cmd: add eval command for lightweight model evals

.gitattributes

@@ -19,8 +19,6 @@ ml/backend/**/*.comp linguist-vendored
 ml/backend/**/*.glsl linguist-vendored
 ml/backend/**/CMakeLists.txt linguist-vendored
 
-app/webview linguist-vendored
-
 llama/build-info.cpp linguist-generated
 ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.s linguist-generated
 

.github/workflows/release.yaml

@@ -16,15 +16,13 @@ jobs:
     outputs:
       GOFLAGS: ${{ steps.goflags.outputs.GOFLAGS }}
       VERSION: ${{ steps.goflags.outputs.VERSION }}
-      vendorsha: ${{ steps.changes.outputs.vendorsha }}
     steps:
       - uses: actions/checkout@v4
       - name: Set environment
         id: goflags
         run: |
-          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" | tee -a $GITHUB_OUTPUT
-          echo VERSION="${GITHUB_REF_NAME#v}" | tee -a $GITHUB_OUTPUT
-          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_OUTPUT
+          echo VERSION="${GITHUB_REF_NAME#v}" >>$GITHUB_OUTPUT
 
   darwin-build:
     runs-on: macos-14-xlarge
@@ -55,9 +53,6 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
-          cache-dependency-path: |
-            go.sum
-            Makefile.sync
       - run: |
           ./scripts/build_darwin.sh
       - name: Log build results
@@ -190,7 +185,7 @@ jobs:
       - uses: actions/cache@v4
         with:
           path: ${{ github.workspace }}\.ccache
-          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}-${{ needs.setup-environment.outputs.vendorsha }}
+          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
       - name: Build target "${{ matrix.preset }}"
         run: |
           Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
@@ -254,9 +249,6 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
-          cache-dependency-path: |
-            go.sum
-            Makefile.sync
       - name: Verify gcc is actually clang
         run: |
           $ErrorActionPreference='Continue'
@@ -310,9 +302,6 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
-          cache-dependency-path: |
-            go.sum
-            Makefile.sync
       - uses: actions/download-artifact@v4
         with:
           pattern: depends-windows*

.github/workflows/test.yaml

@@ -22,7 +22,6 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       changed: ${{ steps.changes.outputs.changed }}
-      vendorsha: ${{ steps.changes.outputs.vendorsha }}
     steps:
       - uses: actions/checkout@v4
         with:
@@ -38,7 +37,6 @@ jobs:
           }
 
           echo changed=$(changed 'llama/llama.cpp/**/*' 'ml/backend/ggml/ggml/**/*') | tee -a $GITHUB_OUTPUT
-          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
 
   linux:
     needs: [changes]
@@ -85,7 +83,7 @@ jobs:
       - uses: actions/cache@v4
         with:
           path: /github/home/.cache/ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
       - run: |
           cmake --preset ${{ matrix.preset }} ${{ matrix.flags }}
           cmake --build --preset ${{ matrix.preset }} --parallel
@@ -180,7 +178,7 @@ jobs:
       - uses: actions/cache@v4
         with:
           path: ${{ github.workspace }}\.ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}-${{ needs.changes.outputs.vendorsha }}
+          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
       - run: |
           Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
           Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
@@ -208,9 +206,6 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: 'go.mod'
-          cache-dependency-path: |
-            go.sum
-            Makefile.sync
       - uses: actions/setup-node@v4
         with:
           node-version: '20'

.golangci.yaml

@@ -1,51 +1,77 @@
 version: "2"
 linters:
+  default: none
   enable:
     - asasalint
     - bidichk
     - bodyclose
     - containedctx
+    - copyloopvar
+    - errcheck
+    - errorlint
+    - exptostd
     - gocheckcompilerdirectives
+    - gocritic
+    - govet
+    - ineffassign
     - intrange
     - makezero
     - misspell
+    - modernize
     - nilerr
+    - nilnil
     - nolintlint
     - nosprintfhostport
+    - perfsprint
+    - prealloc
+    - sloglint
+    - staticcheck
     - unconvert
+    - unused
+    - usestdlibvars
     - usetesting
     - wastedassign
     - whitespace
-  disable:
-    - errcheck
-    - usestdlibvars
   settings:
-    govet:
-      disable:
-        - unusedresult
+    errcheck:
+      exclude-functions:
+        - fmt.Fprintf
+    perfsprint:
+      strconcat: false
+      concat-loop: false
     staticcheck:
       checks:
         - all
-        - -QF* # disable quick fix suggestions
+        # Using a deprecated function, variable, constant or field.
+        # https://staticcheck.dev/docs/checks/#SA1019
         - -SA1019
-        - -ST1000 # package comment format
-        - -ST1003 # underscores in package names
-        - -ST1005 # error strings should not be capitalized
-        - -ST1012 # error var naming (ErrFoo)
-        - -ST1016 # receiver name consistency
-        - -ST1020 # comment on exported function format
-        - -ST1021 # comment on exported type format
-        - -ST1022 # comment on exported var format
-        - -ST1023 # omit type from declaration
-severity:
-  default: error
-  rules:
-    - linters:
-        - gofmt
-        - goimports
-        - intrange
-      severity: info
+        # Incorrect or missing package comment.
+        # https://staticcheck.dev/docs/checks/#ST1000
+        - -ST1000
+        # Poorly chosen identifier.
+        # https://staticcheck.dev/docs/checks/#ST1003
+        - -ST1003
+        # The documentation of an exported function should start with the function's name.
+        # https://staticcheck.dev/docs/checks/#ST1020
+        - -ST1020
+        # The documentation of an exported type should start with type's name.
+        # https://staticcheck.dev/docs/checks/#ST1021
+        - -ST1021
+        # The documentation of an exported variable or constant should start with variable's name.
+        # https://staticcheck.dev/docs/checks/#ST1022
+        - -ST1022
+    usestdlibvars:
+      http-method: false
+      http-status-code: false
+
 formatters:
   enable:
+    - gci
     - gofmt
     - gofumpt
+  settings:
+    gci:
+      sections:
+        - standard
+        - default
+        - localmodule

CMakeLists.txt

@@ -12,7 +12,7 @@ set(BUILD_SHARED_LIBS ON)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
+set(CMAKE_CXX_EXTENSIONS OFF)
 
 set(GGML_BUILD ON)
 set(GGML_SHARED ON)
@@ -54,13 +54,6 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cp
 
 add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
 
-# Define GGML version variables for shared library SOVERSION
-# These are required by ggml/src/CMakeLists.txt for proper library versioning
-set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 0)
-set(GGML_VERSION_PATCH 0)
-set(GGML_VERSION "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
-
 set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)
@@ -147,30 +140,14 @@ if(CMAKE_HIP_COMPILER)
     endif()
 endif()
 
-if(NOT APPLE)
-    find_package(Vulkan)
-    if(Vulkan_FOUND)
-        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-        install(TARGETS ggml-vulkan
-            RUNTIME_DEPENDENCIES
-                PRE_INCLUDE_REGEXES vulkan
-                PRE_EXCLUDE_REGEXES ".*"
-            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        )
-    endif()
-endif()
-
-option(MLX_ENGINE "Enable MLX backend" OFF)
-
-if(MLX_ENGINE)
-    message(STATUS "Setting up MLX (this takes a while...)")
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/ml/backend/mlx)
-    install(TARGETS mlx mlxc
+find_package(Vulkan)
+if(Vulkan_FOUND)
+    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
+    install(TARGETS ggml-vulkan
         RUNTIME_DEPENDENCIES
+            PRE_INCLUDE_REGEXES vulkan
             PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/lib/ollama COMPONENT MLX
-        LIBRARY DESTINATION ${CMAKE_INSTALL_PREFIX}/lib/ollama COMPONENT MLX
-        FRAMEWORK DESTINATION ${CMAKE_INSTALL_PREFIX}/lib/ollama COMPONENT MLX
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
     )
-endif()
+endif()

CMakePresets.json

@@ -83,14 +83,6 @@
       "cacheVariables": {
         "OLLAMA_RUNNER_DIR": "vulkan"
       }
-    },
-    {
-      "name": "MLX",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "MLX_ENGINE": "ON",
-        "OLLAMA_RUNNER_DIR": "mlx"
-      }
     }
   ],
   "buildPresets": [
@@ -148,11 +140,6 @@
       "name": "Vulkan",
       "targets": [ "ggml-vulkan" ],
       "configurePreset": "Vulkan"
-    },
-    {
-      "name": "MLX",
-      "targets": [ "mlx", "mlxc" ],
-      "configurePreset": "MLX"
     }
   ]
 }

Dockerfile

@@ -131,36 +131,7 @@ COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'Vulkan' \
         && cmake --build --parallel --preset 'Vulkan' \
-        && cmake --install build --component Vulkan --strip --parallel 8
-
-FROM base AS mlx
-ARG CUDA13VERSION=13.0
-RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
-    && dnf install -y openblas-devel lapack-devel
-ENV PATH=/usr/local/cuda-13/bin:$PATH
-ENV BLAS_INCLUDE_DIRS=/usr/include/openblas
-ENV LAPACK_INCLUDE_DIRS=/usr/include/openblas
-ARG PARALLEL
-WORKDIR /go/src/github.com/ollama/ollama
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-COPY x/ml/backend/mlx x/ml/backend/mlx
-COPY go.mod go.sum .
-RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-$(case $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
-ENV PATH=/usr/local/go/bin:$PATH
-RUN go mod download
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'MLX' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
-        && cmake --build --parallel ${PARALLEL} --preset 'MLX' \
-        && cmake --install build --component MLX --strip --parallel ${PARALLEL}
-COPY . .
-ARG GOFLAGS="'-ldflags=-w -s'"
-ENV CGO_ENABLED=1
-ARG CGO_CFLAGS
-ARG CGO_CXXFLAGS
-# TODO wire up the actual MLX engine here instead of building the main binary...
-RUN go build -tags mlx -trimpath -buildmode=pie -o /bin/ollama-mlx-engine .
-RUN go build -trimpath -buildmode=pie -o /bin/imagegen ./x/imagegen/cmd/engine
+        && cmake --install build --component Vulkan --strip --parallel 8 
 
 
 FROM base AS build
@@ -182,8 +153,6 @@ FROM --platform=linux/amd64 scratch AS amd64
 COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
 COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
 COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
-COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/
-COPY --from=mlx     /bin/ /bin/
 
 FROM --platform=linux/arm64 scratch AS arm64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=ec98e2002
+FETCH_HEAD=3cfa9c3f125763305b4226bc032f1954f08990dc
 
 .PHONY: help
 help:
@@ -57,7 +57,7 @@ checkout: $(WORKDIR)
 $(WORKDIR):
 	git clone $(UPSTREAM) $(WORKDIR)
 
-.PHONY: format-patches
+.PHONE: format-patches
 format-patches: llama/patches
 	git -C $(WORKDIR) format-patch \
 		--no-signature \
@@ -66,11 +66,7 @@ format-patches: llama/patches
 		-o $(realpath $<) \
 		$(FETCH_HEAD)
 
-.PHONY: clean
+.PHONE: clean
 clean: checkout
 	@git -C $(WORKDIR) am --abort || true
 	$(RM) llama/patches/.*.patched
-
-.PHONY: print-base
-print-base:
-	@echo $(FETCH_HEAD)

README.md

@@ -555,7 +555,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Parakeet](https://github.com/parakeet-nest/parakeet) is a GoLang library, made to simplify the development of small generative AI applications with Ollama.
 - [Haverscript](https://github.com/andygill/haverscript) with [examples](https://github.com/andygill/haverscript/tree/main/examples)
 - [Ollama for Swift](https://github.com/mattt/ollama-swift)
-- [Swollama for Swift](https://github.com/guitaripod/Swollama) with [DocC](https://guitaripod.github.io/Swollama/documentation/swollama)
+- [Swollama for Swift](https://github.com/marcusziade/Swollama) with [DocC](https://marcusziade.github.io/Swollama/documentation/swollama/)
 - [GoLamify](https://github.com/prasad89/golamify)
 - [Ollama for Haskell](https://github.com/tusharad/ollama-haskell)
 - [multi-llm-ts](https://github.com/nbonamy/multi-llm-ts) (A Typescript/JavaScript library allowing access to different LLM in a unified API)

api/client.go

@@ -226,14 +226,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 
 		bts := scanner.Bytes()
 		if err := json.Unmarshal(bts, &errorResponse); err != nil {
-			if response.StatusCode >= http.StatusBadRequest {
-				return StatusError{
-					StatusCode:   response.StatusCode,
-					Status:       response.Status,
-					ErrorMessage: string(bts),
-				}
-			}
-			return errors.New(string(bts))
+			return fmt.Errorf("unmarshal: %w", err)
 		}
 
 		if response.StatusCode == http.StatusUnauthorized {
@@ -347,7 +340,7 @@ type CreateProgressFunc func(ProgressResponse) error
 // Create creates a model from a [Modelfile]. fn is a progress function that
 // behaves similarly to other methods (see [Client.Pull]).
 //
-// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.mdx
+// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.md
 func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/create", req, func(bts []byte) error {
 		var resp ProgressResponse

api/client_test.go

@@ -55,7 +55,6 @@ func TestClientFromEnvironment(t *testing.T) {
 type testError struct {
 	message    string
 	statusCode int
-	raw        bool // if true, write message as-is instead of JSON encoding
 }
 
 func (e testError) Error() string {
@@ -112,20 +111,6 @@ func TestClientStream(t *testing.T) {
 				},
 			},
 		},
-		{
-			name: "plain text error response",
-			responses: []any{
-				"internal server error",
-			},
-			wantErr: "internal server error",
-		},
-		{
-			name: "HTML error page",
-			responses: []any{
-				"<html><body>404 Not Found</body></html>",
-			},
-			wantErr: "404 Not Found",
-		},
 	}
 
 	for _, tc := range testCases {
@@ -150,12 +135,6 @@ func TestClientStream(t *testing.T) {
 						return
 					}
 
-					if str, ok := resp.(string); ok {
-						fmt.Fprintln(w, str)
-						flusher.Flush()
-						continue
-					}
-
 					if err := json.NewEncoder(w).Encode(resp); err != nil {
 						t.Fatalf("failed to encode response: %v", err)
 					}
@@ -194,10 +173,9 @@ func TestClientStream(t *testing.T) {
 
 func TestClientDo(t *testing.T) {
 	testCases := []struct {
-		name           string
-		response       any
-		wantErr        string
-		wantStatusCode int
+		name     string
+		response any
+		wantErr  string
 	}{
 		{
 			name: "immediate error response",
@@ -205,8 +183,7 @@ func TestClientDo(t *testing.T) {
 				message:    "test error message",
 				statusCode: http.StatusBadRequest,
 			},
-			wantErr:        "test error message",
-			wantStatusCode: http.StatusBadRequest,
+			wantErr: "test error message",
 		},
 		{
 			name: "server error response",
@@ -214,8 +191,7 @@ func TestClientDo(t *testing.T) {
 				message:    "internal error",
 				statusCode: http.StatusInternalServerError,
 			},
-			wantErr:        "internal error",
-			wantStatusCode: http.StatusInternalServerError,
+			wantErr: "internal error",
 		},
 		{
 			name: "successful response",
@@ -227,26 +203,6 @@ func TestClientDo(t *testing.T) {
 				Success: true,
 			},
 		},
-		{
-			name: "plain text error response",
-			response: testError{
-				message:    "internal server error",
-				statusCode: http.StatusInternalServerError,
-				raw:        true,
-			},
-			wantErr:        "internal server error",
-			wantStatusCode: http.StatusInternalServerError,
-		},
-		{
-			name: "HTML error page",
-			response: testError{
-				message:    "<html><body>404 Not Found</body></html>",
-				statusCode: http.StatusNotFound,
-				raw:        true,
-			},
-			wantErr:        "<html><body>404 Not Found</body></html>",
-			wantStatusCode: http.StatusNotFound,
-		},
 	}
 
 	for _, tc := range testCases {
@@ -254,16 +210,11 @@ func TestClientDo(t *testing.T) {
 			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 				if errResp, ok := tc.response.(testError); ok {
 					w.WriteHeader(errResp.statusCode)
-					if !errResp.raw {
-						err := json.NewEncoder(w).Encode(map[string]string{
-							"error": errResp.message,
-						})
-						if err != nil {
-							t.Fatal("failed to encode error response:", err)
-						}
-					} else {
-						// Write raw message (simulates non-JSON error responses)
-						fmt.Fprint(w, errResp.message)
+					err := json.NewEncoder(w).Encode(map[string]string{
+						"error": errResp.message,
+					})
+					if err != nil {
+						t.Fatal("failed to encode error response:", err)
 					}
 					return
 				}
@@ -290,15 +241,6 @@ func TestClientDo(t *testing.T) {
 				if err.Error() != tc.wantErr {
 					t.Errorf("error message mismatch: got %q, want %q", err.Error(), tc.wantErr)
 				}
-				if tc.wantStatusCode != 0 {
-					if statusErr, ok := err.(StatusError); ok {
-						if statusErr.StatusCode != tc.wantStatusCode {
-							t.Errorf("status code mismatch: got %d, want %d", statusErr.StatusCode, tc.wantStatusCode)
-						}
-					} else {
-						t.Errorf("expected StatusError, got %T", err)
-					}
-				}
 				return
 			}
 

api/examples/chat/main.go

@@ -15,19 +15,19 @@ func main() {
 	}
 
 	messages := []api.Message{
-		{
+		api.Message{
 			Role:    "system",
 			Content: "Provide very brief, concise responses",
 		},
-		{
+		api.Message{
 			Role:    "user",
 			Content: "Name some unusual animals",
 		},
-		{
+		api.Message{
 			Role:    "assistant",
 			Content: "Monotreme, platypus, echidna",
 		},
-		{
+		api.Message{
 			Role:    "user",
 			Content: "which of these is the most dangerous?",
 		},

api/types.go

@@ -3,7 +3,6 @@ package api
 import (
 	"encoding/json"
 	"fmt"
-	"iter"
 	"log/slog"
 	"math"
 	"os"
@@ -15,7 +14,6 @@ import (
 	"github.com/google/uuid"
 
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/internal/orderedmap"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -229,79 +227,13 @@ type ToolCallFunction struct {
 	Arguments ToolCallFunctionArguments `json:"arguments"`
 }
 
-// ToolCallFunctionArguments holds tool call arguments in insertion order.
-type ToolCallFunctionArguments struct {
-	om *orderedmap.Map[string, any]
-}
-
-// NewToolCallFunctionArguments creates a new empty ToolCallFunctionArguments.
-func NewToolCallFunctionArguments() ToolCallFunctionArguments {
-	return ToolCallFunctionArguments{om: orderedmap.New[string, any]()}
-}
-
-// Get retrieves a value by key.
-func (t *ToolCallFunctionArguments) Get(key string) (any, bool) {
-	if t == nil || t.om == nil {
-		return nil, false
-	}
-	return t.om.Get(key)
-}
-
-// Set sets a key-value pair, preserving insertion order.
-func (t *ToolCallFunctionArguments) Set(key string, value any) {
-	if t == nil {
-		return
-	}
-	if t.om == nil {
-		t.om = orderedmap.New[string, any]()
-	}
-	t.om.Set(key, value)
-}
-
-// Len returns the number of arguments.
-func (t *ToolCallFunctionArguments) Len() int {
-	if t == nil || t.om == nil {
-		return 0
-	}
-	return t.om.Len()
-}
-
-// All returns an iterator over all key-value pairs in insertion order.
-func (t *ToolCallFunctionArguments) All() iter.Seq2[string, any] {
-	if t == nil || t.om == nil {
-		return func(yield func(string, any) bool) {}
-	}
-	return t.om.All()
-}
-
-// ToMap returns a regular map (order not preserved).
-func (t *ToolCallFunctionArguments) ToMap() map[string]any {
-	if t == nil || t.om == nil {
-		return nil
-	}
-	return t.om.ToMap()
-}
+type ToolCallFunctionArguments map[string]any
 
 func (t *ToolCallFunctionArguments) String() string {
-	if t == nil || t.om == nil {
-		return "{}"
-	}
-	bts, _ := json.Marshal(t.om)
+	bts, _ := json.Marshal(t)
 	return string(bts)
 }
 
-func (t *ToolCallFunctionArguments) UnmarshalJSON(data []byte) error {
-	t.om = orderedmap.New[string, any]()
-	return json.Unmarshal(data, t.om)
-}
-
-func (t ToolCallFunctionArguments) MarshalJSON() ([]byte, error) {
-	if t.om == nil {
-		return []byte("{}"), nil
-	}
-	return json.Marshal(t.om)
-}
-
 type Tool struct {
 	Type     string       `json:"type"`
 	Items    any          `json:"items,omitempty"`
@@ -350,78 +282,12 @@ func (pt PropertyType) String() string {
 	return fmt.Sprintf("%v", []string(pt))
 }
 
-// ToolPropertiesMap holds tool properties in insertion order.
-type ToolPropertiesMap struct {
-	om *orderedmap.Map[string, ToolProperty]
-}
-
-// NewToolPropertiesMap creates a new empty ToolPropertiesMap.
-func NewToolPropertiesMap() *ToolPropertiesMap {
-	return &ToolPropertiesMap{om: orderedmap.New[string, ToolProperty]()}
-}
-
-// Get retrieves a property by name.
-func (t *ToolPropertiesMap) Get(key string) (ToolProperty, bool) {
-	if t == nil || t.om == nil {
-		return ToolProperty{}, false
-	}
-	return t.om.Get(key)
-}
-
-// Set sets a property, preserving insertion order.
-func (t *ToolPropertiesMap) Set(key string, value ToolProperty) {
-	if t == nil {
-		return
-	}
-	if t.om == nil {
-		t.om = orderedmap.New[string, ToolProperty]()
-	}
-	t.om.Set(key, value)
-}
-
-// Len returns the number of properties.
-func (t *ToolPropertiesMap) Len() int {
-	if t == nil || t.om == nil {
-		return 0
-	}
-	return t.om.Len()
-}
-
-// All returns an iterator over all properties in insertion order.
-func (t *ToolPropertiesMap) All() iter.Seq2[string, ToolProperty] {
-	if t == nil || t.om == nil {
-		return func(yield func(string, ToolProperty) bool) {}
-	}
-	return t.om.All()
-}
-
-// ToMap returns a regular map (order not preserved).
-func (t *ToolPropertiesMap) ToMap() map[string]ToolProperty {
-	if t == nil || t.om == nil {
-		return nil
-	}
-	return t.om.ToMap()
-}
-
-func (t ToolPropertiesMap) MarshalJSON() ([]byte, error) {
-	if t.om == nil {
-		return []byte("null"), nil
-	}
-	return json.Marshal(t.om)
-}
-
-func (t *ToolPropertiesMap) UnmarshalJSON(data []byte) error {
-	t.om = orderedmap.New[string, ToolProperty]()
-	return json.Unmarshal(data, t.om)
-}
-
 type ToolProperty struct {
-	AnyOf       []ToolProperty     `json:"anyOf,omitempty"`
-	Type        PropertyType       `json:"type,omitempty"`
-	Items       any                `json:"items,omitempty"`
-	Description string             `json:"description,omitempty"`
-	Enum        []any              `json:"enum,omitempty"`
-	Properties  *ToolPropertiesMap `json:"properties,omitempty"`
+	AnyOf       []ToolProperty `json:"anyOf,omitempty"`
+	Type        PropertyType   `json:"type,omitempty"`
+	Items       any            `json:"items,omitempty"`
+	Description string         `json:"description,omitempty"`
+	Enum        []any          `json:"enum,omitempty"`
 }
 
 // ToTypeScriptType converts a ToolProperty to a TypeScript type string
@@ -470,11 +336,11 @@ func mapToTypeScriptType(jsonType string) string {
 }
 
 type ToolFunctionParameters struct {
-	Type       string             `json:"type"`
-	Defs       any                `json:"$defs,omitempty"`
-	Items      any                `json:"items,omitempty"`
-	Required   []string           `json:"required,omitempty"`
-	Properties *ToolPropertiesMap `json:"properties"`
+	Type       string                  `json:"type"`
+	Defs       any                     `json:"$defs,omitempty"`
+	Items      any                     `json:"items,omitempty"`
+	Required   []string                `json:"required,omitempty"`
+	Properties map[string]ToolProperty `json:"properties"`
 }
 
 func (t *ToolFunctionParameters) String() string {
@@ -687,9 +553,6 @@ type CreateRequest struct {
 	Renderer string `json:"renderer,omitempty"`
 	Parser   string `json:"parser,omitempty"`
 
-	// Requires is the minimum version of Ollama required by the model.
-	Requires string `json:"requires,omitempty"`
-
 	// Info is a map of additional information for the model
 	Info map[string]any `json:"info,omitempty"`
 
@@ -740,7 +603,6 @@ type ShowResponse struct {
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`
 	ModifiedAt    time.Time          `json:"modified_at,omitempty"`
-	Requires      string             `json:"requires,omitempty"`
 }
 
 // CopyRequest is the request passed to [Client.Copy].

api/types_test.go

@@ -11,24 +11,6 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-// testPropsMap creates a ToolPropertiesMap from a map (convenience function for tests, order not preserved)
-func testPropsMap(m map[string]ToolProperty) *ToolPropertiesMap {
-	props := NewToolPropertiesMap()
-	for k, v := range m {
-		props.Set(k, v)
-	}
-	return props
-}
-
-// testArgs creates ToolCallFunctionArguments from a map (convenience function for tests, order not preserved)
-func testArgs(m map[string]any) ToolCallFunctionArguments {
-	args := NewToolCallFunctionArguments()
-	for k, v := range m {
-		args.Set(k, v)
-	}
-	return args
-}
-
 func TestKeepAliveParsingFromJSON(t *testing.T) {
 	tests := []struct {
 		name string
@@ -327,9 +309,9 @@ func TestToolFunctionParameters_MarshalJSON(t *testing.T) {
 			input: ToolFunctionParameters{
 				Type:     "object",
 				Required: []string{"name"},
-				Properties: testPropsMap(map[string]ToolProperty{
+				Properties: map[string]ToolProperty{
 					"name": {Type: PropertyType{"string"}},
-				}),
+				},
 			},
 			expected: `{"type":"object","required":["name"],"properties":{"name":{"type":"string"}}}`,
 		},
@@ -337,9 +319,9 @@ func TestToolFunctionParameters_MarshalJSON(t *testing.T) {
 			name: "no required",
 			input: ToolFunctionParameters{
 				Type: "object",
-				Properties: testPropsMap(map[string]ToolProperty{
+				Properties: map[string]ToolProperty{
 					"name": {Type: PropertyType{"string"}},
-				}),
+				},
 			},
 			expected: `{"type":"object","properties":{"name":{"type":"string"}}}`,
 		},
@@ -357,7 +339,7 @@ func TestToolFunctionParameters_MarshalJSON(t *testing.T) {
 func TestToolCallFunction_IndexAlwaysMarshals(t *testing.T) {
 	fn := ToolCallFunction{
 		Name:      "echo",
-		Arguments: testArgs(map[string]any{"message": "hi"}),
+		Arguments: ToolCallFunctionArguments{"message": "hi"},
 	}
 
 	data, err := json.Marshal(fn)
@@ -522,116 +504,6 @@ func TestThinking_UnmarshalJSON(t *testing.T) {
 	}
 }
 
