Add experimental MLX backend and engine with imagegen support (#13648)

* WIP - MLX backend with gemma3

* MLX: add cmake and go tag build toggles

To build the new MLX backend code:
  cmake --preset MLX
  cmake --build --preset MLX --parallel
  cmake --install build --component MLX
  go build -tags mlx .

Note: the main.go entrypoint for the MLX engine will change in a follow up commit.

* add experimental image generation runtime

* add experimental image generation runtime

* MLX: wire up cuda build for linux

* MLX: get dependencies correct and dedup

This is still too large for a unified github artifact, but is now "correct" for the mlx_cuda_v13
directory.

* fix relative link bug in dedup

* Add darwin build and readme

* add go build tag for mlx dependent code and wire up build_darwin.sh

* lint cleanup

* macos: build mlx for x86

This will be CPU only.

* cuda build instructions and fix drift from mlx bump

* stale comment

* Delete agent helper doc

* Clean up readme.md

* Revise README for tokenizer clarity and details

Updated README to clarify tokenizer functionality and removed correctness section.

---------
Co-authored-by: jmorganca <jmorganca@gmail.com>

CMakeLists.txt

@@ -2,6 +2,22 @@ cmake_minimum_required(VERSION 3.21)
 
 project(Ollama C CXX)
 
+# Handle cross-compilation on macOS: when CMAKE_OSX_ARCHITECTURES is set to a
+# single architecture different from the host, override CMAKE_SYSTEM_PROCESSOR
+# to match. This is necessary because CMAKE_SYSTEM_PROCESSOR defaults to the
+# host architecture, but downstream projects (like MLX) use it to detect the
+# target architecture.
+if(CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES ";")
+    # Single architecture specified
+    if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+        message(STATUS "Cross-compiling for x86_64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to x86_64")
+        set(CMAKE_SYSTEM_PROCESSOR "x86_64")
+    elseif(CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        message(STATUS "Cross-compiling for arm64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to arm64")
+        set(CMAKE_SYSTEM_PROCESSOR "arm64")
+    endif()
+endif()
+
 include(CheckLanguage)
 include(GNUInstallDirs)
 
@@ -12,7 +28,7 @@ set(BUILD_SHARED_LIBS ON)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
 
 set(GGML_BUILD ON)
 set(GGML_SHARED ON)
@@ -147,14 +163,48 @@ if(CMAKE_HIP_COMPILER)
     endif()
 endif()
 
-find_package(Vulkan)
-if(Vulkan_FOUND)
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-    install(TARGETS ggml-vulkan
+if(NOT APPLE)
+    find_package(Vulkan)
+    if(Vulkan_FOUND)
+        add_subdirectory(${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)
+
+    # Find CUDA toolkit if MLX is built with CUDA support
+    find_package(CUDAToolkit)
+
+    install(TARGETS mlx mlxc
         RUNTIME_DEPENDENCIES
-            PRE_INCLUDE_REGEXES vulkan
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
+            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc cudnn nccl
             PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
     )
-endif()
+
+    # Manually install cudart and cublas since they might not be picked up as direct dependencies
+    if(CUDAToolkit_FOUND)
+        file(GLOB CUDART_LIBS
+            "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
+            "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*")
+        if(CUDART_LIBS)
+            install(FILES ${CUDART_LIBS}
+                DESTINATION ${OLLAMA_INSTALL_DIR}
+                COMPONENT MLX)
+        endif()
+    endif()
+endif()
\ No newline at end of file

CMakePresets.json

@@ -41,7 +41,7 @@
       "inherits": [ "CUDA" ],
       "cacheVariables": {
         "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual",
-        "CMAKE_CUDA_FLAGS": "-t 2",
+        "CMAKE_CUDA_FLAGS": "-t 4",
         "OLLAMA_RUNNER_DIR": "cuda_v13"
       }
     },
@@ -83,6 +83,28 @@
       "cacheVariables": {
         "OLLAMA_RUNNER_DIR": "vulkan"
       }
+    },
+    {
+      "name": "MLX",
+      "inherits": [ "Default" ],
+      "cacheVariables": {
+        "MLX_ENGINE": "ON",
+        "OLLAMA_RUNNER_DIR": "mlx"
+      }
+    },
+    {
+      "name": "MLX CUDA 12",
+      "inherits": [ "MLX", "CUDA 12" ],
+      "cacheVariables": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v12"
+      }
+    },
+    {
+      "name": "MLX CUDA 13",
+      "inherits": [ "MLX", "CUDA 13" ],
+      "cacheVariables": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+      }
     }
   ],
   "buildPresets": [
@@ -140,6 +162,21 @@
       "name": "Vulkan",
       "targets": [ "ggml-vulkan" ],
       "configurePreset": "Vulkan"
+    },
+    {
+      "name": "MLX",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX"
+    },
+    {
+      "name": "MLX CUDA 12",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX CUDA 12"
+    },
+    {
+      "name": "MLX CUDA 13",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX CUDA 13"
     }
   ]
 }

