fs/ggml: add multiplier in graph estimates (#11208)

* update patches

* cherry pick metal mean kernel

* cherry pick cuda mean kernel

* gemma3n

.github/workflows/release.yaml

@@ -103,21 +103,18 @@ jobs:
         arch: [amd64]
         preset: ['CPU']
         include:
-          - os: windows
-            arch: amd64
-            preset: 'CUDA 11'
-            install: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
-            cuda-version: '11.3'
           - os: windows
             arch: amd64
             preset: 'CUDA 12'
             install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
             cuda-version: '12.8'
+            flags: ''
           - os: windows
             arch: amd64
             preset: 'ROCm 6'
             install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
             rocm-version: '6.2'
+            flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     env:
@@ -160,6 +157,9 @@ jobs:
           echo "$hipPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CC=$hipPath\bin\clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
           echo "CXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "HIPCXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "HIP_PLATFORM=amd" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "CMAKE_PREFIX_PATH=$hipPath" | Out-File -FilePath $env:GITHUB_ENV -Append
       - if: matrix.preset == 'CPU'
         run: |
           echo "CC=clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
@@ -178,9 +178,9 @@ jobs:
           key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}
       - name: Build target "${{ matrix.preset }}"
         run: |
-          Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
-          cmake --preset "${{ matrix.preset }}"
+          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'
+          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
           cmake --build --parallel --preset "${{ matrix.preset }}"
           cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || 'CPU' }}" --strip --parallel 8
         env:
@@ -246,7 +246,7 @@ jobs:
             dist\${{ matrix.os }}-${{ matrix.arch }}-app.exe
 
   windows-sign:
-    runs-on: windows-2022
+    runs-on: windows
     environment: release
     needs: [windows-depends, windows-build]
     steps:
@@ -322,16 +322,21 @@ jobs:
       - run: |
           for COMPONENT in bin/* lib/ollama/*; do
             case "$COMPONENT" in
-              bin/ollama)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/*.so)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_v11)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_v12)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
-              lib/ollama/cuda_jetpack5) echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
-              lib/ollama/cuda_jetpack6) echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
-              lib/ollama/rocm)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
+              bin/ollama)                echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_sbsa)      echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
+              lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
+              lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
             esac
           done
         working-directory: dist/${{ matrix.os }}-${{ matrix.arch }}
+      - run: |
+          echo "Manifests"
+          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in ; do
+            echo $ARCHIVE
+            cat $ARCHIVE
+          done
       - run: |
           for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
             tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz);
@@ -470,8 +475,18 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           pattern: dist-linux-*
-          path: dist
-          merge-multiple: true
+          path: stage
+          merge-multiple: false
+      - name: Merge linux amd64 payload
+        working-directory: stage/dist-linux-amd64-archive
+        run: |
+          tar zxf ollama-linux-amd64.tgz
+          tar zxf ../dist-linux-amd64-rocm/ollama-linux-amd64.tgz
+          rm -f ollama-linux-amd64.tgz ../dist-linux-amd64-rocm/ollama-linux-amd64.tgz
+          tar -c -f- --owner 0 --group 0 . | pigz -9vc > ../ollama-linux-amd64.tgz
+      - name: Cleanup linux payloads
+        run: |
+          find stage -name ollama-linux\*.tgz -exec mv {} dist/ \;
       - run: find . -type f -not -name 'sha256sum.txt' | xargs sha256sum | tee sha256sum.txt
         working-directory: dist
       - name: Create or update Release

.github/workflows/test.yaml

@@ -36,7 +36,7 @@ jobs:
               | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
           }
 
-          echo changed=$(changed 'llama/llama.cpp/**' 'ml/backend/ggml/ggml/**') | tee -a $GITHUB_OUTPUT
+          echo changed=$(changed 'llama/llama.cpp/**/*' 'ml/backend/ggml/ggml/**/*') | tee -a $GITHUB_OUTPUT
 
   linux:
     needs: [changes]
@@ -46,7 +46,7 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            container: nvidia/cuda:11.8.0-devel-ubuntu22.04
+            container: nvidia/cuda:12.8.1-devel-ubuntu22.04
             flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
           - preset: ROCm
             container: rocm/dev-ubuntu-22.04:6.1.2
@@ -78,11 +78,11 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
+            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
             flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
           - preset: ROCm
             install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            flags: '-DAMDGPU_TARGETS=gfx1010'
+            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
     runs-on: windows
     steps:
       - run: |
@@ -102,7 +102,7 @@ jobs:
           $ErrorActionPreference = "Stop"
           if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
             Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_11.3", "nvcc_11.3", "cublas_11.3", "cublas_dev_11.3")) -NoNewWindow -Wait
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_12.8", "nvcc_12.8", "cublas_12.8", "cublas_dev_12.8")) -NoNewWindow -Wait
           }
 
           $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
@@ -120,6 +120,9 @@ jobs:
           echo "$hipPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CC=$hipPath\bin\clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
           echo "CXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "HIPCXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "HIP_PLATFORM=amd" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "CMAKE_PREFIX_PATH=$hipPath" | Out-File -FilePath $env:GITHUB_ENV -Append
       - if: ${{ !cancelled() && steps.cache-install.outputs.cache-hit != 'true' }}
         uses: actions/cache/save@v4
         with:
@@ -133,8 +136,8 @@ jobs:
           path: ${{ github.workspace }}\.ccache
           key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
       - run: |
-          Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          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'
           cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
           cmake --build --parallel --preset "${{ matrix.preset }}"
         env:

CMakeLists.txt

@@ -78,14 +78,13 @@ if(CMAKE_CUDA_COMPILER)
 
     find_package(CUDAToolkit)
     add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
-    set(OLLAMA_CUDA_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/cuda_v${CUDAToolkit_VERSION_MAJOR})
     install(TARGETS ggml-cuda
         RUNTIME_DEPENDENCIES
             DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_LIBRARY_DIR}
             PRE_INCLUDE_REGEXES cublas cublasLt cudart
             PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
-        LIBRARY DESTINATION ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
     )
 endif()
 
@@ -116,7 +115,11 @@ if(CMAKE_HIP_COMPILER)
 
         set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
         install(TARGETS ggml-hip
-            RUNTIME_DEPENDENCIES
+            RUNTIME_DEPENDENCY_SET rocm
+            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
+            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
+        )
+        install(RUNTIME_DEPENDENCY_SET rocm
                 DIRECTORIES ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR}
                 PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register drm drm_amdgpu numa elf
                 PRE_EXCLUDE_REGEXES ".*"

CMakePresets.json

@@ -17,20 +17,12 @@
       "name": "CUDA",
       "inherits": [ "Default" ]
     },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
-      }
-    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
         "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120",
-        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets -t 2"
       }
     },
     {
@@ -58,6 +50,7 @@
       "name": "ROCm 6",
       "inherits": [ "ROCm" ],
       "cacheVariables": {
+        "CMAKE_HIP_FLAGS": "-parallel-jobs=4",
         "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
       }
     }
@@ -78,11 +71,6 @@
       "configurePreset": "CUDA",
       "targets": [ "ggml-cuda" ]
     },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 11"
-    },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],

Dockerfile

@@ -7,12 +7,13 @@ ARG JETPACK5VERSION=r35.4.1
 ARG JETPACK6VERSION=r36.4.0
 ARG CMAKEVERSION=3.31.2
 
-# CUDA v11 requires gcc v10.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
+# We require gcc v10 minimum.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
 FROM --platform=linux/amd64 rocm/dev-almalinux-8:${ROCMVERSION}-complete AS base-amd64
 RUN yum install -y yum-utils \
     && yum-config-manager --add-repo https://dl.rockylinux.org/vault/rocky/8.5/AppStream/\$basearch/os/ \
     && rpm --import https://dl.rockylinux.org/pub/rocky/RPM-GPG-KEY-Rocky-8 \
     && dnf install -y yum-utils ccache gcc-toolset-10-gcc-10.2.1-8.2.el8 gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 gcc-toolset-10-binutils-2.35-11.el8 \
+    && dnf install -y ccache \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
 ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 
@@ -38,15 +39,6 @@ RUN --mount=type=cache,target=/root/.ccache \
         && cmake --build --parallel --preset 'CPU' \
         && cmake --install build --component CPU --strip --parallel 8
 
-FROM base AS cuda-11
-ARG CUDA11VERSION=11.3
-RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
-ENV PATH=/usr/local/cuda-11/bin:$PATH
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 11' \
-        && cmake --build --parallel --preset 'CUDA 11' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
 FROM base AS cuda-12
 ARG CUDA12VERSION=12.8
 RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
@@ -98,17 +90,15 @@ RUN --mount=type=cache,target=/root/.cache/go-build \
     go build -trimpath -buildmode=pie -o /bin/ollama .
 
 FROM --platform=linux/amd64 scratch AS amd64
-COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
-COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama
 
 FROM --platform=linux/arm64 scratch AS arm64
-COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
-COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
-COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_jetpack5
-COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_jetpack6
+COPY --from=cuda-12 dist/lib/ollama /lib/ollama/cuda_sbsa
+COPY --from=jetpack-5 dist/lib/ollama /lib/ollama/cuda_jetpack5
+COPY --from=jetpack-6 dist/lib/ollama /lib/ollama/cuda_jetpack6
 
 FROM scratch AS rocm
-COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm
+COPY --from=rocm-6 dist/lib/ollama /lib/ollama
 
 FROM ${FLAVOR} AS archive
 COPY --from=cpu dist/lib/ollama /lib/ollama

README.md

@@ -40,10 +40,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3.2](https://ollama.com/library/llama3.2):
+To run and chat with [Gemma 3](https://ollama.com/library/gemma3):
 
 ```shell
-ollama run llama3.2
+ollama run gemma3
 ```
 
 ## Model library
@@ -407,6 +407,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
 - [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
 - [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
+- [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
+- [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
+- [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
 
 ### Cloud
 
@@ -451,6 +454,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull, and download models from Ollama Registry in your terminal.
 - [GGUF-to-Ollama](https://github.com/jonathanhecl/gguf-to-ollama) - Importing GGUF to Ollama made easy (multiplatform)
 - [AWS-Strands-With-Ollama](https://github.com/rapidarchitect/ollama_strands) - AWS Strands Agents with Ollama Examples
+- [ollama-multirun](https://github.com/attogram/ollama-multirun) - A bash shell script to run a single prompt against any or all of your locally installed ollama models, saving the output and performance statistics as easily navigable web pages. ([Demo](https://attogram.github.io/ai_test_zone/))
 
 ### Apple Vision Pro
 
@@ -587,6 +591,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Simple-Discord-AI](https://github.com/zyphixor/simple-discord-ai)
 - [LLM Telegram Bot](https://github.com/innightwolfsleep/llm_telegram_bot) (telegram bot, primary for RP. Oobabooga-like buttons, [A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) API integration e.t.c)
 - [mcp-llm](https://github.com/sammcj/mcp-llm) (MCP Server to allow LLMs to call other LLMs)
+- [SimpleOllamaUnity](https://github.com/HardCodeDev777/SimpleOllamaUnity) (Unity Engine extension for communicating with Ollama in a few lines of code. Also works at runtime)
 - [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)
 
 ### Supported backends

api/types.go

@@ -83,6 +83,12 @@ type GenerateRequest struct {
 	// Options lists model-specific options. For example, temperature can be
 	// set through this field, if the model supports it.
 	Options map[string]any `json:"options"`
+
+	// Think controls whether thinking/reasoning models will think before
+	// responding. Needs to be a pointer so we can distinguish between false
+	// (request that thinking _not_ be used) and unset (use the old behavior
+	// before this option was introduced)
+	Think *bool `json:"think,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -108,6 +114,10 @@ type ChatRequest struct {
 
 	// Options lists model-specific options.
 	Options map[string]any `json:"options"`
+
+	// Think controls whether thinking/reasoning models will think before
+	// responding
+	Think *bool `json:"think,omitempty"`
 }
 
 type Tools []Tool
@@ -126,8 +136,11 @@ func (t Tool) String() string {
 // role ("system", "user", or "assistant"), the content and an optional list
 // of images.
 type Message struct {
-	Role      string      `json:"role"`
-	Content   string      `json:"content"`
+	Role    string `json:"role"`
+	Content string `json:"content"`
+	// Thinking contains the text that was inside thinking tags in the
+	// original model output when ChatRequest.Think is enabled.
+	Thinking  string      `json:"thinking,omitempty"`
 	Images    []ImageData `json:"images,omitempty"`
 	ToolCalls []ToolCall  `json:"tool_calls,omitempty"`
 }
@@ -478,6 +491,10 @@ type GenerateResponse struct {
 	// Response is the textual response itself.
 	Response string `json:"response"`
 
+	// Thinking contains the text that was inside thinking tags in the
+	// original model output when ChatRequest.Think is enabled.
+	Thinking string `json:"thinking,omitempty"`
+
 	// Done specifies if the response is complete.
 	Done bool `json:"done"`
 

api/types_test.go

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

benchmark/server_benchmark_test.go

@@ -1,178 +0,0 @@
-package benchmark
-
-import (
-	"context"
-	"flag"
-	"fmt"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-// Command line flags
-var modelFlag string
-
-func init() {
-	flag.StringVar(&modelFlag, "m", "", "Name of the model to benchmark")
-	flag.Lookup("m").DefValue = "model"
-}
-
-// modelName returns the model name from flags, failing the test if not set
-func modelName(b *testing.B) string {
-	if modelFlag == "" {
-		b.Fatal("Error: -m flag is required for benchmark tests")
-	}
-	return modelFlag
-}
-
-type TestCase struct {
-	name      string
-	prompt    string
-	maxTokens int
-}
-
-// runGenerateBenchmark contains the common generate and metrics logic
-func runGenerateBenchmark(b *testing.B, ctx context.Context, client *api.Client, req *api.GenerateRequest) {
-	start := time.Now()
-	var ttft time.Duration
-	var metrics api.Metrics
-
-	err := client.Generate(ctx, req, func(resp api.GenerateResponse) error {
-		if ttft == 0 && resp.Response != "" {
-			ttft = time.Since(start)
-		}
-		if resp.Done {
-			metrics = resp.Metrics
-		}
-		return nil
-	})
-
-	// Report custom metrics as part of the benchmark results
-	b.ReportMetric(float64(ttft.Milliseconds()), "ttft_ms")
-	b.ReportMetric(float64(metrics.LoadDuration.Milliseconds()), "load_ms")
-
-	// Token throughput metrics
-	promptThroughput := float64(metrics.PromptEvalCount) / metrics.PromptEvalDuration.Seconds()
-	genThroughput := float64(metrics.EvalCount) / metrics.EvalDuration.Seconds()
-	b.ReportMetric(promptThroughput, "prompt_tok/s")
-	b.ReportMetric(genThroughput, "gen_tok/s")
-
-	// Token counts
-	b.ReportMetric(float64(metrics.PromptEvalCount), "prompt_tokens")
-	b.ReportMetric(float64(metrics.EvalCount), "gen_tokens")
-	if err != nil {
-		b.Fatal(err)
-	}
-}
-
-// BenchmarkColdStart runs benchmarks with model loading from cold state
-func BenchmarkColdStart(b *testing.B) {
-	client := setup(b)
-	tests := []TestCase{
-		{"short_prompt", "Write a long story", 100},
-		{"medium_prompt", "Write a detailed economic analysis", 500},
-		{"long_prompt", "Write a comprehensive AI research paper", 1000},
-	}
-	m := modelName(b)
-
-	for _, tt := range tests {
-		b.Run(fmt.Sprintf("%s/cold/%s", m, tt.name), func(b *testing.B) {
-			ctx := b.Context()
-
-			// Set number of tokens as our throughput metric
-			b.SetBytes(int64(tt.maxTokens))
-
-			for b.Loop() {
-				b.StopTimer()
-				// Ensure model is unloaded before each iteration
-				unload(client, m, b)
-				b.StartTimer()
-
-				req := &api.GenerateRequest{
-					Model:   m,
-					Prompt:  tt.prompt,
-					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
-				}
-
-				runGenerateBenchmark(b, ctx, client, req)
-			}
-		})
-	}
-}
-
-// BenchmarkWarmStart runs benchmarks with pre-loaded model
-func BenchmarkWarmStart(b *testing.B) {
-	client := setup(b)
-	tests := []TestCase{
-		{"short_prompt", "Write a long story", 100},
-		{"medium_prompt", "Write a detailed economic analysis", 500},
-		{"long_prompt", "Write a comprehensive AI research paper", 1000},
-	}
-	m := modelName(b)
-
-	for _, tt := range tests {
-		b.Run(fmt.Sprintf("%s/warm/%s", m, tt.name), func(b *testing.B) {
-			ctx := b.Context()
-
-			// Pre-warm the model
-			warmup(client, m, tt.prompt, b)
-
-			// Set number of tokens as our throughput metric
-			b.SetBytes(int64(tt.maxTokens))
-
-			for b.Loop() {
-				req := &api.GenerateRequest{
-					Model:   m,
-					Prompt:  tt.prompt,
-					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
-				}
-
-				runGenerateBenchmark(b, ctx, client, req)
-			}
-		})
-	}
-}
-
-// setup verifies server and model availability
-func setup(b *testing.B) *api.Client {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		b.Fatal(err)
-	}
-	if _, err := client.Show(b.Context(), &api.ShowRequest{Model: modelName(b)}); err != nil {
-		b.Fatalf("Model unavailable: %v", err)
-	}
-
-	return client
-}
-
-// warmup ensures the model is loaded and warmed up
-func warmup(client *api.Client, model string, prompt string, b *testing.B) {
-	for range 3 {
-		err := client.Generate(
-			context.Background(),
-			&api.GenerateRequest{
-				Model:   model,
-				Prompt:  prompt,
-				Options: map[string]any{"num_predict": 50, "temperature": 0.1},
-			},
-			func(api.GenerateResponse) error { return nil },
-		)
-		if err != nil {
-			b.Logf("Error during model warm-up: %v", err)
-		}
-	}
-}
-
-// unload forces model unloading using KeepAlive: 0 parameter
-func unload(client *api.Client, model string, b *testing.B) {
-	req := &api.GenerateRequest{
-		Model:     model,
-		KeepAlive: &api.Duration{Duration: 0},
-	}
-	if err := client.Generate(context.Background(), req, func(api.GenerateResponse) error { return nil }); err != nil {
-		b.Logf("Unload error: %v", err)
-	}
-	time.Sleep(1 * time.Second)
-}

cmd/cmd.go

@@ -39,6 +39,7 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/runner"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
@@ -46,6 +47,23 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
+// ensureThinkingSupport emits a warning if the model does not advertise thinking support
+func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
+	if name == "" {
+		return
+	}
+	resp, err := client.Show(ctx, &api.ShowRequest{Model: name})
+	if err != nil {
+		return
+	}
+	for _, cap := range resp.Capabilities {
+		if cap == model.CapabilityThinking {
+			return
+		}
+	}
+	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
+}
+
 var errModelfileNotFound = errors.New("specified Modelfile wasn't found")
 
 func getModelfileName(cmd *cobra.Command) (string, error) {
@@ -265,6 +283,9 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 	req := &api.GenerateRequest{
 		Model:     opts.Model,
 		KeepAlive: opts.KeepAlive,
+
+		// pass Think here so we fail before getting to the chat prompt if the model doesn't support it
+		Think: opts.Think,
 	}
 
 	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
@@ -299,6 +320,22 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.Format = format
 
+	thinkFlag := cmd.Flags().Lookup("think")
+	if thinkFlag.Changed {
+		think, err := cmd.Flags().GetBool("think")
+		if err != nil {
+			return err
+		}
+		opts.Think = &think
+	} else {
+		opts.Think = nil
+	}
+	hidethinking, err := cmd.Flags().GetBool("hidethinking")
+	if err != nil {
+		return err
+	}
+	opts.HideThinking = hidethinking
+
 	keepAlive, err := cmd.Flags().GetString("keepalive")
 	if err != nil {
 		return err
@@ -362,6 +399,11 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	opts.Think, err = inferThinkingOption(&info.Capabilities, &opts, thinkFlag.Changed)
+	if err != nil {
+		return err
+	}
+
 	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision)
 
 	// TODO: remove the projector info and vision info checks below,
@@ -923,17 +965,19 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 type generateContextKey string
 
 type runOptions struct {
-	Model       string
-	ParentModel string
-	Prompt      string
-	Messages    []api.Message
-	WordWrap    bool
-	Format      string
-	System      string
-	Images      []api.ImageData
-	Options     map[string]any
-	MultiModal  bool
-	KeepAlive   *api.Duration
+	Model        string
+	ParentModel  string
+	Prompt       string
+	Messages     []api.Message
+	WordWrap     bool
+	Format       string
+	System       string
+	Images       []api.ImageData
+	Options      map[string]any
+	MultiModal   bool
+	KeepAlive    *api.Duration
+	Think        *bool
+	HideThinking bool
 }
 
 type displayResponseState struct {
@@ -989,6 +1033,26 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	}
 }
 
+func thinkingOutputOpeningText(plainText bool) string {
+	text := "Thinking...\n"
+
+	if plainText {
+		return text
+	}
+
+	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault + readline.ColorGrey
+}
+
+func thinkingOutputClosingText(plainText bool) string {
+	text := "...done thinking.\n\n"
+
+	if plainText {
+		return text
+	}
+
+	return readline.ColorGrey + readline.ColorBold + text + readline.ColorDefault
+}
+
 func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -1016,14 +1080,34 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 	var latest api.ChatResponse
 	var fullResponse strings.Builder
 	var role string
+	var thinkTagOpened bool = false
+	var thinkTagClosed bool = false
 
 	fn := func(response api.ChatResponse) error {
-		p.StopAndClear()
+		if response.Message.Content != "" || !opts.HideThinking {
+			p.StopAndClear()
+		}
 
 		latest = response
 
 		role = response.Message.Role
+		if response.Message.Thinking != "" && !opts.HideThinking {
+			if !thinkTagOpened {
+				fmt.Print(thinkingOutputOpeningText(false))
+				thinkTagOpened = true
+			}
+			displayResponse(response.Message.Thinking, opts.WordWrap, state)
+		}
+
 		content := response.Message.Content
+		if thinkTagOpened && !thinkTagClosed && content != "" {
+			fmt.Print(thinkingOutputClosingText(false))
+			thinkTagClosed = true
+		}
+		// purposefully not putting thinking blocks in the response, which would
+		// only be needed if we later added tool calling to the cli (they get
+		// filtered out anyway since current models don't expect them unless you're
+		// about to finish some tool calls)
 		fullResponse.WriteString(content)
 
 		displayResponse(content, opts.WordWrap, state)
@@ -1040,6 +1124,7 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		Messages: opts.Messages,
 		Format:   json.RawMessage(opts.Format),
 		Options:  opts.Options,
+		Think:    opts.Think,
 	}
 
 	if opts.KeepAlive != nil {
@@ -1101,13 +1186,32 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	}()
 
 	var state *displayResponseState = &displayResponseState{}
+	var thinkTagOpened bool = false
+	var thinkTagClosed bool = false
+
+	plainText := !term.IsTerminal(int(os.Stdout.Fd()))
 
 	fn := func(response api.GenerateResponse) error {
-		p.StopAndClear()
-
 		latest = response
 		content := response.Response
 
+		if response.Response != "" || !opts.HideThinking {
+			p.StopAndClear()
+		}
+
+		if response.Thinking != "" && !opts.HideThinking {
+			if !thinkTagOpened {
+				fmt.Print(thinkingOutputOpeningText(plainText))
+				thinkTagOpened = true
+			}
+			displayResponse(response.Thinking, opts.WordWrap, state)
+		}
+
+		if thinkTagOpened && !thinkTagClosed && content != "" {
+			fmt.Print(thinkingOutputClosingText(plainText))
+			thinkTagClosed = true
+		}
+
 		displayResponse(content, opts.WordWrap, state)
 
 		return nil
@@ -1133,6 +1237,7 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		System:    opts.System,
 		Options:   opts.Options,
 		KeepAlive: opts.KeepAlive,
+		Think:     opts.Think,
 	}
 
 	if err := client.Generate(ctx, &request, fn); err != nil {
@@ -1348,6 +1453,8 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 	runCmd.Flags().String("format", "", "Response format (e.g. json)")
+	runCmd.Flags().Bool("think", false, "Whether to use thinking mode for supported models")
+	runCmd.Flags().Bool("hidethinking", false, "Hide thinking output (if provided)")
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",
@@ -1399,7 +1506,6 @@ func NewCLI() *cobra.Command {
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListRunningHandler,
 	}
-
 	copyCmd := &cobra.Command{
 		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
@@ -1488,3 +1594,45 @@ func NewCLI() *cobra.Command {
 
 	return rootCmd
 }
+
+// If the user has explicitly set thinking options, either through the CLI or
+// through the `/set think` or `set nothink` interactive options, then we
+// respect them. Otherwise, we check model capabilities to see if the model
+// supports thinking. If the model does support thinking, we enable it.
+// Otherwise, we unset the thinking option (which is different than setting it
+// to false).
+//
+// If capabilities are not provided, we fetch them from the server.
+func inferThinkingOption(caps *[]model.Capability, runOpts *runOptions, explicitlySetByUser bool) (*bool, error) {
+	if explicitlySetByUser {
+		return runOpts.Think, nil
+	}
+
+	if caps == nil {
+		client, err := api.ClientFromEnvironment()
+		if err != nil {
+			return nil, err
+		}
+		ret, err := client.Show(context.Background(), &api.ShowRequest{
+			Model: runOpts.Model,
+		})
+		if err != nil {
+			return nil, err
+		}
+		caps = &ret.Capabilities
+	}
+
+	thinkingSupported := false
+	for _, cap := range *caps {
+		if cap == model.CapabilityThinking {
+			thinkingSupported = true
+		}
+	}
+
+	if thinkingSupported {
+		thinking := true
+		return &thinking, nil
+	}
+
+	return nil, nil
+}

cmd/interactive.go

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

cmd/start_darwin.go

@@ -5,7 +5,7 @@ import (
 	"errors"
 	"os"
 	"os/exec"
-	"strings"
+	"regexp"
 
