @@ func streamResponse(c *gin.Context, ch chan any) {
 	})
 }
 
-func (s *Server) WhoamiHandler(c *gin.Context) {
-	// todo allow other hosts
-	u, err := url.Parse("https://ollama.com")
-	if err != nil {
-		slog.Error(err.Error())
-		c.JSON(http.StatusInternalServerError, gin.H{"error": "URL parse error"})
-		return
-	}
-
-	client := api.NewClient(u, http.DefaultClient)
-	user, err := client.Whoami(c)
-	if err != nil {
-		slog.Error(err.Error())
-	}
-
-	// user isn't signed in
-	if user != nil && user.Name == "" {
-		sURL, sErr := signinURL()
-		if sErr != nil {
-			slog.Error(sErr.Error())
-			c.JSON(http.StatusInternalServerError, gin.H{"error": "error getting authorization details"})
-			return
-		}
-
-		c.JSON(http.StatusUnauthorized, gin.H{"error": "unauthorized", "signin_url": sURL})
-		return
-	}
-
-	c.JSON(http.StatusOK, user)
-}
-
-func (s *Server) SignoutHandler(c *gin.Context) {
-	pubKey, err := auth.GetPublicKey()
-	if err != nil {
-		slog.Error("couldn't get public key", "error", err)
-		c.JSON(http.StatusInternalServerError, gin.H{"error": "there was an error signing out"})
-		return
-	}
-
-	encKey := base64.RawURLEncoding.EncodeToString([]byte(pubKey))
-
-	// todo allow other hosts
-	u, err := url.Parse("https://ollama.com")
-	if err != nil {
-		slog.Error(err.Error())
-		c.JSON(http.StatusInternalServerError, gin.H{"error": "URL parse error"})
-		return
-	}
-
-	client := api.NewClient(u, http.DefaultClient)
-	err = client.Disconnect(c, encKey)
-	if err != nil {
-		var authError api.AuthorizationError
-		if errors.As(err, &authError) {
-			c.JSON(http.StatusUnauthorized, gin.H{"error": "you are not currently signed in"})
-			return
-		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": "there was an error signing out"})
-		return
-	}
-
-	c.JSON(http.StatusOK, nil)
-}
-
 func (s *Server) PsHandler(c *gin.Context) {
 	models := []api.ProcessModelResponse{}
 
@@ -1863,34 +1552,21 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	name := model.ParseName(req.Model)
-	if !name.IsValid() {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "model is required"})
-		return
-	}
-
-	name, err := getExistingName(name)
-	if err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "model is required"})
-		return
-	}
-
-	m, err := GetModel(req.Model)
-	if err != nil {
-		switch {
-		case os.IsNotExist(err):
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
-		case err.Error() == errtypes.InvalidModelNameErrMsg:
-			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		default:
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		}
-		return
-	}
-
 	// expire the runner
 	if len(req.Messages) == 0 && req.KeepAlive != nil && req.KeepAlive.Duration == 0 {
-		s.sched.expireRunner(m)
+		model, err := GetModel(req.Model)
+		if err != nil {
+			switch {
+			case os.IsNotExist(err):
+				c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
+			case err.Error() == errtypes.InvalidModelNameErrMsg:
+				c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+			default:
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			}
+			return
+		}
+		s.sched.expireRunner(model)
 