-func TestToolPropertyNestedProperties(t *testing.T) {
-	tests := []struct {
-		name     string
-		input    string
-		expected ToolProperty
-	}{
-		{
-			name: "nested object properties",
-			input: `{
-				"type": "object",
-				"description": "Location details",
-				"properties": {
-					"address": {
-						"type": "string",
-						"description": "Street address"
-					},
-					"city": {
-						"type": "string",
-						"description": "City name"
-					}
-				}
-			}`,
-			expected: ToolProperty{
-				Type:        PropertyType{"object"},
-				Description: "Location details",
-				Properties: testPropsMap(map[string]ToolProperty{
-					"address": {
-						Type:        PropertyType{"string"},
-						Description: "Street address",
-					},
-					"city": {
-						Type:        PropertyType{"string"},
-						Description: "City name",
-					},
-				}),
-			},
-		},
-		{
-			name: "deeply nested properties",
-			input: `{
-				"type": "object",
-				"description": "Event",
-				"properties": {
-					"location": {
-						"type": "object",
-						"description": "Location",
-						"properties": {
-							"coordinates": {
-								"type": "object",
-								"description": "GPS coordinates",
-								"properties": {
-									"lat": {"type": "number", "description": "Latitude"},
-									"lng": {"type": "number", "description": "Longitude"}
-								}
-							}
-						}
-					}
-				}
-			}`,
-			expected: ToolProperty{
-				Type:        PropertyType{"object"},
-				Description: "Event",
-				Properties: testPropsMap(map[string]ToolProperty{
-					"location": {
-						Type:        PropertyType{"object"},
-						Description: "Location",
-						Properties: testPropsMap(map[string]ToolProperty{
-							"coordinates": {
-								Type:        PropertyType{"object"},
-								Description: "GPS coordinates",
-								Properties: testPropsMap(map[string]ToolProperty{
-									"lat": {Type: PropertyType{"number"}, Description: "Latitude"},
-									"lng": {Type: PropertyType{"number"}, Description: "Longitude"},
-								}),
-							},
-						}),
-					},
-				}),
-			},
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			var prop ToolProperty
-			err := json.Unmarshal([]byte(tt.input), &prop)
-			require.NoError(t, err)
-
-			// Compare JSON representations since pointer comparison doesn't work
-			expectedJSON, err := json.Marshal(tt.expected)
-			require.NoError(t, err)
-			actualJSON, err := json.Marshal(prop)
-			require.NoError(t, err)
-			assert.JSONEq(t, string(expectedJSON), string(actualJSON))
-
-			// Round-trip test: marshal and unmarshal again
-			data, err := json.Marshal(prop)
-			require.NoError(t, err)
-
-			var prop2 ToolProperty
-			err = json.Unmarshal(data, &prop2)
-			require.NoError(t, err)
-
-			prop2JSON, err := json.Marshal(prop2)
-			require.NoError(t, err)
-			assert.JSONEq(t, string(expectedJSON), string(prop2JSON))
-		})
-	}
-}
-
 func TestToolFunctionParameters_String(t *testing.T) {
 	tests := []struct {
 		name     string
@@ -643,12 +515,12 @@ func TestToolFunctionParameters_String(t *testing.T) {
 			params: ToolFunctionParameters{
 				Type:     "object",
 				Required: []string{"name"},
-				Properties: testPropsMap(map[string]ToolProperty{
+				Properties: map[string]ToolProperty{
 					"name": {
 						Type:        PropertyType{"string"},
 						Description: "The name of the person",
 					},
-				}),
+				},
 			},
 			expected: `{"type":"object","required":["name"],"properties":{"name":{"type":"string","description":"The name of the person"}}}`,
 		},
@@ -665,7 +537,7 @@ func TestToolFunctionParameters_String(t *testing.T) {
 					s.Self = s
 					return s
 				}(),
-				Properties: testPropsMap(map[string]ToolProperty{}),
+				Properties: map[string]ToolProperty{},
 			},
 			expected: "",
 		},
@@ -678,235 +550,3 @@ func TestToolFunctionParameters_String(t *testing.T) {
 		})
 	}
 }
-
-func TestToolCallFunctionArguments_OrderPreservation(t *testing.T) {
-	t.Run("marshal preserves insertion order", func(t *testing.T) {
-		args := NewToolCallFunctionArguments()
-		args.Set("zebra", "z")
-		args.Set("apple", "a")
-		args.Set("mango", "m")
-
-		data, err := json.Marshal(args)
-		require.NoError(t, err)
-
-		// Should preserve insertion order, not alphabetical
-		assert.Equal(t, `{"zebra":"z","apple":"a","mango":"m"}`, string(data))
-	})
-
-	t.Run("unmarshal preserves JSON order", func(t *testing.T) {
-		jsonData := `{"zebra":"z","apple":"a","mango":"m"}`
-
-		var args ToolCallFunctionArguments
-		err := json.Unmarshal([]byte(jsonData), &args)
-		require.NoError(t, err)
-
-		// Verify iteration order matches JSON order
-		var keys []string
-		for k := range args.All() {
-			keys = append(keys, k)
-		}
-		assert.Equal(t, []string{"zebra", "apple", "mango"}, keys)
-	})
-
-	t.Run("round trip preserves order", func(t *testing.T) {
-		original := `{"z":1,"a":2,"m":3,"b":4}`
-
-		var args ToolCallFunctionArguments
-		err := json.Unmarshal([]byte(original), &args)
-		require.NoError(t, err)
-
-		data, err := json.Marshal(args)
-		require.NoError(t, err)
-
-		assert.Equal(t, original, string(data))
-	})
-
-	t.Run("String method returns ordered JSON", func(t *testing.T) {
-		args := NewToolCallFunctionArguments()
-		args.Set("c", 3)
-		args.Set("a", 1)
-		args.Set("b", 2)
-
-		assert.Equal(t, `{"c":3,"a":1,"b":2}`, args.String())
-	})
-
-	t.Run("Get retrieves correct values", func(t *testing.T) {
-		args := NewToolCallFunctionArguments()
-		args.Set("key1", "value1")
-		args.Set("key2", 42)
-
-		v, ok := args.Get("key1")
-		assert.True(t, ok)
-		assert.Equal(t, "value1", v)
-
-		v, ok = args.Get("key2")
-		assert.True(t, ok)
-		assert.Equal(t, 42, v)
-
-		_, ok = args.Get("nonexistent")
-		assert.False(t, ok)
-	})
-
-	t.Run("Len returns correct count", func(t *testing.T) {
-		args := NewToolCallFunctionArguments()
-		assert.Equal(t, 0, args.Len())
-
-		args.Set("a", 1)
-		assert.Equal(t, 1, args.Len())
-
-		args.Set("b", 2)
-		assert.Equal(t, 2, args.Len())
-	})
-
-	t.Run("empty args marshal to empty object", func(t *testing.T) {
-		args := NewToolCallFunctionArguments()
-		data, err := json.Marshal(args)
-		require.NoError(t, err)
-		assert.Equal(t, `{}`, string(data))
-	})
-
-	t.Run("zero value args marshal to empty object", func(t *testing.T) {
-		var args ToolCallFunctionArguments
-		assert.Equal(t, "{}", args.String())
-	})
-}
-
-func TestToolPropertiesMap_OrderPreservation(t *testing.T) {
-	t.Run("marshal preserves insertion order", func(t *testing.T) {
-		props := NewToolPropertiesMap()
-		props.Set("zebra", ToolProperty{Type: PropertyType{"string"}})
-		props.Set("apple", ToolProperty{Type: PropertyType{"number"}})
-		props.Set("mango", ToolProperty{Type: PropertyType{"boolean"}})
-
-		data, err := json.Marshal(props)
-		require.NoError(t, err)
-
-		// Should preserve insertion order, not alphabetical
-		expected := `{"zebra":{"type":"string"},"apple":{"type":"number"},"mango":{"type":"boolean"}}`
-		assert.Equal(t, expected, string(data))
-	})
-
-	t.Run("unmarshal preserves JSON order", func(t *testing.T) {
-		jsonData := `{"zebra":{"type":"string"},"apple":{"type":"number"},"mango":{"type":"boolean"}}`
-
-		var props ToolPropertiesMap
-		err := json.Unmarshal([]byte(jsonData), &props)
-		require.NoError(t, err)
-
-		// Verify iteration order matches JSON order
-		var keys []string
-		for k := range props.All() {
-			keys = append(keys, k)
-		}
-		assert.Equal(t, []string{"zebra", "apple", "mango"}, keys)
-	})
-
-	t.Run("round trip preserves order", func(t *testing.T) {
-		original := `{"z":{"type":"string"},"a":{"type":"number"},"m":{"type":"boolean"}}`
-
-		var props ToolPropertiesMap
-		err := json.Unmarshal([]byte(original), &props)
-		require.NoError(t, err)
-
-		data, err := json.Marshal(props)
-		require.NoError(t, err)
-
-		assert.Equal(t, original, string(data))
-	})
-
-	t.Run("Get retrieves correct values", func(t *testing.T) {
-		props := NewToolPropertiesMap()
-		props.Set("name", ToolProperty{Type: PropertyType{"string"}, Description: "The name"})
-		props.Set("age", ToolProperty{Type: PropertyType{"integer"}, Description: "The age"})
-
-		v, ok := props.Get("name")
-		assert.True(t, ok)
-		assert.Equal(t, "The name", v.Description)
-
-		v, ok = props.Get("age")
-		assert.True(t, ok)
-		assert.Equal(t, "The age", v.Description)
-
-		_, ok = props.Get("nonexistent")
-		assert.False(t, ok)
-	})
-
-	t.Run("Len returns correct count", func(t *testing.T) {
-		props := NewToolPropertiesMap()
-		assert.Equal(t, 0, props.Len())
-
-		props.Set("a", ToolProperty{})
-		assert.Equal(t, 1, props.Len())
-
-		props.Set("b", ToolProperty{})
-		assert.Equal(t, 2, props.Len())
-	})
-
-	t.Run("nil props marshal to null", func(t *testing.T) {
-		var props *ToolPropertiesMap
-		data, err := json.Marshal(props)
-		require.NoError(t, err)
-		assert.Equal(t, `null`, string(data))
-	})
-
-	t.Run("ToMap returns regular map", func(t *testing.T) {
-		props := NewToolPropertiesMap()
-		props.Set("a", ToolProperty{Type: PropertyType{"string"}})
-		props.Set("b", ToolProperty{Type: PropertyType{"number"}})
-
-		m := props.ToMap()
-		assert.Equal(t, 2, len(m))
-		assert.Equal(t, PropertyType{"string"}, m["a"].Type)
-		assert.Equal(t, PropertyType{"number"}, m["b"].Type)
-	})
-}
-
-func TestToolCallFunctionArguments_ComplexValues(t *testing.T) {
-	t.Run("nested objects preserve order", func(t *testing.T) {
-		jsonData := `{"outer":{"z":1,"a":2},"simple":"value"}`
-
-		var args ToolCallFunctionArguments
-		err := json.Unmarshal([]byte(jsonData), &args)
-		require.NoError(t, err)
-
-		// Outer keys should be in order
-		var keys []string
-		for k := range args.All() {
-			keys = append(keys, k)
-		}
-		assert.Equal(t, []string{"outer", "simple"}, keys)
-	})
-
-	t.Run("arrays as values", func(t *testing.T) {
-		args := NewToolCallFunctionArguments()
-		args.Set("items", []string{"a", "b", "c"})
-		args.Set("numbers", []int{1, 2, 3})
-
-		data, err := json.Marshal(args)
-		require.NoError(t, err)
-
-		assert.Equal(t, `{"items":["a","b","c"],"numbers":[1,2,3]}`, string(data))
-	})
-}
-
-func TestToolPropertiesMap_NestedProperties(t *testing.T) {
-	t.Run("nested properties preserve order", func(t *testing.T) {
-		props := NewToolPropertiesMap()
-
-		nestedProps := NewToolPropertiesMap()
-		nestedProps.Set("z_field", ToolProperty{Type: PropertyType{"string"}})
-		nestedProps.Set("a_field", ToolProperty{Type: PropertyType{"number"}})
-
-		props.Set("outer", ToolProperty{
-			Type:       PropertyType{"object"},
-			Properties: nestedProps,
-		})
-
-		data, err := json.Marshal(props)
-		require.NoError(t, err)
-
-		// Both outer and inner should preserve order
-		expected := `{"outer":{"type":"object","properties":{"z_field":{"type":"string"},"a_field":{"type":"number"}}}}`
-		assert.Equal(t, expected, string(data))
-	})
-}

app/cmd/app/app.go

@@ -273,6 +273,10 @@ func main() {
 		Handler: uiServer.Handler(),
 	}
 
+	if _, err := uiServer.UserData(ctx); err != nil {
+		slog.Warn("failed to load user data", "error", err)
+	}
+
 	// Start the UI server
 	slog.Info("starting ui server", "port", port)
 	go func() {
@@ -316,17 +320,6 @@ func main() {
 		slog.Debug("no URL scheme request to handle")
 	}
 
-	go func() {
-		slog.Debug("waiting for ollama server to be ready")
-		if err := ui.WaitForServer(ctx, 10*time.Second); err != nil {
-			slog.Warn("ollama server not ready, continuing anyway", "error", err)
-		}
-
-		if _, err := uiServer.UserData(ctx); err != nil {
-			slog.Warn("failed to load user data", "error", err)
-		}
-	}()
-
 	osRun(cancel, hasCompletedFirstRun, startHidden)
 
 	slog.Info("shutting down desktop server")
@@ -368,7 +361,7 @@ func checkUserLoggedIn(uiServerPort int) bool {
 		return false
 	}
 
-	resp, err := http.Post(fmt.Sprintf("http://127.0.0.1:%d/api/me", uiServerPort), "application/json", nil)
+	resp, err := http.Get(fmt.Sprintf("http://127.0.0.1:%d/api/v1/me", uiServerPort))
 	if err != nil {
 		slog.Debug("failed to call local auth endpoint", "error", err)
 		return false

app/cmd/app/app_darwin.go

@@ -191,6 +191,13 @@ func LaunchNewApp() {
 	C.launchApp(appName)
 }
 
+// Send a request to the main app thread to load a UI page
+func sendUIRequestMessage(path string) {
+	p := C.CString(path)
+	defer C.free(unsafe.Pointer(p))
+	C.uiRequest(p)
+}
+
 func registerLaunchAgent(hasCompletedFirstRun bool) {
 	// Remove any stale Login Item registrations
 	C.unregisterSelfFromLoginItem()

app/cmd/app/app_windows.go

@@ -263,6 +263,11 @@ func createLoginShortcut() error {
 	return nil
 }
 
+// Send a request to the main app thread to load a UI page
+func sendUIRequestMessage(path string) {
+	wintray.SendUIRequestMessage(path)
+}
+
 func LaunchNewApp() {
 }
 

app/dialog/cocoa/dlg.m

@@ -169,47 +169,37 @@ DlgResult fileDlg(FileDlgParams* params) {
 	}
 	
 	NSArray* urls = [panel URLs];
-	if([urls count] == 0) {
-		return DLG_CANCEL;
-	}
-	
-	if(self->params->allowMultiple) {
+	if(self->params->allowMultiple && [urls count] >= 1) {
 		// For multiple files, we need to return all paths separated by null bytes
 		char* bufPtr = self->params->buf;
 		int remainingBuf = self->params->nbuf;
 		
-		// Calculate total required buffer size first
-		int totalSize = 0;
-		for(NSURL* url in urls) {
-			char tempBuf[PATH_MAX];
-			if(![url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX]) {
-				return DLG_URLFAIL;
-			}
-			totalSize += strlen(tempBuf) + 1; // +1 for null terminator
-		}
-		totalSize += 1; // Final null terminator
+  // Calculate total required buffer size first
+  int totalSize = 0;
+  for(NSURL* url in urls) {
+      char tempBuf[PATH_MAX];
+      if(![url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX]) {
+          return DLG_URLFAIL;
+      }
+      totalSize += strlen(tempBuf) + 1; // +1 for null terminator
+  }
+  totalSize += 1; // Final null terminator
 
-		if(totalSize > self->params->nbuf) {
-			// Not enough buffer space
-			return DLG_URLFAIL;
-		}
+  if(totalSize > self->params->nbuf) {
+      // Not enough buffer space
+      return DLG_URLFAIL;
+  }
 
-		// Now actually copy the paths (we know we have space)
-		bufPtr = self->params->buf;
-		for(NSURL* url in urls) {
-			char tempBuf[PATH_MAX];
-			[url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX];
-			int pathLen = strlen(tempBuf);
-			strcpy(bufPtr, tempBuf);
-			bufPtr += pathLen + 1;
-		}
-		*bufPtr = '\0'; // Final null terminator
-	} else {
-		// Single file/directory selection - write path to buffer
-		NSURL* url = [urls firstObject];
-		if(![url getFileSystemRepresentation:self->params->buf maxLength:self->params->nbuf]) {
-			return DLG_URLFAIL;
-		}
+  // Now actually copy the paths (we know we have space)
+  bufPtr = self->params->buf;
+  for(NSURL* url in urls) {
+      char tempBuf[PATH_MAX];
+      [url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX];
+      int pathLen = strlen(tempBuf);
+      strcpy(bufPtr, tempBuf);
+      bufPtr += pathLen + 1;
+  }
+  *bufPtr = '\0'; // Final null terminator
 	}
 	
 	return DLG_OK;

app/dialog/dlgs_windows.go

@@ -15,7 +15,7 @@ const multiFileBufferSize = w32.MAX_PATH * 10
 type WinDlgError int
 
 func (e WinDlgError) Error() string {
-	return fmt.Sprintf("CommDlgExtendedError: %#x", int(e))
+	return fmt.Sprintf("CommDlgExtendedError: %#x", e)
 }
 
 func err() error {

app/server/server.go

@@ -224,7 +224,9 @@ func (s *Server) cmd(ctx context.Context) (*exec.Cmd, error) {
 		if _, err := os.Stat(settings.Models); err == nil {
 			env["OLLAMA_MODELS"] = settings.Models
 		} else {
-			slog.Warn("models path not accessible, using default", "path", settings.Models, "err", err)
+			slog.Warn("models path not accessible, clearing models setting", "path", settings.Models, "err", err)
+			settings.Models = ""
+			s.store.SetSettings(settings)
 		}
 	}
 	if settings.ContextLength > 0 {

app/ui/app/codegen/gotypes.gen.ts

@@ -469,24 +469,26 @@ export class HealthResponse {
 }
 export class User {
     id: string;
-    email: string;
     name: string;
-    bio?: string;
-    avatarurl?: string;
-    firstname?: string;
-    lastname?: string;
-    plan?: string;
+    email: string;
+    avatarURL: string;
+    plan: string;
+    bio: string;
+    firstName: string;
+    lastName: string;
+    overThreshold: boolean;
 
     constructor(source: any = {}) {
         if ('string' === typeof source) source = JSON.parse(source);
         this.id = source["id"];
-        this.email = source["email"];
         this.name = source["name"];
-        this.bio = source["bio"];
-        this.avatarurl = source["avatarurl"];
-        this.firstname = source["firstname"];
-        this.lastname = source["lastname"];
+        this.email = source["email"];
+        this.avatarURL = source["avatarURL"];
         this.plan = source["plan"];
+        this.bio = source["bio"];
+        this.firstName = source["firstName"];
+        this.lastName = source["lastName"];
+        this.overThreshold = source["overThreshold"];
     }
 }
 export class Attachment {

app/ui/app/src/api.ts

@@ -15,7 +15,7 @@ import {
 import { parseJsonlFromResponse } from "./util/jsonl-parsing";
 import { ollamaClient as ollama } from "./lib/ollama-client";
 import type { ModelResponse } from "ollama/browser";
-import { API_BASE, OLLAMA_DOT_COM } from "./lib/config";
+import { API_BASE } from "./lib/config";
 
 // Extend Model class with utility methods
 declare module "@/gotypes" {
@@ -27,6 +27,7 @@ declare module "@/gotypes" {
 Model.prototype.isCloud = function (): boolean {
   return this.model.endsWith("cloud");
 };
+
 // Helper function to convert Uint8Array to base64
 function uint8ArrayToBase64(uint8Array: Uint8Array): string {
   const chunkSize = 0x8000; // 32KB chunks to avoid stack overflow
@@ -41,50 +42,44 @@ function uint8ArrayToBase64(uint8Array: Uint8Array): string {
 }
 
 export async function fetchUser(): Promise<User | null> {
-  const response = await fetch(`${API_BASE}/api/me`, {
-    method: "POST",
-    headers: {
-      "Content-Type": "application/json",
-    },
-  });
+  try {
+    const response = await fetch(`${API_BASE}/api/v1/me`, {
+      method: "GET",
+      headers: {
+        "Content-Type": "application/json",
+      },
+    });
 
-  if (response.ok) {
-    const userData: User = await response.json();
-
-    if (userData.avatarurl && !userData.avatarurl.startsWith("http")) {
-      userData.avatarurl = `${OLLAMA_DOT_COM}${userData.avatarurl}`;
+    if (response.ok) {
+      const userData: User = await response.json();
+      return userData;
     }
 
-    return userData;
-  }
-
-  if (response.status === 401 || response.status === 403) {
+    return null;
+  } catch (error) {
+    console.error("Error fetching user:", error);
     return null;
   }
-
-  throw new Error(`Failed to fetch user: ${response.status}`);
 }
 
 export async function fetchConnectUrl(): Promise<string> {
-  const response = await fetch(`${API_BASE}/api/me`, {
-    method: "POST",
+  const response = await fetch(`${API_BASE}/api/v1/connect`, {
+    method: "GET",
     headers: {
       "Content-Type": "application/json",
     },
   });
 
-  if (response.status === 401) {
-    const data = await response.json();
-    if (data.signin_url) {
-      return data.signin_url;
-    }
+  if (!response.ok) {
+    throw new Error("Failed to fetch connect URL");
   }
 
-  throw new Error("Failed to fetch connect URL");
+  const data = await response.json();
+  return data.connect_url;
 }
 
 export async function disconnectUser(): Promise<void> {
-  const response = await fetch(`${API_BASE}/api/signout`, {
+  const response = await fetch(`${API_BASE}/api/v1/disconnect`, {
     method: "POST",
     headers: {
       "Content-Type": "application/json",
@@ -209,10 +204,12 @@ export async function* sendMessage(
     data: uint8ArrayToBase64(att.data),
   }));
 
-  // Send think parameter when it's explicitly set (true, false, or a non-empty string).
+  // Only send think parameter when actually requesting thinking
+  // Don't send false as it causes issues with some providers
   const shouldSendThink =
     think !== undefined &&
-    (typeof think === "boolean" || (typeof think === "string" && think !== ""));
+    ((typeof think === "boolean" && think) ||
+      (typeof think === "string" && think !== ""));
 
   const response = await fetch(`${API_BASE}/api/v1/chat/${chatId}`, {
     method: "POST",
@@ -394,8 +391,7 @@ export async function getInferenceCompute(): Promise<InferenceCompute[]> {
 
 export async function fetchHealth(): Promise<boolean> {
   try {
-    // Use the /api/version endpoint as a health check
-    const response = await fetch(`${API_BASE}/api/version`, {
+    const response = await fetch(`${API_BASE}/api/v1/health`, {
       method: "GET",
       headers: {
         "Content-Type": "application/json",
@@ -404,8 +400,7 @@ export async function fetchHealth(): Promise<boolean> {
 
     if (response.ok) {
       const data = await response.json();
-      // If we get a version back, the server is healthy
-      return !!data.version;
+      return data.healthy || false;
     }
 
     return false;

app/ui/app/src/components/Settings.tsx

@@ -299,9 +299,9 @@ export default function Settings() {
                         </Button>
                       </div>
                     </div>
-                    {user?.avatarurl && (
+                    {user?.avatarURL && (
                       <img
-                        src={user.avatarurl}
+                        src={user.avatarURL}
                         alt={user?.name}
                         className="h-10 w-10 rounded-full bg-neutral-200 dark:bg-neutral-700 flex-shrink-0"
                         onError={(e) => {

app/ui/app/src/components/Thinking.tsx

@@ -50,33 +50,21 @@ export default function Thinking({
   // Position content to show bottom when collapsed
   useEffect(() => {
     if (isCollapsed && contentRef.current && wrapperRef.current) {
-      requestAnimationFrame(() => {
-        if (!contentRef.current || !wrapperRef.current) return;
-
-        const contentHeight = contentRef.current.scrollHeight;
-        const wrapperHeight = wrapperRef.current.clientHeight;
-        if (contentHeight > wrapperHeight) {
-          const translateY = -(contentHeight - wrapperHeight);
-          contentRef.current.style.transform = `translateY(${translateY}px)`;
-          setHasOverflow(true);
-        } else {
-          contentRef.current.style.transform = "translateY(0)";
-          setHasOverflow(false);
-        }
-      });
+      const contentHeight = contentRef.current.scrollHeight;
+      const wrapperHeight = wrapperRef.current.clientHeight;
+      if (contentHeight > wrapperHeight) {
+        const translateY = -(contentHeight - wrapperHeight);
+        contentRef.current.style.transform = `translateY(${translateY}px)`;
+        setHasOverflow(true);
+      } else {
+        setHasOverflow(false);
+      }
     } else if (contentRef.current) {
       contentRef.current.style.transform = "translateY(0)";
       setHasOverflow(false);
     }
   }, [thinking, isCollapsed]);
 
-  useEffect(() => {
-    if (activelyThinking && wrapperRef.current && !isCollapsed) {
-      // When expanded and actively thinking, scroll to bottom
-      wrapperRef.current.scrollTop = wrapperRef.current.scrollHeight;
-    }
-  }, [thinking, activelyThinking, isCollapsed]);
-
   const handleToggle = () => {
     setIsCollapsed(!isCollapsed);
     setHasUserInteracted(true);

app/ui/app/src/hooks/useChats.ts

@@ -7,7 +7,6 @@ import { createQueryBatcher } from "./useQueryBatcher";
 import { useRefetchModels } from "./useModels";
 import { useStreamingContext } from "@/contexts/StreamingContext";
 import { useSettings } from "./useSettings";
-import { getModelCapabilities } from "@/api";
 
 export const useChats = () => {
   return useQuery({
@@ -607,24 +606,6 @@ export const useSendMessage = (chatId: string) => {
               queryClient.setQueryData(["staleModels"], newStaleMap);
 
               queryClient.invalidateQueries({ queryKey: ["models"] });
-
-              // Fetch fresh capabilities for the downloaded model
-              getModelCapabilities(selectedModel.model)
-                .then((capabilities) => {
-                  queryClient.setQueryData(
-                    ["modelCapabilities", selectedModel.model],
-                    capabilities,
-                  );
-                })
-                .catch((error) => {
-                  console.error(
-                    "Failed to fetch capabilities after download:",
-                    error,
-                  );
-                  queryClient.invalidateQueries({
-                    queryKey: ["modelCapabilities", selectedModel.model],
-                  });
-                });
             }
             break;
           }

app/ui/app/src/hooks/useDownloadModel.ts

@@ -0,0 +1,114 @@
+import { useMutation, useQueryClient } from "@tanstack/react-query";
+import { useState } from "react";
+import { pullModel } from "@/api";
+import { useSelectedModel } from "./useSelectedModel";
+import { useSettings } from "./useSettings";
+
+interface DownloadProgress {
+  status: string;
+  digest?: string;
+  total?: number;
+  completed?: number;
+  done?: boolean;
+}
+
+export function useDownloadModel(chatId?: string) {
+  const queryClient = useQueryClient();
+  const { selectedModel } = useSelectedModel(chatId);
+  const { setSettings } = useSettings();
+  const [downloadProgress, setDownloadProgress] =
+    useState<DownloadProgress | null>(null);
+  const [abortController, setAbortController] =
+    useState<AbortController | null>(null);
+  const [downloadingChatIds, setDownloadingChatIds] = useState<Set<string>>(
+    new Set(),
+  );
+
+  const mutation = useMutation({
+    mutationFn: async (modelName: string) => {
+      const controller = new AbortController();
+      setAbortController(controller);
+      setDownloadProgress({ status: "Starting download..." });
+      if (chatId) {
+        setDownloadingChatIds((prev) => new Set(prev).add(chatId));
+      }
+
+      try {
+        for await (const progress of pullModel(modelName, controller.signal)) {
+          setDownloadProgress(progress);
+
+          if (progress.status === "success") {
+            // Update selected model to indicate it's now available locally
+            if (selectedModel && selectedModel.model === modelName) {
+              setSettings({ SelectedModel: modelName });
+            }
+            // Invalidate models query to refresh the list
+            await queryClient.invalidateQueries({ queryKey: ["models"] });
+            break;
+          }
+        }
+      } finally {
+        setAbortController(null);
+        if (chatId) {
+          setDownloadingChatIds((prev) => {
+            const newSet = new Set(prev);
+            newSet.delete(chatId);
+            return newSet;
+          });
+        }
+      }
+    },
+    onSuccess: () => {
+      setDownloadProgress(null);
+      if (chatId) {
+        setDownloadingChatIds((prev) => {
+          const newSet = new Set(prev);
+          newSet.delete(chatId);
+          return newSet;
+        });
+      }
+    },
+    onError: (error: Error) => {
+      const status =
+        error.name === "AbortError" ? "Download cancelled" : "Download failed";
+      setDownloadProgress({ status, done: true });
+
+      // Clear error message after delay
+      const delay = error.name === "AbortError" ? 1500 : 3000;
+      setTimeout(() => {
+        setDownloadProgress(null);
+        if (chatId) {
+          setDownloadingChatIds((prev) => {
+            const newSet = new Set(prev);
+            newSet.delete(chatId);
+            return newSet;
+          });
+        }
+      }, delay);
+    },
+  });
+
+  const cancelDownload = () => {
+    if (abortController) {
+      abortController.abort();
+      setAbortController(null);
+      if (chatId) {
+        setDownloadingChatIds((prev) => {
+          const newSet = new Set(prev);
+          newSet.delete(chatId);
+          return newSet;
+        });
+      }
+    }
+  };
+
+  return {
+    downloadModel: mutation.mutate,
+    isDownloading:
+      mutation.isPending && chatId ? downloadingChatIds.has(chatId) : false,
+    downloadProgress:
+      chatId && downloadingChatIds.has(chatId) ? downloadProgress : null,
+    error: mutation.error,
+    cancelDownload,
+  };
+}

app/ui/app/src/hooks/useUser.ts

@@ -1,20 +1,29 @@
 import { useQuery, useMutation, useQueryClient } from "@tanstack/react-query";
+import { useEffect, useState } from "react";
 import { fetchUser, fetchConnectUrl, disconnectUser } from "@/api";
 
 export function useUser() {
   const queryClient = useQueryClient();
+  const [initialDataLoaded, setInitialDataLoaded] = useState(false);
+
+  // Wait for initial data to be loaded
+  useEffect(() => {
+    const initialPromise = window.__initialUserDataPromise;
+    if (initialPromise) {
+      initialPromise.finally(() => {
+        setInitialDataLoaded(true);
+      });
+    } else {
+      setInitialDataLoaded(true);
+    }
+  }, []);
 
   const userQuery = useQuery({
     queryKey: ["user"],
-    queryFn: async () => {
-      const result = await fetchUser();
-      return result;
-    },
+    queryFn: () => fetchUser(),
     staleTime: 5 * 60 * 1000, // Consider data stale after 5 minutes
     gcTime: 10 * 60 * 1000, // Keep in cache for 10 minutes
-    retry: 10,
-    retryDelay: (attemptIndex) => Math.min(500 * attemptIndex, 2000),
-    refetchOnMount: true, // Always fetch when component mounts
+    initialData: null, // Start with null to prevent flashing
   });
 
