wip

This package provides a way to convert JSON schemas to equivalent EBNF.
It is intended to be a replacement to llama.cpp's schema_to_grammar.

This is still an early version and does not yet support all JSON schema
features. The to-do list includes:

- minumum/maximum constraints on integer types
- minLength/maxLength constraints on string types
- defs and refs

.dockerignore

@@ -3,9 +3,7 @@ ollama
 app
 macapp
 dist
-build
 .env
 .cache
 test_data
-.git
-
+llama/build

.gitattributes

@@ -15,10 +15,6 @@ ml/backend/**/*.cu linguist-vendored
 ml/backend/**/*.cuh linguist-vendored
 ml/backend/**/*.m linguist-vendored
 ml/backend/**/*.metal linguist-vendored
-ml/backend/**/CMakeLists.txt linguist-vendored
-
-llama/build-info.cpp linguist-generated
-ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.s linguist-generated
 
 * text=auto
 *.go text eol=lf

.github/ISSUE_TEMPLATE/10_bug_report.yml

@@ -9,14 +9,6 @@ body:
       description: What happened? What did you expect to happen?
     validations:
       required: true
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) for details.
-      render: shell
-    validations:
-      required: false
   - type: dropdown
     id: os
     attributes:

.github/workflows/test.yaml

@@ -40,113 +40,28 @@ jobs:
 
   linux:
     needs: [changes]
-    if: needs.changes.outputs.changed == 'True'
+    if: ${{ needs.changes.outputs.changed == 'True' }}
     strategy:
       matrix:
         include:
-          - preset: CPU
-          - preset: CUDA
-            container: nvidia/cuda:11.8.0-devel-ubuntu22.04
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
-          - preset: ROCm
-            container: rocm/dev-ubuntu-22.04:6.1.2
+          - container: nvidia/cuda:11.8.0-devel-ubuntu22.04
+            preset: CUDA
+          - container: rocm/dev-ubuntu-22.04:6.1.2
+            preset: ROCm
             extra-packages: rocm-libs
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_PREFIX_PATH=/opt/rocm'
-    runs-on: linux
+    runs-on: ubuntu-latest
     container: ${{ matrix.container }}
     steps:
       - uses: actions/checkout@v4
       - run: |
-          [ -n "${{ matrix.container }}" ] || sudo=sudo
-          $sudo apt-get update
-          $sudo apt-get install -y cmake ccache ${{ matrix.extra-packages }}
+          apt-get update
+          apt-get install -y cmake pkg-config ${{ matrix.extra-packages }}
         env:
           DEBIAN_FRONTEND: noninteractive
-      - uses: actions/cache@v4
-        with:
-          path: /github/home/.cache/ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
       - run: |
-          cmake --preset ${{ matrix.preset }} ${{ matrix.flags }}
+          cmake --preset ${{ matrix.preset }}
           cmake --build --preset ${{ matrix.preset }} --parallel
 
-  windows:
-    needs: [changes]
-    if: needs.changes.outputs.changed == 'True'
-    strategy:
-      matrix:
-        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
-            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'
-    runs-on: windows
-    steps:
-      - run: |
-          choco install -y --no-progress ccache ninja
-          ccache -o cache_dir=${{ github.workspace }}\.ccache
-      - if: matrix.preset == 'CUDA' || matrix.preset == 'ROCm'
-        id: cache-install
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
-            C:\Program Files\AMD\ROCm
-          key: ${{ matrix.install }}
-      - if: matrix.preset == 'CUDA'
-        name: Install CUDA ${{ matrix.cuda-version }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          if ("${{ 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
-          }
-
-          $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
-          echo "$cudaPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-      - if: matrix.preset == 'ROCm'
-        name: Install ROCm ${{ matrix.rocm-version }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
-            Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList '-install' -NoNewWindow -Wait
-          }
-
-          $hipPath = (Resolve-Path "C:\Program Files\AMD\ROCm\*").path
-          echo "$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
-      - if: ${{ !cancelled() && steps.cache-install.outputs.cache-hit != 'true' }}
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
-            C:\Program Files\AMD\ROCm
-          key: ${{ matrix.install }}
-      - uses: actions/checkout@v4
-      - uses: actions/cache@v4
-        with:
-          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'
-          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }}
-          cmake --build --parallel --preset "${{ matrix.preset }}"
-        env:
-          CMAKE_GENERATOR: Ninja
-
-  go_mod_tidy:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-      - name: check that 'go mod tidy' is clean
-        run: go mod tidy --diff || (echo "Please run 'go mod tidy'." && exit 1)
-
   test:
     strategy:
       matrix:
@@ -154,82 +69,15 @@ jobs:
     runs-on: ${{ matrix.os }}
     env:
       CGO_ENABLED: '1'
-      GOEXPERIMENT: 'synctest'
     steps:
-      - name: checkout
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # 4.2.2
-
-      - name: cache restore
-        uses: actions/cache/restore@1bd1e32a3bdc45362d1e726936510720a7c30a57 # v4.2.0
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
         with:
-          # Note: unlike the other setups, this is only grabbing the mod download
-          # cache, rather than the whole mod directory, as the download cache
-          # contains zips that can be unpacked in parallel faster than they can be
-          # fetched and extracted by tar
-          path: |
-            ~/.cache/go-build
-            ~/go/pkg/mod/cache
-            ~\AppData\Local\go-build
-          # NOTE: The -3- here should be incremented when the scheme of data to be
-          # cached changes (e.g. path above changes).
-          key: ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-${{ hashFiles('**/go.sum') }}-${{ github.run_id }}
-          restore-keys: |
-            ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-${{ hashFiles('**/go.sum') }}
-            ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-
-
-      - name: Setup Go
-        uses: actions/setup-go@v5
-        with:
-          # The caching strategy of setup-go is less than ideal, and wastes
-          # time by not saving artifacts due to small failures like the linter
-          # complaining, etc. This means subsequent have to rebuild their world
-          # again until all checks pass. For instance, if you mispell a word,
-          # you're punished until you fix it. This is more hostile than
-          # helpful.
-          cache: false
-
           go-version-file: go.mod
-
-      # It is tempting to run this in a platform independent way, but the past
-      # shows this codebase will see introductions of platform specific code
-      # generation, and so we need to check this per platform to ensure we
-      # don't abuse go generate on specific platforms.
-      - name: check that 'go generate' is clean
-        if: always()
-        run: |
-          go generate ./...
-          git diff --name-only --exit-code || (echo "Please run 'go generate ./...'." && exit 1)
-
-      - name: go test
-        if: always()
-        run: go test -count=1 -benchtime=1x ./...
-
-      # TODO(bmizerany): replace this heavy tool with just the
-      # tools/checks/binaries we want and then make them all run in parallel
-      # across jobs, not on a single tiny vm on Github Actions.
       - uses: golangci/golangci-lint-action@v6
         with:
           args: --timeout 10m0s -v
-
-      - name: cache save
-        # Always save the cache, even if the job fails. The artifacts produced
-        # during the building of test binaries are not all for naught. They can
-        # be used to speed up subsequent runs.
-        if: always()
-
-        uses: actions/cache/save@1bd1e32a3bdc45362d1e726936510720a7c30a57 # v4.2.0
-        with:
-          # Note: unlike the other setups, this is only grabbing the mod download
-          # cache, rather than the whole mod directory, as the download cache
-          # contains zips that can be unpacked in parallel faster than they can be
-          # fetched and extracted by tar
-          path: |
-            ~/.cache/go-build
-            ~/go/pkg/mod/cache
-            ~\AppData\Local\go-build
-          # NOTE: The -3- here should be incremented when the scheme of data to be
-          # cached changes (e.g. path above changes).
-          key: ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-${{ hashFiles('**/go.sum') }}-${{ github.run_id }}
+      - run: go test ./...
 
   patches:
     runs-on: ubuntu-latest
@@ -237,5 +85,5 @@ jobs:
       - uses: actions/checkout@v4
       - name: Verify patches apply cleanly and do not change files
         run: |
-          make -f Makefile.sync clean checkout apply-patches sync
-          git diff --compact-summary --exit-code
+          make -f Makefile2 clean checkout sync
+          git diff --compact-summary --exit-code

.gitignore

@@ -4,13 +4,12 @@
 .venv
 .swp
 dist
-build
+ollama
 .cache
 *.exe
 .idea
 test_data
 *.crt
-__debug_bin*
 llama/build
-llama/vendor
-/ollama
+__debug_bin*
+llama/vendor

.golangci.yaml

@@ -6,6 +6,8 @@ linters:
     - bidichk
     - bodyclose
     - containedctx
+    - contextcheck
+    - errcheck
     - gocheckcompilerdirectives
     - gofmt
     - gofumpt
@@ -21,11 +23,10 @@ linters:
     - staticcheck
     - tenv
     - unconvert
+    - unused
+    - usestdlibvars
     - wastedassign
     - whitespace
-  disable:
-    - usestdlibvars
-    - errcheck
 linters-settings:
   staticcheck:
     checks:
@@ -38,4 +39,5 @@ severity:
         - gofmt
         - goimports
         - intrange
+        - usestdlibvars
       severity: info

CMakeLists.txt

@@ -19,32 +19,11 @@ set(GGML_CCACHE ON)
 set(GGML_BACKEND_DL ON)
 set(GGML_BACKEND_SHARED ON)
 set(GGML_SCHED_MAX_COPIES 4)
-
-set(GGML_LLAMAFILE ON)
+set(GGML_CPU_ALL_VARIANTS ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
-set(GGML_CUDA_GRAPHS ON)
-set(GGML_CUDA_FA ON)
-set(GGML_CUDA_COMPRESSION_MODE default)
+set(GGML_LLAMAFILE ON)
 
-if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
-    OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
-    set(GGML_CPU_ALL_VARIANTS ON)
-endif()
-
-if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
-    set(CMAKE_BUILD_RPATH "@loader_path")
-    set(CMAKE_INSTALL_RPATH "@loader_path")
-endif()
-
-set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama)
-
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include)
@@ -55,79 +34,12 @@ set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)
 
-get_target_property(CPU_VARIANTS ggml-cpu MANUALLY_ADDED_DEPENDENCIES)
-if(NOT CPU_VARIANTS)
-    set(CPU_VARIANTS "ggml-cpu")
-endif()
-
-install(TARGETS ggml-base ${CPU_VARIANTS}
-    RUNTIME_DEPENDENCIES
-        PRE_EXCLUDE_REGEXES ".*"
-    RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-    FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
-)
-
 check_language(CUDA)
 if(CMAKE_CUDA_COMPILER)
-    if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24" AND NOT CMAKE_CUDA_ARCHITECTURES)
-        set(CMAKE_CUDA_ARCHITECTURES "native")
-    endif()
-
-    find_package(CUDAToolkit)
     add_subdirectory(${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
-    )
 endif()
 
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a|1200|1201):xnack[+-]$"
-    CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a|1200|1201):xnack[+-]$\"."
-)
-
 check_language(HIP)
 if(CMAKE_HIP_COMPILER)
-    set(HIP_PLATFORM "amd")
-
-    find_package(hip REQUIRED)
-    if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
-    elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
-        list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
-    endif()
-
-    if(AMDGPU_TARGETS)
-        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
-
-        if (WIN32)
-            target_compile_definitions(ggml-hip PRIVATE GGML_CUDA_NO_PEER_COPY)
-        endif()
-
-        target_compile_definitions(ggml-hip PRIVATE GGML_HIP_NO_VMM)
-
-        set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
-        install(TARGETS ggml-hip
-            RUNTIME_DEPENDENCIES
-                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 ".*"
-                POST_EXCLUDE_REGEXES "system32"
-            RUNTIME DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
-            LIBRARY DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP
-        )
-
-        foreach(HIP_LIB_BIN_INSTALL_DIR IN ITEMS ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR})
-            if(EXISTS ${HIP_LIB_BIN_INSTALL_DIR}/rocblas)
-                install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_HIP_INSTALL_DIR} COMPONENT HIP)
-                break()
-            endif()
-        endforeach()
-    endif()
+    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
 endif()

CMakePresets.json

@@ -4,15 +4,10 @@
     {
       "name": "Default",
       "binaryDir": "${sourceDir}/build",
-      "installDir": "${sourceDir}/dist",
       "cacheVariables": {
         "CMAKE_BUILD_TYPE": "Release"
       }
     },
-    {
-      "name": "CPU",
-      "inherits": [ "Default" ]
-    },
     {
       "name": "CUDA",
       "inherits": [ "Default" ]
@@ -21,14 +16,14 @@
       "name": "CUDA 11",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86"
+        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;62;70;72;75;80;86"
       }
     },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120"
+        "CMAKE_CUDA_ARCHITECTURES": "60;61;62;70;72;75;80;86;87;89;90;90a"
       }
     },
     {
@@ -47,29 +42,20 @@
     },
     {
       "name": "ROCm",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "CMAKE_HIP_PLATFORM": "amd"
-      }
+      "inherits": [ "Default" ]
     },
     {
       "name": "ROCm 6",
       "inherits": [ "ROCm" ],
       "cacheVariables": {
-        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
+        "CMAKE_HIP_ARCHITECTURES": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
       }
     }
   ],
   "buildPresets": [
     {
       "name": "Default",
-      "configurePreset": "Default",
-      "configuration": "Release"
-    },
-    {
-      "name": "CPU",
-      "configurePreset": "Default",
-      "targets": [ "ggml-cpu" ]
+      "configurePreset": "Default"
     },
     {
       "name": "CUDA",

CONTRIBUTING.md

@@ -6,6 +6,8 @@ Thank you for your interest in contributing to Ollama! Here are a few guidelines
 
 See the [development documentation](./docs/development.md) for instructions on how to build and run Ollama locally.
 
+## Pull requests
+
 ### Ideal issues
 
 * [Bugs](https://github.com/ollama/ollama/issues?q=is%3Aissue+is%3Aopen+label%3Abug): issues where Ollama stops working or where it results in an unexpected error.
@@ -24,64 +26,11 @@ See the [development documentation](./docs/development.md) for instructions on h
 * Changes that add significant friction to the user experience
 * Changes that create a large future maintenance burden for maintainers and contributors
 
-## Proposing a (non-trivial) change
+### Best practices
 
-> By "non-trivial", we mean a change that is not a bug fix or small
-> documentation update. If you are unsure, please ask us on our [Discord
-> server](https://discord.gg/ollama).
-
-Before opening a non-trivial Pull Request, please open an issue to discuss the change and
-get feedback from the maintainers. This helps us understand the context of the
-change and how it fits into Ollama's roadmap and prevents us from duplicating
-work or you from spending time on a change that we may not be able to accept.
-
-Tips for proposals:
-
-* Explain the problem you are trying to solve, not what you are trying to do.
-* Explain why the change is important.
-* Explain how the change will be used.
-* Explain how the change will be tested.
-
-Additionally, for bonus points: Provide draft documentation you would expect to
-see if the change were accepted.
-
-## Pull requests
-
-**Commit messages**
-
-The title should look like:
-
-    <package>: <short description>
-
-The package is the most affected Go package. If the change does not affect Go
-code, then use the directory name instead. Changes to a single well-known
-file in the root directory may use the file name.
-
-The short description should start with a lowercase letter and be a
-continuation of the sentence:
-
-      "This changes Ollama to..."
-
-Examples:
-
-      llm/backend/mlx: support the llama architecture
-      CONTRIBUTING: provide clairity on good commit messages, and bad
-
-Bad Examples:
-
-      feat: add more emoji
-      fix: was not using famous web framework
-      chore: generify code
-
-**Tests**
-
-Please include tests. Strive to test behavior, not implementation.
-
-**New dependencies**
-
-Dependencies should be added sparingly. If you are adding a new dependency,
-please explain why it is necessary and what other ways you attempted that
-did not work without it.
+* Commit messages: please leave both a title and a description in your commit messages. The title should be a short summary of the changes, with a leading word that explains the section of the code being changed (e.g. `api: fix parsing of prompt field`) . In the description, leave a short 2-3 sentences that explain more about the change and its impact.
+* Tests: please add test coverage to changes where possible.
+* Minimize dependencies: avoid adding new dependencies unless absolutely necessary.
 
 ## Need help?
 

Dockerfile

@@ -1,131 +1,201 @@
-# vim: filetype=dockerfile
+ARG GOLANG_VERSION=1.22.8
+ARG CUDA_VERSION_11=11.3.1
+ARG CUDA_VERSION_12=12.4.0
+ARG ROCM_VERSION=6.1.2
+ARG JETPACK_6=r36.2.0
+ARG JETPACK_5=r35.4.1
 
-ARG FLAVOR=${TARGETARCH}
+### To create a local image for building linux binaries on mac or windows with efficient incremental builds
+#
+# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile --target unified-builder-amd64 .
+# docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
+#
+### Then incremental builds will be much faster in this container
+#
+# make -j 10 dist
+#
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
+ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
+COPY ./scripts/rh_linux_deps.sh /
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+RUN GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+# TODO intel oneapi goes here...
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
 
-ARG ROCMVERSION=6.3.3
-ARG JETPACK5VERSION=r35.4.1
-ARG JETPACK6VERSION=r36.4.0
-ARG CMAKEVERSION=3.31.2
+### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
+# Note: this does not contain jetson variants
+#
+# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile --target unified-builder-arm64 .
+# docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
+#
+FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
+ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
+COPY ./scripts/rh_linux_deps.sh /
+RUN GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
+    dnf config-manager --set-enabled appstream && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
+ENV GOARCH arm64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
 
-# CUDA v11 requires gcc v10.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
-FROM --platform=linux/amd64 rocm/dev-almalinux-8:${ROCMVERSION}-complete AS base-amd64
-RUN yum install -y yum-utils \
-    && yum-config-manager --add-repo https://dl.rockylinux.org/vault/rocky/8.5/AppStream/\$basearch/os/ \
-    && rpm --import https://dl.rockylinux.org/pub/rocky/RPM-GPG-KEY-Rocky-8 \
-    && dnf install -y yum-utils ccache gcc-toolset-10-gcc-10.2.1-8.2.el8 gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 gcc-toolset-10-binutils-2.35-11.el8 \
-    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-
-FROM --platform=linux/arm64 almalinux:8 AS base-arm64
-# install epel-release for ccache
-RUN yum install -y yum-utils epel-release \
-    && dnf install -y clang ccache \
-    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
-ENV CC=clang CXX=clang++
-
-FROM base-${TARGETARCH} AS base
-ARG CMAKEVERSION
-RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-ENV LDFLAGS=-s
-
-FROM base AS cpu
-RUN dnf install -y gcc-toolset-11-gcc gcc-toolset-11-gcc-c++
-ENV PATH=/opt/rh/gcc-toolset-11/root/usr/bin:$PATH
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CPU' \
-        && cmake --build --parallel --preset 'CPU' \
-        && cmake --install build --component CPU --strip --parallel 8
-
-FROM base AS cuda-11
-ARG CUDA11VERSION=11.3
-RUN dnf install -y cuda-toolkit-${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//./-}
-ENV PATH=/usr/local/cuda-12/bin:$PATH
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'CUDA 12' \
-        && cmake --build --parallel --preset 'CUDA 12' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
-FROM base AS rocm-6
-ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'ROCm 6' \
-        && cmake --build --parallel --preset 'ROCm 6' \
-        && cmake --install build --component HIP --strip --parallel 8
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK5VERSION} AS jetpack-5
-ARG CMAKEVERSION
-RUN apt-get update && apt-get install -y curl ccache \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 5' \
-        && cmake --build --parallel --preset 'JetPack 5' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
-FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK6VERSION} AS jetpack-6
-ARG CMAKEVERSION
-RUN apt-get update && apt-get install -y curl ccache \
-    && curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKEVERSION}/cmake-${CMAKEVERSION}-linux-$(uname -m).tar.gz | tar xz -C /usr/local --strip-components 1
-COPY CMakeLists.txt CMakePresets.json .
-COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
-RUN --mount=type=cache,target=/root/.ccache \
-    cmake --preset 'JetPack 6' \
-        && cmake --build --parallel --preset 'JetPack 6' \
-        && cmake --install build --component CUDA --strip --parallel 8
-
-FROM base AS build
-WORKDIR /go/src/github.com/ollama/ollama
-COPY go.mod go.sum .
-RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-$(case $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
-ENV PATH=/usr/local/go/bin:$PATH
-RUN go mod download
+FROM --platform=linux/amd64 unified-builder-amd64 AS build-amd64
 COPY . .
-ARG GOFLAGS="'-ldflags=-w -s'"
-ENV CGO_ENABLED=1
-RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -trimpath -buildmode=pie -o /bin/ollama .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_ROCM_GENERATE
+ARG OLLAMA_FAST_BUILD
+ARG VERSION
+ARG CUSTOM_CPU_FLAGS
+RUN --mount=type=cache,target=/root/.ccache \
+    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
+        make -j $(nproc) dist ; \
+    else \
+        make -j 5 dist ; \
+    fi
+RUN cd dist/linux-$GOARCH && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
+    cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
+    fi
 
-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
+# Jetsons need to be built in discrete stages
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_5} AS runners-jetpack5-arm64
+ARG GOLANG_VERSION
+RUN apt-get update && apt-get install -y git curl ccache && \
+    curl -s -L https://dl.google.com/go/go${GOLANG_VERSION}.linux-arm64.tar.gz | tar xz -C /usr/local && \
+    ln -s /usr/local/go/bin/go /usr/local/bin/go && \
+    ln -s /usr/local/go/bin/gofmt /usr/local/bin/gofmt && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+WORKDIR /go/src/github.com/ollama/ollama/
+COPY . .
+ARG CGO_CFLAGS
+ENV GOARCH arm64
+ARG VERSION
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 dist_cuda_v11 \
+        CUDA_ARCHITECTURES="72;87" \
+        GPU_RUNNER_VARIANT=_jetpack5 \
+        DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama \
+        DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama/cuda_jetpack5
 
-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
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_6} AS runners-jetpack6-arm64
+ARG GOLANG_VERSION
+RUN apt-get update && apt-get install -y git curl ccache && \
+    curl -s -L https://dl.google.com/go/go${GOLANG_VERSION}.linux-arm64.tar.gz | tar xz -C /usr/local && \
+    ln -s /usr/local/go/bin/go /usr/local/bin/go && \
+    ln -s /usr/local/go/bin/gofmt /usr/local/bin/gofmt && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+WORKDIR /go/src/github.com/ollama/ollama/
+COPY . .
+ARG CGO_CFLAGS
+ENV GOARCH arm64
+ARG VERSION
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 dist_cuda_v12 \
+        CUDA_ARCHITECTURES="87" \
+        GPU_RUNNER_VARIANT=_jetpack6 \
+        DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama \
+        DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama/cuda_jetpack6
 
-FROM scratch AS rocm
-COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm
+FROM --platform=linux/arm64 unified-builder-arm64 AS build-arm64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_FAST_BUILD
+ARG VERSION
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 dist
+COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+RUN cd dist/linux-$GOARCH && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN cd dist/linux-$GOARCH-jetpack5 && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack5.tgz
+RUN cd dist/linux-$GOARCH-jetpack6 && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack6.tgz
 
-FROM ${FLAVOR} AS archive
-COPY --from=cpu dist/lib/ollama /lib/ollama
-COPY --from=build /bin/ollama /bin/ollama
+FROM --platform=linux/amd64 scratch AS dist-amd64
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM --platform=linux/arm64 scratch AS dist-arm64
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM dist-$TARGETARCH AS dist
 
-FROM ubuntu:20.04
-RUN apt-get update \
-    && apt-get install -y ca-certificates \
-    && apt-get clean \
-    && rm -rf /var/lib/apt/lists/*
-COPY --from=archive /bin /usr/bin
+
+# For amd64 container images, filter out cuda/rocm to minimize size
+FROM build-amd64 AS runners-cuda-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
+    ./dist/linux-amd64/lib/ollama/runners/rocm*
+
+FROM build-amd64 AS runners-rocm-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
+    ./dist/linux-amd64/lib/ollama/libcu*.so* \
+    ./dist/linux-amd64/lib/ollama/runners/cuda*
+
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ /lib/
+COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ /lib/
+
+
+# ROCm libraries larger so we keep it distinct from the CPU/CUDA image
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
+# Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
+# across releases
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
+
+ENTRYPOINT ["/bin/ollama"]
+CMD ["serve"]
+
+FROM runtime-$TARGETARCH
+EXPOSE 11434
+ENV OLLAMA_HOST 0.0.0.0
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
-COPY --from=archive /lib/ollama /usr/lib/ollama
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
 ENV NVIDIA_VISIBLE_DEVICES=all
-ENV OLLAMA_HOST=0.0.0.0:11434
-EXPOSE 11434
+
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]

Dockerfile2

@@ -0,0 +1,66 @@
+ARG CUDA_11_VERSION=11.3
+ARG CUDA_12_VERSION=12.4
+ARG ROCM_VERSION=6.1.2
+ARG JETPACK_5_VERSION=r35.4.1
+ARG JETPACK_6_VERSION=r36.2.0
+ARG CMAKE_VERSION=3.31.2
+
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS base
+ARG CMAKE_VERSION
+RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-x86_64.tar.gz | tar xz -C /usr --strip-components 1
+RUN sed -i -e 's/mirror.centos.org/vault.centos.org/g' -e 's/^#.*baseurl=http/baseurl=http/g' -e 's/^mirrorlist=http/#mirrorlist=http/g' /etc/yum.repos.d/*.repo \
+    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
+
+# FROM --platform=linux/arm64 rockylinux:8 AS base
+# ARG CMAKE_VERSION
+# RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-aarch64.tar.gz | tar xz -C /usr --strip-components 1
+# RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
+
+FROM base AS amd64
+ARG CUDA_11_VERSION
+ARG CUDA_12_VERSION
+RUN yum install -y cuda-toolkit-${CUDA_11_VERSION//./-} \
+    && yum install -y cuda-toolkit-${CUDA_12_VERSION//./-}
+COPY CMakeLists.txt CMakeLists.txt
+COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+
+FROM --platform=linux/amd64 amd64 AS cuda_11
+ENV PATH=/usr/local/cuda-${CUDA_11_VERSION}/bin:$PATH
+RUN cmake -S . -B build -DCMAKE_CUDA_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+RUN cmake --build build --target ggml-cuda -j
+
+FROM --platform=linux/amd64 amd64 AS cuda_12
+ENV PATH=/usr/local/cuda-${CUDA_12_VERSION}/bin:$PATH
+RUN cmake -S . -B build -DCMAKE_CUDA_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
+RUN cmake --build build --target ggml-cuda -j
+
+FROM --platform=linux/amd64 amd64 AS rocm
+RUN cmake -S . -B build -DCMAKE_HIP_ARCHITECTURES="gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
+RUN cmake --build build --target ggml-hip -j
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_5_VERSION} AS jetpack_5
+ARG CMAKE_VERSION
+RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-aarch64.tar.gz | tar xz -C /usr --strip-components 1
+COPY CMakeLists.txt .
+COPY ml/backend/ggml/ggml .
+RUN cmake -S . -B build \
+    -DCMAKE_CUDA_ARCHITECTURES="72;87"
+RUN cmake --build build --target ggml-cuda
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_6_VERSION} AS jetpack_6
+ARG CMAKE_VERSION
+RUN curl -fsSL https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-aarch64.tar.gz | tar xz -C /usr --strip-components 1
+COPY CMakeLists.txt .
+COPY ml/backend/ggml/ggml .
+RUN cmake -S . -B build \
+    -DCMAKE_CUDA_ARCHITECTURES="87"
+RUN cmake --build build --target ggml-cuda
+
+FROM --platform=linux/amd64 golang:1.23
+COPY --from=cuda_11 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-11.so
+COPY --from=cuda_12 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-12.so
+COPY --from=rocm build/ml/backend/ggml/ggml/src/ggml-hip/libggml-hip.so libggml-hip.so
+
+# FROM --platform=linux/arm64 golang:1.23
+# COPY --from=jetpack_5 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-jetpack-5.so
+# COPY --from=jetpack_6 build/ml/backend/ggml/ggml/src/ggml-cuda/libggml-cuda.so libggml-cuda-jetpack-6.so

Makefile.sync

@@ -1,60 +0,0 @@
-UPSTREAM=https://github.com/ggerganov/llama.cpp.git
-WORKDIR=llama/vendor
-FETCH_HEAD=2016f07bd106c73699ecbaace80f55db5ed95dac
-
-.PHONY: help
-help:
-	@echo "Available targets:"
-	@echo "    sync                 Sync with upstream repositories"
-	@echo "    checkout             Checkout upstream repository"
-	@echo "    apply-patches        Apply patches to local repository"
-	@echo "    format-patches       Format patches from local repository"
-	@echo "    clean                Clean local repository"
-	@echo
-	@echo "Example:"
-	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean sync"
-
-.PHONY: sync
-sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml
-
-.PHONY: llama/build-info.cpp
-llama/build-info.cpp: llama/build-info.cpp.in
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' $< > $@
-
-.PHONY: llama/llama.cpp
-llama/llama.cpp: llama/vendor/
-	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
-
-.PHONY: ml/backend/ggml/ggml
-ml/backend/ggml/ggml: llama/vendor/ggml/
-	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
-
-PATCHES=$(wildcard llama/patches/*.patch)
-
-.PHONY: apply-patches
-.NOTPARALLEL:
-apply-patches: $(addsuffix ed, $(PATCHES))
-
-%.patched: %.patch
-	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then touch $@; else git -C $(WORKDIR) am --abort; exit 1; fi
-
-.PHONY: checkout
-checkout: $(WORKDIR)
-	git -C $(WORKDIR) fetch
-	git -C $(WORKDIR) checkout -f $(FETCH_HEAD)
-
-$(WORKDIR):
-	git clone $(UPSTREAM) $(WORKDIR)
-
-.PHONE: format-patches
-format-patches: llama/patches
-	git -C $(WORKDIR) format-patch \
-		--no-signature \
-		--no-numbered \
-		--zero-commit \
-		-o $(realpath $<) \
-		$(FETCH_HEAD)
-
-.PHONE: clean
-clean: checkout
-	$(RM) $(addsuffix ed, $(PATCHES))

Makefile2

@@ -0,0 +1,46 @@
+UPSTREAM=https://github.com/ggerganov/llama.cpp.git
+WORKDIR=llama/vendor
+FETCH_HEAD=46e3556e01b824e52395fb050b29804b6cff2a7c
+
+all: sync
+
+.PHONY: sync
+sync: llama/llama.cpp ml/backend/ggml/ggml
+
+.PHONY: llama/llama.cpp
+llama/llama.cpp: llama/vendor/ apply_patches
+	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
+
+.PHONY: ml/backend/ggml/ggml apply_patches
+ml/backend/ggml/ggml: llama/vendor/ggml/ apply_patches
+	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
+
+PATCHES=$(wildcard llama/patches/*.patch)
+
+.PHONY: apply_patches
+.NOTPARALLEL:
+apply_patches: $(addsuffix ed, $(PATCHES))
+
+%.patched: %.patch
+	@if git -c user.name=nobody -c 'user.email=<>' -C $(WORKDIR) am -3 $(realpath $<); then touch $@; else git -C $(WORKDIR) am --abort; exit 1; fi
+
+.PHONY: checkout
+checkout: $(WORKDIR)
+	git -C $(WORKDIR) fetch
+	git -C $(WORKDIR) checkout -f $(FETCH_HEAD)
+
+$(WORKDIR):
+	git clone $(UPSTREAM) $(WORKDIR)
+
+.PHONE: format_patches
+format_patches: llama/patches
+	git -C $(WORKDIR) format-patch \
+		--no-signature \
+		--no-numbered \
+		--zero-commit \
+		-o $(realpath $<) \
+		$(FETCH_HEAD)
+
+.PHONE: clean
+clean: checkout
+	$(RM) $(addsuffix ed, $(PATCHES))

README.md

@@ -1,5 +1,5 @@
 <div align="center">
-  <a href="https://ollama.com">
+  <a href="https://ollama.com" />
     <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
   </a>
 </div>
@@ -18,7 +18,7 @@ Get up and running with large language models.
 
 ### Linux
 
-```shell
+```
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
@@ -42,7 +42,7 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 To run and chat with [Llama 3.2](https://ollama.com/library/llama3.2):
 
-```shell
+```
 ollama run llama3.2
 ```
 
@@ -54,13 +54,6 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                         |
 | ------------------ | ---------- | ----- | -------------------------------- |
-| Gemma 3            | 1B         | 815MB | `ollama run gemma3:1b`           |
-| Gemma 3            | 4B         | 3.3GB | `ollama run gemma3`              |
-| Gemma 3            | 12B        | 8.1GB | `ollama run gemma3:12b`          |
-| Gemma 3            | 27B        | 17GB  | `ollama run gemma3:27b`          |
-| QwQ                | 32B        | 20GB  | `ollama run qwq`                 |
-| DeepSeek-R1        | 7B         | 4.7GB | `ollama run deepseek-r1`         |
-| DeepSeek-R1        | 671B       | 404GB | `ollama run deepseek-r1:671b`    |
 | Llama 3.3          | 70B        | 43GB  | `ollama run llama3.3`            |
 | Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`            |
 | Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`         |
@@ -69,7 +62,10 @@ Here are some example models that can be downloaded:
 | Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`            |
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`       |
 | Phi 4              | 14B        | 9.1GB | `ollama run phi4`                |
-| Phi 4 Mini         | 3.8B       | 2.5GB | `ollama run phi4-mini`           |
+| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`                |
+| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`           |
+| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`              |
+| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`          |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`             |
 | Moondream 2        | 1.4B       | 829MB | `ollama run moondream`           |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`         |
@@ -77,7 +73,7 @@ Here are some example models that can be downloaded:
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`           |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored`   |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`               |
-| Granite-3.2         | 8B         | 4.9GB | `ollama run granite3.2`          |
+| Solar              | 10.7B      | 6.1GB | `ollama run solar`               |
 
 > [!NOTE]
 > You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -96,13 +92,13 @@ Ollama supports importing GGUF models in the Modelfile:
 
 2. Create the model in Ollama
 
-   ```shell
+   ```
    ollama create example -f Modelfile
    ```
 
 3. Run the model
 
-   ```shell
+   ```
    ollama run example
    ```
 
@@ -114,7 +110,7 @@ See the [guide](docs/import.md) on importing models for more information.
 
 Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.2` model:
 
-```shell
+```
 ollama pull llama3.2
 ```
 
@@ -149,13 +145,13 @@ For more information on working with a Modelfile, see the [Modelfile](docs/model
 
 `ollama create` is used to create a model from a Modelfile.
 
