change name to context_length

Co-authored-by: Michael Yang <mxyng@pm.me>

.dockerignore

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

.gitattributes

@@ -7,18 +7,5 @@ llama/**/*.cuh linguist-vendored
 llama/**/*.m linguist-vendored
 llama/**/*.metal linguist-vendored
 
-ml/backend/**/*.c linguist-vendored
-ml/backend/**/*.h linguist-vendored
-ml/backend/**/*.cpp linguist-vendored
-ml/backend/**/*.hpp linguist-vendored
-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

@@ -1,5 +1,11 @@
 name: test
 
+env:
+  ROCM_WINDOWS_URL: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe
+  MSYS2_URL: https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-x86_64-20240727.exe
+  CUDA_12_WINDOWS_URL: https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe
+  CUDA_12_WINDOWS_VER: 12.4
+
 concurrency:
   # For PRs, later CI runs preempt previous ones. e.g. a force push on a PR
   # cancels running CI jobs and starts all new ones.
@@ -21,7 +27,7 @@ jobs:
   changes:
     runs-on: ubuntu-latest
     outputs:
-      changed: ${{ steps.changes.outputs.changed }}
+      RUNNERS: ${{ steps.changes.outputs.RUNNERS }}
     steps:
       - uses: actions/checkout@v4
         with:
@@ -29,213 +35,309 @@ jobs:
       - id: changes
         run: |
           changed() {
-            local BASE=${{ github.event.pull_request.base.sha }}
-            local HEAD=${{ github.event.pull_request.head.sha }}
-            local MERGE_BASE=$(git merge-base $BASE $HEAD)
-            git diff-tree -r --no-commit-id --name-only "$MERGE_BASE" "$HEAD" \
+            git diff-tree -r --no-commit-id --name-only \
+              $(git merge-base ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }}) \
+              ${{ github.event.pull_request.head.sha }} \
               | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
           }
 
-          echo changed=$(changed 'llama/llama.cpp/**' 'ml/backend/ggml/ggml/**') | tee -a $GITHUB_OUTPUT
+          {
+            echo RUNNERS=$(changed 'llama/**')
+          } >>$GITHUB_OUTPUT
 
-  linux:
+  runners-linux-cuda:
     needs: [changes]
-    if: needs.changes.outputs.changed == 'True'
+    if: ${{ needs.changes.outputs.RUNNERS == '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
-            extra-packages: rocm-libs
-            flags: '-DAMDGPU_TARGETS=gfx1010 -DCMAKE_PREFIX_PATH=/opt/rocm'
+        cuda-version:
+          - '11.8.0'
     runs-on: linux
-    container: ${{ matrix.container }}
+    container: nvidia/cuda:${{ matrix.cuda-version }}-devel-ubuntu20.04
     steps:
-      - uses: actions/checkout@v4
       - run: |
-          [ -n "${{ matrix.container }}" ] || sudo=sudo
-          $sudo apt-get update
-          $sudo apt-get install -y cmake ccache ${{ matrix.extra-packages }}
+          apt-get update && apt-get install -y git build-essential curl
         env:
           DEBIAN_FRONTEND: noninteractive
-      - uses: actions/cache@v4
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v4
         with:
-          path: /github/home/.cache/ccache
-          key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ matrix.preset }}
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
       - run: |
-          cmake --preset ${{ matrix.preset }} ${{ matrix.flags }}
-          cmake --build --preset ${{ matrix.preset }} --parallel
-
-  windows:
+          git config --global --add safe.directory /__w/ollama/ollama
+          cores=$(grep '^core id' /proc/cpuinfo |sort -u|wc -l)
+          make -j $cores cuda_v11
+  runners-linux-rocm:
     needs: [changes]
-    if: needs.changes.outputs.changed == 'True'
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     strategy:
       matrix:
-        include:
-          - preset: CPU
-          - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
-          - preset: ROCm
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            flags: '-DAMDGPU_TARGETS=gfx1010'
+        rocm-version:
+          - '6.1.2'
+    runs-on: linux
+    container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
+    steps:
+      - run: |
+          apt-get update && apt-get install -y git build-essential curl rocm-libs
+        env:
+          DEBIAN_FRONTEND: noninteractive
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v4
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
+      - run: |
+          git config --global --add safe.directory /__w/ollama/ollama
+          cores=$(grep '^core id' /proc/cpuinfo |sort -u|wc -l)
+          make -j $cores rocm
+
+  # ROCm generation step
+  runners-windows-rocm:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     runs-on: windows
     steps:
-      - run: |
-          choco install -y --no-progress ccache ninja
-          ccache -o cache_dir=${{ 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
+      - uses: actions/setup-go@v5
         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-version-file: go.mod
+          cache: true
+      - name: Set make jobs default
+        run: |
+          echo "MAKEFLAGS=--jobs=$((Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
 