 		c.JSON(http.StatusOK, api.ChatResponse{
 			Model:      req.Model,
@@ -1902,83 +1578,6 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	if m.Config.RemoteHost != "" && m.Config.RemoteModel != "" {
-		origModel := req.Model
-
-		remoteURL, err := url.Parse(m.Config.RemoteHost)
-		if err != nil {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-
-		if !slices.Contains(envconfig.Remotes(), remoteURL.Hostname()) {
-			slog.Info("remote model", "remotes", envconfig.Remotes(), "remoteURL", m.Config.RemoteHost, "hostname", remoteURL.Hostname())
-			c.JSON(http.StatusBadRequest, gin.H{"error": "this server cannot run this remote model"})
-			return
-		}
-
-		req.Model = m.Config.RemoteModel
-		if req.Options == nil {
-			req.Options = map[string]any{}
-		}
-
-		msgs := append(m.Messages, req.Messages...)
-		if req.Messages[0].Role != "system" && m.System != "" {
-			msgs = append([]api.Message{{Role: "system", Content: m.System}}, msgs...)
-		}
-		msgs = filterThinkTags(msgs, m)
-		req.Messages = msgs
-
-		for k, v := range m.Options {
-			if _, ok := req.Options[k]; !ok {
-				req.Options[k] = v
-			}
-		}
-
-		fn := func(resp api.ChatResponse) error {
-			resp.Model = origModel
-			resp.RemoteModel = m.Config.RemoteModel
-			resp.RemoteHost = m.Config.RemoteHost
-
-			data, err := json.Marshal(resp)
-			if err != nil {
-				return err
-			}
-
-			if _, err = c.Writer.Write(append(data, '\n')); err != nil {
-				return err
-			}
-			c.Writer.Flush()
-			return nil
-		}
-
-		client := api.NewClient(remoteURL, http.DefaultClient)
-		err = client.Chat(c, &req, fn)
-		if err != nil {
-			var authError api.AuthorizationError
-			if errors.As(err, &authError) {
-				sURL, sErr := signinURL()
-				if sErr != nil {
-					slog.Error(sErr.Error())
-					c.JSON(http.StatusInternalServerError, gin.H{"error": "error getting authorization details"})
-					return
-				}
-
-				c.JSON(authError.StatusCode, gin.H{"error": "unauthorized", "signin_url": sURL})
-				return
-			}
-			var apiError api.StatusError
-			if errors.As(err, &apiError) {
-				c.JSON(apiError.StatusCode, apiError)
-				return
-			}
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-
-		return
-	}
-
 	caps := []model.Capability{model.CapabilityCompletion}
 	if len(req.Tools) > 0 {
 		caps = append(caps, model.CapabilityTools)
@@ -1987,6 +1586,17 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		caps = append(caps, model.CapabilityThinking)
 	}
 
+	name := model.ParseName(req.Model)
+	if !name.IsValid() {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "model is required"})
+		return
+	}
+	name, err := getExistingName(name)
+	if err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "model is required"})
+		return
+	}
+
 	r, m, opts, err := s.scheduleRunner(c.Request.Context(), name.String(), caps, req.Options, req.KeepAlive)
 	if errors.Is(err, errCapabilityCompletion) {
 		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%q does not support chat", req.Model)})
@@ -2015,23 +1625,27 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 	msgs = filterThinkTags(msgs, m)
 
-	if shouldUseHarmony(m) && m.Config.Parser == "" {
-		m.Config.Parser = "harmony"
-	}
+	var harmonyMessageHandler *harmony.HarmonyMessageHandler
+	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
+
+	useHarmony := shouldUseHarmony(m)
 
-	var builtinParser parsers.Parser
 	processedTools := req.Tools
+	if useHarmony {
+		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
+		var lastMessage *api.Message
+		if len(msgs) > 0 {
+			lastMessage = &msgs[len(msgs)-1]
+		}
+		harmonyMessageHandler.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
+		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
 
-	if m.Config.Parser != "" {
-		builtinParser = parsers.ParserForName(m.Config.Parser)
-		if builtinParser != nil {
-			// Determine last message for chat prefill
-			var lastMessage *api.Message
-			if len(msgs) > 0 {
-				lastMessage = &msgs[len(msgs)-1]
-			}
-			// Initialize parser and get processed tools
-			processedTools = builtinParser.Init(req.Tools, lastMessage)
+		// make a copy of tools to pass to the chat prompt. Function names may be
+		// renamed to be valid Harmony function names.
+		processedTools = make([]api.Tool, len(req.Tools))
+		copy(processedTools, req.Tools)
+		for i, tool := range processedTools {
+			processedTools[i].Function.Name = harmonyMessageHandler.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
 		}
 	}
 
@@ -2044,10 +1658,10 @@ func (s *Server) ChatHandler(c *gin.Context) {
 
 	// If debug mode is enabled, return the rendered template instead of calling the model
 	if req.DebugRenderOnly {
-		c.JSON(http.StatusOK, api.ChatResponse{
+		c.JSON(http.StatusOK, api.DebugTemplateResponse{
 			Model:     req.Model,
 			CreatedAt: time.Now().UTC(),
-			DebugInfo: &api.DebugInfo{
+			DebugInfo: api.DebugInfo{
 				RenderedTemplate: prompt,
 				ImageCount:       len(images),
 			},
@@ -2055,8 +1669,8 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	// Validate Think value: string values currently only allowed for harmony/gptoss models
-	if req.Think != nil && req.Think.IsString() && m.Config.Parser != "harmony" {
+	// Validate Think value: string values currently only allowed for gptoss models
+	if req.Think != nil && req.Think.IsString() && !useHarmony {
 		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("think value %q is not supported for this model", req.Think.String())})
 		return
 	}
@@ -2075,7 +1689,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 
 	var toolParser *tools.Parser
-	if len(req.Tools) > 0 && (builtinParser == nil || !builtinParser.HasToolSupport()) {
+	if len(req.Tools) > 0 && !useHarmony {
 		toolParser = tools.NewParser(m.Template.Template, req.Tools)
 	}
 