-```shell
+```
 ollama create mymodel -f ./Modelfile
 ```
 
 ### Pull a model
 
-```shell
+```
 ollama pull llama3.2
 ```
 
@@ -163,13 +159,13 @@ ollama pull llama3.2
 
 ### Remove a model
 
-```shell
+```
 ollama rm llama3.2
 ```
 
 ### Copy a model
 
-```shell
+```
 ollama cp llama3.2 my-model
 ```
 
@@ -188,39 +184,37 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 
 ```
 ollama run llava "What's in this image? /Users/jmorgan/Desktop/smile.png"
+The image features a yellow smiley face, which is likely the central focus of the picture.
 ```
 
-> **Output**: The image features a yellow smiley face, which is likely the central focus of the picture.
-
 ### Pass the prompt as an argument
 
-```shell
-ollama run llama3.2 "Summarize this file: $(cat README.md)"
 ```
-
-> **Output**: Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
+$ ollama run llama3.2 "Summarize this file: $(cat README.md)"
+ Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
+```
 
 ### Show model information
 
-```shell
+```
 ollama show llama3.2
 ```
 
 ### List models on your computer
 
-```shell
+```
 ollama list
 ```
 
 ### List which models are currently loaded
 
-```shell
+```
 ollama ps
 ```
 
 ### Stop a model which is currently running
 
-```shell
+```
 ollama stop llama3.2
 ```
 
@@ -236,13 +230,13 @@ See the [developer guide](https://github.com/ollama/ollama/blob/main/docs/develo
 
 Next, start the server:
 
-```shell
+```
 ./ollama serve
 ```
 
 Finally, in a separate shell, run a model:
 
-```shell
+```
 ./ollama run llama3.2
 ```
 
@@ -252,7 +246,7 @@ Ollama has a REST API for running and managing models.
 
 ### Generate a response
 
-```shell
+```
 curl http://localhost:11434/api/generate -d '{
   "model": "llama3.2",
   "prompt":"Why is the sky blue?"
@@ -261,7 +255,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ### Chat with a model
 
-```shell
+```
 curl http://localhost:11434/api/chat -d '{
   "model": "llama3.2",
   "messages": [
@@ -277,7 +271,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Web & Desktop
 
 - [Open WebUI](https://github.com/open-webui/open-webui)
-- [SwiftChat (macOS with ReactNative)](https://github.com/aws-samples/swift-chat)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
 - [Hollama](https://github.com/fmaclen/hollama)
 - [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
@@ -285,13 +278,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
 - [Saddle](https://github.com/jikkuatwork/saddle)
-- [TagSpaces](https://www.tagspaces.org) (A platform for file based apps, [utilizing Ollama](https://docs.tagspaces.org/ai/) for the generation of tags and descriptions)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
 - [Chatbot UI v2](https://github.com/mckaywrigley/chatbot-ui)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
-- [big-AGI](https://github.com/enricoros/big-AGI)
+- [big-AGI](https://github.com/enricoros/big-AGI/blob/main/docs/config-local-ollama.md)
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)
@@ -325,7 +317,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [RWKV-Runner](https://github.com/josStorer/RWKV-Runner) (RWKV offline LLM deployment tool, also usable as a client for ChatGPT and Ollama)
 - [Ollama Grid Search](https://github.com/dezoito/ollama-grid-search) (app to evaluate and compare models)
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
-- [Casibase](https://casibase.org) (An open source AI knowledge base and dialogue system combining the latest RAG, SSO, ollama support and multiple large language models.)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
 - [Shinkai Desktop](https://github.com/dcSpark/shinkai-apps) (Two click install Local AI using Ollama + Files + RAG)
@@ -348,7 +339,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
 - [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
 - [PyOllaMx](https://github.com/kspviswa/pyOllaMx) - macOS application capable of chatting with both Ollama and Apple MLX models.
-- [Cline](https://github.com/cline/cline) - Formerly known as Claude Dev is a VSCode extension for multi-file/whole-repo coding
+- [Claude Dev](https://github.com/saoudrizwan/claude-dev) - VSCode extension for multi-file/whole-repo coding
 - [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 - [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
@@ -362,7 +353,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Web management](https://github.com/lemonit-eric-mao/ollama-web-management) (Web management page)
 - [Promptery](https://github.com/promptery/promptery) (desktop client for Ollama.)
 - [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
-- [chat-ollama](https://github.com/annilq/chat-ollama) (a React Native client for Ollama)
 - [SpaceLlama](https://github.com/tcsenpai/spacellama) (Firefox and Chrome extension to quickly summarize web pages with ollama in a sidebar)
 - [YouLama](https://github.com/tcsenpai/youlama) (Webapp to quickly summarize any YouTube video, supporting Invidious as well)
 - [DualMind](https://github.com/tcsenpai/dualmind) (Experimental app allowing two models to talk to each other in the terminal or in a web interface)
@@ -379,26 +369,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Minima](https://github.com/dmayboroda/minima) (RAG with on-premises or fully local workflow)
 - [aidful-ollama-model-delete](https://github.com/AidfulAI/aidful-ollama-model-delete) (User interface for simplified model cleanup)
 - [Perplexica](https://github.com/ItzCrazyKns/Perplexica) (An AI-powered search engine & an open-source alternative to Perplexity AI)
-- [Ollama Chat WebUI for Docker ](https://github.com/oslook/ollama-webui) (Support for local docker deployment, lightweight ollama webui)
-- [AI Toolkit for Visual Studio Code](https://aka.ms/ai-tooklit/ollama-docs) (Microsoft-official VSCode extension to chat, test, evaluate models with Ollama support, and use them in your AI applications.)
-- [MinimalNextOllamaChat](https://github.com/anilkay/MinimalNextOllamaChat) (Minimal Web UI for Chat and Model Control)
-- [Chipper](https://github.com/TilmanGriesel/chipper) AI interface for tinkerers (Ollama, Haystack RAG, Python)
-- [ChibiChat](https://github.com/CosmicEventHorizon/ChibiChat) (Kotlin-based Android app to chat with Ollama and Koboldcpp API endpoints)
-- [LocalLLM](https://github.com/qusaismael/localllm) (Minimal Web-App to run ollama models on it with a GUI)
-- [Ollamazing](https://github.com/buiducnhat/ollamazing) (Web extension to run Ollama models)
-- [OpenDeepResearcher-via-searxng](https://github.com/benhaotang/OpenDeepResearcher-via-searxng) (A Deep Research equivent endpoint with Ollama support for running locally)
-- [AntSK](https://github.com/AIDotNet/AntSK) (Out-of-the-box & Adaptable RAG Chatbot)
-- [MaxKB](https://github.com/1Panel-dev/MaxKB/) (Ready-to-use & flexible RAG Chatbot)
-- [yla](https://github.com/danielekp/yla) (Web interface to freely interact with your customized models)
-- [LangBot](https://github.com/RockChinQ/LangBot) (LLM-based instant messaging bots platform, with Agents, RAG features, supports multiple platforms)
-- [1Panel](https://github.com/1Panel-dev/1Panel/) (Web-based Linux Server Management Tool)
-- [AstrBot](https://github.com/Soulter/AstrBot/) (User-friendly LLM-based multi-platform chatbot with a WebUI, supporting RAG, LLM agents, and plugins integration)
-- [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
-- [Ellama](https://github.com/zeozeozeo/ellama) (Friendly native app to chat with an Ollama instance)
-- [screenpipe](https://github.com/mediar-ai/screenpipe) Build agents powered by your screen history
-- [Ollamb](https://github.com/hengkysteen/ollamb) (Simple yet rich in features, cross-platform built with Flutter and designed for Ollama. Try the [web demo](https://hengkysteen.github.io/demo/ollamb/).)
-- [Writeopia](https://github.com/Writeopia/Writeopia) (Text editor with integration with Ollama)
-- [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 
 ### Cloud
 
@@ -438,14 +408,10 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [SwollamaCLI](https://github.com/marcusziade/Swollama) bundled with the Swollama Swift package. [Demo](https://github.com/marcusziade/Swollama?tab=readme-ov-file#cli-usage)
 - [aichat](https://github.com/sigoden/aichat) All-in-one LLM CLI tool featuring Shell Assistant, Chat-REPL, RAG, AI tools & agents, with access to OpenAI, Claude, Gemini, Ollama, Groq, and more.
 - [PowershAI](https://github.com/rrg92/powershai) PowerShell module that brings AI to terminal on Windows, including support for Ollama
-- [DeepShell](https://github.com/Abyss-c0re/deepshell) Your self-hosted AI assistant. Interactive Shell, Files and Folders analysis.
 - [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.
-- [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull and download models from Ollama Registry in your terminal.
-- [GGUF-to-Ollama](https://github.com/jonathanhecl/gguf-to-ollama) - Importing GGUF to Ollama made easy (multiplatform)
 
 ### Apple Vision Pro
 
-- [SwiftChat](https://github.com/aws-samples/swift-chat) (Cross-platform AI chat app supporting Apple Vision Pro via "Designed for iPad")
 - [Enchanted](https://github.com/AugustDev/enchanted)
 
 ### Database
@@ -460,10 +426,9 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
 - [Gentoo](https://github.com/gentoo/guru/tree/master/app-misc/ollama)
-- [Homebrew](https://formulae.brew.sh/formula/ollama)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
 - [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
-- [Nix package](https://search.nixos.org/packages?show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
+- [Nix package](https://search.nixos.org/packages?channel=24.05&show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
 - [Flox](https://flox.dev/blog/ollama-part-one)
 
 ### Libraries
@@ -516,21 +481,13 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [GoLamify](https://github.com/prasad89/golamify)
 - [Ollama for Haskell](https://github.com/tusharad/ollama-haskell)
 - [multi-llm-ts](https://github.com/nbonamy/multi-llm-ts) (A Typescript/JavaScript library allowing access to different LLM in unified API)
-- [LlmTornado](https://github.com/lofcz/llmtornado) (C# library providing a unified interface for major FOSS & Commercial inference APIs)
-- [Ollama for Zig](https://github.com/dravenk/ollama-zig)
-- [Abso](https://github.com/lunary-ai/abso) (OpenAI-compatible TypeScript SDK for any LLM provider)
-- [Nichey](https://github.com/goodreasonai/nichey) is a Python package for generating custom wikis for your research topic
-- [Ollama for D](https://github.com/kassane/ollama-d)
 
 ### Mobile
 
-- [SwiftChat](https://github.com/aws-samples/swift-chat) (Lightning-fast Cross-platform AI chat app with native UI for Android, iOS and iPad)
 - [Enchanted](https://github.com/AugustDev/enchanted)
 - [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
 - [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
-- [Ollama Android Chat](https://github.com/sunshine0523/OllamaServer) (No need for Termux, start the Ollama service with one click on an Android device)
-- [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
 
 ### Extensions & Plugins
 
@@ -574,18 +531,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [TextCraft](https://github.com/suncloudsmoon/TextCraft) (Copilot in Word alternative using Ollama)
 - [Alfred Ollama](https://github.com/zeitlings/alfred-ollama) (Alfred Workflow)
 - [TextLLaMA](https://github.com/adarshM84/TextLLaMA) A Chrome Extension that helps you write emails, correct grammar, and translate into any language
-- [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)
 
 ### Supported backends
 
 - [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.
 
 ### Observability
-- [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native intergration to Ollama.
-- [Lunary](https://lunary.ai/docs/integrations/ollama) is the leading open-source LLM observability platform. It provides a variety of enterprise-grade features such as real-time analytics, prompt templates management, PII masking, and comprehensive agent tracing.
+
 - [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
-- [HoneyHive](https://docs.honeyhive.ai/integrations/ollama) is an AI observability and evaluation platform for AI agents. Use HoneyHive to evaluate agent performance, interrogate failures, and monitor quality in production.
-- [Langfuse](https://langfuse.com/docs/integrations/ollama) is an open source LLM observability platform that enables teams to collaboratively monitor, evaluate and debug AI applications.
-- [MLflow Tracing](https://mlflow.org/docs/latest/llms/tracing/index.html#automatic-tracing) is an open source LLM observability tool with a convenient API to log and visualize traces, making it easy to debug and evaluate GenAI applications.
+- [HoneyHive](https://docs.honeyhive.ai/integrations/ollama) is an AI observability and evaluation platform for AI agents. Use HoneyHive to evaluate agent performance, interrogate failures, and monitor quality in production. 

api/client.go

@@ -10,7 +10,7 @@
 // repository].
 //
 // [the API documentation]: https://github.com/ollama/ollama/blob/main/docs/api.md
-// [in the GitHub repository]: https://github.com/ollama/ollama/tree/main/api/examples
+// [in the GitHub repository]: https://github.com/ollama/ollama/tree/main/examples
 package api
 
 import (
@@ -132,7 +132,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 const maxBufferSize = 512 * format.KiloByte
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
-	var buf io.Reader
+	var buf *bytes.Buffer
 	if data != nil {
 		bts, err := json.Marshal(data)
 		if err != nil {

api/client_test.go

@@ -1,13 +1,6 @@
 package api
 
 import (
-	"context"
-	"encoding/json"
-	"fmt"
-	"net/http"
-	"net/http/httptest"
-	"net/url"
-	"strings"
 	"testing"
 )
 
@@ -50,206 +43,3 @@ func TestClientFromEnvironment(t *testing.T) {
 		})
 	}
 }
-
-// testError represents an internal error type with status code and message
-// this is used since the error response from the server is not a standard error struct
-type testError struct {
-	message    string
-	statusCode int
-}
-
-func (e testError) Error() string {
-	return e.message
-}
-
-func TestClientStream(t *testing.T) {
-	testCases := []struct {
-		name      string
-		responses []any
-		wantErr   string
-	}{
-		{
-			name: "immediate error response",
-			responses: []any{
-				testError{
-					message:    "test error message",
-					statusCode: http.StatusBadRequest,
-				},
-			},
-			wantErr: "test error message",
-		},
-		{
-			name: "error after successful chunks, ok response",
-			responses: []any{
-				ChatResponse{Message: Message{Content: "partial response 1"}},
-				ChatResponse{Message: Message{Content: "partial response 2"}},
-				testError{
-					message:    "mid-stream error",
-					statusCode: http.StatusOK,
-				},
-			},
-			wantErr: "mid-stream error",
-		},
-		{
-			name: "successful stream completion",
-			responses: []any{
-				ChatResponse{Message: Message{Content: "chunk 1"}},
-				ChatResponse{Message: Message{Content: "chunk 2"}},
-				ChatResponse{
-					Message:    Message{Content: "final chunk"},
-					Done:       true,
-					DoneReason: "stop",
-				},
-			},
-		},
-	}
-
-	for _, tc := range testCases {
-		t.Run(tc.name, func(t *testing.T) {
-			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-				flusher, ok := w.(http.Flusher)
-				if !ok {
-					t.Fatal("expected http.Flusher")
-				}
-
-				w.Header().Set("Content-Type", "application/x-ndjson")
-
-				for _, resp := range tc.responses {
-					if errResp, ok := resp.(testError); ok {
-						w.WriteHeader(errResp.statusCode)
-						err := json.NewEncoder(w).Encode(map[string]string{
-							"error": errResp.message,
-						})
-						if err != nil {
-							t.Fatal("failed to encode error response:", err)
-						}
-						return
-					}
-
-					if err := json.NewEncoder(w).Encode(resp); err != nil {
-						t.Fatalf("failed to encode response: %v", err)
-					}
-					flusher.Flush()
-				}
-			}))
-			defer ts.Close()
-
-			client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)
-
-			var receivedChunks []ChatResponse
-			err := client.stream(context.Background(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
-				var resp ChatResponse
-				if err := json.Unmarshal(chunk, &resp); err != nil {
-					return fmt.Errorf("failed to unmarshal chunk: %w", err)
-				}
-				receivedChunks = append(receivedChunks, resp)
-				return nil
-			})
-
-			if tc.wantErr != "" {
-				if err == nil {
-					t.Fatal("expected error but got nil")
-				}
-				if !strings.Contains(err.Error(), tc.wantErr) {
-					t.Errorf("expected error containing %q, got %v", tc.wantErr, err)
-				}
-				return
-			}
-			if err != nil {
-				t.Errorf("unexpected error: %v", err)
-			}
-		})
-	}
-}
-
-func TestClientDo(t *testing.T) {
-	testCases := []struct {
-		name     string
-		response any
-		wantErr  string
-	}{
-		{
-			name: "immediate error response",
-			response: testError{
-				message:    "test error message",
-				statusCode: http.StatusBadRequest,
-			},
-			wantErr: "test error message",
-		},
-		{
-			name: "server error response",
-			response: testError{
-				message:    "internal error",
-				statusCode: http.StatusInternalServerError,
-			},
-			wantErr: "internal error",
-		},
-		{
-			name: "successful response",
-			response: struct {
-				ID      string `json:"id"`
-				Success bool   `json:"success"`
-			}{
-				ID:      "msg_123",
-				Success: true,
-			},
-		},
-	}
-
-	for _, tc := range testCases {
-		t.Run(tc.name, func(t *testing.T) {
-			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-				if errResp, ok := tc.response.(testError); ok {
-					w.WriteHeader(errResp.statusCode)
-					err := json.NewEncoder(w).Encode(map[string]string{
-						"error": errResp.message,
-					})
-					if err != nil {
-						t.Fatal("failed to encode error response:", err)
-					}
-					return
-				}
-
-				w.Header().Set("Content-Type", "application/json")
-				if err := json.NewEncoder(w).Encode(tc.response); err != nil {
-					t.Fatalf("failed to encode response: %v", err)
-				}
-			}))
-			defer ts.Close()
-
-			client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)
-
-			var resp struct {
-				ID      string `json:"id"`
-				Success bool   `json:"success"`
-			}
-			err := client.do(context.Background(), http.MethodPost, "/v1/messages", nil, &resp)
-
-			if tc.wantErr != "" {
-				if err == nil {
-					t.Fatalf("got nil, want error %q", tc.wantErr)
-				}
-				if err.Error() != tc.wantErr {
-					t.Errorf("error message mismatch: got %q, want %q", err.Error(), tc.wantErr)
-				}
-				return
-			}
-
-			if err != nil {
-				t.Fatalf("got error %q, want nil", err)
-			}
-
-			if expectedResp, ok := tc.response.(struct {
-				ID      string `json:"id"`
-				Success bool   `json:"success"`
-			}); ok {
-				if resp.ID != expectedResp.ID {
-					t.Errorf("response ID mismatch: got %q, want %q", resp.ID, expectedResp.ID)
-				}
-				if resp.Success != expectedResp.Success {
-					t.Errorf("response Success mismatch: got %v, want %v", resp.Success, expectedResp.Success)
-				}
-			}
-		})
-	}
-}

api/examples/README.md

@@ -2,10 +2,9 @@
 
 Run the examples in this directory with:
 
-```shell
+```
 go run example_name/main.go
 ```
-
 ## Chat - Chat with a model
 - [chat/main.go](chat/main.go)
 

api/types.go

@@ -10,9 +10,6 @@ import (
 	"strconv"
 	"strings"
 	"time"
-
-	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/types/model"
 )
 
 // StatusError is an error with an HTTP status code and message.
@@ -76,13 +73,13 @@ type GenerateRequest struct {
 	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
-	// Images is an optional list of raw image bytes accompanying this
+	// Images is an optional list of base64-encoded images accompanying this
 	// request, for multimodal models.
 	Images []ImageData `json:"images,omitempty"`
 
 	// 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"`
+	Options map[string]interface{} `json:"options"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -107,7 +104,7 @@ type ChatRequest struct {
 	Tools `json:"tools,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]any `json:"options"`
+	Options map[string]interface{} `json:"options"`
 }
 
 type Tools []Tool
@@ -163,65 +160,19 @@ func (t *ToolCallFunctionArguments) String() string {
 
 type Tool struct {
 	Type     string       `json:"type"`
-	Items    any          `json:"items,omitempty"`
 	Function ToolFunction `json:"function"`
 }
 
-// PropertyType can be either a string or an array of strings
-type PropertyType []string
-
-// UnmarshalJSON implements the json.Unmarshaler interface
-func (pt *PropertyType) UnmarshalJSON(data []byte) error {
-	// Try to unmarshal as a string first
-	var s string
-	if err := json.Unmarshal(data, &s); err == nil {
-		*pt = []string{s}
-		return nil
-	}
-
-	// If that fails, try to unmarshal as an array of strings
-	var a []string
-	if err := json.Unmarshal(data, &a); err != nil {
-		return err
-	}
-	*pt = a
-	return nil
-}
-
-// MarshalJSON implements the json.Marshaler interface
-func (pt PropertyType) MarshalJSON() ([]byte, error) {
-	if len(pt) == 1 {
-		// If there's only one type, marshal as a string
-		return json.Marshal(pt[0])
-	}
-	// Otherwise marshal as an array
-	return json.Marshal([]string(pt))
-}
-
-// String returns a string representation of the PropertyType
-func (pt PropertyType) String() string {
-	if len(pt) == 0 {
-		return ""
-	}
-	if len(pt) == 1 {
-		return pt[0]
-	}
-	return fmt.Sprintf("%v", []string(pt))
-}
-
 type ToolFunction struct {
 	Name        string `json:"name"`
 	Description string `json:"description"`
 	Parameters  struct {
 		Type       string   `json:"type"`
-		Defs       any      `json:"$defs,omitempty"`
-		Items      any      `json:"items,omitempty"`
 		Required   []string `json:"required"`
 		Properties map[string]struct {
-			Type        PropertyType `json:"type"`
-			Items       any          `json:"items,omitempty"`
-			Description string       `json:"description"`
-			Enum        []any        `json:"enum,omitempty"`
+			Type        string   `json:"type"`
+			Description string   `json:"description"`
+			Enum        []string `json:"enum,omitempty"`
 		} `json:"properties"`
 	} `json:"parameters"`
 }
@@ -307,7 +258,7 @@ type EmbedRequest struct {
 	Truncate *bool `json:"truncate,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]any `json:"options"`
+	Options map[string]interface{} `json:"options"`
 }
 
 // EmbedResponse is the response from [Client.Embed].
@@ -333,7 +284,7 @@ type EmbeddingRequest struct {
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]any `json:"options"`
+	Options map[string]interface{} `json:"options"`
 }
 
 // EmbeddingResponse is the response from [Client.Embeddings].
@@ -379,7 +330,7 @@ type ShowRequest struct {
 	Template string `json:"template"`
 	Verbose  bool   `json:"verbose"`
 
-	Options map[string]any `json:"options"`
+	Options map[string]interface{} `json:"options"`
 
 	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
@@ -387,18 +338,16 @@ type ShowRequest struct {
 
 // ShowResponse is the response returned from [Client.Show].
 type ShowResponse struct {
-	License       string             `json:"license,omitempty"`
-	Modelfile     string             `json:"modelfile,omitempty"`
-	Parameters    string             `json:"parameters,omitempty"`
-	Template      string             `json:"template,omitempty"`
-	System        string             `json:"system,omitempty"`
-	Details       ModelDetails       `json:"details,omitempty"`
-	Messages      []Message          `json:"messages,omitempty"`
-	ModelInfo     map[string]any     `json:"model_info,omitempty"`
-	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
-	Tensors       []Tensor           `json:"tensors,omitempty"`
-	Capabilities  []model.Capability `json:"capabilities,omitempty"`
-	ModifiedAt    time.Time          `json:"modified_at,omitempty"`
+	License       string         `json:"license,omitempty"`
+	Modelfile     string         `json:"modelfile,omitempty"`
+	Parameters    string         `json:"parameters,omitempty"`
+	Template      string         `json:"template,omitempty"`
+	System        string         `json:"system,omitempty"`
+	Details       ModelDetails   `json:"details,omitempty"`
+	Messages      []Message      `json:"messages,omitempty"`
+	ModelInfo     map[string]any `json:"model_info,omitempty"`
+	ProjectorInfo map[string]any `json:"projector_info,omitempty"`
+	ModifiedAt    time.Time      `json:"modified_at,omitempty"`
 }
 
 // CopyRequest is the request passed to [Client.Copy].
@@ -410,9 +359,9 @@ type CopyRequest struct {
 // PullRequest is the request passed to [Client.Pull].
 type PullRequest struct {
 	Model    string `json:"model"`
-	Insecure bool   `json:"insecure,omitempty"` // Deprecated: ignored
-	Username string `json:"username"`           // Deprecated: ignored
-	Password string `json:"password"`           // Deprecated: ignored
+	Insecure bool   `json:"insecure,omitempty"`
+	Username string `json:"username"`
+	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`
 
 	// Deprecated: set the model name with Model instead
@@ -516,13 +465,6 @@ type ModelDetails struct {
 	QuantizationLevel string   `json:"quantization_level"`
 }
 
-// Tensor describes the metadata for a given tensor.
-type Tensor struct {
-	Name  string   `json:"name"`
-	Type  string   `json:"type"`
-	Shape []uint64 `json:"shape"`
-}
-
 func (m *Metrics) Summary() {
 	if m.TotalDuration > 0 {
 		fmt.Fprintf(os.Stderr, "total duration:       %v\n", m.TotalDuration)
@@ -551,7 +493,7 @@ func (m *Metrics) Summary() {
 	}
 }
 
-func (opts *Options) FromMap(m map[string]any) error {
+func (opts *Options) FromMap(m map[string]interface{}) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct
 
@@ -608,12 +550,12 @@ func (opts *Options) FromMap(m map[string]any) error {
 				}
 				field.SetString(val)
 			case reflect.Slice:
-				// JSON unmarshals to []any, not []string
-				val, ok := val.([]any)
+				// JSON unmarshals to []interface{}, not []string
+				val, ok := val.([]interface{})
 				if !ok {
 					return fmt.Errorf("option %q must be of type array", key)
 				}
-				// convert []any to []string
+				// convert []interface{} to []string
 				slice := make([]string, len(val))
 				for i, item := range val {
 					str, ok := item.(string)
@@ -667,7 +609,7 @@ func DefaultOptions() Options {
 
 		Runner: Runner{
 			// options set when the model is loaded
-			NumCtx:    int(envconfig.ContextLength()),
+			NumCtx:    2048,
 			NumBatch:  512,
 			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
 			NumThread: 0,  // let the runtime decide
@@ -720,7 +662,7 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 }
 
 // FormatParams converts specified parameter options to their correct types
-func FormatParams(params map[string][]string) (map[string]any, error) {
+func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 	opts := Options{}
 	valueOpts := reflect.ValueOf(&opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts)           // types of the fields in the options struct
@@ -734,7 +676,7 @@ func FormatParams(params map[string][]string) (map[string]any, error) {
 		}
 	}
 
-	out := make(map[string]any)
+	out := make(map[string]interface{})
 	// iterate params and set values based on json struct tags
 	for key, vals := range params {
 		if opt, ok := jsonOpts[key]; !ok {

api/types_test.go

@@ -134,7 +134,7 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
-			var oMap map[string]any
+			var oMap map[string]interface{}
 			err := json.Unmarshal([]byte(test.req), &oMap)
 			require.NoError(t, err)
 			opts := DefaultOptions()
@@ -231,144 +231,3 @@ func TestMessage_UnmarshalJSON(t *testing.T) {
 		}
 	}
 }
-
-func TestToolFunction_UnmarshalJSON(t *testing.T) {
-	tests := []struct {
-		name    string
-		input   string
-		wantErr string
-	}{
-		{
-			name: "valid enum with same types",
-			input: `{
-				"name": "test",
-				"description": "test function",
-				"parameters": {
-					"type": "object",
-					"required": ["test"],
-					"properties": {
-						"test": {
-							"type": "string",
-							"description": "test prop",
-							"enum": ["a", "b", "c"]
-						}
-					}
-				}
-			}`,
-			wantErr: "",
-		},
-		{
-			name: "empty enum array",
-			input: `{
-				"name": "test",
-				"description": "test function",
-				"parameters": {
-					"type": "object",
-					"required": ["test"],
-					"properties": {
-						"test": {
-							"type": "string",
-							"description": "test prop",
-							"enum": []
-						}
-					}
-				}
-			}`,
-			wantErr: "",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			var tf ToolFunction
-			err := json.Unmarshal([]byte(tt.input), &tf)
-
-			if tt.wantErr != "" {
-				require.Error(t, err)
-				assert.Contains(t, err.Error(), tt.wantErr)
-			} else {
-				require.NoError(t, err)
-			}
-		})
-	}
-}
-
-func TestPropertyType_UnmarshalJSON(t *testing.T) {
-	tests := []struct {
-		name     string
-		input    string
-		expected PropertyType
-	}{
-		{
-			name:     "string type",
-			input:    `"string"`,
-			expected: PropertyType{"string"},
-		},
-		{
-			name:     "array of types",
-			input:    `["string", "number"]`,
-			expected: PropertyType{"string", "number"},
-		},
-		{
-			name:     "array with single type",
-			input:    `["string"]`,
-			expected: PropertyType{"string"},
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			var pt PropertyType
-			if err := json.Unmarshal([]byte(test.input), &pt); err != nil {
-				t.Errorf("Unexpected error: %v", err)
-			}
-
-			if len(pt) != len(test.expected) {
-				t.Errorf("Length mismatch: got %v, expected %v", len(pt), len(test.expected))
-			}
-
-			for i, v := range pt {
-				if v != test.expected[i] {
-					t.Errorf("Value mismatch at index %d: got %v, expected %v", i, v, test.expected[i])
-				}
-			}
-		})
-	}
-}
-
-func TestPropertyType_MarshalJSON(t *testing.T) {
-	tests := []struct {
-		name     string
-		input    PropertyType
-		expected string
-	}{
-		{
-			name:     "single type",
-			input:    PropertyType{"string"},
-			expected: `"string"`,
-		},
-		{
-			name:     "multiple types",
-			input:    PropertyType{"string", "number"},
-			expected: `["string","number"]`,
-		},
-		{
-			name:     "empty type",
-			input:    PropertyType{},
-			expected: `[]`,
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			data, err := json.Marshal(test.input)
-			if err != nil {
-				t.Errorf("Unexpected error: %v", err)
-			}
-
-			if string(data) != test.expected {
-				t.Errorf("Marshaled data mismatch: got %v, expected %v", string(data), test.expected)
-			}
-		})
-	}
-}

app/README.md

@@ -17,6 +17,6 @@ If you want to build the installer, youll need to install
 In the top directory of this repo, run the following powershell script
 to build the ollama CLI, ollama app, and ollama installer.
 