-  go_mod_tidy:
-    runs-on: ubuntu-latest
+      # ROCM installation steps
+      - name: 'Cache ROCm installer'
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: rocm-install.exe
+          key: ${{ env.ROCM_WINDOWS_URL }}
+      - name: 'Conditionally Download ROCm'
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        run: |
+          $ErrorActionPreference = "Stop"
+          Invoke-WebRequest -Uri "${env:ROCM_WINDOWS_URL}" -OutFile "rocm-install.exe"
+      - name: 'Install ROCm'
+        run: |
+          Start-Process "rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+      - name: 'Verify ROCm'
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+          echo "HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path | select -first 1)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+
+      - name: Add msys paths
+        run: |
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+
+      - name: make rocm runner
+        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'
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
+          make -C llama print-HIP_PATH print-HIP_LIB_DIR
+          make rocm
+
+  # CUDA generation step
+  runners-windows-cuda:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    runs-on: windows
     steps:
       - uses: actions/checkout@v4
-      - name: check that 'go mod tidy' is clean
-        run: go mod tidy --diff || (echo "Please run 'go mod tidy'." && exit 1)
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: Set make jobs default
+        run: |
+          echo "MAKEFLAGS=--jobs=$((Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
 
-  test:
+      # CUDA installation steps
+      - name: 'Cache CUDA installer'
+        id: cache-cuda
+        uses: actions/cache@v4
+        with:
+          path: cuda-install.exe
+          key: ${{ env.CUDA_12_WINDOWS_URL }}
+      - name: 'Conditionally Download CUDA'
+        if: steps.cache-cuda.outputs.cache-hit != 'true'
+        run: |
+          $ErrorActionPreference = "Stop"
+          Invoke-WebRequest -Uri "${env:CUDA_12_WINDOWS_URL}" -OutFile "cuda-install.exe"
+      - name: 'Install CUDA'
+        run: |
+          $subpackages = @("cudart", "nvcc", "cublas", "cublas_dev") | foreach-object {"${_}_${{ env.CUDA_12_WINDOWS_VER }}"}
+          Start-Process "cuda-install.exe" -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
+      - name: 'Verify CUDA'
+        run: |
+          & (resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0] --version
+          $cudaPath=((resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0].path | split-path | split-path)
+          $cudaVer=($cudaPath | split-path -leaf ) -replace 'v(\d+).(\d+)', '$1_$2' 
+          echo "$cudaPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=$cudaPath" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          echo "CUDA_PATH_V${cudaVer}=$cudaPath" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          echo "CUDA_PATH_VX_Y=CUDA_PATH_V${cudaVer}" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+
+      - name: Add msys paths
+        run: |
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+      - name: make cuda runner
+        run: |
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
+          make cuda_v$(($env:CUDA_PATH | split-path -leaf) -replace 'v(\d+).*', '$1')
+
+  runners-cpu:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
+        os: [ubuntu-latest, macos-latest, windows-2019]
+        arch: [amd64, arm64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-2019
+            arch: arm64
     runs-on: ${{ matrix.os }}
     env:
+      GOARCH: ${{ matrix.arch }}
+      ARCH: ${{ matrix.arch }}
       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()
+          cache: true
+      - name: Add msys paths
+        if: ${{ startsWith(matrix.os, 'windows-') }}
         run: |
-          go generate ./...
-          git diff --name-only --exit-code || (echo "Please run 'go generate ./...'." && exit 1)
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+      - name: 'Build Windows Go Runners'
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          $gopath=(get-command go).source | split-path -parent
+          $gccpath=(get-command gcc).source | split-path -parent
+          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'
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$gccpath;$env:PATH"
+          echo $env:PATH
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
+          make -j 4
+      - name: 'Build Unix Go Runners'
+        if: ${{ ! startsWith(matrix.os, 'windows-') }}
+        run: make -j 4
+      - run: go build .
 