   // Mutation to refresh user data
@@ -40,15 +49,14 @@ export function useUser() {
     },
   });
 
-  const isLoading = userQuery.isLoading || userQuery.isFetching;
-  const isAuthenticated = Boolean(userQuery.data?.name);
-
   return {
     user: userQuery.data,
-    isLoading,
+    isLoading:
+      !initialDataLoaded ||
+      (userQuery.isLoading && userQuery.data === undefined), // Show loading until initial data is loaded
     isError: userQuery.isError,
     error: userQuery.error,
-    isAuthenticated,
+    isAuthenticated: Boolean(userQuery.data?.name),
     refreshUser: refreshUser.mutate,
     isRefreshing: refreshUser.isPending,
     refetchUser: userQuery.refetch,

app/ui/app/src/lib/config.ts

@@ -8,6 +8,3 @@ export const API_BASE = import.meta.env.DEV ? DEV_API_URL : "";
 export const OLLAMA_HOST = import.meta.env.DEV
   ? DEV_API_URL
   : window.location.origin;
-
-export const OLLAMA_DOT_COM =
-  import.meta.env.VITE_OLLAMA_DOT_COM_URL || "https://ollama.com";

app/ui/app/src/lib/highlighter.ts

@@ -147,7 +147,6 @@ export const highlighterPromise = createHighlighter({
     "c",
     "cpp",
     "sql",
-    "swift",
     "yaml",
     "markdown",
   ],

app/ui/app/src/main.tsx

@@ -5,6 +5,13 @@ import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
 import { routeTree } from "./routeTree.gen";
 import { fetchUser } from "./api";
 import { StreamingProvider } from "./contexts/StreamingContext";
+import { User } from "@/gotypes";
+
+declare global {
+  interface Window {
+    __initialUserDataPromise?: Promise<User | null>;
+  }
+}
 
 const queryClient = new QueryClient({
   defaultOptions: {
@@ -17,11 +24,27 @@ const queryClient = new QueryClient({
   },
 });
 
-fetchUser().then((userData) => {
-  if (userData) {
+// Track initial user data fetch
+let initialUserDataPromise: Promise<User | null> | null = null;
+
+// Initialize user data on app startup
+const initializeUserData = async () => {
+  try {
+    const userData = await fetchUser();
     queryClient.setQueryData(["user"], userData);
+    return userData;
+  } catch (error) {
+    console.error("Error initializing user data:", error);
+    queryClient.setQueryData(["user"], null);
+    return null;
   }
-});
+};
+
+// Start initialization immediately and track the promise
+initialUserDataPromise = initializeUserData();
+
+// Export the promise so hooks can await it
+window.__initialUserDataPromise = initialUserDataPromise;
 
 const router = createRouter({
   routeTree,

app/ui/responses/types.go

@@ -101,14 +101,15 @@ type HealthResponse struct {
 }
 
 type User struct {
-	ID        string `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"`
+	ID            string `json:"id"`
+	Name          string `json:"name"`
+	Email         string `json:"email"`
+	AvatarURL     string `json:"avatarURL"`
+	Plan          string `json:"plan"`
+	Bio           string `json:"bio"`
+	FirstName     string `json:"firstName"`
+	LastName      string `json:"lastName"`
+	OverThreshold bool   `json:"overThreshold"`
 }
 
 type Attachment struct {

app/ui/ui.go

@@ -12,17 +12,18 @@ import (
 	"log/slog"
 	"net/http"
 	"net/http/httputil"
+	"net/url"
 	"os"
 	"runtime"
 	"runtime/debug"
 	"slices"
 	"strconv"
 	"strings"
-	"sync"
 	"time"
 
 	"github.com/google/uuid"
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/app/auth"
 	"github.com/ollama/ollama/app/server"
 	"github.com/ollama/ollama/app/store"
 	"github.com/ollama/ollama/app/tools"
@@ -117,66 +118,40 @@ func (s *Server) log() *slog.Logger {
 
 // ollamaProxy creates a reverse proxy handler to the Ollama server
 func (s *Server) ollamaProxy() http.Handler {
-	var (
-		proxy   http.Handler
-		proxyMu sync.Mutex
-	)
+	ollamaHost := os.Getenv("OLLAMA_HOST")
+	if ollamaHost == "" {
+		ollamaHost = "http://127.0.0.1:11434"
+	}
 
-	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		proxyMu.Lock()
-		p := proxy
-		proxyMu.Unlock()
+	if !strings.HasPrefix(ollamaHost, "http://") && !strings.HasPrefix(ollamaHost, "https://") {
+		ollamaHost = "http://" + ollamaHost
+	}
 
-		if p == nil {
-			proxyMu.Lock()
-			if proxy == nil {
-				var err error
-				for i := range 2 {
-					if i > 0 {
-						s.log().Warn("ollama server not ready, retrying", "attempt", i+1)
-						time.Sleep(1 * time.Second)
-					}
+	target, err := url.Parse(ollamaHost)
+	if err != nil {
+		s.log().Error("failed to parse OLLAMA_HOST", "error", err, "host", ollamaHost)
+		return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			http.Error(w, "failed to configure proxy", http.StatusInternalServerError)
+		})
+	}
 
-					err = WaitForServer(context.Background(), 10*time.Second)
-					if err == nil {
-						break
-					}
-				}
+	s.log().Info("configuring ollama proxy", "target", target.String())
 
-				if err != nil {
-					proxyMu.Unlock()
-					s.log().Error("ollama server not ready after retries", "error", err)
-					http.Error(w, "Ollama server is not ready", http.StatusServiceUnavailable)
-					return
-				}
+	proxy := httputil.NewSingleHostReverseProxy(target)
 
-				target := envconfig.Host()
-				s.log().Info("configuring ollama proxy", "target", target.String())
+	originalDirector := proxy.Director
+	proxy.Director = func(req *http.Request) {
+		originalDirector(req)
+		req.Host = target.Host
+		s.log().Debug("proxying request", "method", req.Method, "path", req.URL.Path, "target", target.Host)
+	}
 
-				newProxy := httputil.NewSingleHostReverseProxy(target)
+	proxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
+		s.log().Error("proxy error", "error", err, "path", r.URL.Path, "target", target.String())
+		http.Error(w, "proxy error: "+err.Error(), http.StatusBadGateway)
+	}
 
-				originalDirector := newProxy.Director
-				newProxy.Director = func(req *http.Request) {
-					originalDirector(req)
-					req.Host = target.Host
-					s.log().Debug("proxying request", "method", req.Method, "path", req.URL.Path, "target", target.Host)
-				}
-
-				newProxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
-					s.log().Error("proxy error", "error", err, "path", r.URL.Path, "target", target.String())
-					http.Error(w, "proxy error: "+err.Error(), http.StatusBadGateway)
-				}
-
-				proxy = newProxy
-				p = newProxy
-			} else {
-				p = proxy
-			}
-			proxyMu.Unlock()
-		}
-
-		p.ServeHTTP(w, r)
-	})
+	return proxy
 }
 
 type errHandlerFunc func(http.ResponseWriter, *http.Request) error
@@ -289,10 +264,11 @@ func (s *Server) Handler() http.Handler {
 	ollamaProxy := s.ollamaProxy()
 	mux.Handle("GET /api/tags", ollamaProxy)
 	mux.Handle("POST /api/show", ollamaProxy)
-	mux.Handle("GET /api/version", ollamaProxy)
-	mux.Handle("HEAD /api/version", ollamaProxy)
-	mux.Handle("POST /api/me", ollamaProxy)
-	mux.Handle("POST /api/signout", ollamaProxy)
+
+	mux.Handle("GET /api/v1/me", handle(s.me))
+	mux.Handle("POST /api/v1/disconnect", handle(s.disconnect))
+	mux.Handle("GET /api/v1/connect", handle(s.connectURL))
+	mux.Handle("GET /api/v1/health", handle(s.health))
 
 	// React app - catch all non-API routes and serve the React app
 	mux.Handle("GET /", s.appHandler())
@@ -362,7 +338,7 @@ func (s *Server) doSelfSigned(ctx context.Context, method, path string) (*http.R
 }
 
 // UserData fetches user data from ollama.com API for the current ollama key
-func (s *Server) UserData(ctx context.Context) (*api.UserResponse, error) {
+func (s *Server) UserData(ctx context.Context) (*responses.User, error) {
 	resp, err := s.doSelfSigned(ctx, http.MethodPost, "/api/me")
 	if err != nil {
 		return nil, fmt.Errorf("failed to call ollama.com/api/me: %w", err)
@@ -373,7 +349,7 @@ func (s *Server) UserData(ctx context.Context) (*api.UserResponse, error) {
 		return nil, fmt.Errorf("unexpected status code: %d", resp.StatusCode)
 	}
 
-	var user api.UserResponse
+	var user responses.User
 	if err := json.NewDecoder(resp.Body).Decode(&user); err != nil {
 		return nil, fmt.Errorf("failed to parse user response: %w", err)
 	}
@@ -392,27 +368,29 @@ func (s *Server) UserData(ctx context.Context) (*api.UserResponse, error) {
 	return &user, nil
 }
 
-// WaitForServer waits for the Ollama server to be ready
-func WaitForServer(ctx context.Context, timeout time.Duration) error {
-	deadline := time.Now().Add(timeout)
-	for time.Now().Before(deadline) {
+func waitForServer(ctx context.Context) error {
+	timeout := time.Now().Add(10 * time.Second)
+	// TODO: this avoids an error on first load of the app
+	// however we should either show a loading state or
+	// wait for the Ollama server to be ready before redirecting
+	for {
 		c, err := api.ClientFromEnvironment()
 		if err != nil {
 			return err
 		}
 		if _, err := c.Version(ctx); err == nil {
-			slog.Debug("ollama server is ready")
-			return nil
+			break
+		}
+		if time.Now().After(timeout) {
+			return fmt.Errorf("timeout waiting for Ollama server to be ready")
 		}
 		time.Sleep(10 * time.Millisecond)
 	}
-	return errors.New("timeout waiting for Ollama server to be ready")
+	return nil
 }
 
 func (s *Server) createChat(w http.ResponseWriter, r *http.Request) error {
-	if err := WaitForServer(r.Context(), 10*time.Second); err != nil {
-		return err
-	}
+	waitForServer(r.Context())
 
 	id, err := uuid.NewV7()
 	if err != nil {
@@ -997,7 +975,7 @@ func (s *Server) chat(w http.ResponseWriter, r *http.Request) error {
 				for _, toolCall := range res.Message.ToolCalls {
 					// continues loop as tools were executed
 					toolsExecuted = true
-					result, content, err := registry.Execute(ctx, toolCall.Function.Name, toolCall.Function.Arguments.ToMap())
+					result, content, err := registry.Execute(ctx, toolCall.Function.Name, toolCall.Function.Arguments)
 					if err != nil {
 						errContent := fmt.Sprintf("Error: %v", err)
 						toolErrMsg := store.NewMessage("tool", errContent, nil)
@@ -1460,6 +1438,129 @@ func (s *Server) settings(w http.ResponseWriter, r *http.Request) error {
 	})
 }
 
+func (s *Server) me(w http.ResponseWriter, r *http.Request) error {
+	if r.Method != http.MethodGet {
+		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
+		return nil
+	}
+
+	user, err := s.UserData(r.Context())
+	if err != nil {
+		// If fetching from API fails, try to return cached user data if available
+		if cachedUser, cacheErr := s.Store.User(); cacheErr == nil && cachedUser != nil {
+			s.log().Info("API request failed, returning cached user data", "error", err)
+			responseUser := &responses.User{
+				Name:  cachedUser.Name,
+				Email: cachedUser.Email,
+				Plan:  cachedUser.Plan,
+			}
+			w.Header().Set("Content-Type", "application/json")
+			w.WriteHeader(http.StatusOK)
+			return json.NewEncoder(w).Encode(responseUser)
+		}
+
+		s.log().Error("failed to get user data", "error", err)
+		w.WriteHeader(http.StatusInternalServerError)
+		return json.NewEncoder(w).Encode(responses.Error{
+			Error: "failed to get user data",
+		})
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	return json.NewEncoder(w).Encode(user)
+}
+
+func (s *Server) disconnect(w http.ResponseWriter, r *http.Request) error {
+	if r.Method != http.MethodPost {
+		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
+		return nil
+	}
+
+	if err := s.Store.ClearUser(); err != nil {
+		s.log().Warn("failed to clear cached user data", "error", err)
+	}
+
+	// Get the SSH public key to encode for the delete request
+	pubKey, err := ollamaAuth.GetPublicKey()
+	if err != nil {
+		s.log().Error("failed to get public key", "error", err)
+		w.WriteHeader(http.StatusInternalServerError)
+		return json.NewEncoder(w).Encode(responses.Error{
+			Error: "failed to get public key",
+		})
+	}
+
+	// Encode the key using base64 URL encoding
+	encodedKey := base64.RawURLEncoding.EncodeToString([]byte(pubKey))
+
+	// Call the /api/user/keys/{encodedKey} endpoint with DELETE
+	resp, err := s.doSelfSigned(r.Context(), http.MethodDelete, fmt.Sprintf("/api/user/keys/%s", encodedKey))
+	if err != nil {
+		s.log().Error("failed to call ollama.com/api/user/keys", "error", err)
+		w.WriteHeader(http.StatusInternalServerError)
+		return json.NewEncoder(w).Encode(responses.Error{
+			Error: "failed to disconnect from ollama.com",
+		})
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		s.log().Error("disconnect request failed", "status", resp.StatusCode)
+		w.WriteHeader(http.StatusInternalServerError)
+		return json.NewEncoder(w).Encode(responses.Error{
+			Error: "failed to disconnect from ollama.com",
+		})
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	return json.NewEncoder(w).Encode(map[string]string{"status": "disconnected"})
+}
+
+func (s *Server) connectURL(w http.ResponseWriter, r *http.Request) error {
+	if r.Method != http.MethodGet {
+		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
+		return nil
+	}
+
+	connectURL, err := auth.BuildConnectURL(OllamaDotCom)
+	if err != nil {
+		s.log().Error("failed to build connect URL", "error", err)
+		w.WriteHeader(http.StatusInternalServerError)
+		return json.NewEncoder(w).Encode(responses.Error{
+			Error: "failed to build connect URL",
+		})
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	return json.NewEncoder(w).Encode(map[string]string{
+		"connect_url": connectURL,
+	})
+}
+
+func (s *Server) health(w http.ResponseWriter, r *http.Request) error {
+	if r.Method != http.MethodGet {
+		http.Error(w, "Method Not Allowed", http.StatusMethodNotAllowed)
+		return nil
+	}
+
+	healthy := false
+	c, err := api.ClientFromEnvironment()
+	if err == nil {
+		if _, err := c.Version(r.Context()); err == nil {
+			healthy = true
+		}
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	return json.NewEncoder(w).Encode(responses.HealthResponse{
+		Healthy: healthy,
+	})
+}
+
 func (s *Server) getInferenceCompute(w http.ResponseWriter, r *http.Request) error {
 	ctx, cancel := context.WithTimeout(r.Context(), 500*time.Millisecond)
 	defer cancel()
@@ -1558,13 +1659,13 @@ func convertToOllamaTool(toolSchema map[string]any) api.Tool {
 
 	tool.Function.Parameters.Type = "object"
 	tool.Function.Parameters.Required = []string{}
-	tool.Function.Parameters.Properties = api.NewToolPropertiesMap()
+	tool.Function.Parameters.Properties = make(map[string]api.ToolProperty)
 
 	if schemaProps, ok := toolSchema["schema"].(map[string]any); ok {
 		tool.Function.Parameters.Type = getStringFromMap(schemaProps, "type", "object")
 
 		if props, ok := schemaProps["properties"].(map[string]any); ok {
-			tool.Function.Parameters.Properties = api.NewToolPropertiesMap()
+			tool.Function.Parameters.Properties = make(map[string]api.ToolProperty)
 
 			for propName, propDef := range props {
 				if propMap, ok := propDef.(map[string]any); ok {
@@ -1572,7 +1673,7 @@ func convertToOllamaTool(toolSchema map[string]any) api.Tool {
 						Type:        api.PropertyType{getStringFromMap(propMap, "type", "string")},
 						Description: getStringFromMap(propMap, "description", ""),
 					}
-					tool.Function.Parameters.Properties.Set(propName, prop)
+					tool.Function.Parameters.Properties[propName] = prop
 				}
 			}
 		}

app/wintray/eventloop.go

@@ -158,16 +158,16 @@ func (t *winTray) wndProc(hWnd windows.Handle, message uint32, wParam, lParam ui
 	case uint32(UI_REQUEST_MSG_ID):
 		// Requests for the UI must always come from the main event thread
 		l := int(wParam)
-		path := unsafe.String((*byte)(unsafe.Pointer(lParam)), l) //nolint:govet,gosec
+		path := unsafe.String((*byte)(unsafe.Pointer(lParam)), l)
 		t.app.UIRun(path)
 	case WM_COPYDATA:
 		// Handle URL scheme requests from other instances
 		if lParam != 0 {
-			cds := (*COPYDATASTRUCT)(unsafe.Pointer(lParam)) //nolint:govet,gosec
-			if cds.DwData == 1 {                             // Our identifier for URL scheme messages
+			cds := (*COPYDATASTRUCT)(unsafe.Pointer(lParam))
+			if cds.DwData == 1 { // Our identifier for URL scheme messages
 				// Convert the data back to string
 				data := make([]byte, cds.CbData)
-				copy(data, (*[1 << 30]byte)(unsafe.Pointer(cds.LpData))[:cds.CbData:cds.CbData]) //nolint:govet,gosec
+				copy(data, (*[1 << 30]byte)(unsafe.Pointer(cds.LpData))[:cds.CbData:cds.CbData])
 				urlScheme := string(data)
 				handleURLSchemeRequest(urlScheme)
 				lResult = 1 // Return non-zero to indicate success

cmd/bench/README.md

@@ -15,7 +15,7 @@ A Go-based command-line tool for benchmarking Ollama models with configurable pa
 
 ```
 go build -o ollama-bench bench.go
-./ollama-bench -model gpt-oss:20b -epochs 6 -format csv
+./bench -model gpt-oss:20b -epochs 6 -format csv
 ```
 
 Using Go Run (without building)
@@ -29,32 +29,31 @@ go run bench.go -model gpt-oss:20b -epochs 3
 ### Basic Example
 
 ```
-./ollama-bench -model gemma3 -epochs 6
+./bench -model gemma3 -epochs 6
 ```
 
 ### Benchmark Multiple Models
 
 ```
-./ollama-bench -model gemma3,gemma3n -epochs 6 -max-tokens 100 -p "Write me a short story" | tee gemma.bench
+./bench -model gemma3,gemma3n -epochs 6 -max-tokens 100 -p "Write me a short story" | tee gemma.bench
 benchstat -col /name gemma.bench
 ```
 
 ### With Image Prompt
 
 ```
-./ollama-bench -model qwen3-vl -image photo.jpg -epochs 6 -max-tokens 100 -p "Describe this image"
+./bench -model qwen3-vl -image photo.jpg -epochs 6 -max-tokens 100 -p "Describe this image"
 ```
 
 ### Advanced Example
 
 ```
-./ollama-bench -model llama3 -epochs 10 -temperature 0.7 -max-tokens 500 -seed 42 -format csv -output results.csv
+./bench -model llama3 -epochs 10 -temperature 0.7 -max-tokens 500 -seed 42 -format csv -output results.csv
 ```
 
 ## Command Line Options
 
 | Option  	| Description | Default |
-|----------|-------------|---------|
 | -model	| Comma-separated list of models to benchmark	| (required)		|
 | -epochs	| Number of iterations per model		| 1			|
 | -max-tokens	| Maximum tokens for model response		| 0 (unlimited)		|

cmd/bench/bench.go

@@ -48,8 +48,8 @@ func OutputMetrics(w io.Writer, format string, metrics []Metrics, verbose bool)
 	case "benchstat":
 		if verbose {
 			printHeader := func() {
-				fmt.Fprintf(w, "sysname: %s\n", runtime.GOOS)
-				fmt.Fprintf(w, "machine: %s\n", runtime.GOARCH)
+				fmt.Printf("sysname: %s\n", runtime.GOOS)
+				fmt.Printf("machine: %s\n", runtime.GOARCH)
 			}
 			once.Do(printHeader)
 		}
@@ -147,17 +147,6 @@ func BenchmarkChat(fOpt flagOptions) error {
 		return err
 	}
 
-	var out io.Writer = os.Stdout
-	if fOpt.outputFile != nil && *fOpt.outputFile != "" {
-		f, err := os.OpenFile(*fOpt.outputFile, os.O_CREATE|os.O_WRONLY, 0o644)
-		if err != nil {
-			fmt.Fprintf(os.Stderr, "ERROR: cannot open output file %s: %v\n", *fOpt.outputFile, err)
-			return err
-		}
-		defer f.Close()
-		out = f
-	}
-
 	for _, model := range models {
 		for range *fOpt.epochs {
 			options := make(map[string]interface{})
@@ -252,14 +241,13 @@ func BenchmarkChat(fOpt flagOptions) error {
 				},
 			}
 
-			OutputMetrics(out, *fOpt.format, metrics, *fOpt.verbose)
+			OutputMetrics(os.Stdout, *fOpt.format, metrics, *fOpt.verbose)
 
 			if *fOpt.keepAlive > 0 {
 				time.Sleep(time.Duration(*fOpt.keepAlive*float64(time.Second)) + 200*time.Millisecond)
 			}
 		}
 	}
-
 	return nil
 }
 

cmd/cmd.go

@@ -45,7 +45,6 @@ import (
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
-	xcmd "github.com/ollama/ollama/x/cmd"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -518,10 +517,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return generateEmbedding(cmd, name, opts.Prompt, opts.KeepAlive, truncate, dimensions)
 	}
 
-	// Check for experimental flag
-	isExperimental, _ := cmd.Flags().GetBool("experimental")
-	yoloMode, _ := cmd.Flags().GetBool("yolo")
-
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
 			var sErr api.AuthorizationError
@@ -548,11 +543,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 			}
 		}
 
-		// Use experimental agent loop with tools
-		if isExperimental {
-			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode)
-		}
-
 		return generateInteractive(cmd, opts)
 	}
 	return generate(cmd, opts)
@@ -953,9 +943,6 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
 		}
 		rows = append(rows, []string{"", "quantization", resp.Details.QuantizationLevel})
-		if resp.Requires != "" {
-			rows = append(rows, []string{"", "requires", resp.Requires})
-		}
 		return
 	})
 
@@ -1443,7 +1430,7 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		latest.Summary()
 	}
 
-	return &api.Message{Role: role, Thinking: thinkingContent.String(), Content: fullResponse.String()}, nil
+	return &api.Message{Role: role, Content: fullResponse.String()}, nil
 }
 
 func generate(cmd *cobra.Command, opts runOptions) error {
@@ -1764,8 +1751,6 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("hidethinking", false, "Hide thinking output (if provided)")
 	runCmd.Flags().Bool("truncate", false, "For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead")
 	runCmd.Flags().Int("dimensions", 0, "Truncate output embeddings to specified dimension (embedding models only)")
-	runCmd.Flags().Bool("experimental", false, "Enable experimental agent loop with tools")
-	runCmd.Flags().BoolP("yolo", "y", false, "Skip all tool approval prompts (use with caution)")
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",

cmd/cmd_test.go

@@ -291,31 +291,6 @@ Weigh anchor!
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
 		}
 	})
-
-	t.Run("min version", func(t *testing.T) {
-		var b bytes.Buffer
-		if err := showInfo(&api.ShowResponse{
-			Details: api.ModelDetails{
-				Family:            "test",
-				ParameterSize:     "7B",
-				QuantizationLevel: "FP16",
-			},
-			Requires: "0.14.0",
-		}, false, &b); err != nil {
-			t.Fatal(err)
-		}
-
-		expect := `  Model
-    architecture    test      
-    parameters      7B        
-    quantization    FP16      
-    requires        0.14.0    
-
-`
-		if diff := cmp.Diff(expect, b.String()); diff != "" {
-			t.Errorf("unexpected output (-want +got):\n%s", diff)
-		}
-	})
 }
 
 func TestDeleteHandler(t *testing.T) {

cmd/eval/README.md

@@ -0,0 +1,50 @@
+# eval
+
+Evaluation tool for testing Ollama models.
+
+## Usage
+
+Run all tests:
+
+```bash
+go run . -model llama3.2:latest
+```
+
+Run specific suite:
+
+```bash
+go run . -model llama3.2:latest -suite tool-calling-basic -v
+```
+
+List available suites:
+
+```bash
+go run . -list
+```
+
+## Adding Tests
+
+Edit `suites.go` to add new test suites. Each test needs:
+
+- `Name`: test identifier
+- `Prompt`: what to send to the model
+- `Check`: function to validate the response
+
+Example:
+
+```go
+{
+    Name:   "my-test",
+    Prompt: "What is 2+2?",
+    Check:  Contains("4"),
+}
+```
+
+Available check functions:
+
+- `HasResponse()` - response is non-empty
+- `Contains(s)` - response contains substring
+- `CallsTool(name)` - model called specific tool
+- `NoTools()` - model called no tools
+- `MinTools(n)` - model called at least n tools
+- `All(checks...)` - all checks pass

cmd/eval/eval.go

@@ -0,0 +1,151 @@
+package main
+
+import (
+	"context"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Test is a single evaluation test
+type Test struct {
+	Name    string
+	Prompt  string
+	System  string
+	Tools   []api.Tool
+	Think   bool
+	Options map[string]any
+	Check   func(response string, tools []api.ToolCall) bool
+}
+
+// Suite is a collection of tests
+type Suite struct {
+	Name  string
+	Tests []Test
+}
+
+// Result holds test execution results
+type Result struct {
+	Name      string
+	Passed    bool
+	Error     error
+	Duration  time.Duration
+	Response  string
+	Tools     []string
+	ToolCalls []api.ToolCall
+	Thinking  bool
+}
+
+// Run executes a test against a model
+func Run(ctx context.Context, client *api.Client, model string, test Test) Result {
+	result := Result{Name: test.Name}
+
+	req := &api.ChatRequest{
+		Model: model,
+		Messages: []api.Message{
+			{Role: "user", Content: test.Prompt},
+		},
+		Options: test.Options,
+	}
+
+	if test.System != "" {
+		req.Messages = append([]api.Message{
+			{Role: "system", Content: test.System},
+		}, req.Messages...)
+	}
+
+	if len(test.Tools) > 0 {
+		req.Tools = test.Tools
+	}
+
+	if test.Think {
+		req.Think = &api.ThinkValue{Value: true}
+	}
+
+	var resp strings.Builder
+	var toolCalls []api.ToolCall
+
+	start := time.Now()
+	err := client.Chat(ctx, req, func(r api.ChatResponse) error {
+		resp.WriteString(r.Message.Content)
+		if r.Message.Thinking != "" {
+			result.Thinking = true
+		}
+		toolCalls = append(toolCalls, r.Message.ToolCalls...)
+		return nil
+	})
+	result.Duration = time.Since(start)
+
+	if err != nil {
+		result.Error = err
+		return result
+	}
+
+	result.Response = resp.String()
+	result.Tools = uniqueToolNames(toolCalls)
+	result.ToolCalls = toolCalls
+	result.Passed = test.Check(result.Response, toolCalls)
+
+	return result
+}
+
+func uniqueToolNames(calls []api.ToolCall) []string {
+	seen := make(map[string]bool)
+	var names []string
+	for _, c := range calls {
+		if !seen[c.Function.Name] {
+			seen[c.Function.Name] = true
+			names = append(names, c.Function.Name)
+		}
+	}
+	return names
+}
+
+// Check functions for common test patterns
+
+func HasResponse() func(string, []api.ToolCall) bool {
+	return func(resp string, _ []api.ToolCall) bool {
+		return strings.TrimSpace(resp) != ""
+	}
+}
+
+func Contains(s string) func(string, []api.ToolCall) bool {
+	return func(resp string, _ []api.ToolCall) bool {
+		return strings.Contains(strings.ToLower(resp), strings.ToLower(s))
+	}
+}
+
+func CallsTool(name string) func(string, []api.ToolCall) bool {
+	return func(_ string, tools []api.ToolCall) bool {
+		for _, t := range tools {
+			if t.Function.Name == name {
+				return true
+			}
+		}
+		return false
+	}
+}
+
+func NoTools() func(string, []api.ToolCall) bool {
+	return func(_ string, tools []api.ToolCall) bool {
+		return len(tools) == 0
+	}
+}
+
+func MinTools(n int) func(string, []api.ToolCall) bool {
+	return func(_ string, tools []api.ToolCall) bool {
+		return len(tools) >= n
+	}
+}
+
+func All(checks ...func(string, []api.ToolCall) bool) func(string, []api.ToolCall) bool {
+	return func(resp string, tools []api.ToolCall) bool {
+		for _, check := range checks {
+			if !check(resp, tools) {
+				return false
+			}
+		}
+		return true
+	}
+}