Dockerfile

@@ -131,7 +131,39 @@ 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 
+        && cmake --install build --component Vulkan --strip --parallel 8
+
+FROM base AS mlx
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
+    && dnf install -y openblas-devel lapack-devel \
+    && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
+    && dnf install -y libnccl libnccl-devel
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+ENV BLAS_INCLUDE_DIRS=/usr/include/openblas
+ENV LAPACK_INCLUDE_DIRS=/usr/include/openblas
+ENV CGO_LDFLAGS="-L/usr/local/cuda-13/lib64 -L/usr/local/cuda-13/targets/x86_64-linux/lib/stubs"
+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 CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
+        && cmake --build --parallel ${PARALLEL} --preset 'MLX CUDA 13' \
+        && 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 mkdir -p dist/bin
+RUN go build -tags mlx -trimpath -buildmode=pie -o dist/bin/imagegen ./x/imagegen/cmd/engine
 
 
 FROM base AS build
@@ -153,6 +185,8 @@ 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     /go/src/github.com/ollama/ollama/dist/bin/ /bin/
 
 FROM --platform=linux/arm64 scratch AS arm64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/

convert/convert.go

@@ -6,11 +6,14 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"iter"
 	"log/slog"
+	"maps"
 	"os"
 	"slices"
 	"strings"
 
+	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,8 +21,13 @@ 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"`
+		VocabSize  uint32 `json:"vocab_size"`
+		HiddenSize uint32 `json:"hidden_size"`
+		ModelType  string `json:"model_type"`
 	} `json:"text_config"`
 }
 
@@ -33,8 +41,94 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) ggml.KV {
-	kv := ggml.KV{
+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{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -63,7 +157,7 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() ggml.KV {
+func (p AdapterParameters) KV() KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -71,7 +165,7 @@ func (p AdapterParameters) KV() ggml.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := ggml.KV{
+	kv := KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -88,9 +182,14 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-type ModelConverter interface {
+type ModelKV interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) ggml.KV
+	KV(*Tokenizer) KV
+}
+
+type ModelConverter interface {
+	ModelKV
+
 	// 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.
@@ -107,7 +206,7 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(ggml.KV) ggml.KV
+	KV(ofs.Config) 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.
@@ -115,7 +214,7 @@ type AdapterConverter interface {
 	Replacements() []string
 }
 
-func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -126,8 +225,8 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 		return err
 	}
 
-	arch, ok := baseKV["general.architecture"]
-	if !ok {
+	arch := baseKV.Architecture()
+	if arch == "" {
 		return errors.New("architecture not set for the base model")
 	}
 
@@ -153,23 +252,19 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 	return writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
 }
 
-// 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 {
+func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if len(p.Architectures) < 1 {
-		return errors.New("unknown architecture")
+		return nil, nil, errors.New("unknown architecture")
 	}
 
 	var conv ModelConverter
@@ -217,22 +312,22 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
 	default:
-		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
+		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if t, ok := conv.(moreParser); ok {
 		if err := t.parseMore(fsys); err != nil {
-			return err
+			return nil, nil, err
 		}
 	}
 
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
@@ -254,6 +349,19 @@ func ConvertModel(fsys fs.FS, f *os.File) 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 {
@@ -263,7 +371,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	return writeFile(f, conv.KV(t), conv.Tensors(ts))
 }
 
-func writeFile(f *os.File, kv ggml.KV, ts []*ggml.Tensor) error {
+func writeFile(f *os.File, kv 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) ggml.KV {
+func (p *bertModel) KV(t *Tokenizer) 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) ggml.KV {
+func (p *commandrModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "command-r"
 	kv["general.name"] = "command-r"

convert/convert_deepseek2.go

@@ -47,7 +47,7 @@ type deepseek2Model struct {
 	Architecture string
 }
 
-func (p *deepseek2Model) KV(t *Tokenizer) ggml.KV {
+func (p *deepseek2Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseek2"
 	kv["general.type"] = "model"

convert/convert_deepseekocr.go

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

convert/convert_gemma2.go

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

convert/convert_gemma2_adapter.go

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

convert/convert_gemma3.go

@@ -3,8 +3,6 @@ package convert
 import (
 	"cmp"
 	"slices"
-
-	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemma3Model struct {
@@ -55,7 +53,7 @@ const (
 	gemma27BLayerCount = 62
 )
 
-func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3"
 

convert/convert_gemma3n.go

@@ -38,7 +38,7 @@ type gemma3nModel struct {
 	VisionModel struct{} `json:"vision_config"`
 }
 
-func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
+func (m *gemma3nModel) KV(t *Tokenizer) 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) ggml.KV {
+func (m *gptossModel) KV(t *Tokenizer) 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) ggml.KV {
+func (p *llamaModel) KV(t *Tokenizer) 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) ggml.KV {
+func (p *llama4Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama4"
 

convert/convert_llama_adapter.go

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

convert/convert_mistral.go

@@ -60,7 +60,7 @@ type mistral3Model struct {
 	ProjectorHiddenAct      string `json:"projector_hidden_act"`
 }
 
-func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.TextModel.VocabSize

convert/convert_mistral_causal.go

@@ -39,7 +39,7 @@ type mistral3CausalModel struct {
 	} `json:"rope_parameters"`
 }
 
-func (p *mistral3CausalModel) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3CausalModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.VocabSize

convert/convert_mixtral.go

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