-      - 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.
+  lint:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-2019]
+        arch: [amd64, arm64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-2019
+            arch: arm64
+          - os: macos-latest
+            arch: amd64
+    runs-on: ${{ matrix.os }}
+    env:
+      GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Add msys paths
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: false
+      - run: |
+          case ${{ matrix.arch }} in
+            amd64) echo ARCH=x86_64 ;;
+            arm64) echo ARCH=arm64 ;;
+          esac >>$GITHUB_ENV
+        shell: bash
       - 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
+  test:
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-2019]
+        arch: [amd64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-2019
+            arch: arm64
+    runs-on: ${{ matrix.os }}
+    env:
+      GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: '1'
+    steps:
+      - uses: actions/checkout@v4
         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 }}
+          submodules: recursive
+      - name: Add msys paths
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: |
+          case ${{ matrix.arch }} in
+            amd64) echo ARCH=amd64 ;;
+            arm64) echo ARCH=arm64 ;;
+          esac >>$GITHUB_ENV
+        shell: bash
+      - run: go test ./...
 
   patches:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
-      - name: Verify patches apply cleanly and do not change files
+        with:
+          submodules: recursive
+      - name: Verify patches carry all the changes
         run: |
-          make -f Makefile.sync clean sync
-          git diff --compact-summary --exit-code
+          make apply-patches sync && git diff --compact-summary --exit-code llama

.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

@@ -1,132 +0,0 @@
-cmake_minimum_required(VERSION 3.21)
-
-project(Ollama C CXX)
-
-include(CheckLanguage)
-
-find_package(Threads REQUIRED)
-
-set(CMAKE_BUILD_TYPE Release)
-set(BUILD_SHARED_LIBS ON)
-
-set(CMAKE_CXX_STANDARD 17)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS OFF)
-
-set(GGML_BUILD ON)
-set(GGML_SHARED ON)
-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_CUDA_PEER_MAX_BATCH_SIZE 128)
-set(GGML_CUDA_GRAPHS ON)
-set(GGML_CUDA_FA 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 ${OLLAMA_BUILD_DIR})
-
-include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
-include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include)
-include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu)
-include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx)
-
-set(GGML_CPU ON)
-add_subdirectory(${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 ${OLLAMA_CUDA_INSTALL_DIR} COMPONENT CUDA
-    )
-endif()
-
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a):xnack[+-]$"
-    CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a):xnack[+-]$\"."
-)
-
-check_language(HIP)
-if(CMAKE_HIP_COMPILER)
-    set(HIP_PLATFORM "amd")
-
-    find_package(hip REQUIRED)
-    if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012])$")
-    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()
-endif()

CMakePresets.json

@@ -1,110 +0,0 @@
-{
-  "version": 3,
-  "configurePresets": [
-    {
-      "name": "Default",
-      "binaryDir": "${sourceDir}/build",
-      "installDir": "${sourceDir}/dist",
-      "cacheVariables": {
-        "CMAKE_BUILD_TYPE": "Release"
-      }
-    },
-    {
-      "name": "CPU",
-      "inherits": [ "Default" ]
-    },
-    {
-      "name": "CUDA",
-      "inherits": [ "Default" ]
-    },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86"
-      }
-    },
-    {
-      "name": "CUDA 12",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120"
-      }
-    },
-    {
-      "name": "JetPack 5",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "72;87"
-      }
-    },
-    {
-      "name": "JetPack 6",
-      "inherits": [ "CUDA" ],
-      "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "87"
-      }
-    },
-    {
-      "name": "ROCm",
-      "inherits": [ "Default" ],
-      "cacheVariables": {
-        "CMAKE_HIP_PLATFORM": "amd"
-      }
-    },
-    {
-      "name": "ROCm 6",
-      "inherits": [ "ROCm" ],
-      "cacheVariables": {
-        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
-      }
-    }
-  ],
-  "buildPresets": [
-    {
-      "name": "Default",
-      "configurePreset": "Default",
-      "configuration": "Release"
-    },
-    {
-      "name": "CPU",
-      "configurePreset": "Default",
-      "targets": [ "ggml-cpu" ]
-    },
-    {
-      "name": "CUDA",
-      "configurePreset": "CUDA",
-      "targets": [ "ggml-cuda" ]
-    },
-    {
-      "name": "CUDA 11",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 11"
-    },
-    {
-      "name": "CUDA 12",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "CUDA 12"
-    },
-    {
-      "name": "JetPack 5",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "JetPack 5"
-    },
-    {
-      "name": "JetPack 6",
-      "inherits": [ "CUDA" ],
-      "configurePreset": "JetPack 6"
-    },
-    {
-      "name": "ROCm",
-      "configurePreset": "ROCm",
-      "targets": [ "ggml-hip" ]
-    },
-    {
-      "name": "ROCm 6",
-      "inherits": [ "ROCm" ],
-      "configurePreset": "ROCm 6"
-    }
-  ]
-}