cmd/eval/main.go

@@ -0,0 +1,217 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	model := flag.String("model", "", "model to evaluate")
+	suite := flag.String("suite", "", "comma-separated list of suites to run (empty runs all)")
+	list := flag.Bool("list", false, "list available suites")
+	verbose := flag.Bool("v", false, "verbose output")
+	timeout := flag.Int("timeout", 60, "timeout per test in seconds")
+	export := flag.String("export", "eval-results.json", "export results to file")
+	flag.Parse()
+
+	if *list {
+		for _, s := range suites {
+			fmt.Printf("%s (%d tests)\n", s.Name, len(s.Tests))
+		}
+		return
+	}
+
+	if *model == "" {
+		fmt.Fprintf(os.Stderr, "error: -model parameter is required\n")
+		os.Exit(1)
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "error: %v\n", err)
+		os.Exit(1)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	if err := client.Heartbeat(ctx); err != nil {
+		cancel()
+		fmt.Fprintf(os.Stderr, "error: cannot connect to ollama\n")
+		os.Exit(1)
+	}
+	cancel()
+
+	selected := suites
+	if *suite != "" {
+		suiteNames := strings.Split(*suite, ",")
+		selected = []Suite{}
+		var notFound []string
+
+		for _, name := range suiteNames {
+			name = strings.TrimSpace(name)
+			if name == "" {
+				continue
+			}
+
+			found := false
+			for _, s := range suites {
+				if s.Name == name {
+					selected = append(selected, s)
+					found = true
+					break
+				}
+			}
+			if !found {
+				notFound = append(notFound, name)
+			}
+		}
+
+		if len(notFound) > 0 {
+			fmt.Fprintf(os.Stderr, "error: suite(s) not found: %s\n", strings.Join(notFound, ", "))
+			os.Exit(1)
+		}
+	}
+
+	var results []Result
+	for _, s := range selected {
+		if *verbose {
+			fmt.Printf("\n%s (%d tests)\n", s.Name, len(s.Tests))
+		}
+		for i, test := range s.Tests {
+			if test.Options == nil {
+				test.Options = map[string]any{"temperature": 0.1}
+			}
+			if test.Check == nil {
+				test.Check = HasResponse()
+			}
+
+			if *verbose {
+				fmt.Printf("  [%d/%d] %s... ", i+1, len(s.Tests), test.Name)
+			}
+
+			ctx, cancel := context.WithTimeout(context.Background(), time.Duration(*timeout)*time.Second)
+			result := Run(ctx, client, *model, test)
+			cancel()
+
+			results = append(results, result)
+
+			if *verbose {
+				if result.Error != nil {
+					fmt.Printf("ERROR: %v\n", result.Error)
+				} else if result.Passed {
+					fmt.Printf("PASS (%.2fs)", result.Duration.Seconds())
+					if len(result.Tools) > 0 || result.Thinking {
+						fmt.Printf(" [")
+						if len(result.Tools) > 0 {
+							fmt.Printf("tools: %s", strings.Join(result.Tools, ","))
+						}
+						if result.Thinking {
+							if len(result.Tools) > 0 {
+								fmt.Printf(", ")
+							}
+							fmt.Printf("thinking")
+						}
+						fmt.Printf("]")
+					}
+					fmt.Println()
+
+					// Print tool calls with details
+					if len(result.ToolCalls) > 0 {
+						fmt.Printf("    Tool Calls:\n")
+						for _, tc := range result.ToolCalls {
+							argsJSON, _ := json.Marshal(tc.Function.Arguments)
+							fmt.Printf("      - %s: %s\n", tc.Function.Name, string(argsJSON))
+						}
+					}
+
+					// Print response if there is one
+					if result.Response != "" {
+						fmt.Printf("    Response: %s\n", result.Response)
+					}
+				} else {
+					fmt.Printf("FAIL (%.2fs)\n", result.Duration.Seconds())
+
+					// Print tool calls with details even on failure
+					if len(result.ToolCalls) > 0 {
+						fmt.Printf("    Tool Calls:\n")
+						for _, tc := range result.ToolCalls {
+							argsJSON, _ := json.Marshal(tc.Function.Arguments)
+							fmt.Printf("      - %s: %s\n", tc.Function.Name, string(argsJSON))
+						}
+					}
+
+					// Print response even on failure
+					if result.Response != "" {
+						fmt.Printf("    Response: %s\n", result.Response)
+					}
+				}
+			}
+		}
+	}
+
+	printSummary(results)
+
+	if *export != "" {
+		if err := writeJSON(*export, results); err != nil {
+			fmt.Fprintf(os.Stderr, "warning: export failed: %v\n", err)
+		} else if *verbose {
+			fmt.Printf("\nResults: %s\n", *export)
+		}
+	}
+
+	if anyFailed(results) {
+		os.Exit(1)
+	}
+}
+
+func printSummary(results []Result) {
+	var passed, failed, errors int
+	for _, r := range results {
+		if r.Error != nil {
+			errors++
+		} else if r.Passed {
+			passed++
+		} else {
+			failed++
+		}
+	}
+
+	total := len(results)
+	rate := 0.0
+	if total > 0 {
+		rate = float64(passed) / float64(total) * 100
+	}
+
+	fmt.Printf("\n%d/%d passed (%.1f%%)", passed, total, rate)
+	if errors > 0 {
+		fmt.Printf(", %d errors", errors)
+	}
+	fmt.Println()
+}
+
+func anyFailed(results []Result) bool {
+	for _, r := range results {
+		if !r.Passed || r.Error != nil {
+			return true
+		}
+	}
+	return false
+}
+
+func writeJSON(path string, results []Result) error {
+	f, err := os.Create(path)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+
+	enc := json.NewEncoder(f)
+	enc.SetIndent("", "  ")
+	return enc.Encode(results)
+}

cmd/eval/suites.go

@@ -0,0 +1,178 @@
+package main
+
+import "github.com/ollama/ollama/api"
+
+var suites = []Suite{
+	{
+		Name: "basic-qa",
+		Tests: []Test{
+			{
+				Name:   "simple-math",
+				Prompt: "What is 2+2? Reply with just the number.",
+				Check:  Contains("4"),
+			},
+			{
+				Name:   "capital-city",
+				Prompt: "What is the capital of France? Reply with just the city name.",
+				Check:  Contains("Paris"),
+			},
+			{
+				Name:   "greeting",
+				Prompt: "Say hello",
+				Check:  HasResponse(),
+			},
+		},
+	},
+	{
+		Name: "reasoning",
+		Tests: []Test{
+			{
+				Name:   "logic-puzzle",
+				Prompt: "If all roses are flowers and some flowers fade quickly, can we conclude that some roses fade quickly? Answer yes or no.",
+				Check:  Contains("no"),
+			},
+			{
+				Name:   "counting",
+				Prompt: "How many letters are in the word 'HELLO'?",
+				Check:  Contains("5"),
+			},
+		},
+	},
+	{
+		Name: "instruction-following",
+		Tests: []Test{
+			{
+				Name:   "json-output",
+				Prompt: "Reply with a JSON object containing a 'status' field set to 'ok'.",
+				Check:  All(Contains("status"), Contains("ok")),
+			},
+			{
+				Name:   "system-prompt",
+				Prompt: "What is your name?",
+				System: "You are a helpful assistant named TestBot. When asked your name, always respond with 'TestBot'.",
+				Check:  Contains("TestBot"),
+			},
+		},
+	},
+	{
+		Name: "tool-calling-basic",
+		Tests: []Test{
+			{
+				Name:   "single-tool",
+				Prompt: "What's the weather like in San Francisco?",
+				Tools:  []api.Tool{weatherTool},
+				Check:  CallsTool("get_weather"),
+			},
+			{
+				Name:   "tool-selection",
+				Prompt: "What time is it in Tokyo?",
+				Tools:  []api.Tool{weatherTool, timeTool},
+				Check:  CallsTool("get_time"),
+			},
+			{
+				Name:   "no-tool-needed",
+				Prompt: "What is 2+2?",
+				Tools:  []api.Tool{weatherTool, timeTool},
+				Check:  NoTools(),
+			},
+		},
+	},
+	{
+		Name: "tool-calling-advanced",
+		Tests: []Test{
+			{
+				Name:   "parallel-calls",
+				Prompt: "Get the weather in both New York and Los Angeles.",
+				Tools:  []api.Tool{weatherTool},
+				Check:  All(CallsTool("get_weather"), MinTools(2)),
+			},
+			{
+				Name:   "multi-param",
+				Prompt: "Search for Italian restaurants with prices between $20 and $40.",
+				Tools:  []api.Tool{restaurantTool},
+				Check:  CallsTool("search_restaurants"),
+			},
+		},
+	},
+	{
+		Name: "tool-calling-thinking",
+		Tests: []Test{
+			{
+				Name:   "thinking-before-tool",
+				Prompt: "I need to know the weather in Paris before I decide what to pack.",
+				Tools:  []api.Tool{weatherTool},
+				Think:  true,
+				Check:  CallsTool("get_weather"),
+			},
+			{
+				Name:   "thinking-multi-tool",
+				Prompt: "I'm planning a trip to London. I need to know what time it is there and what the weather is like.",
+				Tools:  []api.Tool{weatherTool, timeTool},
+				Think:  true,
+				Check:  MinTools(1),
+			},
+		},
+	},
+}
+
+var weatherTool = api.Tool{
+	Type: "function",
+	Function: api.ToolFunction{
+		Name:        "get_weather",
+		Description: "Get the current weather in a given location",
+		Parameters: api.ToolFunctionParameters{
+			Type:     "object",
+			Required: []string{"location"},
+			Properties: map[string]api.ToolProperty{
+				"location": {
+					Type:        api.PropertyType{"string"},
+					Description: "The city and state",
+				},
+			},
+		},
+	},
+}
+
+var timeTool = api.Tool{
+	Type: "function",
+	Function: api.ToolFunction{
+		Name:        "get_time",
+		Description: "Get the current time in a timezone",
+		Parameters: api.ToolFunctionParameters{
+			Type:     "object",
+			Required: []string{"timezone"},
+			Properties: map[string]api.ToolProperty{
+				"timezone": {
+					Type:        api.PropertyType{"string"},
+					Description: "The timezone name",
+				},
+			},
+		},
+	},
+}
+
+var restaurantTool = api.Tool{
+	Type: "function",
+	Function: api.ToolFunction{
+		Name:        "search_restaurants",
+		Description: "Search for restaurants",
+		Parameters: api.ToolFunctionParameters{
+			Type:     "object",
+			Required: []string{"cuisine"},
+			Properties: map[string]api.ToolProperty{
+				"cuisine": {
+					Type:        api.PropertyType{"string"},
+					Description: "Type of cuisine",
+				},
+				"min_price": {
+					Type:        api.PropertyType{"number"},
+					Description: "Minimum price",
+				},
+				"max_price": {
+					Type:        api.PropertyType{"number"},
+					Description: "Maximum price",
+				},
+			},
+		},
+	},
+}

cmd/interactive.go

@@ -40,7 +40,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /bye            Exit")
 		fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 		fmt.Fprintln(os.Stderr, "  /? shortcuts    Help for keyboard shortcuts")
-
 		fmt.Fprintln(os.Stderr, "")
 		fmt.Fprintln(os.Stderr, "Use \"\"\" to begin a multi-line message.")
 

convert/convert.go

@@ -6,14 +6,11 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
-	"iter"
 	"log/slog"
-	"maps"
 	"os"
 	"slices"
 	"strings"
 
-	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -21,28 +18,9 @@ type ModelParameters struct {
 	Architectures []string `json:"architectures"`
 	VocabSize     uint32   `json:"vocab_size"`
 
-	// TODO is this needed?
-	ModelType string `json:"model_type"`
-
 	TextModel struct {
-		VocabSize  uint32 `json:"vocab_size"`
-		HiddenSize uint32 `json:"hidden_size"`
-		ModelType  string `json:"model_type"`
+		VocabSize uint32 `json:"vocab_size"`
 	} `json:"text_config"`
-
-	// TODO vision config
-	/*
-		"vision_config": {
-			"hidden_size": 1152,
-			"image_size": 896,
-			"intermediate_size": 4304,
-			"model_type": "siglip_vision_model",
-			"num_attention_heads": 16,
-			"num_hidden_layers": 27,
-			"patch_size": 14,
-			"vision_use_head": false
-		}
-	*/
 }
 
 type AdapterParameters struct {
@@ -55,91 +33,8 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-type KV map[string]any
-
-func (kv KV) Architecture() string {
-	return kv.String("general.architecture", "unknown")
-}
-
-type valueTypes interface {
-	uint8 | int8 | uint16 | int16 |
-		uint32 | int32 | uint64 | int64 |
-		string | float32 | float64 | bool
-}
-
-type arrayValueTypes interface {
-	[]uint8 | []int8 | []uint16 | []int16 |
-		[]uint32 | []int32 | []uint64 | []int64 |
-		[]string | []float32 | []float64 | []bool
-}
-
-func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) (T, bool) {
-	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
-		key = kv.Architecture() + "." + key
-	}
-
-	if val, ok := kv[key].(T); ok {
-		return val, true
-	}
-	return defaultValue[0], false
-}
-
-func (kv KV) String(key string, defaultValue ...string) string {
-	val, _ := keyValue(kv, key, append(defaultValue, "")...)
-	return val
-}
-func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
-	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
-	return val
-}
-
-func (kv KV) Float(key string, defaultValue ...float32) float32 {
-	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
-	return val
-}
-
-func (kv KV) Bool(key string, defaultValue ...bool) bool {
-	val, _ := keyValue(kv, key, append(defaultValue, false)...)
-	return val
-}
-
-func (kv KV) Strings(key string, defaultValue ...[]string) []string {
-	val, _ := keyValue(kv, key, append(defaultValue, []string{""})...)
-	return val
-}
-func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
-	val, _ := keyValue(kv, key, append(defaultValue, []int32{0})...)
-	return val
-}
-
-func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
-	val, _ := keyValue(kv, key, append(defaultValue, []uint32{0})...)
-	return val
-}
-
-func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
-	val, _ := keyValue(kv, key, append(defaultValue, []float32{0})...)
-	return val
-}
-
-func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
-	val, _ := keyValue(kv, key, append(defaultValue, []bool{false})...)
-	return val
-}
-func (kv KV) Len() int {
-	return len(kv)
-}
-
-func (kv KV) Keys() iter.Seq[string] {
-	return maps.Keys(kv)
-}
-
-func (kv KV) Value(key string) any {
-	return kv[key]
-}
-
-func (ModelParameters) KV(t *Tokenizer) KV {
-	kv := KV{
+func (ModelParameters) KV(t *Tokenizer) ggml.KV {
+	kv := ggml.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -168,7 +63,7 @@ func (ModelParameters) KV(t *Tokenizer) KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() KV {
+func (p AdapterParameters) KV() ggml.KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -176,7 +71,7 @@ func (p AdapterParameters) KV() KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := KV{
+	kv := ggml.KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -193,14 +88,9 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-type ModelKV interface {
-	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) KV
-}
-
 type ModelConverter interface {
-	ModelKV
-
+	// KV maps parameters to LLM key-values
+	KV(*Tokenizer) ggml.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -217,7 +107,7 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(ofs.Config) KV
+	KV(ggml.KV) ggml.KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -225,7 +115,7 @@ type AdapterConverter interface {
 	Replacements() []string
 }
 
-func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
+func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -236,8 +126,8 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
 		return err
 	}
 
-	arch := baseKV.Architecture()
-	if arch == "" {
+	arch, ok := baseKV["general.architecture"]
+	if !ok {
 		return errors.New("architecture not set for the base model")
 	}
 
@@ -263,19 +153,23 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
 	return writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
 }
 
-func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
+// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
+// and files it finds in the input path.
+// Supported input model formats include safetensors.
+// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
+func ConvertModel(fsys fs.FS, f *os.File) error {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
-		return nil, nil, err
+		return err
 	}
 
 	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
-		return nil, nil, err
+		return err
 	}
 
 	if len(p.Architectures) < 1 {
-		return nil, nil, errors.New("unknown architecture")
+		return errors.New("unknown architecture")
 	}
 
 	var conv ModelConverter
@@ -288,8 +182,6 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &llama4Model{}
 	case "Mistral3ForConditionalGeneration":
 		conv = &mistral3Model{}
-	case "Ministral3ForCausalLM":
-		conv = &mistral3CausalModel{}
 	case "MixtralForCausalLM":
 		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
@@ -308,37 +200,31 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 		conv = &qwen25VLModel{}
 	case "Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration":
 		conv = &qwen3VLModel{}
-	case "Olmo3ForCausalLM":
-		conv = &olmoModel{}
 	case "BertModel":
 		conv = &bertModel{}
-	case "NomicBertModel", "NomicBertMoEModel":
-		conv = &nomicbertModel{}
 	case "CohereForCausalLM":
 		conv = &commandrModel{}
 	case "GptOssForCausalLM":
 		conv = &gptossModel{}
 	case "DeepseekOCRForCausalLM":
 		conv = &deepseekocr{}
-	case "DeepseekV3ForCausalLM":
-		conv = &deepseek2Model{}
 	default:
-		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
+		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
-		return nil, nil, err
+		return err
 	}
 
 	if t, ok := conv.(moreParser); ok {
 		if err := t.parseMore(fsys); err != nil {
-			return nil, nil, err
+			return err
 		}
 	}
 
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
-		return nil, nil, err
+		return err
 	}
 
 	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
@@ -360,19 +246,6 @@ func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
-	return conv, t, nil
-}
-
-// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
-// and files it finds in the input path.
-// Supported input model formats include safetensors.
-// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func ConvertModel(fsys fs.FS, f *os.File) error {
-	kv, t, err := LoadModelMetadata(fsys)
-	if err != nil {
-		return err
-	}
-	conv := kv.(ModelConverter)
 
 	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
@@ -382,7 +255,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	return writeFile(f, conv.KV(t), conv.Tensors(ts))
 }
 
-func writeFile(f *os.File, kv KV, ts []*ggml.Tensor) error {
+func writeFile(f *os.File, kv ggml.KV, ts []*ggml.Tensor) error {
 	for i := range ts {
 		ts[i].Shape = slices.Clone(ts[i].Shape)
 		slices.Reverse(ts[i].Shape)

convert/convert_bert.go

@@ -88,7 +88,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) KV {
+func (p *bertModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false

convert/convert_commandr.go

@@ -24,7 +24,7 @@ type commandrModel struct {
 
 var _ ModelConverter = (*commandrModel)(nil)
 
-func (p *commandrModel) KV(t *Tokenizer) KV {
+func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "command-r"
 	kv["general.name"] = "command-r"

convert/convert_deepseek2.go

@@ -1,173 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"fmt"
-	"log/slog"
-	"regexp"
-	"strconv"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type deepseek2Model struct {
-	ModelParameters               // architectures, vocab_size
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	HiddenLayers          uint32  `json:"num_hidden_layers"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	NumAttentionHeads     uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-	RMSNormEPS            float32 `json:"rms_norm_eps"`
-
-	RopeTheta     float32 `json:"rope_theta"`
-	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
-	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
-	KVLoraRank    uint32  `json:"kv_lora_rank"`
-	QLoraRank     uint32  `json:"q_lora_rank"`
-	VHeadDim      uint32  `json:"v_head_dim"`
-
-	ExpertCount            uint32  `json:"n_routed_experts"`
-	ExpertSharedCount      uint32  `json:"n_shared_experts"`
-	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
-	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
-	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
-	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
-
-	ScoringFunc            string `json:"scoring_func"`
-	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
-
-	RopeScaling struct {
-		Factor                        float32 `json:"factor"`
-		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
-		Type                          string  `json:"type"`
-		MScaleAllDim                  float32 `json:"mscale_all_dim"`
-	} `json:"rope_scaling"`
-
-	Architecture string
-}
-
-func (p *deepseek2Model) KV(t *Tokenizer) KV {
-	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "deepseek2"
-	kv["general.type"] = "model"
-	kv["deepseek2.block_count"] = p.HiddenLayers
-
-	numHeads := p.NumAttentionHeads
-	numKVHeads := p.NumKeyValueHeads
-
-	kv["deepseek2.attention.head_count"] = numHeads
-	kv["deepseek2.attention.head_count_kv"] = numKVHeads
-	kv["deepseek2.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
-	kv["deepseek2.attention.kv_lora_rank"] = p.KVLoraRank
-	kv["deepseek2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
-	kv["deepseek2.attention.q_lora_rank"] = p.QLoraRank
-	kv["deepseek2.attention.value_length"] = p.VHeadDim
-	kv["deepseek2.context_length"] = p.MaxPositionEmbeddings
-	kv["deepseek2.embedding_length"] = p.HiddenSize
-	kv["deepseek2.expert_count"] = p.ExpertCount
-	kv["deepseek2.expert_feed_forward_length"] = p.ExpertIntermediateSize
-	kv["deepseek2.expert_shared_count"] = p.ExpertSharedCount
-
-	var scoringFunc uint32
-	switch p.ScoringFunc {
-	case "softmax":
-		// not currently supported in the model, but needed for Deepseek-OCR
-		scoringFunc = 1
-	case "sigmoid":
-		scoringFunc = 2
-	}
-	kv["deepseek2.expert_gating_func"] = scoringFunc
-	kv["deepseek2.expert_used_count"] = p.ExpertUsedCount
-	kv["deepseek2.expert_weights_norm"] = p.ExpertWeightsNorm
-	kv["deepseek2.expert_weights_scale"] = p.ExpertWeightsScale
-	kv["deepseek2.feed_forward_length"] = p.IntermediateSize
-	kv["deepseek2.leading_dense_block_count"] = p.LeadingDenseBlockCount
-
-	kv["deepseek2.rope.dimension_count"] = p.QKRopeHeadDim
-	kv["deepseek2.rope.freq_base"] = cmp.Or(p.RopeTheta, 10000.0)
-	kv["deepseek2.rope.scaling.factor"] = p.RopeScaling.Factor
-	kv["deepseek2.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeddings
-	kv["deepseek2.rope.scaling.type"] = p.RopeScaling.Type
-	kv["deepseek2.rope.scaling.yarn_log_multiplier"] = 0.1 * p.RopeScaling.MScaleAllDim
-
-	kv["tokenizer.ggml.pre"] = "deepseek-v3"
-
-	return kv
-}
-
-func (p *deepseek2Model) Replacements() []string {
-	return []string{
-		"lm_head", "output",
-		"model.embed_tokens", "token_embd",
-		"model.norm", "output_norm",
-		"language_model.", "",
-		"model.layers", "blk",
-		"input_layernorm", "attn_norm",
-		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
-		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
-		"self_attn.kv_b_proj", "attn_kv_b",
-		"self_attn.q_a_proj", "attn_q_a",
-		"self_attn.q_a_layernorm", "attn_q_a_norm",
-		"self_attn.q_b_proj", "attn_q_b",
-		"self_attn.o_proj", "attn_output",
-		"post_attention_layernorm", "ffn_norm",
-		"mlp.shared_experts.down_proj", "ffn_down_shexp",
-		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
-		"mlp.shared_experts.up_proj", "ffn_up_shexp",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.down_proj", "ffn_down",
-		"mlp.up_proj", "ffn_up",
-		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
-		"mlp.gate", "ffn_gate_inp",
-	}
-}
-
-func (p *deepseek2Model) Tensors(s []Tensor) (out []*ggml.Tensor) {
-	merges := make([]merge, p.HiddenLayers*3)
-	for i := range p.HiddenLayers {
-		merges[i*3+0] = merge{
-			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
-			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
-		}
-		merges[i*3+1] = merge{
-			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
-			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
-		}
-		merges[i*3+2] = merge{
-			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
-			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
-		}
-	}
-
-	skipLayer := func(n string, minValue uint32) bool {
-		re := regexp.MustCompile(`^blk\.(\d+)`)
-		matches := re.FindStringSubmatch(n)
-		if matches == nil {
-			return false
-		}
-
-		blkNum, err := strconv.Atoi(matches[1])
-		if err != nil {
-			return false
-		}
-
-		return uint32(blkNum) >= minValue
-	}
-
-	out, s = mergeTensors(s, merges...)
-	for _, t := range s {
-		// skip any additional layers (such as the Multi-Token Prediction layer)
-		if skipLayer(t.Name(), p.HiddenLayers) {
-			slog.Debug("skipping layer", "name", t.Name())
-			continue
-		}
-		out = append(out, &ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-	return out
-}

convert/convert_deepseekocr.go

@@ -41,7 +41,7 @@ type deepseekocr struct {
 	} `json:"vision_config"`
 }
 
-func (m *deepseekocr) KV(t *Tokenizer) KV {
+func (m *deepseekocr) KV(t *Tokenizer) ggml.KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseekocr"
 	kv["block_count"] = m.LanguageConfig.HiddenLayers

convert/convert_gemma.go

@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) KV {
+func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2.go

@@ -1,5 +1,7 @@
 package convert
 
+import "github.com/ollama/ollama/fs/ggml"
+
 type gemma2Model struct {
 	gemmaModel
 	SlidingWindow         uint32  `json:"sliding_window"`
@@ -7,7 +9,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) KV {
+func (p *gemma2Model) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,7 +6,6 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -16,7 +15,7 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV fs.Config) KV {
+func (p *gemma2Adapter) KV(baseKV ggml.KV) ggml.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv

convert/convert_gemma3.go

@@ -2,7 +2,8 @@ package convert
 
 import (
 	"cmp"
-	"slices"
+
+	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemma3Model struct {
@@ -25,26 +26,16 @@ type gemma3Model struct {
 		NumChannels       uint32  `json:"num_channels"`        // num_channels 3
 		PatchSize         uint32  `json:"patch_size"`          // patch_size 14
 	} `json:"vision_config"`
-	MaxPositionEmbeddings    uint32   `json:"max_position_embeddings"`
-	NumAttentionHeads        uint32   `json:"num_attention_heads"`
-	NumKeyValueHeads         uint32   `json:"num_key_value_heads"`
-	RMSNormEPS               float32  `json:"rms_norm_eps"`
-	HeadDim                  uint32   `json:"head_dim"`
-	FinalLogitSoftcap        float32  `json:"final_logit_softcapping"`
-	RopeLocalTheta           float32  `json:"rope_local_base_freq"`
-	RopeTheta                float32  `json:"rope_theta"`
-	SlidingWindow            uint32   `json:"sliding_window"`
-	SlidingWindowPattern     *uint32  `json:"sliding_window_pattern"`
-	LayerTypes               []string `json:"layer_types"`
-	MultiModalTokensPerImage uint32   `json:"mm_tokens_per_image"`
-	RopeScaling              *struct {
-		Type                          string  `json:"rope_type"`
-		Factor                        float32 `json:"factor"`
-		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
-		ExtrapolationFactor           float32 `json:"extrapolation_factor"`
-		BetaFast                      float32 `json:"beta_fast"`
-		BetaSlow                      float32 `json:"beta_slow"`
-	} `json:"rope_scaling"`
+	MaxPositionEmbeddings    uint32  `json:"max_position_embeddings"`
+	NumAttentionHeads        uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads         uint32  `json:"num_key_value_heads"`
+	RMSNormEPS               float32 `json:"rms_norm_eps"`
+	HeadDim                  uint32  `json:"head_dim"`
+	FinalLogitSoftcap        float32 `json:"final_logit_softcapping"`
+	RopeLocalTheta           float32 `json:"rope_local_base_freq"`
+	RopeGlobalTheta          float32 `json:"rope_global_base_freq"`
+	SlidingWindow            uint32  `json:"sliding_window"`
+	MultiModalTokensPerImage uint32  `json:"mm_tokens_per_image"`
 }
 
 const (
@@ -53,7 +44,7 @@ const (
 	gemma27BLayerCount = 62
 )
 
-func (p *gemma3Model) KV(t *Tokenizer) KV {
+func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3"
 