-```powershell
+```
 powershell -ExecutionPolicy Bypass -File .\scripts\build_windows.ps1
 ```

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 := context.Background()
-
-			// 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 := context.Background()
-
-			// 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(context.Background(), &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)
-}

cache/cache.go

@@ -0,0 +1,63 @@
+package cache
+
+import (
+	"github.com/ollama/ollama/ml"
+)
+
+type Options struct {
+	Position int
+}
+
+type Cache interface {
+	Sub(i int) Cache
+	Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor)
+}
+
+type Simple struct {
+	DType    ml.DType
+	Capacity int
+
+	keys, values []ml.Tensor
+}
+
+func (c *Simple) Sub(i int) Cache {
+	if i >= len(c.keys) {
+		c.keys = append(c.keys, make([]ml.Tensor, i-len(c.keys)+1)...)
+		c.values = append(c.values, make([]ml.Tensor, i-len(c.values)+1)...)
+	}
+
+	return &Simple{
+		keys:     c.keys[i : i+1],
+		values:   c.values[i : i+1],
+		Capacity: c.Capacity,
+		DType:    c.DType,
+	}
+}
+
+func (c *Simple) Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor) {
+	if c.keys[0] == nil || c.values[0] == nil {
+		c.keys[0] = ctx.Zeros(c.DType, int(key.Dim(0)*key.Dim(1))*c.Capacity)
+		c.values[0] = ctx.Zeros(c.DType, int(value.Dim(0)*value.Dim(1))*c.Capacity)
+	}
+
+	ctx.Forward(key.Copy(ctx, c.keys[0].View(ctx, int(key.Stride(2))*opts.Position, int(key.Dim(0)*key.Dim(1)*key.Dim(2)))))
+	ctx.Forward(value.Copy(ctx, c.values[0].View(ctx, int(value.Stride(2))*opts.Position, int(value.Dim(0)*value.Dim(1)*value.Dim(2)))))
+
+	n := min(c.Capacity, int(key.Dim(2))+opts.Position)
+
+	key = c.keys[0].View(ctx, 0,
+		int(key.Dim(0)), int(key.Stride(1)),
+		int(key.Dim(1)), int(key.Stride(2)),
+		n,
+	)
+
+	value = c.values[0].View(ctx, 0,
+		int(value.Dim(0)), int(value.Stride(1)),
+		int(value.Dim(1)), int(value.Stride(2)),
+		n,
+	)
+
+	// TODO shift context if necessary
+
+	return key, value
+}

cmd/cmd.go

@@ -18,11 +18,8 @@ import (
 	"os/signal"
 	"path/filepath"
 	"runtime"
-	"slices"
-	"sort"
 	"strconv"
 	"strings"
-	"sync"
 	"sync/atomic"
 	"syscall"
 	"time"
@@ -32,15 +29,15 @@ import (
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
-	"golang.org/x/sync/errgroup"
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/llama"
+	"github.com/ollama/ollama/llama/runner"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
-	"github.com/ollama/ollama/runner"
 	"github.com/ollama/ollama/server"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -62,7 +59,7 @@ func getModelfileName(cmd *cobra.Command) (string, error) {
 
 	_, err = os.Stat(absName)
 	if err != nil {
-		return "", err
+		return filename, err
 	}
 
 	return absName, nil
@@ -108,7 +105,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Model = args[0]
+	req.Name = args[0]
 	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
@@ -119,43 +116,26 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	var mu sync.Mutex
-	var g errgroup.Group
-	g.SetLimit(max(runtime.GOMAXPROCS(0)-1, 1))
-	// copy files since we'll be modifying the map
-	temp := req.Files
-	req.Files = make(map[string]string, len(temp))
-	for f, digest := range temp {
-		g.Go(func() error {
+	if len(req.Files) > 0 {
+		fileMap := map[string]string{}
+		for f, digest := range req.Files {
 			if _, err := createBlob(cmd, client, f, digest, p); err != nil {
 				return err
 			}
-
-			mu.Lock()
-			req.Files[filepath.Base(f)] = digest
-			mu.Unlock()
-			return nil
-		})
+			fileMap[filepath.Base(f)] = digest
+		}
+		req.Files = fileMap
 	}
 
-	// copy files since we'll be modifying the map
-	temp = req.Adapters
-	req.Adapters = make(map[string]string, len(temp))
-	for f, digest := range temp {
-		g.Go(func() error {
+	if len(req.Adapters) > 0 {
+		fileMap := map[string]string{}
+		for f, digest := range req.Adapters {
 			if _, err := createBlob(cmd, client, f, digest, p); err != nil {
 				return err
 			}
-
-			mu.Lock()
-			req.Adapters[filepath.Base(f)] = digest
-			mu.Unlock()
-			return nil
-		})
-	}
-
-	if err := g.Wait(); err != nil {
-		return err
+			fileMap[filepath.Base(f)] = digest
+		}
+		req.Adapters = fileMap
 	}
 
 	bars := make(map[string]*progress.Bar)
@@ -232,7 +212,7 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string, digest stri
 		}
 	}()
 
-	if err := client.CreateBlob(cmd.Context(), digest, io.TeeReader(bin, &pw)); err != nil {
+	if err = client.CreateBlob(cmd.Context(), digest, io.TeeReader(bin, &pw)); err != nil {
 		return "", err
 	}
 	return digest, nil
@@ -276,7 +256,6 @@ func StopHandler(cmd *cobra.Command, args []string) error {
 		if strings.Contains(err.Error(), "not found") {
 			return fmt.Errorf("couldn't find model \"%s\" to stop", args[0])
 		}
-		return err
 	}
 	return nil
 }
@@ -287,7 +266,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	opts := runOptions{
 		Model:    args[0],
 		WordWrap: os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]any{},
+		Options:  map[string]interface{}{},
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -359,21 +338,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision)
-
-	// TODO: remove the projector info and vision info checks below,
-	// these are left in for backwards compatibility with older servers
-	// that don't have the capabilities field in the model info
-	if len(info.ProjectorInfo) != 0 {
-		opts.MultiModal = true
-	}
-	for k := range info.ModelInfo {
-		if strings.Contains(k, ".vision.") {
-			opts.MultiModal = true
-			break
-		}
-	}
-
+	opts.MultiModal = len(info.ProjectorInfo) != 0
 	opts.ParentModel = info.Details.ParentModel
 
 	if interactive {
@@ -594,9 +559,8 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 	parameters, errParams := cmd.Flags().GetBool("parameters")
 	system, errSystem := cmd.Flags().GetBool("system")
 	template, errTemplate := cmd.Flags().GetBool("template")
-	verbose, errVerbose := cmd.Flags().GetBool("verbose")
 
-	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate, errVerbose} {
+	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate} {
 		if boolErr != nil {
 			return errors.New("error retrieving flags")
 		}
@@ -634,7 +598,7 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
 	}
 
-	req := api.ShowRequest{Name: args[0], Verbose: verbose}
+	req := api.ShowRequest{Name: args[0]}
 	resp, err := client.Show(cmd.Context(), &req)
 	if err != nil {
 		return err
@@ -657,10 +621,10 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	return showInfo(resp, verbose, os.Stdout)
+	return showInfo(resp, os.Stdout)
 }
 
-func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
+func showInfo(resp *api.ShowResponse, w io.Writer) error {
 	tableRender := func(header string, rows func() [][]string) {
 		fmt.Fprintln(w, " ", header)
 		table := tablewriter.NewWriter(w)
@@ -694,15 +658,6 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 		return
 	})
 
-	if len(resp.Capabilities) > 0 {
-		tableRender("Capabilities", func() (rows [][]string) {
-			for _, capability := range resp.Capabilities {
-				rows = append(rows, []string{"", capability.String()})
-			}
-			return
-		})
-	}
-
 	if resp.ProjectorInfo != nil {
 		tableRender("Projector", func() (rows [][]string) {
 			arch := resp.ProjectorInfo["general.architecture"].(string)
@@ -726,47 +681,6 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 		})
 	}
 
-	if resp.ModelInfo != nil && verbose {
-		tableRender("Metadata", func() (rows [][]string) {
-			keys := make([]string, 0, len(resp.ModelInfo))
-			for k := range resp.ModelInfo {
-				keys = append(keys, k)
-			}
-			sort.Strings(keys)
-
-			for _, k := range keys {
-				var v string
-				switch vData := resp.ModelInfo[k].(type) {
-				case bool:
-					v = fmt.Sprintf("%t", vData)
-				case string:
-					v = vData
-				case float64:
-					v = fmt.Sprintf("%g", vData)
-				case []any:
-					n := 3
-					if len(vData) < n {
-						n = len(vData)
-					}
-					v = fmt.Sprintf("%v", vData[:n])
-				default:
-					v = fmt.Sprintf("%T", vData)
-				}
-				rows = append(rows, []string{"", k, v})
-			}
-			return
-		})
-	}
-
-	if len(resp.Tensors) > 0 && verbose {
-		tableRender("Tensors", func() (rows [][]string) {
-			for _, t := range resp.Tensors {
-				rows = append(rows, []string{"", t.Name, t.Type, fmt.Sprint(t.Shape)})
-			}
-			return
-		})
-	}
-
 	head := func(s string, n int) (rows [][]string) {
 		scanner := bufio.NewScanner(strings.NewReader(s))
 		for scanner.Scan() && (len(rows) < n || n < 0) {
@@ -827,38 +741,13 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 
 	fn := func(resp api.ProgressResponse) error {
 		if resp.Digest != "" {
-			if resp.Completed == 0 {
-				// This is the initial status update for the
-				// layer, which the server sends before
-				// beginning the download, for clients to
-				// compute total size and prepare for
-				// downloads, if needed.
-				//
-				// Skipping this here to avoid showing a 0%
-				// progress bar, which *should* clue the user
-				// into the fact that many things are being
-				// downloaded and that the current active
-				// download is not that last. However, in rare
-				// cases it seems to be triggering to some, and
-				// it isn't worth explaining, so just ignore
-				// and regress to the old UI that keeps giving
-				// you the "But wait, there is more!" after
-				// each "100% done" bar, which is "better."
-				return nil
-			}
-
 			if spinner != nil {
 				spinner.Stop()
 			}
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				name, isDigest := strings.CutPrefix(resp.Digest, "sha256:")
-				name = strings.TrimSpace(name)
-				if isDigest {
-					name = name[:min(12, len(name))]
-				}
-				bar = progress.NewBar(fmt.Sprintf("pulling %s:", name), resp.Total, resp.Completed)
+				bar = progress.NewBar(fmt.Sprintf("pulling %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -878,7 +767,11 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	request := api.PullRequest{Name: args[0], Insecure: insecure}
-	return client.Pull(cmd.Context(), &request, fn)
+	if err := client.Pull(cmd.Context(), &request, fn); err != nil {
+		return err
+	}
+
+	return nil
 }
 
 type generateContextKey string
@@ -892,7 +785,7 @@ type runOptions struct {
 	Format      string
 	System      string
 	Images      []api.ImageData
-	Options     map[string]any
+	Options     map[string]interface{}
 	MultiModal  bool
 	KeepAlive   *api.Duration
 }
@@ -1294,7 +1187,6 @@ func NewCLI() *cobra.Command {
 	showCmd.Flags().Bool("parameters", false, "Show parameters of a model")
 	showCmd.Flags().Bool("template", false, "Show template of a model")
 	showCmd.Flags().Bool("system", false, "Show system message of a model")
-	showCmd.Flags().BoolP("verbose", "v", false, "Show detailed model information")
 
 	runCmd := &cobra.Command{
 		Use:     "run MODEL [PROMPT]",
@@ -1379,6 +1271,7 @@ func NewCLI() *cobra.Command {
 
 	runnerCmd := &cobra.Command{
 		Use:    "runner",
+		Short:  llama.PrintSystemInfo(),
 		Hidden: true,
 		RunE: func(cmd *cobra.Command, args []string) error {
 			return runner.Execute(os.Args[1:])
@@ -1421,12 +1314,12 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_NOPRUNE"],
 				envVars["OLLAMA_ORIGINS"],
 				envVars["OLLAMA_SCHED_SPREAD"],
+				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_KV_CACHE_TYPE"],
 				envVars["OLLAMA_LLM_LIBRARY"],
 				envVars["OLLAMA_GPU_OVERHEAD"],
 				envVars["OLLAMA_LOAD_TIMEOUT"],
-				envVars["OLLAMA_CONTEXT_LENGTH"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

cmd/cmd_test.go

@@ -10,13 +10,11 @@ import (
 	"os"
 	"strings"
 	"testing"
-	"time"
 
 	"github.com/google/go-cmp/cmp"
 	"github.com/spf13/cobra"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/types/model"
 )
 
 func TestShowInfo(t *testing.T) {
@@ -28,7 +26,7 @@ func TestShowInfo(t *testing.T) {
 				ParameterSize:     "7B",
 				QuantizationLevel: "FP16",
 			},
-		}, false, &b); err != nil {
+		}, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -58,7 +56,7 @@ func TestShowInfo(t *testing.T) {
 				ParameterSize:     "7B",
 				QuantizationLevel: "FP16",
 			},
-		}, false, &b); err != nil {
+		}, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -69,60 +67,6 @@ func TestShowInfo(t *testing.T) {
     embedding length    0       
     quantization        FP16    
 
-`
-		if diff := cmp.Diff(expect, b.String()); diff != "" {
-			t.Errorf("unexpected output (-want +got):\n%s", diff)
-		}
-	})
-
-	t.Run("verbose model", func(t *testing.T) {
-		var b bytes.Buffer
-		if err := showInfo(&api.ShowResponse{
-			Details: api.ModelDetails{
-				Family:            "test",
-				ParameterSize:     "8B",
-				QuantizationLevel: "FP16",
-			},
-			Parameters: `
-			stop up`,
-			ModelInfo: map[string]any{
-				"general.architecture":    "test",
-				"general.parameter_count": float64(8_000_000_000),
-				"some.true_bool":          true,
-				"some.false_bool":         false,
-				"test.context_length":     float64(1000),
-				"test.embedding_length":   float64(11434),
-			},
-			Tensors: []api.Tensor{
-				{Name: "blk.0.attn_k.weight", Type: "BF16", Shape: []uint64{42, 3117}},
-				{Name: "blk.0.attn_q.weight", Type: "FP16", Shape: []uint64{3117, 42}},
-			},
-		}, true, &b); err != nil {
-			t.Fatal(err)
-		}
-
-		expect := `  Model
-    architecture        test     
-    parameters          8B       
-    context length      1000     
-    embedding length    11434    
-    quantization        FP16     
-
-  Parameters
-    stop    up    
-
-  Metadata
-    general.architecture       test     
-    general.parameter_count    8e+09    
-    some.false_bool            false    
-    some.true_bool             true     
-    test.context_length        1000     
-    test.embedding_length      11434    
-
-  Tensors
-    blk.0.attn_k.weight    BF16    [42 3117]    
-    blk.0.attn_q.weight    FP16    [3117 42]    
-
 `
 		if diff := cmp.Diff(expect, b.String()); diff != "" {
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
@@ -144,7 +88,7 @@ func TestShowInfo(t *testing.T) {
 			stop you
 			stop up
 			temperature 99`,
-		}, false, &b); err != nil {
+		}, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -181,7 +125,7 @@ func TestShowInfo(t *testing.T) {
 				"clip.vision.embedding_length": float64(0),
 				"clip.vision.projection_dim":   float64(0),
 			},
-		}, false, &b); err != nil {
+		}, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -214,7 +158,7 @@ func TestShowInfo(t *testing.T) {
 Ahoy, matey!
 Weigh anchor!
 			`,
-		}, false, &b); err != nil {
+		}, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -243,7 +187,7 @@ Weigh anchor!
 				QuantizationLevel: "FP16",
 			},
 			License: license,
-		}, false, &b); err != nil {
+		}, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -261,34 +205,6 @@ Weigh anchor!
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
 		}
 	})
-
-	t.Run("capabilities", func(t *testing.T) {
-		var b bytes.Buffer
-		if err := showInfo(&api.ShowResponse{
-			Details: api.ModelDetails{
-				Family:            "test",
-				ParameterSize:     "7B",
-				QuantizationLevel: "FP16",
-			},
-			Capabilities: []model.Capability{model.CapabilityVision, model.CapabilityTools},
-		}, false, &b); err != nil {
-			t.Fatal(err)
-		}
-
-		expect := "  Model\n" +
-			"    architecture    test    \n" +
-			"    parameters      7B      \n" +
-			"    quantization    FP16    \n" +
-			"\n" +
-			"  Capabilities\n" +
-			"    vision    \n" +
-			"    tools     \n" +
-			"\n"
-
-		if diff := cmp.Diff(expect, b.String()); diff != "" {
-			t.Errorf("unexpected output (-want +got):\n%s", diff)
-		}
-	})
 }
 
 func TestDeleteHandler(t *testing.T) {
@@ -363,7 +279,7 @@ func TestGetModelfileName(t *testing.T) {
 			name:          "no modelfile specified, no modelfile exists",
 			modelfileName: "",
 			fileExists:    false,
-			expectedName:  "",
+			expectedName:  "Modelfile",
 			expectedErr:   os.ErrNotExist,
 		},
 		{
@@ -377,7 +293,7 @@ func TestGetModelfileName(t *testing.T) {
 			name:          "modelfile specified, no modelfile exists",
 			modelfileName: "crazyfile",
 			fileExists:    false,
-			expectedName:  "",
+			expectedName:  "crazyfile",
 			expectedErr:   os.ErrNotExist,
 		},
 		{
@@ -415,7 +331,6 @@ func TestGetModelfileName(t *testing.T) {
 				if err != nil {
 					t.Fatalf("temp modelfile creation failed: %v", err)
 				}
-				defer tempFile.Close()
 
 				expectedFilename = tempFile.Name()
 				err = cmd.Flags().Set("file", expectedFilename)
@@ -575,96 +490,6 @@ func TestPushHandler(t *testing.T) {
 	}
 }
 
-func TestListHandler(t *testing.T) {
-	tests := []struct {
-		name           string
-		args           []string
-		serverResponse []api.ListModelResponse
-		expectedError  string
-		expectedOutput string
-	}{
-		{
-			name: "list all models",
-			args: []string{},
-			serverResponse: []api.ListModelResponse{
-				{Name: "model1", Digest: "sha256:abc123", Size: 1024, ModifiedAt: time.Now().Add(-24 * time.Hour)},
-				{Name: "model2", Digest: "sha256:def456", Size: 2048, ModifiedAt: time.Now().Add(-48 * time.Hour)},
-			},
-			expectedOutput: "NAME      ID              SIZE      MODIFIED     \n" +
-				"model1    sha256:abc12    1.0 KB    24 hours ago    \n" +
-				"model2    sha256:def45    2.0 KB    2 days ago      \n",
-		},
-		{
-			name: "filter models by prefix",
-			args: []string{"model1"},
-			serverResponse: []api.ListModelResponse{
-				{Name: "model1", Digest: "sha256:abc123", Size: 1024, ModifiedAt: time.Now().Add(-24 * time.Hour)},
-				{Name: "model2", Digest: "sha256:def456", Size: 2048, ModifiedAt: time.Now().Add(-24 * time.Hour)},
-			},
-			expectedOutput: "NAME      ID              SIZE      MODIFIED     \n" +
-				"model1    sha256:abc12    1.0 KB    24 hours ago    \n",
-		},
-		{
-			name:          "server error",
-			args:          []string{},
-			expectedError: "server error",
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			mockServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-				if r.URL.Path != "/api/tags" || r.Method != http.MethodGet {
-					t.Errorf("unexpected request to %s %s", r.Method, r.URL.Path)
-					http.Error(w, "not found", http.StatusNotFound)
-					return
-				}
-
-				if tt.expectedError != "" {
-					http.Error(w, tt.expectedError, http.StatusInternalServerError)
-					return
-				}
-
-				response := api.ListResponse{Models: tt.serverResponse}
-				if err := json.NewEncoder(w).Encode(response); err != nil {
-					t.Fatal(err)
-				}
-			}))
-			defer mockServer.Close()
-
-			t.Setenv("OLLAMA_HOST", mockServer.URL)
-
-			cmd := &cobra.Command{}
-			cmd.SetContext(context.TODO())
-
-			// Capture stdout
-			oldStdout := os.Stdout
-			r, w, _ := os.Pipe()
-			os.Stdout = w
-
-			err := ListHandler(cmd, tt.args)
-
-			// Restore stdout and get output
-			w.Close()
-			os.Stdout = oldStdout
-			output, _ := io.ReadAll(r)
-
-			if tt.expectedError == "" {
-				if err != nil {
-					t.Errorf("expected no error, got %v", err)
-				}
-				if got := string(output); got != tt.expectedOutput {
-					t.Errorf("expected output:\n%s\ngot:\n%s", tt.expectedOutput, got)
-				}
-			} else {
-				if err == nil || !strings.Contains(err.Error(), tt.expectedError) {
-					t.Errorf("expected error containing %q, got %v", tt.expectedError, err)
-				}
-			}
-		})
-	}
-}
-
 func TestCreateHandler(t *testing.T) {
 	tests := []struct {
 		name           string
@@ -690,7 +515,7 @@ func TestCreateHandler(t *testing.T) {
 						return
 					}
 
-					if req.Model != "test-model" {
+					if req.Name != "test-model" {
 						t.Errorf("expected model name 'test-model', got %s", req.Name)
 					}
 
@@ -791,132 +616,3 @@ func TestCreateHandler(t *testing.T) {
 		})
 	}
 }
-
-func TestNewCreateRequest(t *testing.T) {
-	tests := []struct {
-		name     string
-		from     string
-		opts     runOptions
-		expected *api.CreateRequest
-	}{
-		{
-			"basic test",
-			"newmodel",
-			runOptions{
-				Model:       "mymodel",
-				ParentModel: "",
-				Prompt:      "You are a fun AI agent",
-				Messages:    []api.Message{},
-				WordWrap:    true,
-			},
-			&api.CreateRequest{
-				From:  "mymodel",
-				Model: "newmodel",
-			},
-		},
-		{
-			"parent model test",
-			"newmodel",
-			runOptions{
-				Model:       "mymodel",
-				ParentModel: "parentmodel",
-				Messages:    []api.Message{},
-				WordWrap:    true,
-			},
-			&api.CreateRequest{
-				From:  "parentmodel",
-				Model: "newmodel",
-			},
-		},
-		{
-			"parent model as filepath test",
-			"newmodel",
-			runOptions{
-				Model:       "mymodel",
-				ParentModel: "/some/file/like/etc/passwd",
-				Messages:    []api.Message{},
-				WordWrap:    true,
-			},
-			&api.CreateRequest{
-				From:  "mymodel",
-				Model: "newmodel",
-			},
-		},
-		{
-			"parent model as windows filepath test",
-			"newmodel",
-			runOptions{
-				Model:       "mymodel",
-				ParentModel: "D:\\some\\file\\like\\etc\\passwd",
-				Messages:    []api.Message{},
-				WordWrap:    true,
-			},
-			&api.CreateRequest{
-				From:  "mymodel",
-				Model: "newmodel",
-			},
-		},
-		{
-			"options test",
-			"newmodel",
-			runOptions{
-				Model:       "mymodel",
-				ParentModel: "parentmodel",
-				Options: map[string]any{
-					"temperature": 1.0,
-				},
-			},
-			&api.CreateRequest{
-				From:  "parentmodel",
-				Model: "newmodel",
-				Parameters: map[string]any{
-					"temperature": 1.0,
-				},
-			},
-		},
-		{
-			"messages test",
-			"newmodel",
-			runOptions{
-				Model:       "mymodel",
-				ParentModel: "parentmodel",
-				System:      "You are a fun AI agent",
-				Messages: []api.Message{
-					{
-						Role:    "user",
-						Content: "hello there!",
-					},
-					{
-						Role:    "assistant",
-						Content: "hello to you!",
-					},
-				},
-				WordWrap: true,
-			},
-			&api.CreateRequest{
-				From:   "parentmodel",
-				Model:  "newmodel",
-				System: "You are a fun AI agent",
-				Messages: []api.Message{
-					{
-						Role:    "user",
-						Content: "hello there!",
-					},
-					{
-						Role:    "assistant",
-						Content: "hello to you!",
-					},
-				},
-			},
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			actual := NewCreateRequest(tt.from, tt.opts)
-			if !cmp.Equal(actual, tt.expected) {
-				t.Errorf("expected output %#v, got %#v", tt.expected, actual)
-			}
-		})
-	}
-}

cmd/interactive.go

@@ -18,7 +18,6 @@ import (
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/types/errtypes"
-	"github.com/ollama/ollama/types/model"
 )
 
 type MultilineState int
@@ -196,10 +195,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
-				if strings.Contains(err.Error(), "not found") {
-					fmt.Printf("error: %v\n", err)
-					continue
-				}
 				return err
 			}
 			continue
@@ -348,7 +343,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					_ = showInfo(resp, false, os.Stderr)
+					_ = showInfo(resp, os.Stderr)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")
@@ -460,16 +455,9 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 }
 
 func NewCreateRequest(name string, opts runOptions) *api.CreateRequest {
-	parentModel := opts.ParentModel
-
-	modelName := model.ParseName(parentModel)
-	if !modelName.IsValid() {
-		parentModel = ""
-	}
-
 	req := &api.CreateRequest{
-		Model: name,
-		From:  cmp.Or(parentModel, opts.Model),
+		Name: name,
+		From: cmp.Or(opts.ParentModel, opts.Model),
 	}
 
 	if opts.System != "" {
@@ -503,7 +491,6 @@ func normalizeFilePath(fp string) string {
 		"\\\\", "\\", // Escaped backslash
 		"\\*", "*", // Escaped asterisk
 		"\\?", "?", // Escaped question mark
-		"\\~", "~", // Escaped tilde
 	).Replace(fp)
 }
 

cmd/runner/main.go

@@ -4,7 +4,7 @@ import (
 	"fmt"
 	"os"
 
-	"github.com/ollama/ollama/runner"
+	"github.com/ollama/ollama/llama/runner"
 )
 
 func main() {

convert/convert.go

@@ -4,22 +4,17 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"io"
 	"io/fs"
 	"log/slog"
-	"os"
 	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
 
 type ModelParameters struct {
-	Architectures []string       `json:"architectures"`
-	VocabSize     uint32         `json:"vocab_size"`
-	TextModel     TextParameters `json:"text_config"`
-}
-
-type TextParameters struct {
-	VocabSize uint32 `json:"vocab_size"`
+	Architectures []string `json:"architectures"`
+	VocabSize     uint32   `json:"vocab_size"`
 }
 
 type AdapterParameters struct {
@@ -84,12 +79,12 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-func (ModelParameters) writeFile(f *os.File, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(f, kv, ts)
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
+	return ggml.WriteGGUF(ws, kv, ts)
 }
 
-func (AdapterParameters) writeFile(f *os.File, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(f, kv, ts)
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
+	return ggml.WriteGGUF(ws, kv, ts)
 }
 
 type ModelConverter interface {
@@ -104,7 +99,7 @@ type ModelConverter interface {
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
 	// writeFile writes the model to the provided io.WriteSeeker
-	writeFile(*os.File, ggml.KV, []ggml.Tensor) error
+	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
 type moreParser interface {
@@ -120,10 +115,10 @@ type AdapterConverter interface {
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
 
-	writeFile(*os.File, ggml.KV, []ggml.Tensor) error
+	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
-func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -158,14 +153,14 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 		return err
 	}
 
-	return conv.writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
+	return conv.writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
 }
 
 // Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
 // and files it finds in the input path.
 // Supported input model formats include safetensors.
 // Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func ConvertModel(fsys fs.FS, f *os.File) error {
+func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
 		return err
@@ -182,28 +177,22 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 
 	var conv ModelConverter
 	switch p.Architectures[0] {
-	case "LlamaForCausalLM":
+	case "LlamaForCausalLM", "MistralForCausalLM":
 		conv = &llamaModel{}
-	case "Mistral3ForConditionalGeneration":
-		conv = &mistral3Model{}
 	case "MixtralForCausalLM":
 		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
 		conv = &gemmaModel{}
 	case "Gemma2ForCausalLM":
 		conv = &gemma2Model{}
-	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
-		conv = &gemma3Model{Architecture: p.Architectures[0]}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":
 		conv = &qwen2Model{}
 	case "BertModel":
 		conv = &bertModel{}
-	case "CohereForCausalLM":
-		conv = &commandrModel{}
 	default:
-		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
+		return errors.New("unsupported architecture")
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
@@ -222,14 +211,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	}
 
 	vocabSize := int(p.VocabSize)
-	if vocabSize == 0 {
-		tVocabSize := int(p.TextModel.VocabSize)
-		vocabSize = tVocabSize
-	}
-
 	switch {
-	case vocabSize == 0:
-		slog.Warn("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
 	case vocabSize > len(t.Vocabulary.Tokens):
 		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
 		for i := range vocabSize - len(t.Vocabulary.Tokens) {
@@ -248,5 +230,5 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		return err
 	}
 
-	return conv.writeFile(f, conv.KV(t), conv.Tensors(ts))
+	return conv.writeFile(ws, conv.KV(t), conv.Tensors(ts))
 }

convert/convert_commandr.go

@@ -1,76 +0,0 @@
-package convert
-
-import (
-	"cmp"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type commandrModel struct {
-	ModelParameters
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	HiddenLayers          uint32  `json:"num_hidden_layers"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	NumAttentionHeads     uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-	LayerNormEPS          float32 `json:"layer_norm_eps"`
-	RopeTheta             float32 `json:"rope_theta"`
-	UseQKNorm             bool    `json:"use_qk_norm"`
-	MaxLength             uint32  `json:"model_max_length"`
-	LogitScale            float32 `json:"logit_scale"`
-	NCtx                  uint32  `json:"n_ctx"`
-}
-
-var _ ModelConverter = (*commandrModel)(nil)
-
-func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
-	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "command-r"
-	kv["general.name"] = "command-r"
-	kv["command-r.context_length"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings, p.NCtx)
-	kv["command-r.embedding_length"] = p.HiddenSize
-	kv["command-r.block_count"] = p.HiddenLayers
-	kv["command-r.feed_forward_length"] = p.IntermediateSize
-	kv["command-r.attention.head_count"] = p.NumAttentionHeads
-	kv["command-r.attention.head_count_kv"] = p.NumKeyValueHeads
-	kv["command-r.attention.layer_norm_epsilon"] = p.LayerNormEPS
-	kv["command-r.rope.freq_base"] = p.RopeTheta
-	kv["command-r.max_position_embeddings"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings)
-	kv["command-r.logit_scale"] = p.LogitScale
-	kv["command-r.rope.scaling.type"] = "none"
-
-	return kv
-}
-
-func (p *commandrModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
-	for _, t := range ts {
-		out = append(out, ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *commandrModel) Replacements() []string {
-	return []string{
-		"self_attn.q_norm", "attn_q_norm",
-		"self_attn.k_norm", "attn_k_norm",
-		"model.layers", "blk",
-		"input_layernorm", "attn_norm",
-		"mlp.down_proj", "ffn_down",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.up_proj", "ffn_up",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.o_proj", "attn_output",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.v_proj", "attn_v",
-		"model.norm", "output_norm",
-		"model.embed_tokens", "token_embd",
-	}
-}

convert/convert_gemma.go

@@ -45,7 +45,7 @@ func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
 func (p *gemmaModel) Tensors(ts []Tensor) []ggml.Tensor {
 	var out []ggml.Tensor
 	for _, t := range ts {
-		if !strings.HasPrefix(t.Name(), "v.") && strings.HasSuffix(t.Name(), "_norm.weight") {
+		if strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 

convert/convert_gemma3.go

@@ -1,142 +0,0 @@
-package convert
-
-import (
-	"cmp"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type gemma3Model struct {
-	gemmaModel
-	Architecture string
-	TextModel    struct {
-		HeadDim          uint32 `json:"head_dim"`
-		HiddenSize       uint32 `json:"hidden_size"`
-		HiddenLayers     uint32 `json:"num_hidden_layers"`
-		IntermediateSize uint32 `json:"intermediate_size"`
-		SlidingWindow    uint32 `json:"sliding_window"`
-	} `json:"text_config"`
-	VisionModel struct {
-		NumAttentionHeads uint32  `json:"num_attention_heads"` // attention.head_count 16
-		LayerNormEpsilon  float32 `json:"layer_norm_eps"`      // attention.layer_norm_epsilon 1e-05
-		NumHiddenLayers   uint32  `json:"num_hidden_layers"`   // block_count 32
-		HiddenSize        uint32  `json:"hidden_size"`         // embedding_length 1280
-		IntermediateSize  uint32  `json:"intermediate_size"`   // feed_forward_length 5120
-		ImageSize         uint32  `json:"image_size"`          // image_size 560
-		NumChannels       uint32  `json:"num_channels"`        // num_channels 3
-		PatchSize         uint32  `json:"patch_size"`          // patch_size 14
-	} `json:"vision_config"`
-	MaxPositionEmbeddings    uint32  `json:"max_position_embeddings"`
-	NumAttentionHeads        uint32  `json:"num_attention_heads"`
-	NumKeyValueHeads         uint32  `json:"num_key_value_heads"`
-	RMSNormEPS               float32 `json:"rms_norm_eps"`
-	HeadDim                  uint32  `json:"head_dim"`
-	FinalLogitSoftcap        float32 `json:"final_logit_softcapping"`
-	RopeLocalTheta           float32 `json:"rope_local_base_freq"`
-	RopeGlobalTheta          float32 `json:"rope_global_base_freq"`
-	SlidingWindow            uint32  `json:"sliding_window"`
-	MultiModalTokensPerImage uint32  `json:"mm_tokens_per_image"`
-}
-
-const (
-	gemma4BLayerCount  = 34
-	gemma12BLayerCount = 48
-	gemma27BLayerCount = 62
-)
-
-func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
-	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "gemma3"
-
-	numBlocks := cmp.Or(p.HiddenLayers, p.TextModel.HiddenLayers)
-	kv["gemma3.block_count"] = numBlocks
-
-	var (
-		numHeads   uint32
-		numKVHeads uint32
-	)
-
-	switch numBlocks {
-	case gemma4BLayerCount:
-		numHeads = 8
-		numKVHeads = 4
-	case gemma12BLayerCount:
-		numHeads = 16
-		numKVHeads = 8
-	case gemma27BLayerCount:
-		numHeads = 32
-		numKVHeads = 16
-	default:
-		numHeads = p.NumAttentionHeads
-		numKVHeads = p.NumKeyValueHeads
-	}
-
-	kv["gemma3.attention.head_count"] = numHeads
-	kv["gemma3.attention.head_count_kv"] = numKVHeads
-
-	switch p.Architecture {
-	case "Gemma3ForCausalLM":
-		kv["gemma3.context_length"] = p.MaxPositionEmbeddings
-		kv["gemma3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
-		kv["gemma3.attention.key_length"] = p.HeadDim
-		kv["gemma3.attention.value_length"] = p.HeadDim
-		kv["gemma3.attention.sliding_window"] = p.SlidingWindow
-		kv["gemma3.final_logit_softcapping"] = cmp.Or(p.FinalLogitSoftcap, 30)
-		kv["gemma3.rope.local.freq_base"] = cmp.Or(p.RopeLocalTheta, 10000.0)
-		kv["gemma3.rope.global.freq_base"] = cmp.Or(p.RopeGlobalTheta, 1000000.0)
-		kv["gemma3.embedding_length"] = p.HiddenSize
-		kv["gemma3.feed_forward_length"] = p.IntermediateSize
-	default:
-		kv["gemma3.context_length"] = cmp.Or(p.MaxPositionEmbeddings, 131072)
-		kv["gemma3.embedding_length"] = p.TextModel.HiddenSize
-		kv["gemma3.feed_forward_length"] = p.TextModel.IntermediateSize
-		kv["gemma3.attention.sliding_window"] = p.TextModel.SlidingWindow
-		kv["gemma3.vision.block_count"] = p.VisionModel.NumHiddenLayers
-		kv["gemma3.vision.embedding_length"] = p.VisionModel.HiddenSize
-		kv["gemma3.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
-		kv["gemma3.vision.image_size"] = p.VisionModel.ImageSize
-		kv["gemma3.vision.patch_size"] = p.VisionModel.PatchSize
-		kv["gemma3.vision.num_channels"] = cmp.Or(p.VisionModel.NumChannels, 3)
-		kv["gemma3.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
-		kv["gemma3.vision.attention.layer_norm_epsilon"] = cmp.Or(p.VisionModel.LayerNormEpsilon, 1e-6)
-		kv["gemma3.attention.key_length"] = cmp.Or(p.TextModel.HeadDim, 256)
-		kv["gemma3.attention.value_length"] = cmp.Or(p.TextModel.HeadDim, 256)
-	}
-
-	if p.MultiModalTokensPerImage > 0 {
-		kv["gemma3.mm.tokens_per_image"] = p.MultiModalTokensPerImage
-	}
-
-	return kv
-}
-
-func (p *gemma3Model) Replacements() []string {
-	return []string{
-		"lm_head", "output",
-		"model.embed_tokens", "token_embd",
-		"model.norm", "output_norm",
-		"vision_tower.vision_model.embeddings", "v",
-		"vision_tower.vision_model", "v",
-		"vision_model.vision_model.embeddings", "v",
-		"vision_model.vision_model", "v",
-		"language_model.", "",
-		"model.layers", "blk",
-		"encoder.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",
-		"self_attn.out_proj", "attn_output",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.down_proj", "ffn_down",
-		"mlp.up_proj", "ffn_up",
-		"post_attention_layernorm", "post_attention_norm",
-		"pre_feedforward_layernorm", "ffn_norm",
-		"post_feedforward_layernorm", "post_ffw_norm",
-		"input_projection_weight", "input_projection.weight",
-		"multi_modal_projector", "mm",
-	}
-}

convert/convert_llama.go

@@ -28,12 +28,12 @@ type llamaModel struct {
 	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
 	RopeTheta             float32 `json:"rope_theta"`
 	RopeScaling           struct {
-		Type                          string  `json:"type"`
-		RopeType                      string  `json:"rope_type"`
-		Factor                        float32 `json:"factor"`
-		LowFrequencyFactor            float32 `json:"low_freq_factor"`
-		HighFrequencyFactor           float32 `json:"high_freq_factor"`
-		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+		Type                            string  `json:"type"`
+		RopeType                        string  `json:"rope_type"`
+		Factor                          float32 `json:"factor"`
+		LowFrequencyFactor              float32 `json:"low_freq_factor"`
+		HighFrequencyFactor             float32 `json:"high_freq_factor"`
+		OriginalMaxPositionalEmbeddings uint32  `json:"original_max_positional_embeddings"`
 
 		factors ropeFactor
 	} `json:"rope_scaling"`
@@ -84,7 +84,7 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 			factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
 			factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
 
-			original := cmp.Or(p.RopeScaling.OriginalMaxPositionEmbeddings, 8192)
+			original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
 			lambdaLow := float32(original) / factorLow
 			lambdaHigh := float32(original) / factorHigh
 

convert/convert_mistral.go

@@ -1,190 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"fmt"
-	"strings"
-
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-
-	"github.com/ollama/ollama/fs/ggml"
-)
-
-type mistral3Model struct {
-	ModelParameters
-	ImageTokenIndex    uint32 `json:"image_token_index"`
-	SpatialMergeSize   uint32 `json:"spatial_merge_size"`
-	VisionFeatureLayer int32  `json:"vision_feature_layer"`
-	TextModel          struct {
-		NumHiddenLayers       uint32  `json:"num_hidden_layers"`
-		MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-		HiddenSize            uint32  `json:"hidden_size"`
-		IntermediateSize      uint32  `json:"intermediate_size"`
-		NumAttentionHeads     uint32  `json:"num_attention_heads"`
-		NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
-		RopeTheta             float32 `json:"rope_theta"`
-		RMSNormEPS            float32 `json:"rms_norm_eps"`
-		HeadDim               uint32  `json:"head_dim"`
-		SlidingWindow         *uint32 `json:"sliding_window"`
-		HiddenAct             string  `json:"hidden_act"`
-		VocabSize             uint32  `json:"vocab_size"`
-	} `json:"text_config"`
-	VisionModel struct {
-		NumAttentionHeads uint32  `json:"num_attention_heads"`
-		NumHiddenLayers   uint32  `json:"num_hidden_layers"`
-		HiddenSize        uint32  `json:"hidden_size"`
-		IntermediateSize  uint32  `json:"intermediate_size"`
-		ImageSize         uint32  `json:"image_size"`
-		NumChannels       uint32  `json:"num_channels"`
-		PatchSize         uint32  `json:"patch_size"`
-		HeadDim           uint32  `json:"head_dim"`
-		HiddenAct         string  `json:"hidden_act"`
-		RopeTheta         float32 `json:"rope_theta"`
-	} `json:"vision_config"`
-	MultiModalProjectorBias bool   `json:"multimodal_projector_bias"`
-	ProjectorHiddenAct      string `json:"projector_hidden_act"`
-}
-
-func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
-	kv := p.ModelParameters.KV(t)
-	kv["general.architecture"] = "mistral3"
-	kv["mistral3.vocab_size"] = p.TextModel.VocabSize
-
-	// Text configuration
-	kv["mistral3.block_count"] = p.TextModel.NumHiddenLayers
-	kv["mistral3.context_length"] = p.TextModel.MaxPositionEmbeddings
-	kv["mistral3.embedding_length"] = p.TextModel.HiddenSize
-	kv["mistral3.feed_forward_length"] = p.TextModel.IntermediateSize
-	kv["mistral3.attention.head_count"] = p.TextModel.NumAttentionHeads
-	kv["mistral3.attention.head_count_kv"] = p.TextModel.NumKeyValueHeads
-	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.TextModel.RMSNormEPS
-	kv["mistral3.attention.key_length"] = p.TextModel.HeadDim
-	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
-	kv["mistral3.rope.dimension_count"] = p.TextModel.HiddenSize / p.TextModel.NumHiddenLayers
-	kv["mistral3.rope.freq_base"] = p.TextModel.RopeTheta
-
-	// Vision configuration
-	kv["mistral3.vision.block_count"] = p.VisionModel.NumHiddenLayers
-	kv["mistral3.vision.embedding_length"] = p.VisionModel.HiddenSize
-	kv["mistral3.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
-	kv["mistral3.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
-	kv["mistral3.vision.attention.key_length"] = p.VisionModel.HeadDim
-	kv["mistral3.vision.image_size"] = p.VisionModel.ImageSize
-	kv["mistral3.vision.patch_size"] = p.VisionModel.PatchSize
-	kv["mistral3.vision.num_channels"] = p.VisionModel.NumChannels
-	// kv["mistral3.vision.attention.layer_norm_epsilon"] = 1e-05 // Default value
-	kv["mistral3.vision.rope.freq_base"] = p.VisionModel.RopeTheta
-
-	// Multimodal configuration
-	kv["mistral3.image_token_index"] = p.ImageTokenIndex
-	kv["mistral3.spatial_merge_size"] = p.SpatialMergeSize
-
-	kv["mistral3.mm.projector_bias"] = p.MultiModalProjectorBias
-
-	if p.ProjectorHiddenAct != "" {
-		kv["mistral3.mm.projector_hidden_act"] = p.ProjectorHiddenAct
-	}
-
-	return kv
-}
-
-func (p *mistral3Model) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
-
-	for _, t := range ts {
-		if !strings.HasPrefix(t.Name(), "v.") {
-			if strings.HasSuffix(t.Name(), ".attn_q.weight") ||
-				strings.HasSuffix(t.Name(), ".attn_k.weight") {
-				t.SetRepacker(p.repack)
-			}
-		}
-
-		out = append(out, ggml.Tensor{
-			Name:     t.Name(),
-			Kind:     t.Kind(),
-			Shape:    t.Shape(),
-			WriterTo: t,
-		})
-	}
-
-	return out
-}
-
-func (p *mistral3Model) Replacements() []string {
-	return []string{
-		"language_model.model.norm", "output_norm",
-		"language_model.model.", "",
-		"language_model.", "",
-		"layers", "blk",
-		"transformer.layers", "blk",
-		"vision_tower", "v",
-		"ln_pre", "encoder_norm",
-		"input_layernorm", "attn_norm",
-		"post_attention_layernorm", "ffn_norm",
-		"embed_tokens", "token_embd",
-		"self_attn.q_proj", "attn_q",
-		"self_attn.k_proj", "attn_k",
-		"self_attn.v_proj", "attn_v",
-		"self_attn.o_proj", "attn_output",
-		"mlp.down_proj", "ffn_down",
-		"mlp.gate_proj", "ffn_gate",
-		"mlp.up_proj", "ffn_up",
-		"attention.q_proj", "attn_q",
-		"attention.k_proj", "attn_k",
-		"attention.v_proj", "attn_v",
-		"attention.o_proj", "attn_output",
-		"attention_norm", "attn_norm",
-		"feed_forward.gate_proj", "ffn_gate",
-		"feed_forward.down_proj", "ffn_down",
-		"feed_forward.up_proj", "ffn_up",
-		"multi_modal_projector", "mm",
-		"ffn_norm", "ffn_norm",
-		"lm_head", "output",
-	}
-}
-
-func (p *mistral3Model) repack(name string, data []float32, shape []uint64) ([]float32, error) {
-	var dims []int
-	for _, dim := range shape {
-		dims = append(dims, int(dim))
-	}
-
-	var heads uint32
-	if strings.HasSuffix(name, ".attn_q.weight") {
-		heads = p.TextModel.NumAttentionHeads
-	} else if strings.HasSuffix(name, ".attn_k.weight") {
-		heads = cmp.Or(p.TextModel.NumKeyValueHeads, p.TextModel.NumAttentionHeads)
-	} else {
-		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
-	}
-
-	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(dims...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-
-	ts, err := native.SelectF32(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var f32s []float32
-	for _, t := range ts {
-		f32s = append(f32s, t...)
-	}
-
-	return f32s, nil
-}

convert/convert_phi3.go

@@ -118,5 +118,6 @@ func (p *phi3Model) Replacements() []string {
 type ropeFactor []float32
 
 func (r ropeFactor) WriteTo(w io.Writer) (int64, error) {
-	return 0, binary.Write(w, binary.LittleEndian, r)
+	err := binary.Write(w, binary.LittleEndian, r)
+	return 0, err
 }

convert/convert_qwen2.go

@@ -2,6 +2,7 @@ package convert
 
 import "github.com/ollama/ollama/fs/ggml"
 
+
 type qwen2Model struct {
 	ModelParameters
 	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`

convert/convert_test.go

@@ -109,7 +109,6 @@ func TestConvertModel(t *testing.T) {
 		"all-MiniLM-L6-v2",
 		"gemma-2-9b-it",
 		"Qwen2.5-0.5B-Instruct",
-		"c4ai-command-r-v01",
 	}
 
 	for i := range cases {

convert/reader.go

@@ -62,7 +62,10 @@ func parseTensors(fsys fs.FS, replacer *strings.Replacer) ([]Tensor, error) {
 		Pattern string
 		Func    func(fs.FS, *strings.Replacer, ...string) ([]Tensor, error)
 	}{
-		{"*.safetensors", parseSafetensors},
+		{"model-*-of-*.safetensors", parseSafetensors},
+		{"model.safetensors", parseSafetensors},
+		{"adapters.safetensors", parseSafetensors},
+		{"adapter_model.safetensors", parseSafetensors},
 		{"pytorch_model-*-of-*.bin", parseTorch},
 		{"pytorch_model.bin", parseTorch},
 		{"consolidated.*.pth", parseTorch},

convert/sentencepiece/sentencepiece_model.pb.go

@@ -1360,7 +1360,7 @@ func file_sentencepiece_model_proto_rawDescGZIP() []byte {
 
 var file_sentencepiece_model_proto_enumTypes = make([]protoimpl.EnumInfo, 2)
 var file_sentencepiece_model_proto_msgTypes = make([]protoimpl.MessageInfo, 6)
-var file_sentencepiece_model_proto_goTypes = []any{
+var file_sentencepiece_model_proto_goTypes = []interface{}{
 	(TrainerSpec_ModelType)(0),         // 0: sentencepiece.TrainerSpec.ModelType
 	(ModelProto_SentencePiece_Type)(0), // 1: sentencepiece.ModelProto.SentencePiece.Type
 	(*TrainerSpec)(nil),                // 2: sentencepiece.TrainerSpec
@@ -1392,7 +1392,7 @@ func file_sentencepiece_model_proto_init() {
 		return
 	}
 	if !protoimpl.UnsafeEnabled {
-		file_sentencepiece_model_proto_msgTypes[0].Exporter = func(v any, i int) any {
+		file_sentencepiece_model_proto_msgTypes[0].Exporter = func(v interface{}, i int) interface{} {
 			switch v := v.(*TrainerSpec); i {
 			case 0:
 				return &v.state
@@ -1406,7 +1406,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[1].Exporter = func(v any, i int) any {
+		file_sentencepiece_model_proto_msgTypes[1].Exporter = func(v interface{}, i int) interface{} {
 			switch v := v.(*NormalizerSpec); i {
 			case 0:
 				return &v.state
@@ -1420,7 +1420,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[2].Exporter = func(v any, i int) any {
+		file_sentencepiece_model_proto_msgTypes[2].Exporter = func(v interface{}, i int) interface{} {
 			switch v := v.(*SelfTestData); i {
 			case 0:
 				return &v.state
@@ -1434,7 +1434,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[3].Exporter = func(v any, i int) any {
+		file_sentencepiece_model_proto_msgTypes[3].Exporter = func(v interface{}, i int) interface{} {
 			switch v := v.(*ModelProto); i {
 			case 0:
 				return &v.state
@@ -1448,7 +1448,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[4].Exporter = func(v any, i int) any {
+		file_sentencepiece_model_proto_msgTypes[4].Exporter = func(v interface{}, i int) interface{} {
 			switch v := v.(*SelfTestData_Sample); i {
 			case 0:
 				return &v.state
@@ -1460,7 +1460,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[5].Exporter = func(v any, i int) any {
+		file_sentencepiece_model_proto_msgTypes[5].Exporter = func(v interface{}, i int) interface{} {
 			switch v := v.(*ModelProto_SentencePiece); i {
 			case 0:
 				return &v.state

convert/testdata/c4ai-command-r-v01.json

@@ -1,344 +0,0 @@
-{
-    "general.architecture": "command-r",
-    "general.name": "command-r",
-    "command-r.attention.head_count": "64",
-    "command-r.attention.head_count_kv": "64",
-    "command-r.attention.layer_norm_epsilon": "1e-05",
-    "command-r.block_count": "40",
-    "command-r.context_length": "131072",