 	"github.com/ollama/ollama/api"
 )
@@ -19,11 +19,12 @@ func startApp(ctx context.Context, client *api.Client) error {
 	if err != nil {
 		return err
 	}
-	if !strings.Contains(link, "Ollama.app") {
+	r := regexp.MustCompile(`^.*/Ollama\s?\d*.app`)
+	m := r.FindStringSubmatch(link)
+	if len(m) != 1 {
 		return errors.New("could not find ollama app")
 	}
-	path := strings.Split(link, "Ollama.app")
-	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
+	if err := exec.Command("/usr/bin/open", "-j", "-a", m[0], "--args", "--fast-startup").Run(); err != nil {
 		return err
 	}
 	return waitForServer(ctx, client)

cmd/start_windows.go

@@ -45,14 +45,11 @@ func startApp(ctx context.Context, client *api.Client) error {
 			}
 		}
 	}
-	// log.Printf("XXX attempting to start app %s", appExe)
 
 	cmd_path := "c:\\Windows\\system32\\cmd.exe"
-	cmd := exec.Command(cmd_path, "/c", appExe)
-	// TODO - these hide flags aren't working - still pops up a command window for some reason
+	cmd := exec.Command(cmd_path, "/c", appExe, "--hide", "--fast-startup")
 	cmd.SysProcAttr = &syscall.SysProcAttr{CreationFlags: 0x08000000, HideWindow: true}
 
-	// TODO this didn't help either...
 	cmd.Stdin = strings.NewReader("")
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
@@ -74,7 +71,16 @@ func isProcRunning(procName string) []uint32 {
 		slog.Debug("failed to check for running installers", "error", err)
 		return nil
 	}
-	pids = pids[:ret]
+	if ret > uint32(len(pids)) {
+		pids = make([]uint32, ret+10)
+		if err := windows.EnumProcesses(pids, &ret); err != nil || ret == 0 {
+			slog.Debug("failed to check for running installers", "error", err)
+			return nil
+		}
+	}
+	if ret < uint32(len(pids)) {
+		pids = pids[:ret]
+	}
 	var matches []uint32
 	for _, pid := range pids {
 		if pid == 0 {

cmd/warn_thinking_test.go

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

convert/convert.go

@@ -190,6 +190,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &gemma2Model{}
 	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
 		conv = &gemma3Model{Architecture: p.Architectures[0]}
+	case "Gemma3nForConditionalGeneration":
+		conv = &gemma3nModel{}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":

convert/convert_gemma3n.go

@@ -0,0 +1,168 @@
+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"gonum.org/v1/gonum/stat/distuv"
+)
+
+type gemma3nModel struct {
+	ModelParameters
+
+	TextModel struct {
+		ActivationSparsityPattern []float32 `json:"activation_sparsity_pattern"`
+		AltupActiveIdx            uint32    `json:"altup_active_idx"`
+		AltupCoefClip             float32   `json:"altup_coef_clip"`
+		AltupCorrectScale         bool      `json:"altup_correct_scale"`
+		AltupLRMultiplier         float32   `json:"altup_lr_multiplier"`
+		AltupNumInputs            uint32    `json:"altup_num_inputs"`
+		HeadDim                   uint32    `json:"head_dim"`
+		HiddenSize                uint32    `json:"hidden_size"`
+		HiddenSizePerLayerInput   uint32    `json:"hidden_size_per_layer_input"`
+		IntermediateSize          uint32    `json:"intermediate_size"`
+		LaurelRank                uint32    `json:"laurel_rank"`
+		MaxPositionEmbeddings     uint32    `json:"max_position_embeddings"`
+		NumAttentionHeads         uint32    `json:"num_attention_heads"`
+		NumHiddenLayers           uint32    `json:"num_hidden_layers"`
+		NumKeyValueHeads          uint32    `json:"num_key_value_heads"`
+		NumKVSharedLayers         uint32    `json:"num_kv_shared_layers"`
+		RMSNormEPS                float32   `json:"rms_norm_eps"`
+		RopeLocalBaseFreq         float32   `json:"rope_local_base_freq"`
+		RopeTheta                 float32   `json:"rope_theta"`
+		SlidingWindow             uint32    `json:"sliding_window"`
+		LayerTypes                []string  `json:"layer_types"`
+	} `json:"text_config"`
+	VisionModel struct{} `json:"vision_config"`
+}
+
+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) {
+		norm := distuv.Normal{Mu: 0, Sigma: 1}
+		for _, v := range m.TextModel.ActivationSparsityPattern {
+			if !yield(float32(norm.Quantile(float64(v)))) {
+				break
+			}
+		}
+	})
+	kv["gemma3n.altup.active_idx"] = m.TextModel.AltupActiveIdx
+	kv["gemma3n.altup.correct_scale"] = m.TextModel.AltupCorrectScale
+	kv["gemma3n.altup.lr_multiplier"] = m.TextModel.AltupLRMultiplier
+	kv["gemma3n.altup.num_inputs"] = m.TextModel.AltupNumInputs
+	kv["gemma3n.attention.head_count_kv"] = m.TextModel.NumKeyValueHeads
+	kv["gemma3n.attention.head_count"] = m.TextModel.NumAttentionHeads
+	kv["gemma3n.attention.layer_norm_rms_epsilon"] = m.TextModel.RMSNormEPS
+	kv["gemma3n.attention.sliding_window"] = m.TextModel.SlidingWindow
+	kv["gemma3n.attention.sliding_window_pattern"] = slices.Collect(func(yield func(bool) bool) {
+		for _, t := range m.TextModel.LayerTypes {
+			if !yield(t == "sliding_attention") {
+				break
+			}
+		}
+	})
+	kv["gemma3n.attention.shared_kv_layers"] = m.TextModel.NumKVSharedLayers
+	kv["gemma3n.block_count"] = m.TextModel.NumHiddenLayers
+	kv["gemma3n.context_length"] = m.TextModel.MaxPositionEmbeddings
+	kv["gemma3n.embedding_length_per_layer_input"] = m.TextModel.HiddenSizePerLayerInput
+	kv["gemma3n.embedding_length"] = m.TextModel.HiddenSize
+	kv["gemma3n.feed_forward_length"] = m.TextModel.IntermediateSize
+	kv["gemma3n.head_dim"] = m.TextModel.HeadDim
+	kv["gemma3n.laurel_rank"] = m.TextModel.LaurelRank
+	kv["gemma3n.num_kv_shared_layers"] = m.TextModel.NumKVSharedLayers
+	kv["gemma3n.rope.freq_base_local"] = m.TextModel.RopeLocalBaseFreq
+	kv["gemma3n.rope.freq_base"] = m.TextModel.RopeTheta
+	return kv
+}
+
+func (m *gemma3nModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	out, ts := mergeTensors(ts,
+		merge{"altup_proj.*.weight", "altup_proj.weight"},
+		merge{"altup_unembd_proj.*.weight", "altup_unembd_proj.weight"},
+	)
+
+	for _, t := range ts {
+		switch {
+		case strings.Contains(t.Name(), "audio_tower"),
+			strings.Contains(t.Name(), "embed_audio"),
+			strings.Contains(t.Name(), "vision_tower"),
+			strings.Contains(t.Name(), "embed_vision"):
+			// TODO: handle audio and vision towers
+			continue
+		case strings.Contains(t.Name(), "altup_predict_coef"),
+			strings.Contains(t.Name(), "altup_correct_coef"):
+			if m.TextModel.AltupCoefClip > 0 {
+				t.SetRepacker(func(name string, data []float32, shape []uint64) (_ []float32, err error) {
+					dims := make([]int, len(shape))
+					for i := range shape {
+						dims[i] = int(shape[i])
+					}
+
+					var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+					t, err = tensor.Clamp(t, -m.TextModel.AltupCoefClip, m.TextModel.AltupCoefClip)
+					if err != nil {
+						return nil, err
+					}
+
+					if err := t.Reshape(t.Shape().TotalSize()); err != nil {
+						return nil, err
+					}
+
+					return native.VectorF32(t.(*tensor.Dense))
+				})
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (m *gemma3nModel) Replacements() []string {
+	return []string{
+		"model.language_model.embed_tokens_per_layer", "per_layer_token_embd",
+		"model.language_model.embed_tokens", "token_embd",
+		"model.language_model.per_layer_model_projection", "per_layer_model_proj",
+		"model.language_model.per_layer_projection_norm", "per_layer_proj_norm", "model.language_model.altup_projections", "altup_proj",
+		"model.language_model.altup_unembed_projections", "altup_unembd_proj",
+		"model.language_model.norm", "output_norm",
+		"model.language_model.layers", "blk",
+
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.k_norm", "attn_k_norm",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"mlp.down_proj", "ffn_down",
+		"post_feedforward_layernorm", "post_ffw_norm",
+		"per_layer_input_gate", "inp_gate",
+		"per_layer_projection", "proj",
+		"post_per_layer_input_norm", "post_norm",
+		"altup.", "altup_",
+		"modality_router", "router",
+		"prediction_coefs", "predict_coef",
+		"correction_coefs", "correct_coef",
+		"correct_output_scale", "correct_scale.weight",
+		"laurel.", "laurel_",
+		"linear_left", "l",
+		"linear_right", "r",
+		"post_laurel_norm", "post_norm",
+	}
+}

convert/convert_mixtral.go

@@ -2,9 +2,6 @@ package convert
 
 import (
 	"fmt"
-	"io"
-	"slices"
-	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
@@ -30,65 +27,38 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 }
 
 func (p *mixtralModel) Tensors(ts []Tensor) []*ggml.Tensor {
-	oldnew := []string{
-		"model.layers", "blk",
-		"w1", "ffn_gate_exps",
-		"w2", "ffn_down_exps",
-		"w3", "ffn_up_exps",
-	}
-
-	for i := range p.NumLocalExperts {
-		oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
-	}
-
-	// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
-	namer := strings.NewReplacer(oldnew...)
-	experts := make(map[string]experts)
-
-	// merge experts into a single tensor while removing them from ts
-	ts = slices.DeleteFunc(ts, func(t Tensor) bool {
-		if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
-			return false
-		}
-
-		name := namer.Replace(t.Name())
-		experts[name] = append(experts[name], t)
-		return true
-	})
-
-	var out []*ggml.Tensor
-	for n, e := range experts {
-		// TODO(mxyng): sanity check experts
-		out = append(out, &ggml.Tensor{
-			Name:     n,
-			Kind:     e[0].Kind(),
-			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
-			WriterTo: e,
+	merges := make([]merge, 0, p.NumHiddenLayers*6)
+	for i := range p.NumHiddenLayers {
+		merges = append(merges, merge{
+			fmt.Sprintf("blk.%d.*.w1.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w1.bias", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.bias", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w2.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w2.bias", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.bias", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w3.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}, merge{
+			fmt.Sprintf("blk.%d.*.w3.bias", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.bias", i),
 		})
 	}
 
+	out, ts := mergeTensors(ts, merges...)
 	return append(out, p.llamaModel.Tensors(ts)...)
 }
 
 func (p *mixtralModel) Replacements() []string {
 	return append(
 		p.llamaModel.Replacements(),
+		"model.layers", "blk",
 		"block_sparse_moe.gate", "ffn_gate_inp",
+		"block_sparse_moe.experts.", ".",
 	)
 }
-
-type experts []Tensor
-
-func (e experts) WriteTo(w io.Writer) (int64, error) {
-	// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
-	for _, t := range e {
-		// the canonical merged experts tensor stacks all experts along a new, 0 axis,
-		// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
-		// this accomplishes the same thing by writing each expert tensor in sequence
-		if _, err := t.WriteTo(w); err != nil {
-			return 0, err
-		}
-	}
-
-	return 0, nil
-}

convert/convert_qwen25vl.go

@@ -65,17 +65,17 @@ func (q *qwen25VLModel) Tensors(ts []Tensor) []*ggml.Tensor {
 	for _, t := range ts {
 		if strings.Contains(t.Name(), "patch_embed.proj") {
 			for t := range splitDim(t, 2,
-				strings.NewReplacer("patch_embed.proj", "patch_embd_0"),
-				strings.NewReplacer("patch_embed.proj", "patch_embd_1"),
+				split{Replacer: strings.NewReplacer("patch_embed.proj", "patch_embd_0")},
+				split{Replacer: strings.NewReplacer("patch_embed.proj", "patch_embd_1")},
 			) {
 				t.Shape = slices.DeleteFunc(t.Shape, func(i uint64) bool { return i == 1 })
 				out = append(out, t)
 			}
 		} else if strings.Contains(t.Name(), "attn.qkv") {
 			out = append(out, slices.Collect(splitDim(t, 0,
-				strings.NewReplacer("attn.qkv", "attn_q"),
-				strings.NewReplacer("attn.qkv", "attn_k"),
-				strings.NewReplacer("attn.qkv", "attn_v"),
+				split{Replacer: strings.NewReplacer("attn.qkv", "attn_q")},
+				split{Replacer: strings.NewReplacer("attn.qkv", "attn_k")},
+				split{Replacer: strings.NewReplacer("attn.qkv", "attn_v")},
 			))...)
 		} else {
 			out = append(out, &ggml.Tensor{

convert/tensor.go

@@ -1,56 +1,129 @@
 package convert
 
 import (
+	"cmp"
+	"io"
 	"iter"
+	"path"
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/fs/ggml"
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
 )
 
+type split struct {
+	*strings.Replacer
+	dim int
+
+	// fn is an optional function to apply to the tensor after slicing
+	fn func(tensor.Tensor) (tensor.Tensor, error)
+}
+
 // splitDim splits a tensor along a specified dimension into multiple tensors. The dimension
-// is split evenly based on the number of replacers provided.
-func splitDim(t Tensor, dim int, replacers ...*strings.Replacer) iter.Seq[*ggml.Tensor] {
+// is split evenly based on the number of replacers provided unless a specific count is given.
+func splitDim(t Tensor, dim int, splits ...split) iter.Seq[*ggml.Tensor] {
 	return func(yield func(*ggml.Tensor) bool) {
-		for i, replacer := range replacers {
+		var offset int
+		for _, split := range splits {
+			t := t.Clone()
 			shape := slices.Clone(t.Shape())
-			shape[dim] = shape[dim] / uint64(len(replacers))
+			shape[dim] = cmp.Or(uint64(split.dim), shape[dim]/uint64(len(splits)))
 
 			slice := slices.Repeat([]tensor.Slice{nil}, len(shape))
-			slice[dim] = tensor.S(i*int(shape[dim]), (i+1)*int(shape[dim]))
+			slice[dim] = tensor.S(offset, offset+int(shape[dim]))
+			offset += int(shape[dim])
 
-			tt := t.Clone()
-			tt.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
+			t.SetRepacker(func(_ string, data []float32, shape []uint64) ([]float32, error) {
 				dims := make([]int, len(shape))
 				for i := range shape {
 					dims[i] = int(shape[i])
 				}
 
-				var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-				t, err := t.Slice(slice...)
+				var tt tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+				tt, err := tt.Slice(slice...)
 				if err != nil {
 					return nil, err
 				}
 
-				t = tensor.Materialize(t)
+				tt = tensor.Materialize(tt)
+
+				if split.fn != nil {
+					tt, err = split.fn(tt)
+					if err != nil {
+						return nil, err
+					}
+				}
+
 				// flatten tensor so it can be written as a vector
-				if err := t.Reshape(t.Shape().TotalSize()); err != nil {
+				if err := tt.Reshape(tt.Shape().TotalSize()); err != nil {
 					return nil, err
 				}
 
-				return native.VectorF32(t.(*tensor.Dense))
+				return native.VectorF32(tt.(*tensor.Dense))
 			})
 
 			if !yield(&ggml.Tensor{
-				Name:     replacer.Replace(t.Name()),
+				Name:     split.Replace(t.Name()),
 				Kind:     t.Kind(),
 				Shape:    shape,
-				WriterTo: tt,
+				WriterTo: t,
 			}) {
 				break
 			}
 		}
 	}
 }
+
+type merge struct {
+	pattern, name string
+}
+
+// mergeTensors merges tensors that match a given pattern into a single tensor.
+func mergeTensors(unmatched []Tensor, merges ...merge) (out []*ggml.Tensor, _ []Tensor) {
+	var matched []Tensor
+	for i := range merges {
+		matched, unmatched = slicesSplitFunc(unmatched, func(t Tensor) bool {
+			matched, _ := path.Match(merges[i].pattern, t.Name())
+			return matched
+		})
+
+		if len(matched) > 0 {
+			out = append(out, &ggml.Tensor{
+				Name:     merges[i].name,
+				Kind:     matched[0].Kind(),
+				Shape:    append([]uint64{uint64(len(matched))}, matched[0].Shape()...),
+				WriterTo: mergeGroup(matched),
+			})
+		}
+	}
+
+	return out, unmatched
+}
+
+// slicesSplitFunc splits a slice into two slices based on a predicate function.
+func slicesSplitFunc[S ~[]E, E comparable](s S, fn func(e E) bool) (matched, unmatched S) {
+	for _, e := range s {
+		if fn(e) {
+			matched = append(matched, e)
+		} else {
+			unmatched = append(unmatched, e)
+		}
+	}
+
+	return matched, unmatched
+}
+
+type mergeGroup []Tensor
+
+func (g mergeGroup) WriteTo(w io.Writer) (int64, error) {
+	for _, t := range g {
+		if _, err := t.WriteTo(w); err != nil {
+			return 0, err
+		}
+	}
+
+	return 0, nil
+}

convert/tensor_test.go

@@ -0,0 +1,402 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"io"
+	"iter"
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/pdevine/tensor"
+)
+
+type fakeTensor struct {
+	name  string
+	shape []uint64
+	data  []float32
+
+	repacker Repacker
+}
+
+func (f fakeTensor) Name() string {
+	return f.name
+}
+
+func (f fakeTensor) Shape() []uint64 {
+	return f.shape
+}
+
+func (f fakeTensor) Kind() uint32 {
+	return 0
+}
+
+func (f *fakeTensor) SetRepacker(fn Repacker) {
+	f.repacker = fn
+}
+
+func (f fakeTensor) Clone() Tensor {
+	return &fakeTensor{
+		name:     f.name,
+		shape:    slices.Clone(f.shape),
+		data:     slices.Clone(f.data),
+		repacker: f.repacker,
+	}
+}
+
+func (f fakeTensor) WriteTo(w io.Writer) (n int64, err error) {
+	data := f.data
+	if f.repacker != nil {
+		data, err = f.repacker(f.name, data, f.shape)
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	if err := binary.Write(w, binary.LittleEndian, data); err != nil {
+		return 0, err
+	}
+
+	return int64(len(data) * 4), nil
+}
+
+func mul(shape []uint64) int {
+	n := 1
+	for _, dim := range shape {
+		n *= int(dim)
+	}
+	return n
+}
+
+func TestSplitDim(t *testing.T) {
+	r := fakeTensor{
+		name:  "a.b",
+		shape: []uint64{3, 4},
+		data:  []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11},
+	}
+
+	t.Run("no split", func(t *testing.T) {
+		for tt := range splitDim(&r, 0, split{Replacer: strings.NewReplacer("a", "x")}) {
+			if tt.Name != "x.b" {
+				t.Fatalf("expected name 'x', got '%s'", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 4}) {
+				t.Fatalf("expected shape [3, 4], got %v", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}) {
+				t.Fatalf("expected data [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], got %v", f32s)
+			}
+		}
+	})
+
+	t.Run("even split", func(t *testing.T) {
+		next, stop := iter.Pull(splitDim(&r, 1,
+			split{Replacer: strings.NewReplacer("a", "x")},
+			split{Replacer: strings.NewReplacer("b", "y")},
+		))
+		defer stop()
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "x.b" {
+				t.Fatal("expected name 'x.b', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 4, 5, 8, 9}) {
+				t.Fatal("expected data [0, 1, 4, 5, 8, 9], got", f32s)
+			}
+		}
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "a.y" {
+				t.Fatal("expected name 'a.y', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{2, 3, 6, 7, 10, 11}) {
+				t.Fatal("expected data [2, 3, 6, 7, 10, 11], got", f32s)
+			}
+		}
+	})
+
+	t.Run("uneven split", func(t *testing.T) {
+		next, stop := iter.Pull(splitDim(&r, 0,
+			split{Replacer: strings.NewReplacer("a", "x"), dim: 2},
+			split{Replacer: strings.NewReplacer("b", "y"), dim: 1},
+		))
+		defer stop()
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "x.b" {
+				t.Fatal("expected name 'x.b', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{2, 4}) {
+				t.Fatal("expected shape [2, 4], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 2, 3, 4, 5, 6, 7}) {
+				t.Fatal("expected data [0, 1, 2, 3, 4, 5, 6, 7], got", f32s)
+			}
+		}
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "a.y" {
+				t.Fatal("expected name 'a.y', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{1, 4}) {
+				t.Fatal("expected shape [1, 4], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{8, 9, 10, 11}) {
+				t.Fatal("expected data [8, 9, 10, 11], got", f32s)
+			}
+		}
+	})
+
+	t.Run("split with transpose", func(t *testing.T) {
+		next, stop := iter.Pull(splitDim(&r, 1,
+			split{Replacer: strings.NewReplacer("a", "x")},
+			split{Replacer: strings.NewReplacer("b", "y"), fn: func(tt tensor.Tensor) (tensor.Tensor, error) {
+				return tensor.Transpose(tt, 1, 0)
+			}},
+		))
+		defer stop()
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "x.b" {
+				t.Fatal("expected name 'x.b', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{0, 1, 4, 5, 8, 9}) {
+				t.Fatal("expected data [0, 1, 4, 5, 8, 9], got", f32s)
+			}
+		}
+
+		{
+			tt, ok := next()
+			if !ok {
+				t.Fatal("expected at least one split")
+			}
+
+			if tt.Name != "a.y" {
+				t.Fatal("expected name 'a.y', got", tt.Name)
+			}
+
+			if !slices.Equal(tt.Shape, []uint64{3, 2}) {
+				t.Fatal("expected shape [3, 2], got", tt.Shape)
+			}
+
+			var b bytes.Buffer
+			if _, err := tt.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, mul(tt.Shape))
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(f32s, []float32{2, 6, 10, 3, 7, 11}) {
+				t.Fatal("expected data [2, 6, 10, 3, 7, 11], got", f32s)
+			}
+		}
+	})
+}
+
+func TestMerge(t *testing.T) {
+	unmatched := []Tensor{
+		&fakeTensor{
+			name:  "a.0.b",
+			shape: []uint64{5, 2},
+			data:  []float32{10, 11, 12, 13, 14, 15, 16, 17, 18, 19},
+		},
+		&fakeTensor{
+			name:  "a.1.b",
+			shape: []uint64{5, 2},
+			data:  []float32{20, 21, 22, 23, 24, 25, 26, 27, 28, 29},
+		},
+		&fakeTensor{
+			name:  "c.0.d",
+			shape: []uint64{5, 2},
+			data:  []float32{30, 31, 32, 33, 34, 35, 36, 37, 38, 39},
+		},
+		&fakeTensor{
+			name:  "c.1.d",
+			shape: []uint64{5, 2},
+			data:  []float32{40, 41, 42, 43, 44, 45, 46, 47, 48, 49},
+		},
+		&fakeTensor{
+			name:  "e.0.f",
+			shape: []uint64{5, 2},
+			data:  []float32{50, 51, 52, 53, 54, 55, 56, 57, 58, 59},
+		},
+	}
+
+	checkMatched := func(t *testing.T, n int, matched []*ggml.Tensor) {
+		for i := range n {
+			got := matched[i]
+			if diff := cmp.Diff([]uint64{2, 5, 2}, got.Shape); diff != "" {
+				t.Errorf("unexpected (-want +got):\n%s", diff)
+			}
+
+			var b bytes.Buffer
+			if _, err := got.WriteTo(&b); err != nil {
+				t.Fatal(err)
+			}
+
+			f32s := make([]float32, 20)
+			if err := binary.Read(&b, binary.LittleEndian, &f32s); err != nil {
+				t.Fatal(err)
+			}
+
+			offset := 10 + (i * 20)
+			want := make([]float32, 20)
+			for j := range 20 {
+				want[j] = float32(offset + j)
+			}
+
+			if diff := cmp.Diff(want, f32s); diff != "" {
+				t.Errorf("unexpected data (-want +got):\n%s", diff)
+			}
+		}
+	}
+
+	t.Run("single merge", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"})
+		if len(unmatched) != 3 {
+			t.Error("expected 3 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 1 {
+			t.Error("expected 1 merged tensor, got", len(matched))
+		}
+
+		checkMatched(t, 1, matched)
+	})
+
+	t.Run("multiple merges", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"a.*.b", "a.b"}, merge{"c.*.d", "c.d"})
+		if len(unmatched) != 1 {
+			t.Error("expected 1 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 2 {
+			t.Error("expected 2 merged tensor, got", len(matched))
+		}
+
+		checkMatched(t, 2, matched)
+	})
+
+	t.Run("no match", func(t *testing.T) {
+		matched, unmatched := mergeTensors(unmatched, merge{"x.*.y", "x.y"})
+		if len(unmatched) != 5 {
+			t.Error("expected 5 remaining tensors, got", len(unmatched))
+		}
+
+		if len(matched) != 0 {
+			t.Error("expected no merged tensors, got", len(matched))
+		}
+	})
+}

discover/cuda_common.go

@@ -3,6 +3,7 @@
 package discover
 
 import (
+	"fmt"
 	"log/slog"
 	"os"
 	"regexp"
@@ -55,10 +56,13 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 				}
 			}
 		}
+		return "sbsa"
 	}
 
 	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
 	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
+		// The detected driver is older than Feb 2023
+		slog.Warn("old CUDA driver detected - please upgrade to a newer driver", "version", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor))
 		return "v11"
 	}
 	return "v12"

discover/path.go

@@ -12,7 +12,7 @@ import (
 // '../lib/ollama' on Linux and the executable's directory on macOS
 // note: distribution builds, additional GPU-specific libraries are
 // found in subdirectories of the returned path, such as
-// 'cuda_v11', 'cuda_v12', 'rocm', etc.
+// 'cuda_v12', 'rocm', etc.
 var LibOllamaPath string = func() string {
 	exe, err := os.Executable()
 	if err != nil {

docs/api.md

@@ -43,6 +43,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 - `prompt`: the prompt to generate a response for
 - `suffix`: the text after the model response
 - `images`: (optional) a list of base64-encoded images (for multimodal models such as `llava`)
+- `think`: (for thinking models) should the model think before responding?
 