@@ -90,38 +81,9 @@ func (p *gemma3Model) KV(t *Tokenizer) KV {
 		kv["gemma3.attention.key_length"] = p.HeadDim
 		kv["gemma3.attention.value_length"] = p.HeadDim
 		kv["gemma3.attention.sliding_window"] = p.SlidingWindow
-
-		// The sliding window pattern is either provided as the sliding_window_pattern
-		// key (an int) or as the layer_types key (a list of strings).
-		if p.SlidingWindowPattern != nil || len(p.LayerTypes) > 0 {
-			kv["gemma3.attention.sliding_window_pattern"] = slices.Collect(func(yield func(bool) bool) {
-				for i := range numBlocks {
-					var isLocal bool
-					if len(p.LayerTypes) > 0 && int(i) < len(p.LayerTypes) {
-						isLocal = p.LayerTypes[i] == "sliding_attention"
-					} else if p.SlidingWindowPattern != nil && *p.SlidingWindowPattern > 0 {
-						isLocal = (i+1)%*p.SlidingWindowPattern != 0
-					}
-					if !yield(isLocal) {
-						break
-					}
-				}
-			})
-		}
-		if p.FinalLogitSoftcap > 0 {
-			kv["gemma3.final_logit_softcapping"] = p.FinalLogitSoftcap
-		}
+		kv["gemma3.final_logit_softcapping"] = cmp.Or(p.FinalLogitSoftcap, 30)
 		kv["gemma3.rope.local.freq_base"] = cmp.Or(p.RopeLocalTheta, 10000.0)
-		kv["gemma3.rope.freq_base"] = cmp.Or(p.RopeTheta, 1000000.0)
-		if p.RopeScaling != nil && p.RopeScaling.Type == "yarn" && p.RopeScaling.Factor > 0 {
-			kv["gemma3.rope.scaling.type"] = "yarn"
-			kv["gemma3.rope.scaling.factor"] = p.RopeScaling.Factor
-			kv["gemma3.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeddings
-			kv["gemma3.rope.scaling.extrapolation_factor"] = cmp.Or(p.RopeScaling.ExtrapolationFactor, float32(1.0))
-			kv["gemma3.rope.scaling.beta_fast"] = cmp.Or(p.RopeScaling.BetaFast, float32(64.0))
-			kv["gemma3.rope.scaling.beta_slow"] = cmp.Or(p.RopeScaling.BetaSlow, float32(1.0))
-		}
-
+		kv["gemma3.rope.global.freq_base"] = cmp.Or(p.RopeGlobalTheta, 1000000.0)
 		kv["gemma3.embedding_length"] = p.HiddenSize
 		kv["gemma3.feed_forward_length"] = p.IntermediateSize
 	default:

convert/convert_gemma3n.go

@@ -38,7 +38,7 @@ type gemma3nModel struct {
 	VisionModel struct{} `json:"vision_config"`
 }
 
-func (m *gemma3nModel) KV(t *Tokenizer) KV {
+func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3n"
 	kv["gemma3n.activation_sparsity_scale"] = slices.Collect(func(yield func(float32) bool) {

convert/convert_gptoss.go

@@ -37,7 +37,7 @@ type gptossModel struct {
 
 var _ ModelConverter = (*gptossModel)(nil)
 
-func (m *gptossModel) KV(t *Tokenizer) KV {
+func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gptoss"
 	kv["general.file_type"] = uint32(4)

convert/convert_llama.go

@@ -48,7 +48,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) KV {
+func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize

convert/convert_llama4.go

@@ -35,7 +35,7 @@ type llama4Model struct {
 }
 
 // KV implements ModelConverter.
-func (p *llama4Model) KV(t *Tokenizer) KV {
+func (p *llama4Model) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama4"
 

convert/convert_llama_adapter.go

@@ -7,7 +7,6 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -19,13 +18,13 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV fs.Config) KV {
+func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
-	kv["llama.attention.head_count"] = baseKV.Value("llama.attention.head_count")
-	kv["llama.attention.head_count_kv"] = baseKV.Value("llama.attention.head_count_kv")
+	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
+	kv["llama.attention.head_count_kv"] = baseKV["llama.attention.head_count_kv"]
 
-	p.NumAttentionHeads = baseKV.Value("llama.attention.head_count").(uint32)
+	p.NumAttentionHeads = baseKV["llama.attention.head_count"].(uint32)
 
 	return kv
 }

convert/convert_mistral.go

@@ -29,17 +29,6 @@ type mistral3Model struct {
 		SlidingWindow         *uint32 `json:"sliding_window"`
 		HiddenAct             string  `json:"hidden_act"`
 		VocabSize             uint32  `json:"vocab_size"`
-		RopeParameters        struct {
-			BetaFast                  float32  `json:"beta_fast"`
-			BetaSlow                  float32  `json:"beta_slow"`
-			Factor                    float32  `json:"factor"`
-			Llama4ScalingBeta         *float32 `json:"llama_4_scaling_beta"`
-			OrigMaxPositionEmbeddings uint32   `json:"original_max_position_embeddings"`
-			RopeType                  string   `json:"rope_type"`
-			RopeTheta                 float32  `json:"rope_theta"`
-			Mscale                    *float32 `json:"mscale"`
-			MscaleAllDim              *float32 `json:"mscale_all_dim"`
-		} `json:"rope_parameters"`
 	} `json:"text_config"`
 	VisionModel struct {
 		NumAttentionHeads uint32  `json:"num_attention_heads"`
@@ -52,15 +41,12 @@ type mistral3Model struct {
 		HeadDim           uint32  `json:"head_dim"`
 		HiddenAct         string  `json:"hidden_act"`
 		RopeTheta         float32 `json:"rope_theta"`
-		RopeParameters    struct {
-			RopeTheta float32 `json:"rope_theta"`
-		} `json:"rope_parameters"`
 	} `json:"vision_config"`
 	MultiModalProjectorBias bool   `json:"multimodal_projector_bias"`
 	ProjectorHiddenAct      string `json:"projector_hidden_act"`
 }
 
-func (p *mistral3Model) KV(t *Tokenizer) KV {
+func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.TextModel.VocabSize
@@ -75,25 +61,8 @@ func (p *mistral3Model) KV(t *Tokenizer) KV {
 	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.TextModel.RMSNormEPS
 	kv["mistral3.attention.key_length"] = p.TextModel.HeadDim
 	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
-	kv["mistral3.rope.dimension_count"] = cmp.Or(p.TextModel.HeadDim, p.TextModel.HiddenSize/p.TextModel.NumAttentionHeads)
-	kv["mistral3.rope.freq_base"] = cmp.Or(p.TextModel.RopeTheta, p.TextModel.RopeParameters.RopeTheta)
-	kv["mistral3.rope.scaling.factor"] = p.TextModel.RopeParameters.Factor
-	kv["mistral3.rope.scaling.type"] = p.TextModel.RopeParameters.RopeType
-	kv["mistral3.rope.scaling.beta_fast"] = p.TextModel.RopeParameters.BetaFast
-	kv["mistral3.rope.scaling.beta_slow"] = p.TextModel.RopeParameters.BetaSlow
-
-	if p.TextModel.RopeParameters.Mscale != nil {
-		kv["mistral3.rope.scaling.mscale"] = *p.TextModel.RopeParameters.Mscale
-	}
-	if p.TextModel.RopeParameters.MscaleAllDim != nil {
-		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.TextModel.RopeParameters.MscaleAllDim
-	}
-	if p.TextModel.RopeParameters.OrigMaxPositionEmbeddings > 0 {
-		kv["mistral3.rope.scaling.original_context_length"] = p.TextModel.RopeParameters.OrigMaxPositionEmbeddings
-	}
-	if p.TextModel.RopeParameters.Llama4ScalingBeta != nil {
-		kv["mistral3.rope.scaling_beta"] = *p.TextModel.RopeParameters.Llama4ScalingBeta
-	}
+	kv["mistral3.rope.dimension_count"] = p.TextModel.HiddenSize / p.TextModel.NumHiddenLayers
+	kv["mistral3.rope.freq_base"] = p.TextModel.RopeTheta
 
 	// Vision configuration
 	kv["mistral3.vision.block_count"] = p.VisionModel.NumHiddenLayers
@@ -105,7 +74,7 @@ func (p *mistral3Model) KV(t *Tokenizer) KV {
 	kv["mistral3.vision.patch_size"] = p.VisionModel.PatchSize
 	kv["mistral3.vision.num_channels"] = p.VisionModel.NumChannels
 	// kv["mistral3.vision.attention.layer_norm_epsilon"] = 1e-05 // Default value
-	kv["mistral3.vision.rope.freq_base"] = cmp.Or(p.VisionModel.RopeTheta, p.VisionModel.RopeParameters.RopeTheta)
+	kv["mistral3.vision.rope.freq_base"] = p.VisionModel.RopeTheta
 
 	// Multimodal configuration
 	kv["mistral3.image_token_index"] = p.ImageTokenIndex

convert/convert_mistral_causal.go

@@ -1,181 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"fmt"
-	"strings"
-
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type mistral3CausalModel struct {
-	ModelParameters
-
-	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	NumAttentionHeads     uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-	RopeTheta             float32 `json:"rope_theta"`
-	RMSNormEPS            float32 `json:"rms_norm_eps"`
-	HeadDim               uint32  `json:"head_dim"`
-	SlidingWindow         *uint32 `json:"sliding_window"`
-	HiddenAct             string  `json:"hidden_act"`
-	VocabSize             uint32  `json:"vocab_size"`
-	RopeParameters        struct {
-		BetaFast                  float32  `json:"beta_fast"`
-		BetaSlow                  float32  `json:"beta_slow"`
-		Factor                    float32  `json:"factor"`
-		Llama4ScalingBeta         *float32 `json:"llama_4_scaling_beta"`
-		OrigMaxPositionEmbeddings uint32   `json:"original_max_position_embeddings"`
-		RopeType                  string   `json:"rope_type"`
-		RopeTheta                 float32  `json:"rope_theta"`
-		Mscale                    *float32 `json:"mscale"`
-		MscaleAllDim              *float32 `json:"mscale_all_dim"`
-	} `json:"rope_parameters"`
-}
-
-func (p *mistral3CausalModel) KV(t *Tokenizer) KV {
-	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "mistral3"
-	kv["mistral3.vocab_size"] = p.VocabSize
-
-	// Text configuration
-	kv["mistral3.block_count"] = p.NumHiddenLayers
-	kv["mistral3.context_length"] = p.MaxPositionEmbeddings
-	kv["mistral3.embedding_length"] = p.HiddenSize
-	kv["mistral3.feed_forward_length"] = p.IntermediateSize
-	kv["mistral3.attention.head_count"] = p.NumAttentionHeads
-	kv["mistral3.attention.head_count_kv"] = p.NumKeyValueHeads
-	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
-	kv["mistral3.attention.key_length"] = p.HeadDim
-	kv["mistral3.attention.value_length"] = p.HeadDim
-	kv["mistral3.rope.dimension_count"] = cmp.Or(p.HeadDim, p.HiddenSize/p.NumAttentionHeads)
-	kv["mistral3.rope.freq_base"] = cmp.Or(p.RopeTheta, p.RopeParameters.RopeTheta)
-	kv["mistral3.rope.scaling.factor"] = p.RopeParameters.Factor
-	kv["mistral3.rope.scaling.type"] = p.RopeParameters.RopeType
-	kv["mistral3.rope.scaling.beta_fast"] = p.RopeParameters.BetaFast
-	kv["mistral3.rope.scaling.beta_slow"] = p.RopeParameters.BetaSlow
-
-	if p.RopeParameters.Mscale != nil {
-		kv["mistral3.rope.scaling.mscale"] = *p.RopeParameters.Mscale
-	}
-
-	if p.RopeParameters.MscaleAllDim != nil {
-		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.RopeParameters.MscaleAllDim
-	}
-
-	if p.RopeParameters.OrigMaxPositionEmbeddings > 0 {
-		kv["mistral3.rope.scaling.original_context_length"] = p.RopeParameters.OrigMaxPositionEmbeddings
-		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
-	}
-
-	if p.RopeParameters.Llama4ScalingBeta != nil {
-		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
-	}
-
-	return kv
-}
-
-func (p *mistral3CausalModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	var out []*ggml.Tensor
-
-	for _, t := range ts {
-		if !strings.HasPrefix(t.Name(), "v.") {
-			if strings.HasSuffix(t.Name(), ".attn_q.weight") ||
-				strings.HasSuffix(t.Name(), ".attn_k.weight") {
-				t.SetRepacker(p.repack)
-			}
-		}
-
-		out = append(out, &ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *mistral3CausalModel) Replacements() []string {
-	return []string{
-		"model.norm", "output_norm",
-		"model.", "",
-		"layers", "blk",
-		"transformer.layers", "blk",
-		"vision_tower", "v",
-		"ln_pre", "encoder_norm",
-		"input_layernorm", "attn_norm",
-		"post_attention_layernorm", "ffn_norm",
-		"embed_tokens", "token_embd",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.v_proj", "attn_v",
-		"self_attn.o_proj", "attn_output",
-		"mlp.down_proj", "ffn_down",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.up_proj", "ffn_up",
-		"attention.q_proj", "attn_q",
-		"attention.k_proj", "attn_k",
-		"attention.v_proj", "attn_v",
-		"attention.o_proj", "attn_output",
-		"attention_norm", "attn_norm",
-		"feed_forward.gate_proj", "ffn_gate",
-		"feed_forward.down_proj", "ffn_down",
-		"feed_forward.up_proj", "ffn_up",
-		"multi_modal_projector", "mm",
-		"ffn_norm", "ffn_norm",
-		"lm_head", "output",
-	}
-}
-
-func (p *mistral3CausalModel) repack(name string, data []float32, shape []uint64) ([]float32, error) {
-	var dims []int
-	for _, dim := range shape {
-		dims = append(dims, int(dim))
-	}
-
-	var heads uint32
-	if strings.HasSuffix(name, ".attn_q.weight") {
-		heads = p.NumAttentionHeads
-	} else if strings.HasSuffix(name, ".attn_k.weight") {
-		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
-	} else {
-		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
-	}
-
-	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(dims...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-
-	ts, err := native.SelectF32(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var f32s []float32
-	for _, t := range ts {
-		f32s = append(f32s, t...)
-	}
-
-	return f32s, nil
-}

convert/convert_mixtral.go

@@ -12,7 +12,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) KV {
+func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {

convert/convert_mllama.go

@@ -34,7 +34,7 @@ type mllamaModel struct {
 	} `json:"vision_config"`
 }
 
-func (m *mllamaModel) KV(t *Tokenizer) KV {
+func (m *mllamaModel) KV(t *Tokenizer) ggml.KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "mllama"
 

convert/convert_nomicbert.go

@@ -1,213 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"encoding/json"
-	"io/fs"
-	"path/filepath"
-	"slices"
-	"strings"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type nomicbertModel struct {
-	ModelParameters
-	NLayers               uint32  `json:"n_layers"`
-	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	NumAttentionHeads     uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-	LayerNormEPS          float32 `json:"layer_norm_eps"`
-	LayerNormEpsilon      float32 `json:"layer_norm_epsilon"`
-	RopeFreqBase          float32 `json:"rope_theta"`
-	normalizeEmbeddings   bool
-	PoolingType           uint32
-
-	// MoE parameters (only present in v2 models)
-	NumExperts      uint32 `json:"num_local_experts"`
-	NumExpertsUsed  uint32 `json:"num_experts_per_tok"`
-	MoEEveryNLayers uint32 `json:"moe_every_n_layers"`
-}
-
-var (
-	_ ModelConverter = (*nomicbertModel)(nil)
-	_ moreParser     = (*nomicbertModel)(nil)
-)
-
-func (p *nomicbertModel) parseMore(fsys fs.FS) error {
-	bts, err := fs.ReadFile(fsys, "modules.json")
-	if err != nil {
-		return err
-	}
-
-	var modules []struct {
-		Type string `json:"type"`
-		Path string `json:"path"`
-	}
-
-	if err := json.Unmarshal(bts, &modules); err != nil {
-		return err
-	}
-
-	var pooling string
-	for _, m := range modules {
-		switch m.Type {
-		case "sentence_transformers.models.Pooling":
-			pooling = m.Path
-		case "sentence_transformers.models.Normalize":
-			p.normalizeEmbeddings = true
-		}
-	}
-
-	if pooling != "" {
-		bts, err := fs.ReadFile(fsys, filepath.Join(pooling, "config.json"))
-		if err != nil {
-			return err
-		}
-
-		var pc struct {
-			PoolingModeCLSToken   bool `json:"pooling_mode_cls_token"`
-			PoolingModeMeanTokens bool `json:"pooling_mode_mean_tokens"`
-		}
-
-		if err := json.Unmarshal(bts, &pc); err != nil {
-			return err
-		}
-
-		if pc.PoolingModeMeanTokens {
-			p.PoolingType = 1
-		} else if pc.PoolingModeCLSToken {
-			p.PoolingType = 2
-		}
-	}
-
-	return nil
-}
-
-func (p *nomicbertModel) KV(t *Tokenizer) KV {
-	kv := p.ModelParameters.KV(t)
-
-	// Determine architecture based on MoE parameters (following qwen3 pattern)
-	arch := "nomic-bert"
-	if p.MoEEveryNLayers > 0 {
-		arch += "-moe"
-	}
-
-	kv["general.architecture"] = arch
-	kv["attention.causal"] = false
-	kv["pooling_type"] = p.PoolingType
-	kv["normalize_embeddings"] = p.normalizeEmbeddings
-
-	kv["block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers)
-
-	if contextLength := p.MaxPositionEmbeddings; contextLength > 0 {
-		kv["context_length"] = contextLength
-	}
-
-	if embeddingLength := p.HiddenSize; embeddingLength > 0 {
-		kv["embedding_length"] = p.HiddenSize
-	}
-
-	if feedForwardLength := p.IntermediateSize; feedForwardLength > 0 {
-		kv["feed_forward_length"] = p.IntermediateSize
-	}
-
-	if headCount := p.NumAttentionHeads; headCount > 0 {
-		kv["attention.head_count"] = p.NumAttentionHeads
-	}
-
-	if kvHeadCount := p.NumKeyValueHeads; kvHeadCount > 0 {
-		kv["attention.head_count_kv"] = p.NumKeyValueHeads
-	}
-
-	if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon); layerNormEpsilon > 0 {
-		kv["attention.layer_norm_epsilon"] = layerNormEpsilon
-	}
-
-	if p.RopeFreqBase > 0 {
-		kv["rope.freq_base"] = p.RopeFreqBase
-	}
-
-	// MoE specific parameters (only if MoE is enabled)
-	if p.NumExperts > 0 {
-		kv["expert_count"] = p.NumExperts
-	}
-
-	if p.NumExpertsUsed > 0 {
-		kv["expert_used_count"] = p.NumExpertsUsed
-	}
-
-	if p.MoEEveryNLayers > 0 {
-		kv["moe_every_n_layers"] = p.MoEEveryNLayers
-	}
-
-	kv["tokenizer.ggml.model"] = "bert"
-	kv["tokenizer.ggml.token_type_count"] = uint32(2)
-
-	// convert to phantom space tokens
-	for i, e := range t.Tokens {
-		switch {
-		case strings.HasPrefix(e, "[") && strings.HasSuffix(e, "]"):
-			// noop - keep special tokens as-is
-		case strings.HasPrefix(e, "##"):
-			t.Tokens[i] = e[2:]
-		default:
-			t.Tokens[i] = "\u2581" + e
-		}
-	}
-
-	kv["tokenizer.ggml.tokens"] = t.Tokens
-
-	return kv
-}
-
-func (p *nomicbertModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	out := make([]*ggml.Tensor, 0, len(ts))
-	for _, t := range ts {
-		if slices.Contains([]string{
-			"embeddings.position_ids",
-			"pooler.dense.weight",
-			"pooler.dense.bias",
-		}, t.Name()) {
-			continue
-		}
-
-		out = append(out, &ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (nomicbertModel) Replacements() []string {
-	return []string{
-		"encoder.layer", "blk",
-		"encoder.layers", "blk",
-		"embeddings.word_embeddings", "token_embd",
-		"embeddings.token_type_embeddings", "token_types",
-		"embeddings.LayerNorm", "token_embd_norm",
-
-		"attention.self.qkv", "attn_qkv",
-
-		"attention.output.dense", "attn_output",
-		"attention.output.LayerNorm", "attn_output_norm",
-
-		"mlp.up", "ffn_up",
-		"mlp.down", "ffn_down",
-
-		"mlp.router", "ffn_gate_inp",
-		"mlp.experts.up", "ffn_up_exps",
-		"mlp.experts.down", "ffn_down_exps",
-
-		"intermediate.dense", "ffn_up",
-		"output.dense", "ffn_down",
-		"output.LayerNorm", "layer_output_norm",
-	}
-}

convert/convert_olmo.go

@@ -1,117 +0,0 @@
-package convert
-
-import (
-	"cmp"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type ropeScaling struct {
-	Factor                    float32 `json:"factor"`
-	OriginalMaxPositionEmbeds uint32  `json:"original_max_position_embeddings"`
-	AttentionFactor           float32 `json:"attention_factor"`
-	BetaFast                  float32 `json:"beta_fast"`
-	BetaSlow                  float32 `json:"beta_slow"`
-	RopeType                  string  `json:"rope_type"`
-	ExtrapolationFactor       float32 `json:"extrapolation_factor"`
-}
-
-type olmoModel struct {
-	ModelParameters
-
-	HiddenSize            uint32       `json:"hidden_size"`
-	NumHiddenLayers       uint32       `json:"num_hidden_layers"`
-	IntermediateSize      uint32       `json:"intermediate_size"`
-	NumAttentionHeads     uint32       `json:"num_attention_heads"`
-	NumKeyValueHeads      uint32       `json:"num_key_value_heads"`
-	MaxPositionEmbeddings uint32       `json:"max_position_embeddings"`
-	RMSNormEPS            float32      `json:"rms_norm_eps"`
-	RopeTheta             float32      `json:"rope_theta"`
-	RopeScaling           *ropeScaling `json:"rope_scaling"`
-	SlidingWindow         uint32       `json:"sliding_window"`
-	LayerTypes            []string     `json:"layer_types"`
-}
-
-var _ ModelConverter = (*olmoModel)(nil)
-
-func (p *olmoModel) KV(t *Tokenizer) KV {
-	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "olmo3"
-	kv["olmo3.block_count"] = p.NumHiddenLayers
-	kv["olmo3.context_length"] = p.MaxPositionEmbeddings
-	kv["olmo3.embedding_length"] = p.HiddenSize
-	kv["olmo3.feed_forward_length"] = p.IntermediateSize
-	kv["olmo3.attention.head_count"] = p.NumAttentionHeads
-	kv["olmo3.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
-
-	if p.RopeTheta > 0 {
-		kv["olmo3.rope.freq_base"] = p.RopeTheta
-	}
-
-	if p.RopeScaling != nil {
-		if p.RopeScaling.Factor > 0 {
-			kv["olmo3.rope.scaling.factor"] = p.RopeScaling.Factor
-		}
-		if p.RopeScaling.OriginalMaxPositionEmbeds > 0 {
-			kv["olmo3.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeds
-		}
-		if p.RopeScaling.AttentionFactor > 0 {
-			kv["olmo3.rope.scaling.attn_factor"] = p.RopeScaling.AttentionFactor
-		}
-		if p.RopeScaling.RopeType != "" {
-			kv["olmo3.rope.scaling.type"] = p.RopeScaling.RopeType
-		}
-	}
-
-	if p.RMSNormEPS > 0 {
-		kv["olmo3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
-	}
-
-	if p.SlidingWindow > 0 {
-		kv["olmo3.attention.sliding_window"] = p.SlidingWindow
-	}
-
-	if len(p.LayerTypes) > 0 {
-		slidingPattern := make([]bool, len(p.LayerTypes))
-		for i, layerType := range p.LayerTypes {
-			slidingPattern[i] = (layerType == "sliding_attention")
-		}
-		kv["olmo3.attention.sliding_window_pattern"] = slidingPattern
-	}
-
-	return kv
-}
-
-func (p *olmoModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	out := make([]*ggml.Tensor, 0, len(ts))
-	for _, t := range ts {
-		out = append(out, &ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *olmoModel) Replacements() []string {
-	return []string{
-		"lm_head", "output",
-		"model.embed_tokens", "token_embd",
-		"model.layers", "blk",
-		"model.norm", "output_norm",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.v_proj", "attn_v",
-		"self_attn.o_proj", "attn_output",
-		"self_attn.q_norm", "attn_q_norm",
-		"self_attn.k_norm", "attn_k_norm",
-		"post_attention_layernorm", "post_attention_norm",
-		"post_feedforward_layernorm", "post_ffw_norm",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.down_proj", "ffn_down",
-		"mlp.up_proj", "ffn_up",
-	}
-}

convert/convert_phi3.go

@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) KV {
+func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings

convert/convert_qwen2.go

@@ -22,7 +22,7 @@ type qwen2Model struct {
 
 var _ ModelConverter = (*qwen2Model)(nil)
 
-func (q *qwen2Model) KV(t *Tokenizer) KV {
+func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen2"
 	kv["qwen2.block_count"] = q.HiddenLayers

convert/convert_qwen25vl.go

@@ -29,7 +29,7 @@ type qwen25VLModel struct {
 
 var _ ModelConverter = (*qwen25VLModel)(nil)
 
-func (q *qwen25VLModel) KV(t *Tokenizer) KV {
+func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen25vl"
 

convert/convert_qwen3.go

@@ -32,7 +32,7 @@ type qwen3Model struct {
 }
 
 // KV implements ModelConverter.
-func (q *qwen3Model) KV(t *Tokenizer) KV {
+func (q *qwen3Model) KV(t *Tokenizer) ggml.KV {
 	arch := "qwen3"
 	if q.NumExperts > 0 {
 		arch += "moe"

convert/convert_qwen3vl.go

@@ -45,7 +45,7 @@ func (m *qwen3VLModel) parseMore(fsys fs.FS) error {
 	return json.Unmarshal(bts, &m.VisionModel)
 }
 
-func (m *qwen3VLModel) KV(t *Tokenizer) KV {
+func (m *qwen3VLModel) KV(t *Tokenizer) ggml.KV {
 	kv := m.qwen3Model.KV(t)
 
 	arch := "qwen3vl"

convert/convert_test.go

@@ -19,7 +19,6 @@ import (
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
-	fsc "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -29,7 +28,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, fsc.Config, ggml.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -60,10 +59,9 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, fsc.Config, ggml.Tensors)
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv fsc.Config, tensors ggml.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv ggml.KV, tensors ggml.Tensors) map[string]string {
 	actual := make(map[string]string)
-	for k := range kv.Keys() {
-		v := kv.Value(k)
+	for k, v := range kv {
 		if s, ok := v.(json.Marshaler); !ok {
 			actual[k] = fmt.Sprintf("%v", v)
 		} else {
@@ -279,7 +277,7 @@ func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[str
 func TestConvertAdapter(t *testing.T) {
 	type AdapterCase struct {
 		Name     string
-		BaseKV   KV
+		BaseKV   map[string]any
 		Expected map[string]string
 	}
 

convert/tokenizer_spm.go

@@ -49,8 +49,7 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 			tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
 
 			// temporary fix to handle gemma3 broken configs
-			// TODO(parthsareen): allow reading of tokenizer.json to allow managing special tokens when using spm
-			if slices.Contains([]string{"<end_of_turn>", "<start_of_turn>", "<start_function_declaration>", "<end_function_declaration>", "<start_function_call>", "<end_function_call>", "<start_function_response>", "<end_function_response>", "<escape>"}, piece.GetPiece()) {
+			if slices.Contains([]string{"<end_of_turn>", "<start_of_turn>"}, piece.GetPiece()) {
 				tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
 			}
 

discover/runner.go

@@ -65,7 +65,6 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 		}
 
 		slog.Info("discovering available GPUs...")
-		detectIncompatibleLibraries()
 
 		// Warn if any user-overrides are set which could lead to incorrect GPU discovery
 		overrideWarnings()
@@ -99,9 +98,6 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 					continue
 				} else if jetpack != "" && filepath.Base(dir) != "cuda_"+jetpack {
 					continue
-				} else if jetpack == "" && strings.Contains(filepath.Base(dir), "cuda_jetpack") {
-					slog.Debug("jetpack not detected (set JETSON_JETPACK or OLLAMA_LLM_LIBRARY to override), skipping", "libDir", dir)
-					continue
 				} else if !envconfig.EnableVulkan() && strings.Contains(filepath.Base(dir), "vulkan") {
 					slog.Info("experimental Vulkan support disabled.  To enable, set OLLAMA_VULKAN=1")
 					continue
@@ -147,7 +143,7 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 			wg.Add(1)
 			go func(i int) {
 				defer wg.Done()
-				extraEnvs := ml.GetVisibleDevicesEnv(devices[i:i+1], true)
+				extraEnvs := ml.GetVisibleDevicesEnv(devices[i : i+1])
 				devices[i].AddInitValidation(extraEnvs)
 				if len(bootstrapDevices(ctx2ndPass, devices[i].LibraryPath, extraEnvs)) == 0 {
 					slog.Debug("filtering device which didn't fully initialize",
@@ -333,8 +329,7 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 			defer cancel()
 