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"]

Makefile

@@ -0,0 +1,103 @@
+# top level makefile for Ollama
+include make/common-defs.make
+
+
+# Determine which if any GPU runners we should build
+include make/cuda-v11-defs.make
+include make/cuda-v12-defs.make
+include make/rocm-defs.make
+
+ifeq ($(CUSTOM_CPU_FLAGS),)
+ifeq ($(ARCH),amd64)
+	RUNNER_TARGETS=cpu
+endif
+# Without CUSTOM_CPU_FLAGS we default to build both v11 and v12 if present
+ifeq ($(OLLAMA_SKIP_CUDA_GENERATE),)
+ifneq ($(CUDA_11_COMPILER),)
+	RUNNER_TARGETS += cuda_v11
+endif
+ifneq ($(CUDA_12_COMPILER),)
+	RUNNER_TARGETS += cuda_v12
+endif
+endif
+else # CUSTOM_CPU_FLAGS is set, we'll build only the latest cuda version detected
+ifneq ($(CUDA_12_COMPILER),)
+	RUNNER_TARGETS += cuda_v12
+else ifneq ($(CUDA_11_COMPILER),)
+	RUNNER_TARGETS += cuda_v11
+endif
+endif
+
+ifeq ($(OLLAMA_SKIP_ROCM_GENERATE),)
+ifneq ($(HIP_COMPILER),)
+	RUNNER_TARGETS += rocm
+endif
+endif
+
+
+all: runners exe
+
+dist: $(addprefix dist_, $(RUNNER_TARGETS)) dist_exe
+
+dist_%:
+	@$(MAKE) --no-print-directory -f make/Makefile.$* dist
+
+runners: $(RUNNER_TARGETS)
+
+$(RUNNER_TARGETS):
+	@$(MAKE) --no-print-directory -f make/Makefile.$@
+
+exe dist_exe:
+	@$(MAKE) --no-print-directory -f make/Makefile.ollama $@
+
+help-sync apply-patches create-patches sync sync-clean:
+	@$(MAKE) --no-print-directory -f make/Makefile.sync $@
+
+test integration lint:
+	@$(MAKE) --no-print-directory -f make/Makefile.test $@
+
+clean:
+	rm -rf $(BUILD_DIR) $(DIST_LIB_DIR) $(OLLAMA_EXE) $(DIST_OLLAMA_EXE)
+	go clean -cache
+
+help:
+	@echo "The following make targets will help you build Ollama"
+	@echo ""
+	@echo "	make all   		# (default target) Build Ollama llm subprocess runners, and the primary ollama executable"
+	@echo "	make runners		# Build Ollama llm subprocess runners; after you may use 'go build .' to build the primary ollama exectuable"
+	@echo "	make <runner>		# Build specific runners. Enabled: '$(RUNNER_TARGETS)'"
+	@echo "	make dist		# Build the runners and primary ollama executable for distribution"
+	@echo "	make help-sync 		# Help information on vendor update targets"
+	@echo "	make help-runners 	# Help information on runner targets"
+	@echo ""
+	@echo "The following make targets will help you test Ollama"
+	@echo ""
+	@echo "	make test   		# Run unit tests"
+	@echo "	make integration	# Run integration tests.  You must 'make all' first"
+	@echo "	make lint   		# Run lint and style tests"
+	@echo ""
+	@echo "For more information see 'docs/development.md'"
+	@echo ""
+
+
+help-runners:
+	@echo "The following runners will be built based on discovered GPU libraries: '$(RUNNER_TARGETS)'"
+	@echo ""
+	@echo "GPU Runner CPU Flags: '$(GPU_RUNNER_CPU_FLAGS)'  (Override with CUSTOM_CPU_FLAGS)"
+	@echo ""
+	@echo "# CUDA_PATH sets the location where CUDA toolkits are present"
+	@echo "CUDA_PATH=$(CUDA_PATH)"
+	@echo "	CUDA_11_PATH=$(CUDA_11_PATH)"
+	@echo "	CUDA_11_COMPILER=$(CUDA_11_COMPILER)"
+	@echo "	CUDA_12_PATH=$(CUDA_12_PATH)"
+	@echo "	CUDA_12_COMPILER=$(CUDA_12_COMPILER)"
+	@echo ""
+	@echo "# HIP_PATH sets the location where the ROCm toolkit is present"
+	@echo "HIP_PATH=$(HIP_PATH)"
+	@echo "	HIP_COMPILER=$(HIP_COMPILER)"
+
+.PHONY: all exe dist help help-sync help-runners test integration lint runners clean $(RUNNER_TARGETS)
+
+# Handy debugging for make variables
+print-%:
+	@echo '$*=$($*)'