 Advanced parameters (optional):
 
@@ -490,11 +491,13 @@ Generate the next message in a chat with a provided model. This is a streaming e
 - `model`: (required) the [model name](#model-names)
 - `messages`: the messages of the chat, this can be used to keep a chat memory
 - `tools`: list of tools in JSON for the model to use if supported
+- `think`: (for thinking models) should the model think before responding?
 
 The `message` object has the following fields:
 
 - `role`: the role of the message, either `system`, `user`, `assistant`, or `tool`
 - `content`: the content of the message
+- `thinking`: (for thinking models) the model's thinking process
 - `images` (optional): a list of images to include in the message (for multimodal models such as `llava`)
 - `tool_calls` (optional): a list of tools in JSON that the model wants to use
 

docs/benchmark.md

@@ -1,59 +0,0 @@
-# Benchmark
-
-Go benchmark tests that measure end-to-end performance of a running Ollama server. Run these tests to evaluate model inference performance on your hardware and measure the impact of code changes.
-
-## When to use
-
-Run these benchmarks when:
-- Making changes to the model inference engine
-- Modifying model loading/unloading logic
-- Changing prompt processing or token generation code
-- Implementing a new model architecture
-- Testing performance across different hardware setups
-
-## Prerequisites
-- Ollama server running locally with `ollama serve` on `127.0.0.1:11434`
-## Usage and Examples
-
->[!NOTE]
->All commands must be run from the root directory of the Ollama project.
-
-Basic syntax:
-```bash
-go test -bench=. ./benchmark/... -m $MODEL_NAME
-```
-
-Required flags:
-- `-bench=.`: Run all benchmarks
-- `-m`: Model name to benchmark
-
-Optional flags:
-- `-count N`: Number of times to run the benchmark (useful for statistical analysis)
-- `-timeout T`: Maximum time for the benchmark to run (e.g. "10m" for 10 minutes)
-
-Common usage patterns:
-
-Single benchmark run with a model specified:
-```bash
-go test -bench=. ./benchmark/... -m llama3.3
-```
-
-## Output metrics
-
-The benchmark reports several key metrics:
-
-- `gen_tok/s`: Generated tokens per second
-- `prompt_tok/s`: Prompt processing tokens per second
-- `ttft_ms`: Time to first token in milliseconds
-- `load_ms`: Model load time in milliseconds
-- `gen_tokens`: Total tokens generated
-- `prompt_tokens`: Total prompt tokens processed
-
-Each benchmark runs two scenarios:
-- Cold start: Model is loaded from disk for each test
-- Warm start: Model is pre-loaded in memory
-
-Three prompt lengths are tested for each scenario:
-- Short prompt (100 tokens)
-- Medium prompt (500 tokens)
-- Long prompt (1000 tokens)

docs/development.md

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

docs/gpu.md

@@ -1,6 +1,6 @@
 # GPU
 ## Nvidia
-Ollama supports Nvidia GPUs with compute capability 5.0+.
+Ollama supports Nvidia GPUs with compute capability 5.0+ and driver version 531 and newer.
 
 Check your compute compatibility to see if your card is supported:
 [https://developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus)

docs/linux.md

@@ -112,8 +112,8 @@ sudo systemctl status ollama
 > While AMD has contributed the `amdgpu` driver upstream to the official linux
 > kernel source, the version is older and may not support all ROCm features. We
 > recommend you install the latest driver from
-> https://www.amd.com/en/support/linux-drivers for best support of your Radeon
-> GPU.
+> [AMD](https://www.amd.com/en/support/download/linux-drivers.html) for best support
+> of your Radeon GPU.
 
 ## Customizing
 

docs/troubleshooting.md

@@ -43,7 +43,7 @@ Ollama includes multiple LLM libraries compiled for different GPUs and CPU vecto
 In the server log, you will see a message that looks something like this (varies from release to release):
 
 ```
-Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
+Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v12 rocm_v5]
 ```
 
 **Experimental LLM Library Override**

fs/config.go

@@ -10,4 +10,5 @@ type Config interface {
 	Strings(string, ...[]string) []string
 	Ints(string, ...[]int32) []int32
 	Floats(string, ...[]float32) []float32
+	Bools(string, ...[]bool) []bool
 }

fs/ggml/ggml.go

@@ -34,7 +34,8 @@ func (kv KV) Kind() string {
 }
 
 func (kv KV) ParameterCount() uint64 {
-	return keyValue(kv, "general.parameter_count", uint64(0))
+	val, _ := keyValue(kv, "general.parameter_count", uint64(0))
+	return val
 }
 
 func (kv KV) FileType() FileType {
@@ -53,16 +54,27 @@ func (kv KV) EmbeddingLength() uint64 {
 	return uint64(kv.Uint("embedding_length"))
 }
 
-func (kv KV) HeadCount() uint64 {
-	return uint64(kv.Uint("attention.head_count"))
+func (kv KV) HeadCountMax() uint64 {
+	// TODO(drifkin): using the max value can cause an overestimation. In the
+	// future if array values become more popular, we can adapt the more invasive
+	// <https://github.com/ollama/ollama/pull/10225>
+	return uint64(kv.UintOrMaxArrayValue("attention.head_count", 1))
 }
 
-func (kv KV) HeadCountKV() uint64 {
-	return uint64(kv.Uint("attention.head_count_kv", 1))
+func (kv KV) HeadCountMin() uint64 {
+	return uint64(kv.UintOrMinArrayValue("attention.head_count", 1))
 }
 
-func (kv KV) EmbeddingHeadCount() uint64 {
-	if heads := kv.HeadCount(); heads > 0 {
+func (kv KV) HeadCountKVMax() uint64 {
+	return uint64(kv.UintOrMaxArrayValue("attention.head_count_kv", 1))
+}
+
+func (kv KV) HeadCountKVMin() uint64 {
+	return uint64(kv.UintOrMinArrayValue("attention.head_count_kv", 1))
+}
+
+func (kv KV) EmbeddingHeadCountMax() uint64 {
+	if heads := kv.HeadCountMin(); heads > 0 {
 		return kv.EmbeddingLength() / heads
 	}
 
@@ -70,15 +82,11 @@ func (kv KV) EmbeddingHeadCount() uint64 {
 }
 
 func (kv KV) EmbeddingHeadCountK() uint64 {
-	return uint64(kv.Uint("attention.key_length", uint32(kv.EmbeddingHeadCount())))
+	return uint64(kv.Uint("attention.key_length", uint32(kv.EmbeddingHeadCountMax())))
 }
 
 func (kv KV) EmbeddingHeadCountV() uint64 {
-	return uint64(kv.Uint("attention.value_length", uint32(kv.EmbeddingHeadCount())))
-}
-
-func (kv KV) GQA() uint64 {
-	return kv.HeadCount() / kv.HeadCountKV()
+	return uint64(kv.Uint("attention.value_length", uint32(kv.EmbeddingHeadCountMax())))
 }
 
 func (kv KV) ContextLength() uint64 {
@@ -90,40 +98,83 @@ func (kv KV) ChatTemplate() string {
 }
 
 func (kv KV) String(key string, defaultValue ...string) string {
-	return keyValue(kv, key, append(defaultValue, "")...)
+	val, _ := keyValue(kv, key, append(defaultValue, "")...)
+	return val
 }
 
 func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
-	return keyValue(kv, key, append(defaultValue, 0)...)
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
 }
 
 func (kv KV) Float(key string, defaultValue ...float32) float32 {
-	return keyValue(kv, key, append(defaultValue, 0)...)
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
 }
 
 func (kv KV) Bool(key string, defaultValue ...bool) bool {
-	return keyValue(kv, key, append(defaultValue, false)...)
+	val, _ := keyValue(kv, key, append(defaultValue, false)...)
+	return val
+}
+
+func (kv KV) UintOrMaxArrayValue(key string, defaultValue uint32) uint32 {
+	_, max := kv.UintOrArrayValue(key, defaultValue)
+	return max
+}
+
+func (kv KV) UintOrMinArrayValue(key string, defaultValue uint32) uint32 {
+	min, _ := kv.UintOrArrayValue(key, defaultValue)
+	return min
+}
+
+func (kv KV) UintOrArrayValue(key string, defaultValue uint32) (uint32, uint32) {
+	if u32, ok := keyValue(kv, key, uint32(0)); ok {
+		return u32, u32
+	} else if u32s, ok := keyValue(kv, key, &array[uint32]{}); ok {
+		min := slices.Min(u32s.values)
+		max := slices.Max(u32s.values)
+		return min, max
+	} else if i32s, ok := keyValue(kv, key, &array[int32]{}); ok {
+		min := slices.Min(i32s.values)
+		max := slices.Max(i32s.values)
+		if min < 0 || max < 0 {
+			slog.Warn("array values are unexpectedly negative", "key", key, "min", min, "max", max)
+		}
+		return uint32(min), uint32(max)
+	}
+
+	return defaultValue, defaultValue
 }
 
 func (kv KV) Strings(key string, defaultValue ...[]string) []string {
-	return keyValue(kv, key, &array[string]{values: append(defaultValue, []string(nil))[0]}).values
+	val, _ := keyValue(kv, key, &array[string]{values: append(defaultValue, []string(nil))[0]})
+	return val.values
 }
 
 func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
-	return keyValue(kv, key, &array[int32]{values: append(defaultValue, []int32(nil))[0]}).values
+	val, _ := keyValue(kv, key, &array[int32]{values: append(defaultValue, []int32(nil))[0]})
+	return val.values
 }
 
 func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
-	return keyValue(kv, key, &array[uint32]{values: append(defaultValue, []uint32(nil))[0]}).values
+	val, _ := keyValue(kv, key, &array[uint32]{values: append(defaultValue, []uint32(nil))[0]})
+	return val.values
 }
 
 func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
-	return keyValue(kv, key, &array[float32]{values: append(defaultValue, []float32(nil))[0]}).values
+	val, _ := keyValue(kv, key, &array[float32]{values: append(defaultValue, []float32(nil))[0]})
+	return val.values
+}
+
+func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
+	val, _ := keyValue(kv, key, &array[bool]{values: append(defaultValue, []bool(nil))[0]})
+	return val.values
 }
 
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"gemma3",
+		"gemma3n",
 		"mistral3",
 		"llama4",
 		"mllama",
@@ -143,17 +194,17 @@ type arrayValueTypes interface {
 		*array[string] | *array[float32] | *array[float64] | *array[bool]
 }
 
-func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) T {
+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]; ok {
-		return val.(T)
+	if val, ok := kv[key].(T); ok {
+		return val, true
 	}
 
-	slog.Debug("key not found", "key", key, "default", defaultValue[0])
-	return defaultValue[0]
+	slog.Debug("key with type not found", "key", key, "default", defaultValue[0])
+	return defaultValue[0], false
 }
 
 type Tensors struct {
@@ -425,11 +476,11 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 
 func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
 	embedding := f.KV().EmbeddingLength()
-	heads := f.KV().HeadCount()
-	headsKV := f.KV().HeadCountKV()
+	heads := f.KV().HeadCountMax()
+	headsKV := f.KV().HeadCountKVMax()
 	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array[string]).size)
 
-	embeddingHeads := f.KV().EmbeddingHeadCount()
+	embeddingHeads := f.KV().EmbeddingHeadCountMax()
 	embeddingHeadsK := f.KV().EmbeddingHeadCountK()
 	embeddingHeadsV := f.KV().EmbeddingHeadCountV()
 
@@ -504,7 +555,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 			// vocab graph
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
-	case "gemma", "gemma2", "gemma3":
+	case "gemma", "gemma2", "gemma3", "gemma3n":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
 			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
@@ -517,6 +568,11 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 				embedding*embeddingHeadsK*heads*9/16,
 		)
 
+		if f.KV().Architecture() == "gemma3n" {
+			fullOffload *= 4
+			partialOffload *= 4
+		}
+
 		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
 		// engine. Gemma3 always uses the Ollama engine.
 		if f.KV().Architecture() == "gemma3" {

fs/ggml/ggml_test.go

@@ -269,3 +269,33 @@ func TestKeyValue(t *testing.T) {
 		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
 	}
 }
+
+func TestHeadCount(t *testing.T) {
+	valuesArray := []int32{1, 5, 3, 4}
+	cases := []struct {
+		kv   KV
+		want uint64
+	}{
+		{
+			kv: KV{
+				"general.architecture":     "abc",
+				"abc.attention.head_count": &array[int32]{values: valuesArray, size: len(valuesArray)},
+			},
+			want: uint64(5),
+		},
+		{
+			kv: KV{
+				"general.architecture":     "abc",
+				"abc.attention.head_count": uint32(3),
+			},
+			want: uint64(3),
+		},
+	}
+
+	for _, tt := range cases {
+		got := tt.kv.HeadCountMax()
+		if got != tt.want {
+			t.Errorf("unexpected max value: got=%d want=%d", got, tt.want)
+		}
+	}
+}

fs/ggml/gguf.go

@@ -527,23 +527,17 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		return err
 	}
 
-	keys := slices.Collect(maps.Keys(kv))
-	slices.Sort(keys)
-
-	for _, key := range keys {
+	for _, key := range slices.Sorted(maps.Keys(kv)) {
 		if err := ggufWriteKV(f, key, kv[key]); err != nil {
 			return err
 		}
 	}
 
 	slices.SortStableFunc(ts, func(a, b *Tensor) int {
-		if i, j := a.block(), b.block(); i < 0 && j > 0 {
-			return 1
-		} else if i > 0 && j < 0 {
-			return -1
-		} else {
+		if i, j := a.block(), b.block(); i > 0 && j > 0 {
 			return cmp.Compare(i, j)
 		}
+		return cmp.Compare(a.Name, b.Name)
 	})
 
 	var s uint64
@@ -615,6 +609,10 @@ func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
 		err = writeGGUFArray(ws, ggufTypeString, v)
 	case *array[string]:
 		err = writeGGUFArray(ws, ggufTypeString, v.values)
+	case []bool:
+		err = writeGGUFArray(ws, ggufTypeBool, v)
+	case *array[bool]:
+		err = writeGGUFArray(ws, ggufTypeBool, v.values)
 	default:
 		return fmt.Errorf("improper type for '%s'", k)
 	}

fs/ggml/gguf_test.go

@@ -2,62 +2,82 @@ package ggml
 
 import (
 	"bytes"
+	"math/rand/v2"
 	"os"
-	"slices"
+	"strings"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
 )
 
 func TestWriteGGUF(t *testing.T) {
-	w, err := os.CreateTemp(t.TempDir(), "*.bin")
-	if err != nil {
-		t.Fatal(err)
-	}
-	defer w.Close()
+	r := rand.New(rand.NewPCG(0, 0))
+	for range 8 {
+		t.Run("shuffle", func(t *testing.T) {
+			t.Parallel()
 
-	if err := WriteGGUF(w, KV{
-		"general.alignment": uint32(16),
-	}, []*Tensor{
-		{Name: "test.0", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
-		{Name: "test.1", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
-		{Name: "test.2", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
-		{Name: "test.3", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
-		{Name: "test.4", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
-		{Name: "test.5", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(slices.Repeat([]byte{0}, 2*3*4))},
-	}); err != nil {
-		t.Fatal(err)
-	}
+			ts := []*Tensor{
+				{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.1.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.2.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.3.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.4.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.5.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
+				{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
+			}
 
-	r, err := os.Open(w.Name())
-	if err != nil {
-		t.Fatal(err)
-	}
-	defer r.Close()
+			r.Shuffle(len(ts), func(i, j int) {
+				ts[i], ts[j] = ts[j], ts[i]
+			})
 
-	ff, err := Decode(r, 0)
-	if err != nil {
-		t.Fatal(err)
-	}
+			w, err := os.CreateTemp(t.TempDir(), strings.ReplaceAll(t.Name(), "/", "_")+"*.bin")
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer w.Close()
 
-	if diff := cmp.Diff(ff.KV(), KV{
-		"general.alignment":       uint32(16),
-		"general.parameter_count": uint64(36),
-	}); diff != "" {
-		t.Errorf("Mismatch (-want +got):\n%s", diff)
-	}
+			if err := WriteGGUF(w, KV{
+				"general.alignment": uint32(16),
+			}, ts); err != nil {
+				t.Fatal(err)
+			}
 
-	if diff := cmp.Diff(ff.Tensors(), Tensors{
-		Offset: 336,
-		items: []*Tensor{
-			{Name: "test.0", Offset: 0, Shape: []uint64{2, 3}},
-			{Name: "test.1", Offset: 32, Shape: []uint64{2, 3}},
-			{Name: "test.2", Offset: 64, Shape: []uint64{2, 3}},
-			{Name: "test.3", Offset: 96, Shape: []uint64{2, 3}},
-			{Name: "test.4", Offset: 128, Shape: []uint64{2, 3}},
-			{Name: "test.5", Offset: 160, Shape: []uint64{2, 3}},
-		},
-	}, cmp.AllowUnexported(Tensors{})); diff != "" {
-		t.Errorf("Mismatch (-want +got):\n%s", diff)
+			r, err := os.Open(w.Name())
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer r.Close()
+
+			ff, err := Decode(r, 0)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if diff := cmp.Diff(KV{
+				"general.alignment":       uint32(16),
+				"general.parameter_count": uint64(54),
+			}, ff.KV()); diff != "" {
+				t.Errorf("Mismatch (-want +got):\n%s", diff)
+			}
+
+			if diff := cmp.Diff(Tensors{
+				Offset: 608,
+				items: []*Tensor{
+					{Name: "blk.0.attn_norm.weight", Offset: 0, Shape: []uint64{2, 3}},
+					{Name: "blk.1.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
+					{Name: "blk.2.attn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
+					{Name: "blk.3.attn_norm.weight", Offset: 96, Shape: []uint64{2, 3}},
+					{Name: "blk.4.attn_norm.weight", Offset: 128, Shape: []uint64{2, 3}},
+					{Name: "blk.5.attn_norm.weight", Offset: 160, Shape: []uint64{2, 3}},
+					{Name: "output.weight", Offset: 192, Shape: []uint64{3, 2}},
+					{Name: "output_norm.weight", Offset: 224, Shape: []uint64{3, 2}},
+					{Name: "token_embd.weight", Offset: 256, Shape: []uint64{2, 3}},
+				},
+			}, ff.Tensors(), cmp.AllowUnexported(Tensors{})); diff != "" {
+				t.Errorf("Mismatch (-want +got):\n%s", diff)
+			}
+		})
 	}
 }

fs/gguf/gguf.go

@@ -0,0 +1,347 @@
+package gguf
+
+import (
+	"bytes"
+	"cmp"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+	"iter"
+	"os"
+	"slices"
+	"strings"
+)
+
+const (
+	typeUint8 uint32 = iota
+	typeInt8
+	typeUint16
+	typeInt16
+	typeUint32
+	typeInt32
+	typeFloat32
+	typeBool
+	typeString
+	typeArray
+	typeUint64
+	typeInt64
+	typeFloat64
+)
+
+var ErrUnsupported = errors.New("unsupported")
+
+type File struct {
+	Magic   [4]byte
+	Version uint32
+
+	keyValues *lazy[KeyValue]
+	tensors   *lazy[TensorInfo]
+	offset    int64
+
+	file   *os.File
+	reader *bufferedReader
+	bts    []byte
+}
+
+func Open(path string) (f *File, err error) {
+	f = &File{bts: make([]byte, 4096)}
+	f.file, err = os.Open(path)
+	if err != nil {
+		return nil, err
+	}
+
+	f.reader = newBufferedReader(f.file, 32<<10)
+
+	if err := binary.Read(f.reader, binary.LittleEndian, &f.Magic); err != nil {
+		return nil, err
+	}
+
+	if bytes.Equal(f.Magic[:], []byte("gguf")) {
+		return nil, fmt.Errorf("%w file type %v", ErrUnsupported, f.Magic)
+	}
+
+	if err := binary.Read(f.reader, binary.LittleEndian, &f.Version); err != nil {
+		return nil, err
+	}
+
+	if f.Version < 2 {
+		return nil, fmt.Errorf("%w version %v", ErrUnsupported, f.Version)
+	}
+
+	f.tensors, err = newLazy(f, f.readTensor)
+	if err != nil {
+		return nil, err
+	}
+
+	f.tensors.successFunc = func() error {
+		offset := f.reader.offset
+
+		alignment := cmp.Or(f.KeyValue("general.alignment").Int(), 32)
+		f.offset = offset + (alignment-offset%alignment)%alignment
+		return nil
+	}
+
+	f.keyValues, err = newLazy(f, f.readKeyValue)
+	if err != nil {
+		return nil, err
+	}
+
+	return f, nil
+}
+
+func (f *File) readTensor() (TensorInfo, error) {
+	name, err := readString(f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	dims, err := read[uint32](f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	shape := make([]uint64, dims)
+	for i := range dims {
+		shape[i], err = read[uint64](f)
+		if err != nil {
+			return TensorInfo{}, err
+		}
+	}
+
+	type_, err := read[uint32](f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	offset, err := read[uint64](f)
+	if err != nil {
+		return TensorInfo{}, err
+	}
+
+	return TensorInfo{
+		Name:   name,
+		Offset: offset,
+		Shape:  shape,
+		Type:   TensorType(type_),
+	}, nil
+}
+
+func (f *File) readKeyValue() (KeyValue, error) {
+	key, err := readString(f)
+	if err != nil {
+		return KeyValue{}, err
+	}
+
+	t, err := read[uint32](f)
+	if err != nil {
+		return KeyValue{}, err
+	}
+
+	value, err := func() (any, error) {
+		switch t {
+		case typeUint8:
+			return read[uint8](f)
+		case typeInt8:
+			return read[int8](f)
+		case typeUint16:
+			return read[uint16](f)
+		case typeInt16:
+			return read[int16](f)
+		case typeUint32:
+			return read[uint32](f)
+		case typeInt32:
+			return read[int32](f)
+		case typeUint64:
+			return read[uint64](f)
+		case typeInt64:
+			return read[int64](f)
+		case typeFloat32:
+			return read[float32](f)
+		case typeFloat64:
+			return read[float64](f)
+		case typeBool:
+			return read[bool](f)
+		case typeString:
+			return readString(f)
+		case typeArray:
+			return readArray(f)
+		default:
+			return nil, fmt.Errorf("%w type %d", ErrUnsupported, t)
+		}
+	}()
+	if err != nil {
+		return KeyValue{}, err
+	}
+
+	return KeyValue{
+		Key:   key,
+		Value: Value{value},
+	}, nil
+}
+
+func read[T any](f *File) (t T, err error) {
+	err = binary.Read(f.reader, binary.LittleEndian, &t)
+	return t, err
+}
+
+func readString(f *File) (string, error) {
+	n, err := read[uint64](f)
+	if err != nil {
+		return "", err
+	}
+
+	if int(n) > len(f.bts) {
+		f.bts = make([]byte, n)
+	}
+
+	bts := f.bts[:n]
+	if _, err := io.ReadFull(f.reader, bts); err != nil {
+		return "", err
+	}
+	defer clear(bts)
+
+	return string(bts), nil
+}
+
+func readArray(f *File) (any, error) {
+	t, err := read[uint32](f)
+	if err != nil {
+		return nil, err
+	}
+
+	n, err := read[uint64](f)
+	if err != nil {
+		return nil, err
+	}
+
+	switch t {
+	case typeUint8:
+		return readArrayData[uint8](f, n)
+	case typeInt8:
+		return readArrayData[int8](f, n)
+	case typeUint16:
+		return readArrayData[uint16](f, n)
+	case typeInt16:
+		return readArrayData[int16](f, n)
+	case typeUint32:
+		return readArrayData[uint32](f, n)
+	case typeInt32:
+		return readArrayData[int32](f, n)
+	case typeUint64:
+		return readArrayData[uint64](f, n)
+	case typeInt64:
+		return readArrayData[int64](f, n)
+	case typeFloat32:
+		return readArrayData[float32](f, n)
+	case typeFloat64:
+		return readArrayData[float64](f, n)
+	case typeBool:
+		return readArrayData[bool](f, n)
+	case typeString:
+		return readArrayString(f, n)
+	default:
+		return nil, fmt.Errorf("%w type %d", ErrUnsupported, t)
+	}
+}
+
+func readArrayData[T any](f *File, n uint64) (s []T, err error) {
+	s = make([]T, n)
+	for i := range n {
+		e, err := read[T](f)
+		if err != nil {
+			return nil, err
+		}
+
+		s[i] = e
+	}
+
+	return s, nil
+}
+
+func readArrayString(f *File, n uint64) (s []string, err error) {
+	s = make([]string, n)
+	for i := range n {
+		e, err := readString(f)
+		if err != nil {
+			return nil, err
+		}
+
+		s[i] = e
+	}
+
+	return s, nil
+}
+
+func (f *File) Close() error {
+	f.keyValues.stop()
+	f.tensors.stop()
+	return f.file.Close()
+}
+
+func (f *File) KeyValue(key string) KeyValue {
+	if !strings.HasPrefix(key, "general.") && !strings.HasPrefix(key, "tokenizer.") {
+		key = f.KeyValue("general.architecture").String() + "." + key
+	}
+
+	if index := slices.IndexFunc(f.keyValues.values, func(kv KeyValue) bool {
+		return kv.Key == key
+	}); index >= 0 {
+		return f.keyValues.values[index]
+	}
+
+	for keyValue, ok := f.keyValues.next(); ok; keyValue, ok = f.keyValues.next() {
+		if keyValue.Key == key {
+			return keyValue
+		}
+	}
+
+	return KeyValue{}
+}
+
+func (f *File) NumKeyValues() int {
+	return int(f.keyValues.count)
+}
+
+func (f *File) KeyValues() iter.Seq2[int, KeyValue] {
+	return f.keyValues.All()
+}
+
+func (f *File) TensorInfo(name string) TensorInfo {
+	if index := slices.IndexFunc(f.tensors.values, func(t TensorInfo) bool {
+		return t.Name == name
+	}); index >= 0 {
+		return f.tensors.values[index]
+	}
+
+	// fast-forward through key values if we haven't already
+	_ = f.keyValues.rest()
+	for tensor, ok := f.tensors.next(); ok; tensor, ok = f.tensors.next() {
+		if tensor.Name == name {
+			return tensor
+		}
+	}
+
+	return TensorInfo{}
+}
+
+func (f *File) NumTensors() int {
+	return int(f.tensors.count)
+}
+
+func (f *File) TensorInfos() iter.Seq2[int, TensorInfo] {
+	// fast forward through key values if we haven't already
+	f.keyValues.rest()
+	return f.tensors.All()
+}
+
+func (f *File) TensorReader(name string) (TensorInfo, io.Reader, error) {
+	t := f.TensorInfo(name)
+	if t.NumBytes() == 0 {
+		return TensorInfo{}, nil, fmt.Errorf("tensor %s not found", name)
+	}
+
+	// fast forward through tensor info if we haven't already
+	_ = f.tensors.rest()
+	return t, io.NewSectionReader(f.file, f.offset+int64(t.Offset), t.NumBytes()), nil
+}