 			// Apply any dev filters to avoid re-discovering unsupported devices, and get IDs correct
-			// We avoid CUDA filters here to keep ROCm from failing to discover GPUs in a mixed environment
-			devFilter := ml.GetVisibleDevicesEnv(devices, false)
+			devFilter := ml.GetVisibleDevicesEnv(devices)
 
 			for dir := range libDirs {
 				updatedDevices := bootstrapDevices(ctx, []string{ml.LibOllamaPath, dir}, devFilter)
@@ -489,16 +484,3 @@ func overrideWarnings() {
 		slog.Warn("if GPUs are not correctly discovered, unset and try again")
 	}
 }
-
-func detectIncompatibleLibraries() {
-	if runtime.GOOS != "windows" {
-		return
-	}
-	basePath, err := exec.LookPath("ggml-base.dll")
-	if err != nil || basePath == "" {
-		return
-	}
-	if !strings.HasPrefix(basePath, ml.LibOllamaPath) {
-		slog.Warn("potentially incompatible library detected in PATH", "location", basePath)
-	}
-}

docs/api.md

@@ -50,7 +50,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 Advanced parameters (optional):
 
 - `format`: the format to return a response in. Format can be `json` or a JSON schema
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `system`: system message to (overrides what is defined in the `Modelfile`)
 - `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
@@ -507,7 +507,7 @@ The `message` object has the following fields:
 Advanced parameters (optional):
 
 - `format`: the format to return a response in. Format can be `json` or a JSON schema.
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
@@ -895,11 +895,11 @@ curl http://localhost:11434/api/chat -d '{
       "tool_calls": [
         {
           "function": {
-            "name": "get_weather",
+            "name": "get_temperature",
             "arguments": {
               "city": "Toronto"
             }
-          }
+          },
         }
       ]
     },
@@ -907,7 +907,7 @@ curl http://localhost:11434/api/chat -d '{
     {
       "role": "tool",
       "content": "11 degrees celsius",
-      "tool_name": "get_weather"
+      "tool_name": "get_temperature",
     }
   ],
   "stream": false,
@@ -1189,7 +1189,7 @@ If you are creating a model from a safetensors directory or from a GGUF file, yo
 - `template`: (optional) the prompt template for the model
 - `license`: (optional) a string or list of strings containing the license or licenses for the model
 - `system`: (optional) a string containing the system prompt for the model
-- `parameters`: (optional) a dictionary of parameters for the model (see [Modelfile](./modelfile.mdx#valid-parameters-and-values) for a list of parameters)
+- `parameters`: (optional) a dictionary of parameters for the model (see [Modelfile](./modelfile.md#valid-parameters-and-values) for a list of parameters)
 - `messages`: (optional) a list of message objects used to create a conversation
 - `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
 - `quantize` (optional): quantize a non-quantized (e.g. float16) model
@@ -1698,7 +1698,7 @@ Generate embeddings from a model
 Advanced parameters:
 
 - `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 - `dimensions`: number of dimensions for the embedding
 
@@ -1817,7 +1817,7 @@ Generate embeddings from a model
 
 Advanced parameters:
 
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
 ### Examples

docs/capabilities/tool-calling.mdx

@@ -15,7 +15,7 @@ Also known as "single-shot" tool calling.
     ```shell
     curl -s http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
       "model": "qwen3",
-      "messages": [{"role": "user", "content": "What is the temperature in New York?"}],
+      "messages": [{"role": "user", "content": "What's the temperature in New York?"}],
       "stream": false,
       "tools": [
         {
@@ -41,7 +41,7 @@ Also known as "single-shot" tool calling.
     curl -s http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
       "model": "qwen3",
       "messages": [
-        {"role": "user", "content": "What is the temperature in New York?"},
+        {"role": "user", "content": "What's the temperature in New York?"},
         {
           "role": "assistant",
           "tool_calls": [
@@ -90,7 +90,7 @@ Also known as "single-shot" tool calling.
       }
       return temperatures.get(city, "Unknown")
 
-    messages = [{"role": "user", "content": "What is the temperature in New York?"}]
+    messages = [{"role": "user", "content": "What's the temperature in New York?"}]
 
     # pass functions directly as tools in the tools list or as a JSON schema
     response = chat(model="qwen3", messages=messages, tools=[get_temperature], think=True)
@@ -146,7 +146,7 @@ Also known as "single-shot" tool calling.
       },
     ]
 
-    const messages = [{ role: 'user', content: "What is the temperature in New York?" }]
+    const messages = [{ role: 'user', content: "What's the temperature in New York?" }]
 
     const response = await ollama.chat({
       model: 'qwen3',
@@ -609,7 +609,7 @@ def get_temperature(city: str) -> str:
   return temperatures.get(city, 'Unknown')
 
 
-messages = [{'role': 'user', 'content': "What is the temperature in New York?"}]
+messages = [{'role': 'user', 'content': "What's the temperature in New York?"}]
 
 while True:
   stream = chat(
@@ -684,7 +684,7 @@ const getTemperatureTool = {
 }
 
 async function agentLoop() {
-  const messages = [{ role: 'user', content: "What is the temperature in New York?" }]
+  const messages = [{ role: 'user', content: "What's the temperature in New York?" }]
 
   while (true) {
     const stream = await ollama.chat({

docs/capabilities/vision.mdx

@@ -36,6 +36,7 @@ Provide an `images` array. SDKs accept file paths, URLs or raw bytes while the R
         }],
         "stream": false
     }'
+    "
     ```
   </Tab>
   <Tab title="Python">

docs/development.md

@@ -49,8 +49,6 @@ Install prerequisites:
     - [Ninja](https://github.com/ninja-build/ninja/releases)
 - (Optional) NVIDIA GPU support
     - [CUDA SDK](https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_network)
-- (Optional) VULKAN GPU support
-    - [VULKAN SDK](https://vulkan.lunarg.com/sdk/home) - useful for AMD/Intel GPUs
 
 Then, configure and build the project:
 
@@ -59,17 +57,6 @@ cmake -B build
 cmake --build build --config Release
 ```
 
-> Building for Vulkan requires VULKAN_SDK environment variable:
-> 
-> PowerShell
-> ```powershell
-> $env:VULKAN_SDK="C:\VulkanSDK\<version>"
-> ```
-> CMD
-> ```cmd
-> set VULKAN_SDK=C:\VulkanSDK\<version>
-> ```
-
 > [!IMPORTANT]
 > Building for ROCm requires additional flags:
 > ```
@@ -78,7 +65,6 @@ cmake --build build --config Release
 > ```
 
 
-
 Lastly, run Ollama:
 
 ```shell
@@ -98,9 +84,7 @@ Install prerequisites:
     - [ROCm](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/quick-start.html)
 - (Optional) NVIDIA GPU support
     - [CUDA SDK](https://developer.nvidia.com/cuda-downloads)
-- (Optional) VULKAN GPU support
-    - [VULKAN SDK](https://vulkan.lunarg.com/sdk/home) - useful for AMD/Intel GPUs
-    - Or install via package manager: `sudo apt install vulkan-sdk` (Ubuntu/Debian) or `sudo dnf install vulkan-sdk` (Fedora/CentOS)
+
 > [!IMPORTANT]
 > Ensure prerequisites are in `PATH` before running CMake.
 

docs/faq.mdx

@@ -14,11 +14,11 @@ curl -fsSL https://ollama.com/install.sh | sh
 
 ## How can I view the logs?
 
-Review the [Troubleshooting](./troubleshooting) docs for more about using logs.
+Review the [Troubleshooting](./troubleshooting.md) docs for more about using logs.
 
 ## Is my GPU compatible with Ollama?
 
-Please refer to the [GPU docs](./gpu).
+Please refer to the [GPU docs](./gpu.md).
 
 ## How can I specify the context window size?
 
@@ -57,13 +57,8 @@ ollama ps
 ```
 
 <Info>
-
-**Output**:
-
-```
-NAME        ID            SIZE    PROCESSOR   UNTIL
-llama3:70b  bcfb190ca3a7  42 GB   100% GPU    4 minutes from now
-```
+  **Output**: ``` NAME ID SIZE PROCESSOR UNTIL llama3:70b bcfb190ca3a7 42 GB
+  100% GPU 4 minutes from now ```
 </Info>
 
 The `Processor` column will show which memory the model was loaded in to:
@@ -390,4 +385,4 @@ Ollama for Windows and macOS register as a login item during installation.  You
 - In `Task Manager` go to the `Startup apps` tab, search for `ollama` then click `Disable`
 
 **MacOS**
-- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.
+- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.

docs/gpu.mdx

@@ -33,7 +33,7 @@ Check your compute compatibility to see if your card is supported:
 | 5.0                | GeForce GTX         | `GTX 750 Ti` `GTX 750` `NVS 810`                                                                                               |
 |                    | Quadro              | `K2200` `K1200` `K620` `M1200` `M520` `M5000M` `M4000M` `M3000M` `M2000M` `M1000M` `K620M` `M600M` `M500M`                     |
 
-For building locally to support older GPUs, see [developer](./development#linux-cuda-nvidia)
+For building locally to support older GPUs, see [developer.md](./development.md#linux-cuda-nvidia)
 
 ### GPU Selection
 
@@ -54,7 +54,7 @@ sudo modprobe nvidia_uvm`
 
 Ollama supports the following AMD GPUs via the ROCm library:
 
-> **NOTE:**
+> [!NOTE]
 > Additional AMD GPU support is provided by the Vulkan Library - see below.
 
 
@@ -132,9 +132,9 @@ Ollama supports GPU acceleration on Apple devices via the Metal API.
 
 ## Vulkan GPU Support
 
-> **NOTE:**
+> [!NOTE]
 > Vulkan is currently an Experimental feature.  To enable, you must set OLLAMA_VULKAN=1 for the Ollama server as
-described in the [FAQ](faq#how-do-i-configure-ollama-server)
+described in the [FAQ](faq.md#how-do-i-configure-ollama-server)
 
 Additional GPU support on Windows and Linux is provided via
 [Vulkan](https://www.vulkan.org/). On Windows most GPU vendors drivers come
@@ -161,6 +161,6 @@ sudo setcap cap_perfmon+ep /usr/local/bin/ollama
 
 To select specific Vulkan GPU(s), you can set the environment variable
 `GGML_VK_VISIBLE_DEVICES` to one or more numeric IDs on the Ollama server as
-described in the [FAQ](faq#how-do-i-configure-ollama-server). If you
+described in the [FAQ](faq.md#how-do-i-configure-ollama-server). If you
 encounter any problems with Vulkan based GPUs, you can disable all Vulkan GPUs
 by setting `GGML_VK_VISIBLE_DEVICES=-1` 

docs/modelfile.mdx

@@ -41,7 +41,6 @@ INSTRUCTION arguments
 | [`ADAPTER`](#adapter)               | Defines the (Q)LoRA adapters to apply to the model.            |
 | [`LICENSE`](#license)               | Specifies the legal license.                                   |
 | [`MESSAGE`](#message)               | Specify message history.                                       |
-| [`REQUIRES`](#requires)             | Specify the minimum version of Ollama required by the model.   |
 
 ## Examples
 
@@ -249,16 +248,6 @@ MESSAGE user Is Ontario in Canada?
 MESSAGE assistant yes
 ```
 
-### REQUIRES
-
-The `REQUIRES` instruction allows you to specify the minimum version of Ollama required by the model.
-
-```
-REQUIRES <version>
-```
-
-The version should be a valid Ollama version (e.g. 0.14.0).
-
 ## Notes
 
 - the **`Modelfile` is not case sensitive**. In the examples, uppercase instructions are used to make it easier to distinguish it from arguments.

docs/tools/extract-examples/README.md

@@ -1,46 +0,0 @@
-# extract-examples
-
-Extracts code examples from MDX files to a temp directory so you can run them.
-
-## Usage
-
-```shell
-go run docs/tools/extract-examples/main.go <mdx-file>
-```
-
-## Example
-
-```shell
-go run docs/tools/extract-examples/main.go docs/api/openai-compatibility.mdx
-```
-
-Output:
-
-```
-Extracting code examples to: /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
-
-  - 01_basic.py
-  - 01_basic.js
-  - 01_basic.sh
-  - 02_responses.py
-  - 02_responses.js
-  - 02_responses.sh
-  - 03_vision.py
-  - 03_vision.js
-  - 03_vision.sh
-
-Extracted 9 file(s) to /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
-
-To run examples:
-
-  cd /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
-  npm install   # for JS examples
-
-then run individual files with `node file.js`, `python file.py`, `bash file.sh`
-```
-
-## How it works
-
-- Parses MDX files looking for fenced code blocks with filenames (e.g., ` ```python basic.py `)
-- Groups examples by their `<CodeGroup>` and prefixes filenames with `01_`, `02_`, etc.
-- Writes all extracted files to a temp directory

docs/tools/extract-examples/main.go

@@ -1,137 +0,0 @@
-package main
-
-import (
-	"bufio"
-	"fmt"
-	"os"
-	"path/filepath"
-	"regexp"
-	"strings"
-)
-
-func main() {
-	if len(os.Args) < 2 {
-		fmt.Fprintln(os.Stderr, "Usage: go run extract-examples.go <mdx-file>")
-		os.Exit(1)
-	}
-
-	mdxFile := os.Args[1]
-
-	f, err := os.Open(mdxFile)
-	if err != nil {
-		fmt.Fprintf(os.Stderr, "Error: %v\n", err)
-		os.Exit(1)
-	}
-	defer f.Close()
-
-	// Create temp directory
-	tempDir, err := os.MkdirTemp("", "mdx-examples-*")
-	if err != nil {
-		fmt.Fprintf(os.Stderr, "Error creating temp dir: %v\n", err)
-		os.Exit(1)
-	}
-
-	fmt.Printf("Extracting code examples to: %s\n\n", tempDir)
-
-	// Patterns
-	codeBlockStart := regexp.MustCompile("^```([a-zA-Z0-9_-]+)\\s+([^\\s]+)$")
-	codeGroupStart := regexp.MustCompile("^<CodeGroup")
-	codeGroupEnd := regexp.MustCompile("^</CodeGroup>")
-
-	scanner := bufio.NewScanner(f)
-	inCodeBlock := false
-	inCodeGroup := false
-	var currentFile string
-	var content strings.Builder
-	count := 0
-	codeGroupNum := 0
-
-	for scanner.Scan() {
-		line := scanner.Text()
-
-		// Track CodeGroup boundaries
-		if codeGroupStart.MatchString(line) {
-			inCodeGroup = true
-			codeGroupNum++
-			continue
-		}
-		if codeGroupEnd.MatchString(line) {
-			inCodeGroup = false
-			continue
-		}
-
-		if inCodeBlock {
-			if line == "```" {
-				// End of code block - write file
-				if currentFile != "" {
-					outPath := filepath.Join(tempDir, currentFile)
-					if err := os.WriteFile(outPath, []byte(content.String()), 0o644); err != nil {
-						fmt.Fprintf(os.Stderr, "Error writing %s: %v\n", currentFile, err)
-					} else {
-						fmt.Printf("  - %s\n", currentFile)
-						count++
-					}
-				}
-				inCodeBlock = false
-				currentFile = ""
-				content.Reset()
-			} else {
-				content.WriteString(line)
-				content.WriteString("\n")
-			}
-		} else {
-			if matches := codeBlockStart.FindStringSubmatch(line); matches != nil {
-				inCodeBlock = true
-				filename := matches[2]
-				// Prefix with CodeGroup number if inside a CodeGroup
-				if inCodeGroup {
-					currentFile = fmt.Sprintf("%02d_%s", codeGroupNum, filename)
-				} else {
-					currentFile = filename
-				}
-				content.Reset()
-			}
-		}
-	}
-
-	if err := scanner.Err(); err != nil {
-		fmt.Fprintf(os.Stderr, "Error reading file: %v\n", err)
-		os.Exit(1)
-	}
-
-	// Write package.json for JavaScript dependencies
-	packageJSON := `{
-  "name": "mdx-examples",
-  "type": "module",
-  "dependencies": {
-    "openai": "^4",
-    "ollama": "^0.5"
-  }
-}
-`
-	if err := os.WriteFile(filepath.Join(tempDir, "package.json"), []byte(packageJSON), 0o644); err != nil {
-		fmt.Fprintf(os.Stderr, "Error writing package.json: %v\n", err)
-	}
-
-	// Write pyproject.toml for Python dependencies
-	pyprojectTOML := `[project]
-name = "mdx-examples"
-version = "0.0.0"
-dependencies = [
-    "openai",
-    "ollama",
-]
-`
-	if err := os.WriteFile(filepath.Join(tempDir, "pyproject.toml"), []byte(pyprojectTOML), 0o644); err != nil {
-		fmt.Fprintf(os.Stderr, "Error writing pyproject.toml: %v\n", err)
-	}
-
-	fmt.Printf("\n")
-	fmt.Printf("Extracted %d file(s) to %s\n", count, tempDir)
-	fmt.Printf("\n")
-	fmt.Printf("To run examples:\n")
-	fmt.Printf("\n")
-	fmt.Printf("  cd %s\n  npm install   # for JS examples\n", tempDir)
-	fmt.Printf("\n")
-	fmt.Printf("then run individual files with `node file.js`, `python file.py`, `bash file.sh`\n")
-}

docs/troubleshooting.mdx

@@ -87,7 +87,7 @@ When Ollama starts up, it takes inventory of the GPUs present in the system to d
 
 ### Linux NVIDIA Troubleshooting
 
-If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker](./docker)
+If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
 
 Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
 

fs/config.go

@@ -1,7 +1,5 @@
 package fs
 
-import "iter"
-
 type Config interface {
 	Architecture() string
 	String(string, ...string) string
@@ -13,8 +11,4 @@ type Config interface {
 	Ints(string, ...[]int32) []int32
 	Floats(string, ...[]float32) []float32
 	Bools(string, ...[]bool) []bool
-
-	Len() int
-	Keys() iter.Seq[string]
-	Value(key string) any
 }

fs/ggml/ggml.go

@@ -6,16 +6,13 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"iter"
 	"log/slog"
-	"maps"
 	"math"
 	"slices"
 	"strings"
 
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/util/bufioutil"
-	"github.com/ollama/ollama/ml"
 )
 
 type GGML struct {
@@ -241,34 +238,20 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
 	return val.values
 }
 
-func (kv KV) Len() int {
-	return len(kv)
-}
-
-func (kv KV) Keys() iter.Seq[string] {
-	return maps.Keys(kv)
-}
-
-func (kv KV) Value(key string) any {
-	return kv[key]
-}
-
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
-		"bert",
-		"deepseek2",
-		"deepseekocr",
 		"gemma3",
 		"gemma3n",
 		"gptoss", "gpt-oss",
 		"llama4",
 		"mistral3",
 		"mllama",
-		"nomic-bert",
-		"olmo3",
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
+		"deepseekocr",
+		"deepseek2",
+		"nomic-bert",
 	}, kv.Architecture())
 }
 
@@ -567,7 +550,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	}, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention ml.FlashAttentionType) (kv []uint64, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention bool) (kv []uint64, partialOffload, fullOffload uint64) {
 	context *= uint64(numParallel)
 
 	embedding := f.KV().EmbeddingLength()
@@ -808,7 +791,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 		}
 
 		partialOffload = 2 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
-		if useFlashAttention == ml.FlashAttentionEnabled {
+		if useFlashAttention {
 			// rough estimate of graph size with flash attention on
 			partialOffload = (4*uint64(numParallel) + context>>10 + 110) * format.MebiByte
 		}
@@ -826,14 +809,6 @@ func (f GGML) SupportsKVCacheType(cacheType string) bool {
 	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }
 
-// KVCacheTypeIsQuantized checks if the requested cache type is a quantized type
-func (f GGML) KVCacheTypeIsQuantized(cacheType string) bool {
-	if cacheType == "" || cacheType == "f16" || cacheType == "f32" || cacheType == "bf16" {
-		return false
-	}
-	return true
-}
-
 // SupportsFlashAttention checks if the model supports flash attention
 func (f GGML) SupportsFlashAttention() bool {
 	_, isEmbedding := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]
@@ -854,11 +829,8 @@ func (f GGML) SupportsFlashAttention() bool {
 // FlashAttention checks if the model should enable flash attention
 func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
-		"bert",
 		"gemma3",
 		"gptoss", "gpt-oss",
-		"mistral3",
-		"olmo3",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
 	}, f.KV().String("general.architecture"))

fs/ggml/gguf.go

@@ -8,12 +8,12 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"maps"
 	"os"
 	"runtime"
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/fs"
 	"golang.org/x/sync/errgroup"
 )
 
@@ -508,7 +508,7 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func WriteGGUF(f *os.File, kv fs.Config, ts []*Tensor) error {
+func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 	arch := kv.String("general.architecture")
 	if arch == "" {
 		return fmt.Errorf("architecture not set")
@@ -526,12 +526,12 @@ func WriteGGUF(f *os.File, kv fs.Config, ts []*Tensor) error {
 		return err
 	}
 
-	if err := binary.Write(f, binary.LittleEndian, uint64(kv.Len())); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(len(kv))); err != nil {
 		return err
 	}
 
-	for _, key := range slices.Sorted(kv.Keys()) {
-		if err := ggufWriteKV(f, arch, key, kv.Value(key)); err != nil {
+	for _, key := range slices.Sorted(maps.Keys(kv)) {
+		if err := ggufWriteKV(f, arch, key, kv[key]); err != nil {
 			return err
 		}
 	}
@@ -597,10 +597,6 @@ func ggufWriteKV(ws io.WriteSeeker, arch, k string, v any) error {
 
 	var err error
 	switch v := v.(type) {
-	case int32:
-		err = writeGGUF(ws, ggufTypeInt32, v)
-	case int64:
-		err = writeGGUF(ws, ggufTypeInt64, v)
 	case uint32, FileType:
 		err = writeGGUF(ws, ggufTypeUint32, v)
 	case uint64:
@@ -615,10 +611,6 @@ func ggufWriteKV(ws io.WriteSeeker, arch, k string, v any) error {
 		err = writeGGUFArray(ws, ggufTypeInt32, v)
 	case *array[int32]:
 		err = writeGGUFArray(ws, ggufTypeInt32, v.values)
-	case []int64:
-		err = writeGGUFArray(ws, ggufTypeInt64, v)
-	case *array[int64]:
-		err = writeGGUFArray(ws, ggufTypeInt64, v.values)
 	case []uint32:
 		err = writeGGUFArray(ws, ggufTypeUint32, v)
 	case *array[uint32]:

fs/ggml/gguf_test.go

@@ -42,10 +42,6 @@ func TestWriteGGUF(t *testing.T) {
 				"general.architecture": "test",
 				"general.alignment":    uint32(16),
 				"test.key":             "value",
-				"test.int32_key":       int32(-42),
-				"test.int64_key":       int64(-9223372036854775808),
-				"test.int32_array":     []int32{-1, 0, 1, 2147483647, -2147483648},
-				"test.int64_array":     []int64{-1, 0, 1, 9223372036854775807, -9223372036854775808},
 				"attention.key":        "value2",
 				"tokenizer.key":        "value3",
 				"adapter.key":          "value4",
@@ -59,7 +55,7 @@ func TestWriteGGUF(t *testing.T) {
 			}
 			defer r.Close()
 
-			ff, err := Decode(r, -1)
+			ff, err := Decode(r, 0)
 			if err != nil {
 				t.Fatal(err)
 			}
@@ -69,19 +65,15 @@ func TestWriteGGUF(t *testing.T) {
 				"general.alignment":       uint32(16),
 				"general.parameter_count": uint64(54),
 				"test.key":                "value",
-				"test.int32_key":          int32(-42),
-				"test.int64_key":          int64(-9223372036854775808),
-				"test.int32_array":        &array[int32]{size: 5, values: []int32{-1, 0, 1, 2147483647, -2147483648}},
-				"test.int64_array":        &array[int64]{size: 5, values: []int64{-1, 0, 1, 9223372036854775807, -9223372036854775808}},
 				"test.attention.key":      "value2",
 				"tokenizer.key":           "value3",
 				"adapter.key":             "value4",
-			}, ff.KV(), cmp.AllowUnexported(array[int32]{}, array[int64]{})); diff != "" {
+			}, ff.KV()); diff != "" {
 				t.Errorf("Mismatch (-want +got):\n%s", diff)
 			}
 
 			if diff := cmp.Diff(Tensors{
-				Offset: 992,
+				Offset: 800,
 				items: []*Tensor{
 					{Name: "blk.0.attn_k.weight", Offset: 0, Shape: []uint64{2, 3}},
 					{Name: "blk.0.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},

go.mod

@@ -15,8 +15,8 @@ require (
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
-	golang.org/x/sync v0.17.0
-	golang.org/x/sys v0.37.0
+	golang.org/x/sync v0.12.0
+	golang.org/x/sys v0.36.0
 )
 
 require (
@@ -28,17 +28,13 @@ require (
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	github.com/tkrajina/typescriptify-golang-structs v0.2.0
-	github.com/wk8/go-ordered-map/v2 v2.1.8
 	golang.org/x/image v0.22.0
-	golang.org/x/mod v0.30.0
-	golang.org/x/tools v0.38.0
+	golang.org/x/tools v0.30.0
 	gonum.org/v1/gonum v0.15.0
 )
 