Makefile.sync

@@ -1,60 +0,0 @@
-UPSTREAM=https://github.com/ggerganov/llama.cpp.git
-WORKDIR=llama/vendor
-FETCH_HEAD=d7cfe1ffe0f435d0048a6058d529daf76e072d9c
-
-.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 $(MAKEFILE_LIST)) clean sync"
-
-.PHONY: sync
-sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml apply-patches
-
-.PHONY: llama/build-info.cpp
-llama/build-info.cpp: llama/build-info.cpp.in
-	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' $< > $@
-
-.PHONY: llama/llama.cpp
-llama/llama.cpp: llama/vendor/ 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
 ```
 
@@ -54,9 +54,6 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                         |
 | ------------------ | ---------- | ----- | -------------------------------- |
-| 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`         |
@@ -65,7 +62,7 @@ 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`          |
@@ -76,7 +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.
@@ -95,13 +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
    ```
 
@@ -113,7 +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
 ```
 
@@ -148,13 +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
 ```
 
@@ -162,13 +159,13 @@ ollama pull llama3.2
 
 ### Remove a model
 
-```shell
+```
 ollama rm llama3.2
 ```
 
 ### Copy a model
 
-```shell
+```
 ollama cp llama3.2 my-model
 ```
 
@@ -187,39 +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
 ```
 
@@ -235,13 +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
 ```
 
@@ -251,7 +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?"
@@ -260,7 +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": [
@@ -276,7 +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)
@@ -359,7 +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)
@@ -376,21 +369,7 @@ 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.)
 
 ### Cloud
 
@@ -434,7 +413,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### 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
@@ -449,10 +427,9 @@ 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
@@ -506,19 +483,13 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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
 
 ### 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
 
@@ -562,18 +533,13 @@ 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.

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

@@ -10,8 +10,6 @@ import (
 	"strconv"
 	"strings"
 	"time"
-
-	"github.com/ollama/ollama/envconfig"
 )
 
 // StatusError is an error with an HTTP status code and message.
@@ -361,9 +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
@@ -611,7 +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

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
 ```

cmd/cmd.go

@@ -34,9 +34,10 @@ import (
 	"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"
@@ -255,7 +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
 }
@@ -338,16 +338,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	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 {
@@ -1280,6 +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:])

cmd/cmd_test.go

@@ -10,7 +10,6 @@ import (
 	"os"
 	"strings"
 	"testing"
-	"time"
 
 	"github.com/google/go-cmp/cmp"
 	"github.com/spf13/cobra"
@@ -491,96 +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