@@ -2107,24 +1721,30 @@ func (s *Server) ChatHandler(c *gin.Context) {
 				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
 			}
 
-			if builtinParser != nil {
-				slog.Log(context.TODO(), logutil.LevelTrace, "builtin parser input", "parser", m.Config.Parser, "content", r.Content)
-
-				content, thinking, toolCalls, err := builtinParser.Add(r.Content, r.Done)
-				if err != nil {
-					ch <- gin.H{"error": err.Error()}
-					return
-				}
-
+			if useHarmony {
+				content, thinking, toolContent := harmonyMessageHandler.AddContent(r.Content, harmonyToolParser)
 				res.Message.Content = content
 				res.Message.Thinking = thinking
-				res.Message.ToolCalls = toolCalls
+				harmonyToolParser.Add(toolContent)
 
-				if res.Message.Content != "" || res.Message.Thinking != "" || len(res.Message.ToolCalls) > 0 || r.Done {
-					slog.Log(context.TODO(), logutil.LevelTrace, "builtin parser output", "parser", m.Config.Parser, "content", content, "thinking", thinking, "toolCalls", toolCalls, "done", r.Done)
+				if r.Done {
+					toolName, toolContent := harmonyToolParser.Drain()
+					if toolName != nil {
+						*toolName = strings.TrimPrefix(*toolName, "functions.")
+						*toolName = harmonyMessageHandler.FunctionNameMap.OriginalFromConverted(*toolName)
+						var args api.ToolCallFunctionArguments
+						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
+							errStr := fmt.Sprintf("error parsing tool call: raw='%s', err=%s", toolContent, err.Error())
+							ch <- gin.H{"error": errStr}
+							return
+						}
+						res.Message.ToolCalls = []api.ToolCall{{Function: api.ToolCallFunction{Name: *toolName, Arguments: args}}}
+					}
+				}
+
+				// only send messages with meaningful content (empty messages confuse clients)
+				if res.Message.Content != "" || res.Message.Thinking != "" || len(res.Message.ToolCalls) > 0 || res.Done {
 					ch <- res
-				} else {
-					slog.Log(context.TODO(), logutil.LevelTrace, "builtin parser empty output", "parser", m.Config.Parser)
 				}
 
 				return

server/routes_create_test.go

@@ -11,7 +11,6 @@ import (
 	"net/http/httptest"
 	"os"
 	"path/filepath"
-	"reflect"
 	"slices"
 	"strings"
 	"testing"
@@ -21,7 +20,6 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/fs/ggml"
-	"github.com/ollama/ollama/types/model"
 )
 
 var stream bool = false
@@ -617,78 +615,6 @@ func TestCreateTemplateSystem(t *testing.T) {
 	})
 }
 
-func TestCreateAndShowRemoteModel(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-
-	var s Server
-
-	w := createRequest(t, s.CreateHandler, api.CreateRequest{
-		Model:      "test",
-		From:       "bob",
-		RemoteHost: "https://ollama.com",
-		Info: map[string]any{
-			"capabilities":       []string{"completion", "tools", "thinking"},
-			"model_family":       "gptoss",
-			"context_length":     131072,
-			"embedding_length":   2880,
-			"quantization_level": "MXFP4",
-			"parameter_size":     "20.9B",
-		},
-		Stream: &stream,
-	})
-
-	if w.Code != http.StatusOK {
-		t.Fatalf("exected status code 200, actual %d", w.Code)
-	}
-
-	w = createRequest(t, s.ShowHandler, api.ShowRequest{Model: "test"})
-	if w.Code != http.StatusOK {
-		t.Fatalf("exected status code 200, actual %d", w.Code)
-	}
-
-	var resp api.ShowResponse
-	if err := json.NewDecoder(w.Body).Decode(&resp); err != nil {
-		t.Fatal(err)
-	}
-
-	expectedDetails := api.ModelDetails{
-		ParentModel:       "",
-		Format:            "",
-		Family:            "gptoss",
-		Families:          []string{"gptoss"},
-		ParameterSize:     "20.9B",
-		QuantizationLevel: "MXFP4",
-	}
-
-	if !reflect.DeepEqual(resp.Details, expectedDetails) {
-		t.Errorf("model details: expected %#v, actual %#v", expectedDetails, resp.Details)
-	}
-
-	expectedCaps := []model.Capability{
-		model.Capability("completion"),
-		model.Capability("tools"),
-		model.Capability("thinking"),
-	}
-
-	if !slices.Equal(resp.Capabilities, expectedCaps) {
-		t.Errorf("capabilities: expected %#v, actual %#v", expectedCaps, resp.Capabilities)
-	}
-
-	v, ok := resp.ModelInfo["gptoss.context_length"]
-	ctxlen := v.(float64)
-	if !ok || int(ctxlen) != 131072 {
-		t.Errorf("context len: expected %d, actual %d", 131072, int(ctxlen))
-	}
-
-	v, ok = resp.ModelInfo["gptoss.embedding_length"]
-	embedlen := v.(float64)
-	if !ok || int(embedlen) != 2880 {
-		t.Errorf("embed len: expected %d, actual %d", 2880, int(embedlen))
-	}
-
-	fmt.Printf("resp = %#v\n", resp)
-}
-
 func TestCreateLicenses(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 