fs/gguf/gguf_test.go

@@ -0,0 +1,249 @@
+package gguf_test
+
+import (
+	"bytes"
+	"os"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+	"github.com/google/go-cmp/cmp/cmpopts"
+	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/fs/gguf"
+)
+
+func createBinFile(tb testing.TB) string {
+	tb.Helper()
+	f, err := os.CreateTemp(tb.TempDir(), "")
+	if err != nil {
+		tb.Fatal(err)
+	}
+	defer f.Close()
+
+	kv := ggml.KV{
+		"general.architecture":                   "llama",
+		"llama.block_count":                      uint32(8),
+		"llama.embedding_length":                 uint32(3),
+		"llama.attention.head_count":             uint32(2),
+		"llama.attention.head_count_kv":          uint32(2),
+		"llama.attention.key_length":             uint32(3),
+		"llama.rope.dimension_count":             uint32(4),
+		"llama.rope.freq_base":                   float32(10000.0),
+		"llama.rope.freq_scale":                  float32(1.0),
+		"llama.attention.layer_norm_rms_epsilon": float32(1e-6),
+		"tokenizer.ggml.eos_token_id":            uint32(0),
+		"tokenizer.ggml.eos_token_ids":           []int32{1, 2, 3},
+		"tokenizer.ggml.tokens":                  []string{"hello", "world"},
+		"tokenizer.ggml.scores":                  []float32{0, 1},
+	}
+
+	tensors := []*ggml.Tensor{
+		{
+			Name:     "token_embd.weight",
+			Kind:     0,
+			Shape:    []uint64{2, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*2*3)),
+		},
+		{
+			Name:     "output.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 2},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*2)),
+		},
+	}
+
+	for i := range 8 {
+		tensors = append(tensors, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_q.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		}, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_k.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		}, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_v.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		}, &ggml.Tensor{
+			Name:     "blk." + strconv.Itoa(i) + ".attn_output.weight",
+			Kind:     0,
+			Shape:    []uint64{3, 3},
+			WriterTo: bytes.NewBuffer(make([]byte, 4*3*3)),
+		})
+	}
+
+	if err := ggml.WriteGGUF(f, kv, tensors); err != nil {
+		tb.Fatal(err)
+	}
+
+	return f.Name()
+}
+
+func TestRead(t *testing.T) {
+	f, err := gguf.Open(createBinFile(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	if got := f.KeyValue("does.not.exist").Valid(); got {
+		t.Errorf(`KeyValue("does.not.exist").Exists() = %v, want false`, got)
+	}
+
+	if got := f.KeyValue("general.architecture").String(); got != "llama" {
+		t.Errorf(`KeyValue("general.architecture").String() = %q, want %q`, got, "llama")
+	}
+
+	if got := f.TensorInfo("token_embd.weight"); got.Name != "token_embd.weight" {
+		t.Errorf(`TensorInfo("token_embd.weight").Name = %q, want %q`, got.Name, "token_embd.weight")
+	} else if diff := cmp.Diff(got.Shape, []uint64{2, 3}); diff != "" {
+		t.Errorf(`TensorInfo("token_embd.weight").Shape mismatch (-got +want):\n%s`, diff)
+	} else if got.Type != gguf.TensorTypeF32 {
+		t.Errorf(`TensorInfo("token_embd.weight").Type = %d, want %d`, got.Type, gguf.TensorTypeF32)
+	}
+
+	if got := f.KeyValue("block_count").Uint(); got != 8 {
+		t.Errorf(`KeyValue("block_count").Uint() = %d, want %d`, got, 8)
+	}
+
+	if diff := cmp.Diff(f.KeyValue("tokenizer.ggml.tokens").Strings(), []string{"hello", "world"}); diff != "" {
+		t.Errorf("KeyValue(\"tokenizer.ggml.tokens\").Strings() mismatch (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(f.KeyValue("tokenizer.ggml.scores").Floats(), []float64{0, 1}); diff != "" {
+		t.Errorf("KeyValue(\"tokenizer.ggml.scores\").Ints() mismatch (-got +want):\n%s", diff)
+	}
+
+	var kvs []string
+	for _, kv := range f.KeyValues() {
+		if !kv.Valid() {
+			t.Error("found invalid key-value pair:", kv)
+		}
+
+		kvs = append(kvs, kv.Key)
+	}
+
+	if len(kvs) != f.NumKeyValues() {
+		t.Errorf("iterated key count = %d, want %d", len(kvs), f.NumKeyValues())
+	}
+
+	if diff := cmp.Diff(kvs, []string{
+		"general.architecture",
+		"llama.block_count",
+		"llama.embedding_length",
+		"llama.attention.head_count",
+		"llama.attention.head_count_kv",
+		"llama.attention.key_length",
+		"llama.rope.dimension_count",
+		"llama.rope.freq_base",
+		"llama.rope.freq_scale",
+		"llama.attention.layer_norm_rms_epsilon",
+		"tokenizer.ggml.eos_token_id",
+		"tokenizer.ggml.eos_token_ids",
+		"tokenizer.ggml.tokens",
+		"tokenizer.ggml.scores",
+	}, cmpopts.SortSlices(strings.Compare)); diff != "" {
+		t.Errorf("KeyValues() mismatch (-got +want):\n%s", diff)
+	}
+
+	var tis []string
+	for _, ti := range f.TensorInfos() {
+		if !ti.Valid() {
+			t.Error("found invalid tensor info:", ti)
+		}
+
+		tis = append(tis, ti.Name)
+	}
+
+	if len(tis) != f.NumTensors() {
+		t.Errorf("iterated tensor count = %d, want %d", len(tis), f.NumTensors())
+	}
+
+	if diff := cmp.Diff(tis, []string{
+		"token_embd.weight",
+		"output.weight",
+		"blk.0.attn_q.weight",
+		"blk.0.attn_k.weight",
+		"blk.0.attn_v.weight",
+		"blk.0.attn_output.weight",
+		"blk.1.attn_q.weight",
+		"blk.1.attn_k.weight",
+		"blk.1.attn_v.weight",
+		"blk.1.attn_output.weight",
+		"blk.2.attn_q.weight",
+		"blk.2.attn_k.weight",
+		"blk.2.attn_v.weight",
+		"blk.2.attn_output.weight",
+		"blk.3.attn_q.weight",
+		"blk.3.attn_k.weight",
+		"blk.3.attn_v.weight",
+		"blk.3.attn_output.weight",
+		"blk.4.attn_q.weight",
+		"blk.4.attn_k.weight",
+		"blk.4.attn_v.weight",
+		"blk.4.attn_output.weight",
+		"blk.5.attn_q.weight",
+		"blk.5.attn_k.weight",
+		"blk.5.attn_v.weight",
+		"blk.5.attn_output.weight",
+		"blk.6.attn_q.weight",
+		"blk.6.attn_k.weight",
+		"blk.6.attn_v.weight",
+		"blk.6.attn_output.weight",
+		"blk.7.attn_q.weight",
+		"blk.7.attn_k.weight",
+		"blk.7.attn_v.weight",
+		"blk.7.attn_output.weight",
+	}, cmpopts.SortSlices(strings.Compare)); diff != "" {
+		t.Errorf("TensorInfos() mismatch (-got +want):\n%s", diff)
+	}
+
+	ti, r, err := f.TensorReader("output.weight")
+	if err != nil {
+		t.Fatalf(`TensorReader("output.weight") error: %v`, err)
+	}
+
+	if ti.Name != "output.weight" {
+		t.Errorf(`TensorReader("output.weight").Name = %q, want %q`, ti.Name, "output.weight")
+	} else if diff := cmp.Diff(ti.Shape, []uint64{3, 2}); diff != "" {
+		t.Errorf(`TensorReader("output.weight").Shape mismatch (-got +want):\n%s`, diff)
+	} else if ti.Type != gguf.TensorTypeF32 {
+		t.Errorf(`TensorReader("output.weight").Type = %d, want %d`, ti.Type, gguf.TensorTypeF32)
+	}
+
+	var b bytes.Buffer
+	if _, err := b.ReadFrom(r); err != nil {
+		t.Fatalf(`ReadFrom TensorReader("output.weight") error: %v`, err)
+	}
+
+	if b.Len() != int(ti.NumBytes()) {
+		t.Errorf(`ReadFrom TensorReader("output.weight") length = %d, want %d`, b.Len(), ti.NumBytes())
+	}
+}
+
+func BenchmarkRead(b *testing.B) {
+	b.ReportAllocs()
+
+	p := createBinFile(b)
+	for b.Loop() {
+		f, err := gguf.Open(p)
+		if err != nil {
+			b.Fatal(err)
+		}
+
+		if got := f.KeyValue("general.architecture").String(); got != "llama" {
+			b.Errorf("got = %q, want %q", got, "llama")
+		}
+
+		// Iterate through some tensors
+		for range f.TensorInfos() {
+		}
+
+		f.Close()
+	}
+}

fs/gguf/keyvalue.go

@@ -0,0 +1,90 @@
+package gguf
+
+import (
+	"reflect"
+	"slices"
+)
+
+type KeyValue struct {
+	Key string
+	Value
+}
+
+func (kv KeyValue) Valid() bool {
+	return kv.Key != "" && kv.Value.value != nil
+}
+
+type Value struct {
+	value any
+}
+
+func value[T any](v Value, kinds ...reflect.Kind) (t T) {
+	vv := reflect.ValueOf(v.value)
+	if slices.Contains(kinds, vv.Kind()) {
+		t = vv.Convert(reflect.TypeOf(t)).Interface().(T)
+	}
+	return
+}
+
+func values[T any](v Value, kinds ...reflect.Kind) (ts []T) {
+	switch vv := reflect.ValueOf(v.value); vv.Kind() {
+	case reflect.Slice:
+		if slices.Contains(kinds, vv.Type().Elem().Kind()) {
+			ts = make([]T, vv.Len())
+			for i := range vv.Len() {
+				ts[i] = vv.Index(i).Convert(reflect.TypeOf(ts[i])).Interface().(T)
+			}
+		}
+	}
+	return
+}
+
+// Int returns Value as a signed integer. If it is not a signed integer, it returns 0.
+func (v Value) Int() int64 {
+	return value[int64](v, reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64)
+}
+
+// Ints returns Value as a signed integer slice. If it is not a signed integer slice, it returns nil.
+func (v Value) Ints() (i64s []int64) {
+	return values[int64](v, reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64)
+}
+
+// Uint converts an unsigned integer value to uint64. If the value is not a unsigned integer, it returns 0.
+func (v Value) Uint() uint64 {
+	return value[uint64](v, reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64)
+}
+
+// Uints returns Value as a unsigned integer slice. If it is not a unsigned integer slice, it returns nil.
+func (v Value) Uints() (u64s []uint64) {
+	return values[uint64](v, reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64)
+}
+
+// Float returns Value as a float. If it is not a float, it returns 0.
+func (v Value) Float() float64 {
+	return value[float64](v, reflect.Float32, reflect.Float64)
+}
+
+// Floats returns Value as a float slice. If it is not a float slice, it returns nil.
+func (v Value) Floats() (f64s []float64) {
+	return values[float64](v, reflect.Float32, reflect.Float64)
+}
+
+// Bool returns Value as a boolean. If it is not a boolean, it returns false.
+func (v Value) Bool() bool {
+	return value[bool](v, reflect.Bool)
+}
+
+// Bools returns Value as a boolean slice. If it is not a boolean slice, it returns nil.
+func (v Value) Bools() (bools []bool) {
+	return values[bool](v, reflect.Bool)
+}
+
+// String returns Value as a string. If it is not a string, it returns an empty string.
+func (v Value) String() string {
+	return value[string](v, reflect.String)
+}
+
+// Strings returns Value as a string slice. If it is not a string slice, it returns nil.
+func (v Value) Strings() (strings []string) {
+	return values[string](v, reflect.String)
+}

fs/gguf/keyvalue_test.go

@@ -0,0 +1,208 @@
+package gguf
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+func split(name string, values map[string][]any) (matched []any, unmatched []any) {
+	for key, value := range values {
+		if key == name {
+			matched = value
+		} else {
+			unmatched = append(unmatched, value...)
+		}
+	}
+	return
+}
+
+func TestValue(t *testing.T) {
+	values := map[string][]any{
+		"int64":   {int(42), int8(42), int16(42), int32(42), int64(42)},
+		"uint64":  {uint(42), uint8(42), uint16(42), uint32(42), uint64(42)},
+		"float64": {float32(42), float64(42)},
+		"string":  {"42", "hello"},
+		"bool":    {true, false},
+	}
+
+	t.Run("int64", func(t *testing.T) {
+		matched, unmatched := split("int64", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if i64 := kv.Int(); i64 != 42 {
+				t.Errorf("expected 42, got %d", i64)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if i64 := kv.Int(); i64 != 0 {
+				t.Errorf("expected 42, got %d", i64)
+			}
+		}
+	})
+
+	t.Run("uint64", func(t *testing.T) {
+		matched, unmatched := split("uint64", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if u64 := kv.Uint(); u64 != 42 {
+				t.Errorf("expected 42, got %d", u64)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if u64 := kv.Uint(); u64 != 0 {
+				t.Errorf("expected 42, got %d", u64)
+			}
+		}
+	})
+
+	t.Run("float64", func(t *testing.T) {
+		matched, unmatched := split("float64", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if f64 := kv.Float(); f64 != 42 {
+				t.Errorf("expected 42, got %f", f64)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if f64 := kv.Float(); f64 != 0 {
+				t.Errorf("expected 42, got %f", f64)
+			}
+		}
+	})
+
+	t.Run("string", func(t *testing.T) {
+		matched, unmatched := split("string", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if s := kv.String(); s != v {
+				t.Errorf("expected 42, got %s", s)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if s := kv.String(); s != "" {
+				t.Errorf("expected 42, got %s", s)
+			}
+		}
+	})
+
+	t.Run("bool", func(t *testing.T) {
+		matched, unmatched := split("bool", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if b := kv.Bool(); b != v {
+				t.Errorf("expected true, got %v", b)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if b := kv.Bool(); b != false {
+				t.Errorf("expected false, got %v", b)
+			}
+		}
+	})
+}
+
+func TestValues(t *testing.T) {
+	values := map[string][]any{
+		"int64s":   {[]int{42}, []int8{42}, []int16{42}, []int32{42}, []int64{42}},
+		"uint64s":  {[]uint{42}, []uint8{42}, []uint16{42}, []uint32{42}, []uint64{42}},
+		"float64s": {[]float32{42}, []float64{42}},
+		"strings":  {[]string{"42"}, []string{"hello"}},
+		"bools":    {[]bool{true}, []bool{false}},
+	}
+
+	t.Run("int64s", func(t *testing.T) {
+		matched, unmatched := split("int64s", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Ints(), []int64{42}); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if i64s := kv.Ints(); i64s != nil {
+				t.Errorf("expected nil, got %v", i64s)
+			}
+		}
+	})
+
+	t.Run("uint64s", func(t *testing.T) {
+		matched, unmatched := split("uint64s", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Uints(), []uint64{42}); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if u64s := kv.Uints(); u64s != nil {
+				t.Errorf("expected nil, got %v", u64s)
+			}
+		}
+	})
+
+	t.Run("float64s", func(t *testing.T) {
+		matched, unmatched := split("float64s", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Floats(), []float64{42}); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if f64s := kv.Floats(); f64s != nil {
+				t.Errorf("expected nil, got %v", f64s)
+			}
+		}
+	})
+
+	t.Run("strings", func(t *testing.T) {
+		matched, unmatched := split("strings", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Strings(), v); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if s := kv.Strings(); s != nil {
+				t.Errorf("expected nil, got %v", s)
+			}
+		}
+	})
+
+	t.Run("bools", func(t *testing.T) {
+		matched, unmatched := split("bools", values)
+		for _, v := range matched {
+			kv := KeyValue{"key", Value{v}}
+			if diff := cmp.Diff(kv.Bools(), v); diff != "" {
+				t.Errorf("diff: %s", diff)
+			}
+		}
+
+		for _, v := range unmatched {
+			kv := KeyValue{"key", Value{v}}
+			if b := kv.Bools(); b != nil {
+				t.Errorf("expected nil, got %v", b)
+			}
+		}
+	})
+}

fs/gguf/lazy.go

@@ -0,0 +1,89 @@
+package gguf
+
+import (
+	"encoding/binary"
+	"iter"
+	"log/slog"
+)
+
+type lazy[T any] struct {
+	count  uint64
+	next   func() (T, bool)
+	stop   func()
+	values []T
+
+	// successFunc is called when all values have been successfully read.
+	successFunc func() error
+}
+
+func newLazy[T any](f *File, fn func() (T, error)) (*lazy[T], error) {
+	it := lazy[T]{}
+	if err := binary.Read(f.reader, binary.LittleEndian, &it.count); err != nil {
+		return nil, err
+	}
+
+	it.values = make([]T, 0)
+	it.next, it.stop = iter.Pull(func(yield func(T) bool) {
+		for i := range it.count {
+			t, err := fn()
+			if err != nil {
+				slog.Error("error reading tensor", "index", i, "error", err)
+				return
+			}
+
+			it.values = append(it.values, t)
+			if !yield(t) {
+				break
+			}
+		}
+
+		if it.successFunc != nil {
+			it.successFunc()
+		}
+	})
+
+	return &it, nil
+}
+
+func (g *lazy[T]) Values() iter.Seq[T] {
+	return func(yield func(T) bool) {
+		for _, v := range g.All() {
+			if !yield(v) {
+				break
+			}
+		}
+	}
+}
+
+func (g *lazy[T]) All() iter.Seq2[int, T] {
+	return func(yield func(int, T) bool) {
+		for i := range int(g.count) {
+			if i < len(g.values) {
+				if !yield(i, g.values[i]) {
+					break
+				}
+			} else {
+				t, ok := g.next()
+				if !ok {
+					break
+				}
+
+				if !yield(i, t) {
+					break
+				}
+			}
+		}
+	}
+}
+
+func (g *lazy[T]) rest() (collected bool) {
+	for {
+		_, ok := g.next()
+		collected = collected || ok
+		if !ok {
+			break
+		}
+	}
+
+	return collected
+}

fs/gguf/reader.go

@@ -0,0 +1,23 @@
+package gguf
+
+import (
+	"bufio"
+	"io"
+)
+
+type bufferedReader struct {
+	offset int64
+	*bufio.Reader
+}
+
+func newBufferedReader(rs io.ReadSeeker, size int) *bufferedReader {
+	return &bufferedReader{
+		Reader: bufio.NewReaderSize(rs, size),
+	}
+}
+
+func (rs *bufferedReader) Read(p []byte) (n int, err error) {
+	n, err = rs.Reader.Read(p)
+	rs.offset += int64(n)
+	return n, err
+}

fs/gguf/tensor.go

@@ -0,0 +1,288 @@
+package gguf
+
+import (
+	"log/slog"
+	"strings"
+)
+
+type TensorInfo struct {
+	Name   string
+	Offset uint64
+	Shape  []uint64
+	Type   TensorType
+}
+
+func (ti TensorInfo) Valid() bool {
+	return ti.Name != "" && ti.NumBytes() > 0
+}
+
+func (ti TensorInfo) NumValues() int64 {
+	var numItems int64 = 1
+	for _, dim := range ti.Shape {
+		numItems *= int64(dim)
+	}
+	return numItems
+}
+
+// NumBytes returns the number of bytes in the tensor.
+func (ti TensorInfo) NumBytes() int64 {
+	return int64(float64(ti.NumValues()) * ti.Type.NumBytes())
+}
+
+func (ti TensorInfo) LogValue() slog.Value {
+	return slog.GroupValue(
+		slog.String("name", ti.Name),
+		slog.Int64("offset", int64(ti.Offset)),
+		slog.Any("shape", ti.Shape),
+		slog.Int64("num_values", ti.NumValues()),
+		slog.Int64("num_bytes", ti.NumBytes()),
+		slog.Any("type", ti.Type),
+	)
+}
+
+type TensorType uint32
+
+const (
+	TensorTypeF32 TensorType = iota
+	TensorTypeF16
+	TensorTypeQ4_0
+	TensorTypeQ4_1
+
+	// unexported // unused in gguf
+	tensorTypeQ4_2
+	tensorTypeQ4_3
+
+	TensorTypeQ5_0
+	TensorTypeQ5_1
+	TensorTypeQ8_0
+	TensorTypeQ8_1
+	TensorTypeQ2_K
+	TensorTypeQ3_K
+	TensorTypeQ4_K
+	TensorTypeQ5_K
+	TensorTypeQ6_K
+	TensorTypeQ8_K
+
+	// unexported // unquantizable by ollama
+	tensorTypeIQ2_XXS
+	tensorTypeIQ2_XS
+	tensorTypeIQ3_XXS
+	tensorTypeIQ1_S
+	tensorTypeIQ4_NL
+	tensorTypeIQ3_S
+	tensorTypeIQ2_S
+	tensorTypeIQ4_XS
+
+	TensorTypeI8
+	TensorTypeI16
+	TensorTypeI32
+	TensorTypeI64
+	TensorTypeF64
+
+	// unexported // unquantizable by ollama
+	tensorTypeIQ1_M
+
+	TensorTypeBF16
+
+	// unexported // unused in gguf
+	tensorTypeQ4_0_4_4
+	tensorTypeQ4_0_4_8
+	tensorTypeQ4_0_8_8
+
+	// unexported // unquantizable by ollama
+	tensorTypeTQ1_0
+	tensorTypeTQ2_0
+
+	// unexported // unused in gguf
+	tensorTypeIQ4_NL_4_4
+	tensorTypeIQ4_NL_4_8
+	tensorTypeIQ4_NL_8_8
+)
+
+func (tt TensorType) NumBytes() float64 {
+	return float64(tt.typeSize()) / float64(tt.blockSize())
+}
+
+func (tt TensorType) typeSize() int64 {
+	switch tt {
+	case TensorTypeF32:
+		return 4
+	case TensorTypeF16:
+		return 2
+	case TensorTypeQ4_0:
+		return 2 + tt.blockSize()/2
+	case TensorTypeQ4_1:
+		return 2 + 2 + tt.blockSize()/2
+	case TensorTypeQ5_0:
+		return 2 + 4 + tt.blockSize()/2
+	case TensorTypeQ5_1:
+		return 2 + 2 + 4 + tt.blockSize()/2
+	case TensorTypeQ8_0:
+		return 2 + tt.blockSize()
+	case TensorTypeQ8_1:
+		return 2 + 2 + tt.blockSize()
+	case TensorTypeQ2_K:
+		return tt.blockSize()/16 + tt.blockSize()/4 + 2 + 2
+	case TensorTypeQ3_K:
+		return tt.blockSize()/8 + tt.blockSize()/4 + 12 + 2
+	case TensorTypeQ4_K:
+		return 2 + 2 + 12 + tt.blockSize()/2
+	case TensorTypeQ5_K:
+		return 2 + 2 + 12 + tt.blockSize()/8 + tt.blockSize()/2
+	case TensorTypeQ6_K:
+		return tt.blockSize()/2 + tt.blockSize()/4 + tt.blockSize()/16 + 2
+	case TensorTypeQ8_K:
+		return 4 + tt.blockSize() + 2*tt.blockSize()/16
+	case tensorTypeIQ2_XXS:
+		return 2 + 2*tt.blockSize()/8
+	case tensorTypeIQ2_XS:
+		return 2 + 2*tt.blockSize()/8 + tt.blockSize()/32
+	case tensorTypeIQ3_XXS:
+		return 2 + tt.blockSize()/4 + tt.blockSize()/8
+	case tensorTypeIQ1_S:
+		return 2 + tt.blockSize()/8 + tt.blockSize()/16
+	case tensorTypeIQ4_NL:
+		return 2 + tt.blockSize()/2
+	case tensorTypeIQ3_S:
+		return 2 + tt.blockSize()/4 + tt.blockSize()/8 + tt.blockSize()/32 + 4
+	case tensorTypeIQ2_S:
+		return 2 + tt.blockSize()/4 + tt.blockSize()/16
+	case tensorTypeIQ4_XS:
+		return 2 + 2 + tt.blockSize()/2 + tt.blockSize()/64
+	case TensorTypeI8:
+		return 1
+	case TensorTypeI16:
+		return 2
+	case TensorTypeI32:
+		return 4
+	case TensorTypeI64:
+		return 8
+	case TensorTypeF64:
+		return 8
+	case tensorTypeIQ1_M:
+		return tt.blockSize()/8 + tt.blockSize()/16 + tt.blockSize()/32
+	case TensorTypeBF16:
+		return 2
+	default:
+		return 0
+	}
+}
+
+func (tt TensorType) blockSize() int64 {
+	switch tt {
+	case TensorTypeF32,
+		TensorTypeF16,
+		TensorTypeI8,
+		TensorTypeI16,
+		TensorTypeI32,
+		TensorTypeI64,
+		TensorTypeF64,
+		TensorTypeBF16:
+		return 1
+	case TensorTypeQ4_0,
+		TensorTypeQ4_1,
+		TensorTypeQ5_0,
+		TensorTypeQ5_1,
+		TensorTypeQ8_0,
+		TensorTypeQ8_1,
+		tensorTypeIQ4_NL:
+		return 32
+	default:
+		return 256
+	}
+}
+
+func (tt TensorType) String() string {
+	switch tt {
+	case TensorTypeF32:
+		return "f32"
+	case TensorTypeF16:
+		return "f16"
+	case TensorTypeQ4_0:
+		return "q4_0"
+	case TensorTypeQ4_1:
+		return "q4_1"
+	case tensorTypeQ4_2:
+		return "q4_2"
+	case tensorTypeQ4_3:
+		return "q4_3"
+	case TensorTypeQ5_0:
+		return "q5_0"
+	case TensorTypeQ5_1:
+		return "q5_1"
+	case TensorTypeQ8_0:
+		return "q8_0"
+	case TensorTypeQ8_1:
+		return "q8_1"
+	case TensorTypeQ2_K:
+		return "q2_k"
+	case TensorTypeQ3_K:
+		return "q3_k"
+	case TensorTypeQ4_K:
+		return "q4_k"
+	case TensorTypeQ5_K:
+		return "q5_k"
+	case TensorTypeQ6_K:
+		return "q6_k"
+	case TensorTypeQ8_K:
+		return "q8_k"
+	case tensorTypeIQ2_XXS:
+		return "iq2_xxs"
+	case tensorTypeIQ2_XS:
+		return "iq2_xs"
+	case tensorTypeIQ3_XXS:
+		return "iq3_xxs"
+	case tensorTypeIQ1_S:
+		return "iq1_s"
+	case tensorTypeIQ4_NL:
+		return "iq4_nl"
+	case tensorTypeIQ3_S:
+		return "iq3_s"
+	case tensorTypeIQ2_S:
+		return "iq2_s"
+	case tensorTypeIQ4_XS:
+		return "iq4_xs"
+	case TensorTypeI8:
+		return "i8"
+	case TensorTypeI16:
+		return "i16"
+	case TensorTypeI32:
+		return "i32"
+	case TensorTypeI64:
+		return "i64"
+	case TensorTypeF64:
+		return "f64"
+	case tensorTypeIQ1_M:
+		return "iq1_m"
+	case TensorTypeBF16:
+		return "bf16"
+	case tensorTypeQ4_0_4_4:
+		return "q4_0_4_4"
+	case tensorTypeQ4_0_4_8:
+		return "q4_0_4_8"
+	case tensorTypeQ4_0_8_8:
+		return "q4_0_8_8"
+	case tensorTypeTQ1_0:
+		return "tq1_0"
+	case tensorTypeTQ2_0:
+		return "tq2_0"
+	case tensorTypeIQ4_NL_4_4:
+		return "iq4_nl_4_4"
+	case tensorTypeIQ4_NL_4_8:
+		return "iq4_nl_4_8"
+	case tensorTypeIQ4_NL_8_8:
+		return "iq4_nl_8_8"
+	default:
+		return "unknown"
+	}
+}
+
+func (tt TensorType) LogValue() slog.Value {
+	return slog.GroupValue(
+		slog.Uint64("value", uint64(tt)),
+		slog.String("name", strings.ToUpper(tt.String())),
+		slog.Int64("size", tt.typeSize()),
+		slog.Int64("block_size", tt.blockSize()),
+		slog.Float64("num_bytes", tt.NumBytes()),
+	)
+}