-    "command-r.embedding_length": "8192",
-    "command-r.feed_forward_length": "22528",
-    "command-r.logit_scale": "0.0625",
-    "command-r.rope.freq_base": "8e+06",
-    "command-r.rope.scaling.type": "none",
-    "tokenizer.ggml.add_bos_token": "true",
-    "tokenizer.ggml.add_eos_token": "false",
-    "tokenizer.ggml.bos_token_id": "5",
-    "tokenizer.ggml.eos_token_id": "255001",
-    "tokenizer.ggml.merges": "902a060cac8884a5793d2a857dd2e53a259de46c8d08c4deb243c239671e1350",
-    "tokenizer.ggml.model": "gpt2",
-    "tokenizer.ggml.padding_token_id": "0",
-    "tokenizer.ggml.token_type": "b7a352ccd1c99d4413bcf452c2db707b0526d0e1216616b865560fab80296462",
-    "tokenizer.ggml.tokens": "815ac90ff23565081522d7258f46648c8a0619eb847a9c7c31b238a9b984e4ae",
-    "blk.0.attn_k.weight": "6fcfdb466f9ceb1229404ce4ec4e480751b8d00da12707a11783dad7256cb864",
-    "blk.0.attn_norm.weight": "6063317f731371864049c7704a70772f1eb632194201ebdc2ed0f8e483507c72",
-    "blk.0.attn_output.weight": "920f49716a1e2fc73b6794ec777947f1c122701e63ed302422ac89e90f06e9da",
-    "blk.0.attn_q.weight": "ddbcd7cde197e632564ac58e4f25d9e3a8ca52917329eeb6081eb41a797932ab",
-    "blk.0.attn_v.weight": "318fc02a189d87420f0cbf57f47f11e00c21ec1ed472ce0a2a895b44f7fa0fca",
-    "blk.0.ffn_down.weight": "aa71975b6eb1f4c77b03d2ac4a194cf8d95718efac741bb12f0f3ff79a27f9bc",
-    "blk.0.ffn_gate.weight": "42967702fa0bc738b88dc50007ace26dbe74a5a9e0978124dd093f818241a9e1",
-    "blk.0.ffn_up.weight": "5282c8788b086bd30f46525e7995a17464882a72703fd27165491afdd8bfd4af",
-    "blk.1.attn_k.weight": "cd248882e64fd2c3402c44790ebe12440133dc671b6893fdad0564c461973adc",
-    "blk.1.attn_norm.weight": "ba84e1c8fd30af6ec94208db4078befac8c921aad3acb887812887f3282ea2be",
-    "blk.1.attn_output.weight": "2efa3ef7c5666ccceb05e339b83ad680cc0d2c3ec78203f5da5959f23a80e14f",
-    "blk.1.attn_q.weight": "5106f2e255358a1303c22e8b5f0ec044852bb30a866c52cabefd30017a7a6b7d",
-    "blk.1.attn_v.weight": "a211a634a1a5df1d5f973645438be0461dd922210f9747c6b04e386c7f1ebe95",
-    "blk.1.ffn_down.weight": "37093afe48d32c578ec956c9ed85242cd000d6aa979e60526aafa10c822dbb10",
-    "blk.1.ffn_gate.weight": "469860819e9159caefb1aad0bc66db790f3393f05fd87b08e52256a7ed256543",
-    "blk.1.ffn_up.weight": "736742c97d35d1a011f9cafd3c0ce947ad559bb2fba6da73c816f6bfd0fa9aeb",
-    "blk.2.attn_k.weight": "92c219d92804d832ab404bd6dc7339c90877bb7cf405dd030c121f8b27757739",
-    "blk.2.attn_norm.weight": "61e4466069474b76b6d1e702566420eb669faf3556b00ff7b824784aca13a2d6",
-    "blk.2.attn_output.weight": "d2fb38a2b2171fd91caf037faa585a62225819aa232d86fd4f7f9d2c3c8a45e9",
-    "blk.2.attn_q.weight": "f6faf5cc6844e3daa4f9f68d90f5458c64879de68a7728860e38374e30c3429d",
-    "blk.2.attn_v.weight": "f340ef8f7341d987a6f37c0e9afe0aef5be67be00c0ce5f57612daf73319cce1",
-    "blk.2.ffn_down.weight": "c7be61a701d779860b621b143fb6365b607bf99ec7c0f153b07908ac8120885a",
-    "blk.2.ffn_gate.weight": "b64f0878187bd3392abfa4c3e8ad2f8b4c133903e54246747ff8f3b4639ad83e",
-    "blk.2.ffn_up.weight": "50b11c712652e90ee7428dbb45cffebb80662ac982bc72bd9eafff361b5eb5a8",
-    "blk.3.attn_k.weight": "2b7bcbe9ee5c9c630c8c8d7483887e78b73581016f4cbb6933db2a147a25f431",
-    "blk.3.attn_norm.weight": "0181dac7f4eee7252980323e8032cf339bef2046ce0a16c0fd72af7c98a8a37b",
-    "blk.3.attn_output.weight": "aef8843b636ce231da9e7c9acbee197883cc15df0e2887709324c6a50f16da7b",
-    "blk.3.attn_q.weight": "55404130fa10e81322d33eb378aa0de31a92990ce7730f1338c0ace0406bb1b1",
-    "blk.3.attn_v.weight": "76f7fb8040d82b957d689ce34fea2302a6640ad5bbaa0052ad2b7ebce270c33d",
-    "blk.3.ffn_down.weight": "648628933eff3b357c3729c33c5b1ae51c28e59b9c19acd1601a2ff7c5d5d9a5",
-    "blk.3.ffn_gate.weight": "6a588885d16e98d5f50ebed05af089154f680085ca9c97691e5b489088630a4a",
-    "blk.3.ffn_up.weight": "e12455a1d702f4986e1a663493e3d5102b367af74d45557522002a35d63ecac2",
-    "blk.4.attn_k.weight": "40d943380a8a85e4eab147934bf6e16f23cc8ab753f6636526382c074d182288",
-    "blk.4.attn_norm.weight": "4ab2c098983d4599fe540eef624c4df954adb7473faebda7471ef0ba4134814c",
-    "blk.4.attn_output.weight": "d14b91e40f58bf4a3c8c2eca0b12bb541de406574af39027d56f6c588a147082",
-    "blk.4.attn_q.weight": "e1224960a3562107488589f883fa32414bae41712fa8dbd47c5f3e3a7801452f",
-    "blk.4.attn_v.weight": "063f297bc4aa6e709fc32c4c32e35af7d07d80e83cb939b76adbba858006c03d",
-    "blk.4.ffn_down.weight": "f88a18020c5e1caaa29596895eb348e76ee5bfad27ed57651a86cd8cd1f9b5aa",
-    "blk.4.ffn_gate.weight": "48e7e1eed3fb52e92e61d3557dd0ec002418327090e034ce4322fd68542266f8",
-    "blk.4.ffn_up.weight": "1ca8a7aa17355b6ce0d9ad5539fdad3899fa47fd359c285fbfb31f19f47bf073",
-    "blk.5.attn_k.weight": "2bdf15f8e73d068d972380f25d207004cf0bf3b5bfa46946803ba6fba07d9175",
-    "blk.5.attn_norm.weight": "60448d7cde6e1b6467aa31bdea012e39cdb08c88081cee7d102dca4f93f766ef",
-    "blk.5.attn_output.weight": "f9f687d7c457537f9fca8a4087a59f1c3bebfaf5537b94e42c831a13224f7799",
-    "blk.5.attn_q.weight": "987db7a2ad68657a92625e1980effbb1f79697c2183f2b9f3b3a0570c51b0ab9",
-    "blk.5.attn_v.weight": "cf696891148f3e4783ad1d20f93462ae091eb8651c656bba9b662253b6263e02",
-    "blk.5.ffn_down.weight": "c0662b0bd0929136005fb9d691fdd9b2c33867d9ce9622339a6a456b720b059a",
-    "blk.5.ffn_gate.weight": "200bbdfab615d7a3a84719b6ced7751e3ce52757ef212d96f87798bc1de5e987",
-    "blk.5.ffn_up.weight": "df5d23e7e035fb1b9d163da7ddfdfe38da6a37e86e96534dc02ad20f011b55b3",
-    "blk.6.attn_k.weight": "c0dae2d272a7c5a2fa004bbb8475dbab362fc1f6d008e73d5a4434a9382ac6ba",
-    "blk.6.attn_norm.weight": "51c57ac8b55e04354d5dca6bb9c0cf4177639d3b038e80209e33036209688f64",
-    "blk.6.attn_output.weight": "229d97892c62f85bcdf431675250e01c976ad69ffa450b01fb543bf88f14a2fb",
-    "blk.6.attn_q.weight": "c20e49621821bd46ed156e6823864a5bda4f317750e71ab8dc54e44eb48cf7c2",
-    "blk.6.attn_v.weight": "53ceb1a2ee43fce3c7b5b33c58a9fc5ee7f44dc1c6f29bc9dbefc37582102dc9",
-    "blk.6.ffn_down.weight": "7923c943b7629d560a032d1efa210d1d75c6692140f1be94464ee7ed24f44ed0",
-    "blk.6.ffn_gate.weight": "57593d350361af753a6a39f53b066282634c0fb44f396f6f2966a574b01d8f8c",
-    "blk.6.ffn_up.weight": "327b6a7a387098b8899d3ded04a4d4e7c658ca61b80d4e7b17594be232721602",
-    "blk.7.attn_k.weight": "9ca48b87a10116fd8868e62b76f211d4bb91f166096be9061439ee2e1c3a5c20",
-    "blk.7.attn_norm.weight": "cd56cfcc4e2ad6b96e23ea7b0d32b4caf236107d99a0b22c56760b62e63c8cfd",
-    "blk.7.attn_output.weight": "7352b509a03cae2491ffc060e577d189341a0f861233f18c96f9d275dc4234bf",
-    "blk.7.attn_q.weight": "2b3791c8c008c33ddbe12bedba8191322ceea2dcce5cf0eb7a93d40ad254e672",
-    "blk.7.attn_v.weight": "3ae721d52466487a3d48150581e57f6d64ea1e83ab929f23b28c3d777422eeb6",
-    "blk.7.ffn_down.weight": "3b6fa8ececdb3c34af3a5363863d6f94289c1c95bf47fce3a3ddcf184c5f0848",
-    "blk.7.ffn_gate.weight": "dbd7df6c5ae5eb4adb859f0d36453813a4e289a359a1ba8f72d67fcbf21c3e22",
-    "blk.7.ffn_up.weight": "de68380a334b4c5cfd4c318b0e9854aec59bd79aa0f0c30af3f56414f83482b0",
-    "blk.8.attn_k.weight": "7303c4e4480abc72a7ee271811311199245fb5c2ea27a2bd3b8cad3a53a03c27",
-    "blk.8.attn_norm.weight": "2e3d1921898d1b943ce1a1b6818546c8b471d6d542da24f51a8b514b8c3dd4ef",
-    "blk.8.attn_output.weight": "30421520887b66bf97a18dbcdc283bc8d0b60590b612fd638a319a6eae923227",
-    "blk.8.attn_q.weight": "73e064d5433c9b500068a1c31744dbd53f4ade298fb450a0e8c97f62cf1f8a8d",
-    "blk.8.attn_v.weight": "27e21f8b9a9a8533e8178ca34a72aa1d786393d57302b7806dcdf3e51de511a8",
-    "blk.8.ffn_down.weight": "bf694bd8e00047982108000e7b3dee7b225db8b19abc595e5697b6bbefd92e7c",
-    "blk.8.ffn_gate.weight": "d55fdbf8606d9141b774b0500c58944fd1253b9e69d1f765eaa9a680b9f2ca40",
-    "blk.8.ffn_up.weight": "1ae3f580655e7c8e8dd6c34fa4ac574fdfc5e3f1a8536da0c5442d3a2976f0e7",
-    "blk.9.attn_k.weight": "b18080626012d8aabcf78542d6c7bf31c712bf55a70172fbfe173fcf34481036",
-    "blk.9.attn_norm.weight": "2e3620620dc09998c6d3063a7d5de5433fbbae8c11e5b00d13f145d39140e162",
-    "blk.9.attn_output.weight": "69c3c0e27ef1c0fc933eeb7b612b70909f18cde238873c0d576a2ba9714ef174",
-    "blk.9.attn_q.weight": "68330e5aa28a28873c9a6e67f032186ef651df2df5844e0f27094ba349fbe4ab",
-    "blk.9.attn_v.weight": "3df8d45a102be082d0793a51cb82aa62a43cd0e9d047ba4115ca0f2414b39325",
-    "blk.9.ffn_down.weight": "1d6cc162b73745b135b4f040a0aac3c06d5135a3dc5b2421e7ee2af48662fd7f",
-    "blk.9.ffn_gate.weight": "034a9d40fb1e32b534b45f4bccd65cbe43c4a6a3f5d01132bd245ca0005de5fc",
-    "blk.9.ffn_up.weight": "c838c38d0e1a0ac0da17eb2a66023ed31929f07d8fcfe1cc546df26096c91f0c",
-    "blk.10.attn_k.weight": "a78507cb72f744b86ceaa032596e74e5571c822d0226d334881169addb32cbd5",
-    "blk.10.attn_norm.weight": "35f48d0b28ee0e6b4cad4e983925737562d64824be5b168b3e26df3d6b260cf1",
-    "blk.10.attn_output.weight": "53712db06796de39b131323e7abf9a58551b6d52da6db66a471580386d396252",
-    "blk.10.attn_q.weight": "efe08429ba196026b81cd1c471e1c7418afd9e966659feb3936b674aa0803b58",
-    "blk.10.attn_v.weight": "7ec6055e134f89da0cbe79ec9f13ef2e442ac584b1f03c3e13e7d0cdad0078bd",
-    "blk.10.ffn_down.weight": "37e66af4bcd1f3079e841e892255b8255070655901864ea3a8c602a7f681a640",
-    "blk.10.ffn_gate.weight": "1825282bc34830d371c6edcc3c1e73e6ecc1e10f4aea0122dbb7acc1d6f7b1bc",
-    "blk.10.ffn_up.weight": "819b3b276a4d4c14a35ed6682d5ef18a5e8ed468e5ce3f12e8c75ec18ac20ec4",
-    "blk.11.attn_k.weight": "5327e6a2af82dfff0619a14971f5864a15553c36fead84e1af42c7630f2729c6",
-    "blk.11.attn_norm.weight": "fec363b3c4a43036d2c635fb8aa9e122dd87ee79811839f2f6cd955be3373e7b",
-    "blk.11.attn_output.weight": "ccf7b38f18ee8798b8a6a35018e2df3eb3e007de62876befb68025dd66c79763",
-    "blk.11.attn_q.weight": "da8c4a1c824ffe174e39f126cd72f7ef83c56aff1259d452a1212de80f98f5e9",
-    "blk.11.attn_v.weight": "d17ae6bb77f03982b55d341eb67acb5969e9ad3da5994b96eafc09793dcfe3a0",
-    "blk.11.ffn_down.weight": "a6bac521e2791345f22c57205fa1c2f2f687794dfd24d0e98d50ae0d0eb6088a",
-    "blk.11.ffn_gate.weight": "5ed902c488cb51ba5635f3df08258c5f84f31a679a00211ea5f9d8b824ef6d9d",
-    "blk.11.ffn_up.weight": "ee9f1437eb890d2cf9df2574afa1cecf20aafdd847cd75b152d7eb74419afd34",
-    "blk.12.attn_k.weight": "5a069c06e1019b0f889088e67458f7a11ec77fa190ada6069e46211f62219947",
-    "blk.12.attn_norm.weight": "194d7e5fcc8c49aea62daf1940532419cf3c505afdce6be377286b677db5db8f",
-    "blk.12.attn_output.weight": "6534995fd4d6fecb55e317add4b1723aba4d825e1e9471d0b08813dfdc247176",
-    "blk.12.attn_q.weight": "4ab51ca519b5995581fa34f846276feca3b907ef2b51f192f6cc0b3263c3f5a2",
-    "blk.12.attn_v.weight": "5652ca3fa81ef9a1ac1543d71fc6813f8517f8ec54b25c701f6f98061614830f",
-    "blk.12.ffn_down.weight": "4b2c263f54c88516b8eb273bb8d9615b01c5c8b484dc70358adb91b50b300edd",
-    "blk.12.ffn_gate.weight": "8f50c3c3e3e8568991d6c1b0e74b500cf4f208e7700bbb8e87c3f6a6d359b6b5",
-    "blk.12.ffn_up.weight": "1c1a581fec1fbe959e1427fa513f400100b5e1ee9d83932630be9905fb49c231",
-    "blk.13.attn_k.weight": "efd7a38c46f08d8376d82974f33c644e3a02220e142d63b1704718699a8a884c",
-    "blk.13.attn_norm.weight": "d28fa4f1bd75abbd063b0e622e08f579c89cd0c0c5ce63c1952ec9f944f8ee13",
-    "blk.13.attn_output.weight": "71e0068a639288718bdb70a6cfdefd50bc8b3ec3993347a65129e70001ca5827",
-    "blk.13.attn_q.weight": "b97077adc92cff07a2e07d80ee38f214ad8713571c69cd5c70ebd43dc501ac87",
-    "blk.13.attn_v.weight": "79b3e2749ab4b459c81e96e322b215f1e8af645eb346e176c326bd00cf6ed2fd",
-    "blk.13.ffn_down.weight": "9f8687d11effa1db7cfecf7bec5631734bcf2962aad74a9f519144491e08ec85",
-    "blk.13.ffn_gate.weight": "7d14dfa0543852e7777fe8fff29ca533744cbcf1ebcf10067e5adfc4eb345e65",
-    "blk.13.ffn_up.weight": "852b9527b97fdab211ff3f832a660ee1d93ccb56906144c50f01319a6e8ee615",
-    "blk.14.attn_k.weight": "79e926b20f36f66d58226cb358881f2f68ae7b468787d33cafae5110287a14a0",
-    "blk.14.attn_norm.weight": "97d481b63deb0df6142c2c6cd23043720c62eb609e390f47a7113751c79974ec",
-    "blk.14.attn_output.weight": "aa6e94d7176d5c79fbb89b96e5f13ce75702ce3dd23ee52986446da436a6c3d6",
-    "blk.14.attn_q.weight": "214becb6d1bb460da9fb8ace0f99b9a5afa9edf7aa7acc19606c7401b11d6305",
-    "blk.14.attn_v.weight": "488b0e6d7f1a7a2ed0972aaa6d10ef9c775ee5373460324efcf5b3e3da9311df",
-    "blk.14.ffn_down.weight": "29c7ad16cf9542e30996a1a01ab95b844533b28051f04cc7949c371afb796471",
-    "blk.14.ffn_gate.weight": "b7ef208f2b054803665b377f5a5980c122c026841809cf855c6ba06d1c3a885a",
-    "blk.14.ffn_up.weight": "76a5cc28100748d79c4398ce7b9176aab4d661548b6293a82f99144812e5b70e",
-    "blk.15.attn_k.weight": "a6b8f9e98ab878fa7ebc5d080978ebf2d050acc2ab2fa8ea9188eb10e27702c8",
-    "blk.15.attn_norm.weight": "a26d07a9752d6dccb68e3a8a2a49fd0752cdd0a415e05547819bc37d9ba63d5e",
-    "blk.15.attn_output.weight": "c63616c69048ccbee801e05be4f56d21fda21aa0cc470f41d57c31b4d9283a4d",
-    "blk.15.attn_q.weight": "fd595a67bf96c6ba16eb148a9d02fa52fa3c1d33ed10be28a08f851409fd6e64",
-    "blk.15.attn_v.weight": "1c5c9d33fa07c05d5f4ed0032c6c4aa83d863f0d31c94a66109d239dcd03cea3",
-    "blk.15.ffn_down.weight": "585ea62ab8aff7d7d212ea5c1a03226fda6b68370c890b776834af70c948dcbc",
-    "blk.15.ffn_gate.weight": "a13c63f86f879b03a573d5dd2a25cfd1f4dc73e8132e6454ecc23e538b4cdf6f",
-    "blk.15.ffn_up.weight": "f7112450f57c12fcd511f049e0dc0b541625a107a7901c3261ed9e984299f65c",
-    "blk.16.attn_k.weight": "2d2c8b11dd71fba6d1c106aa1673c113a5448653cca7eab897c8739212ed5003",
-    "blk.16.attn_norm.weight": "95c2ec7be9469690e18a9a1779684acb3e9da44b13e263a0da840305646fbf8a",
-    "blk.16.attn_output.weight": "31a65046e677f54dae654ded4e733479fcc0f7283d83076b7dc7cbcae8528230",
-    "blk.16.attn_q.weight": "bfc6292b9c6d49b7118d08060242a138182eb182d136ba5dfaf469437c16081d",
-    "blk.16.attn_v.weight": "68f81d037340217d87c7853ff4d6edfbc46d9e827ee6d5bff7c3f6238e3a95ad",
-    "blk.16.ffn_down.weight": "bbd6629691950cef4d5113e1c6670e91b216a9b872cb92cee02dfda4d6c4f7b8",
-    "blk.16.ffn_gate.weight": "63cb56f282b7401ed6c76e5bb6fdf1bf68a64f9af0c82c014209b55bcb5191d0",
-    "blk.16.ffn_up.weight": "b54f39a2541063cbfb6f713aa81c3b69a04100e999aa2ebbeec195dc382eceec",
-    "blk.17.attn_k.weight": "3d9ba49799cc56664ec30a002bcad61eb651294212a68c3ddb573eb042aef5a4",
-    "blk.17.attn_norm.weight": "42ee0db4b9d63257bca0012a30b12737ead1caafeb5ed3d93c8f48ffec4b46de",
-    "blk.17.attn_output.weight": "a38fd100f05c9041c592bc739e287de0b10d08ef2bda41a879225bdca9002f71",
-    "blk.17.attn_q.weight": "8a3bee285b0180a9eb35662e449ee4cbe16d992bdd48fb3a94bc4a347728cfa2",
-    "blk.17.attn_v.weight": "d7f8f1b8b863494ed4392a1656775912e9b264ad36016547b12e832a1d6757d6",
-    "blk.17.ffn_down.weight": "bb7ee58f61da8630972e25b621996fbe8ec06f4dc9ab1e268ab5b120c526ca28",
-    "blk.17.ffn_gate.weight": "6b652dbf167fee09a45ebfd78d500ff6548fb2756dbe5343ffec3f7e6207179f",
-    "blk.17.ffn_up.weight": "3b67f727e55e742715de978fab80457781e7a3762bc48f79d13b45dcb8de664c",
-    "blk.18.attn_k.weight": "ff7fe57c57b90c6fcc0aefc39ec24593c3a7d1ea1c23770480075a015450e0f5",
-    "blk.18.attn_norm.weight": "1d40faca082d2633ef0ccf19e121870dd6c7c3e2154607c7f3543fa96e99cb2d",
-    "blk.18.attn_output.weight": "9adfecaaa397a92db4687efd5fcabfa0daef9e6b0493763b7ff5ebc185c43a6c",
-    "blk.18.attn_q.weight": "ad1803eb9b291948639277afe981e666b07167eb3fcae903ba5b73bf86d8f50b",
-    "blk.18.attn_v.weight": "308cf23399adccf27401a4ab60d74dac6fb9d4cd4b9c5940d9145118d1881b34",
-    "blk.18.ffn_down.weight": "7de4ac9a561fb580619b745687dfd7ca8a69ef70471dee978741b80e9ff7bead",
-    "blk.18.ffn_gate.weight": "0c66970f696b33bd5ee8f1f2fbcb41fd78fa5ccabdc927e11a4d5a4089f19c69",
-    "blk.18.ffn_up.weight": "66a42e988e8a1f468fabf976c48e9e4bb045eaac6916ef16555ac101cd674abc",
-    "blk.19.attn_k.weight": "a928ab50390bacbcebe2e4b66922498134ce22d7b93beaa87d6cf4ab52eb7174",
-    "blk.19.attn_norm.weight": "b4a02c55b46c2a96aec9c64a254087cf48e6c1d4b6f31782c77a46fc4daebad1",
-    "blk.19.attn_output.weight": "b768319c641dff1eac5d1f8ceb960c9899c795bf2b24c1d6bf70aa24fda45f77",
-    "blk.19.attn_q.weight": "79ef3f57d187d3954a26362096e1b6c222d76f537dff73e034d6e9999935b8bc",
-    "blk.19.attn_v.weight": "ce13d6b13e24fcb2d5bc6a2662e5bd295b31b12db10a6d0307f86cf29b8d5001",
-    "blk.19.ffn_down.weight": "cf90d7e2137482cfd50934a8223ad774621d08554969da80a9712df5e6227eb0",
-    "blk.19.ffn_gate.weight": "71ce30150f003b6eeb3bf7464e05b6ae615f135110d8e47f0a47fd973e537c0f",
-    "blk.19.ffn_up.weight": "7f92aca0cc29866633feec701ec01a85a8ee2fd4e2b9630173a6cffb1d9d50ee",
-    "blk.20.attn_k.weight": "a2df23159d6fb74ef28e14b61028fe8b00a693a2fc9234a980be74f20b958682",
-    "blk.20.attn_norm.weight": "c6cd5f1b096fc5efa4eb59ca1c8c4bd28730f3dcedd59a63601663eccc6724ed",
-    "blk.20.attn_output.weight": "896a8a166d0f006d4b09867ae4345426303cbc3fb13a18d3d4e1bde00f16dbdf",
-    "blk.20.attn_q.weight": "01eb79588fe61baea0da43e99f4dc5939590e1bafd01e12dadb8326f102bfea2",
-    "blk.20.attn_v.weight": "bd39630fdd5a7c859ac1addaf53e63faf524c3f32f5f4896d86b6e746b1d5c06",
-    "blk.20.ffn_down.weight": "0304a5d39957a0e3f031c4bcc4549a135d396c8d97c8d276fd1c823ce86560c2",
-    "blk.20.ffn_gate.weight": "117b79d595b1dca0c8b37586beaecc4d84411507276212dc286cde7fc36c9bef",
-    "blk.20.ffn_up.weight": "6e799346db145c125f01783539749d3828fcc451cd4f10c5352f047a47e28714",
-    "blk.21.attn_k.weight": "1c37e4c0664147e775bb006b226b9553e3421140cd96288ea755f81731ab80ba",
-    "blk.21.attn_norm.weight": "00ae783a29000ccda5e4bdbff03df0752fb82805dc3f9b987500ebd80714476e",
-    "blk.21.attn_output.weight": "7588b84f9fb19f15095b5265c60b4a4e7ae74bcc47d4607dfa5d0bfab6f136cb",
-    "blk.21.attn_q.weight": "a65f1c0dd06d45bb97532d3e932689c1eecfe7359089b39174a96a149335cbc1",
-    "blk.21.attn_v.weight": "4220b77e7d5e8709b4eef33a679b5dad11f297085ef44c9977f9e54ef08f7a2d",
-    "blk.21.ffn_down.weight": "b8c082a0530d4b5328e67db0df84c5498f2af956de23c639fa0198ffea853950",
-    "blk.21.ffn_gate.weight": "cd1b656ee72d00e9835ef667c19ef89a88de261eb8eb7c0e936e0f9ddf83ef9f",
-    "blk.21.ffn_up.weight": "dc445f73e36ec7a3bd86884186b728f8e0187f32848c3b8b69d4d41f8571bf31",
-    "blk.22.attn_k.weight": "e37cf0b893ec8b9ee8c78dd139b8d9c45cb997a3bc0c3d93a70ca1c3f6af8859",
-    "blk.22.attn_norm.weight": "248a27838d3c46cc03a5c312facc84e2e0e2c990ef8401e93da25918497f88d1",
-    "blk.22.attn_output.weight": "fc191a18f6d18332c66761f7ab28008bfe295dd1f5c8741a2488442f9e00d0f5",
-    "blk.22.attn_q.weight": "4b193a2ab8bc2b085db18f2bf3eeba26e02b537b2cdd738160c8f14b165d0f5a",
-    "blk.22.attn_v.weight": "7a60ce5ccac7e045e55ba1e1e85bd2a0f93f8c781daee96c5223665e22f0c666",
-    "blk.22.ffn_down.weight": "e0a34fb4244e2c7168f3dbaa1904c15d339ec39999cdf27128bbaf619ee0a237",
-    "blk.22.ffn_gate.weight": "8bac872d4b8549c8812f927efa309f1792b524f33601095fff61b826de5a5615",
-    "blk.22.ffn_up.weight": "b67fa2b94dd901b6ec64c0853ce8ca2d86fe9cb1cc6d2f15fbbbe0e691c0c648",
-    "blk.23.attn_k.weight": "2c32e66ad01942b819ac09a197c71579fe66f02226a264fdd72ad1e02c67a27e",
-    "blk.23.attn_norm.weight": "825fdc94deb439cb93c713eeb077c1052b90ed658d6d464fc4ad3d611e911d48",
-    "blk.23.attn_output.weight": "95ca6707a95b8750b0c7c5d379d368f0f2e7ebef631954e7d4d8ec0f41f13a3a",
-    "blk.23.attn_q.weight": "6eccc84faca5fac015d1b26e2854501edcfd292a302228fe14cf99f5eb59a34b",
-    "blk.23.attn_v.weight": "b343ac3d226040f1033ee049668aa1d89b1774bc18431965682e5dbdce78ccdc",
-    "blk.23.ffn_down.weight": "9fc599befea8d3b1e342d564a110074f66d2542df406c4b90b6bdc5828fbb2b2",
-    "blk.23.ffn_gate.weight": "488556c1b0c9f0b20b0c99b4bac2e0f4046b81edb601d7b91e7e5b3bab47d667",
-    "blk.23.ffn_up.weight": "1088e291d7008dd9c7c2dd6830af686a8a84b724d123a016209bd5156d6898f1",
-    "blk.24.attn_k.weight": "a923fbe35e61e009a53927d7828818e0592bb737d6a1106c4b0b5a1efc367e07",
-    "blk.24.attn_norm.weight": "9b51aaaa939cefafdd9b13a7e5b74ac7fa2d603427e55a16a909d6f3f353750a",
-    "blk.24.attn_output.weight": "1beb2baba56f8409466434b037771248c2f620ec5f53e15f44c271d5a2d9ecf4",
-    "blk.24.attn_q.weight": "4b0194fe5bfae0c6bf6131dcf8cb6e2b994f6ea10b27cb03574f0f4f8cc0c950",
-    "blk.24.attn_v.weight": "6ac34b1ab0f66226d85bca1194a7c212cd93d384ecbc8b8395de48aec0970a61",
-    "blk.24.ffn_down.weight": "5508f74cb732a662c2936b32ac5e90742d172b9f961a747b0e5cba0e5906a89d",
-    "blk.24.ffn_gate.weight": "095e39b8584403835f9bb1ac33e0e81f54175575e4800273d281b845bff381e7",
-    "blk.24.ffn_up.weight": "2d43ec21637dda12973de367b0113ee9840b0d815bf6fce042f7c3f270b0b530",
-    "blk.25.attn_k.weight": "9e2aee029f3d2c7f67dfc7926e72c8228fb978382c8e5a4701bbf82c93801419",
-    "blk.25.attn_norm.weight": "220cd7164fb4cdbe22d26058e4153b26c27c7b5ce2bec8e95bf2c0ea08d23103",
-    "blk.25.attn_output.weight": "a17f4a5dc6aa51f03dbd75602d98e9491767c205cdc2c3a5f8667fc54bbf7c64",
-    "blk.25.attn_q.weight": "f60827496835c440c794bf57ce9780704d10a59d8229886bf75ebb18900ba4ef",
-    "blk.25.attn_v.weight": "9cac217e9e9f4f4c85f14ee51165a77c580165bd4a34b202389169bbe61a1ced",
-    "blk.25.ffn_down.weight": "a0f36949b663e80849581dfb71e7babcc73580793bbcb0c80ab26d5a6e000359",
-    "blk.25.ffn_gate.weight": "df4d1be4d50d6afe5ad3ef0d0e0fac76a33e85c963dea769641d612dd53e7d13",
-    "blk.25.ffn_up.weight": "992da76be762632e25ebc5ef4d03728eece1b43f7c4e31827df19ca724aea694",
-    "blk.26.attn_k.weight": "34199ff856ac32a500c754539d070258574192a34ecba87a182897cb59fdff52",
-    "blk.26.attn_norm.weight": "a8e9dfb2dae5d22b5c0aec5f3675991c0e3c3e6a44153db2579136b73f456e00",
-    "blk.26.attn_output.weight": "1c4f257ffb0d7db0f11cfb275e38b4af736917b43ad82de1badce3f1d227da4d",
-    "blk.26.attn_q.weight": "33d55786274c2e718cf61e8fbecf3dfa5ee0c208f0b716d42b061f55459acb3c",
-    "blk.26.attn_v.weight": "684b636939cd4ffcfec5a6238a0790ffa43d853c95783af9b9e8275e74071a7a",
-    "blk.26.ffn_down.weight": "89d0bf066db154e6d312b5433aed1714f6a28b40f4c52e3e1530ee07703303c8",
-    "blk.26.ffn_gate.weight": "393d649bebe5e2940e1b043649f6c860b4b8b9f380f30e9da1744a830f358156",
-    "blk.26.ffn_up.weight": "179edc85ababd9d8440cc6093eecd1004290aa1cb96434b26ecf7585b6cca17b",
-    "blk.27.attn_k.weight": "334841445a7f1e14731b08f56eb0b1f0938c63823d28bc6d078c4c5f05b36f19",
-    "blk.27.attn_norm.weight": "57344471bbda2e9deffdfdb2dd05a07aa47f8761e24de53525588639145bf551",
-    "blk.27.attn_output.weight": "506126af9ee54b535d49f97e36f630e74834f480329f098d6d62e96246d8d65a",
-    "blk.27.attn_q.weight": "dd984df1acb4783849e25ba7ae378bfd385cd9efc540fb798cd5bdd873f0118f",
-    "blk.27.attn_v.weight": "b4b3fe9a4455d34c297ff20a2f537b647cef424741d840a747b265f23d320ac0",
-    "blk.27.ffn_down.weight": "621fdb185ba0d35ba5476dae73d2c81ec1482a0e878d5bfd5c3b29fe837af013",
-    "blk.27.ffn_gate.weight": "e4fbab45f2ec506fa374103251a0bdb7baa6f576080bdd796f3e9db92098e08f",
-    "blk.27.ffn_up.weight": "a0c57e463e988002bbd6a6c6792baa21a65e6f89ae303a2c301951b0ae6e4bbe",
-    "blk.28.attn_k.weight": "bac36cbd52ec5056841663865e1291ddab4b47ef9a2544dd285d4503bfb0e4a0",
-    "blk.28.attn_norm.weight": "5774a9df2bbb2e86d1f70179c7b92d81e1f401160148b3328fb64db6646a5425",
-    "blk.28.attn_output.weight": "e8712622d1569557000c75f26c3f55fad267fd300463c2c2cfe3afbfa1c8f908",
-    "blk.28.attn_q.weight": "11677751fddee52cc739699c02836f7be54d96038be4240be5d4f53d00161608",
-    "blk.28.attn_v.weight": "e5ee459b8958d65e1445997b9aa1e90e2f5d17761ebcf5357313119a45322507",
-    "blk.28.ffn_down.weight": "3934518f9f85292da8475fe38a8edcbfc4e24ac56c351b472d6351f98750871e",
-    "blk.28.ffn_gate.weight": "6ba735d57e98d0847e487f25ffaa25256deaa8abec76f428cb70bd9774279d83",
-    "blk.28.ffn_up.weight": "977fae6e1e5353114fc645dd98429464749758765cbc6e6457593d596e57850c",
-    "blk.29.attn_k.weight": "8122a457307d580ad6f1e0acea09a2f593d97f595ba0d6737f5fea16d2433642",
-    "blk.29.attn_norm.weight": "d626f721e05aa1202439b01027031d4caf1adace61ed37870a277cb6297c77cc",
-    "blk.29.attn_output.weight": "7fb7122ab1b6b1e6615ca746897da27bc52c92cb70d3147183cdde61795b72b3",
-    "blk.29.attn_q.weight": "be43e94ff6b6e391024dc824101efa0ddf4005d5b002ac26cb03765c0c73c2fa",
-    "blk.29.attn_v.weight": "af93c85ebff908f74f9935b81bde0516ca487c84139868a1ce079c3ae20036b1",
-    "blk.29.ffn_down.weight": "39dae12340ed3120bd19c495fe0872b559613641e41fde69d02d8631900b84c0",
-    "blk.29.ffn_gate.weight": "36fd482439840ef197c9f3b8905d86acfcea49bcf018544106ca465d4bf8d5c7",
-    "blk.29.ffn_up.weight": "5243fbdfdc1e2a1dd84b6210a9869d18a014db9088897e345240cdc99990bd5d",
-    "blk.30.attn_k.weight": "948f263616bd3788b2b968baafd69b9c5bd1b77578665f096c4b7e247b4cea42",
-    "blk.30.attn_norm.weight": "e168df981e744874ff303faf2eb470e5f6868c2040ba5f383f6c5148669975e7",
-    "blk.30.attn_output.weight": "4cf0ccca04b792573b756655a24fc89cfb1f272da8305633f0bc66ef14990b93",
-    "blk.30.attn_q.weight": "21e07d6cba6c50d65350289258209717174a13c42be57e8141d69712cbaf32c1",
-    "blk.30.attn_v.weight": "65a8ca29c7237b3182ccf03e2fc94e84f9a53d0e160fb679ab401c853170dd9c",
-    "blk.30.ffn_down.weight": "8b00500a6d00d84058f6658ee1d6f06fb4fcae2f90d4341792259362923b3c13",
-    "blk.30.ffn_gate.weight": "5bc0e19ab7a31b50ac2118ad1b36e31055271a322cd8ff661d47c3ac0210703c",
-    "blk.30.ffn_up.weight": "f37a0561955725bd59ee2d064fa9f4e00a12a1b620b624db3bc3add5330bc321",
-    "blk.31.attn_k.weight": "9a5663edda227f5d87533897146764f8e8a7481b9e71fae197c39204f8463221",
-    "blk.31.attn_norm.weight": "060a4f438a1ee5e220b5b5278ad2f5c085a428bf38c515766781815597c87529",
-    "blk.31.attn_output.weight": "6ada5d3cad9dea4780ffbb43302bb6ccc2f24eddd0fc4f5f84c9ce0fc0c6e5dd",
-    "blk.31.attn_q.weight": "bb5d08c08603907981ad388d5d8b70fcc9b98034ba264b8474c8890cc0297af0",
-    "blk.31.attn_v.weight": "e01b4252ea9c6a889c32b21144b441a347464d04536ef4f6572425be55759796",
-    "blk.31.ffn_down.weight": "8ba4d679c36e93ba65ba03180385ef35ea86b3b7cdf2fded9df59369f1c09630",
-    "blk.31.ffn_gate.weight": "e5b41dc93645f8b5e8eebae3ada3ea43a18f97ce2654228655170b07b463ccb0",
-    "blk.31.ffn_up.weight": "25b88cdddc8b547af294ed107d3d1312e90b983cae87936fa6062ecd8ea02539",
-    "blk.32.attn_k.weight": "4bcf86dc0858c8ca2fbdf6aa76674d43eb698f78979fdc1a38f556a7af1facc4",
-    "blk.32.attn_norm.weight": "cdcc12f3b8b9773c6722736bfb748a2729230b21478cbcc4104859d3148df815",
-    "blk.32.attn_output.weight": "d43f1196822995ed89a9365c97054753a8b30ce20b6e273c8edcc42673a1e141",
-    "blk.32.attn_q.weight": "ebf2972bb3865cbc5be4840113a322089752038344beab2a0122c7cb4fb399b6",
-    "blk.32.attn_v.weight": "714db81704ff34fa137512903c1013acee7877467473e46600728b9240582eb7",
-    "blk.32.ffn_down.weight": "2cde3da1258bb170a79d5d3cdfe10c86a71eb34b77da46b74c5ed71e7f4fe274",
-    "blk.32.ffn_gate.weight": "c7e1ed792532613ff9d4e5834b6536e2e0f47df2303bc0fdaa90aac0c1f4e8db",
-    "blk.32.ffn_up.weight": "d8d6f13fe66a716e28f79101a29817f0c0d6f99969a6f017d51bafd1a16c600c",
-    "blk.33.attn_k.weight": "a0a28f6cbca88da00cab2ca37094d9b0503bf9defdae77b91895b911c408cbb6",
-    "blk.33.attn_norm.weight": "0251200c24cc8445607ace6dc8c5aa0566567997262b7cca53a11ac23cc564b2",
-    "blk.33.attn_output.weight": "b2423205bdf6a1096d43c44d8d12f1a84fcd4e1bb70fcf6dc8542b8b8a71a13c",
-    "blk.33.attn_q.weight": "00b425c3ef71065ce5e0234e702bf38143b4952da78a85f52ab2c2e3073d97ab",
-    "blk.33.attn_v.weight": "035edd2335df816c42c765a5e66b9d9b9e15a822a8dc1863508145499c942c14",
-    "blk.33.ffn_down.weight": "4894a923a3db75bae4496ba3ce5f28796ad31fe33996a066271fb8654964310e",
-    "blk.33.ffn_gate.weight": "8f6c819b8bbfbe3357fae89e1ac5a3d58be85b3b04be3bacf7b62775869046ff",
-    "blk.33.ffn_up.weight": "257c3544b5b544fd5d839665bf5caf107a329b59dbc3751efcaa24ae63c56179",
-    "blk.34.attn_k.weight": "b6cd8bba892e38dac4a2ebc3ba1bce49e71b967fc436fde30c6d76f54a18935f",
-    "blk.34.attn_norm.weight": "2b3c8e60a064cba9955752bbbbdd92c71ba5c2f1bd721097bdbe88b5abc68787",
-    "blk.34.attn_output.weight": "8cc272551c9aaca9db5a660c6927bab94a0243d74a30b2bc165f06bd577714ea",
-    "blk.34.attn_q.weight": "74b561eb4792484e6a94b58fe2583848c3ae28ff2f1bf3d02939a0cfdfa49990",
-    "blk.34.attn_v.weight": "dba19e24ff05154dc5a1f55c023729303a583d13d68732ce22ea74d4410dc8f0",
-    "blk.34.ffn_down.weight": "76eca5dfeb274c35774e0bf9f22ee420ed9085c8e99aa2cd5a236e4918b44c61",
-    "blk.34.ffn_gate.weight": "9af0862d5fcbc24732846488e653db8242a467765c0cdbc00332b3a40256b4a6",
-    "blk.34.ffn_up.weight": "2a03126bf73587eaba99ece2066103d12e47bcd4ce30ff6c17b2f383b81d40df",
-    "blk.35.attn_k.weight": "52513fc0cd4e997a842729af7d21dd09399bce0a339558374738be266d0fa2f0",
-    "blk.35.attn_norm.weight": "e5281fa911964263ccf1630b14762edbd41d0b9472d6ec695fc600fed4892c35",
-    "blk.35.attn_output.weight": "b391d6705d5dc6f48326b5fd16573f679edf64109d86fb729a498819676590ca",
-    "blk.35.attn_q.weight": "d16446921966db9b0e0539626ad22a2511ace780e59379d6a4162d8c5441440b",
-    "blk.35.attn_v.weight": "9d8cdf23ffdb0c5c74106843390b94b24c9f33ef0eb9998d39f78c73390101ea",
-    "blk.35.ffn_down.weight": "938eb6301f7bbf162d7dd965682a5ed11d0a4a530c6fedd7e5469ce80012fc17",
-    "blk.35.ffn_gate.weight": "5ad84f5a0c8edcfea1ecf1a3e3d21d85ceda0c4ad9e3c6ca68885eeff8ed3c2f",
-    "blk.35.ffn_up.weight": "1c4330d9dc71bf4c98812c34356c51f520f47610a534152aa6d29284b758090d",
-    "blk.36.attn_k.weight": "ef720655e5ca2465f13db2dfc4732fb4ef2c9d53acde52f514fd4f301e974081",
-    "blk.36.attn_norm.weight": "88f4b9310b3c8c2644e3029160cd35678c79dfa59280430e03f5c29a6fe84a58",
-    "blk.36.attn_output.weight": "aec6f915fffd7bb72cd783273e871b4f09605950089d45e72059d1316b6c4b01",
-    "blk.36.attn_q.weight": "72f9408a2405d42f8db6ce5fcf1d26a3660b6f225fc60e77d0277109cfcb82ed",
-    "blk.36.attn_v.weight": "0f3b3d851dc44b3893ef53f6cca5b4acc9658bacfe1cc2d13c3d704ddd409b67",
-    "blk.36.ffn_down.weight": "470aec48ce8c5129a6654d9fd26fcae72776f9fc1429a8bb05818072a876475d",
-    "blk.36.ffn_gate.weight": "7f5f296d09cf55679767b5d15de3eff489c456782119f25204be4b1647f18dcf",
-    "blk.36.ffn_up.weight": "b7ef74a1f7ffb4982711d93f1787be3a70edc3d2358d5203c41d8900508037d4",
-    "blk.37.attn_k.weight": "c4ffa5412e4ff2dcfe1aed991c1f54169fd171a4c7638e4b9f21a1ca64c5e1d6",
-    "blk.37.attn_norm.weight": "4eb6c888d841cccfacf5b963f8611120f6ff24b84af0b5714fd9ab36dcda422f",
-    "blk.37.attn_output.weight": "db2a7bbf9682f9f6eea672dae8e150738f1bf74dbc80edc7022017a3f040c8ac",
-    "blk.37.attn_q.weight": "e38c0462aff139afcbab289189823527e453abc9e541154adde5e7af88cacf0b",
-    "blk.37.attn_v.weight": "952eb2492ed452a72f96bcc12d4b2affad9dfdf46ee39ce4a5d7b57a5dc301e5",
-    "blk.37.ffn_down.weight": "25f23a8fbc44febf6dc4848fd7fe03a580e2822bd3b3b5a51f4990826bfe3e4e",
-    "blk.37.ffn_gate.weight": "707da5eb40118b035305d3262444382351f170a20a537386a70e90c5a83a7817",
-    "blk.37.ffn_up.weight": "d2d2ba5cfc4ef47338dd7384219e22bf030a5a2209e0354d88f5bbaaafd20e87",
-    "blk.38.attn_k.weight": "abc4bb189dedf7ce661e79028427623a4f91ac091c2cd60e31b58bc62b1cda71",
-    "blk.38.attn_norm.weight": "9f4803a7d03fd40fcb83d85f84eb1d5682ea4e5bb084f210c02850675d804c3d",
-    "blk.38.attn_output.weight": "77cb66007f1a41df7135d0e7f900ceb499c2f667dfc3f1a6ac01a3203bbd3ccf",
-    "blk.38.attn_q.weight": "d94a8b26cd375bf2bcaa76597e314aa8268ee50a479d00931e5e0e021feadb5d",
-    "blk.38.attn_v.weight": "660c907888bc5016dc69b7d35fe6f55c7ded697c93be0e2d332a2f17aff88758",
-    "blk.38.ffn_down.weight": "6f06173bae5b00ffaf88ef383619a8b9c6a8d0d5c6494695d17f6c1de1a68a13",
-    "blk.38.ffn_gate.weight": "89f99be149d03f116527bfcabe073c50001c874de40fb6e817f6619027f3cd05",
-    "blk.38.ffn_up.weight": "8d57557c8d5e2d2688b73f01dddf1ce8d5194990cda6358153320aea88aac7f8",
-    "blk.39.attn_k.weight": "21be09c988b46c8393e6c2ec9230f3b5136eb7607dd1953ba92d0811c2f0dd75",
-    "blk.39.attn_norm.weight": "ba7c1912dd1c4e2d16917201f62396fd0600e4a451137eaddff255548c209abd",
-    "blk.39.attn_output.weight": "acfaf4abb3fd27fd899b5563c3877f176b597d8f6cdb2f2fd3f3a0bd4da15ed6",
-    "blk.39.attn_q.weight": "e8adbc140d4c8f0db2a27ca584c5531d5b1e080555fe627e34d80d0814a92bed",
-    "blk.39.attn_v.weight": "92f96b0e1f724e73a0f90a76c145654418844c04a6d4b14c05eb5af8a62bf8dc",
-    "blk.39.ffn_down.weight": "4d9ee7c65fc16fe95d10c47b79ac6a525741947600a64b5fcea5d300a82c50de",
-    "blk.39.ffn_gate.weight": "7e18507989f39b32191133d2657c2ee3b74f42f070579204d727eb72215793d1",
-    "blk.39.ffn_up.weight": "22cda752269c9757ba918abede1df95bb0f83a5c772dea13c8deea3d5f2723d9",
-    "output_norm.weight": "2858cf0e39d32caf52b7861378ace076000241e147f10b9eb21d8a5cd149e3cb"
-}

convert/tokenizer_spm.go

@@ -6,9 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
-	"log/slog"
 	"os"
-	"reflect"
 	"slices"
 
 	"google.golang.org/protobuf/proto"
@@ -17,8 +15,6 @@ import (
 )
 
 func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
-	slog.Debug("using spm vocabulary")
-
 	ast, err := parseAdditionalSpecialTokens(fsys)
 	if err != nil {
 		return nil, err
@@ -47,19 +43,10 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 			v.Types = append(v.Types, int32(t))
 		default:
 			tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
-
-			// temporary fix to handle gemma3 broken configs
-			if slices.Contains([]string{"<end_of_turn>", "<start_of_turn>"}, piece.GetPiece()) {
+			if slices.Contains(ast, piece.GetPiece()) {
 				tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
 			}
 
-			for _, t := range ast {
-				if t.Content == piece.GetPiece() {
-					tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
-					break
-				}
-			}
-
 			v.Types = append(v.Types, tt)
 		}
 	}
@@ -91,16 +78,10 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 		return cmp.Compare(i.id, j.id)
 	})
 
-	for _, t := range ts {
-		if t.id < len(v.Tokens) {
-			if v.Tokens[t.id] == t.content {
-				slog.Warn("tokenizer", "duplicate token", t.content, "id", t.id)
-				continue
-			}
-			return nil, fmt.Errorf("token mismatch: %s != %s at pos [%d]", t.content, v.Tokens[t.id], t.id)
-		}
-		if t.id != len(v.Tokens) {
-			return nil, fmt.Errorf("invalid token id: [%d] as pos [%d]", t.id, len(v.Tokens))
+	n := len(v.Tokens)
+	for i, t := range ts {
+		if t.id != i+n {
+			return nil, fmt.Errorf("invalid token id: %d", t.id)
 		}
 
 		v.Tokens = append(v.Tokens, t.content)
@@ -111,15 +92,7 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 	return &v, nil
 }
 
-type specialToken struct {
-	Content    string `json:"content"`
-	Lstrip     bool   `json:"lstrip"`
-	Normalized bool   `json:"normalized"`
-	Rstrip     bool   `json:"rstrip"`
-	SingleWord bool   `json:"single_word"`
-}
-
-func parseAdditionalSpecialTokens(fsys fs.FS) ([]specialToken, error) {
+func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
 	f, err := fsys.Open("special_tokens_map.json")
 	if errors.Is(err, os.ErrNotExist) {
 		return nil, nil
@@ -129,43 +102,12 @@ func parseAdditionalSpecialTokens(fsys fs.FS) ([]specialToken, error) {
 	defer f.Close()
 
 	var m struct {
-		AdditionalSpecialTokens any `json:"additional_special_tokens"`
+		AdditionalSpecialTokens []string `json:"additional_special_tokens"`
 	}
 
 	if err := json.NewDecoder(f).Decode(&m); err != nil {
 		return nil, err
 	}
 
-	var ast []specialToken
-
-	switch st := m.AdditionalSpecialTokens.(type) {
-	case []string:
-		for _, s := range st {
-			ast = append(ast, specialToken{Content: s})
-		}
-	case []any:
-		for _, s := range st {
-			// marshal and unmarshal the object to get the special token
-			tMap := s.(map[string]any)
-			data, err := json.Marshal(tMap)
-			if err != nil {
-				return nil, err
-			}
-
-			var token specialToken
-			err = json.Unmarshal(data, &token)
-			if err != nil {
-				return nil, err
-			}
-
-			ast = append(ast, token)
-		}
-
-	default:
-		slog.Warn("special token", "unknown token", reflect.TypeOf(st))
-	}
-
-	slog.Debug("spm tokenizer", "additional tokens", ast)
-
-	return ast, nil
+	return m.AdditionalSpecialTokens, nil
 }

discover/amd_common.go

@@ -9,6 +9,8 @@ import (
 	"path/filepath"
 	"runtime"
 	"strings"
+
+	"github.com/ollama/ollama/envconfig"
 )
 
 // Determine if the given ROCm lib directory is usable by checking for existence of some glob patterns
@@ -39,10 +41,13 @@ func commonAMDValidateLibDir() (string, error) {
 	// Favor our bundled version
 
 	// Installer payload location if we're running the installed binary
-	rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
-	if rocmLibUsable(rocmTargetDir) {
-		slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-		return rocmTargetDir, nil
+	exe, err := os.Executable()
+	if err == nil {
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), envconfig.LibRelativeToExe(), "lib", "ollama")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
 	}
 
 	// Prefer explicit HIP env var

discover/amd_linux.go

@@ -77,7 +77,8 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 
 	gfxOverride := envconfig.HsaOverrideGfxVersion()
 	var supported []string
-	var libDir string
+	depPaths := LibraryDirs()
+	libDir := ""
 
 	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
 	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
@@ -352,8 +353,9 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				})
 				return nil, err
 			}
+			depPaths = append(depPaths, libDir)
 		}
-		gpuInfo.DependencyPath = []string{libDir}
+		gpuInfo.DependencyPath = depPaths
 
 		if gfxOverride == "" {
 			// Only load supported list once

discover/amd_windows.go

@@ -5,6 +5,7 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"os"
 	"path/filepath"
 	"slices"
 	"strconv"
@@ -49,13 +50,14 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		slog.Info(err.Error())
 		return nil, err
 	}
-
+	depPaths := LibraryDirs()
 	libDir, err := AMDValidateLibDir()
 	if err != nil {
 		err = fmt.Errorf("unable to verify rocm library: %w", err)
 		slog.Warn(err.Error())
 		return nil, err
 	}
+	depPaths = append(depPaths, libDir)
 
 	var supported []string
 	gfxOverride := envconfig.HsaOverrideGfxVersion()
@@ -111,7 +113,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				UnreliableFreeMemory: true,
 
 				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
-				DependencyPath: []string{libDir},
+				DependencyPath: depPaths,
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
 				Compute:        gfx,
@@ -162,7 +164,9 @@ func AMDValidateLibDir() (string, error) {
 	}
 
 	// Installer payload (if we're running from some other location)
-	rocmTargetDir := filepath.Join(LibOllamaPath, "rocm")
+	localAppData := os.Getenv("LOCALAPPDATA")
+	appDir := filepath.Join(localAppData, "Programs", "Ollama")
+	rocmTargetDir := filepath.Join(appDir, envconfig.LibRelativeToExe(), "lib", "ollama")
 	if rocmLibUsable(rocmTargetDir) {
 		slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
 		return rocmTargetDir, nil

discover/cpu_common.go

@@ -12,7 +12,7 @@ func IsNUMA() bool {
 		// numa support in llama.cpp is linux only
 		return false
 	}
-	ids := map[string]any{}
+	ids := map[string]interface{}{}
 	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
 	for _, packageId := range packageIds {
 		id, err := os.ReadFile(packageId)

discover/cuda_common.go

@@ -57,8 +57,7 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 		}
 	}
 
-	// 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) {
+	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
 		return "v11"
 	}
 	return "v12"

discover/gpu.go

@@ -23,6 +23,7 @@ import (
 
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )
 
 type cudaHandles struct {
@@ -100,7 +101,15 @@ func initCudaHandles() *cudaHandles {
 
 	// Aligned with driver, we can't carry as payloads
 	nvcudaMgmtPatterns := NvcudaGlobs
-	cudartMgmtPatterns = append(cudartMgmtPatterns, filepath.Join(LibOllamaPath, "cuda_v*", CudartMgmtName))
+
+	if runtime.GOOS == "windows" {
+		localAppData := os.Getenv("LOCALAPPDATA")
+		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
+	}
+	libDirs := LibraryDirs()
+	for _, d := range libDirs {
+		cudartMgmtPatterns = append(cudartMgmtPatterns, filepath.Join(d, CudartMgmtName))
+	}
 	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)
 
 	if len(NvmlGlobs) > 0 {
@@ -231,7 +240,7 @@ func GetGPUInfo() GpuInfoList {
 		if err != nil {
 			slog.Warn("error looking up system memory", "error", err)
 		}
-
+		depPaths := LibraryDirs()
 		details, err := GetCPUDetails()
 		if err != nil {
 			slog.Warn("failed to lookup CPU details", "error", err)
@@ -239,9 +248,11 @@ func GetGPUInfo() GpuInfoList {
 		cpus = []CPUInfo{
 			{
 				GpuInfo: GpuInfo{
-					memInfo: mem,
-					Library: "cpu",
-					ID:      "0",
+					memInfo:        mem,
+					Library:        "cpu",
+					Variant:        runners.GetCPUCapability().String(),
+					ID:             "0",
+					DependencyPath: depPaths,
 				},
 				CPUs: details,
 			},
@@ -283,13 +294,17 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
 				variant := cudaVariant(gpuInfo)
-
-				// Start with our bundled libraries
-				if variant != "" {
-					variantPath := filepath.Join(LibOllamaPath, "cuda_"+variant)
-					if _, err := os.Stat(variantPath); err == nil {
-						// Put the variant directory first in the search path to avoid runtime linking to the wrong library
-						gpuInfo.DependencyPath = append([]string{variantPath}, gpuInfo.DependencyPath...)
+				if depPaths != nil {
+					gpuInfo.DependencyPath = depPaths
+					// Check for variant specific directory
+					if variant != "" {
+						for _, d := range depPaths {
+							if _, err := os.Stat(filepath.Join(d, "cuda_"+variant)); err == nil {
+								// Put the variant directory first in the search path to avoid runtime linking to the wrong library
+								gpuInfo.DependencyPath = append([]string{filepath.Join(d, "cuda_"+variant)}, gpuInfo.DependencyPath...)
+								break
+							}
+						}
 					}
 				}
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
@@ -361,7 +376,7 @@ func GetGPUInfo() GpuInfoList {
 						gpuInfo.FreeMemory = uint64(memInfo.free)
 						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-						gpuInfo.DependencyPath = []string{LibOllamaPath}
+						gpuInfo.DependencyPath = depPaths
 						oneapiGPUs = append(oneapiGPUs, gpuInfo)
 					}
 				}
@@ -497,30 +512,33 @@ func GetGPUInfo() GpuInfoList {
 
 func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
+	var ldPaths []string
 	gpuLibPaths := []string{}
 	slog.Debug("Searching for GPU library", "name", baseLibName)
 
-	// search our bundled libraries first
-	patterns := []string{filepath.Join(LibOllamaPath, baseLibName)}
-
-	var ldPaths []string
-	switch runtime.GOOS {
-	case "windows":
-		ldPaths = strings.Split(os.Getenv("PATH"), string(os.PathListSeparator))
-	case "linux":
-		ldPaths = strings.Split(os.Getenv("LD_LIBRARY_PATH"), string(os.PathListSeparator))
+	// Start with our bundled libraries
+	patterns := []string{}
+	for _, d := range LibraryDirs() {
+		patterns = append(patterns, filepath.Join(d, baseLibName))
 	}
 
-	// then search the system's LD_LIBRARY_PATH
-	for _, p := range ldPaths {
-		p, err := filepath.Abs(p)
+	switch runtime.GOOS {
+	case "windows":
+		ldPaths = strings.Split(os.Getenv("PATH"), ";")
+	case "linux":
+		ldPaths = strings.Split(os.Getenv("LD_LIBRARY_PATH"), ":")
+	default:
+		return gpuLibPaths
+	}
+
+	// Then with whatever we find in the PATH/LD_LIBRARY_PATH
+	for _, ldPath := range ldPaths {
+		d, err := filepath.Abs(ldPath)
 		if err != nil {
 			continue
 		}
-		patterns = append(patterns, filepath.Join(p, baseLibName))
+		patterns = append(patterns, filepath.Join(d, baseLibName))
 	}
-
-	// finally, search the default patterns provided by the caller
 	patterns = append(patterns, defaultPatterns...)
 	slog.Debug("gpu library search", "globs", patterns)
 	for _, pattern := range patterns {
@@ -697,6 +715,23 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 	}
 }
 
+func LibraryDirs() []string {
+	// dependencies can exist wherever we found the runners (e.g. build tree for developers) and relative to the executable
+	// This can be simplified once we no longer carry runners as payloads
+	exe, err := os.Executable()
+	if err != nil {
+		slog.Warn("failed to lookup executable path", "error", err)
+		return nil
+	}
+
+	lib := filepath.Join(filepath.Dir(exe), envconfig.LibRelativeToExe(), "lib", "ollama")
+	if _, err := os.Stat(lib); err != nil {
+		return nil
+	}
+
+	return []string{lib}
+}
+
 func GetSystemInfo() SystemInfo {
 	gpus := GetGPUInfo()
 	gpuMutex.Lock()

discover/gpu_darwin.go

@@ -15,6 +15,7 @@ import (
 	"syscall"
 
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )
 
 const (
@@ -27,6 +28,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
+				Variant: runners.GetCPUCapability().String(),
 				memInfo: mem,
 			},
 		}
@@ -49,6 +51,7 @@ func GetCPUInfo() GpuInfoList {
 	return []GpuInfo{
 		{
 			Library: "cpu",
+			Variant: runners.GetCPUCapability().String(),
 			memInfo: mem,
 		},
 	}

discover/gpu_linux.go

@@ -111,7 +111,6 @@ func GetCPUDetails() ([]CPU, error) {
 	if err != nil {
 		return nil, err
 	}
-	defer file.Close()
 	return linuxCPUDetails(file)
 }
 
@@ -169,11 +168,13 @@ func linuxCPUDetails(file io.Reader) ([]CPU, error) {
 	for id, s := range socketByID {
 		s.CoreCount = len(coreBySocket[id])
 		s.ThreadCount = 0
+		for _, tc := range threadsByCoreBySocket[id] {
+			s.ThreadCount += tc
+		}
 
 		// This only works if HT is enabled, consider a more reliable model, maybe cache size comparisons?
 		efficiencyCoreCount := 0
 		for _, threads := range threadsByCoreBySocket[id] {
-			s.ThreadCount += threads
 			if threads == 1 {
 				efficiencyCoreCount++
 			}

discover/path.go

@@ -1,56 +0,0 @@
-package discover
-
-import (
-	"os"
-	"path/filepath"
-	"runtime"
-)
-
-// LibPath is a path to lookup dynamic libraries
-// in development it's usually 'build/lib/ollama'
-// in distribution builds it's 'lib/ollama' on Windows
-// '../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.
-var LibOllamaPath string = func() string {
-	exe, err := os.Executable()
-	if err != nil {
-		return ""
-	}
-
-	if eval, err := filepath.EvalSymlinks(exe); err == nil {
-		exe = eval
-	}
-
-	var libPath string
-	switch runtime.GOOS {
-	case "windows":
-		libPath = filepath.Join(filepath.Dir(exe), "lib", "ollama")
-	case "linux":
-		libPath = filepath.Join(filepath.Dir(exe), "..", "lib", "ollama")
-	case "darwin":
-		libPath = filepath.Dir(exe)
-	}
-
-	cwd, err := os.Getwd()
-	if err != nil {
-		return ""
-	}
-
-	paths := []string{
-		libPath,
-
-		// build paths for development
-		filepath.Join(filepath.Dir(exe), "build", "lib", "ollama"),
-		filepath.Join(cwd, "build", "lib", "ollama"),
-	}
-
-	for _, p := range paths {
-		if _, err := os.Stat(p); err == nil {
-			return p
-		}
-	}
-
-	return filepath.Dir(exe)
-}()

discover/types.go

@@ -5,6 +5,7 @@ import (
 	"log/slog"
 
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/runners"
 )
 
 type memInfo struct {
@@ -106,7 +107,7 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != "" {
+		if info.Variant != runners.CPUCapabilityNone.String() {
 			requested += "_" + info.Variant
 		}
 		for i, lib := range libs {

docs/README.md

@@ -2,7 +2,7 @@
 
 ### Getting Started
 * [Quickstart](../README.md#quickstart)
-* [Examples](./examples.md)
+* [Examples](../examples)
 * [Importing models](./import.md)
 * [Linux Documentation](./linux.md)
 * [Windows Documentation](./windows.md)

docs/api.md

@@ -31,7 +31,7 @@ Certain endpoints stream responses as JSON objects. Streaming can be disabled by
 
 ## Generate a completion
 
-```
+```shell
 POST /api/generate
 ```
 