server/routes_debug_test.go

@@ -180,7 +180,7 @@ func TestGenerateDebugRenderOnly(t *testing.T) {
 						t.Errorf("expected status %d, got %d, body: %s", http.StatusOK, w.Code, w.Body.String())
 					}
 
-					var response api.GenerateResponse
+					var response api.DebugTemplateResponse
 					if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
 						t.Fatalf("failed to unmarshal response: %v", err)
 					}
@@ -385,7 +385,7 @@ func TestChatDebugRenderOnly(t *testing.T) {
 						t.Errorf("expected status %d, got %d, body: %s", http.StatusOK, w.Code, w.Body.String())
 					}
 
-					var response api.ChatResponse
+					var response api.DebugTemplateResponse
 					if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
 						t.Fatalf("failed to unmarshal response: %v", err)
 					}

server/routes_test.go

@@ -13,7 +13,6 @@ import (
 	"net"
 	"net/http"
 	"net/http/httptest"
-	"net/url"
 	"os"
 	"path/filepath"
 	"reflect"
@@ -93,9 +92,6 @@ func (t *panicTransport) RoundTrip(r *http.Request) (*http.Response, error) {
 var panicOnRoundTrip = &http.Client{Transport: &panicTransport{}}
 
 func TestRoutes(t *testing.T) {
-	// Disable authentication for tests to avoid issues with missing private keys
-	t.Setenv("OLLAMA_AUTH", "false")
-
 	type testCase struct {
 		Name     string
 		Method   string
@@ -130,24 +126,11 @@ func TestRoutes(t *testing.T) {
 			t.Fatalf("failed to create model: %v", err)
 		}
 
-		config := &ConfigV2{
-			OS:           "linux",
-			Architecture: "amd64",
-			RootFS: RootFS{
-				Type: "layers",
-			},
-		}
-
-		if err := createModel(r, modelName, baseLayers, config, fn); err != nil {
+		if err := createModel(r, modelName, baseLayers, fn); err != nil {
 			t.Fatal(err)
 		}
 	}
 