go.mod

@@ -19,12 +19,13 @@ require (
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/dlclark/regexp2 v1.11.4
 	github.com/emirpasic/gods/v2 v2.0.0-alpha
-	github.com/google/go-cmp v0.6.0
+	github.com/google/go-cmp v0.7.0
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	golang.org/x/image v0.22.0
 	golang.org/x/tools v0.30.0
+	gonum.org/v1/gonum v0.15.0
 )
 
 require (
@@ -44,7 +45,6 @@ require (
 	github.com/xtgo/set v1.0.0 // indirect
 	go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6 // indirect
 	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect
-	gonum.org/v1/gonum v0.15.0 // indirect
 	gorgonia.org/vecf32 v0.9.0 // indirect
 	gorgonia.org/vecf64 v0.9.0 // indirect
 )

go.sum

@@ -112,8 +112,8 @@ github.com/google/go-cmp v0.4.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/go-cmp v0.5.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.6/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
-github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
-github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
+github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
 github.com/google/uuid v1.1.2/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=

integration/model_arch_test.go

@@ -45,6 +45,8 @@ var (
 		"qwen2.5-coder:latest",
 		"qwen:latest",
 		"solar-pro:latest",
+		"codellama:latest",
+		"nous-hermes:latest",
 	}
 )
 

llama/patches/0005-solar-pro.patch

@@ -150,7 +150,7 @@ index 4cce5166..7f6617fa 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 3a4e72a3..831b68c0 100644
+index 3a4e72a3..db62973f 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
 @@ -1402,6 +1402,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {

llama/patches/0008-ensure-KV-cache-is-fully-defragmented.patch

@@ -22,10 +22,10 @@ multiple batches of processing until everything is complete.
  4 files changed, 59 insertions(+), 79 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index c22687e4..c5948e8f 100644
+index dca22d8b..1f3a3956 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -950,9 +950,12 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -947,9 +947,12 @@ int llama_context::decode(llama_batch & inp_batch) {
  
          // find KV slot
          if (!kv_self->find_slot(ubatch)) {
@@ -41,7 +41,7 @@ index c22687e4..c5948e8f 100644
          }
  
          ggml_backend_sched_reset(sched.get());
-@@ -1967,9 +1970,12 @@ void llama_context::opt_epoch_iter(
+@@ -1965,9 +1968,12 @@ void llama_context::opt_epoch_iter(
  
              // TODO: not sure if this is needed
              if (!kv_self->find_slot(ubatch)) {

llama/patches/0015-add-argsort-and-cuda-copy-for-i32.patch

@@ -10,10 +10,10 @@ Subject: [PATCH] add argsort and cuda copy for i32
  3 files changed, 192 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index becdae07..7a44b6cf 100644
+index 955fec59..654e2f28 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -6890,6 +6890,45 @@ static void ggml_compute_forward_argsort_f32(
+@@ -6822,6 +6822,45 @@ static void ggml_compute_forward_argsort_f32(
      }
  }
  
@@ -59,7 +59,7 @@ index becdae07..7a44b6cf 100644
  void ggml_compute_forward_argsort(
      const ggml_compute_params * params,
      ggml_tensor * dst) {
-@@ -6901,6 +6940,10 @@ void ggml_compute_forward_argsort(
+@@ -6833,6 +6872,10 @@ void ggml_compute_forward_argsort(
              {
                  ggml_compute_forward_argsort_f32(params, dst);
              } break;
@@ -195,7 +195,7 @@ index 607ded85..53b02634 100644
 +    }
  }
 diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
-index 2d46176e..47383486 100644
+index d027271f..4abd01d7 100644
 --- a/ggml/src/ggml-cuda/cpy.cu
 +++ b/ggml/src/ggml-cuda/cpy.cu
 @@ -38,6 +38,13 @@ static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
@@ -257,7 +257,7 @@ index 2d46176e..47383486 100644
  static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
      const float * xi = (const float *) cxi;
      block_q8_0 * dsti = (block_q8_0 *) cdsti;
-@@ -631,6 +676,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
+@@ -633,6 +678,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
          ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
      } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
          ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
@@ -266,7 +266,7 @@ index 2d46176e..47383486 100644
      } else {
          GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                  ggml_type_name(src0->type), ggml_type_name(src1->type));
-@@ -686,6 +733,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
+@@ -688,6 +735,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
          return (void*) cpy_f32_f16<cpy_1_f32_f16>;
      } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
          return (void*) cpy_f32_f16<cpy_1_f16_f32>;

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

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

llama/patches/0018-temporary-prevent-rocm-cuda-mixed-loading.patch

@@ -0,0 +1,32 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Daniel Hiltgen <daniel@ollama.com>
+Date: Sun, 22 Jun 2025 09:22:05 -0700
+Subject: [PATCH] temporary prevent rocm+cuda mixed loading
+
+---
+ ggml/src/ggml-backend-reg.cpp | 12 ++++++++++--
+ 1 file changed, 10 insertions(+), 2 deletions(-)
+
+diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
+index 4e67d243..8f49f084 100644
+--- a/ggml/src/ggml-backend-reg.cpp
++++ b/ggml/src/ggml-backend-reg.cpp
+@@ -573,8 +573,16 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
+ 
+     ggml_backend_load_best("blas", silent, dir_path);
+     ggml_backend_load_best("cann", silent, dir_path);
+-    ggml_backend_load_best("cuda", silent, dir_path);
+-    ggml_backend_load_best("hip", silent, dir_path);
++
++    // Avoid mixed hip+cuda configurations
++    const char * hip_devices = std::getenv("HIP_VISIBLE_DEVICES");
++    const char * rocr_devices = std::getenv("ROCR_VISIBLE_DEVICES"); 
++    if (!hip_devices && !rocr_devices) {
++        ggml_backend_load_best("cuda", silent, dir_path);
++    } else {
++        ggml_backend_load_best("hip", silent, dir_path);
++    }
++    
+     ggml_backend_load_best("kompute", silent, dir_path);
+     ggml_backend_load_best("metal", silent, dir_path);
+     ggml_backend_load_best("rpc", silent, dir_path);

llama/patches/0019-metal-add-mean-kernel-14267.patch

@@ -0,0 +1,169 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Georgi Gerganov <ggerganov@gmail.com>
+Date: Thu, 19 Jun 2025 08:05:21 +0300
+Subject: [PATCH] metal : add mean kernel (#14267)
+
+* metal : add mean kernel
+
+ggml-ci
+
+* cont : dedup implementation
+
+ggml-ci
+---
+ ggml/src/ggml-metal/ggml-metal.m     | 33 ++++++++++++++++---
+ ggml/src/ggml-metal/ggml-metal.metal | 48 ++++++++++++++++++++++------
+ 2 files changed, 67 insertions(+), 14 deletions(-)
+
+diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
+index ee4f2dcb..f20f5615 100644
+--- a/ggml/src/ggml-metal/ggml-metal.m
++++ b/ggml/src/ggml-metal/ggml-metal.m
+@@ -489,6 +489,7 @@ enum ggml_metal_kernel_type {
+     GGML_METAL_KERNEL_TYPE_COS,
+     GGML_METAL_KERNEL_TYPE_NEG,
+     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
++    GGML_METAL_KERNEL_TYPE_MEAN,
+     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
+     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
+     GGML_METAL_KERNEL_TYPE_ARGMAX,
+@@ -1436,6 +1437,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
++        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,                 pool_2d_max_f32,                 true);
+@@ -1634,6 +1636,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
+         case GGML_OP_LOG:
+             return false; // TODO: implement
+         case GGML_OP_SUM_ROWS:
++        case GGML_OP_MEAN:
+         case GGML_OP_SOFT_MAX:
+         case GGML_OP_GROUP_NORM:
+             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
+@@ -2362,11 +2365,30 @@ static bool ggml_metal_encode_node(
+                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+             } break;
+         case GGML_OP_SUM_ROWS:
++        case GGML_OP_MEAN:
+             {
+                 GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+ 
+-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
++                id<MTLComputePipelineState> pipeline = nil;
++
++                switch (dst->op) {
++                    case GGML_OP_SUM_ROWS:
++                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
++                        break;
++                    case GGML_OP_MEAN:
++                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MEAN].pipeline;
++                        break;
++                    default:
++                        GGML_ABORT("fatal error");
++                }
++
++                int nth = 32; // SIMD width
++
++                while (nth < ne00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
++                    nth *= 2;
++                }
+ 
++                nth = MIN(nth, ne00);
+ 
+                 ggml_metal_kargs_sum_rows args = {
+                    /*.ne00 =*/ ne00,
+@@ -2396,11 +2418,12 @@ static bool ggml_metal_encode_node(
+                 };
+ 
+                 [encoder setComputePipelineState:pipeline];
+-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+-                [encoder setBytes:&args length:sizeof(args) atIndex:2];
++                [encoder setBytes:&args length:sizeof(args) atIndex:0];
++                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
++                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
++                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
+ 
+-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
++                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+             } break;
+         case GGML_OP_SOFT_MAX:
+             {
+diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
+index 9cfddf45..08e8d807 100644
+--- a/ggml/src/ggml-metal/ggml-metal.metal
++++ b/ggml/src/ggml-metal/ggml-metal.metal
+@@ -956,31 +956,61 @@ kernel void kernel_neg(
+     dst[tpig] = -src0[tpig];
+ }
+ 
++template <bool norm>
+ kernel void kernel_sum_rows(
++        constant ggml_metal_kargs_sum_rows & args,
+         device const float * src0,
+         device       float * dst,
+-        constant ggml_metal_kargs_sum_rows & args,
+-        uint3 tpig[[thread_position_in_grid]]) {
+-    int64_t i3 = tpig.z;
+-    int64_t i2 = tpig.y;
+-    int64_t i1 = tpig.x;
++        threadgroup  float * shmem_f32 [[threadgroup(0)]],
++        uint3   tgpig[[threadgroup_position_in_grid]],
++        ushort3 tpitg[[thread_position_in_threadgroup]],
++        ushort  sgitg[[simdgroup_index_in_threadgroup]],
++        ushort  tiisg[[thread_index_in_simdgroup]],
++        ushort3   ntg[[threads_per_threadgroup]]) {
++    int64_t i3 = tgpig.z;
++    int64_t i2 = tgpig.y;
++    int64_t i1 = tgpig.x;
+ 
+     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+         return;
+     }
+ 
++    if (sgitg == 0) {
++        shmem_f32[tiisg] = 0.0f;
++    }
++
+     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+ 
+-    float row_sum = 0;
++    float sumf = 0;
+ 
+-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
+-        row_sum += src_row[i0];
++    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
++        sumf += src_row[i0];
+     }
+ 
+-    dst_row[0] = row_sum;
++    sumf = simd_sum(sumf);
++
++    threadgroup_barrier(mem_flags::mem_threadgroup);
++
++    if (tiisg == 0) {
++        shmem_f32[sgitg] = sumf;
++    }
++
++    threadgroup_barrier(mem_flags::mem_threadgroup);
++
++    sumf = shmem_f32[tiisg];
++    sumf = simd_sum(sumf);
++
++    if (tpitg.x == 0) {
++        dst_row[0] = norm ? sumf / args.ne00 : sumf;
++    }
+ }
+ 
++typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
++
++template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
++template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
++
+ template<typename T>
+ kernel void kernel_soft_max(
+         device const  char * src0,

llm/memory.go

@@ -151,7 +151,12 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 	}
 
 	if graphPartialOffload == 0 {
-		graphPartialOffload = f.KV().GQA() * kvTotal / 6
+		headsKV := f.KV().HeadCountKVMin()
+		if headsKV == 0 {
+			headsKV = 1
+		}
+		gqa := f.KV().HeadCountMax() / headsKV
+		graphPartialOffload = gqa * kvTotal / 6
 	}
 	if graphFullOffload == 0 {
 		graphFullOffload = graphPartialOffload

llm/server.go

@@ -139,6 +139,13 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		gpus = discover.GetCPUInfo()
 	}
 
+	// Verify the requested context size is <= the model training size
+	trainCtx := f.KV().ContextLength()
+	if opts.NumCtx/numParallel > int(trainCtx) && trainCtx > 0 {
+		slog.Warn("requested context size too large for model", "num_ctx", opts.NumCtx, "num_parallel", numParallel, "n_ctx_train", trainCtx)
+		opts.NumCtx = int(trainCtx) * numParallel
+	}
+
 	estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 	if len(gpus) > 1 || gpus[0].Library != "cpu" {
 		switch {
@@ -311,7 +318,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		params = append(params, "--mmproj", projectors[0])
 	}
 
-	// iterate through compatible GPU libraries such as 'cuda_v12', 'cuda_v11', 'rocm', etc.
+	// iterate through compatible GPU libraries such as 'cuda_v12', 'rocm', etc.
 	// adding each library's respective path to the LD_LIBRARY_PATH, until finally running
 	// without any LD_LIBRARY_PATH flags
 	for {
@@ -797,7 +804,8 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 
 	res, err := http.DefaultClient.Do(serverReq)
 	if err != nil {
-		return fmt.Errorf("POST predict: %v", err)
+		slog.Error("post predict", "error", err)
+		return errors.New("model runner has unexpectedly stopped, this may be due to resource limitations or an internal error, check ollama server logs for details")
 	}
 	defer res.Body.Close()
 

ml/backend.go

@@ -124,6 +124,10 @@ type DeviceMemory struct {
 	// may not be persistent across instances of the runner.
 	Name string
 
+	// UUID is a unique persistent identifier for the device for matching
+	// with system management libraries
+	UUID string
+
 	// Weights is the per-layer memory needed for the model weights.
 	Weights []Memory
 
@@ -152,6 +156,10 @@ func (m DeviceMemory) LogValue() slog.Value {
 		attrs = append(attrs, slog.Any("Graph", m.Graph))
 	}
 
+	if len(attrs) > 0 && m.UUID != "" {
+		attrs = append([]slog.Attr{slog.String("UUID", m.UUID)}, attrs...)
+	}
+
 	return slog.GroupValue(attrs...)
 }
 
@@ -245,6 +253,7 @@ type Tensor interface {
 
 	Neg(ctx Context) Tensor
 	Add(ctx Context, t2 Tensor) Tensor
+	Sub(ctx Context, t2 Tensor) Tensor
 	Mul(ctx Context, t2 Tensor) Tensor
 	Div(ctx Context, t2 Tensor) Tensor
 
@@ -268,6 +277,7 @@ type Tensor interface {
 	Tanh(ctx Context) Tensor
 	GELU(ctx Context) Tensor
 	SILU(ctx Context) Tensor
+	RELU(ctx Context) Tensor
 	Sigmoid(ctx Context) Tensor
 
 	Reshape(ctx Context, shape ...int) Tensor
@@ -289,6 +299,12 @@ type Tensor interface {
 
 	TopK(ctx Context, k int) Tensor
 	Argsort(ctx Context) Tensor
+	Mean(ctx Context) Tensor
+	Variance(ctx Context) Tensor
+	Stddev(ctx Context) Tensor
+	Sqr(ctx Context) Tensor
+	Sqrt(ctx Context) Tensor
+	Clamp(ctx Context, min, max float32) Tensor
 }
 
 // ScaledDotProductAttention implements a fused attention

ml/backend/ggml/ggml.go

@@ -136,6 +136,9 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	}
 
 	requiredMemory.CPU.Name = C.GoString(C.ggml_backend_dev_name(cpuDeviceBufferType.d))
+	var props C.struct_ggml_backend_dev_props
+	C.ggml_backend_dev_get_props(cpuDeviceBufferType.d, &props)
+	requiredMemory.CPU.UUID = C.GoString(props.uuid)
 	requiredMemory.CPU.Weights = make([]ml.Memory, blocks+1)
 	requiredMemory.CPU.Cache = make([]ml.Memory, blocks+1)
 
@@ -150,6 +153,9 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 		})
 		btDeviceMemory[bt] = &requiredMemory.GPUs[i]
 		requiredMemory.GPUs[i].Name = C.GoString(C.ggml_backend_dev_name(d))
+		var props C.struct_ggml_backend_dev_props
+		C.ggml_backend_dev_get_props(d, &props)
+		requiredMemory.GPUs[i].UUID = C.GoString(props.uuid)
 		requiredMemory.GPUs[i].Weights = make([]ml.Memory, blocks+1)
 		requiredMemory.GPUs[i].Cache = make([]ml.Memory, blocks+1)
 	}
@@ -291,7 +297,9 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
 				createTensor(tensor{source: t, target: "output.weight"}, output.bts, blocks)
 			}
-		case contains(t.Name, "cls", "output", "output_norm"):
+		case contains(t.Name, "cls", "output", "output_norm",
+			"altup_proj", "altup_unembd_proj",
+			"per_layer_token_embd", "per_layer_model_proj", "per_layer_proj_norm"):
 			createTensor(tensor{source: t}, output.bts, blocks)
 		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
 			// TODO: assign vision tensors to the gpu if possible
@@ -596,7 +604,9 @@ func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
 }
 
 func (c *Context) Compute(tensors ...ml.Tensor) {
-	C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph)
+	if status := C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph); status != C.GGML_STATUS_SUCCESS {
+		panic(fmt.Errorf("error computing ggml graph: %v", status))
+	}
 	C.ggml_backend_sched_reset(c.b.sched)
 
 	needSync := true
@@ -885,6 +895,13 @@ func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	}
 }
 
+func (t *Tensor) Sub(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sub(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) Repeat(ctx ml.Context, dim, n int) ml.Tensor {
 	if dim < 0 || dim >= C.GGML_MAX_DIMS {
 		panic("invalid dimension")
@@ -1192,6 +1209,13 @@ func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
 	}
 }
 
+func (t *Tensor) RELU(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_relu_inplace(ctx.(*Context).ctx, t.t),
+	}
+}
+
 func (t *Tensor) Conv2D(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
 	return &Tensor{
 		b: t.b,
@@ -1267,3 +1291,42 @@ func (t *Tensor) Argsort(ctx ml.Context) ml.Tensor {
 		t: C.ggml_argsort(ctx.(*Context).ctx, t.t, C.GGML_SORT_ORDER_ASC),
 	}
 }
+
+func (t *Tensor) Mean(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_mean(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Variance(ctx ml.Context) ml.Tensor {
+	return t.Add(ctx, t.Mean(ctx).Scale(ctx, -1)).
+		Sqr(ctx).
+		SumRows(ctx).
+		Scale(ctx, 1/float64(t.Dim(0)))
+}
+
+func (t *Tensor) Stddev(ctx ml.Context) ml.Tensor {
+	return t.Variance(ctx).Sqrt(ctx)
+}
+
+func (t *Tensor) Sqr(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sqr(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Sqrt(ctx ml.Context) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_sqrt(ctx.(*Context).ctx, t.t),
+	}
+}
+
+func (t *Tensor) Clamp(ctx ml.Context, min, max float32) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_clamp(ctx.(*Context).ctx, t.t, C.float(min), C.float(max)),
+	}
+}

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

@@ -152,6 +152,7 @@ extern "C" {
     struct ggml_backend_dev_props {
         const char * name;
         const char * description;
+        const char * uuid;
         size_t memory_free;
         size_t memory_total;
         enum ggml_backend_dev_type type;

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

@@ -573,8 +573,16 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
 
     ggml_backend_load_best("blas", silent, dir_path);
     ggml_backend_load_best("cann", silent, dir_path);
-    ggml_backend_load_best("cuda", silent, dir_path);
-    ggml_backend_load_best("hip", silent, dir_path);
+
+    // Avoid mixed hip+cuda configurations
+    const char * hip_devices = std::getenv("HIP_VISIBLE_DEVICES");
+    const char * rocr_devices = std::getenv("ROCR_VISIBLE_DEVICES"); 
+    if (!hip_devices && !rocr_devices) {
+        ggml_backend_load_best("cuda", silent, dir_path);
+    } else {
+        ggml_backend_load_best("hip", silent, dir_path);
+    }
+    
     ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);

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

@@ -362,6 +362,26 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template<bool norm>
+static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll

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

@@ -35,6 +35,7 @@
 #include "ggml-cuda/ssm-scan.cuh"
 #include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
+#include "ggml-cuda/mean.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
@@ -2322,6 +2323,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SUM_ROWS:
             ggml_cuda_op_sum_rows(ctx, dst);
             break;
+        case GGML_OP_MEAN:
+            ggml_cuda_op_mean(ctx, dst);
+            break;
         case GGML_OP_SSM_CONV:
             ggml_cuda_op_ssm_conv(ctx, dst);
             break;
@@ -2884,6 +2888,7 @@ struct ggml_backend_cuda_device_context {
     int device;
     std::string name;
     std::string description;
+    std::string uuid;
 };
 
 static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -2896,6 +2901,11 @@ static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t
     return ctx->description.c_str();
 }
 
+static const char * ggml_backend_cuda_device_get_uuid(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ctx->uuid.c_str();
+}
+
 static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
     ggml_cuda_set_device(ctx->device);
@@ -2910,6 +2920,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
 static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
     props->name        = ggml_backend_cuda_device_get_name(dev);
     props->description = ggml_backend_cuda_device_get_description(dev);
+    props->uuid        = ggml_backend_cuda_device_get_uuid(dev);
     props->type        = ggml_backend_cuda_device_get_type(dev);
     ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
 
@@ -3204,6 +3215,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;
@@ -3458,6 +3470,32 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;
 
+                #if !defined(GGML_USE_HIP)
+                char uuid[64];
+                snprintf(uuid, sizeof(uuid),
+                    "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
+                    (unsigned char)prop.uuid.bytes[0],
+                    (unsigned char)prop.uuid.bytes[1],
+                    (unsigned char)prop.uuid.bytes[2],
+                    (unsigned char)prop.uuid.bytes[3],
+                    (unsigned char)prop.uuid.bytes[4],
+                    (unsigned char)prop.uuid.bytes[5],
+                    (unsigned char)prop.uuid.bytes[6],
+                    (unsigned char)prop.uuid.bytes[7],
+                    (unsigned char)prop.uuid.bytes[8],
+                    (unsigned char)prop.uuid.bytes[9],
+                    (unsigned char)prop.uuid.bytes[10],
+                    (unsigned char)prop.uuid.bytes[11],
+                    (unsigned char)prop.uuid.bytes[12],
+                    (unsigned char)prop.uuid.bytes[13],
+                    (unsigned char)prop.uuid.bytes[14],
+                    (unsigned char)prop.uuid.bytes[15]
+                  );
+                dev_ctx->uuid = uuid;
+                #else
+                dev_ctx->uuid = "GPU-" + std::string(prop.uuid.bytes, 16);
+                #endif
+
                 ggml_backend_dev_t dev = new ggml_backend_device {
                     /* .iface   = */ ggml_backend_cuda_device_interface,
                     /* .reg     = */ &reg,