@@ -173,7 +173,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ##### Response
 
-```json5
+```json
 {
   "model": "codellama:code",
   "created_at": "2024-07-22T20:47:51.147561Z",
@@ -306,7 +306,7 @@ curl http://localhost:11434/api/generate -d '{
 
 #### Response
 
-```json
+```
 {
   "model": "llava",
   "created_at": "2023-11-03T15:36:02.583064Z",
@@ -485,7 +485,7 @@ A single JSON object is returned:
 
 ## Generate a chat completion
 
-```
+```shell
 POST /api/chat
 ```
 
@@ -495,14 +495,14 @@ 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
+- `tools`: tools for the model to use if supported. Requires `stream` to be set to `false`
 
 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
 - `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
+- `tool_calls` (optional): a list of tools the model wants to use
 
 Advanced parameters (optional):
 
@@ -558,10 +558,6 @@ Final response:
 {
   "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
-  "message": {
-    "role": "assistant",
-    "content": ""
-  },
   "done": true,
   "total_duration": 4883583458,
   "load_duration": 1334875,
@@ -799,7 +795,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ##### Request
 
-```shell
+```
 curl http://localhost:11434/api/chat -d '{
   "model": "llama3.2",
   "messages": [
@@ -874,7 +870,7 @@ If the messages array is empty, the model will be loaded into memory.
 
 ##### Request
 
-```shell
+```
 curl http://localhost:11434/api/chat -d '{
   "model": "llama3.2",
   "messages": []
@@ -882,7 +878,6 @@ curl http://localhost:11434/api/chat -d '{
 ```
 
 ##### Response
-
 ```json
 {
   "model": "llama3.2",
@@ -902,7 +897,7 @@ If the messages array is empty and the `keep_alive` parameter is set to `0`, a m
 
 ##### Request
 
-```shell
+```
 curl http://localhost:11434/api/chat -d '{
   "model": "llama3.2",
   "messages": [],
@@ -929,7 +924,7 @@ A single JSON object is returned:
 
 ## Create a Model
 
-```
+```shell
 POST /api/create
 ```
 
@@ -1025,7 +1020,7 @@ curl http://localhost:11434/api/create -d '{
 
 A stream of JSON objects is returned:
 
-```json
+```
 {"status":"quantizing F16 model to Q4_K_M"}
 {"status":"creating new layer sha256:667b0c1932bc6ffc593ed1d03f895bf2dc8dc6df21db3042284a6f4416b06a29"}
 {"status":"using existing layer sha256:11ce4ee3e170f6adebac9a991c22e22ab3f8530e154ee669954c4bc73061c258"}
@@ -1056,7 +1051,7 @@ curl http://localhost:11434/api/create -d '{
 
 A stream of JSON objects is returned:
 
-```json
+```
 {"status":"parsing GGUF"}
 {"status":"using existing layer sha256:432f310a77f4650a88d0fd59ecdd7cebed8d684bafea53cbff0473542964f0c3"}
 {"status":"writing manifest"}
@@ -1123,7 +1118,7 @@ Return 200 OK if the blob exists, 404 Not Found if it does not.
 
 ## Push a Blob
 
-```
+```shell
 POST /api/blobs/:digest
 ```
 
@@ -1147,7 +1142,7 @@ Return 201 Created if the blob was successfully created, 400 Bad Request if the
 
 ## List Local Models
 
-```
+```shell
 GET /api/tags
 ```
 
@@ -1200,7 +1195,7 @@ A single JSON object will be returned.
 
 ## Show Model Information
 
-```
+```shell
 POST /api/show
 ```
 
@@ -1217,13 +1212,13 @@ Show information about a model including details, modelfile, template, parameter
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "model": "llava"
+  "model": "llama3.2"
 }'
 ```
 
 #### Response
 
-```json5
+```json
 {
   "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"",
   "parameters": "num_keep                       24\nstop                           \"<|start_header_id|>\"\nstop                           \"<|end_header_id|>\"\nstop                           \"<|eot_id|>\"",
@@ -1260,17 +1255,13 @@ curl http://localhost:11434/api/show -d '{
     "tokenizer.ggml.pre": "llama-bpe",
     "tokenizer.ggml.token_type": [],        // populates if `verbose=true`
     "tokenizer.ggml.tokens": []             // populates if `verbose=true`
-  },
-  "capabilities": [
-    "completion",
-    "vision"
-  ],
+  }
 }
 ```
 
 ## Copy a Model
 
-```
+```shell
 POST /api/copy
 ```
 
@@ -1293,7 +1284,7 @@ Returns a 200 OK if successful, or a 404 Not Found if the source model doesn't e
 
 ## Delete a Model
 
-```
+```shell
 DELETE /api/delete
 ```
 
@@ -1319,7 +1310,7 @@ Returns a 200 OK if successful, 404 Not Found if the model to be deleted doesn't
 
 ## Pull a Model
 
-```
+```shell
 POST /api/pull
 ```
 
@@ -1391,7 +1382,7 @@ if `stream` is set to false, then the response is a single JSON object:
 
 ## Push a Model
 
-```
+```shell
 POST /api/push
 ```
 
@@ -1456,7 +1447,7 @@ If `stream` is set to `false`, then the response is a single JSON object:
 
 ## Generate Embeddings
 
-```
+```shell
 POST /api/embed
 ```
 
@@ -1524,7 +1515,7 @@ curl http://localhost:11434/api/embed -d '{
 ```
 
 ## List Running Models
-```
+```shell
 GET /api/ps
 ```
 
@@ -1571,7 +1562,7 @@ A single JSON object will be returned.
 
 > Note: this endpoint has been superseded by `/api/embed`
 
-```
+```shell
 POST /api/embeddings
 ```
 
@@ -1611,7 +1602,7 @@ curl http://localhost:11434/api/embeddings -d '{
 
 ## Version
 
-```
+```shell
 GET /api/version
 ```
 

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

@@ -1,159 +1,165 @@
 # Development
 
-Install prerequisites:
+Install required tools:
 
-- [Go](https://go.dev/doc/install)
-- C/C++ Compiler e.g. Clang on macOS, [TDM-GCC](https://github.com/jmeubank/tdm-gcc/releases/latest) (Windows amd64) or [llvm-mingw](https://github.com/mstorsjo/llvm-mingw) (Windows arm64), GCC/Clang on Linux.
+- go version 1.22 or higher
+- OS specific C/C++ compiler (see below)
+- GNU Make
 
-Then build and run Ollama from the root directory of the repository:
 
-```shell
-go run . serve
+## Overview
+
+Ollama uses a mix of Go and C/C++ code to interface with GPUs.  The C/C++ code is compiled with both CGO and GPU library specific compilers.  A set of GNU Makefiles are used to compile the project.  GPU Libraries are auto-detected based on the typical environment variables used by the respective libraries, but can be overridden if necessary.  The default make target will build the runners and primary Go Ollama application that will run within the repo directory.  Throughout the examples below `-j 5` is suggested for 5 parallel jobs to speed up the build.  You can adjust the job count based on your CPU Core count to reduce build times.  If you want to relocate the built binaries, use the `dist` target and recursively copy the files in `./dist/$OS-$ARCH/` to your desired location. To learn more about the other make targets use `make help`
+
+Once you have built the GPU/CPU runners, you can compile the main application with `go build .` 
+
+### MacOS
+
+[Download Go](https://go.dev/dl/)
+
+```bash
+make -j 5
 ```
 
-## macOS (Apple Silicon)
+Now you can run `ollama`:
 
-macOS Apple Silicon supports Metal which is built-in to the Ollama binary. No additional steps are required.
-
-## macOS (Intel)
-
-Install prerequisites:
-
-- [CMake](https://cmake.org/download/) or `brew install cmake`
-
-Then, configure and build the project:
-
-```shell
-cmake -B build
-cmake --build build
+```bash
+./ollama
 ```
 
-Lastly, run Ollama:
+#### Xcode 15 warnings
 
-```shell
-go run . serve
+If you are using Xcode newer than version 14, you may see a warning during `go build` about `ld: warning: ignoring duplicate libraries: '-lobjc'` due to Golang issue https://github.com/golang/go/issues/67799 which can be safely ignored.  You can suppress the warning with `export CGO_LDFLAGS="-Wl,-no_warn_duplicate_libraries"`
+
+### Linux
+
+#### Linux CUDA (NVIDIA)
+
+_Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
+
+Install `make`, `gcc` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
+development and runtime packages.
+
+Typically the makefile will auto-detect CUDA, however, if your Linux distro
+or installation approach uses alternative paths, you can specify the location by
+overriding `CUDA_PATH` to the location of the CUDA toolkit. You can customize
+a set of target CUDA architectures by setting `CUDA_ARCHITECTURES` (e.g. `CUDA_ARCHITECTURES=50;60;70`)
+
+```
+make -j 5
 ```
 
-## Windows
+If both v11 and v12 tookkits are detected, runners for both major versions will be built by default.  You can build just v12 with `make cuda_v12`
 
-Install prerequisites:
+#### Older Linux CUDA (NVIDIA)
 
-- [CMake](https://cmake.org/download/)
-- [Visual Studio 2022](https://visualstudio.microsoft.com/downloads/) including the Native Desktop Workload
-- (Optional) AMD GPU support
-    - [ROCm](https://rocm.docs.amd.com/en/latest/)
-    - [Ninja](https://github.com/ninja-build/ninja/releases)
-- (Optional) NVIDIA GPU support
-    - [CUDA SDK](https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_network)
+To support older GPUs with Compute Capability 3.5 or 3.7, you will need to use an older version of the Driver from [Unix Driver Archive](https://www.nvidia.com/en-us/drivers/unix/) (tested with 470) and [CUDA Toolkit Archive](https://developer.nvidia.com/cuda-toolkit-archive) (tested with cuda V11).  When you build Ollama, you will need to set two make variable to adjust the minimum compute capability Ollama supports via `make -j 5 CUDA_ARCHITECTURES="35;37;50;52" EXTRA_GOLDFLAGS="\"-X=github.com/ollama/ollama/discover.CudaComputeMajorMin=3\" \"-X=github.com/ollama/ollama/discover.CudaComputeMinorMin=5\""`.  To find the Compute Capability of your older GPU, refer to [GPU Compute Capability](https://developer.nvidia.com/cuda-gpus).
 
-Then, configure and build the project:
+#### Linux ROCm (AMD)
 
-```shell
-cmake -B build
-cmake --build build --config Release
+_Your operating system distribution may already have packages for AMD ROCm. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
+
+Install [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `make`, `gcc`, and `golang`.
+
+Typically the build scripts will auto-detect ROCm, however, if your Linux distro
+or installation approach uses unusual paths, you can specify the location by
+specifying an environment variable `HIP_PATH` to the location of the ROCm
+install (typically `/opt/rocm`). You can also customize
+the AMD GPU targets by setting HIP_ARCHS (e.g. `HIP_ARCHS=gfx1101;gfx1102`)
+
+```
+make -j 5
 ```
 
-> [!IMPORTANT]
-> Building for ROCm requires additional flags:
-> ```
-> cmake -B build -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++
-> cmake --build build --config Release
-> ```
+ROCm requires elevated privileges to access the GPU at runtime. On most distros you can add your user account to the `render` group, or run as root.
 