-	var (
-		searchRequests []api.WebSearchRequest
-		fetchRequests  []api.WebFetchRequest
-	)
-
 	testCases := []testCase{
 		{
 			Name:   "Version Handler",
@@ -464,69 +447,6 @@ func TestRoutes(t *testing.T) {
 				}
 			},
 		},
-		{
-			Name:   "Web Search Handler",
-			Method: http.MethodPost,
-			Path:   "/api/web_search",
-			Setup: func(t *testing.T, req *http.Request) {
-				searchRequests = nil
-				payload := api.WebSearchRequest{Query: "cats", MaxResults: 2}
-				data, err := json.Marshal(payload)
-				if err != nil {
-					t.Fatalf("failed to marshal request: %v", err)
-				}
-				req.Body = io.NopCloser(bytes.NewReader(data))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, resp *http.Response) {
-				if resp.StatusCode != http.StatusOK {
-					t.Fatalf("expected status 200, got %d", resp.StatusCode)
-				}
-				var out api.WebSearchResponse
-				if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
-					t.Fatalf("failed to decode response: %v", err)
-				}
-				if len(out.Results) != 1 || out.Results[0].Title != "Result" {
-					t.Fatalf("unexpected response: %+v", out)
-				}
-				if len(searchRequests) != 1 {
-					t.Fatalf("expected 1 forwarded request, got %d", len(searchRequests))
-				}
-				if searchRequests[0].Query != "cats" || searchRequests[0].MaxResults != 2 {
-					t.Fatalf("unexpected forwarded request: %+v", searchRequests[0])
-				}
-			},
-		},
-		{
-			Name:   "Web Fetch Handler",
-			Method: http.MethodPost,
-			Path:   "/api/web_fetch",
-			Setup: func(t *testing.T, req *http.Request) {
-				fetchRequests = nil
-				payload := api.WebFetchRequest{URL: "https://example.com"}
-				data, err := json.Marshal(payload)
-				if err != nil {
-					t.Fatalf("failed to marshal request: %v", err)
-				}
-				req.Body = io.NopCloser(bytes.NewReader(data))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, resp *http.Response) {
-				if resp.StatusCode != http.StatusOK {
-					t.Fatalf("expected status 200, got %d", resp.StatusCode)
-				}
-				var out api.WebFetchResponse
-				if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
-					t.Fatalf("failed to decode response: %v", err)
-				}
-				if out.Title != "Example" || len(out.Links) != 1 {
-					t.Fatalf("unexpected response: %+v", out)
-				}
-				if len(fetchRequests) != 1 || fetchRequests[0].URL != "https://example.com" {
-					t.Fatalf("unexpected forwarded request: %+v", fetchRequests)
-				}
-			},
-		},
 		{
 			Name: "openai retrieve model handler",
 			Setup: func(t *testing.T, req *http.Request) {
@@ -585,40 +505,7 @@ func TestRoutes(t *testing.T) {
 		HTTPClient: panicOnRoundTrip,
 	}
 
-	remoteSrv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		switch r.URL.Path {
-		case "/api/web_search":
-			var req api.WebSearchRequest
-			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-				w.WriteHeader(http.StatusBadRequest)
-				return
-			}
-			searchRequests = append(searchRequests, req)
-			resp := api.WebSearchResponse{Results: []api.WebSearchResult{{Title: "Result", URL: "https://example.com", Content: "snippet"}}}
-			_ = json.NewEncoder(w).Encode(resp)
-		case "/api/web_fetch":
-			var req api.WebFetchRequest
-			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-				w.WriteHeader(http.StatusBadRequest)
-				return
-			}
-			fetchRequests = append(fetchRequests, req)
-			resp := api.WebFetchResponse{Title: "Example", Content: "content", Links: []string{"https://example.com"}}
-			_ = json.NewEncoder(w).Encode(resp)
-		default:
-			w.WriteHeader(http.StatusNotFound)
-		}
-	}))
-	defer remoteSrv.Close()
-
-	remoteURL, err := url.Parse(remoteSrv.URL)
-	if err != nil {
-		t.Fatalf("parse remote server URL: %v", err)
-	}
-
-	s := &Server{
-		cloudBaseURL: remoteURL,
-	}
+	s := &Server{}
 	router, err := s.GenerateRoutes(rc)
 	if err != nil {
 		t.Fatalf("failed to generate routes: %v", err)

server/sched.go

@@ -382,7 +382,10 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 // load creates a new model based on req and loads it. If requireFull is true then the model must be loaded fully onto GPUs
 // (if any). Returns whether the scheduler needs to evict a model to make this one fit.
 func (s *Scheduler) load(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, requireFull bool) bool {
-	numParallel := max(int(envconfig.NumParallel()), 1)
+	numParallel := int(envconfig.NumParallel())
+	if numParallel < 1 {
+		numParallel = 1
+	}
 
 	// Embedding models should always be loaded with parallel=1
 	if req.model.CheckCapabilities(model.CapabilityCompletion) != nil {

tools/tools.go

@@ -273,21 +273,9 @@ func findArguments(buffer []byte) (map[string]any, int) {
 						if args, ok := obj["arguments"].(map[string]any); ok {
 							return args, true
 						}
-						if argsStr, ok := obj["arguments"].(string); ok {
-							var argsData map[string]interface{}
-							if err := json.Unmarshal([]byte(argsStr), &argsData); err == nil {
-								return argsData, ok
-							}
-						}
 						if args, ok := obj["parameters"].(map[string]any); ok {
 							return args, true
 						}
-						if argsStr, ok := obj["parameters"].(string); ok {
-							var argsData map[string]interface{}
-							if err := json.Unmarshal([]byte(argsStr), &argsData); err == nil {
-								return argsData, ok
-							}
-						}
 						return nil, true
 					}