 require (
 	github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 // indirect
-	github.com/bahlo/generic-list-go v0.2.0 // indirect
-	github.com/buger/jsonparser v1.1.1 // indirect
 	github.com/bytedance/sonic/loader v0.1.1 // indirect
 	github.com/chewxy/hm v1.0.0 // indirect
 	github.com/chewxy/math32 v1.11.0 // indirect
@@ -48,7 +44,6 @@ require (
 	github.com/gogo/protobuf v1.3.2 // indirect
 	github.com/google/flatbuffers v24.3.25+incompatible // indirect
 	github.com/kr/text v0.2.0 // indirect
-	github.com/mailru/easyjson v0.7.7 // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/pmezard/go-difflib v1.0.0 // indirect
 	github.com/rivo/uniseg v0.2.0 // indirect
@@ -81,11 +76,11 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/crypto v0.43.0
+	golang.org/x/crypto v0.36.0
 	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa // indirect
-	golang.org/x/net v0.46.0 // indirect
-	golang.org/x/term v0.36.0
-	golang.org/x/text v0.30.0
+	golang.org/x/net v0.38.0 // indirect
+	golang.org/x/term v0.30.0
+	golang.org/x/text v0.23.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -14,11 +14,7 @@ github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 h1:q4dksr6IC
 github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40/go.mod h1:Q7yQnSMnLvcXlZ8RV+jwz/6y1rQTqbX6C82SndT52Zs=
 github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0 h1:jfIu9sQUG6Ig+0+Ap1h4unLjW6YQJpKZVmUzxsD4E/Q=
 github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0/go.mod h1:t2tdKJDJF9BV14lnkjHmOQgcvEKgtqs5a1N3LNdJhGE=
-github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=
-github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=
 github.com/boombuler/barcode v1.0.0/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
-github.com/buger/jsonparser v1.1.1 h1:2PnMjfWD7wBILjqQbt530v576A/cAbQvEW9gGIpYMUs=
-github.com/buger/jsonparser v1.1.1/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=
 github.com/bytedance/sonic v1.11.6 h1:oUp34TzMlL+OY1OUWxHqsdkgC/Zfc85zGqw9siXjrc0=
 github.com/bytedance/sonic v1.11.6/go.mod h1:LysEHSvpvDySVdC2f87zGWf6CIKJcAvqab1ZaiQtds4=
 github.com/bytedance/sonic/loader v0.1.1 h1:c+e5Pt1k/cy5wMveRDyk2X4B9hF4g7an8N3zCYjJFNM=
@@ -127,7 +123,6 @@ github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+
 github.com/grpc-ecosystem/grpc-gateway v1.16.0/go.mod h1:BDjrQk3hbvj6Nolgz8mAMFbcEtjT1g+wF4CSlocrBnw=
 github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
 github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
-github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y=
 github.com/json-iterator/go v1.1.12 h1:PV8peI4a0ysnczrg+LtxykD8LfKY9ML6u2jnxaEnrnM=
 github.com/json-iterator/go v1.1.12/go.mod h1:e30LSqwooZae/UwlEbR2852Gd8hjQvJoHmT4TnhNGBo=
 github.com/jung-kurt/gofpdf v1.0.0/go.mod h1:7Id9E/uU8ce6rXgefFLlgrJj/GYY22cpxn+r32jIOes=
@@ -148,8 +143,6 @@ github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728 h1:QwWKgMY28TAXaDl+
 github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728/go.mod h1:1fEHWurg7pvf5SG6XNE5Q8UZmOwex51Mkx3SLhrW5B4=
 github.com/leodido/go-urn v1.4.0 h1:WT9HwE9SGECu3lg4d/dIA+jxlljEa1/ffXKmRjqdmIQ=
 github.com/leodido/go-urn v1.4.0/go.mod h1:bvxc+MVxLKB4z00jd1z+Dvzr47oO32F/QSNjSBOlFxI=
-github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0=
-github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=
 github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/mattn/go-runewidth v0.0.9/go.mod h1:H031xJmbD/WCDINGzjvQ9THkh0rPKHF+m2gUSrubnMI=
@@ -214,8 +207,6 @@ github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
 github.com/ugorji/go/codec v1.2.12 h1:9LC83zGrHhuUA9l16C9AHXAqEV/2wBQ4nkvumAE65EE=
 github.com/ugorji/go/codec v1.2.12/go.mod h1:UNopzCgEMSXjBc6AOMqYvWC1ktqTAfzJZUZgYf6w6lg=
-github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/fJgbpc=
-github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=
 github.com/x448/float16 v0.8.4/go.mod h1:14CWIYCyZA/cWjXOioeEpHeN/83MdbZDRQHoFcYsOfg=
 github.com/xtgo/set v1.0.0 h1:6BCNBRv3ORNDQ7fyoJXRv+tstJz3m1JVFQErfeZz2pY=
@@ -233,8 +224,8 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
-golang.org/x/crypto v0.43.0 h1:dduJYIi3A3KOfdGOHX8AVZ/jGiyPa3IbBozJ5kNuE04=
-golang.org/x/crypto v0.43.0/go.mod h1:BFbav4mRNlXJL4wNeejLpWxB7wMbc79PdRGhWKncxR0=
+golang.org/x/crypto v0.36.0 h1:AnAEvhDddvBdpY+uR+MyHmuZzzNqXSe/GvuDeob5L34=
+golang.org/x/crypto v0.36.0/go.mod h1:Y4J0ReaxCR1IMaabaSMugxJES1EpwhBHhv2bDHklZvc=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
@@ -264,8 +255,6 @@ golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzB
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.4.2/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
-golang.org/x/mod v0.30.0 h1:fDEXFVZ/fmCKProc/yAXXUijritrDzahmwwefnjoPFk=
-golang.org/x/mod v0.30.0/go.mod h1:lAsf5O2EvJeSFMiBxXDki7sCgAxEUcZHXoXMKT4GJKc=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20180826012351-8a410e7b638d/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
@@ -278,8 +267,8 @@ golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81R
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
-golang.org/x/net v0.46.0 h1:giFlY12I07fugqwPuWJi68oOnpfqFnJIJzaIIm2JVV4=
-golang.org/x/net v0.46.0/go.mod h1:Q9BGdFy1y4nkUwiLvT5qtyhAnEHgnQ/zd8PfU6nc210=
+golang.org/x/net v0.38.0 h1:vRMAPTMaeGqVhG5QyLJHqNDwecKTomGeqbnfZyKlBI8=
+golang.org/x/net v0.38.0/go.mod h1:ivrbrMbzFq5J41QOQh0siUuly180yBYtLp+CKbEaFx8=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -289,8 +278,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.17.0 h1:l60nONMj9l5drqw6jlhIELNv9I0A4OFgRsG9k2oT9Ug=
-golang.org/x/sync v0.17.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
+golang.org/x/sync v0.12.0 h1:MHc5BpPuC30uJk597Ri8TV3CNZcTLu6B6z4lJy+g6Jw=
+golang.org/x/sync v0.12.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -306,17 +295,17 @@ golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.37.0 h1:fdNQudmxPjkdUTPnLn5mdQv7Zwvbvpaxqs831goi9kQ=
-golang.org/x/sys v0.37.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+golang.org/x/sys v0.36.0 h1:KVRy2GtZBrk1cBYA7MKu5bEZFxQk4NIDV6RLVcC8o0k=
+golang.org/x/sys v0.36.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.36.0 h1:zMPR+aF8gfksFprF/Nc/rd1wRS1EI6nDBGyWAvDzx2Q=
-golang.org/x/term v0.36.0/go.mod h1:Qu394IJq6V6dCBRgwqshf3mPF85AqzYEzofzRdZkWss=
+golang.org/x/term v0.30.0 h1:PQ39fJZ+mfadBm0y5WlL4vlM7Sx1Hgf13sMIY2+QS9Y=
+golang.org/x/term v0.30.0/go.mod h1:NYYFdzHoI5wRh/h5tDMdMqCqPJZEuNqVR5xJLd/n67g=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.30.0 h1:yznKA/E9zq54KzlzBEAWn1NXSQ8DIp/NYMy88xJjl4k=
-golang.org/x/text v0.30.0/go.mod h1:yDdHFIX9t+tORqspjENWgzaCVXgk0yYnYuSZ8UzzBVM=
+golang.org/x/text v0.23.0 h1:D71I7dUrlY+VX0gQShAThNGHFxZ13dGLBHQLVl1mJlY=
+golang.org/x/text v0.23.0/go.mod h1:/BLNzu4aZCJ1+kcD0DNRotWKage4q2rGVAg4o22unh4=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
@@ -330,8 +319,8 @@ golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapK
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
 golang.org/x/tools v0.1.4/go.mod h1:o0xws9oXOQQZyjljx8fwUC0k7L1pTE6eaCbjGeHmOkk=
-golang.org/x/tools v0.38.0 h1:Hx2Xv8hISq8Lm16jvBZ2VQf+RLmbd7wVUsALibYI/IQ=
-golang.org/x/tools v0.38.0/go.mod h1:yEsQ/d/YK8cjh0L6rZlY8tgtlKiBNTL14pGDJPJpYQs=
+golang.org/x/tools v0.30.0 h1:BgcpHewrV5AUp2G9MebG4XPFI1E2W41zU1SaqVA9vJY=
+golang.org/x/tools v0.30.0/go.mod h1:c347cR/OJfw5TI+GfX7RUPNMdDRRbjvYTS0jPyvsVtY=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

integration/embed_test.go

@@ -4,9 +4,7 @@ package integration
 
 import (
 	"context"
-	"errors"
 	"math"
-	"strings"
 	"testing"
 	"time"
 
@@ -206,8 +204,8 @@ func TestAllMiniLMEmbed(t *testing.T) {
 		t.Fatalf("expected %v, got %v (similarity: %f)", expected[0:5], res.Embeddings[0][0:5], sim)
 	}
 
-	if res.PromptEvalCount != 8 {
-		t.Fatalf("expected 8 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 6 {
+		t.Fatalf("expected 6 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 
@@ -253,8 +251,8 @@ func TestAllMiniLMBatchEmbed(t *testing.T) {
 		t.Fatalf("expected %v, got %v (similarity: %f)", expected[1][0:5], res.Embeddings[1][0:5], sim)
 	}
 
-	if res.PromptEvalCount != 16 {
-		t.Fatalf("expected 16 prompt tokens, got %d", res.PromptEvalCount)
+	if res.PromptEvalCount != 12 {
+		t.Fatalf("expected 12 prompt tokens, got %d", res.PromptEvalCount)
 	}
 }
 
@@ -277,7 +275,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 	cases := []struct {
 		name    string
 		request api.EmbedRequest
-		check   func(*testing.T, *api.EmbedResponse, error)
+		check   func(*api.EmbedResponse, error)
 	}{
 		{
 			name: "target truncation",
@@ -285,7 +283,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Model: "all-minilm",
 				Input: "why",
 			},
-			check: func(t *testing.T, got *api.EmbedResponse, err error) {
+			check: func(got *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
@@ -302,11 +300,10 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Input:   "why is the sky blue?",
 				Options: map[string]any{"num_ctx": 3},
 			},
-			check: func(t *testing.T, got *api.EmbedResponse, err error) {
+			check: func(got *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
-				t.Logf("PromptEvalCount: want=%d got=%d", want.PromptEvalCount, got.PromptEvalCount)
 				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
 					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
 				}
@@ -320,11 +317,10 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Truncate: &truncTrue,
 				Options:  map[string]any{"num_ctx": 3},
 			},
-			check: func(t *testing.T, got *api.EmbedResponse, err error) {
+			check: func(got *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
-				t.Logf("PromptEvalCount: want=%d got=%d", want.PromptEvalCount, got.PromptEvalCount)
 				if diff := cmp.Diff(want.Embeddings[0], got.Embeddings[0]); diff != "" {
 					t.Errorf("embedding mismatch (-want +got):\n%s", diff)
 				}
@@ -338,21 +334,21 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Truncate: &truncFalse,
 				Options:  map[string]any{"num_ctx": 3},
 			},
-			check: func(t *testing.T, res *api.EmbedResponse, err error) {
-				if err.Error() != "the input length exceeds the context length" {
+			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 with context length of 1",
+			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": 1},
 			},
-			check: func(t *testing.T, res *api.EmbedResponse, err error) {
+			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)
 				}
@@ -366,7 +362,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Truncate: &truncTrue,
 				Options:  map[string]any{"num_ctx": 0},
 			},
-			check: func(t *testing.T, res *api.EmbedResponse, err error) {
+			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)
 				}
@@ -379,7 +375,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 				Input:   "why is the sky blue? Why is the sky blue? hi there my",
 				Options: map[string]any{"num_ctx": 16},
 			},
-			check: func(t *testing.T, res *api.EmbedResponse, err error) {
+			check: func(res *api.EmbedResponse, err error) {
 				if err != nil {
 					t.Fatal(err)
 				}
@@ -389,8 +385,7 @@ func TestAllMiniLMEmbedTruncate(t *testing.T) {
 
 	for _, req := range cases {
 		t.Run(req.name, func(t *testing.T) {
-			resp, err := embedTestHelper(ctx, client, t, req.request)
-			req.check(t, resp, err)
+			req.check(embedTestHelper(ctx, client, t, req.request))
 		})
 	}
 }
@@ -414,230 +409,3 @@ func embedTestHelper(ctx context.Context, client *api.Client, t *testing.T, req
 
 	return client.Embed(ctx, &req)
 }
-
-func TestEmbedTruncation(t *testing.T) {
-	// Use test deadline if set, otherwise default to 2 minutes
-	timeout := 2 * time.Minute
-	if deadline, ok := t.Deadline(); ok {
-		timeout = time.Until(deadline) - 10*time.Second // Reserve 10s buffer
-	}
-	ctx, cancel := context.WithTimeout(context.Background(), timeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	for _, model := range libraryEmbedModels {
-		model := model
-		t.Run(model, func(t *testing.T) {
-			// Check if we're running out of time (reserve 20s for current model)
-			if deadline, ok := t.Deadline(); ok && time.Until(deadline) < 20*time.Second {
-				t.Skip("skipping remaining tests to avoid timeout")
-			}
-
-			// Give each model its own budget to account for first-time pulls/loads
-			mctx, mcancel := context.WithTimeout(ctx, 3*time.Minute)
-			defer mcancel()
-
-			t.Run("truncation batch", func(t *testing.T) {
-				truncTrue := true
-				req := api.EmbedRequest{
-					Model:    model,
-					Input:    []string{"short", strings.Repeat("long ", 100), "medium text"},
-					Truncate: &truncTrue,
-					Options:  map[string]any{"num_ctx": 30},
-				}
-
-				res, err := embedTestHelper(mctx, client, t, req)
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				if len(res.Embeddings) != 3 {
-					t.Fatalf("expected 3 embeddings, got %d", len(res.Embeddings))
-				}
-
-				if res.PromptEvalCount > 90 {
-					t.Fatalf("expected tokens <= 90 (3 × 30 max), got %d", res.PromptEvalCount)
-				}
-			})
-
-			t.Run("runner token count accuracy", func(t *testing.T) {
-				baseline := api.EmbedRequest{Model: model, Input: "test"}
-				baseRes, err := embedTestHelper(mctx, client, t, baseline)
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				batch := api.EmbedRequest{
-					Model: model,
-					Input: []string{"test", "test", "test"},
-				}
-				batchRes, err := embedTestHelper(mctx, client, t, batch)
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				expectedCount := baseRes.PromptEvalCount * 3
-				if batchRes.PromptEvalCount < expectedCount-2 || batchRes.PromptEvalCount > expectedCount+2 {
-					t.Fatalf("expected ~%d tokens (3 × %d), got %d",
-						expectedCount, baseRes.PromptEvalCount, batchRes.PromptEvalCount)
-				}
-			})
-		})
-	}
-}
-
-// TestEmbedLargeInput tests that embedding models can handle large inputs that would exceed typical batch sizes.
-func TestEmbedLargeInput(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	for _, model := range libraryEmbedModels {
-		model := model
-		t.Run(model, func(t *testing.T) {
-			mctx, mcancel := context.WithTimeout(ctx, 2*time.Minute)
-			defer mcancel()
-
-			// Test with progressively larger inputs
-			testCases := []struct {
-				name       string
-				inputWords int
-			}{
-				{"medium_input_256_words", 256},
-				{"large_input_512_words", 512},
-				{"very_large_input_800_words", 800},
-			}
-
-			for _, tc := range testCases {
-				t.Run(tc.name, func(t *testing.T) {
-					words := make([]string, tc.inputWords)
-					for i := range words {
-						words[i] = "word"
-					}
-					input := strings.Join(words, " ")
-
-					req := api.EmbedRequest{
-						Model:     model,
-						Input:     input,
-						KeepAlive: &api.Duration{Duration: 30 * time.Second},
-					}
-
-					res, err := embedTestHelper(mctx, client, t, req)
-					if err != nil {
-						t.Fatalf("embedding failed for %d words: %v", tc.inputWords, err)
-					}
-
-					if len(res.Embeddings) != 1 {
-						t.Fatalf("expected 1 embedding, got %d", len(res.Embeddings))
-					}
-
-					if len(res.Embeddings[0]) == 0 {
-						t.Fatal("expected non-empty embedding")
-					}
-
-					t.Logf("Successfully embedded %d words (%d tokens)", tc.inputWords, res.PromptEvalCount)
-				})
-			}
-		})
-	}
-}
-
-// TestEmbedStatusCode tests that errors from the embedding endpoint
-// properly preserve their HTTP status codes when returned to the client.
-// This test specifically checks the error handling path in EmbedHandler
-// where api.StatusError errors should maintain their original status code.
-func TestEmbedStatusCode(t *testing.T) {
-	// Use test deadline if set, otherwise default to 2 minutes
-	timeout := 2 * time.Minute
-	if deadline, ok := t.Deadline(); ok {
-		timeout = time.Until(deadline) - 10*time.Second // Reserve 10s buffer
-	}
-	ctx, cancel := context.WithTimeout(context.Background(), timeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	for _, model := range libraryEmbedModels {
-		model := model
-		t.Run(model, func(t *testing.T) {
-			// Check if we're running out of time (reserve 20s for current model)
-			if deadline, ok := t.Deadline(); ok && time.Until(deadline) < 20*time.Second {
-				t.Skip("skipping remaining tests to avoid timeout")
-			}
-
-			mctx, mcancel := context.WithTimeout(ctx, 3*time.Minute)
-			defer mcancel()
-
-			// Pull the model if needed
-			if err := PullIfMissing(mctx, client, model); err != nil {
-				t.Fatal(err)
-			}
-
-			t.Run("truncation error status code", func(t *testing.T) {
-				truncFalse := false
-				longInput := strings.Repeat("word ", 100)
-
-				req := api.EmbedRequest{
-					Model:    model,
-					Input:    longInput,
-					Truncate: &truncFalse,
-					Options:  map[string]any{"num_ctx": 10},
-				}
-
-				_, err := embedTestHelper(mctx, client, t, req)
-				if err == nil {
-					t.Fatal("expected error when truncate=false with long input")
-				}
-
-				// Check that it's a StatusError with the correct status code
-				var statusErr api.StatusError
-				if !errors.As(err, &statusErr) {
-					t.Fatalf("expected api.StatusError, got %T: %v", err, err)
-				}
-
-				// The error should be a 4xx client error (likely 400 Bad Request)
-				// not a 500 Internal Server Error
-				if statusErr.StatusCode < 400 || statusErr.StatusCode >= 500 {
-					t.Errorf("expected 4xx status code, got %d", statusErr.StatusCode)
-				}
-
-				// Verify the error message is meaningful
-				if !strings.Contains(err.Error(), "context length") {
-					t.Errorf("expected error message to mention context length, got: %v", err)
-				}
-			})
-
-			t.Run("batch truncation error status code", func(t *testing.T) {
-				truncFalse := false
-				req := api.EmbedRequest{
-					Model: model,
-					Input: []string{
-						"short input",
-						strings.Repeat("very long input ", 100),
-						"another short input",
-					},
-					Truncate: &truncFalse,
-					Options:  map[string]any{"num_ctx": 10},
-				}
-
-				_, err := embedTestHelper(mctx, client, t, req)
-				if err == nil {
-					t.Fatal("expected error when one input exceeds context with truncate=false")
-				}
-
-				// Check that it's a StatusError with the correct status code
-				var statusErr api.StatusError
-				if !errors.As(err, &statusErr) {
-					t.Fatalf("expected api.StatusError, got %T: %v", err, err)
-				}
-
-				// The error should be a 4xx client error, not a 500 Internal Server Error
-				if statusErr.StatusCode < 400 || statusErr.StatusCode >= 500 {
-					t.Errorf("expected 4xx status code, got %d", statusErr.StatusCode)
-				}
-			})
-		})
-	}
-}

integration/llm_image_test.go

@@ -33,9 +33,6 @@ func TestVisionModels(t *testing.T) {
 			// Qwen 3 VL mixture of experts
 			model: "qwen3-vl:30b",
 		},
-		{
-			model: "ministral-3",
-		},
 	}
 
 	for _, v := range testCases {

integration/tools_test.go

@@ -11,15 +11,6 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
-// testPropsMap creates a ToolPropertiesMap from a map (convenience function for tests)
-func testPropsMap(m map[string]api.ToolProperty) *api.ToolPropertiesMap {
-	props := api.NewToolPropertiesMap()
-	for k, v := range m {
-		props.Set(k, v)
-	}
-	return props
-}
-
 func TestAPIToolCalling(t *testing.T) {
 	initialTimeout := 60 * time.Second
 	streamTimeout := 60 * time.Second
@@ -39,7 +30,6 @@ func TestAPIToolCalling(t *testing.T) {
 		"mistral":       6,
 		"qwen2.5":       6,
 		"qwen2":         6,
-		"ministral-3":   20,
 		"mistral-nemo":  9,
 		"mistral-small": 16,
 		"mixtral:8x22b": 80,
@@ -66,12 +56,12 @@ func TestAPIToolCalling(t *testing.T) {
 						Parameters: api.ToolFunctionParameters{
 							Type:     "object",
 							Required: []string{"location"},
-							Properties: testPropsMap(map[string]api.ToolProperty{
+							Properties: map[string]api.ToolProperty{
 								"location": {
 									Type:        api.PropertyType{"string"},
 									Description: "The city and state, e.g. San Francisco, CA",
 								},
-							}),
+							},
 						},
 					},
 				},

integration/utils_test.go

@@ -38,7 +38,6 @@ var (
 
 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
-		"ministral-3",
 		"qwen3-coder:30b",
 		"gpt-oss:20b",
 		"gemma3n:e2b",
@@ -168,7 +167,6 @@ var (
 		"medllama2",
 		"megadolphin",
 		"minicpm-v",
-		"ministral-3",
 		"mistral-large",
 		"mistral-nemo",
 		"mistral-openorca",
@@ -272,7 +270,6 @@ var (
 		"mistral",
 		"qwen2.5",
 		"qwen2",
-		"ministral-3",
 		"mistral-nemo",
 		"mistral-small",
 		"mixtral:8x22b",

internal/orderedmap/orderedmap.go

@@ -1,94 +0,0 @@
-// Package orderedmap provides a generic ordered map that maintains insertion order.
-// It wraps github.com/wk8/go-ordered-map/v2 to encapsulate the dependency.
-package orderedmap
-
-import (
-	"encoding/json"
-	"iter"
-
-	orderedmap "github.com/wk8/go-ordered-map/v2"
-)
-
-// Map is a generic ordered map that maintains insertion order.
-type Map[K comparable, V any] struct {
-	om *orderedmap.OrderedMap[K, V]
-}
-
-// New creates a new empty ordered map.
-func New[K comparable, V any]() *Map[K, V] {
-	return &Map[K, V]{
-		om: orderedmap.New[K, V](),
-	}
-}
-
-// Get retrieves a value by key.
-func (m *Map[K, V]) Get(key K) (V, bool) {
-	if m == nil || m.om == nil {
-		var zero V
-		return zero, false
-	}
-	return m.om.Get(key)
-}
-
-// Set sets a key-value pair. If the key already exists, its value is updated
-// but its position in the iteration order is preserved. If the key is new,
-// it is appended to the end.
-func (m *Map[K, V]) Set(key K, value V) {
-	if m == nil {
-		return
-	}
-	if m.om == nil {
-		m.om = orderedmap.New[K, V]()
-	}
-	m.om.Set(key, value)
-}
-
-// Len returns the number of entries.
-func (m *Map[K, V]) Len() int {
-	if m == nil || m.om == nil {
-		return 0
-	}
-	return m.om.Len()
-}
-
-// All returns an iterator over all key-value pairs in insertion order.
-func (m *Map[K, V]) All() iter.Seq2[K, V] {
-	return func(yield func(K, V) bool) {
-		if m == nil || m.om == nil {
-			return
-		}
-		for pair := m.om.Oldest(); pair != nil; pair = pair.Next() {
-			if !yield(pair.Key, pair.Value) {
-				return
-			}
-		}
-	}
-}
-
-// ToMap converts to a regular Go map.
-// Note: The resulting map does not preserve order.
-func (m *Map[K, V]) ToMap() map[K]V {
-	if m == nil || m.om == nil {
-		return nil
-	}
-	result := make(map[K]V, m.om.Len())
-	for pair := m.om.Oldest(); pair != nil; pair = pair.Next() {
-		result[pair.Key] = pair.Value
-	}
-	return result
-}
-
-// MarshalJSON implements json.Marshaler. The JSON output preserves key order.
-func (m *Map[K, V]) MarshalJSON() ([]byte, error) {
-	if m == nil || m.om == nil {
-		return []byte("null"), nil
-	}
-	return json.Marshal(m.om)
-}
-
-// UnmarshalJSON implements json.Unmarshaler. The insertion order matches the
-// order of keys in the JSON input.
-func (m *Map[K, V]) UnmarshalJSON(data []byte) error {
-	m.om = orderedmap.New[K, V]()
-	return json.Unmarshal(data, &m.om)
-}

internal/orderedmap/orderedmap_test.go

@@ -1,348 +0,0 @@
-package orderedmap
-
-import (
-	"encoding/json"
-	"slices"
-	"testing"
-)
-
-func TestMap_BasicOperations(t *testing.T) {
-	m := New[string, int]()
-
-	// Test empty map
-	if m.Len() != 0 {
-		t.Errorf("expected Len() = 0, got %d", m.Len())
-	}
-	v, ok := m.Get("a")
-	if ok {
-		t.Error("expected Get on empty map to return false")
-	}
-	if v != 0 {
-		t.Errorf("expected zero value, got %d", v)
-	}
-
-	// Test Set and Get
-	m.Set("a", 1)
-	m.Set("b", 2)
-	m.Set("c", 3)
-
-	if m.Len() != 3 {
-		t.Errorf("expected Len() = 3, got %d", m.Len())
-	}
-
-	v, ok = m.Get("a")
-	if !ok || v != 1 {
-		t.Errorf("expected Get(a) = (1, true), got (%d, %v)", v, ok)
-	}
-
-	v, ok = m.Get("b")
-	if !ok || v != 2 {
-		t.Errorf("expected Get(b) = (2, true), got (%d, %v)", v, ok)
-	}
-
-	v, ok = m.Get("c")
-	if !ok || v != 3 {
-		t.Errorf("expected Get(c) = (3, true), got (%d, %v)", v, ok)
-	}
-
-	// Test updating existing key preserves position
-	m.Set("a", 10)
-	v, ok = m.Get("a")
-	if !ok || v != 10 {
-		t.Errorf("expected Get(a) = (10, true), got (%d, %v)", v, ok)
-	}
-	if m.Len() != 3 {
-		t.Errorf("expected Len() = 3 after update, got %d", m.Len())
-	}
-}
-
-func TestMap_InsertionOrderPreserved(t *testing.T) {
-	m := New[string, int]()
-
-	// Insert in non-alphabetical order
-	m.Set("z", 1)
-	m.Set("a", 2)
-	m.Set("m", 3)
-	m.Set("b", 4)
-
-	// Verify iteration order matches insertion order
-	var keys []string
-	var values []int
-	for k, v := range m.All() {
-		keys = append(keys, k)
-		values = append(values, v)
-	}
-
-	expectedKeys := []string{"z", "a", "m", "b"}
-	expectedValues := []int{1, 2, 3, 4}
-
-	if !slices.Equal(keys, expectedKeys) {
-		t.Errorf("expected keys %v, got %v", expectedKeys, keys)
-	}
-	if !slices.Equal(values, expectedValues) {
-		t.Errorf("expected values %v, got %v", expectedValues, values)
-	}
-}
-
-func TestMap_UpdatePreservesPosition(t *testing.T) {
-	m := New[string, int]()
-
-	m.Set("first", 1)
-	m.Set("second", 2)
-	m.Set("third", 3)
-
-	// Update middle element
-	m.Set("second", 20)
-
-	var keys []string
-	for k := range m.All() {
-		keys = append(keys, k)
-	}
-
-	// Order should still be first, second, third
-	expected := []string{"first", "second", "third"}
-	if !slices.Equal(keys, expected) {
-		t.Errorf("expected keys %v, got %v", expected, keys)
-	}
-}
-
-func TestMap_MarshalJSON_PreservesOrder(t *testing.T) {
-	m := New[string, int]()
-
-	// Insert in non-alphabetical order
-	m.Set("z", 1)
-	m.Set("a", 2)
-	m.Set("m", 3)
-
-	data, err := json.Marshal(m)
-	if err != nil {
-		t.Fatalf("Marshal failed: %v", err)
-	}
-
-	// JSON should preserve insertion order, not alphabetical
-	expected := `{"z":1,"a":2,"m":3}`
-	if string(data) != expected {
-		t.Errorf("expected %s, got %s", expected, string(data))
-	}
-}
-
-func TestMap_UnmarshalJSON_PreservesOrder(t *testing.T) {
-	// JSON with non-alphabetical key order
-	jsonData := `{"z":1,"a":2,"m":3}`
-
-	m := New[string, int]()
-	if err := json.Unmarshal([]byte(jsonData), m); err != nil {
-		t.Fatalf("Unmarshal failed: %v", err)
-	}
-
-	// Verify iteration order matches JSON order
-	var keys []string
-	for k := range m.All() {
-		keys = append(keys, k)
-	}
-
-	expected := []string{"z", "a", "m"}
-	if !slices.Equal(keys, expected) {
-		t.Errorf("expected keys %v, got %v", expected, keys)
-	}
-}
-
-func TestMap_JSONRoundTrip(t *testing.T) {
-	// Test that unmarshal -> marshal produces identical JSON
-	original := `{"zebra":"z","apple":"a","mango":"m","banana":"b"}`
-
-	m := New[string, string]()
-	if err := json.Unmarshal([]byte(original), m); err != nil {
-		t.Fatalf("Unmarshal failed: %v", err)
-	}
-
-	data, err := json.Marshal(m)
-	if err != nil {
-		t.Fatalf("Marshal failed: %v", err)
-	}
-
-	if string(data) != original {
-		t.Errorf("round trip failed: expected %s, got %s", original, string(data))
-	}
-}
-
-func TestMap_ToMap(t *testing.T) {
-	m := New[string, int]()
-	m.Set("a", 1)
-	m.Set("b", 2)
-
-	regular := m.ToMap()
-
-	if len(regular) != 2 {
-		t.Errorf("expected len 2, got %d", len(regular))
-	}
-	if regular["a"] != 1 {
-		t.Errorf("expected regular[a] = 1, got %d", regular["a"])
-	}
-	if regular["b"] != 2 {
-		t.Errorf("expected regular[b] = 2, got %d", regular["b"])
-	}
-}
-
-func TestMap_NilSafety(t *testing.T) {
-	var m *Map[string, int]
-
-	// All operations should be safe on nil
-	if m.Len() != 0 {
-		t.Errorf("expected Len() = 0 on nil map, got %d", m.Len())
-	}
-
-	v, ok := m.Get("a")
-	if ok {
-		t.Error("expected Get on nil map to return false")
-	}
-	if v != 0 {
-		t.Errorf("expected zero value from nil map, got %d", v)
-	}
-
-	// Set on nil is a no-op
-	m.Set("a", 1)
-	if m.Len() != 0 {
-		t.Errorf("expected Len() = 0 after Set on nil, got %d", m.Len())
-	}
-
-	// All returns empty iterator
-	var keys []string
-	for k := range m.All() {
-		keys = append(keys, k)
-	}
-	if len(keys) != 0 {
-		t.Errorf("expected empty iteration on nil map, got %v", keys)
-	}
-
-	// ToMap returns nil
-	if m.ToMap() != nil {
-		t.Error("expected ToMap to return nil on nil map")
-	}
-
-	// MarshalJSON returns null
-	data, err := json.Marshal(m)
-	if err != nil {
-		t.Fatalf("Marshal failed: %v", err)
-	}
-	if string(data) != "null" {
-		t.Errorf("expected null, got %s", string(data))
-	}
-}
-
-func TestMap_EmptyMapMarshal(t *testing.T) {
-	m := New[string, int]()
-
-	data, err := json.Marshal(m)
-	if err != nil {
-		t.Fatalf("Marshal failed: %v", err)
-	}
-	if string(data) != "{}" {
-		t.Errorf("expected {}, got %s", string(data))
-	}
-}
-
-func TestMap_NestedValues(t *testing.T) {
-	m := New[string, any]()
-	m.Set("string", "hello")
-	m.Set("number", 42)
-	m.Set("bool", true)
-	m.Set("nested", map[string]int{"x": 1})
-
-	data, err := json.Marshal(m)
-	if err != nil {
-		t.Fatalf("Marshal failed: %v", err)
-	}
-
-	expected := `{"string":"hello","number":42,"bool":true,"nested":{"x":1}}`
-	if string(data) != expected {
-		t.Errorf("expected %s, got %s", expected, string(data))
-	}
-}
-
-func TestMap_AllIteratorEarlyExit(t *testing.T) {
-	m := New[string, int]()
-	m.Set("a", 1)
-	m.Set("b", 2)
-	m.Set("c", 3)
-	m.Set("d", 4)
-
-	// Collect only first 2
-	var keys []string
-	for k := range m.All() {
-		keys = append(keys, k)
-		if len(keys) == 2 {
-			break
-		}
-	}
-
-	expected := []string{"a", "b"}
-	if !slices.Equal(keys, expected) {
-		t.Errorf("expected %v, got %v", expected, keys)
-	}
-}
-
-func TestMap_IntegerKeys(t *testing.T) {
-	m := New[int, string]()
-	m.Set(3, "three")
-	m.Set(1, "one")
-	m.Set(2, "two")
-
-	var keys []int
-	for k := range m.All() {
-		keys = append(keys, k)
-	}
-
-	// Should preserve insertion order, not numerical order
-	expected := []int{3, 1, 2}
-	if !slices.Equal(keys, expected) {
-		t.Errorf("expected %v, got %v", expected, keys)
-	}
-}
-
-func TestMap_UnmarshalIntoExisting(t *testing.T) {
-	m := New[string, int]()
-	m.Set("existing", 999)
-
-	// Unmarshal should replace contents
-	if err := json.Unmarshal([]byte(`{"new":1}`), m); err != nil {
-		t.Fatalf("Unmarshal failed: %v", err)
-	}
-
-	_, ok := m.Get("existing")
-	if ok {
-		t.Error("existing key should be gone after unmarshal")
-	}
-
-	v, ok := m.Get("new")
-	if !ok || v != 1 {
-		t.Errorf("expected Get(new) = (1, true), got (%d, %v)", v, ok)
-	}
-}
-
-func TestMap_LargeOrderPreservation(t *testing.T) {
-	m := New[string, int]()
-
-	// Create many keys in specific order
-	keys := make([]string, 100)
-	for i := range 100 {
-		keys[i] = string(rune('a' + (99 - i))) // reverse order: 'd', 'c', 'b', 'a' (extended)
-		if i >= 26 {
-			keys[i] = string(rune('A'+i-26)) + string(rune('a'+i%26))
-		}
-	}
-
-	for i, k := range keys {
-		m.Set(k, i)
-	}
-
-	// Verify order preserved
-	var resultKeys []string
-	for k := range m.All() {
-		resultKeys = append(resultKeys, k)
-	}
-
-	if !slices.Equal(keys, resultKeys) {
-		t.Error("large map should preserve insertion order")
-	}
-}

kvcache/causal.go

@@ -140,6 +140,10 @@ func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity
 		c.config.CachePadding = 1
 	}
 