+#### Containerized Linux Build
 
-Lastly, run Ollama:
+If you have Docker and buildx available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting artifacts are placed in `./dist`  and by default the script builds both arm64 and amd64 binaries.  If you want to build only amd64, you can build with `PLATFORM=linux/amd64 ./scripts/build_linux.sh`
 
-```shell
-go run . serve
+### Windows
+
+The following tools are required as a minimal development environment to build CPU inference support.
+
+- Go version 1.22 or higher
+  - https://go.dev/dl/
+- Git
+  - https://git-scm.com/download/win
+- clang with gcc compat and Make.  There are multiple options on how to go about installing these tools on Windows.  We have verified the following, but others may work as well:  
+  - [MSYS2](https://www.msys2.org/)
+    - After installing, from an MSYS2 terminal, run `pacman -S mingw-w64-clang-x86_64-gcc-compat mingw-w64-clang-x86_64-clang make` to install the required tools
+  - Assuming you used the default install prefix for msys2 above, add `C:\msys64\clang64\bin` and `c:\msys64\usr\bin` to your environment variable `PATH` where you will perform the build steps below (e.g. system-wide, account-level, powershell, cmd, etc.)
+
+> [!NOTE]  
+> Due to bugs in the GCC C++ library for unicode support, Ollama should be built with clang on windows.
+
+```
+make -j 5
 ```
 
-## Windows (ARM)
+#### GPU Support
 
-Windows ARM does not support additional acceleration libraries at this time.  Do not use cmake, simply `go run` or `go build`.
+The GPU tools require the Microsoft native build tools.  To build either CUDA or ROCm, you must first install MSVC via Visual Studio:
 
-## Linux
+- Make sure to select `Desktop development with C++` as a Workload during the Visual Studio install
+- You must complete the Visual Studio install and run it once **BEFORE** installing CUDA or ROCm for the tools to properly register
+- Add the location of the **64 bit (x64)** compiler (`cl.exe`) to your `PATH`
+- Note: the default Developer Shell may configure the 32 bit (x86) compiler which will lead to build failures.  Ollama requires a 64 bit toolchain.
 
-Install prerequisites:
+#### Windows CUDA (NVIDIA)
 
-- [CMake](https://cmake.org/download/) or `sudo apt install cmake` or `sudo dnf install cmake`
-- (Optional) AMD GPU support
-    - [ROCm](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/quick-start.html)
-- (Optional) NVIDIA GPU support
-    - [CUDA SDK](https://developer.nvidia.com/cuda-downloads)
+In addition to the common Windows development tools and MSVC described above:
 
-> [!IMPORTANT]
-> Ensure prerequisites are in `PATH` before running CMake.
+- [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)
 
+#### Windows ROCm (AMD Radeon)
 
-Then, configure and build the project:
+In addition to the common Windows development tools and MSVC described above:
 
-```shell
-cmake -B build
-cmake --build build
+- [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
+
+#### Windows arm64
+
+The default `Developer PowerShell for VS 2022` may default to x86 which is not what you want.  To ensure you get an arm64 development environment, start a plain PowerShell terminal and run:
+
+```powershell
+import-module 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\Common7\\Tools\\Microsoft.VisualStudio.DevShell.dll'
+Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community' -skipautomaticlocation
 ```
 
-Lastly, run Ollama:
+You can confirm with `write-host $env:VSCMD_ARG_TGT_ARCH`
 
-```shell
-go run . serve
+Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64 msys2 environment.  Ollama requires gcc and mingw32-make to compile, which is not currently available on Windows arm64, but a gcc compatibility adapter is available via `mingw-w64-clang-aarch64-gcc-compat`. At a minimum you will need to install the following:
+
+```
+pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
 ```
 
-## Docker
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)
 
-```shell
-docker build .
+
+## Advanced CPU Vector Settings
+
+On x86, running `make` will compile several CPU runners which can run on different CPU families. At runtime, Ollama will auto-detect the best variation to load.  If GPU libraries are present at build time, Ollama also compiles GPU runners with the `AVX` CPU vector feature enabled.  This provides a good performance balance when loading large models that split across GPU and CPU with broad compatibility.  Some users may prefer no vector extensions (e.g. older Xeon/Celeron processors, or hypervisors that mask the vector features) while other users may prefer turning on many more vector extensions to further improve performance for split model loads.
+
+To customize the set of CPU vector features enabled for a CPU runner and all GPU runners, use CUSTOM_CPU_FLAGS during the build.
+
+To build without any vector flags:
+
+```
+make CUSTOM_CPU_FLAGS=""
 ```
 
-### ROCm
-
-```shell
-docker build --build-arg FLAVOR=rocm .
+To build with both AVX and AVX2:
+```
+make CUSTOM_CPU_FLAGS=avx,avx2
 ```
 
-## Running tests
+To build with AVX512 features turned on:
 
-To run tests, use `go test`:
-
-```shell
-go test ./...
+```
+make CUSTOM_CPU_FLAGS=avx,avx2,avx512,avx512vbmi,avx512vnni,avx512bf16
 ```
 
-> NOTE: In rare cirumstances, you may nedd 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
-> test failures resulting from your change(s), if any, locally, but CI will
-> break.
->
-> If you see failures in CI, you can either keep pushing changes to see if the
-> CI build passes, or you can enable the "synctest" package locally to see the
-> failures before pushing.
->
-> To enable the "synctest" package for testing, run the following command:
->
-> ```shell
-> GOEXPERIMENT=synctest go test ./...
-> ```
->
-> If you wish to enable synctest for all go commands, you can set the
-> `GOEXPERIMENT` environment variable in your shell profile or by using:
->
-> ```shell
-> go env -w GOEXPERIMENT=synctest
-> ```
->
-> Which will enable the "synctest" package for all go commands without needing
-> to set it for all shell sessions.
->
-> The synctest package is not required for production builds.
-
-## Library detection
-
-Ollama looks for acceleration libraries in the following paths relative to the `ollama` executable:
-
-* `./lib/ollama` (Windows)
-* `../lib/ollama` (Linux)
-* `.` (macOS)
-* `build/lib/ollama` (for development)
-
-If the libraries are not found, Ollama will not run with any acceleration libraries.
+> [!NOTE]  
+> If you are experimenting with different flags, make sure to do a `make clean` between each change to ensure everything is rebuilt with the new compiler flags

docs/docker.md

@@ -2,7 +2,7 @@
 
 ### CPU only
 
-```shell
+```bash
 docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
 ```
 
@@ -11,46 +11,42 @@ Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-
 
 #### Install with Apt
 1.  Configure the repository
-
-    ```shell
-    curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
-        | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
-    curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
-        | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
-        | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
-    sudo apt-get update
-    ```
-
+```bash
+curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
+    | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
+    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
+    | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
+sudo apt-get update
+```
 2.  Install the NVIDIA Container Toolkit packages
-
-    ```shell
-    sudo apt-get install -y nvidia-container-toolkit
-    ```
+```bash
+sudo apt-get install -y nvidia-container-toolkit
+```
 
 #### Install with Yum or Dnf
 1.  Configure the repository
 
-    ```shell
-    curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
-        | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
-    ```
+```bash
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
+    | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
+```
 
 2. Install the NVIDIA Container Toolkit packages
 
-    ```shell
-    sudo yum install -y nvidia-container-toolkit
-    ```
+```bash
+sudo yum install -y nvidia-container-toolkit
+```
 
 #### Configure Docker to use Nvidia driver
-
-```shell
+```
 sudo nvidia-ctk runtime configure --runtime=docker
 sudo systemctl restart docker
 ```
 
 #### Start the container
 
-```shell
+```bash
 docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
 ```
 
@@ -61,7 +57,7 @@ docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ol
 
 To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:
 
-```shell
+```
 docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm
 ```
 
@@ -69,7 +65,7 @@ docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 114
 
 Now you can run a model:
 
-```shell
+```
 docker exec -it ollama ollama run llama3.2
 ```
 

docs/faq.md

@@ -20,18 +20,12 @@ Please refer to the [GPU docs](./gpu.md).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 4096 tokens, unless you have a single GPU with <= 4 GB of VRAM, in which case it will default to 2048 tokens. 
-
-This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use: 
-
-```shell
-OLLAMA_CONTEXT_LENGTH=8192 ollama serve
-```
+By default, Ollama uses a context window size of 2048 tokens.
 
 To change this when using `ollama run`, use `/set parameter`:
 
-```shell
-/set parameter num_ctx 8192
+```
+/set parameter num_ctx 4096
 ```
 
 When using the API, specify the `num_ctx` parameter:
@@ -41,7 +35,7 @@ curl http://localhost:11434/api/generate -d '{
   "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "options": {
-    "num_ctx": 8192
+    "num_ctx": 4096
   }
 }'
 ```
@@ -52,15 +46,10 @@ Use the `ollama ps` command to see what models are currently loaded into memory.
 
 ```shell
 ollama ps
+NAME      	ID          	SIZE 	PROCESSOR	UNTIL
+llama3:70b	bcfb190ca3a7	42 GB	100% GPU 	4 minutes from now
 ```
 
-> **Output**:
->
-> ```
-> NAME      	ID          	SIZE 	PROCESSOR	UNTIL
-> llama3:70b	bcfb190ca3a7	42 GB	100% GPU 	4 minutes from now
-> ```
-
 The `Processor` column will show which memory the model was loaded in to:
 * `100% GPU` means the model was loaded entirely into the GPU
 * `100% CPU` means the model was loaded entirely in system memory
@@ -77,7 +66,7 @@ If Ollama is run as a macOS application, environment variables should be set usi
 1. For each environment variable, call `launchctl setenv`.
 
     ```bash
-    launchctl setenv OLLAMA_HOST "0.0.0.0:11434"
+    launchctl setenv OLLAMA_HOST "0.0.0.0"
     ```
 
 2. Restart Ollama application.
@@ -92,14 +81,14 @@ If Ollama is run as a systemd service, environment variables should be set using
 
     ```ini
     [Service]
-    Environment="OLLAMA_HOST=0.0.0.0:11434"
+    Environment="OLLAMA_HOST=0.0.0.0"
     ```
 
 3. Save and exit.
 
 4. Reload `systemd` and restart Ollama:
 
-   ```shell
+   ```bash
    systemctl daemon-reload
    systemctl restart ollama
    ```
@@ -193,13 +182,6 @@ cloudflared tunnel --url http://localhost:11434 --http-host-header="localhost:11
 
 Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
 
-For browser extensions, you'll need to explicitly allow the extension's origin pattern. Set `OLLAMA_ORIGINS` to include `chrome-extension://*`, `moz-extension://*`, and `safari-web-extension://*` if you wish to allow all browser extensions access, or specific extensions as needed:
-
-```
-# Allow all Chrome, Firefox, and Safari extensions
-OLLAMA_ORIGINS=chrome-extension://*,moz-extension://*,safari-web-extension://* ollama serve
-```
-
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## Where are models stored?
@@ -239,19 +221,16 @@ properties.
 If you are using the API you can preload a model by sending the Ollama server an empty request. This works with both the `/api/generate` and `/api/chat` API endpoints.
 
 To preload the mistral model using the generate endpoint, use:
-
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "mistral"}'
 ```
 
 To use the chat completions endpoint, use:
-
 ```shell
 curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 ```
 
 To preload a model using the CLI, use the command:
-
 ```shell
 ollama run llama3.2 ""
 ```
@@ -271,13 +250,11 @@ If you're using the API, use the `keep_alive` parameter with the `/api/generate`
 * '0' which will unload the model immediately after generating a response
 
 For example, to preload a model and leave it in memory use:
-
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": -1}'
 ```
 
 To unload the model and free up memory use:
-
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": 0}'
 ```

docs/gpu.md

@@ -7,7 +7,7 @@ Check your compute compatibility to see if your card is supported:
 
 | Compute Capability | Family              | Cards                                                                                                       |
 | ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- |
-| 9.0                | NVIDIA              | `H200` `H100`                                                                                               |
+| 9.0                | NVIDIA              | `H100`                                                                                                      |
 | 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060`  |
 |                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
 | 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` `RTX 3050 Ti` `RTX 3050`   |
@@ -38,7 +38,7 @@ Numeric IDs may be used, however ordering may vary, so UUIDs are more reliable.
 You can discover the UUID of your GPUs by running `nvidia-smi -L` If you want to
 ignore the GPUs and force CPU usage, use an invalid GPU ID (e.g., "-1")
 
-### Linux Suspend Resume
+### Laptop Suspend Resume
 
 On linux, after a suspend/resume cycle, sometimes Ollama will fail to discover
 your NVIDIA GPU, and fallback to running on the CPU.  You can workaround this

docs/import.md

@@ -20,13 +20,13 @@ Make sure that you use the same base model in the `FROM` command as you used to
 
 Now run `ollama create` from the directory where the `Modelfile` was created:
 
-```shell
+```bash
 ollama create my-model
 ```
 
 Lastly, test the model:
 
-```shell
+```bash
 ollama run my-model
 ```
 

docs/linux.md

@@ -75,7 +75,7 @@ RestartSec=3
 Environment="PATH=$PATH"
 
 [Install]
-WantedBy=multi-user.target
+WantedBy=default.target
 ```
 
 Then start the service:
@@ -119,7 +119,7 @@ sudo systemctl status ollama
 
 To customize the installation of Ollama, you can edit the systemd service file or the environment variables by running:
 
-```shell
+```
 sudo systemctl edit ollama
 ```
 
@@ -152,7 +152,7 @@ Use `OLLAMA_VERSION` environment variable with the install script to install a s
 For example:
 
 ```shell
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.3.9 sh
 ```
 
 ## Viewing logs
@@ -186,9 +186,3 @@ sudo rm -r /usr/share/ollama
 sudo userdel ollama
 sudo groupdel ollama
 ```
-
-Remove installed libraries:
-
-```shell
-sudo rm -rf /usr/local/lib/ollama
-```

docs/modelfile.md

@@ -28,7 +28,7 @@ A model file is the blueprint to create and share models with Ollama.
 
 The format of the `Modelfile`:
 
-```
+```modelfile
 # comment
 INSTRUCTION arguments
 ```
@@ -49,7 +49,7 @@ INSTRUCTION arguments
 
 An example of a `Modelfile` creating a mario blueprint:
 
-```
+```modelfile
 FROM llama3.2
 # sets the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -67,32 +67,28 @@ To use this:
 3. `ollama run choose-a-model-name`
 4. Start using the model!
 
+More examples are available in the [examples directory](../examples).
+
 To view the Modelfile of a given model, use the `ollama show --modelfile` command.
 
-```shell
-ollama show --modelfile llama3.2
-```
+  ```bash
+  > ollama show --modelfile llama3.2
+  # Modelfile generated by "ollama show"
+  # To build a new Modelfile based on this one, replace the FROM line with:
+  # FROM llama3.2:latest
+  FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
+  TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
 
-> **Output**:
->
-> ```
-> # Modelfile generated by "ollama show"
-> # To build a new Modelfile based on this one, replace the FROM line with:
-> # FROM llama3.2:latest
-> FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
-> TEMPLATE """{{ 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|>
->
-> {{ .Response }}<|eot_id|>"""
-> PARAMETER stop "<|start_header_id|>"
-> PARAMETER stop "<|end_header_id|>"
-> PARAMETER stop "<|eot_id|>"
-> PARAMETER stop "<|reserved_special_token"
-> ```
+  {{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
 
+  {{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+  {{ .Response }}<|eot_id|>"""
+  PARAMETER stop "<|start_header_id|>"
+  PARAMETER stop "<|end_header_id|>"
+  PARAMETER stop "<|eot_id|>"
+  PARAMETER stop "<|reserved_special_token"
+  ```
 
 ## Instructions
 
@@ -100,13 +96,13 @@ ollama show --modelfile llama3.2
 
 The `FROM` instruction defines the base model to use when creating a model.
 
-```
+```modelfile
 FROM <model name>:<tag>
 ```
 
 #### Build from existing model
 
-```
+```modelfile
 FROM llama3.2
 ```
 
@@ -117,7 +113,7 @@ Additional models can be found at:
 
 #### Build from a Safetensors model
 
-```
+```modelfile
 FROM <model directory>
 ```
 
@@ -131,7 +127,7 @@ Currently supported model architectures:
 
 #### Build from a GGUF file
 
-```
+```modelfile
 FROM ./ollama-model.gguf
 ```
 
@@ -142,7 +138,7 @@ The GGUF file location should be specified as an absolute path or relative to th
 
 The `PARAMETER` instruction defines a parameter that can be set when the model is run.
 
-```
+```modelfile
 PARAMETER <parameter> <parametervalue>
 ```
 
@@ -159,6 +155,7 @@ PARAMETER <parameter> <parametervalue>
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
 | seed           | Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt. (Default: 0)                                                                                       | int        | seed 42              |
 | stop           | Sets the stop sequences to use. When this pattern is encountered the LLM will stop generating text and return. Multiple stop patterns may be set by specifying multiple separate `stop` parameters in a modelfile.                                      | string     | stop "AI assistant:" |
+| tfs_z          | Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting. (default: 1)                                               | float      | tfs_z 1              |
 | num_predict    | Maximum number of tokens to predict when generating text. (Default: -1, infinite generation)                                                                                                                                   | int        | num_predict 42       |
 | top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
 | top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
@@ -189,7 +186,7 @@ TEMPLATE """{{ if .System }}<|im_start|>system
 
 The `SYSTEM` instruction specifies the system message to be used in the template, if applicable.
 
-```
+```modelfile
 SYSTEM """<system message>"""
 ```
 
@@ -199,7 +196,7 @@ The `ADAPTER` instruction specifies a fine tuned LoRA adapter that should apply
 
 #### Safetensor adapter
 
-```
+```modelfile
 ADAPTER <path to safetensor adapter>
 ```
 
@@ -210,7 +207,7 @@ Currently supported Safetensor adapters:
 
 #### GGUF adapter
 
-```
+```modelfile
 ADAPTER ./ollama-lora.gguf
 ```
 
@@ -218,7 +215,7 @@ ADAPTER ./ollama-lora.gguf
 
 The `LICENSE` instruction allows you to specify the legal license under which the model used with this Modelfile is shared or distributed.
 
-```
+```modelfile
 LICENSE """
 <license text>
 """
@@ -228,7 +225,7 @@ LICENSE """
 
 The `MESSAGE` instruction allows you to specify a message history for the model to use when responding. Use multiple iterations of the MESSAGE command to build up a conversation which will guide the model to answer in a similar way.
 
-```
+```modelfile
 MESSAGE <role> <message>
 ```
 
@@ -243,7 +240,7 @@ MESSAGE <role> <message>
 
 #### Example conversation
 
-```
+```modelfile
 MESSAGE user Is Toronto in Canada?
 MESSAGE assistant yes
 MESSAGE user Is Sacramento in Canada?

docs/openai.md

@@ -1,7 +1,6 @@
 # OpenAI compatibility
 
-> [!NOTE]
-> OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/ollama/ollama/blob/main/docs/api.md).
+> **Note:** OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/ollama/ollama/blob/main/docs/api.md).
 
 Ollama provides experimental compatibility with parts of the [OpenAI API](https://platform.openai.com/docs/api-reference) to help connect existing applications to Ollama.
 
@@ -60,10 +59,8 @@ embeddings = client.embeddings.create(
     input=["why is the sky blue?", "why is the grass green?"],
 )
 ```
-
 #### Structured outputs
-
-```python
+```py
 from pydantic import BaseModel
 from openai import OpenAI
 
@@ -147,7 +144,7 @@ const embedding = await openai.embeddings.create({
 
 ### `curl`
 
-```shell
+``` shell
 curl http://localhost:11434/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
@@ -322,7 +319,7 @@ ollama pull llama3.2
 
 For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:
 
-```shell
+```
 ollama cp llama3.2 gpt-3.5-turbo
 ```
 
@@ -346,7 +343,7 @@ curl http://localhost:11434/v1/chat/completions \
 
 The OpenAI API does not have a way of setting the context size for a model. If you need to change the context size, create a `Modelfile` which looks like:
 
-```
+```modelfile
 FROM <some model>
 PARAMETER num_ctx <context size>
 ```

docs/template.md

@@ -12,7 +12,7 @@ A basic Go template consists of three main parts:
 
 Here's an example of a simple chat template:
 
-```go
+```gotmpl
 {{- range .Messages }}
 {{ .Role }}: {{ .Content }}
 {{- end }}
@@ -162,6 +162,6 @@ CodeLlama [7B](https://ollama.com/library/codellama:7b-code) and [13B](https://o
 
 Codestral [22B](https://ollama.com/library/codestral:22b) supports fill-in-middle.
 
-```go
+```gotmpl
 [SUFFIX]{{ .Suffix }}[PREFIX] {{ .Prompt }}
 ```

docs/troubleshooting.md

@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:
 
 ```shell
-journalctl -u ollama --no-pager --follow --pager-end 
+journalctl -u ollama --no-pager
 ```
 
 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
@@ -17,7 +17,6 @@ When you run Ollama in a **container**, the logs go to stdout/stderr in the cont
 ```shell
 docker logs <container-name>
 ```
-
 (Use `docker ps` to find the container name)
 
 If manually running `ollama serve` in a terminal, the logs will be on that terminal.
@@ -26,9 +25,9 @@ When you run Ollama on **Windows**, there are a few different locations. You can
 - `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
+- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
 
 To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
-
 ```powershell
 $env:OLLAMA_DEBUG="1"
 & "ollama app.exe"
@@ -50,13 +49,12 @@ Dynamic LLM libraries [rocm_v6 cpu cpu_avx cpu_avx2 cuda_v11 rocm_v5]
 
 You can set OLLAMA_LLM_LIBRARY to any of the available LLM libraries to bypass autodetection, so for example, if you have a CUDA card, but want to force the CPU LLM library with AVX2 vector support, use:
 
-```shell
+```
 OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
 ```
 