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

@@ -0,0 +1,19 @@
+#include "mean.cuh"
+
+void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0   = dst->src[0];
+    const float *       src0_d = (const float *) src0->data;
+    float *             dst_d  = (float *) dst->data;
+    cudaStream_t        stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+}

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

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -1,25 +1,9 @@
 #include "sumrows.cuh"
 
-static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col == 0) {
-        dst[row] = sum;
-    }
-}
-
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
-    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
 }
 
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -35,5 +19,8 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+
+    reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
 }

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

@@ -1,5 +1,4 @@
 #include "common.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream);
-
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -3434,31 +3434,61 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+template <bool norm>
 kernel void kernel_sum_rows(
+        constant ggml_metal_kargs_sum_rows & args,
         device const float * src0,
         device       float * dst,
-        constant ggml_metal_kargs_sum_rows & args,
-        uint3 tpig[[thread_position_in_grid]]) {
-    int64_t i3 = tpig.z;
-    int64_t i2 = tpig.y;
-    int64_t i1 = tpig.x;
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    int64_t i3 = tgpig.z;
+    int64_t i2 = tgpig.y;
+    int64_t i1 = tgpig.x;
 
     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
         return;
     }
 
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
 
-    float row_sum = 0;
+    float sumf = 0;
 
-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
-        row_sum += src_row[i0];
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
     }
 
-    dst_row[0] = row_sum;
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        dst_row[0] = norm ? sumf / args.ne00 : sumf;
+    }
 }
 
+typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
+template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
+
 template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,

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

@@ -489,6 +489,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COS,
     GGML_METAL_KERNEL_TYPE_NEG,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
+    GGML_METAL_KERNEL_TYPE_MEAN,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
     GGML_METAL_KERNEL_TYPE_ARGMAX,
@@ -1436,6 +1437,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,                 pool_2d_max_f32,                 true);
@@ -1634,6 +1636,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_LOG:
             return false; // TODO: implement
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
@@ -2362,11 +2365,30 @@ static bool ggml_metal_encode_node(
                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
             } break;
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
             {
                 GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                id<MTLComputePipelineState> pipeline = nil;
 
+                switch (dst->op) {
+                    case GGML_OP_SUM_ROWS:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                        break;
+                    case GGML_OP_MEAN:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MEAN].pipeline;
+                        break;
+                    default:
+                        GGML_ABORT("fatal error");
+                }
+
+                int nth = 32; // SIMD width
+
+                while (nth < ne00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                nth = MIN(nth, ne00);
 
                 ggml_metal_kargs_sum_rows args = {
                    /*.ne00 =*/ ne00,
@@ -2396,11 +2418,12 @@ static bool ggml_metal_encode_node(
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                [encoder setBytes:&args length:sizeof(args) atIndex:2];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_SOFT_MAX:
             {
@@ -5703,6 +5726,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
 static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
     props->name        = ggml_backend_metal_device_get_name(dev);
     props->description = ggml_backend_metal_device_get_description(dev);
+    props->uuid        = "0";
     props->type        = ggml_backend_metal_device_get_type(dev);
     ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
     props->caps = (struct ggml_backend_dev_caps) {

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

@@ -956,31 +956,61 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+template <bool norm>
 kernel void kernel_sum_rows(
+        constant ggml_metal_kargs_sum_rows & args,
         device const float * src0,
         device       float * dst,
-        constant ggml_metal_kargs_sum_rows & args,
-        uint3 tpig[[thread_position_in_grid]]) {
-    int64_t i3 = tpig.z;
-    int64_t i2 = tpig.y;
-    int64_t i1 = tpig.x;
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    int64_t i3 = tgpig.z;
+    int64_t i2 = tgpig.y;
+    int64_t i1 = tgpig.x;
 
     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
         return;
     }
 
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
 
-    float row_sum = 0;
+    float sumf = 0;
 
-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
-        row_sum += src_row[i0];
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
     }
 
-    dst_row[0] = row_sum;
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        dst_row[0] = norm ? sumf / args.ne00 : sumf;
+    }
 }
 
+typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
+template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
+
 template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,

model/bytepairencoding.go

@@ -3,6 +3,7 @@ package model
 import (
 	"cmp"
 	"context"
+	"fmt"
 	"iter"
 	"log/slog"
 	"strings"
@@ -210,6 +211,14 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
 	return ids, nil
 }
 
+type lazyIdsString struct {
+	ids []int32
+}
+
+func (l lazyIdsString) LogValue() slog.Value {
+	return slog.AnyValue(fmt.Sprint(l.ids))
+}
+
 func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 	var sb strings.Builder
 	for _, id := range ids {
@@ -234,6 +243,6 @@ func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "ids", ids, "string", sb.String())
+	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
 	return sb.String(), nil
 }

model/models/gemma3n/model.go

@@ -0,0 +1,52 @@
+package gemma3n
+
+import (
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Model struct {
+	model.Base
+	model.SentencePieceModel
+
+	*TextModel
+}
+
+// Forward implements model.Model.
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	return m.TextModel.Forward(ctx, batch, m.Cache)
+}
+
+func New(c fs.Config) (model.Model, error) {
+	m := Model{
+		TextModel: newTextModel(c),
+		SentencePieceModel: model.NewSentencePieceModel(
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Scores: c.Floats("tokenizer.ggml.scores"),
+				Types:  c.Ints("tokenizer.ggml.token_type"),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOS: append(
+					[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+					c.Ints("tokenizer.ggml.eos_token_ids")...,
+				),
+			},
+		),
+	}
+
+	// TODO: setup hybrid (local sliding window + global) cache
+	m.Cache = kvcache.NewWrapperCache(
+		kvcache.NewCausalCache(m.Shift),
+		kvcache.NewSWACache(int32(c.Uint("attention.sliding_window")), m.Shift),
+	)
+	return &m, nil
+}
+
+func init() {
+	model.Register("gemma3n", New)
+}

model/models/gemma3n/model_text.go

@@ -0,0 +1,360 @@
+package gemma3n
+
+import (
+	"cmp"
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
+	"github.com/ollama/ollama/model/input"
+)
+
+type TextModel struct {
+	TokenEmbedding *TextScaledWordEmbedding `gguf:"token_embd"`
+
+	*PerLayerProjector
+
+	AltupEmbd   *nn.Linear `gguf:"altup_proj"`
+	AltupUnembd *nn.Linear `gguf:"altup_unembd_proj"`
+
+	TextLayers []TextLayer `gguf:"blk"`
+	OutputNorm *nn.RMSNorm `gguf:"output_norm"`
+	Output     *nn.Linear  `gguf:"output,alt:token_embd"`
+
+	TextOptions
+}
+
+func (m *TextModel) Forward(ctx ml.Context, batch input.Batch, cache kvcache.Cache) (ml.Tensor, error) {
+	positions := ctx.Input().FromIntSlice(batch.Positions, len(batch.Positions))
+	// Create a tensor of a single float32 value of 1.0 to use for altup correction
+	one := ctx.Input().FromFloatSlice([]float32{1.0}, 1)
+
+	inputs := m.TokenEmbedding.Forward(ctx, batch.Inputs, math.Sqrt(float64(m.hiddenSize)))
+	inputsPerLayer := m.PerLayerProjector.Forward(ctx, batch, inputs, &m.TextOptions)
+
+	targetMagnitude := inputs.Sqr(ctx).Mean(ctx).Sqrt(ctx)
+	targetMagnitude = targetMagnitude.Repeat(ctx, 2, m.altupInputs-1)
+
+	hiddenState := inputs.Repeat(ctx, 2, m.altupInputs-1)
+	altupProj := m.AltupEmbd.Forward(ctx, hiddenState)
+	altupProj = altupProj.Mul(ctx, targetMagnitude.Div(ctx, altupProj.Sqr(ctx).Mean(ctx).Sqrt(ctx)))
+
+	hiddenStates := inputs.Concat(ctx, altupProj, 2)
+
+	firstSharedKeyValue := m.hiddenLayers - m.sharedKeyValueLayers
+	for i, layer := range m.TextLayers {
+		if i < firstSharedKeyValue {
+			cache.SetLayer(i)
+		} else if m.isLocal(i) {
+			cache.SetLayer(firstSharedKeyValue - 2)
+		} else {
+			cache.SetLayer(firstSharedKeyValue - 1)
+		}
+
+		var layerType int
+		ropeBase := m.ropeBase
+		if m.isLocal(i) {
+			layerType = 1
+			ropeBase = m.ropeBaseLocal
+		}
+
+		cache.(*kvcache.WrapperCache).SetLayerType(layerType)
+
+		// inputPerLayer = inputsPerLayer[:, i, :]
+		inputPerLayer := inputsPerLayer.View(ctx, i*inputsPerLayer.Stride(1), inputsPerLayer.Dim(0), inputsPerLayer.Stride(2), inputsPerLayer.Dim(2))
+		hiddenStates = layer.Forward(ctx, hiddenStates, inputPerLayer, positions, one, cache, i >= firstSharedKeyValue, ropeBase, float64(m.activationSparsityScale[i]), &m.TextOptions)
+	}
+
+	// hiddenStates = hiddenStates[:, :, 0]
+	hiddenStates0 := hiddenStates.View(ctx, 0, hiddenStates.Dim(0), hiddenStates.Stride(1), hiddenStates.Dim(1))
+	targetMagnitude = hiddenStates0.Sqr(ctx).Mean(ctx).Sqrt(ctx)
+	targetMagnitude = targetMagnitude.Repeat(ctx, 2, m.altupInputs-1)
+
+	// hiddenState = hiddenStates[:, :, 1:]
+	hiddenState = hiddenStates.View(ctx, hiddenStates.Stride(2), hiddenStates.Dim(0), hiddenStates.Stride(1), hiddenStates.Dim(1), hiddenStates.Stride(2), m.altupInputs-1)
+	altupUnembdProj := m.AltupUnembd.Forward(ctx, hiddenState)
+	altupUnembdProj = altupUnembdProj.Mul(ctx, targetMagnitude.Div(ctx, altupUnembdProj.Sqr(ctx).Mean(ctx).Sqrt(ctx)))
+
+	hiddenStates = hiddenStates0.Concat(ctx, altupUnembdProj, 2)
+
+	hiddenStates = hiddenStates.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx).Mean(ctx)
+	hiddenStates = hiddenStates.Permute(ctx, 2, 0, 1, 3).Contiguous(ctx)
+	hiddenStates = hiddenStates.Rows(ctx, ctx.Input().FromIntSlice(batch.Outputs, len(batch.Outputs)))
+
+	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
+	return m.Output.Forward(ctx, hiddenStates), nil
+}
+
+func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	ropeBase := m.ropeBase
+	if m.isLocal(layer) {
+		ropeBase = m.ropeBaseLocal
+	}
+
+	return fast.RoPE(ctx, key, shift, m.headDim(), ropeBase, m.ropeScale, rope.WithTypeNeoX()), nil
+}
+
+type TextScaledWordEmbedding struct {
+	*nn.Embedding
+}
+
+func (e TextScaledWordEmbedding) Forward(ctx ml.Context, inputIDs ml.Tensor, scale float64) ml.Tensor {
+	return e.Embedding.Forward(ctx, inputIDs).Scale(ctx, scale)
+}
+
+type PerLayerProjector struct {
+	TokenEmbedding *TextScaledWordEmbedding `gguf:"per_layer_token_embd"`
+	Projector      *nn.Linear               `gguf:"per_layer_model_proj"`
+	Norm           *nn.RMSNorm              `gguf:"per_layer_proj_norm"`
+}
+
+func (p PerLayerProjector) Forward(ctx ml.Context, batch input.Batch, inputs ml.Tensor, opts *TextOptions) ml.Tensor {
+	inputsPerLayer := p.TokenEmbedding.Forward(ctx, batch.Inputs, math.Sqrt(float64(opts.hiddenSizePerLayerInput)))
+	inputsPerLayer = inputsPerLayer.Reshape(ctx, opts.hiddenSizePerLayerInput, opts.hiddenLayers, batch.Inputs.Dim(0), batch.Inputs.Dim(1))
+
+	perLayerProjection := p.Projector.Forward(ctx, inputs)
+	perLayerProjection = perLayerProjection.Scale(ctx, math.Sqrt(float64(opts.hiddenSize)))
+	perLayerProjection = perLayerProjection.Reshape(ctx, opts.hiddenSizePerLayerInput, opts.hiddenLayers, inputs.Dim(1))
+	perLayerProjection = p.Norm.Forward(ctx, perLayerProjection, opts.eps)
+
+	if inputsPerLayer != nil {
+		perLayerProjection = perLayerProjection.Add(ctx, inputsPerLayer)
+		perLayerProjection = perLayerProjection.Scale(ctx, 1/math.Sqrt(2))
+	}
+
+	return perLayerProjection
+}
+
+type TextLayer struct {
+	*AltUp
+	*Laurel
+
+	AttentionNorm     *nn.RMSNorm `gguf:"attn_norm"`
+	Attention         *TextAttention
+	PostAttentionNorm *nn.RMSNorm `gguf:"post_attention_norm"`
+
+	MLPNorm     *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP         *TextMLP
+	PostMLPNorm *nn.RMSNorm `gguf:"post_ffw_norm"`
+
+	PerLayerInputGate  *nn.Linear  `gguf:"inp_gate"`
+	PerLayerProjection *nn.Linear  `gguf:"proj"`
+	PostPerLayerNorm   *nn.RMSNorm `gguf:"post_norm"`
+}
+
+func (d TextLayer) Forward(ctx ml.Context, hiddenStates, perLayerInput, positions, one ml.Tensor, cache kvcache.Cache, sharedKV bool, ropeBase float32, activationSparsityScale float64, opts *TextOptions) ml.Tensor {
+	predictions := d.Predict(ctx, hiddenStates, opts)
+	active := opts.altupActive(ctx, predictions)
+
+	attn := d.AttentionNorm.Forward(ctx, active, opts.eps)
+	laurel := d.Laurel.Forward(ctx, attn, opts)
+
+	attn = d.Attention.Forward(ctx, attn, positions, cache, sharedKV, ropeBase, opts)
+	attn = d.PostAttentionNorm.Forward(ctx, attn, opts.eps)
+	attn = active.Add(ctx, attn)
+	attn = attn.Add(ctx, laurel).Scale(ctx, 1/math.Sqrt(2))
+
+	mlp := d.MLPNorm.Forward(ctx, attn, opts.eps)
+	mlp = d.MLP.Forward(ctx, mlp, activationSparsityScale)
+	mlp = d.PostMLPNorm.Forward(ctx, mlp, opts.eps)
+	mlp = attn.Add(ctx, mlp)
+
+	predictions = d.Correct(ctx, predictions, mlp, one, opts)
+	active = opts.altupActive(ctx, predictions)
+	if opts.altupCorrectScale {
+		active = d.ScaleCorrectedOutput(ctx, active)
+	}
+
+	active = d.PerLayerInputGate.Forward(ctx, active)
+	active = active.GELU(ctx)
+	active = active.Mul(ctx, perLayerInput)
+
+	active = d.PerLayerProjection.Forward(ctx, active)
+	active = d.PostPerLayerNorm.Forward(ctx, active, opts.eps)
+
+	// inactive := predictions[:, :, 1:]
+	inactive := predictions.View(ctx, predictions.Stride(2), predictions.Dim(0), predictions.Stride(1), predictions.Dim(1), predictions.Stride(2), predictions.Dim(2)-1)
+	active = inactive.Add(ctx, active)
+
+	predictions0 := predictions.View(ctx, 0, predictions.Dim(0), predictions.Stride(1), predictions.Dim(1))
+	return predictions0.Concat(ctx, active, 2)
+}
+
+type AltUp struct {
+	CorrectionScale       ml.Tensor   `gguf:"altup_correct_scale.weight"`
+	PredictionCoefficient *nn.Linear  `gguf:"altup_predict_coef"`
+	CorrectionCoefficient *nn.Linear  `gguf:"altup_correct_coef"`
+	Router                *nn.Linear  `gguf:"altup_router"`
+	RouterNorm            *nn.RMSNorm `gguf:"altup_router_norm"`
+}
+
+func (a AltUp) computeRouterModalities(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
+	routerInputs := a.RouterNorm.Forward(ctx, hiddenStates, opts.eps).Scale(ctx, 1.0/float64(opts.hiddenSize))
+	return a.Router.Forward(ctx, routerInputs).Tanh(ctx)
+}
+
+func (a AltUp) Predict(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
+	modalities := a.computeRouterModalities(ctx, opts.altupActive(ctx, hiddenStates), opts)
+
+	coefficients := a.PredictionCoefficient.Forward(ctx, modalities)
+	coefficients = coefficients.Reshape(ctx, opts.altupInputs, opts.altupInputs, coefficients.Dim(1), coefficients.Dim(2))
+
+	hiddenStates = hiddenStates.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)
+	predictions := coefficients.Mulmat(ctx, hiddenStates)
+	predictions = predictions.Add(ctx, hiddenStates)
+	return predictions.Permute(ctx, 2, 0, 1, 3).Contiguous(ctx)
+}
+
+func (a AltUp) Correct(ctx ml.Context, predictions, activated, one ml.Tensor, opts *TextOptions) ml.Tensor {
+	innovation := activated.Sub(ctx, opts.altupActive(ctx, predictions))
+	innovation = innovation.Repeat(ctx, 2, opts.altupInputs)
+
+	modalities := a.computeRouterModalities(ctx, activated, opts)
+	coefficients := a.CorrectionCoefficient.Forward(ctx, modalities)
+	coefficients = coefficients.Add(ctx, one)
+
+	coefficients = coefficients.Reshape(ctx, 1, coefficients.Dim(0), coefficients.Dim(1))
+	coefficients = coefficients.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
+
+	corrected := innovation.Mul(ctx, coefficients)
+	corrected = corrected.Add(ctx, predictions)
+	return corrected
+}
+
+func (a AltUp) ScaleCorrectedOutput(ctx ml.Context, predictions ml.Tensor) ml.Tensor {
+	return predictions.Mul(ctx, a.CorrectionScale)
+}
+
+type Laurel struct {
+	LinearLeft     *nn.Linear  `gguf:"laurel_l"`
+	LinearRight    *nn.Linear  `gguf:"laurel_r"`
+	PostLaurelNorm *nn.RMSNorm `gguf:"laurel_post_norm"`
+}
+
+func (l Laurel) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOptions) ml.Tensor {
+	residual := hiddenStates
+	hiddenStates = l.LinearLeft.Forward(ctx, hiddenStates)
+	hiddenStates = l.LinearRight.Forward(ctx, hiddenStates)
+	hiddenStates = l.PostLaurelNorm.Forward(ctx, hiddenStates, opts.eps)
+	return hiddenStates.Add(ctx, residual)
+}
+
+type TextAttention struct {
+	Query     *nn.Linear  `gguf:"attn_q"`
+	QueryNorm *nn.RMSNorm `gguf:"attn_q_norm"`
+	Key       *nn.Linear  `gguf:"attn_k"`
+	KeyNorm   *nn.RMSNorm `gguf:"attn_k_norm"`
+	Value     *nn.Linear  `gguf:"attn_v"`
+	Output    *nn.Linear  `gguf:"attn_output"`
+}
+
+func (attn TextAttention) Forward(ctx ml.Context, hiddenStates, positions ml.Tensor, cache kvcache.Cache, sharedKV bool, ropeBase float32, opts *TextOptions) ml.Tensor {
+	batchSize := hiddenStates.Dim(1)
+
+	query := attn.Query.Forward(ctx, hiddenStates)
+	query = query.Reshape(ctx, opts.headDim(), opts.numHeads, batchSize)
+	query = attn.QueryNorm.Forward(ctx, query, opts.eps)
+	query = fast.RoPE(ctx, query, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+
+	var key, value ml.Tensor
+	if !sharedKV {
+		key = attn.Key.Forward(ctx, hiddenStates)
+		key = key.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
+		key = attn.KeyNorm.Forward(ctx, key, opts.eps)
+		key = fast.RoPE(ctx, key, positions, opts.headDim(), ropeBase, opts.ropeScale, rope.WithTypeNeoX())
+
+		value = attn.Value.Forward(ctx, hiddenStates)
+		value = value.Reshape(ctx, opts.headDim(), opts.numKVHeads, batchSize)
+		value = value.RMSNorm(ctx, nil, opts.eps)
+	}
+
+	attention := nn.Attention(ctx, query, key, value, 1., cache)
+	attention = attention.Reshape(ctx, attention.Dim(0)*attention.Dim(1), batchSize)
+	return attn.Output.Forward(ctx, attention)
+}
+
+type TextMLP struct {
+	Gate *nn.Linear `gguf:"ffn_gate"`
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+}
+
+func (mlp TextMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, activationSparsityScale float64) ml.Tensor {
+	upStates := mlp.Up.Forward(ctx, hiddenStates)
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates)
+	if activationSparsityScale > 0 {
+		mean := hiddenStates.Mean(ctx)
+		std := hiddenStates.Stddev(ctx).Scale(ctx, activationSparsityScale)
+		cutoff := mean.Add(ctx, std)
+		hiddenStates = hiddenStates.Sub(ctx, cutoff).RELU(ctx)
+	}
+
+	hiddenStates = hiddenStates.GELU(ctx).Mul(ctx, upStates)
+	hiddenStates = mlp.Down.Forward(ctx, hiddenStates)
+	return hiddenStates
+}
+
+type TextOptions struct {
+	hiddenLayers            int
+	hiddenSize              int
+	hiddenSizePerLayerInput int
+	numHeads, numKVHeads    int
+	keyLength, valueLength  int
+	sharedKeyValueLayers    int
+
+	altupActiveIndex  int
+	altupInputs       int
+	altupCorrectScale bool
+
+	eps           float32
+	ropeBase      float32
+	ropeBaseLocal float32
+	ropeScale     float32
+
+	slidingWindowPattern    []bool
+	activationSparsityScale []float32
+}
+
+func (o *TextOptions) altupActive(ctx ml.Context, t ml.Tensor) ml.Tensor {
+	// t[:, :, o.altupActiveIndex]
+	return t.View(ctx, o.altupActiveIndex*t.Stride(2), t.Dim(0), t.Stride(1), t.Dim(1))
+}
+
+func (o *TextOptions) headDim() int {
+	return cmp.Or(o.keyLength, o.valueLength, o.hiddenSize/o.numHeads)
+}
+
+func (o *TextOptions) isLocal(i int) bool {
+	return o.slidingWindowPattern[i]
+}
+
+func newTextModel(c fs.Config) *TextModel {
+	return &TextModel{
+		TextLayers: make([]TextLayer, c.Uint("block_count")),
+		TextOptions: TextOptions{
+			hiddenLayers:            int(c.Uint("block_count")),
+			hiddenSize:              int(c.Uint("embedding_length")),
+			hiddenSizePerLayerInput: int(c.Uint("embedding_length_per_layer_input")),
+			numHeads:                int(c.Uint("attention.head_count")),
+			numKVHeads:              int(c.Uint("attention.head_count_kv")),
+			keyLength:               int(c.Uint("attention.key_length")),
+			valueLength:             int(c.Uint("attention.value_length")),
+			sharedKeyValueLayers:    int(c.Uint("attention.shared_kv_layers")),
+
+			altupActiveIndex: int(c.Uint("altup.active_idx")),
+			altupInputs:      int(c.Uint("altup.num_inputs")),
+
+			eps:           c.Float("attention.layer_norm_rms_epsilon", 1e-06),
+			ropeBase:      c.Float("rope.freq_base", 1_000_000),
+			ropeBaseLocal: c.Float("rope.freq_base_local", 10_000),
+			ropeScale:     c.Float("rope.freq_scale", 1.0),
+
+			slidingWindowPattern:    c.Bools("attention.sliding_window_pattern"),
+			activationSparsityScale: c.Floats("activation_sparsity_scale"),
+		},
+	}
+}

model/models/llama4/model_text.go

@@ -63,9 +63,9 @@ func (mlp *TextMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *TextOp
 }
 
 type TextExperts struct {
-	Gate ml.Tensor `gguf:"ffn_gate_exps.weight"`
-	Up   ml.Tensor `gguf:"ffn_up_exps.weight"`
-	Down ml.Tensor `gguf:"ffn_down_exps.weight"`
+	Gate *nn.Linear `gguf:"ffn_gate_exps"`
+	Up   *nn.Linear `gguf:"ffn_up_exps"`
+	Down *nn.Linear `gguf:"ffn_down_exps"`
 }
 
 func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tensor, opts *TextOptions) ml.Tensor {
@@ -76,9 +76,9 @@ func (e *TextExperts) Forward(ctx ml.Context, hiddenStates, routerLogits ml.Tens
 	hiddenStates = hiddenStates.Repeat(ctx, 1, opts.numExpertsUsed)
 	hiddenStates = hiddenStates.Mul(ctx, scores)
 
-	upStates := e.Up.MulmatID(ctx, hiddenStates, experts)
-	gateStates := e.Gate.MulmatID(ctx, hiddenStates, experts)
-	downStates := e.Down.MulmatID(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
+	upStates := e.Up.Weight.MulmatID(ctx, hiddenStates, experts)
+	gateStates := e.Gate.Weight.MulmatID(ctx, hiddenStates, experts)
+	downStates := e.Down.Weight.MulmatID(ctx, upStates.Mul(ctx, gateStates.SILU(ctx)), experts)
 
 	nextStates := downStates.View(ctx, 0, hiddenStates.Dim(0), downStates.Stride(2), hiddenStates.Dim(2))
 	for i := 1; i < opts.numExpertsUsed; i++ {

model/models/models.go

@@ -3,6 +3,7 @@ package models
 import (
 	_ "github.com/ollama/ollama/model/models/gemma2"
 	_ "github.com/ollama/ollama/model/models/gemma3"
+	_ "github.com/ollama/ollama/model/models/gemma3n"
 	_ "github.com/ollama/ollama/model/models/llama"
 	_ "github.com/ollama/ollama/model/models/llama4"
 	_ "github.com/ollama/ollama/model/models/mistral3"

model/models/qwen3/model.go

@@ -66,9 +66,9 @@ type MLP interface {
 
 type sparse struct {
 	Router *nn.Linear `gguf:"ffn_gate_inp"`
-	Gate   ml.Tensor  `gguf:"ffn_gate_exps.weight"`
-	Up     ml.Tensor  `gguf:"ffn_up_exps.weight"`
-	Down   ml.Tensor  `gguf:"ffn_down_exps.weight"`
+	Gate   *nn.Linear `gguf:"ffn_gate_exps"`
+	Up     *nn.Linear `gguf:"ffn_up_exps"`
+	Down   *nn.Linear `gguf:"ffn_down_exps"`
 }
 
 func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options) ml.Tensor {
@@ -87,13 +87,13 @@ func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options
 
 	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0), 1, hiddenStates.Dim(1))
 
-	upStates := mlp.Up.MulmatID(ctx, hiddenStates, selectedExperts)
+	upStates := mlp.Up.Weight.MulmatID(ctx, hiddenStates, selectedExperts)
 
-	hiddenStates = mlp.Gate.MulmatID(ctx, hiddenStates, selectedExperts)
+	hiddenStates = mlp.Gate.Weight.MulmatID(ctx, hiddenStates, selectedExperts)
 	hiddenStates = hiddenStates.SILU(ctx)
 	hiddenStates = hiddenStates.Mul(ctx, upStates)
 
-	experts := mlp.Down.MulmatID(ctx, hiddenStates, selectedExperts)
+	experts := mlp.Down.Weight.MulmatID(ctx, hiddenStates, selectedExperts)
 	experts = experts.Mul(ctx, routingWeights)
 
 	nextStates := experts.View(ctx, 0, experts.Dim(0), experts.Stride(2), experts.Dim(2))

model/vocabulary.go

@@ -87,7 +87,7 @@ func (v *Vocabulary) Decode(id int32) string {
 func (v *Vocabulary) SpecialVocabulary() []string {
 	v.specialOnce.Do(func() {
 		for i := range v.Values {
-			if v.Types[i] == TOKEN_TYPE_CONTROL {
+			if v.Types[i] == TOKEN_TYPE_CONTROL || v.Types[i] == TOKEN_TYPE_USER_DEFINED {
 				v.special = append(v.special, v.Values[i])
 			}
 		}

model/vocabulary_test.go

@@ -0,0 +1,16 @@
+package model
+
+import "testing"
+
+func TestVocabulary_SpecialVocabulary(t *testing.T) {
+	vocab := &Vocabulary{
+		Values: []string{"<|startoftext|>", "<|endoftext|>", "<|tool_call_start|>", "<|tool_call_end|>", "hi"},
+		Types:  []int32{TOKEN_TYPE_CONTROL, TOKEN_TYPE_CONTROL, TOKEN_TYPE_USER_DEFINED, TOKEN_TYPE_USER_DEFINED, TOKEN_TYPE_NORMAL},
+	}
+
+	specialVocab := vocab.SpecialVocabulary()
+
+	if len(specialVocab) != 4 {
+		t.Errorf("expected 4 special tokens, got %d", len(specialVocab))
+	}
+}

parser/parser.go

@@ -292,13 +292,18 @@ func filesForModel(path string) ([]string, error) {
 	}
 	files = append(files, js...)
 