+	if c.config.MaskBatchPadding == 0 {
+		c.config.MaskBatchPadding = 1
+	}
+
 	if c.config.MaskDType == ml.DTypeOther {
 		c.config.MaskDType = ml.DTypeF32
 	}
@@ -360,12 +364,15 @@ func roundUp(length, pad int) int {
 // token in the history should apply. This is based on both the sequence and causality (the
 // position of the history is not ahead of the token in the batch).
 func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
+	// Align and pad the two dimensions as required by the backend
+	batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)
+
 	c.curCellRange.min = roundDown(c.curCellRange.min, c.config.CachePadding)
 	c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1
 
 	length := c.curCellRange.max - c.curCellRange.min + 1
 
-	mask := make([]float32, c.curBatchSize*length)
+	mask := make([]float32, batchSize*length)
 
 	for i := range c.curBatchSize {
 		enabled := !slices.Contains(c.opts.Except, i)
@@ -379,7 +386,13 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 		}
 	}
 
-	maskTensor := ctx.Input().FromFloats(mask, length, c.curBatchSize)
+	// Mask out any padding tokens we added. For padding that we added to the cache history, this
+	// has already been masked out because the sequence doesn't match.
+	for i := c.curBatchSize * length; i < len(mask); i++ {
+		mask[i] = float32(math.Inf(-1))
+	}
+
+	maskTensor := ctx.Input().FromFloats(mask, length, batchSize)
 
 	if c.config.MaskDType != ml.DTypeF32 {
 		maskTensor = maskTensor.Cast(ctx, c.config.MaskDType)

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "ec98e2002";
+char const *LLAMA_COMMIT = "3cfa9c3f125763305b4226bc032f1954f08990dc";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -17,17 +17,11 @@ include /tools/mtmd/clip.cpp
 include /tools/mtmd/mtmd.cpp
 include /tools/mtmd/mtmd-audio.cpp
 include /tools/mtmd/mtmd-helper.cpp
-include /tools/mtmd/models/
-include /tools/mtmd/models/*.h
-include /tools/mtmd/models/*.cpp
 include /src/
 include /src/llama.*
 include /src/llama-*.*
 include /src/unicode-data.*
 include /src/unicode.*
-include /src/models/
-include /src/models/*.h
-include /src/models/*.cpp
 include /vendor/
 include /vendor/miniaudio/
 include /vendor/miniaudio/*.h

llama/llama.cpp/common/common.cpp

@@ -8,7 +8,6 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
-#include "sampling.h"
 
 #include <algorithm>
 #include <cinttypes>
@@ -27,6 +26,7 @@
 #include <sstream>
 #include <string>
 #include <thread>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
@@ -60,14 +60,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-common_time_meas::common_time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
-
-common_time_meas::~common_time_meas() {
-    if (t_start_us >= 0) {
-        t_acc += ggml_time_us() - t_start_us;
-    }
-}
-
 //
 // CPU utils
 //
@@ -363,7 +355,11 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
 }
 
 void common_init() {
-    llama_log_set(common_log_default_callback, NULL);
+    llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
+        if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+            common_log_add(common_log_main(), level, "%s", text);
+        }
+    }, NULL);
 
 #ifdef NDEBUG
     const char * build_type = "";
@@ -694,7 +690,7 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
 
 // Validate if a filename is safe to use
 // To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
-bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
+bool fs_validate_filename(const std::string & filename) {
     if (!filename.length()) {
         // Empty filename invalid
         return false;
@@ -754,14 +750,10 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
-            || c == ':' || c == '*' // Illegal characters
+            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
             || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
             return false;
         }
-        if (!allow_subdirs && (c == '/' || c == '\\')) {
-            // Subdirectories not allowed, reject path separators
-            return false;
-        }
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -786,29 +778,11 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
 #include <iostream>
 
 
-#ifdef _WIN32
-static std::wstring utf8_to_wstring(const std::string & str) {
-    if (str.empty()) {
-        return std::wstring();
-    }
-
-    int size = MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0);
-
-    if (size <= 0) {
-        return std::wstring();
-    }
-
-    std::wstring wstr(size, 0);
-    MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), &wstr[0], size);
-
-    return wstr;
-}
-#endif
-
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
-    std::wstring wpath = utf8_to_wstring(path);
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
 
     // if the path already exists, check whether it's a directory
     const DWORD attributes = GetFileAttributesW(wpath.c_str());
@@ -881,11 +855,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
 #endif // _WIN32
 }
 
-bool fs_is_directory(const std::string & path) {
-    std::filesystem::path dir(path);
-    return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
-}
-
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -920,8 +889,6 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
-#elif defined(__EMSCRIPTEN__)
-        GGML_ABORT("not implemented on this platform");
 #else
 #  error Unknown architecture
 #endif
@@ -941,258 +908,34 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
-std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
-    std::vector<common_file_info> files;
-    if (path.empty()) return files;
-
-    std::filesystem::path dir(path);
-    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
-        return files;
-    }
-
-    for (const auto & entry : std::filesystem::directory_iterator(dir)) {
-        try {
-            // Only include regular files (skip directories)
-            const auto & p = entry.path();
-            if (std::filesystem::is_regular_file(p)) {
-                common_file_info info;
-                info.path   = p.string();
-                info.name   = p.filename().string();
-                info.is_dir = false;
-                try {
-                    info.size = static_cast<size_t>(std::filesystem::file_size(p));
-                } catch (const std::filesystem::filesystem_error &) {
-                    info.size = 0;
-                }
-                files.push_back(std::move(info));
-            } else if (include_directories && std::filesystem::is_directory(p)) {
-                common_file_info info;
-                info.path   = p.string();
-                info.name   = p.filename().string();
-                info.size   = 0; // Directories have no size
-                info.is_dir = true;
-                files.push_back(std::move(info));
-            }
-        } catch (const std::filesystem::filesystem_error &) {
-            // skip entries we cannot inspect
-            continue;
-        }
-    }
-
-    return files;
-}
-
-//
-// TTY utils
-//
-
-bool tty_can_use_colors() {
-    // Check NO_COLOR environment variable (https://no-color.org/)
-    if (const char * no_color = std::getenv("NO_COLOR")) {
-        if (no_color[0] != '\0') {
-            return false;
-        }
-    }
-
-    // Check TERM environment variable
-    if (const char * term = std::getenv("TERM")) {
-        if (std::strcmp(term, "dumb") == 0) {
-            return false;
-        }
-    }
-
-    // Check if stdout and stderr are connected to a terminal
-    // We check both because log messages can go to either
-    bool stdout_is_tty = isatty(fileno(stdout));
-    bool stderr_is_tty = isatty(fileno(stderr));
-
-    return stdout_is_tty || stderr_is_tty;
-}
 
 //
 // Model utils
 //
 
-// TODO: move to common/sampling
-static void common_init_sampler_from_model(
-    const llama_model * model,
-    common_params_sampling & sparams) {
-
-    const uint64_t config = sparams.user_sampling_config;
-
-    auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
-        if (config & user_config) {
-            return;
-        }
-
-        char buf[64] = {0};
-        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
-            char * end = nullptr;
-            int32_t v = strtol(buf, &end, 10);
-            if (end && end != buf) {
-                dst = v;
-            }
-        }
-    };
-
-    auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
-        if (config & user_config) {
-            return;
-        }
-
-        char buf[128] = {0};
-        if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
-            char * end = nullptr;
-            float v = strtof(buf, &end);
-            if (end && end != buf) {
-                dst = v;
-            }
-        }
-    };
-
-    // Sampling sequence
-    if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
-        char buf[512] = {0};
-        if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
-            const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
-            if (!sampler_names.empty()) {
-                sparams.samplers = common_sampler_types_from_names(sampler_names, true);
-            }
-        }
-    }
-
-    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K),           sparams.top_k,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P),           sparams.top_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P),           sparams.min_p,           common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD),   sparams.xtc_threshold,   common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP),            sparams.temp,            common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
-    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N),  sparams.penalty_last_n,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT),  sparams.penalty_repeat,  common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
-    get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT),        sparams.mirostat,        common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU),    sparams.mirostat_tau,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
-    get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
-}
-
-struct common_init_result::impl {
-    impl() = default;
-    ~impl() = default;
-
-    llama_model_ptr   model;
-    llama_context_ptr context;
-
-    std::vector<llama_adapter_lora_ptr> lora;
-
-    std::vector<common_sampler_ptr> samplers;
-};
-
-common_init_result::common_init_result(common_params & params) :
-    pimpl(new impl{}) {
+struct common_init_result common_init_from_params(common_params & params) {
+    common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
-    auto cparams = common_context_params_to_llama(params);
-
-    if (params.fit_params) {
-        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\n", __func__);
-        llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
-            params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
-    }
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        return;
+        LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
+        return iparams;
     }
 
-    pimpl->model.reset(model);
-
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    // updates params.sampling
-    // TODO: fix naming
-    common_init_sampler_from_model(model, params.sampling);
-
-    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
-        params.sampling.ignore_eos = false;
-    }
-
-    // initialize once
-    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-        if (llama_vocab_is_eog(vocab, i)) {
-            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY);
-            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
-        }
-    }
-
-    if (params.sampling.ignore_eos) {
-        // add EOG biases to the active set of logit biases
-        params.sampling.logit_bias.insert(
-                params.sampling.logit_bias.end(),
-                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
-    }
-
-    //if (params.sampling.penalty_last_n == -1) {
-    //    LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
-    //}
-
-    //if (params.sampling.dry_penalty_last_n == -1) {
-    //    LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
-    //}
-
-    pimpl->samplers.resize(cparams.n_seq_max);
-
-    for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
-        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
-    }
+    auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
-        return;
+        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
+        llama_model_free(model);
+        return iparams;
     }
 
-    pimpl->context.reset(lctx);
-}
-
-llama_model * common_init_result::model() {
-    return pimpl->model.get();
-}
-
-llama_context * common_init_result::context() {
-    return pimpl->context.get();
-}
-
-common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
-    return pimpl->samplers[seq_id].get();
-}
-
-std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
-    return pimpl->lora;
-}
-
-void common_init_result::free_context() {
-    pimpl->context.reset();
-}
-
-common_init_result_ptr common_init_from_params(common_params & params) {
-    common_init_result_ptr res(new common_init_result(params));
-
-    llama_model * model = res->model();
-    if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
-        return res;
-    }
-
-    llama_context * lctx = res->context();
-    if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
-        return res;
-    }
-
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
@@ -1204,7 +947,10 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
-            return res;
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
         }
 
         int err = llama_apply_adapter_cvec(
@@ -1215,7 +961,10 @@ common_init_result_ptr common_init_from_params(common_params & params) {
                 params.control_vector_layer_start,
                 params.control_vector_layer_end);
         if (err) {
-            return res;
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
         }
     }
 
@@ -1239,7 +988,10 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         }
 
         if (!ok) {
-            return res;
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
         }
     }
 
@@ -1249,7 +1001,9 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
         if (lora == nullptr) {
             LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
-            return res;
+            llama_free(lctx);
+            llama_model_free(model);
+            return iparams;
         }
 
         char buf[1024];
@@ -1258,13 +1012,43 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         la.task_name = buf;
         llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
         la.prompt_prefix = buf;
-        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
+        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
 
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
 
+    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
+        params.sampling.ignore_eos = false;
+    }
+
+    // initialize once
+    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+        if (llama_vocab_is_eog(vocab, i)) {
+            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
+            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
+        }
+    }
+
+    if (params.sampling.ignore_eos) {
+        // add EOG biases to the active set of logit biases
+        params.sampling.logit_bias.insert(
+                params.sampling.logit_bias.end(),
+                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+    }
+
+    if (params.sampling.penalty_last_n == -1) {
+        LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+        params.sampling.penalty_last_n = llama_n_ctx(lctx);
+    }
+
+    if (params.sampling.dry_penalty_last_n == -1) {
+        LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+        params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
+    }
+
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
@@ -1303,10 +1087,11 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         llama_set_warmup(lctx, false);
     }
 
-    return res;
-}
+    iparams.model.reset(model);
+    iparams.context.reset(lctx);
 
-common_init_result::~common_init_result() = default;
+    return iparams;
+}
 
 std::string get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
@@ -1316,9 +1101,7 @@ std::string get_model_endpoint() {
     std::string model_endpoint = "https://huggingface.co/";
     if (endpoint_env) {
         model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') {
-            model_endpoint += '/';
-        }
+        if (model_endpoint.back() != '/') model_endpoint += '/';
     }
     return model_endpoint;
 }

llama/llama.cpp/common/common.h

@@ -2,19 +2,17 @@
 
 #pragma once
 
-#include "ggml-opt.h"
-#include "llama-cpp.h"
-
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
+#include <sstream>
+#include <cmath>
 
-#if defined(_WIN32) && !defined(_WIN32_WINNT)
-#define _WIN32_WINNT 0x0A00
-#endif
+#include "ggml-opt.h"
+#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -30,14 +28,7 @@
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
-struct common_time_meas {
-    common_time_meas(int64_t & t_acc, bool disable = false);
-    ~common_time_meas();
-
-    const int64_t t_start_us;
-
-    int64_t & t_acc;
-};
+#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
 struct common_adapter_lora_info {
     std::string path;
@@ -82,8 +73,7 @@ int32_t cpu_get_num_math();
 enum llama_example {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
-    LLAMA_EXAMPLE_COMPLETION,
-    LLAMA_EXAMPLE_CLI,
+    LLAMA_EXAMPLE_MAIN,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -99,7 +89,6 @@ enum llama_example {
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
     LLAMA_EXAMPLE_FINETUNE,
-    LLAMA_EXAMPLE_FIT_PARAMS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -144,22 +133,6 @@ struct common_grammar_trigger {
     llama_token token = LLAMA_TOKEN_NULL;
 };
 
-enum common_params_sampling_config : uint64_t {
-    COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS        = 1 << 0,
-    COMMON_PARAMS_SAMPLING_CONFIG_TOP_K           = 1 << 1,
-    COMMON_PARAMS_SAMPLING_CONFIG_TOP_P           = 1 << 2,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIN_P           = 1 << 3,
-    COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
-    COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD   = 1 << 5,
-    COMMON_PARAMS_SAMPLING_CONFIG_TEMP            = 1 << 6,
-    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N  = 1 << 7,
-    COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT  = 1 << 8,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT        = 1 << 9,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU    = 1 << 10,
-    COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
-};
-
-
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -192,10 +165,9 @@ struct common_params_sampling {
     bool    no_perf            = false; // disable performance metrics
     bool    timing_per_token   = false;
 
-    uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
-
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
+
     std::vector<enum common_sampler_type> samplers = {
         COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
@@ -216,10 +188,6 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
-    bool has_logit_bias() const {
-        return !logit_bias.empty();
-    }
-
     // print the parameters into a string
     std::string print() const;
 };
@@ -230,7 +198,6 @@ struct common_params_model {
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
     std::string docker_repo = ""; // Docker repo                                            // NOLINT
-    std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
 };
 
 struct common_params_speculative {
@@ -307,8 +274,8 @@ struct lr_opt {
 struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
 
 struct common_params {
-    int32_t n_predict             =    -1; // max. number of new tokens to predict, -1 == no limit
-    int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
+    int32_t n_predict             =    -1; // new tokens to predict
+    int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -329,12 +296,9 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
-    int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
-    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
-    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
-    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
+    int32_t n_gpu_layers      = -1;  // number of layers to store in VRAM (-1 - use default)
+    int32_t main_gpu          = 0;   // the GPU that is used for scratch and small tensors
+    float   tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -380,7 +344,7 @@ struct common_params {
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
-    int32_t verbosity                  = 3;  // LOG_LEVEL_INFO
+    int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
@@ -414,7 +378,6 @@ struct common_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool no_perf           = false; // disable performance metrics
-    bool show_timings      = true;  // show timing information on CLI
     bool ctx_shift         = false; // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
@@ -443,8 +406,6 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
-    int image_min_tokens = -1;
-    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;
@@ -471,7 +432,7 @@ struct common_params {
     std::string public_path   = "";                                                                         // NOLINT
     std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    bool use_jinja = true;                                                                                  // NOLINT
+    bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
@@ -490,22 +451,14 @@ struct common_params {
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
-    // router server configs
-    std::string models_dir    = ""; // directory containing models for the router server
-    std::string models_preset = ""; // directory containing model presets for the router server
-    int models_max = 4;             // maximum number of models to load simultaneously
-    bool models_autoload = true;    // automatically load models when requested via the router server
-
     bool log_json = false;
 
     std::string slot_save_path;
-    std::string media_path; // path to directory for loading media files
 
     float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
-    bool is_pp_shared   = false;
-    bool is_tg_separate = false;
+    bool is_pp_shared = false;
 
     std::vector<int32_t> n_pp;
     std::vector<int32_t> n_tg;
@@ -552,10 +505,6 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
-
-    bool has_speculative() const {
-        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
-    }
 };
 
 // call once at the start of a program if it uses libcommon
@@ -650,55 +599,25 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 // Filesystem utils
 //
 
-bool fs_validate_filename(const std::string & filename, bool allow_subdirs = false);
+bool fs_validate_filename(const std::string & filename);
 bool fs_create_directory_with_parents(const std::string & path);
-bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
 
-struct common_file_info {
-    std::string path;
-    std::string name;
-    size_t      size = 0; // in bytes
-    bool        is_dir = false;
-};
-std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
-
-//
-// TTY utils
-//
-
-// Auto-detect if colors can be enabled based on terminal and environment
-bool tty_can_use_colors();
-
 //
 // Model utils
 //
 
-struct common_sampler;
-
-// note: defines the model, context, samplers, ets. lifetimes
+// note: defines object's lifetime
 struct common_init_result {
-    common_init_result(common_params & params);
-    ~common_init_result();
+    llama_model_ptr   model;
+    llama_context_ptr context;
 
-    llama_model * model();
-    llama_context * context();
-    common_sampler * sampler(llama_seq_id seq_id);
-
-    std::vector<llama_adapter_lora_ptr> & lora();
-
-    void free_context();
-
-private:
-    struct impl;
-    std::unique_ptr<impl> pimpl;
+    std::vector<llama_adapter_lora_ptr> lora;
 };
 
-using common_init_result_ptr = std::unique_ptr<common_init_result>;
-
-common_init_result_ptr common_init_from_params(common_params & params);
+struct common_init_result     common_init_from_params(common_params & params);
 
 struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);

llama/llama.cpp/common/json-schema-to-grammar.cpp

@@ -268,10 +268,10 @@ static bool is_reserved_name(const std::string & name) {
 }
 
 std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
 std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
 std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
-    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
 };
 
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
@@ -303,11 +303,8 @@ static std::string format_literal(const std::string & literal) {
     return "\"" + escaped + "\"";
 }
 
-std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
-
-class common_schema_converter {
+class SchemaConverter {
 private:
-    friend class common_schema_info;
     friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
     std::function<json(const std::string &)> _fetch_json;
     bool _dotall;
@@ -604,10 +601,7 @@ private:
     }
 
     std::string _resolve_ref(const std::string & ref) {
-        auto it = ref.find('#');
-        std::string ref_fragment = it != std::string::npos ? ref.substr(it + 1) : ref;
-        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
-        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
+        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
         if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
             _refs_being_resolved.insert(ref);
             json resolved = _refs[ref];
@@ -730,7 +724,7 @@ private:
     }
 
 public:
-    common_schema_converter(
+    SchemaConverter(
         const std::function<json(const std::string &)> & fetch_json,
         bool dotall)
           : _fetch_json(fetch_json), _dotall(dotall)
@@ -780,24 +774,11 @@ public:
                         std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
-                            if (target.is_object() && target.contains(sel)) {
-                                target = target[sel];
-                            } else if (target.is_array()) {
-                                size_t sel_index;
-                                try {
-                                    sel_index = std::stoul(sel);
-                                } catch (const std::invalid_argument & e) {
-                                    sel_index = target.size();
-                                }
-                                if (sel_index >= target.size()) {
-                                    _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
-                                    return;
-                                }
-                                target = target[sel_index];
-                            } else {
+                            if (target.is_null() || !target.contains(sel)) {
                                 _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
                                 return;
                             }
+                            target = target[sel];
                         }
                         _refs[ref] = target;
                     }
@@ -975,7 +956,7 @@ public:
 
     void check_errors() {
         if (!_errors.empty()) {
-            throw std::invalid_argument("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
+            throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
         }
         if (!_warnings.empty()) {
             fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());
@@ -991,134 +972,6 @@ public:
     }
 };
 
-// common_schema_info implementation (pimpl)
-
-common_schema_info::common_schema_info()
-    : impl_(std::make_unique<common_schema_converter>(
-        [](const std::string &) { return json(); },
-        false)) {}
-
-common_schema_info::~common_schema_info() = default;
-
-common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
-common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
-
-void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
-    impl_->resolve_refs(schema, "");
-}
-
-// Determines if a JSON schema can resolve to a string type through any path.
-// Some models emit raw string values rather than JSON-encoded strings for string parameters.
-// If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
-// true, allowing callers to handle the value as a raw string for simplicity.
-bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
-    std::unordered_set<std::string> visited_refs;
-
-    std::function<bool(const json &)> check = [&](const json & s) -> bool {
-        if (!s.is_object()) {
-            return false;
-        }
-
-        // Handle $ref
-        if (s.contains("$ref")) {
-            const std::string & ref = s["$ref"];
-            if (visited_refs.find(ref) != visited_refs.end()) {
-                // Circular reference, assume not a string to be safe
-                return false;
-            }
-            visited_refs.insert(ref);
-            auto it = impl_->_refs.find(ref);
-            if (it != impl_->_refs.end()) {
-                return check(it->second);
-            }
-            return false;
-        }
-
-        // Check type field
-        if (s.contains("type")) {
-            const json & schema_type = s["type"];
-            if (schema_type.is_string()) {
-                if (schema_type == "string") {
-                    return true;
-                }
-            } else if (schema_type.is_array()) {
-                // Type can be an array like ["string", "null"]
-                for (const auto & t : schema_type) {
-                    if (t == "string") {
-                        return true;
-                    }
-                }
-            }
-        }
-
-        // Check oneOf/anyOf - if any alternative can be a string
-        if (s.contains("oneOf")) {
-            for (const auto & alt : s["oneOf"]) {
-                if (check(alt)) {
-                    return true;
-                }
-            }
-        }
-        if (s.contains("anyOf")) {
-            for (const auto & alt : s["anyOf"]) {
-                if (check(alt)) {
-                    return true;
-                }
-            }
-        }
-
-        // Check allOf - all components must be compatible with string type
-        if (s.contains("allOf")) {
-            bool all_string = true;
-            for (const auto & component : s["allOf"]) {
-                if (!check(component)) {
-                    all_string = false;
-                    break;
-                }
-            }
-            if (all_string) {
-                return true;
-            }
-        }
-
-        // Check const - if the constant value is a string
-        if (s.contains("const")) {
-            if (s["const"].is_string()) {
-                return true;
-            }
-        }
-
-        // Check enum - if any enum value is a string
-        if (s.contains("enum")) {
-            for (const auto & val : s["enum"]) {
-                if (val.is_string()) {
-                    return true;
-                }
-            }
-        }
-
-        // String-specific keywords imply string type
-        if (s.contains("pattern") || s.contains("minLength") || s.contains("maxLength")) {
-            return true;
-        }
-
-        // Check format - many formats imply string
-        if (s.contains("format")) {
-            const std::string & fmt = s["format"];
-            if (fmt == "date" || fmt == "time" || fmt == "date-time" ||
-                fmt == "uri" || fmt == "email" || fmt == "hostname" ||
-                fmt == "ipv4" || fmt == "ipv6" || fmt == "uuid" ||
-                fmt.find("uuid") == 0) {
-                return true;
-            }
-        }
-
-        return false;
-    };
-
-    return check(schema);
-}
-
 std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
@@ -1135,7 +988,7 @@ std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 }
 
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options) {
-    common_schema_converter converter([&](const std::string &) { return json(); }, options.dotall);
+    SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall);
     common_grammar_builder builder {
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);

llama/llama.cpp/common/json-schema-to-grammar.h

@@ -3,31 +3,11 @@
 #include <nlohmann/json_fwd.hpp>
 
 #include <functional>
-#include <memory>
 #include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);
 
-class common_schema_converter;
-
-// Probes a JSON schema to extract information about its structure and type constraints.
-class common_schema_info {
-    std::unique_ptr<common_schema_converter> impl_;
-
-  public:
-    common_schema_info();
-    ~common_schema_info();
-
-    common_schema_info(const common_schema_info &) = delete;
-    common_schema_info & operator=(const common_schema_info &) = delete;
-    common_schema_info(common_schema_info &&) noexcept;
-    common_schema_info & operator=(common_schema_info &&) noexcept;
-
-    void resolve_refs(nlohmann::ordered_json & schema);
-    bool resolves_to_string(const nlohmann::ordered_json & schema);
-};
-
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;
     std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
@@ -38,6 +18,4 @@ struct common_grammar_options {
     bool dotall = false;
 };
 
-std::string gbnf_format_literal(const std::string & literal);
-
 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});