 You can see what features your CPU has with the following.
-
-```shell
+```
 cat /proc/cpuinfo| grep flags | head -1
 ```
 
@@ -64,13 +62,13 @@ cat /proc/cpuinfo| grep flags | head -1
 
 If you run into problems on Linux and want to install an older version, or you'd like to try out a pre-release before it's officially released, you can tell the install script which version to install.
 
-```shell
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh
+```sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 ```
 
-## Linux docker
+## Linux tmp noexec 
 
-If Ollama initially works on the GPU in a docker container, but then switches to running on CPU after some period of time with errors in the server log reporting GPU discovery failures, this can be resolved by disabling systemd cgroup management in Docker.  Edit `/etc/docker/daemon.json` on the host and add `"exec-opts": ["native.cgroupdriver=cgroupfs"]` to the docker configuration.
+If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
 
 ## NVIDIA GPU Discovery
 
@@ -99,6 +97,8 @@ On linux, AMD GPU access typically requires `video` and/or `render` group member
 
 When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -lnd /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the **numeric** group IDs on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.   For example, in the following output `crw-rw---- 1 0  44 226,   0 Sep 16 16:55 /dev/dri/card0` the group ID column is `44` 
 
+If Ollama initially works on the GPU in a docker container, but then switches to running on CPU after some period of time with errors in the server log reporting GPU discovery failures, this can be resolved by disabling systemd cgroup management in Docker.  Edit `/etc/docker/daemon.json` on the host and add `"exec-opts": ["native.cgroupdriver=cgroupfs"]` to the docker configuration.
+
 If you are experiencing problems getting Ollama to correctly discover or use your GPU for inference, the following may help isolate the failure.
 - `AMD_LOG_LEVEL=3` Enable info log levels in the AMD HIP/ROCm libraries.  This can help show more detailed error codes that can help troubleshoot problems
 - `OLLAMA_DEBUG=1` During GPU discovery additional information will be reported

docs/windows.md

@@ -47,7 +47,6 @@ If Ollama is already running, Quit the tray application and relaunch it from the
 ## API Access
 
 Here's a quick example showing API access from `powershell`
-
 ```powershell
 (Invoke-WebRequest -method POST -Body '{"model":"llama3.2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
 ```
@@ -55,13 +54,14 @@ Here's a quick example showing API access from `powershell`
 ## Troubleshooting
 
 Ollama on Windows stores files in a few different locations.  You can view them in
-the explorer window by hitting `<Ctrl>+R` and type in:
+the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
     - *app.log* contains most resent logs from the GUI application
     - *server.log* contains the most recent server logs
     - *upgrade.log* contains log output for upgrades
 - `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
+- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
 
 ## Uninstall
 
@@ -80,11 +80,9 @@ help you keep up to date.
 
 If you'd like to install or integrate Ollama as a service, a standalone
 `ollama-windows-amd64.zip` zip file is available containing only the Ollama CLI
-and GPU library dependencies for Nvidia.  If you have an AMD GPU, also download
-and extract the additional ROCm package `ollama-windows-amd64-rocm.zip` into the
-same directory.  This allows for embedding Ollama in existing applications, or
-running it as a system service via `ollama serve` with tools such as
-[NSSM](https://nssm.cc/). 
+and GPU library dependencies for Nvidia and AMD. This allows for embedding
+Ollama in existing applications, or running it as a system service via `ollama
+serve` with tools such as [NSSM](https://nssm.cc/).
 
 > [!NOTE]  
 > If you are upgrading from a prior version, you should remove the old directories first.

envconfig/config.go

@@ -53,8 +53,8 @@ func Host() *url.URL {
 	}
 }
 
-// AllowedOrigins returns a list of allowed origins. AllowedOrigins can be configured via the OLLAMA_ORIGINS environment variable.
-func AllowedOrigins() (origins []string) {
+// Origins returns a list of allowed origins. Origins can be configured via the OLLAMA_ORIGINS environment variable.
+func Origins() (origins []string) {
 	if s := Var("OLLAMA_ORIGINS"); s != "" {
 		origins = strings.Split(s, ",")
 	}
@@ -73,7 +73,6 @@ func AllowedOrigins() (origins []string) {
 		"file://*",
 		"tauri://*",
 		"vscode-webview://*",
-		"vscode-file://*",
 	)
 
 	return origins
@@ -166,10 +165,6 @@ var (
 	IntelGPU = Bool("OLLAMA_INTEL_GPU")
 	// MultiUserCache optimizes prompt caching for multi-user scenarios
 	MultiUserCache = Bool("OLLAMA_MULTIUSER_CACHE")
-	// Enable the new Ollama engine
-	NewEngine = Bool("OLLAMA_NEW_ENGINE")
-	// ContextLength sets the default context length
-	ContextLength = Int64("OLLAMA_CONTEXT_LENGTH", -1)
 )
 
 func String(s string) func() string {
@@ -227,20 +222,6 @@ func Uint64(key string, defaultValue uint64) func() uint64 {
 	}
 }
 
-func Int64(key string, defaultValue int64) func() int64 {
-	return func() int64 {
-		if s := Var(key); s != "" {
-			if n, err := strconv.ParseInt(s, 10, 64); err != nil {
-				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
-			} else {
-				return n
-			}
-		}
-
-		return defaultValue
-	}
-}
-
 // Set aside VRAM per GPU
 var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
 
@@ -266,11 +247,9 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory(), "Do not preserve readline history"},
 		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
 		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
-		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
+		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default 4096 or 2048 with low VRAM)"},
-		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 
 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
@@ -309,3 +288,12 @@ func Values() map[string]string {
 func Var(key string) string {
 	return strings.Trim(strings.TrimSpace(os.Getenv(key)), "\"'")
 }
+
+// On windows, we keep the binary at the top directory, but
+// other platforms use a "bin" directory, so this returns ".."
+func LibRelativeToExe() string {
+	if runtime.GOOS == "windows" {
+		return "."
+	}
+	return ".."
+}

envconfig/config_test.go

@@ -69,7 +69,6 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
-			"vscode-file://*",
 		}},
 		{"http://10.0.0.1", []string{
 			"http://10.0.0.1",
@@ -89,7 +88,6 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
-			"vscode-file://*",
 		}},
 		{"http://172.16.0.1,https://192.168.0.1", []string{
 			"http://172.16.0.1",
@@ -110,7 +108,6 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
-			"vscode-file://*",
 		}},
 		{"http://totally.safe,http://definitely.legit", []string{
 			"http://totally.safe",
@@ -131,14 +128,13 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
-			"vscode-file://*",
 		}},
 	}
 	for _, tt := range cases {
 		t.Run(tt.value, func(t *testing.T) {
 			t.Setenv("OLLAMA_ORIGINS", tt.value)
 
-			if diff := cmp.Diff(AllowedOrigins(), tt.expect); diff != "" {
+			if diff := cmp.Diff(Origins(), tt.expect); diff != "" {
 				t.Errorf("%s: mismatch (-want +got):\n%s", tt.value, diff)
 			}
 		})
@@ -276,19 +272,3 @@ func TestVar(t *testing.T) {
 		})
 	}
 }
-
-func TestContextLength(t *testing.T) {
-	cases := map[string]int64{
-		"":     -1,
-		"4096": 4096,
-	}
-
-	for k, v := range cases {
-		t.Run(k, func(t *testing.T) {
-			t.Setenv("OLLAMA_CONTEXT_LENGTH", k)
-			if i := ContextLength(); i != v {
-				t.Errorf("%s: expected %d, got %d", k, v, i)
-			}
-		})
-	}
-}

format/bytes.go

@@ -40,6 +40,8 @@ func HumanBytes(b int64) string {
 	}
 
 	switch {
+	case value >= 100:
+		return fmt.Sprintf("%d %s", int(value), unit)
 	case value >= 10:
 		return fmt.Sprintf("%d %s", int(value), unit)
 	case value != math.Trunc(value):

format/bytes_test.go

@@ -1,91 +0,0 @@
-package format
-
-import (
-	"testing"
-)
-
-func TestHumanBytes(t *testing.T) {
-	type testCase struct {
-		input    int64
-		expected string
-	}
-
-	tests := []testCase{
-		// Test bytes (B)
-		{0, "0 B"},
-		{1, "1 B"},
-		{999, "999 B"},
-
-		// Test kilobytes (KB)
-		{1000, "1 KB"},
-		{1500, "1.5 KB"},
-		{999999, "999 KB"},
-
-		// Test megabytes (MB)
-		{1000000, "1 MB"},
-		{1500000, "1.5 MB"},
-		{999999999, "999 MB"},
-
-		// Test gigabytes (GB)
-		{1000000000, "1 GB"},
-		{1500000000, "1.5 GB"},
-		{999999999999, "999 GB"},
-
-		// Test terabytes (TB)
-		{1000000000000, "1 TB"},
-		{1500000000000, "1.5 TB"},
-		{1999999999999, "2.0 TB"},
-
-		// Test fractional values
-		{1234, "1.2 KB"},
-		{1234567, "1.2 MB"},
-		{1234567890, "1.2 GB"},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.expected, func(t *testing.T) {
-			result := HumanBytes(tc.input)
-			if result != tc.expected {
-				t.Errorf("Expected %s, got %s", tc.expected, result)
-			}
-		})
-	}
-}
-
-func TestHumanBytes2(t *testing.T) {
-	type testCase struct {
-		input    uint64
-		expected string
-	}
-
-	tests := []testCase{
-		// Test bytes (B)
-		{0, "0 B"},
-		{1, "1 B"},
-		{1023, "1023 B"},
-
-		// Test kibibytes (KiB)
-		{1024, "1.0 KiB"},
-		{1536, "1.5 KiB"},
-		{1048575, "1024.0 KiB"},
-
-		// Test mebibytes (MiB)
-		{1048576, "1.0 MiB"},
-		{1572864, "1.5 MiB"},
-		{1073741823, "1024.0 MiB"},
-
-		// Test gibibytes (GiB)
-		{1073741824, "1.0 GiB"},
-		{1610612736, "1.5 GiB"},
-		{2147483648, "2.0 GiB"},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.expected, func(t *testing.T) {
-			result := HumanBytes2(tc.input)
-			if result != tc.expected {
-				t.Errorf("Expected %s, got %s", tc.expected, result)
-			}
-		})
-	}
-}

format/format_test.go

@@ -12,9 +12,6 @@ func TestHumanNumber(t *testing.T) {
 
 	testCases := []testCase{
 		{0, "0"},
-		{999, "999"},
-		{1000, "1K"},
-		{1001, "1K"},
 		{1000000, "1M"},
 		{125000000, "125M"},
 		{500500000, "500.50M"},

format/time_test.go

@@ -5,7 +5,7 @@ import (
 	"time"
 )
 
-func assertEqual(t *testing.T, a any, b any) {
+func assertEqual(t *testing.T, a interface{}, b interface{}) {
 	if a != b {
 		t.Errorf("Assert failed, expected %v, got %v", b, a)
 	}

fs/config.go

@@ -1,13 +0,0 @@
-package fs
-
-type Config interface {
-	Architecture() string
-	String(string, ...string) string
-	Uint(string, ...uint32) uint32
-	Float(string, ...float32) float32
-	Bool(string, ...bool) bool
-
-	Strings(string, ...[]string) []string
-	Uints(string, ...[]uint32) []uint32
-	Floats(string, ...[]float32) []float32
-}

fs/ggml/ggml.go

@@ -100,10 +100,6 @@ func (kv KV) Float(key string, defaultValue ...float32) float32 {
 	return keyValue(kv, key, append(defaultValue, 0)...)
 }
 
-func (kv KV) Bool(key string, defaultValue ...bool) bool {
-	return keyValue(kv, key, append(defaultValue, false)...)
-}
-
 func (kv KV) Strings(key string, defaultValue ...[]string) []string {
 	r := keyValue(kv, key, &array{})
 	s := make([]string, r.size)
@@ -124,23 +120,7 @@ func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
 	return s
 }
 
-func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
-	r := keyValue(kv, key, &array{})
-	s := make([]float32, r.size)
-	for i := range r.size {
-		s[i] = float32(r.values[i].(float32))
-	}
-	return s
-}
-
-func (kv KV) OllamaEngineRequired() bool {
-	return slices.Contains([]string{
-		"gemma3",
-		"mistral3",
-	}, kv.Architecture())
-}
-
-func keyValue[T string | uint32 | uint64 | float32 | *array | bool](kv KV, key string, defaultValue ...T) T {
+func keyValue[T string | uint32 | uint64 | float32 | *array](kv KV, key string, defaultValue ...T) T {
 	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
 		key = kv.Architecture() + "." + key
 	}
@@ -173,17 +153,19 @@ func (s Tensors) Items(prefix ...string) []*Tensor {
 	return items
 }
 
-func (ts Tensors) GroupLayers() map[string]Layer {
+func (ts Tensors) Layers() map[string]Layer {
 	layers := make(map[string]Layer)
 	for _, t := range ts.items {
 		parts := strings.Split(t.Name, ".")
-		if index := slices.IndexFunc(parts, func(s string) bool { return s == "blk" || s == "mm" }); index != -1 {
-			if len(parts) > index+2 {
-				// blk and mm should have a number after them, join it
-				parts = append(
-					[]string{strings.Join(parts[:index+2], ".")},
-					parts[index+2:]...)
-			}
+		if i := slices.Index(parts, "blk"); i > 0 {
+			parts = append([]string{
+				strings.Join(parts[:i], "."),
+				strings.Join(parts[i:i+2], "."),
+			}, parts[i+2:]...)
+		} else if i == 0 {
+			parts = append([]string{
+				strings.Join(parts[i:i+2], "."),
+			}, parts[i+2:]...)
 		}
 
 		if _, ok := layers[parts[0]]; !ok {
@@ -227,26 +209,11 @@ func (t Tensor) block() (n int) {
 
 func (t Tensor) blockSize() uint64 {
 	switch t.Kind {
-	case
-		0,  // F32
-		1,  // F16
-		24, // I8
-		25, // I16
-		26, // I32
-		27, // I64
-		28, // F64
-		30: // BF16
+	case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16
 		return 1
-	case
-		2,  // Q4_0
-		3,  // Q4_1
-		6,  // Q5_0
-		7,  // Q5_1
-		8,  // Q8_0
-		9,  // Q8_1
-		20: // IQ4_NL
+	case 2, 3, 4, 5, 6, 7, 8, 9, 20: // Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1, IQ4_NL
 		return 32
-	default:
+	default: // All others
 		return 256
 	}
 }
@@ -270,7 +237,7 @@ func (t Tensor) typeSize() uint64 {
 	case 8: // Q8_0
 		return 2 + blockSize
 	case 9: // Q8_1
-		return 2 + 2 + blockSize
+		return 4 + 4 + blockSize
 	case 10: // Q2_K
 		return blockSize/16 + blockSize/4 + 2 + 2
 	case 11: // Q3_K
@@ -282,7 +249,7 @@ func (t Tensor) typeSize() uint64 {
 	case 14: // Q6_K
 		return blockSize/2 + blockSize/4 + blockSize/16 + 2
 	case 15: // Q8_K
-		return 4 + blockSize + 2*blockSize/16
+		return 2 + blockSize + 2*blockSize/16
 	case 16: // IQ2_XXS
 		return 2 + 2*blockSize/8
 	case 17: // IQ2_XS
@@ -311,8 +278,6 @@ func (t Tensor) typeSize() uint64 {
 		return 8
 	case 29: // IQ1_M
 		return blockSize/8 + blockSize/16 + blockSize/32
-	case 30: // BF16
-		return 2
 	default:
 		return 0
 	}
@@ -330,10 +295,6 @@ func (t Tensor) Size() uint64 {
 	return t.parameters() * t.typeSize() / t.blockSize()
 }
 
-func (t Tensor) Type() string {
-	return fileType(t.Kind).String()
-}
-
 type container interface {
 	Name() string
 	Decode(io.ReadSeeker) (model, error)
@@ -416,25 +377,22 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }
 
-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()
-	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array).size)
+func (llm GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+	embedding := llm.KV().EmbeddingLength()
+	heads := llm.KV().HeadCount()
+	headsKV := llm.KV().HeadCountKV()
+	vocab := uint64(llm.KV()["tokenizer.ggml.tokens"].(*array).size)
 
-	embeddingHeads := f.KV().EmbeddingHeadCount()
-	embeddingHeadsK := f.KV().EmbeddingHeadCountK()
-	embeddingHeadsV := f.KV().EmbeddingHeadCountV()
+	embeddingHeads := llm.KV().EmbeddingHeadCount()
+	embeddingHeadsK := llm.KV().EmbeddingHeadCountK()
+	embeddingHeadsV := llm.KV().EmbeddingHeadCountV()
 
-	layers := f.Tensors().GroupLayers()
+	layers := llm.Tensors().Layers()
 
 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = make([]uint64, f.KV().BlockCount())
-	for i := range kv {
-		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
-	}
+	kv = uint64(float64(context*llm.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
 
-	switch f.KV().Architecture() {
+	switch llm.KV().Architecture() {
 	case "llama":
 		fullOffload = max(
 			4*batch*(1+4*embedding+context*(1+heads)),
@@ -449,7 +407,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 
 		if ffnGateExpsWeight, ok := layers["blk.0"]["ffn_gate_exps.weight"]; ok {
 			// mixtral 8x22b
-			ff := uint64(f.KV()["llama.feed_forward_length"].(uint32))
+			ff := uint64(llm.KV()["llama.feed_forward_length"].(uint32))
 			partialOffload = max(
 				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
 				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
@@ -466,14 +424,16 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4
 
-		crossAttentionLayers := f.KV().Uints("attention.cross_attention_layers")
-		for i := range kv {
-			if slices.Contains(crossAttentionLayers, uint32(i)) {
-				kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
-					4 * // sizeof(float32)
-					visionTokens *
-					tiles
-			}
+		if crossAttentionLayers, ok := llm.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
+			kv = headsKV *
+				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
+				(2* // sizeof(float16)
+					(llm.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
+					context +
+					4* // sizeof(float32)
+						uint64(crossAttentionLayers.size)* // num cross attention layers
+						visionTokens*
+						tiles)
 		}
 
 		fullOffload = max(
@@ -483,7 +443,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 		)
 
 		var ropeFreqsCount uint64
-		if ropeFreqs, ok := f.Tensors().GroupLayers()["rope_freqs"]; ok {
+		if ropeFreqs, ok := llm.Tensors().Layers()["rope_freqs"]; ok {
 			if ropeFreqsWeights, ok := ropeFreqs["weights"]; ok {
 				ropeFreqsCount = ropeFreqsWeights.parameters()
 			}
@@ -497,7 +457,7 @@ 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":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
 			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
@@ -509,20 +469,6 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 				4*embeddingHeadsK*context*8+
 				embedding*embeddingHeadsK*heads*9/16,
 		)
-
-		// 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" {
-			const gemma3GlobalCacheCount = 6
-			slidingWindow := (uint64(numParallel) * uint64(f.KV().Uint("attention.sliding_window"))) + batch
-			for i := range kv {
-				// Every 6th layer is a global layer, which is the full context size that has already been set. The other
-				// layers are the smaller local (sliding) layers.
-				if (i+1)%gemma3GlobalCacheCount != 0 {
-					kv[i] = uint64(float64(slidingWindow*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
-				}
-			}
-		}
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
@@ -600,71 +546,21 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 	return
 }
 
-func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
-	if llm.KV().Uint("vision.block_count") == 0 {
-		return
-	}
-
-	for name, layer := range llm.Tensors().GroupLayers() {
-		if name == "v" || strings.HasPrefix(name, "v.") {
-			for _, tensor := range layer {
-				weights += tensor.Size()
-			}
-		}
-	}
-
-	imageSize := uint64(llm.KV().Uint("vision.image_size"))
-	patchSize := uint64(llm.KV().Uint("vision.patch_size"))
-	if patchSize == 0 {
-		slog.Warn("unknown patch size for vision model")
-		return
-	}
-
-	numChannels := uint64(llm.KV().Uint("vision.num_channels"))
-
-	numPatches := (imageSize / patchSize) * (imageSize / patchSize)
-	if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok {
-		numPatches++
-	}
-
-	headCount := uint64(llm.KV().Uint("vision.attention.head_count"))
-	embeddingLength := uint64(llm.KV().Uint("vision.embedding_length"))
-
-	switch llm.KV().Architecture() {
-	case "mllama":
-		numPaddedPatches := numPatches + 8 - (numPatches%8)%8
-
-		maxNumTiles := uint64(llm.KV().Uint("vision.max_num_tiles"))
-
-		graphSize = 4 * (8 +
-			imageSize*imageSize*numChannels*maxNumTiles +
-			embeddingLength*numPatches*maxNumTiles +
-			9*embeddingLength*numPaddedPatches*maxNumTiles +
-			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
-	case "gemma3", "mistral3":
-		graphSize = 4 * (imageSize*imageSize*numChannels +
-			embeddingLength*patchSize +
-			numPatches*numPatches*headCount)
-	}
-
-	return weights, graphSize
-}
-
 // SupportsKVCacheType checks if the requested cache type is supported
-func (f GGML) SupportsKVCacheType(cacheType string) bool {
+func (llm GGML) SupportsKVCacheType(cacheType string) bool {
 	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)
 }
 
 // SupportsFlashAttention checks if the model supports flash attention
-func (f GGML) SupportsFlashAttention() bool {
-	_, isEmbedding := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]
+func (llm GGML) SupportsFlashAttention() bool {
+	_, isEmbedding := llm.KV()[fmt.Sprintf("%s.pooling_type", llm.KV().Architecture())]
 	if isEmbedding {
 		return false
 	}
 
 	// Check head counts match and are non-zero
-	headCountK := f.KV().EmbeddingHeadCountK()
-	headCountV := f.KV().EmbeddingHeadCountV()
+	headCountK := llm.KV().EmbeddingHeadCountK()
+	headCountV := llm.KV().EmbeddingHeadCountV()
 	return headCountK != 0 && headCountV != 0 && headCountK == headCountV
 }
 

fs/ggml/ggml_test.go

@@ -1,212 +0,0 @@
-package ggml
-
-import (
-	"maps"
-	"slices"
-	"strconv"
-	"strings"
-	"testing"
-
-	"github.com/google/go-cmp/cmp"
-)
-
-func TestTensorLayers(t *testing.T) {
-	tensors := make(map[string]*Tensor)
-	for _, name := range []string{
-		"token_embd.weight",
-		"blk.0.attn_k.weight",
-		"blk.0.attn_output.weight",
-		"blk.0.attn_q.weight",
-		"blk.0.attn_v.weight",
-		"blk.0.attn_norm.weight",
-		"blk.0.ffn_down.weight",
-		"blk.0.ffn_gate.weight",
-		"blk.0.ffn_up.weight",
-		"blk.0.ffn_norm.weight",
-		"output_norm.weight",
-		"mm.0.bias",
-		"mm.0.weight",
-		"v.blk.0.attn_k.weight",
-		"v.blk.0.attn_output.weight",
-		"v.blk.0.attn_q.weight",
-		"v.blk.0.attn_v.weight",
-		"v.blk.0.attn_norm.weight",
-		"v.blk.0.ffn_down.weight",
-		"v.blk.0.ffn_gate.weight",
-		"v.blk.0.ffn_up.weight",
-		"v.blk.0.ffn_norm.weight",
-		"v.patch_embd.weight",
-		"v.position_embd.gate",
-		"v.position_embd.weight",
-	} {
-		tensors[name] = &Tensor{Name: name}
-	}
-
-	cases := []struct {
-		name  string
-		items []*Tensor
-		want  map[string]Layer
-	}{
-		{
-			name: "text",
-			items: slices.Collect(func(yield func(*Tensor) bool) {
-				for k, v := range tensors {
-					if !strings.HasPrefix(k, "mm.") && !strings.HasPrefix(k, "v.") {
-						if !yield(v) {
-							return
-						}
-					}
-				}
-			}),
-			want: map[string]Layer{
-				"blk.0": {
-					"attn_k.weight":      tensors["blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["blk.0.ffn_norm.weight"],
-				},
-				"token_embd":  {"weight": tensors["token_embd.weight"]},
-				"output_norm": {"weight": tensors["output_norm.weight"]},
-			},
-		},
-		{
-			name: "vision",
-			items: slices.Collect(func(yield func(*Tensor) bool) {
-				for k, v := range tensors {
-					if strings.HasPrefix(k, "mm.") || strings.HasPrefix(k, "v.") {
-						if !yield(v) {
-							return
-						}
-					}
-				}
-			}),
-			want: map[string]Layer{
-				"mm.0": {
-					"bias":   tensors["mm.0.bias"],
-					"weight": tensors["mm.0.weight"],
-				},
-				"v.blk.0": {
-					"attn_k.weight":      tensors["v.blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["v.blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["v.blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["v.blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["v.blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["v.blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["v.blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["v.blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["v.blk.0.ffn_norm.weight"],
-				},
-				"v": {
-					"patch_embd.weight":    tensors["v.patch_embd.weight"],
-					"position_embd.gate":   tensors["v.position_embd.gate"],
-					"position_embd.weight": tensors["v.position_embd.weight"],
-				},
-			},
-		},
-		{
-			name:  "vision and text",
-			items: slices.Collect(maps.Values(tensors)),
-			want: map[string]Layer{
-				"blk.0": {
-					"attn_k.weight":      tensors["blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["blk.0.ffn_norm.weight"],
-				},
-				"token_embd":  {"weight": tensors["token_embd.weight"]},
-				"output_norm": {"weight": tensors["output_norm.weight"]},
-				"mm.0": {
-					"bias":   tensors["mm.0.bias"],
-					"weight": tensors["mm.0.weight"],
-				},
-				"v.blk.0": {
-					"attn_k.weight":      tensors["v.blk.0.attn_k.weight"],
-					"attn_q.weight":      tensors["v.blk.0.attn_q.weight"],
-					"attn_v.weight":      tensors["v.blk.0.attn_v.weight"],
-					"attn_output.weight": tensors["v.blk.0.attn_output.weight"],
-					"attn_norm.weight":   tensors["v.blk.0.attn_norm.weight"],
-					"ffn_down.weight":    tensors["v.blk.0.ffn_down.weight"],
-					"ffn_gate.weight":    tensors["v.blk.0.ffn_gate.weight"],
-					"ffn_up.weight":      tensors["v.blk.0.ffn_up.weight"],
-					"ffn_norm.weight":    tensors["v.blk.0.ffn_norm.weight"],
-				},
-				"v": {
-					"patch_embd.weight":    tensors["v.patch_embd.weight"],
-					"position_embd.gate":   tensors["v.position_embd.gate"],
-					"position_embd.weight": tensors["v.position_embd.weight"],
-				},
-			},
-		},
-	}
-
-	for _, tt := range cases {
-		t.Run(tt.name, func(t *testing.T) {
-			got := Tensors{items: tt.items}.GroupLayers()
-			if diff := cmp.Diff(got, tt.want); diff != "" {
-				t.Errorf("unexpected layers (-got +want):\n%s", diff)
-			}
-		})
-	}
-}
-
-// ref: https://github.com/ggml-org/llama.cpp/blob/a82c9e7c23ef6db48cebfa194dc9cebbc4ac3552/ggml/src/ggml.c#L572
-func TestTensorTypes(t *testing.T) {
-	cases := []struct {
-		kind      uint32
-		blockSize uint64
-		typeSize  uint64
-	}{
-		{0, 1, 4},
-		{1, 1, 2},
-		{2, 32, 18},
-		{3, 32, 20},
-		{6, 32, 22},
-		{7, 32, 24},
-		{8, 32, 34},
-		{9, 32, 36},
-		{10, 256, 84},
-		{11, 256, 110},
-		{12, 256, 144},
-		{13, 256, 176},
-		{14, 256, 210},
-		{15, 256, 292},
-		{16, 256, 66},
-		{17, 256, 74},
-		{18, 256, 98},
-		{19, 256, 50},
-		{20, 32, 18},
-		{21, 256, 110},
-		{22, 256, 82},
-		{23, 256, 136},
-		{24, 1, 1},
-		{25, 1, 2},
-		{26, 1, 4},
-		{27, 1, 8},
-		{28, 1, 8},
-		{29, 256, 56},
-		{30, 1, 2},
-	}
-
-	for _, tt := range cases {
-		t.Run(strconv.Itoa(int(tt.kind)), func(t *testing.T) {
-			tensor := Tensor{Kind: tt.kind}
-			if tensor.blockSize() != tt.blockSize {
-				t.Errorf("unexpected block size: got=%d want=%d", tensor.blockSize(), tt.blockSize)
-			}
-
-			if tensor.typeSize() != tt.typeSize {
-				t.Errorf("unexpected type size: got=%d want=%d", tensor.typeSize(), tt.typeSize)
-			}
-		})
-	}
-}

fs/ggml/gguf.go

@@ -9,12 +9,8 @@ import (
 	"io"
 	"log/slog"
 	"maps"
-	"os"
-	"runtime"
 	"slices"
 	"strings"
-
-	"golang.org/x/sync/errgroup"
 )
 
 type containerGGUF struct {
@@ -239,7 +235,10 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	// patch KV with parameter count
 	llm.kv["general.parameter_count"] = llm.parameters
 
-	alignment := llm.kv.Uint("general.alignment", 32)
+	alignment, ok := llm.kv["general.alignment"].(uint32)
+	if !ok {
+		alignment = 32
+	}
 
 	offset, err := rs.Seek(0, io.SeekCurrent)
 	if err != nil {
@@ -506,22 +505,20 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func WriteGGUF(f *os.File, kv KV, ts []Tensor) error {
-	alignment := kv.Uint("general.alignment", 32)
-
-	if err := binary.Write(f, binary.LittleEndian, []byte("GGUF")); err != nil {
+func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
+	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
 		return err
 	}
 
-	if err := binary.Write(f, binary.LittleEndian, uint32(3)); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint32(3)); err != nil {
 		return err
 	}
 
-	if err := binary.Write(f, binary.LittleEndian, uint64(len(ts))); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(ts))); err != nil {
 		return err
 	}
 
-	if err := binary.Write(f, binary.LittleEndian, uint64(len(kv))); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(kv))); err != nil {
 		return err
 	}
 
@@ -529,7 +526,7 @@ func WriteGGUF(f *os.File, kv KV, ts []Tensor) error {
 	slices.Sort(keys)
 
 	for _, key := range keys {
-		if err := ggufWriteKV(f, key, kv[key]); err != nil {
+		if err := ggufWriteKV(ws, key, kv[key]); err != nil {
 			return err
 		}
 	}
@@ -545,34 +542,22 @@ func WriteGGUF(f *os.File, kv KV, ts []Tensor) error {
 	})
 
 	var s uint64
-	for i := range ts {
-		ts[i].Offset = s + uint64(ggufPadding(int64(s), int64(alignment)))
-		if err := ggufWriteTensorInfo(f, ts[i]); err != nil {
+	for _, t := range ts {
+		t.Offset = s
+		if err := ggufWriteTensorInfo(ws, t); err != nil {
 			return err
 		}
-		s += ts[i].Size()
+		s += t.Size()
 	}
 
-	offset, err := f.Seek(0, io.SeekCurrent)
-	if err != nil {
-		return err
-	}
-
-	offset += ggufPadding(offset, int64(alignment))
-	slog.Debug("gguf", "offset", offset, "size", s, "alignment", alignment)
-
-	var g errgroup.Group
-	g.SetLimit(runtime.GOMAXPROCS(0))
+	var alignment int64 = 32
 	for _, t := range ts {
-		t := t
-		w := io.NewOffsetWriter(f, offset+int64(t.Offset))
-		g.Go(func() error {
-			_, err := t.WriteTo(w)
+		if err := ggufWriteTensor(ws, t, alignment); err != nil {
 			return err
-		})
+		}
 	}
 
-	return g.Wait()
+	return nil
 }
 
 func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
@@ -657,6 +642,20 @@ func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
 	return binary.Write(ws, binary.LittleEndian, t.Offset)
 }
 
+func ggufWriteTensor(ws io.WriteSeeker, t Tensor, alignment int64) error {
+	offset, err := ws.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, binary.LittleEndian, bytes.Repeat([]byte{0}, int(ggufPadding(offset, alignment)))); err != nil {
+		return err
+	}
+
+	_, err = t.WriteTo(ws)
+	return err
+}
+
 func ggufPadding(offset, align int64) int64 {
 	return (align - offset%align) % align
 }

fs/ggml/type.go

@@ -32,10 +32,9 @@ const (
 	fileTypeIQ1_S
 	fileTypeIQ4_NL
 	fileTypeIQ3_S
-	fileTypeIQ3_M
 	fileTypeIQ2_S
-	fileTypeIQ2_M
 	fileTypeIQ4_XS
+	fileTypeIQ2_M
 	fileTypeIQ1_M
 	fileTypeBF16
 
@@ -94,14 +93,12 @@ func ParseFileType(s string) (fileType, error) {
 		return fileTypeIQ4_NL, nil
 	case "IQ3_S":
 		return fileTypeIQ3_S, nil
-	case "IQ3_M":
-		return fileTypeIQ3_M, nil
 	case "IQ2_S":
 		return fileTypeIQ2_S, nil
-	case "IQ2_M":
-		return fileTypeIQ2_M, nil
 	case "IQ4_XS":
 		return fileTypeIQ4_XS, nil
+	case "IQ2_M":
+		return fileTypeIQ2_M, nil
 	case "IQ1_M":
 		return fileTypeIQ1_M, nil
 	case "BF16":
@@ -163,8 +160,6 @@ func (t fileType) String() string {
 		return "IQ4_NL"
 	case fileTypeIQ3_S:
 		return "IQ3_S"
-	case fileTypeIQ3_M:
-		return "IQ3_M"
 	case fileTypeIQ2_S:
 		return "IQ2_S"
 	case fileTypeIQ4_XS:

go.mod

@@ -1,6 +1,6 @@
 module github.com/ollama/ollama
 
-go 1.24.0
+go 1.23.4
 
 require (
 	github.com/containerd/console v1.0.3
@@ -11,7 +11,7 @@ require (
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
-	golang.org/x/sync v0.11.0
+	golang.org/x/sync v0.10.0
 )
 
 require (
@@ -24,7 +24,8 @@ require (
 	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
+	golang.org/x/tools v0.28.0
+	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
 )
@@ -70,12 +70,12 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/crypto v0.33.0
-	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa
-	golang.org/x/net v0.35.0 // indirect
-	golang.org/x/sys v0.30.0
-	golang.org/x/term v0.29.0
-	golang.org/x/text v0.22.0
+	golang.org/x/crypto v0.31.0
+	golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa
+	golang.org/x/net v0.32.0 // indirect
+	golang.org/x/sys v0.28.0
+	golang.org/x/term v0.27.0
+	golang.org/x/text v0.21.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -214,16 +214,16 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
-golang.org/x/crypto v0.33.0 h1:IOBPskki6Lysi0lo9qQvbxiQ+FvsCC/YWOecCHAixus=
-golang.org/x/crypto v0.33.0/go.mod h1:bVdXmD7IV/4GdElGPozy6U7lWdRXA4qyRVGJV57uQ5M=
+golang.org/x/crypto v0.31.0 h1:ihbySMvVjLAeSH1IbfcRTkD/iNscyz8rGzjF/E5hV6U=
+golang.org/x/crypto v0.31.0/go.mod h1:kDsLvtWBEx7MV9tJOj9bnXsPbxwJQ6csT/x4KIN4Ssk=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190125153040-c74c464bbbf2/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190306152737-a1d7652674e8/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20191002040644-a1355ae1e2c3/go.mod h1:NOZ3BPKG0ec/BKJQgnvsSFpcKLM5xXVWnvZS97DWHgE=
-golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa h1:t2QcU6V556bFjYgu4L6C+6VrCPyJZ+eyRsABUPs1mz4=
-golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa/go.mod h1:BHOTPb3L19zxehTsLoJXVaTktb06DFgmdW6Wb9s8jqk=
+golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa h1:FRnLl4eNAQl8hwxVVC17teOw8kdjVDVAiFMtgUdTSRQ=
+golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa/go.mod h1:zk2irFbV9DP96SEBUUAy67IdHUaZuSnrz1n472HUCLE=
 golang.org/x/image v0.0.0-20180708004352-c73c2afc3b81/go.mod h1:ux5Hcp/YLpHSI86hEcLt0YII63i6oz57MZXIpbrjZUs=
 golang.org/x/image v0.0.0-20190227222117-0694c2d4d067/go.mod h1:kZ7UVZpmo3dzQBMxlp+ypCbDeSB+sBbTgSJuh5dn5js=
 golang.org/x/image v0.0.0-20190802002840-cff245a6509b/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
@@ -257,8 +257,8 @@ golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81R
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
-golang.org/x/net v0.35.0 h1:T5GQRQb2y08kTAByq9L4/bz8cipCdA8FbRTXewonqY8=
-golang.org/x/net v0.35.0/go.mod h1:EglIi67kWsHKlRzzVMUD93VMSWGFOMSZgxFjparz1Qk=
+golang.org/x/net v0.32.0 h1:ZqPmj8Kzc+Y6e0+skZsuACbx+wzMgo5MQsJh9Qd6aYI=
+golang.org/x/net v0.32.0/go.mod h1:CwU0IoeOlnQQWJ6ioyFrfRuomB8GKF6KbYXZVyeXNfs=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -268,8 +268,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.11.0 h1:GGz8+XQP4FvTTrjZPzNKTMFtSXH80RAzG+5ghFPgK9w=
-golang.org/x/sync v0.11.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
+golang.org/x/sync v0.10.0 h1:3NQrjDixjgGwUOCaF8w2+VYHv0Ve/vGYSbdkTa98gmQ=
+golang.org/x/sync v0.10.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -285,17 +285,17 @@ golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.30.0 h1:QjkSwP/36a20jFYWkSue1YwXzLmsV5Gfq7Eiy72C1uc=
-golang.org/x/sys v0.30.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
+golang.org/x/sys v0.28.0 h1:Fksou7UEQUWlKvIdsqzJmUmCX3cZuD2+P3XyyzwMhlA=
+golang.org/x/sys v0.28.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.29.0 h1:L6pJp37ocefwRRtYPKSWOWzOtWSxVajvz2ldH/xi3iU=
-golang.org/x/term v0.29.0/go.mod h1:6bl4lRlvVuDgSf3179VpIxBF0o10JUpXWOnI7nErv7s=
+golang.org/x/term v0.27.0 h1:WP60Sv1nlK1T6SupCHbXzSaN0b9wUmsPoRS9b61A23Q=
+golang.org/x/term v0.27.0/go.mod h1:iMsnZpn0cago0GOrHO2+Y7u7JPn5AylBrcoWkElMTSM=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.22.0 h1:bofq7m3/HAFvbF51jz3Q9wLg3jkvSPuiZu/pD1XwgtM=
-golang.org/x/text v0.22.0/go.mod h1:YRoo4H8PVmsu+E3Ou7cqLVH8oXWIHVoX0jqUWALQhfY=
+golang.org/x/text v0.21.0 h1:zyQAAkrwaneQ066sspRyJaG9VNi/YJ1NfzcGB3hZ/qo=
+golang.org/x/text v0.21.0/go.mod h1:4IBbMaMmOPCJ8SecivzSH54+73PCFmPWxNTLm+vZkEQ=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
@@ -309,8 +309,8 @@ golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapK
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
 golang.org/x/tools v0.1.4/go.mod h1:o0xws9oXOQQZyjljx8fwUC0k7L1pTE6eaCbjGeHmOkk=
-golang.org/x/tools v0.30.0 h1:BgcpHewrV5AUp2G9MebG4XPFI1E2W41zU1SaqVA9vJY=
-golang.org/x/tools v0.30.0/go.mod h1:c347cR/OJfw5TI+GfX7RUPNMdDRRbjvYTS0jPyvsVtY=
+golang.org/x/tools v0.28.0 h1:WuB6qZ4RPCQo5aP3WdKZS7i595EdWqWR8vqJTlwTVK8=
+golang.org/x/tools v0.28.0/go.mod h1:dcIOrVd3mfQKTgrDVQHqCPMWy6lnhfhtX3hLXYVLfRw=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

grammar/bench_test.go

@@ -0,0 +1,22 @@
+//go:build go1.24
+
+package grammar
+
+import "testing"
+
+func BenchmarkFromSchema(b *testing.B) {
+	for tt := range testCases(b) {
+		b.Run("", func(b *testing.B) {
+			s := []byte(tt.schema)
+
+			b.ReportAllocs()
+			for b.Loop() {
+				_, err := FromSchema(nil, s)
+				if err != nil {
+					b.Fatalf("GrammarFromSchema: %v", err)
+				}
+			}
+		})
+		return
+	}
+}

grammar/grammar.go

@@ -0,0 +1,227 @@
+package grammar
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"iter"
+	"strconv"
+
+	"github.com/ollama/ollama/grammar/jsonschema"
+)
+
+const jsonTerms = `
+# Unicode
+#
+# Unicode characters can be specified directly in the grammar, for example
+# hiragana ::= [ぁ-ゟ], or with escapes: 8-bit (\xXX), 16-bit (\uXXXX) or 32-bit
+# (\UXXXXXXXX).
+unicode ::= \x{hex}{2} | \u{hex}{4} | \U{hex}{8}
+
+# JSON grammar from RFC 7159
+null    ::= "null"
+object  ::= "{" (kv ("," kv)*)? "}"
+array   ::= "[" (value ("," value)*)? "]"
+kv      ::= string ":" value
+integer ::= "0" | [1-9] [0-9]*
+number  ::= "-"? integer frac? exp?
+frac    ::= "." [0-9]+
+exp     ::= ("e" | "E") ("+" | "-") [0-9]+
+string  ::= "\"" char* "\""
+escape  ::= ["/" | "b" | "f" | "n" | "r" | "t" | unicode]
+char    ::= [^"\\] | escape
+space   ::= (" " | "\t" | "\n" | "\r")*
+hex     ::= [0-9] | [a-f] | [A-F]
+boolean ::= "true" | "false"
+value   ::= object | array | string | number | boolean | "null"
+
+# User-defined
+`
+
+// FromSchema generates a grammar from a JSON schema.
+func FromSchema(buf []byte, jsonSchema []byte) ([]byte, error) {
+	var s *jsonschema.Schema
+	if err := json.Unmarshal(jsonSchema, &s); err != nil {
+		return nil, err
+	}
+
+	var g builder
+
+	// "root" is the only rule that is guaranteed to exist, so we start
+	// with its length for padding, and then adjust it as we go.
+	g.pad = len("root")
+	for id := range dependencies("root", s) {
+		g.pad = max(g.pad, len(id))
+	}
+
+	g.b.WriteString(jsonTerms)
+
+	ids := make(map[*jsonschema.Schema]string)
+	for id, s := range dependencies("root", s) {
+		ids[s] = id
+		g.define(id)
+		if err := fromSchema(&g, ids, s); err != nil {
+			return nil, err
+		}
+	}
+	g.define("root")
+	if err := fromSchema(&g, ids, s); err != nil {
+		return nil, err
+	}
+	g.define("") // finalize the last rule
+	return g.b.Bytes(), nil
+}
+
+func fromSchema(g *builder, ids map[*jsonschema.Schema]string, s *jsonschema.Schema) error {
+	switch typ := s.EffectiveType(); typ {
+	case "array":
+		if len(s.PrefixItems) == 0 && s.Items == nil {
+			g.u("array")
+		} else {
+			g.q("[")
+			for i, s := range s.PrefixItems {
+				if i > 0 {
+					g.q(",")
+				}
+				g.u(ids[s])
+			}
+			if s.Items != nil {
+				g.u("(")
+				if len(s.PrefixItems) > 0 {
+					g.q(",")
+				}
+				g.u(ids[s.Items])
+				g.u(")*")
+			}
+			g.q("]")
+		}
+	case "object":
+		if len(s.Properties) == 0 {
+			g.u("object")
+		} else {
+			g.q("{")
+			for i, p := range s.Properties {
+				name := ids[p]
+				if i > 0 {
+					g.q(",")
+				}
+				g.q(p.Name)
+				g.q(":")
+				g.u(name)
+			}
+			g.q("}")
+		}
+	case "number":
+		buildConstrainedNumber(g, s)
+	case "string":
+		if len(s.Enum) == 0 {
+			g.u("string")
+		} else {
+			g.u("(")
+			for i, e := range s.Enum {
+				if i > 0 {
+					g.q("|")
+				}
+				g.q(string(e))
+			}
+			g.u(")")
+		}
+	case "boolean", "value", "null", "integer":
+		g.u(typ)
+	default:
+		return fmt.Errorf("%s: unsupported type %q", s.Name, typ)
+	}
+	return nil
+}
+
+// dependencies returns a sequence of all child dependencies of the schema in
+// post-order.
+//
+// The first value is the id/pointer to the dependency, and the second value
+// is the schema.
+func dependencies(id string, s *jsonschema.Schema) iter.Seq2[string, *jsonschema.Schema] {
+	return func(yield func(string, *jsonschema.Schema) bool) {
+		for i, p := range s.Properties {