-	if tks, _ := glob(filepath.Join(path, "tokenizer.model"), "application/octet-stream"); len(tks) > 0 {
-		// add tokenizer.model if it exists, tokenizer.json is automatically picked up by the previous glob
-		// tokenizer.model might be a unresolved git lfs reference; error if it is
-		files = append(files, tks...)
-	} else if tks, _ := glob(filepath.Join(path, "**/tokenizer.model"), "text/plain"); len(tks) > 0 {
-		// some times tokenizer.model is in a subdirectory (e.g. meta-llama/Meta-Llama-3-8B)
-		files = append(files, tks...)
+	// only include tokenizer.model is tokenizer.json is not present
+	if !slices.ContainsFunc(files, func(s string) bool {
+		return slices.Contains(strings.Split(s, string(os.PathSeparator)), "tokenizer.json")
+	}) {
+		if tks, _ := glob(filepath.Join(path, "tokenizer.model"), "application/octet-stream"); len(tks) > 0 {
+			// add tokenizer.model if it exists, tokenizer.json is automatically picked up by the previous glob
+			// tokenizer.model might be a unresolved git lfs reference; error if it is
+			files = append(files, tks...)
+		} else if tks, _ := glob(filepath.Join(path, "**/tokenizer.model"), "text/plain"); len(tks) > 0 {
+			// some times tokenizer.model is in a subdirectory (e.g. meta-llama/Meta-Llama-3-8B)
+			files = append(files, tks...)
+		}
 	}
 
 	return files, nil

readline/types.go

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

scripts/build_windows.ps1

@@ -27,7 +27,6 @@ function checkEnv() {
         $env:VCToolsRedistDir=(get-item "${MSVC_INSTALL}\VC\Redist\MSVC\*")[0]
     }
     # Locate CUDA versions
-    # Note: this assumes every version found will be built
     $cudaList=(get-item "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*\bin\" -ea 'silentlycontinue')
     if ($cudaList.length -eq 0) {
         $d=(get-command -ea 'silentlycontinue' nvcc).path
@@ -94,19 +93,6 @@ function buildOllama() {
 
         $hashEnv = @{}
         Get-ChildItem env: | foreach { $hashEnv[$_.Name] = $_.Value }
-        if ("$script:CUDA_DIRS".Contains("v11")) {
-            $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V11")) { $v11="$_" }}
-            $env:CUDAToolkit_ROOT=$hashEnv[$v11]
-            write-host "Building CUDA v11 backend libraries"
-            # Note: cuda v11 requires msvc 2019 so force the older generator
-            # to avoid 2022 (or newer) from being used as the default
-            & cmake --fresh --preset "CUDA 11" -G "Visual Studio 16 2019" --install-prefix $script:DIST_DIR
-            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-            & cmake --build --preset "CUDA 11"  --config Release --parallel $script:JOBS
-            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-            & cmake --install build --component "CUDA" --strip
-            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-        }
         if ("$script:CUDA_DIRS".Contains("v12")) {
             $hashEnv.Keys | foreach { if ($_.Contains("CUDA_PATH_V12")) { $v12="$_" }}
             $env:CUDAToolkit_ROOT=$hashEnv[$v12]
@@ -127,12 +113,17 @@ function buildOllama() {
             $env:HIPCXX="${env:HIP_PATH}\bin\clang++.exe"
             $env:HIP_PLATFORM="amd"
             $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-            & cmake --fresh --preset "ROCm 6" -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ --install-prefix $script:DIST_DIR
+            & cmake --fresh --preset "ROCm 6" -G Ninja `
+                -DCMAKE_C_COMPILER=clang `
+                -DCMAKE_CXX_COMPILER=clang++ `
+                -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" `
+                -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" `
+                --install-prefix $script:DIST_DIR
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             $env:HIPCXX=""
             $env:HIP_PLATFORM=""
             $env:CMAKE_PREFIX_PATH=""
-            & cmake --build --preset "ROCm"  --config Release --parallel $script:JOBS
+            & cmake --build --preset "ROCm 6" --config Release --parallel $script:JOBS
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             & cmake --install build --component "HIP" --strip
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}

scripts/env.sh

@@ -10,9 +10,7 @@ OLLAMA_COMMON_BUILD_ARGS="--build-arg=VERSION \
     --build-arg=GOFLAGS \
     --build-arg=OLLAMA_CUSTOM_CPU_DEFS \
     --build-arg=OLLAMA_SKIP_CUDA_GENERATE \
-    --build-arg=OLLAMA_SKIP_CUDA_11_GENERATE \
     --build-arg=OLLAMA_SKIP_CUDA_12_GENERATE \
-    --build-arg=CUDA_V11_ARCHITECTURES \
     --build-arg=CUDA_V12_ARCHITECTURES \
     --build-arg=OLLAMA_SKIP_ROCM_GENERATE \
     --build-arg=OLLAMA_FAST_BUILD \

server/images.go

@@ -23,9 +23,10 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/fs/gguf"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/template"
+	"github.com/ollama/ollama/thinking"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
@@ -37,6 +38,7 @@ var (
 	errCapabilityInsert     = errors.New("insert")
 	errCapabilityVision     = errors.New("vision")
 	errCapabilityEmbedding  = errors.New("embedding")
+	errCapabilityThinking   = errors.New("thinking")
 	errInsecureProtocol     = errors.New("insecure protocol http")
 )
 
@@ -71,22 +73,18 @@ func (m *Model) Capabilities() []model.Capability {
 	capabilities := []model.Capability{}
 
 	// Check for completion capability
-	r, err := os.Open(m.ModelPath)
+	f, err := gguf.Open(m.ModelPath)
 	if err == nil {
-		defer r.Close()
+		defer f.Close()
 
-		f, err := ggml.Decode(r, 1024)
-		if err == nil {
-			if _, ok := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]; ok {
-				capabilities = append(capabilities, model.CapabilityEmbedding)
-			} else {
-				capabilities = append(capabilities, model.CapabilityCompletion)
-			}
-			if _, ok := f.KV()[fmt.Sprintf("%s.vision.block_count", f.KV().Architecture())]; ok {
-				capabilities = append(capabilities, model.CapabilityVision)
-			}
+		if f.KeyValue("pooling_type").Valid() {
+			capabilities = append(capabilities, model.CapabilityEmbedding)
 		} else {
-			slog.Error("couldn't decode ggml", "error", err)
+			// If no embedding is specified, we assume the model supports completion
+			capabilities = append(capabilities, model.CapabilityCompletion)
+		}
+		if f.KeyValue("vision.block_count").Valid() {
+			capabilities = append(capabilities, model.CapabilityVision)
 		}
 	} else {
 		slog.Error("couldn't open model file", "error", err)
@@ -111,6 +109,12 @@ func (m *Model) Capabilities() []model.Capability {
 		capabilities = append(capabilities, model.CapabilityVision)
 	}
 
+	// Check for thinking capability
+	openingTag, closingTag := thinking.InferTags(m.Template.Template)
+	if openingTag != "" && closingTag != "" {
+		capabilities = append(capabilities, model.CapabilityThinking)
+	}
+
 	return capabilities
 }
 
@@ -127,6 +131,7 @@ func (m *Model) CheckCapabilities(want ...model.Capability) error {
 		model.CapabilityInsert:     errCapabilityInsert,
 		model.CapabilityVision:     errCapabilityVision,
 		model.CapabilityEmbedding:  errCapabilityEmbedding,
+		model.CapabilityThinking:   errCapabilityThinking,
 	}
 
 	for _, cap := range want {
@@ -141,11 +146,19 @@ func (m *Model) CheckCapabilities(want ...model.Capability) error {
 		}
 	}
 
+	var err error
 	if len(errs) > 0 {
-		return fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
+		err = fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
 	}
 
-	return nil
+	if slices.Contains(errs, errCapabilityThinking) {
+		if m.Config.ModelFamily == "qwen3" || model.ParseName(m.Name).Model == "deepseek-r1" {
+			// append a message to the existing error
+			return fmt.Errorf("%w. Pull the model again to get the latest version with full thinking support", err)
+		}
+	}
+
+	return err
 }
 
 func (m *Model) String() string {

server/images_test.go

@@ -1,123 +1,42 @@
 package server
 
 import (
-	"bytes"
-	"encoding/binary"
-	"errors"
-	"os"
-	"path/filepath"
 	"strings"
 	"testing"
 
+	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/model"
 )
 
-// Constants for GGUF magic bytes and version
-var (
-	ggufMagic = []byte{0x47, 0x47, 0x55, 0x46} // "GGUF"
-	ggufVer   = uint32(3)                      // Version 3
-)
-
-// Helper function to create mock GGUF data
-func createMockGGUFData(architecture string, vision bool) []byte {
-	var buf bytes.Buffer
-
-	// Write GGUF header
-	buf.Write(ggufMagic)
-	binary.Write(&buf, binary.LittleEndian, ggufVer)
-
-	// Write tensor count (0 for our test)
-	var numTensors uint64 = 0
-	binary.Write(&buf, binary.LittleEndian, numTensors)
-
-	// Calculate number of metadata entries
-	numMetaEntries := uint64(1) // architecture entry
-	if vision {
-		numMetaEntries++
-	}
-	// Add embedding entry if architecture is "bert"
-	if architecture == "bert" {
-		numMetaEntries++
-	}
-	binary.Write(&buf, binary.LittleEndian, numMetaEntries)
-
-	// Write architecture metadata
-	archKey := "general.architecture"
-	keyLen := uint64(len(archKey))
-	binary.Write(&buf, binary.LittleEndian, keyLen)
-	buf.WriteString(archKey)
-
-	// String type (8)
-	var strType uint32 = 8
-	binary.Write(&buf, binary.LittleEndian, strType)
-
-	// String length
-	strLen := uint64(len(architecture))
-	binary.Write(&buf, binary.LittleEndian, strLen)
-	buf.WriteString(architecture)
-
-	if vision {
-		visionKey := architecture + ".vision.block_count"
-		keyLen = uint64(len(visionKey))
-		binary.Write(&buf, binary.LittleEndian, keyLen)
-		buf.WriteString(visionKey)
-
-		// uint32 type (4)
-		var uint32Type uint32 = 4
-		binary.Write(&buf, binary.LittleEndian, uint32Type)
-
-		// uint32 value (1)
-		var countVal uint32 = 1
-		binary.Write(&buf, binary.LittleEndian, countVal)
-	}
-	// Write embedding metadata if architecture is "bert"
-	if architecture == "bert" {
-		poolKey := architecture + ".pooling_type"
-		keyLen = uint64(len(poolKey))
-		binary.Write(&buf, binary.LittleEndian, keyLen)
-		buf.WriteString(poolKey)
-
-		// uint32 type (4)
-		var uint32Type uint32 = 4
-		binary.Write(&buf, binary.LittleEndian, uint32Type)
-
-		// uint32 value (1)
-		var poolingVal uint32 = 1
-		binary.Write(&buf, binary.LittleEndian, poolingVal)
-	}
-
-	return buf.Bytes()
-}
-
 func TestModelCapabilities(t *testing.T) {
-	// Create a temporary directory for test files
-	tempDir := t.TempDir()
+	// Create completion model (llama architecture without vision)
+	completionModelPath, _ := createBinFile(t, ggml.KV{
+		"general.architecture": "llama",
+	}, []*ggml.Tensor{})
 
-	// Create different types of mock model files
-	completionModelPath := filepath.Join(tempDir, "model.bin")
-	visionModelPath := filepath.Join(tempDir, "vision_model.bin")
-	embeddingModelPath := filepath.Join(tempDir, "embedding_model.bin")
-	// Create a simple model file for tests that don't depend on GGUF content
-	simpleModelPath := filepath.Join(tempDir, "simple_model.bin")
+	// Create vision model (llama architecture with vision block count)
+	visionModelPath, _ := createBinFile(t, ggml.KV{
+		"general.architecture":     "llama",
+		"llama.vision.block_count": uint32(1),
+	}, []*ggml.Tensor{})
 
-	if err := errors.Join(
-		os.WriteFile(completionModelPath, createMockGGUFData("llama", false), 0o644),
-		os.WriteFile(visionModelPath, createMockGGUFData("llama", true), 0o644),
-		os.WriteFile(embeddingModelPath, createMockGGUFData("bert", false), 0o644),
-		os.WriteFile(simpleModelPath, []byte("dummy model data"), 0o644),
-	); err != nil {
-		t.Fatalf("Failed to create model files: %v", err)
-	}
+	// Create embedding model (bert architecture with pooling type)
+	embeddingModelPath, _ := createBinFile(t, ggml.KV{
+		"general.architecture": "bert",
+		"bert.pooling_type":    uint32(1),
+	}, []*ggml.Tensor{})
 
 	toolsInsertTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}{{ if .suffix }}{{ .suffix }}{{ end }}")
 	if err != nil {
 		t.Fatalf("Failed to parse template: %v", err)
 	}
+
 	chatTemplate, err := template.Parse("{{ .prompt }}")
 	if err != nil {
 		t.Fatalf("Failed to parse template: %v", err)
 	}
+
 	toolsTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}")
 	if err != nil {
 		t.Fatalf("Failed to parse template: %v", err)
@@ -145,21 +64,13 @@ func TestModelCapabilities(t *testing.T) {
 			},
 			expectedCaps: []model.Capability{model.CapabilityCompletion, model.CapabilityTools, model.CapabilityInsert},
 		},
-		{
-			name: "model with tools and insert capability",
-			model: Model{
-				ModelPath: simpleModelPath,
-				Template:  toolsInsertTemplate,
-			},
-			expectedCaps: []model.Capability{model.CapabilityTools, model.CapabilityInsert},
-		},
 		{
 			name: "model with tools capability",
 			model: Model{
-				ModelPath: simpleModelPath,
+				ModelPath: completionModelPath,
 				Template:  toolsTemplate,
 			},
-			expectedCaps: []model.Capability{model.CapabilityTools},
+			expectedCaps: []model.Capability{model.CapabilityCompletion, model.CapabilityTools},
 		},
 		{
 			name: "model with vision capability",
@@ -224,29 +135,33 @@ func TestModelCapabilities(t *testing.T) {
 }
 
 func TestModelCheckCapabilities(t *testing.T) {
-	// Create a temporary directory for test files
-	tempDir := t.TempDir()
+	// Create simple model file for tests that don't depend on GGUF content
+	completionModelPath, _ := createBinFile(t, ggml.KV{
+		"general.architecture": "llama",
+	}, []*ggml.Tensor{})
 
-	visionModelPath := filepath.Join(tempDir, "vision_model.bin")
-	simpleModelPath := filepath.Join(tempDir, "model.bin")
-	embeddingModelPath := filepath.Join(tempDir, "embedding_model.bin")
+	// Create vision model (llama architecture with vision block count)
+	visionModelPath, _ := createBinFile(t, ggml.KV{
+		"general.architecture":     "llama",
+		"llama.vision.block_count": uint32(1),
+	}, []*ggml.Tensor{})
 
-	if err := errors.Join(
-		os.WriteFile(simpleModelPath, []byte("dummy model data"), 0o644),
-		os.WriteFile(visionModelPath, createMockGGUFData("llama", true), 0o644),
-		os.WriteFile(embeddingModelPath, createMockGGUFData("bert", false), 0o644),
-	); err != nil {
-		t.Fatalf("Failed to create model files: %v", err)
-	}
+	// Create embedding model (bert architecture with pooling type)
+	embeddingModelPath, _ := createBinFile(t, ggml.KV{
+		"general.architecture": "bert",
+		"bert.pooling_type":    uint32(1),
+	}, []*ggml.Tensor{})
 
 	toolsInsertTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}{{ if .suffix }}{{ .suffix }}{{ end }}")
 	if err != nil {
 		t.Fatalf("Failed to parse template: %v", err)
 	}
+
 	chatTemplate, err := template.Parse("{{ .prompt }}")
 	if err != nil {
 		t.Fatalf("Failed to parse template: %v", err)
 	}
+
 	toolsTemplate, err := template.Parse("{{ .prompt }}{{ if .tools }}{{ .tools }}{{ end }}")
 	if err != nil {
 		t.Fatalf("Failed to parse template: %v", err)
@@ -261,7 +176,7 @@ func TestModelCheckCapabilities(t *testing.T) {
 		{
 			name: "completion model without tools capability",
 			model: Model{
-				ModelPath: simpleModelPath,
+				ModelPath: completionModelPath,
 				Template:  chatTemplate,
 			},
 			checkCaps:      []model.Capability{model.CapabilityTools},
@@ -270,7 +185,7 @@ func TestModelCheckCapabilities(t *testing.T) {
 		{
 			name: "model with all needed capabilities",
 			model: Model{
-				ModelPath: simpleModelPath,
+				ModelPath: completionModelPath,
 				Template:  toolsInsertTemplate,
 			},
 			checkCaps: []model.Capability{model.CapabilityTools, model.CapabilityInsert},
@@ -278,7 +193,7 @@ func TestModelCheckCapabilities(t *testing.T) {
 		{
 			name: "model missing insert capability",
 			model: Model{
-				ModelPath: simpleModelPath,
+				ModelPath: completionModelPath,
 				Template:  toolsTemplate,
 			},
 			checkCaps:      []model.Capability{model.CapabilityInsert},
@@ -287,7 +202,7 @@ func TestModelCheckCapabilities(t *testing.T) {
 		{
 			name: "model missing vision capability",
 			model: Model{
-				ModelPath: simpleModelPath,
+				ModelPath: completionModelPath,
 				Template:  toolsTemplate,
 			},
 			checkCaps:      []model.Capability{model.CapabilityVision},
@@ -312,7 +227,7 @@ func TestModelCheckCapabilities(t *testing.T) {
 		{
 			name: "unknown capability",
 			model: Model{
-				ModelPath: simpleModelPath,
+				ModelPath: completionModelPath,
 				Template:  chatTemplate,
 			},
 			checkCaps:      []model.Capability{"unknown"},

server/internal/cache/blob/cache.go

@@ -59,7 +59,7 @@ type DiskCache struct {
 	testHookBeforeFinalWrite func(f *os.File)
 }
 
-// PutString is a convenience function for c.Put(d, strings.NewReader(s), int64(len(s))).
+// PutBytes is a convenience function for c.Put(d, strings.NewReader(s), int64(len(s))).
 func PutBytes[S string | []byte](c *DiskCache, d Digest, data S) error {
 	return c.Put(d, bytes.NewReader([]byte(data)), int64(len(data)))
 }

server/prompt.go

@@ -19,7 +19,7 @@ type tokenizeFunc func(context.Context, string) ([]int, error)
 // chatPrompt accepts a list of messages and returns the prompt and images that should be used for the next chat turn.
 // chatPrompt truncates any messages that exceed the context window of the model, making sure to always include 1) the
 // latest message and 2) system messages
-func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool) (prompt string, images []llm.ImageData, _ error) {
+func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool, think *bool) (prompt string, images []llm.ImageData, _ error) {
 	var system []api.Message
 
 	// TODO: Ideally we would compute this from the projector metadata but some pieces are implementation dependent
@@ -41,8 +41,12 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 			}
 		}
 
+		thinkVal := false
+		if think != nil {
+			thinkVal = *think
+		}
 		var b bytes.Buffer
-		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools}); err != nil {
+		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools, Think: thinkVal, IsThinkSet: think != nil}); err != nil {
 			return "", nil, err
 		}
 
@@ -96,7 +100,11 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 
 	// truncate any messages that do not fit into the context window
 	var b bytes.Buffer
-	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[currMsgIdx:]...), Tools: tools}); err != nil {
+	thinkVal := false
+	if think != nil {
+		thinkVal = *think
+	}
+	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[currMsgIdx:]...), Tools: tools, Think: thinkVal, IsThinkSet: think != nil}); err != nil {
 		return "", nil, err
 	}
 

server/prompt_test.go

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

server/quantization_test.go

@@ -257,16 +257,8 @@ func TestQuantizeModel(t *testing.T) {
 
 	for _, tt := range cases {
 		t.Run(tt.name, func(t *testing.T) {
-			f, err := os.CreateTemp(t.TempDir(), tt.name)
-			if err != nil {
-				t.Fatal(err.Error())
-			}
-			defer f.Close()
-			err = fsggml.WriteGGUF(f, tt.kv, tt.tensors)
-			if err != nil {
-				t.Fatalf("failed to create initial model: %s", err)
-			}
-			fp, err := os.Open(f.Name())
+			p, _ := createBinFile(t, tt.kv, tt.tensors)
+			fp, err := os.Open(p)
 			if err != nil {
 				t.Fatal(err.Error())
 			}

server/routes.go

@@ -17,7 +17,6 @@ import (
 	"net/netip"
 	"os"
 	"os/signal"
-	"regexp"
 	"slices"
 	"strings"
 	"syscall"
@@ -38,6 +37,7 @@ import (
 	"github.com/ollama/ollama/server/internal/client/ollama"
 	"github.com/ollama/ollama/server/internal/registry"
 	"github.com/ollama/ollama/template"
+	"github.com/ollama/ollama/thinking"
 	"github.com/ollama/ollama/tools"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
@@ -186,6 +186,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 	if req.Suffix != "" {
 		caps = append(caps, model.CapabilityInsert)
 	}
+	if req.Think != nil && *req.Think {
+		caps = append(caps, model.CapabilityThinking)
+		// TODO(drifkin): consider adding a warning if it's false and the model
+		// doesn't support thinking. It's not strictly required, but it can be a
+		// hint that the user is on an older qwen3/r1 model that doesn't have an
+		// updated template supporting thinking
+	}
 
 	r, m, opts, err := s.scheduleRunner(c.Request.Context(), name.String(), caps, req.Options, req.KeepAlive)
 	if errors.Is(err, errCapabilityCompletion) {
@@ -254,6 +261,9 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			values.Messages = append(msgs, api.Message{Role: "user", Content: req.Prompt})
 		}
 
+		values.Think = req.Think != nil && *req.Think
+		values.IsThinkSet = req.Think != nil
+
 		var b bytes.Buffer
 		if req.Context != nil {
 			slog.Warn("the context field is deprecated and will be removed in a future version of Ollama")
@@ -273,6 +283,15 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		prompt = b.String()
 	}
 
+	var thinkingState *thinking.Parser
+	openingTag, closingTag := thinking.InferTags(m.Template.Template)
+	if req.Think != nil && *req.Think && openingTag != "" && closingTag != "" {
+		thinkingState = &thinking.Parser{
+			OpeningTag: openingTag,
+			ClosingTag: closingTag,
+		}
+	}
+
 	ch := make(chan any)
 	go func() {
 		// TODO (jmorganca): avoid building the response twice both here and below
@@ -297,6 +316,12 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 				},
 			}
 