+			id := fmt.Sprintf("%s_%d", id, i)
+			for did, d := range dependencies(id, p) {
+				if !yield(did, d) {
+					return
+				}
+			}
+			if !yield(id, p) {
+				return
+			}
+		}
+		for i, p := range s.PrefixItems {
+			id := fmt.Sprintf("tuple_%d", i)
+			for did, d := range dependencies(id, p) {
+				id := fmt.Sprintf("%s_%s", id, did)
+				if !yield(id, d) {
+					return
+				}
+			}
+			if !yield(id, p) {
+				return
+			}
+		}
+		if s.Items != nil {
+			id := fmt.Sprintf("%s_tuple_%d", id, len(s.PrefixItems))
+			for did, d := range dependencies(id, s.Items) {
+				if !yield(did, d) {
+					return
+				}
+			}
+			if !yield(id, s.Items) {
+				return
+			}
+		}
+	}
+}
+
+type builder struct {
+	b     bytes.Buffer
+	pad   int
+	rules int
+	items int
+}
+
+// define terminates the current rule, if any, and then either starts a new
+// rule or does nothing else if the name is empty.
+func (b *builder) define(name string) {
+	if b.rules > 0 {
+		b.b.WriteString(";\n")
+	}
+	if name == "" {
+		return
+	}
+	fmt.Fprintf(&b.b, "% -*s", b.pad, name)
+	b.b.WriteString(" ::=")
+	b.rules++
+	b.items = 0
+}
+
+// quote appends a terminal to the current rule.
+func (b *builder) q(s string) {
+	if b.items > 0 {
+		b.b.WriteString(" ")
+	}
+	b.b.WriteString(" ")
+	b.b.WriteString(strconv.Quote(s))
+}
+
+// u appends a non-terminal to the current rule.
+func (b *builder) u(s string) {
+	if b.items > 0 {
+		b.b.WriteString(" ")
+	}
+	b.b.WriteString(" ")
+	b.b.WriteString(s)
+}
+
+func buildConstrainedNumber(b *builder, s *jsonschema.Schema) {
+	if s.Minimum == 0 && s.Maximum == 0 {
+		b.u("TODO")
+	} else {
+		b.u("number")
+	}
+}

grammar/grammar_test.go

@@ -0,0 +1,75 @@
+package grammar
+
+import (
+	"bufio"
+	"cmp"
+	"iter"
+	"strings"
+	"testing"
+
+	_ "embed"
+
+	"github.com/ollama/ollama/grammar/internal/diff"
+)
+
+func TestFromSchema(t *testing.T) {
+	for tt := range testCases(t) {
+		t.Run(tt.name, func(t *testing.T) {
+			g, err := FromSchema(nil, []byte(tt.schema))
+			if err != nil {
+				t.Fatalf("FromSchema: %v", err)
+			}
+			got := string(g)
+			got = strings.TrimPrefix(got, jsonTerms)
+			if got != tt.want {
+				t.Logf("schema:\n%s", tt.schema)
+				t.Fatal(string(diff.Diff("got", []byte(got), "want", []byte(tt.want))))
+			}
+		})
+	}
+}
+
+type testCase struct {
+	name   string
+	schema string
+	want   string
+}
+
+//go:embed testdata/schemas.txt
+var tests string
+
+func testCases(t testing.TB) iter.Seq[testCase] {
+	t.Helper()
+	return func(yield func(testCase) bool) {
+		t.Helper()
+		sc := bufio.NewScanner(strings.NewReader(tests))
+		name := ""
+		for sc.Scan() {
+			line := strings.TrimSpace(sc.Text())
+			if line == "" {
+				name = ""
+				continue
+			}
+			if line[0] == '#' {
+				name = cmp.Or(name, strings.TrimSpace(line[1:]))
+				continue
+			}
+			s := sc.Text()
+			g := ""
+			for sc.Scan() {
+				line = strings.TrimSpace(sc.Text())
+				if line == "" || line[0] == '#' {
+					break
+				}
+				g += sc.Text() + "\n"
+			}
+			if !yield(testCase{name, s, g}) {
+				return
+			}
+			name = strings.TrimSpace(strings.TrimPrefix(line, "#"))
+		}
+		if err := sc.Err(); err != nil {
+			t.Fatalf("error reading tests: %v", err)
+		}
+	}
+}

grammar/internal/diff/diff.go

@@ -0,0 +1,261 @@
+// Copyright 2022 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package diff
+
+import (
+	"bytes"
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// A pair is a pair of values tracked for both the x and y side of a diff.
+// It is typically a pair of line indexes.
+type pair struct{ x, y int }
+
+// Diff returns an anchored diff of the two texts old and new
+// in the “unified diff” format. If old and new are identical,
+// Diff returns a nil slice (no output).
+//
+// Unix diff implementations typically look for a diff with
+// the smallest number of lines inserted and removed,
+// which can in the worst case take time quadratic in the
+// number of lines in the texts. As a result, many implementations
+// either can be made to run for a long time or cut off the search
+// after a predetermined amount of work.
+//
+// In contrast, this implementation looks for a diff with the
+// smallest number of “unique” lines inserted and removed,
+// where unique means a line that appears just once in both old and new.
+// We call this an “anchored diff” because the unique lines anchor
+// the chosen matching regions. An anchored diff is usually clearer
+// than a standard diff, because the algorithm does not try to
+// reuse unrelated blank lines or closing braces.
+// The algorithm also guarantees to run in O(n log n) time
+// instead of the standard O(n²) time.
+//
+// Some systems call this approach a “patience diff,” named for
+// the “patience sorting” algorithm, itself named for a solitaire card game.
+// We avoid that name for two reasons. First, the name has been used
+// for a few different variants of the algorithm, so it is imprecise.
+// Second, the name is frequently interpreted as meaning that you have
+// to wait longer (to be patient) for the diff, meaning that it is a slower algorithm,
+// when in fact the algorithm is faster than the standard one.
+func Diff(oldName string, old []byte, newName string, new []byte) []byte {
+	if bytes.Equal(old, new) {
+		return nil
+	}
+	x := lines(old)
+	y := lines(new)
+
+	// Print diff header.
+	var out bytes.Buffer
+	fmt.Fprintf(&out, "diff %s %s\n", oldName, newName)
+	fmt.Fprintf(&out, "--- %s\n", oldName)
+	fmt.Fprintf(&out, "+++ %s\n", newName)
+
+	// Loop over matches to consider,
+	// expanding each match to include surrounding lines,
+	// and then printing diff chunks.
+	// To avoid setup/teardown cases outside the loop,
+	// tgs returns a leading {0,0} and trailing {len(x), len(y)} pair
+	// in the sequence of matches.
+	var (
+		done  pair     // printed up to x[:done.x] and y[:done.y]
+		chunk pair     // start lines of current chunk
+		count pair     // number of lines from each side in current chunk
+		ctext []string // lines for current chunk
+	)
+	for _, m := range tgs(x, y) {
+		if m.x < done.x {
+			// Already handled scanning forward from earlier match.
+			continue
+		}
+
+		// Expand matching lines as far as possible,
+		// establishing that x[start.x:end.x] == y[start.y:end.y].
+		// Note that on the first (or last) iteration we may (or definitely do)
+		// have an empty match: start.x==end.x and start.y==end.y.
+		start := m
+		for start.x > done.x && start.y > done.y && x[start.x-1] == y[start.y-1] {
+			start.x--
+			start.y--
+		}
+		end := m
+		for end.x < len(x) && end.y < len(y) && x[end.x] == y[end.y] {
+			end.x++
+			end.y++
+		}
+
+		// Emit the mismatched lines before start into this chunk.
+		// (No effect on first sentinel iteration, when start = {0,0}.)
+		for _, s := range x[done.x:start.x] {
+			ctext = append(ctext, "-"+s)
+			count.x++
+		}
+		for _, s := range y[done.y:start.y] {
+			ctext = append(ctext, "+"+s)
+			count.y++
+		}
+
+		// If we're not at EOF and have too few common lines,
+		// the chunk includes all the common lines and continues.
+		const C = 3 // number of context lines
+		if (end.x < len(x) || end.y < len(y)) &&
+			(end.x-start.x < C || (len(ctext) > 0 && end.x-start.x < 2*C)) {
+			for _, s := range x[start.x:end.x] {
+				ctext = append(ctext, " "+s)
+				count.x++
+				count.y++
+			}
+			done = end
+			continue
+		}
+
+		// End chunk with common lines for context.
+		if len(ctext) > 0 {
+			n := end.x - start.x
+			if n > C {
+				n = C
+			}
+			for _, s := range x[start.x : start.x+n] {
+				ctext = append(ctext, " "+s)
+				count.x++
+				count.y++
+			}
+			done = pair{start.x + n, start.y + n}
+
+			// Format and emit chunk.
+			// Convert line numbers to 1-indexed.
+			// Special case: empty file shows up as 0,0 not 1,0.
+			if count.x > 0 {
+				chunk.x++
+			}
+			if count.y > 0 {
+				chunk.y++
+			}
+			fmt.Fprintf(&out, "@@ -%d,%d +%d,%d @@\n", chunk.x, count.x, chunk.y, count.y)
+			for _, s := range ctext {
+				out.WriteString(s)
+			}
+			count.x = 0
+			count.y = 0
+			ctext = ctext[:0]
+		}
+
+		// If we reached EOF, we're done.
+		if end.x >= len(x) && end.y >= len(y) {
+			break
+		}
+
+		// Otherwise start a new chunk.
+		chunk = pair{end.x - C, end.y - C}
+		for _, s := range x[chunk.x:end.x] {
+			ctext = append(ctext, " "+s)
+			count.x++
+			count.y++
+		}
+		done = end
+	}
+
+	return out.Bytes()
+}
+
+// lines returns the lines in the file x, including newlines.
+// If the file does not end in a newline, one is supplied
+// along with a warning about the missing newline.
+func lines(x []byte) []string {
+	l := strings.SplitAfter(string(x), "\n")
+	if l[len(l)-1] == "" {
+		l = l[:len(l)-1]
+	} else {
+		// Treat last line as having a message about the missing newline attached,
+		// using the same text as BSD/GNU diff (including the leading backslash).
+		l[len(l)-1] += "\n\\ No newline at end of file\n"
+	}
+	return l
+}
+
+// tgs returns the pairs of indexes of the longest common subsequence
+// of unique lines in x and y, where a unique line is one that appears
+// once in x and once in y.
+//
+// The longest common subsequence algorithm is as described in
+// Thomas G. Szymanski, “A Special Case of the Maximal Common
+// Subsequence Problem,” Princeton TR #170 (January 1975),
+// available at https://research.swtch.com/tgs170.pdf.
+func tgs(x, y []string) []pair {
+	// Count the number of times each string appears in a and b.
+	// We only care about 0, 1, many, counted as 0, -1, -2
+	// for the x side and 0, -4, -8 for the y side.
+	// Using negative numbers now lets us distinguish positive line numbers later.
+	m := make(map[string]int)
+	for _, s := range x {
+		if c := m[s]; c > -2 {
+			m[s] = c - 1
+		}
+	}
+	for _, s := range y {
+		if c := m[s]; c > -8 {
+			m[s] = c - 4
+		}
+	}
+
+	// Now unique strings can be identified by m[s] = -1+-4.
+	//
+	// Gather the indexes of those strings in x and y, building:
+	//	xi[i] = increasing indexes of unique strings in x.
+	//	yi[i] = increasing indexes of unique strings in y.
+	//	inv[i] = index j such that x[xi[i]] = y[yi[j]].
+	var xi, yi, inv []int
+	for i, s := range y {
+		if m[s] == -1+-4 {
+			m[s] = len(yi)
+			yi = append(yi, i)
+		}
+	}
+	for i, s := range x {
+		if j, ok := m[s]; ok && j >= 0 {
+			xi = append(xi, i)
+			inv = append(inv, j)
+		}
+	}
+
+	// Apply Algorithm A from Szymanski's paper.
+	// In those terms, A = J = inv and B = [0, n).
+	// We add sentinel pairs {0,0}, and {len(x),len(y)}
+	// to the returned sequence, to help the processing loop.
+	J := inv
+	n := len(xi)
+	T := make([]int, n)
+	L := make([]int, n)
+	for i := range T {
+		T[i] = n + 1
+	}
+	for i := range n {
+		k := sort.Search(n, func(k int) bool {
+			return T[k] >= J[i]
+		})
+		T[k] = J[i]
+		L[i] = k + 1
+	}
+	k := 0
+	for _, v := range L {
+		if k < v {
+			k = v
+		}
+	}
+	seq := make([]pair, 2+k)
+	seq[1+k] = pair{len(x), len(y)} // sentinel at end
+	lastj := n
+	for i := n - 1; i >= 0; i-- {
+		if L[i] == k && J[i] < lastj {
+			seq[k] = pair{xi[i], yi[J[i]]}
+			k--
+		}
+	}
+	seq[0] = pair{0, 0} // sentinel at start
+	return seq
+}

grammar/internal/diff/diff_test.go

@@ -0,0 +1,44 @@
+// Copyright 2022 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package diff
+
+import (
+	"bytes"
+	"path/filepath"
+	"testing"
+
+	"golang.org/x/tools/txtar"
+)
+
+func clean(text []byte) []byte {
+	text = bytes.ReplaceAll(text, []byte("$\n"), []byte("\n"))
+	text = bytes.TrimSuffix(text, []byte("^D\n"))
+	return text
+}
+
+func Test(t *testing.T) {
+	files, _ := filepath.Glob("testdata/*.txt")
+	if len(files) == 0 {
+		t.Fatalf("no testdata")
+	}
+
+	for _, file := range files {
+		t.Run(filepath.Base(file), func(t *testing.T) {
+			a, err := txtar.ParseFile(file)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if len(a.Files) != 3 || a.Files[2].Name != "diff" {
+				t.Fatalf("%s: want three files, third named \"diff\"", file)
+			}
+			diffs := Diff(a.Files[0].Name, clean(a.Files[0].Data), a.Files[1].Name, clean(a.Files[1].Data))
+			want := clean(a.Files[2].Data)
+			if !bytes.Equal(diffs, want) {
+				t.Fatalf("%s: have:\n%s\nwant:\n%s\n%s", file,
+					diffs, want, Diff("have", diffs, "want", want))
+			}
+		})
+	}
+}

grammar/internal/diff/testdata/allnew.txt

@@ -0,0 +1,13 @@
+-- old --
+-- new --
+a
+b
+c
+-- diff --
+diff old new
+--- old
++++ new
+@@ -0,0 +1,3 @@
++a
++b
++c

grammar/internal/diff/testdata/allold.txt

@@ -0,0 +1,13 @@
+-- old --
+a
+b
+c
+-- new --
+-- diff --
+diff old new
+--- old
++++ new
+@@ -1,3 +0,0 @@
+-a
+-b
+-c

grammar/internal/diff/testdata/basic.txt

@@ -0,0 +1,35 @@
+Example from Hunt and McIlroy, “An Algorithm for Differential File Comparison.”
+https://www.cs.dartmouth.edu/~doug/diff.pdf
+
+-- old --
+a
+b
+c
+d
+e
+f
+g
+-- new --
+w
+a
+b
+x
+y
+z
+e
+-- diff --
+diff old new
+--- old
++++ new
+@@ -1,7 +1,7 @@
++w
+ a
+ b
+-c
+-d
++x
++y
++z
+ e
+-f
+-g

grammar/internal/diff/testdata/dups.txt

@@ -0,0 +1,40 @@
+-- old --
+a
+
+b
+
+c
+
+d
+
+e
+
+f
+-- new --
+a
+
+B
+
+C
+
+d
+
+e
+
+f
+-- diff --
+diff old new
+--- old
++++ new
+@@ -1,8 +1,8 @@
+ a
+ $
+-b
+-
+-c
++B
++
++C
+ $
+ d
+ $

grammar/internal/diff/testdata/end.txt

@@ -0,0 +1,38 @@
+-- old --
+1
+2
+3
+4
+5
+6
+7
+eight
+nine
+ten
+eleven
+-- new --
+1
+2
+3
+4
+5
+6
+7
+8
+9
+10
+-- diff --
+diff old new
+--- old
++++ new
+@@ -5,7 +5,6 @@
+ 5
+ 6
+ 7
+-eight
+-nine
+-ten
+-eleven
++8
++9
++10

grammar/internal/diff/testdata/eof.txt

@@ -0,0 +1,9 @@
+-- old --
+a
+b
+c^D
+-- new --
+a
+b
+c^D
+-- diff --

grammar/internal/diff/testdata/eof1.txt

@@ -0,0 +1,18 @@
+-- old --
+a
+b
+c
+-- new --
+a
+b
+c^D
+-- diff --
+diff old new
+--- old
++++ new
+@@ -1,3 +1,3 @@
+ a
+ b
+-c
++c
+\ No newline at end of file

grammar/internal/diff/testdata/eof2.txt

@@ -0,0 +1,18 @@
+-- old --
+a
+b
+c^D
+-- new --
+a
+b
+c
+-- diff --
+diff old new
+--- old
++++ new
+@@ -1,3 +1,3 @@
+ a
+ b
+-c
+\ No newline at end of file
++c

grammar/internal/diff/testdata/long.txt

@@ -0,0 +1,62 @@
+-- old --
+1
+2
+3
+4
+5
+6
+7
+8
+9
+10
+11
+12
+13
+14
+14½
+15
+16
+17
+18
+19
+20
+-- new --
+1
+2
+3
+4
+5
+6
+8
+9
+10
+11
+12
+13
+14
+17
+18
+19
+20
+-- diff --
+diff old new
+--- old
++++ new
+@@ -4,7 +4,6 @@
+ 4
+ 5
+ 6
+-7
+ 8
+ 9
+ 10
+@@ -12,9 +11,6 @@
+ 12
+ 13
+ 14
+-14½
+-15
+-16
+ 17
+ 18
+ 19

grammar/internal/diff/testdata/same.txt

@@ -0,0 +1,5 @@
+-- old --
+hello world
+-- new --
+hello world
+-- diff --

grammar/internal/diff/testdata/start.txt

@@ -0,0 +1,34 @@
+-- old --
+e
+pi
+4
+5
+6
+7
+8
+9
+10
+-- new --
+1
+2
+3
+4
+5
+6
+7
+8
+9
+10
+-- diff --
+diff old new
+--- old
++++ new
+@@ -1,5 +1,6 @@
+-e
+-pi
++1
++2
++3
+ 4
+ 5
+ 6

grammar/internal/diff/testdata/triv.txt

@@ -0,0 +1,40 @@
+Another example from Hunt and McIlroy,
+“An Algorithm for Differential File Comparison.”
+https://www.cs.dartmouth.edu/~doug/diff.pdf
+
+Anchored diff gives up on finding anything,
+since there are no unique lines.
+
+-- old --
+a
+b
+c
+a
+b
+b
+a
+-- new --
+c
+a
+b
+a
+b
+c
+-- diff --
+diff old new
+--- old
++++ new
+@@ -1,7 +1,6 @@
+-a
+-b
+-c
+-a
+-b
+-b
+-a
++c
++a
++b
++a
++b
++c

grammar/jsonschema/decode.go

@@ -0,0 +1,171 @@
+package jsonschema
+
+import (
+	"bytes"
+	"encoding/json"
+	"errors"
+)
+
+// Schema holds a JSON schema.
+type Schema struct {
+	// Name is the name of the property. For the parent/root property, this
+	// is "root". For child properties, this is the name of the property.
+	Name string `json:"-"`
+
+	// Type is the type of the property.
+	//
+	// TODO: Union types (e.g. make this a []string).
+	Type string
+
+	// PrefixItems is a list of schemas for each item in a tuple. By
+	// default, the tuple is "closed." unless Items is set to true or a
+	// valid Schema.
+	PrefixItems []*Schema
+
+	// Items is the schema for each item in a list.
+	//
+	// If it is missing, or its JSON value is "null" or "false", it is nil.
+	// If the JSON value is "true", it is set to the empty Schema. If the
+	// JSON value is an object, it will be decoded as a Schema.
+	Items *Schema
+
+	// MinItems specifies the minimum number of items allowed in a list.
+	MinItems int
+
+	// MaxItems specifies the maximum number of items allowed in a list.
+	MaxItems int
+
+	// Properties is the schema for each property of an object.
+	Properties []*Schema
+
+	// Format is the format of the property. This is used to validate the
+	// property against a specific format.
+	//
+	// It is the callers responsibility to validate the property against
+	// the format.
+	Format string
+
+	// Minimum specifies the minimum value for numeric properties.
+	Minimum float64
+
+	// Maximum specifies the maximum value for numeric properties.
+	Maximum float64
+
+	// Enum is a list of valid values for the property.
+	Enum []json.RawMessage
+}
+
+func (s *Schema) UnmarshalJSON(data []byte) error {
+	type S Schema
+	w := struct {
+		Properties props
+		Items      items
+		*S
+	}{
+		S: (*S)(s),
+	}
+	if err := json.Unmarshal(data, &w); err != nil {
+		return err
+	}
+	if w.Items.set {
+		s.Items = &w.Items.Schema
+	}
+	s.Properties = w.Properties
+	return nil
+}
+
+type items struct {
+	Schema
+	set bool
+}
+
+func (s *items) UnmarshalJSON(data []byte) error {
+	switch b := data[0]; b {
+	case 't':
+		*s = items{set: true}
+	case '{':
+		type I items
+		if err := json.Unmarshal(data, (*I)(s)); err != nil {
+			return err
+		}
+		s.set = true
+	case 'n', 'f':
+	default:
+		return errors.New("invalid Items")
+	}
+	return nil
+}
+
+// EffectiveType returns the effective type of the schema. If the Type field is
+// not empty, it is returned; otherwise:
+//
+//   - If the schema has both Properties and Items, it returns an empty string.
+//   - If the schema has Properties, it returns "object".
+//   - If the schema has Items, it returns "array".
+//   - If the schema has neither Properties nor Items, it returns "value".
+//
+// The returned string is never empty.
+func (d *Schema) EffectiveType() string {
+	if d.Type == "" {
+		if len(d.Properties) > 0 {
+			return "object"
+		}
+		if len(d.PrefixItems) > 0 || d.Items != nil {
+			return "array"
+		}
+		return "value"
+	}
+	return d.Type
+}
+
+// props is an ordered list of properties. The order of the properties
+// is the order in which they were defined in the schema.
+type props []*Schema
+
+var _ json.Unmarshaler = (*props)(nil)
+
+func (v *props) UnmarshalJSON(data []byte) error {
+	if len(data) == 0 {
+		return nil
+	}
+	if data[0] != '{' {
+		return errors.New("expected object")
+	}
+
+	d := json.NewDecoder(bytes.NewReader(data))
+
+	// TODO(bmizerany): Consider DisallowUnknownFields. Currently, we, like
+	// llama.cpp, ignore unknown fields, which could be lead to unexpected
+	// behavior for clients of this package, since they may not be aware
+	// that "additionalFields", "itemsPrefix", etc, are being ignored.
+	//
+	// For now, just do what llama.cpp does.
+
+	t, err := d.Token()
+	if err != nil {
+		return err
+	}
+	if t != json.Delim('{') {
+		return errors.New("expected object")
+	}
+	for d.More() {
+		// Use the first token (map key) as the property name, then
+		// decode the rest of the object fields into a Schema and
+		// append.
+		t, err := d.Token()
+		if err != nil {
+			return err
+		}
+		if t == json.Delim('}') {
+			return nil
+		}
+		s := &Schema{
+			Name: t.(string),
+		}
+		if err := d.Decode(s); err != nil {
+			return err
+		}
+		*v = append(*v, s)
+	}
+	return nil
+}

grammar/jsonschema/decode_test.go

@@ -0,0 +1,104 @@
+package jsonschema
+
+import (
+	"encoding/json"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+const testSchemaBasic = `
+{
+  "properties": {
+    "tupleClosedEmpty":   { "prefixItems": [] },
+    "tupleClosedMissing": { "prefixItems": [{}] },
+    "tupleClosedNull":    { "prefixItems": [{}], "items": null },
+    "tupleClosedFalse":   { "prefixItems": [{}], "items": false },
+    "tupleOpenTrue":      { "prefixItems": [{}], "items": true },
+    "tupleOpenEmpty":     { "prefixItems": [{}], "items": {} },
+    "tupleOpenTyped":     { "prefixItems": [{}], "items": {"type": "boolean"} },
+    "tupleOpenMax":       { "prefixItems": [{}], "items": true, "maxItems": 3},
+
+    "array": { "items": {"type": "number"} },
+
+    "null": { "type": "null" },
+    "string": { "type": "string" },
+    "boolean": { "type": "boolean" }
+  }
+}
+`
+
+func TestSchemaUnmarshal(t *testing.T) {
+	var got *Schema
+	if err := json.Unmarshal([]byte(testSchemaBasic), &got); err != nil {
+		t.Fatalf("Unmarshal: %v", err)
+	}
+	want := &Schema{
+		Properties: []*Schema{
+			{Name: "tupleClosedEmpty", PrefixItems: []*Schema{}, Items: nil},
+			{Name: "tupleClosedMissing", PrefixItems: []*Schema{{}}, Items: nil},
+			{Name: "tupleClosedNull", PrefixItems: []*Schema{{}}, Items: nil},
+			{Name: "tupleClosedFalse", PrefixItems: []*Schema{{}}, Items: nil},
+
+			{Name: "tupleOpenTrue", PrefixItems: []*Schema{{}}, Items: &Schema{}},
+			{Name: "tupleOpenEmpty", PrefixItems: []*Schema{{}}, Items: &Schema{}},
+			{Name: "tupleOpenTyped", PrefixItems: []*Schema{{}}, Items: &Schema{Type: "boolean"}},
+			{Name: "tupleOpenMax", PrefixItems: []*Schema{{}}, Items: &Schema{}, MaxItems: 3},
+
+			{Name: "array", Items: &Schema{Type: "number"}},
+
+			{Name: "null", Type: "null"},
+			{Name: "string", Type: "string"},
+			{Name: "boolean", Type: "boolean"},
+		},
+	}
+
+	if diff := cmp.Diff(want, got); diff != "" {
+		t.Errorf("(-want, +got)\n%s", diff)
+	}
+}
+
+func TestEffectiveType(t *testing.T) {
+	const schema = `
+		{"properties": {
+			"o": {"type": "object"},
+			"a": {"type": "array"},
+			"n": {"type": "number"},
+			"s": {"type": "string"},
+			"z": {"type": "null"},
+			"b": {"type": "boolean"},
+
+			"t0": {"prefixItems": [{}], "items": {"type": "number"}},
+			"t1": {"items": {"type": "number"}, "maxItems": 3},
+
+			"v": {"maxItems": 3}
+		}}
+	`
+
+	var s *Schema
+	if err := json.Unmarshal([]byte(schema), &s); err != nil {
+		t.Fatalf("json.Unmarshal: %v", err)
+	}
+
+	var got []string
+	for _, p := range s.Properties {
+		got = append(got, p.EffectiveType())
+	}
+
+	want := strings.Fields(`
+		object
+		array
+		number
+		string
+		null
+		boolean
+		array
+		array
+		value
+	`)
+	if !reflect.DeepEqual(want, got) {
+		t.Errorf("\ngot:\n\t%v\nwant:\n\t%v", got, want)
+	}
+}

grammar/testdata/schemas.txt

@@ -0,0 +1,76 @@
+# This file holds tests for JSON schema to EBNF grammar conversions.
+#
+# The format is a JSON schema, followed by the expected EBNF grammar. Each test
+# MAY be preceded by a comment that describes the test (e.g. the test name), followed by
+# the JSON schema and the expected EBNF grammar. If no comment is present, the test
+# name the tests number in the file (e.g. "#0", "#1", etc.)
+#
+# Blank lines signify the end or start of a new test. Comments can be added
+# anywhere in the file, but they must be preceded by a '#' character and start at
+# the beginning of the line.
+
+# default
+{}
+root ::= value;
+
+{"properties": {}}
+root ::= value;
+
+# array
+{"properties": {"a": {"type": "array", "items": {"type": "string"}}}}
+root_0_tuple_0 ::= string;
+root_0         ::= "[" ( root_0_tuple_0 )* "]";
+root           ::= "{" "a" ":" root_0 "}";
+
+# array with nested array
+{"type": "array", "items": {"type": "array", "items": {"type": "string"}}}
+root_tuple_0_tuple_0 ::= string;
+root_tuple_0         ::= "[" ( root_tuple_0_tuple_0 )* "]";
+root                 ::= "[" ( root_tuple_0 )* "]";
+
+# object
+{"properties": {"e": {}}}
+root_0 ::= value;
+root   ::= "{" "e" ":" root_0 "}";
+
+# object with nested object
+{"properties": {"o": {"type": "object", "properties": {"e": {}}}}}
+root_0_0 ::= value;
+root_0   ::= "{" "e" ":" root_0_0 "}";
+root     ::= "{" "o" ":" root_0 "}";
+
+# boolean
+{"type": "boolean"}
+root ::= boolean;
+
+# number
+{"properties": {"n": {"type": "number", "minimum": 123, "maximum": 4567}}}
+root_0 ::= number;
+root   ::= "{" "n" ":" root_0 "}";
+
+# string
+{"type": "string"}
+root ::= string;
+
+# string with enum
+{"type": "string", "enum": ["a", "b", "c"]}
+root ::= ( "\"a\"" "|" "\"b\"" "|" "\"c\"" );
+
+# spaces in key
+{"properties": {"a b": {}}}
+root_0 ::= value;
+root   ::= "{" "a b" ":" root_0 "}";
+
+# issue7978
+{ "type": "object", "properties": { "steps": { "type": "array", "items": { "type": "object", "properties": { "explanation": { "type": "string" }, "output": { "type": "string" } }, "required": [ "explanation", "output" ], "additionalProperties": false } }, "final_answer": { "type": "string" } }, "required": [ "steps", "final_answer" ], "additionalProperties": false }
+root_0_tuple_0_0 ::= string;
+root_0_tuple_0_1 ::= string;
+root_0_tuple_0   ::= "{" "explanation" ":" root_0_tuple_0_0 "," "output" ":" root_0_tuple_0_1 "}";
+root_0           ::= "[" ( root_0_tuple_0 )* "]";
+root_1           ::= string;
+root             ::= "{" "steps" ":" root_0 "," "final_answer" ":" root_1 "}";
+
+# !! # special characters in key
+# !! {"properties": {"a!b": {}}}
+# !! !invalid character '!' in key
+# !! 

integration/api_test.go

@@ -1,412 +0,0 @@
-//go:build integration
-
-package integration
-
-import (
-	"bytes"
-	"context"
-	"fmt"
-	"math/rand"
-	"strings"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-func TestAPIGenerate(t *testing.T) {
-	initialTimeout := 60 * time.Second
-	streamTimeout := 30 * time.Second
-	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
-	defer cancel()
-	// Set up the test data
-	req := api.GenerateRequest{
-		Model:  smol,
-		Prompt: "why is the sky blue? be brief",
-		Options: map[string]interface{}{
-			"temperature": 0,
-			"seed":        123,
-		},
-	}
-	anyResp := []string{"rayleigh", "scattering"}
-
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatalf("pull failed %s", err)
-	}
-
-	tests := []struct {
-		name   string
-		stream bool
-	}{
-		{
-			name:   "stream",
-			stream: true,
-		},
-		{
-			name:   "no_stream",
-			stream: false,
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			stallTimer := time.NewTimer(initialTimeout)
-			var buf bytes.Buffer
-			fn := func(response api.GenerateResponse) error {
-				// Fields that must always be present
-				if response.Model == "" {
-					t.Errorf("response missing model: %#v", response)
-				}
-				if response.Done {
-					// Required fields for final updates:
-					if response.DoneReason == "" && *req.Stream {
-						// TODO - is the lack of done reason on non-stream a bug?
-						t.Errorf("final response missing done_reason: %#v", response)
-					}
-					if response.Metrics.TotalDuration == 0 {
-						t.Errorf("final response missing total_duration: %#v", response)
-					}
-					if response.Metrics.LoadDuration == 0 {
-						t.Errorf("final response missing load_duration: %#v", response)
-					}
-					if response.Metrics.PromptEvalDuration == 0 {
-						t.Errorf("final response missing prompt_eval_duration: %#v", response)
-					}
-					if response.Metrics.EvalCount == 0 {
-						t.Errorf("final response missing eval_count: %#v", response)
-					}
-					if response.Metrics.EvalDuration == 0 {
-						t.Errorf("final response missing eval_duration: %#v", response)
-					}
-					if len(response.Context) == 0 {
-						t.Errorf("final response missing context: %#v", response)
-					}
-
-					// Note: caching can result in no prompt eval count, so this can't be verified reliably
-					// if response.Metrics.PromptEvalCount == 0 {
-					// 	t.Errorf("final response missing prompt_eval_count: %#v", response)
-					// }
-
-				} // else incremental response, nothing to check right now...
-				buf.Write([]byte(response.Response))
-				if !stallTimer.Reset(streamTimeout) {
-					return fmt.Errorf("stall was detected while streaming response, aborting")
-				}
-				return nil
-			}
-
-			done := make(chan int)
-			var genErr error
-			go func() {
-				req.Stream = &test.stream
-				req.Options["seed"] = rand.Int() // bust cache for prompt eval results
-				genErr = client.Generate(ctx, &req, fn)
-				done <- 0
-			}()
-
-			select {
-			case <-stallTimer.C:
-				if buf.Len() == 0 {
-					t.Errorf("generate never started.  Timed out after :%s", initialTimeout.String())
-				} else {
-					t.Errorf("generate stalled.  Response so far:%s", buf.String())
-				}
-			case <-done:
-				if genErr != nil {
-					t.Fatalf("failed with %s request prompt %s ", req.Model, req.Prompt)
-				}
-				// Verify the response contains the expected data
-				response := buf.String()
-				atLeastOne := false
-				for _, resp := range anyResp {
-					if strings.Contains(strings.ToLower(response), resp) {
-						atLeastOne = true
-						break
-					}
-				}
-				if !atLeastOne {
-					t.Errorf("none of %v found in %s", anyResp, response)
-				}
-			case <-ctx.Done():
-				t.Error("outer test context done while waiting for generate")
-			}
-		})
-	}
-
-	// Validate PS while we're at it...
-	resp, err := client.ListRunning(ctx)
-	if err != nil {
-		t.Fatalf("list models API error: %s", err)
-	}
-	if resp == nil || len(resp.Models) == 0 {
-		t.Fatalf("list models API returned empty list while model should still be loaded")
-	}
-	// Find the model we just loaded and verify some attributes
-	found := false
-	for _, model := range resp.Models {
-		if strings.Contains(model.Name, req.Model) {
-			found = true
-			if model.Model == "" {
-				t.Errorf("model field omitted: %#v", model)
-			}
-			if model.Size == 0 {
-				t.Errorf("size omitted: %#v", model)
-			}
-			if model.Digest == "" {
-				t.Errorf("digest omitted: %#v", model)
-			}
-			verifyModelDetails(t, model.Details)
-			var nilTime time.Time
-			if model.ExpiresAt == nilTime {
-				t.Errorf("expires_at omitted: %#v", model)
-			}
-			// SizeVRAM could be zero.
-		}
-	}
-	if !found {
-		t.Errorf("unable to locate running model: %#v", resp)
-	}
-}
-
-func TestAPIChat(t *testing.T) {
-	initialTimeout := 60 * time.Second
-	streamTimeout := 30 * time.Second
-	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
-	defer cancel()
-	// Set up the test data
-	req := api.ChatRequest{
-		Model: smol,
-		Messages: []api.Message{
-			{
-				Role:    "user",
-				Content: "why is the sky blue?  be brief",
-			},
-		},
-		Options: map[string]interface{}{
-			"temperature": 0,
-			"seed":        123,
-		},
-	}
-	anyResp := []string{"rayleigh", "scattering"}
-
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatalf("pull failed %s", err)
-	}
-
-	tests := []struct {
-		name   string
-		stream bool
-	}{
-		{
-			name:   "stream",
-			stream: true,
-		},
-		{
-			name:   "no_stream",
-			stream: false,
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			stallTimer := time.NewTimer(initialTimeout)
-			var buf bytes.Buffer
-			fn := func(response api.ChatResponse) error {
-				// Fields that must always be present
-				if response.Model == "" {
-					t.Errorf("response missing model: %#v", response)
-				}
-				if response.Done {
-					// Required fields for final updates:
-					var nilTime time.Time
-					if response.CreatedAt == nilTime {
-						t.Errorf("final response missing total_duration: %#v", response)
-					}
-					if response.DoneReason == "" {
-						t.Errorf("final response missing done_reason: %#v", response)
-					}
-					if response.Metrics.TotalDuration == 0 {
-						t.Errorf("final response missing total_duration: %#v", response)
-					}
-					if response.Metrics.LoadDuration == 0 {
-						t.Errorf("final response missing load_duration: %#v", response)
-					}
-					if response.Metrics.PromptEvalDuration == 0 {
-						t.Errorf("final response missing prompt_eval_duration: %#v", response)
-					}
-					if response.Metrics.EvalCount == 0 {
-						t.Errorf("final response missing eval_count: %#v", response)
-					}
-					if response.Metrics.EvalDuration == 0 {
-						t.Errorf("final response missing eval_duration: %#v", response)
-					}
-
-					if response.Metrics.PromptEvalCount == 0 {
-						t.Errorf("final response missing prompt_eval_count: %#v", response)
-					}
-				} // else incremental response, nothing to check right now...
-				buf.Write([]byte(response.Message.Content))
-				if !stallTimer.Reset(streamTimeout) {
-					return fmt.Errorf("stall was detected while streaming response, aborting")
-				}
-				return nil
-			}
-
-			done := make(chan int)
-			var genErr error
-			go func() {
-				req.Stream = &test.stream
-				req.Options["seed"] = rand.Int() // bust cache for prompt eval results
-				genErr = client.Chat(ctx, &req, fn)
-				done <- 0
-			}()
-
-			select {
-			case <-stallTimer.C:
-				if buf.Len() == 0 {
-					t.Errorf("chat never started.  Timed out after :%s", initialTimeout.String())
-				} else {
-					t.Errorf("chat stalled.  Response so far:%s", buf.String())
-				}
-			case <-done:
-				if genErr != nil {
-					t.Fatalf("failed with %s request prompt %v", req.Model, req.Messages)
-				}
-				// Verify the response contains the expected data
-				response := buf.String()
-				atLeastOne := false
-				for _, resp := range anyResp {
-					if strings.Contains(strings.ToLower(response), resp) {
-						atLeastOne = true
-						break
-					}
-				}
-				if !atLeastOne {
-					t.Errorf("none of %v found in %s", anyResp, response)
-				}
-			case <-ctx.Done():
-				t.Error("outer test context done while waiting for chat")
-			}
-		})
-	}
-}
-
-func TestAPIListModels(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	// Make sure we have at least one model so an empty list can be considered a failure
-	if err := PullIfMissing(ctx, client, smol); err != nil {
-		t.Fatalf("pull failed %s", err)
-	}
-
-	resp, err := client.List(ctx)
-	if err != nil {
-		t.Fatalf("unable to list models: %s", err)
-	}
-	if len(resp.Models) == 0 {
-		t.Fatalf("list should not be empty")
-	}
-	model := resp.Models[0]
-	if model.Name == "" {
-		t.Errorf("first model name empty: %#v", model)
-	}
-	var nilTime time.Time
-	if model.ModifiedAt == nilTime {
-		t.Errorf("first model modified_at empty: %#v", model)
-	}
-	if model.Size == 0 {
-		t.Errorf("first model size empty: %#v", model)
-	}
-	if model.Digest == "" {
-		t.Errorf("first model digest empty: %#v", model)
-	}
-	verifyModelDetails(t, model.Details)
-}
-
-func verifyModelDetails(t *testing.T, details api.ModelDetails) {
-	if details.Format == "" {
-		t.Errorf("first model details.format empty: %#v", details)
-	}
-	if details.Family == "" {
-		t.Errorf("first model details.family empty: %#v", details)
-	}
-	if details.ParameterSize == "" {
-		t.Errorf("first model details.parameter_size empty: %#v", details)
-	}
-	if details.QuantizationLevel == "" {
-		t.Errorf("first model details.quantization_level empty: %#v", details)
-	}
-}
-
-func TestAPIShowModel(t *testing.T) {
-	modelName := "llama3.2"
-	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	if err := PullIfMissing(ctx, client, modelName); err != nil {
-		t.Fatalf("pull failed %s", err)
-	}
-	resp, err := client.Show(ctx, &api.ShowRequest{Name: modelName})
-	if err != nil {
-		t.Fatalf("unable to show model: %s", err)
-	}
-	if resp.License == "" {
-		t.Errorf("%s missing license: %#v", modelName, resp)
-	}
-	if resp.Modelfile == "" {
-		t.Errorf("%s missing modelfile: %#v", modelName, resp)
-	}
-	if resp.Parameters == "" {
-		t.Errorf("%s missing parameters: %#v", modelName, resp)
-	}
-	if resp.Template == "" {
-		t.Errorf("%s missing template: %#v", modelName, resp)
-	}
-	// llama3 omits system
-	verifyModelDetails(t, resp.Details)
-	// llama3 ommits messages
-	if len(resp.ModelInfo) == 0 {
-		t.Errorf("%s missing model_info: %#v", modelName, resp)
-	}
-	// llama3 omits projectors
-	var nilTime time.Time
-	if resp.ModifiedAt == nilTime {
-		t.Errorf("%s missing modified_at: %#v", modelName, resp)
-	}
-}
-
-func TestAPIEmbeddings(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	req := api.EmbeddingRequest{
-		Model:  "orca-mini",
-		Prompt: "why is the sky blue?",
-		Options: map[string]interface{}{
-			"temperature": 0,
-			"seed":        123,
-		},
-	}
-
-	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatalf("pull failed %s", err)
-	}
-
-	resp, err := client.Embeddings(ctx, &req)
-	if err != nil {
-		t.Fatalf("embeddings call failed %s", err)
-	}
-	if len(resp.Embedding) == 0 {
-		t.Errorf("zero length embedding response")
-	}
-}

integration/basic_test.go

@@ -14,15 +14,15 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestBlueSky(t *testing.T) {
+func TestOrcaMiniBlueSky(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
-		Model:  smol,
+		Model:  "orca-mini",
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]any{
+		Options: map[string]interface{}{
 			"temperature": 0,
 			"seed":        123,
 		},
@@ -31,7 +31,6 @@ func TestBlueSky(t *testing.T) {
 }
 
 func TestUnicode(t *testing.T) {
-	skipUnderMinVRAM(t, 6)
 	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
 	// Set up the test data
@@ -40,7 +39,7 @@ func TestUnicode(t *testing.T) {
 		Model:  "deepseek-coder-v2:16b-lite-instruct-q2_K",
 		Prompt: "天空为什么是蓝色的?",
 		Stream: &stream,
-		Options: map[string]any{
+		Options: map[string]interface{}{
 			"temperature": 0,
 			"seed":        123,
 			// Workaround deepseek context shifting bug
@@ -62,7 +61,7 @@ func TestExtendedUnicodeOutput(t *testing.T) {
 		Model:  "gemma2:2b",
 		Prompt: "Output some smily face emoji",
 		Stream: &stream,
-		Options: map[string]any{
+		Options: map[string]interface{}{
 			"temperature": 0,
 			"seed":        123,
 		},
@@ -94,10 +93,10 @@ func TestUnicodeModelDir(t *testing.T) {
 	defer cancel()
 
 	req := api.GenerateRequest{
-		Model:  smol,
+		Model:  "orca-mini",
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]any{
+		Options: map[string]interface{}{
 			"temperature": 0,
 			"seed":        123,
 		},

integration/concurrency_test.go

@@ -21,11 +21,11 @@ func TestMultiModelConcurrency(t *testing.T) {
 	var (
 		req = [2]api.GenerateRequest{
 			{
-				Model:     "llama3.2:1b",
+				Model:     "orca-mini",
 				Prompt:    "why is the ocean blue?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]any{
+				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -34,7 +34,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 				Prompt:    "what is the origin of the us thanksgiving holiday?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]any{
+				Options: map[string]interface{}{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -67,7 +67,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 	wg.Wait()
 }
 
-func TestIntegrationConcurrentPredict(t *testing.T) {
+func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 	req, resp := GenerateRequests()
 	reqLimit := len(req)
 	iterLimit := 5
@@ -117,9 +117,6 @@ func TestMultiModelStress(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if maxVram < 2*format.GibiByte {
-		t.Skip("VRAM less than 2G, skipping model stress tests")
-	}
 
 	type model struct {
 		name string
@@ -128,8 +125,8 @@ func TestMultiModelStress(t *testing.T) {
 
 	smallModels := []model{
 		{
-			name: "llama3.2:1b",
-			size: 2876 * format.MebiByte,
+			name: "orca-mini",
+			size: 2992 * format.MebiByte,
 		},
 		{
 			name: "phi",