+			if thinkingState != nil {
+				thinking, content := thinkingState.AddContent(cr.Content)
+				res.Thinking = thinking
+				res.Response = content
+			}
+
 			if _, err := sb.WriteString(cr.Content); err != nil {
 				ch <- gin.H{"error": err.Error()}
 			}
@@ -324,11 +349,13 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	if req.Stream != nil && !*req.Stream {
 		var r api.GenerateResponse
-		var sb strings.Builder
+		var sbThinking strings.Builder
+		var sbContent strings.Builder
 		for rr := range ch {
 			switch t := rr.(type) {
 			case api.GenerateResponse:
-				sb.WriteString(t.Response)
+				sbThinking.WriteString(t.Thinking)
+				sbContent.WriteString(t.Response)
 				r = t
 			case gin.H:
 				msg, ok := t["error"].(string)
@@ -344,7 +371,9 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 			}
 		}
 
-		r.Response = sb.String()
+		r.Thinking = sbThinking.String()
+		r.Response = sbContent.String()
+
 		c.JSON(http.StatusOK, r)
 		return
 	}
@@ -1436,6 +1465,9 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	if len(req.Tools) > 0 {
 		caps = append(caps, model.CapabilityTools)
 	}
+	if req.Think != nil && *req.Think {
+		caps = append(caps, model.CapabilityThinking)
+	}
 
 	name := model.ParseName(req.Model)
 	if !name.IsValid() {
@@ -1476,21 +1508,25 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 	msgs = filterThinkTags(msgs, m)
 
-	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, req.Tools)
+	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, req.Tools, req.Think)
 	if err != nil {
 		slog.Error("chat prompt error", "error", err)
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
 
+	var thinkingState *thinking.Parser
+	openingTag, closingTag := thinking.InferTags(m.Template.Template)
+	if req.Think != nil && *req.Think && openingTag != "" && closingTag != "" {
+		thinkingState = &thinking.Parser{
+			OpeningTag: openingTag,
+			ClosingTag: closingTag,
+		}
+	}
+
 	var toolParser *tools.Parser
 	if len(req.Tools) > 0 {
-		toolParser, err = tools.NewParser(m.Template.Template)
-		if err != nil {
-			slog.Error("failed to create tool parser", "error", err)
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
+		toolParser = tools.NewParser(m.Template.Template, req.Tools)
 	}
 
 	ch := make(chan any)
@@ -1516,6 +1552,16 @@ func (s *Server) ChatHandler(c *gin.Context) {
 				},
 			}
 
+			if thinkingState != nil {
+				thinkingContent, remainingContent := thinkingState.AddContent(res.Message.Content)
+				if thinkingContent == "" && remainingContent == "" && !r.Done {
+					// need to accumulate more to decide what to send
+					return
+				}
+				res.Message.Content = remainingContent
+				res.Message.Thinking = thinkingContent
+			}
+
 			if r.Done {
 				res.DoneReason = r.DoneReason.String()
 				res.TotalDuration = time.Since(checkpointStart)
@@ -1523,19 +1569,23 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 
 			if len(req.Tools) > 0 {
-				toolCalls, content := toolParser.Add(r.Content)
+				toolCalls, content := toolParser.Add(res.Message.Content)
 				if len(content) > 0 {
 					res.Message.Content = content
 				} else if len(toolCalls) > 0 {
 					res.Message.ToolCalls = toolCalls
 					res.Message.Content = ""
+				} else if res.Message.Thinking != "" {
+					// don't return
 				} else {
 					if r.Done {
+						res.Message.Content = toolParser.Content()
 						ch <- res
 					}
 					return
 				}
 			}
+
 			ch <- res
 		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
@@ -1544,12 +1594,14 @@ func (s *Server) ChatHandler(c *gin.Context) {
 
 	if req.Stream != nil && !*req.Stream {
 		var resp api.ChatResponse
-		var sb strings.Builder
 		var toolCalls []api.ToolCall
+		var sbThinking strings.Builder
+		var sbContent strings.Builder
 		for rr := range ch {
 			switch t := rr.(type) {
 			case api.ChatResponse:
-				sb.WriteString(t.Message.Content)
+				sbThinking.WriteString(t.Message.Thinking)
+				sbContent.WriteString(t.Message.Content)
 				resp = t
 				if len(req.Tools) > 0 {
 					toolCalls = append(toolCalls, t.Message.ToolCalls...)
@@ -1568,7 +1620,9 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 		}
 
-		resp.Message.Content = sb.String()
+		resp.Message.Content = sbContent.String()
+		resp.Message.Thinking = sbThinking.String()
+
 		if len(toolCalls) > 0 {
 			resp.Message.ToolCalls = toolCalls
 		}
@@ -1595,8 +1649,6 @@ func handleScheduleError(c *gin.Context, name string, err error) {
 	}
 }
 
-var thinkTagRegexp = regexp.MustCompile(`<think>(?s).*?</think>(\n)*`)
-
 func filterThinkTags(msgs []api.Message, m *Model) []api.Message {
 	if m.Config.ModelFamily == "qwen3" || model.ParseName(m.Name).Model == "deepseek-r1" {
 		finalUserIndex := -1
@@ -1608,7 +1660,17 @@ func filterThinkTags(msgs []api.Message, m *Model) []api.Message {
 
 		for i, msg := range msgs {
 			if msg.Role == "assistant" && i < finalUserIndex {
-				msgs[i].Content = thinkTagRegexp.ReplaceAllString(msg.Content, "")
+				// TODO(drifkin): this is from before we added proper thinking support.
+				// However, even if thinking is not enabled (and therefore we shouldn't
+				// change the user output), we should probably perform this filtering
+				// for all thinking models (not just qwen3 & deepseek-r1) since it tends
+				// to save tokens and improve quality.
+				thinkingState := &thinking.Parser{
+					OpeningTag: "<think>",
+					ClosingTag: "</think>",
+				}
+				_, content := thinkingState.AddContent(msg.Content)
+				msgs[i].Content = content
 			}
 		}
 	}

server/routes_generate_test.go

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

server/sched.go

@@ -191,7 +191,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					}
 
 					// Load model for fitting
-					ggml, err := llm.LoadModel(pending.model.ModelPath, 0)
+					ggml, err := llm.LoadModel(pending.model.ModelPath, 1024)
 					if err != nil {
 						pending.errCh <- err
 						break

server/sched_test.go

@@ -112,11 +112,7 @@ func newScenarioRequest(t *testing.T, ctx context.Context, modelName string, est
 	b.ctx, b.ctxDone = context.WithCancel(ctx)
 	t.Helper()
 
-	f, err := os.CreateTemp(t.TempDir(), modelName)
-	require.NoError(t, err)
-	defer f.Close()
-
-	require.NoError(t, ggml.WriteGGUF(f, ggml.KV{
+	p, _ := createBinFile(t, ggml.KV{
 		"general.architecture":          "llama",
 		"llama.context_length":          uint32(32),
 		"llama.embedding_length":        uint32(4096),
@@ -129,14 +125,14 @@ func newScenarioRequest(t *testing.T, ctx context.Context, modelName string, est
 	}, []*ggml.Tensor{
 		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
 		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-	}))
-	require.NoError(t, err)
-
-	fname := f.Name()
-	model := &Model{Name: modelName, ModelPath: fname}
-	b.f, err = llm.LoadModel(model.ModelPath, 0)
-	require.NoError(t, err)
+	})
 
+	model := &Model{Name: modelName, ModelPath: p}
+	f, err := llm.LoadModel(model.ModelPath, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+	b.f = f
 	if duration == nil {
 		duration = &api.Duration{Duration: 5 * time.Millisecond}
 	}

template/template.go

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

thinking/parser.go

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

thinking/parser_test.go

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

thinking/template.go

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

thinking/template_test.go

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

tools/template.go

@@ -0,0 +1,156 @@
+package tools
+
+import (
+	"bytes"
+	"log/slog"
+	"slices"
+	"strings"
+	"text/template"
+	"text/template/parse"
+)
+
+// parseTag finds the tool calling tag from a Go template
+// often <tool_call> [TOOL_CALL] or similar by finding the
+// first text node after .ToolCalls and returning the content
+// if no tag is found, return "{" to indicate that json objects
+// should be attempted to be parsed as tool calls
+func parseTag(tmpl *template.Template) string {
+	if tmpl == nil || tmpl.Tree == nil {
+		slog.Debug("template or tree is nil")
+		return "{"
+	}
+
+	tc := findToolCallNode(tmpl.Tree.Root.Nodes)
+	if tc == nil {
+		return "{"
+	}
+
+	tn := findTextNode(tc.List.Nodes)
+	if tn == nil {
+		return "{"
+	}
+
+	tag := string(tn.Text)
+	tag = strings.ReplaceAll(tag, "\r\n", "\n")
+
+	// avoid parsing { onwards as this may be a tool call
+	// however keep '{' as a prefix if there is no tag
+	// so that all json objects will be attempted to
+	// be parsed as tool calls
+	tag, _, _ = strings.Cut(tag, "{")
+	tag = strings.TrimSpace(tag)
+	if tag == "" {
+		tag = "{"
+	}
+
+	return tag
+}
+
+// findToolCallNode searches for and returns an IfNode with .ToolCalls
+func findToolCallNode(nodes []parse.Node) *parse.IfNode {
+	isToolCallsNode := func(n *parse.IfNode) bool {
+		for _, cmd := range n.Pipe.Cmds {
+			for _, arg := range cmd.Args {
+				if field, ok := arg.(*parse.FieldNode); ok {
+					if slices.Contains(field.Ident, "ToolCalls") {
+						return true
+					}
+				}
+			}
+		}
+		return false
+	}
+
+	for _, node := range nodes {
+		switch n := node.(type) {
+		case *parse.IfNode:
+			if isToolCallsNode(n) {
+				return n
+			}
+			// Recursively search in nested IfNodes
+			if result := findToolCallNode(n.List.Nodes); result != nil {
+				return result
+			}
+			if n.ElseList != nil {
+				if result := findToolCallNode(n.ElseList.Nodes); result != nil {
+					return result
+				}
+			}
+		case *parse.ListNode:
+			if result := findToolCallNode(n.Nodes); result != nil {
+				return result
+			}
+		case *parse.RangeNode:
+			if result := findToolCallNode(n.List.Nodes); result != nil {
+				return result
+			}
+			if n.ElseList != nil {
+				if result := findToolCallNode(n.ElseList.Nodes); result != nil {
+					return result
+				}
+			}
+		case *parse.WithNode:
+			if result := findToolCallNode(n.List.Nodes); result != nil {
+				return result
+			}
+			if n.ElseList != nil {
+				if result := findToolCallNode(n.ElseList.Nodes); result != nil {
+					return result
+				}
+			}
+		}
+	}
+	return nil
+}
+
+// findTextNode does a depth-first search for the first text content in nodes,
+// stopping at template constructs to avoid parsing text after the tool calls
+func findTextNode(nodes []parse.Node) *parse.TextNode {
+	for _, node := range nodes {
+		switch n := node.(type) {
+		case *parse.TextNode:
+			// skip whitespace-only text nodes
+			if len(bytes.TrimSpace(n.Text)) == 0 {
+				continue
+			}
+			return n
+		case *parse.IfNode:
+			if text := findTextNode(n.List.Nodes); text != nil {
+				return text
+			}
+			if n.ElseList != nil {
+				if text := findTextNode(n.ElseList.Nodes); text != nil {
+					return text
+				}
+			}
+			return nil
+		case *parse.ListNode:
+			if text := findTextNode(n.Nodes); text != nil {
+				return text
+			}
+		case *parse.RangeNode:
+			if text := findTextNode(n.List.Nodes); text != nil {
+				return text
+			}
+			if n.ElseList != nil {
+				if text := findTextNode(n.ElseList.Nodes); text != nil {
+					return text
+				}
+			}
+			return nil
+		case *parse.WithNode:
+			if text := findTextNode(n.List.Nodes); text != nil {
+				return text
+			}
+			if n.ElseList != nil {
+				if text := findTextNode(n.ElseList.Nodes); text != nil {
+					return text
+				}
+			}
+			return nil
+		case *parse.ActionNode:
+			return nil
+		}
+	}
+	return nil
+}

tools/template_test.go

@@ -0,0 +1,139 @@
+package tools
+
+import (
+	"testing"
+	"text/template"
+)
+
+func TestParseTag(t *testing.T) {
+	cases := []struct {
+		name     string
+		template string
+		want     string
+	}{
+		{
+			name:     "empty",
+			template: "",
+			want:     "{",
+		},
+		{
+			name:     "no tag",
+			template: "{{if .ToolCalls}}{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "no tag with range",
+			template: "{{if .ToolCalls}}{{range .ToolCalls}}{{ . }}{{end}}{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "tool call with json format",
+			template: "{{if .ToolCalls}}```json\n{{end}}",
+			want:     "```json",
+		},
+		{
+			name:     "square brackets",
+			template: "{{if .ToolCalls}}[{{range .ToolCalls}}{{ . }}{{end}}]{{end}}",
+			want:     "[",
+		},
+		{
+			name:     "square brackets with whitespace",
+			template: "{{if .ToolCalls}}\n [ {{range .ToolCalls}}{{ . }}{{end}}]{{end}}",
+			want:     "[",
+		},
+		{
+			name:     "tailing ]",
+			template: "{{if .ToolCalls}}{{range .ToolCalls}}{{ . }}{{end}}]{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "whitespace only",
+			template: "{{if .ToolCalls}} {{range .ToolCalls}}{{ . }}{{end}}{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "whitespace only in range",
+			template: "{{if .ToolCalls}}{{range .ToolCalls}}\n{{ . }}\n{{end}}{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "json objects",
+			template: `{{if .ToolCalls}}{{range .ToolCalls}}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}{{end}}{{end}}`,
+			want:     "{",
+		},
+		{
+			name:     "json objects with whitespace",
+			template: "{{if .ToolCalls}}{{range .ToolCalls}}\n{\"name\": \"{{ .Function.Name }}\", \"arguments\": {{ .Function.Arguments }}}{{end}}{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "json objects with CRLF",
+			template: "{{if .ToolCalls}}{{range .ToolCalls}}\r\n{\"name\": \"{{ .Function.Name }}\", \"arguments\": {{ .Function.Arguments }}}{{end}}{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "json objects with whitespace before and after range",
+			template: "{{if .ToolCalls}}\n{{range .ToolCalls}}\n{\"name\": \"{{ .Function.Name }}\", \"arguments\": {{ .Function.Arguments }}}\r\n{{end}}\r\n{{end}}",
+			want:     "{",
+		},
+		{
+			name:     "before and after range",
+			template: "{{if .ToolCalls}}<|tool▁calls▁begin|>{{range .ToolCalls}}<|tool▁call▁begin|>functionget_current_weather\n```json\n{\"location\": \"Tokyo\"}\n```<|tool▁call▁end|>\n{{end}}<|tool▁calls▁end|>{{end}}",
+			want:     "<|tool▁calls▁begin|>",
+		},
+		{
+			name:     "after range",
+			template: "{{if .ToolCalls}}{{range .ToolCalls}}<tool_call>{\"name\": \"{{ .Function.Name }}\", \"arguments\": {{ .Function.Arguments }}}</tool_call>{{end}}{{end}}",
+			want:     "<tool_call>",
+		},
+		{
+			name:     "after range with leading whitespace before range",
+			template: "{{if .ToolCalls}}\n{{range .ToolCalls}}<tool_call>{\"name\": \"{{ .Function.Name }}\", \"arguments\": {{ .Function.Arguments }}}</tool_call>{{end}}{{end}}",
+			want:     "<tool_call>",
+		},
+		{
+			name:     "tool call in range with {",
+			template: `{{if .ToolCalls}}{{range .ToolCalls}}<tool_call>{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}<tool_call>{{end}}{{end}}`,
+			want:     "<tool_call>",
+		},
+		{
+			name:     "tool call with multiple text nodes",
+			template: "{{if .ToolCalls}}First text{{if .Something}}inner{{end}}Second text{{end}}",
+			want:     "First text",
+		},
+		{
+			name:     "action tag",
+			template: "{{if .ToolCalls}}Action: ```json{{end}}",
+			want:     "Action: ```json",
+		},
+		{
+			name:     "incomplete functools bracket",
+			template: "{{if .ToolCalls}}functools[{{end}}",
+			want:     "functools[",
+		},
+		{
+			name:     "uppercase tool call with incomplete bracket",
+			template: "{{if .ToolCalls}}[TOOL_CALL] [{{end}}",
+			want:     "[TOOL_CALL] [",
+		},
+		{
+			name:     "uppercase tool call with adjacent bracket",
+			template: "{{if .ToolCalls}}[TOOL_CALL][{{end}}",
+			want:     "[TOOL_CALL][",
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			tmpl, err := template.New("test").Parse(tc.template)
+			if err != nil && tc.template != "" {
+				t.Fatalf("failed to parse template: %v", err)
+			}
+
+			got := parseTag(tmpl)
+			if got != tc.want {
+				t.Errorf("got text %q, want %q", got, tc.want)
+			}
+		})
+	}
+}

tools/testdata/command-r-plus.gotmpl

@@ -1,67 +0,0 @@
-{{- if or .Tools .System }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>
-{{- if .Tools }}# Safety Preamble
-The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
-
-# System Preamble
-## Basic Rules
-You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
-
-{{ if .System }}# User Preamble
-{{ .System }}
-{{- end }}
-
-## Available Tools
-Here is a list of tools that you have available to you:
-{{- range .Tools }}
-
-```python
-def {{ .Function.Name }}(
-{{- range $name, $property := .Function.Parameters.Properties }}{{ $name }}: {{ $property.Type }}, {{ end }}) -> List[Dict]:
-    """{{ .Function.Description }}
-
-{{- if .Function.Parameters.Properties }}
-
-    Args:
-{{- range $name, $property := .Function.Parameters.Properties }}
-        {{ $name }} ({{ $property.Type }}): {{ $property.Description }}
-{{- end }}
-{{- end }}
-    """
-    pass
-```
-{{- end }}
-{{- else if .System }}{{ .System }}
-{{- end }}<|END_OF_TURN_TOKEN|>
-{{- end }}
-{{- range .Messages }}
-{{- if eq .Role "system" }}
-{{- continue }}
-{{- end }}<|START_OF_TURN_TOKEN|>
-{{- if eq .Role "user" }}<|USER_TOKEN|>{{ .Content }}
-{{- else if eq .Role "assistant" }}<|CHATBOT_TOKEN|>
-{{- if .Content }}{{ .Content }}
-{{- else if .ToolCalls }}
-Action: ```json
-[
-{{- range .ToolCalls }}
-    {
-        "tool_name": "{{ .Function.Name }}",
-        "parameters": {{ .Function.Arguments }}
-    }
-{{- end }}
-]```
-{{ continue }}
-{{ end }}
-{{- else if eq .Role "tool" }}<|SYSTEM_TOKEN|><results>
-{{ .Content }}</results>
-{{- end }}<|END_OF_TURN_TOKEN|>
-{{- end }}
-{{- if .Tools }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
-```json
-[
-    {
-        "tool_name": title of the tool in the specification,
-        "parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
-    }
-]```
-{{- end }}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

tools/testdata/command-r-plus.out

@@ -1,39 +0,0 @@
-<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|># Safety Preamble
-The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
-
-# System Preamble
-## Basic Rules
-You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
-
-# User Preamble
-You are a knowledgeable assistant. You can answer questions and perform tasks.
-
-## Available Tools
-Here is a list of tools that you have available to you:
-
-```python
-def get_current_weather(format: string, location: string, ) -> List[Dict]:
-    """Get the current weather
-
-    Args:
-        format (string): The temperature unit to use. Infer this from the user's location.
-        location (string): The city and state, e.g. San Francisco, CA
-    """
-    pass
-```<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>What's the weather like today in Paris?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
-Action: ```json
-[
-    {
-        "tool_name": "get_current_weather",
-        "parameters": {"format":"celsius","location":"Paris, France"}
-    }
-]```
-<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|><results>
-22</results><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>The current temperature in Paris, France is 22 degrees Celsius.<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>What's the weather like today in San Francisco and Toronto?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
-```json
-[
-    {
-        "tool_name": title of the tool in the specification,
-        "parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
-    }
-]```<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

tools/testdata/firefunction.gotmpl

@@ -1,31 +0,0 @@
-{{- if or .System .Tools }}<|start_header_id|>system<|end_header_id|>
-{{- if .System }}
-{{ .System }}
-{{- end }}
-In addition to plain text responses, you can chose to call one or more of the provided functions.
-
-Use the following rule to decide when to call a function:
-  * if the response can be generated from your internal knowledge (e.g., as in the case of queries like "What is the capital of Poland?"), do so
-  * if you need external information that can be obtained by calling one or more of the provided functions, generate a function calls
-
-If you decide to call functions:
-  * prefix function calls with functools marker (no closing marker required)
-  * all function calls should be generated in a single JSON list formatted as functools[{"name": [function name], "arguments": [function arguments as JSON]}, ...]
-  * follow the provided JSON schema. Do not hallucinate arguments or values. Do to blindly copy values from the provided samples
-  * respect the argument type formatting. E.g., if the type if number and format is float, write value 7 as 7.0
-  * make sure you pick the right functions that match the user intent
-
-Available functions as JSON spec:
-{{- if .Tools }}
-{{ .Tools }}
-{{- end }}<|eot_id|>
-{{- end }}
-{{- range .Messages }}<|start_header_id|>
-{{- if or (eq .Role "user") (eq .Role "assistant") (eq .Role "tool") }}{{ .Role }}
-{{- end }}<|end_header_id|>
-{{- if .Content }}{{ .Content }}
-{{- else if .ToolCalls }} functools[
-{{- range .ToolCalls }}{{ "{" }}"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}{{ "}" }}
-{{- end }}]
-{{- end }}<|eot_id|>
-{{- end }}<|start_header_id|>assistant<|end_header_id|>

tools/testdata/firefunction.out

@@ -1,17 +0,0 @@
-<|start_header_id|>system<|end_header_id|>
-You are a knowledgeable assistant. You can answer questions and perform tasks.
-In addition to plain text responses, you can chose to call one or more of the provided functions.
-
-Use the following rule to decide when to call a function:
-  * if the response can be generated from your internal knowledge (e.g., as in the case of queries like "What is the capital of Poland?"), do so
-  * if you need external information that can be obtained by calling one or more of the provided functions, generate a function calls
-
-If you decide to call functions:
-  * prefix function calls with functools marker (no closing marker required)
-  * all function calls should be generated in a single JSON list formatted as functools[{"name": [function name], "arguments": [function arguments as JSON]}, ...]
-  * follow the provided JSON schema. Do not hallucinate arguments or values. Do to blindly copy values from the provided samples
-  * respect the argument type formatting. E.g., if the type if number and format is float, write value 7 as 7.0
-  * make sure you pick the right functions that match the user intent
-
-Available functions as JSON spec:
-[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the user's location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}}]<|eot_id|><|start_header_id|><|end_header_id|>You are a knowledgeable assistant. You can answer questions and perform tasks.<|eot_id|><|start_header_id|>user<|end_header_id|>What's the weather like today in Paris?<|eot_id|><|start_header_id|>assistant<|end_header_id|> functools[{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]<|eot_id|><|start_header_id|>tool<|end_header_id|>22<|eot_id|><|start_header_id|>assistant<|end_header_id|>The current temperature in Paris, France is 22 degrees Celsius.<|eot_id|><|start_header_id|>user<|end_header_id|>What's the weather like today in San Francisco and Toronto?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

tools/testdata/llama3-groq-tool-use.gotmpl

@@ -1,43 +0,0 @@
-{{- if .Messages }}
-{{- if or .System .Tools }}<|start_header_id|>system<|end_header_id|>
-
-{{ .System }}
-{{- if .Tools }} You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
-<tool_call>
-{"name": <function-name>,"arguments": <args-dict>}
-</tool_call>
-
-Here are the available tools:
-<tools>
-{{- range .Tools }} {{ .Function }}
-{{- end }} </tools>
-{{- end }}
-{{- end }}<|eot_id|>
-{{- range .Messages }}
-{{- if ne .Role "system" }}<|start_header_id|>{{ .Role }}<|end_header_id|>
-
-{{ if eq .Role "user" }}{{ .Content }}
-{{- else if eq .Role "assistant" }}
-{{- if .Content }}{{ .Content }}
-{{- else if .ToolCalls }}<tool_call>
-{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
-{{- end }}
-</tool_call>
-{{- end }}
-{{- else if eq .Role "tool" }}<tool_response>
-{{ .Content }}
-</tool_response>
-{{- end }}<|eot_id|>
-{{- end }}
-{{- end }}<|start_header_id|>assistant<|end_header_id|>
-
-{{ else }}
-{{ if .System }}<|start_header_id|>system<|end_header_id|>
-
-{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
-
-{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
-
-{{ end }}{{ .Response }}
-{{- if .Response }}<|eot_id|>
-{{- end }}

tools/testdata/llama3-groq-tool-use.out

@@ -1,24 +0,0 @@
-<|start_header_id|>system<|end_header_id|>
-
-You are a knowledgeable assistant. You can answer questions and perform tasks. You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
-<tool_call>
-{"name": <function-name>,"arguments": <args-dict>}
-</tool_call>
-
-Here are the available tools:
-<tools> {"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the user's location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}} </tools><|eot_id|><|start_header_id|>user<|end_header_id|>
-
-What's the weather like today in Paris?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
-
-<tool_call>
-{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}
-</tool_call><|eot_id|><|start_header_id|>tool<|end_header_id|>
-
-<tool_response>
-22
-</tool_response><|eot_id|><|start_header_id|>assistant<|end_header_id|>
-
-The current temperature in Paris, France is 22 degrees Celsius.<|eot_id|><|start_header_id|>user<|end_header_id|>
-
-What's the weather like today in San Francisco and Toronto?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
-