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

@@ -9,17 +9,12 @@ import (
 	"log/slog"
 	"strings"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 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 {
@@ -32,8 +27,8 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) ggml.KV {
-	kv := ggml.KV{
+func (ModelParameters) KV(t *Tokenizer) llm.KV {
+	kv := llm.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -59,7 +54,7 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() ggml.KV {
+func (p AdapterParameters) KV() llm.KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -67,7 +62,7 @@ func (p AdapterParameters) KV() ggml.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := ggml.KV{
+	kv := llm.KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -84,19 +79,19 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-func (ModelParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(ws, kv, ts)
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
 }
 
-func (AdapterParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(ws, kv, ts)
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
 }
 
 type ModelConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) ggml.KV
+	KV(*Tokenizer) llm.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
-	Tensors([]Tensor) []ggml.Tensor
+	Tensors([]Tensor) []llm.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
@@ -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(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }
 
 type moreParser interface {
@@ -113,17 +108,17 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(ggml.KV) ggml.KV
+	KV(llm.KV) llm.KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
-	Tensors([]Tensor) []ggml.Tensor
+	Tensors([]Tensor) []llm.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
 
-	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }
 
-func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV llm.KV) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -190,8 +185,6 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		conv = &gemmaModel{}
 	case "Gemma2ForCausalLM":
 		conv = &gemma2Model{}
-	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
-		conv = &gemma3Model{Architecture: p.Architectures[0]}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":
@@ -220,14 +213,7 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) 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) {

convert/convert_bert.go

@@ -8,7 +8,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type bertModel struct {
@@ -85,7 +85,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) ggml.KV {
+func (p *bertModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false
@@ -132,8 +132,8 @@ func (p *bertModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *bertModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		if slices.Contains([]string{
 			"embeddings.position_ids",
@@ -143,7 +143,7 @@ func (p *bertModel) Tensors(ts []Tensor) []ggml.Tensor {
 			continue
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_commandr.go

@@ -3,7 +3,7 @@ package convert
 import (
 	"cmp"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type commandrModel struct {
@@ -24,7 +24,7 @@ type commandrModel struct {
 
 var _ ModelConverter = (*commandrModel)(nil)
 
-func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
+func (p *commandrModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "command-r"
 	kv["general.name"] = "command-r"
@@ -43,10 +43,10 @@ func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *commandrModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *commandrModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type gemmaModel struct {
@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
+func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings
@@ -42,14 +42,14 @@ func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *gemmaModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *gemmaModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.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)
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma2.go

@@ -1,6 +1,8 @@
 package convert
 
-import "github.com/ollama/ollama/fs/ggml"
+import (
+	"github.com/ollama/ollama/llm"
+)
 
 type gemma2Model struct {
 	gemmaModel
@@ -9,7 +11,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type gemma2Adapter struct {
@@ -15,14 +15,14 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv
 }
 
-func (p *gemma2Adapter) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -31,7 +31,7 @@ func (p *gemma2Adapter) Tensors(ts []Tensor) []ggml.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

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, 8192)
-		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

@@ -9,7 +9,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type llamaModel struct {
@@ -46,7 +46,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
+func (p *llamaModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize
@@ -120,11 +120,11 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *llamaModel) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 
 	if p.RopeScaling.factors != nil {
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     "rope_freqs.weight",
 			Kind:     0,
 			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
@@ -138,7 +138,7 @@ func (p *llamaModel) Tensors(ts []Tensor) []ggml.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama_adapter.go

@@ -7,7 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type llamaAdapter struct {
@@ -18,7 +18,7 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
 	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
@@ -29,8 +29,8 @@ func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
 	return kv
 }
 
-func (p *llamaAdapter) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -41,7 +41,7 @@ func (p *llamaAdapter) Tensors(ts []Tensor) []ggml.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    shape,

convert/convert_mixtral.go

@@ -6,7 +6,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type mixtralModel struct {
@@ -15,7 +15,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
+func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {
@@ -29,7 +29,7 @@ func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *mixtralModel) Tensors(ts []Tensor) []ggml.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -56,10 +56,10 @@ func (p *mixtralModel) Tensors(ts []Tensor) []ggml.Tensor {
 		return true
 	})
 
-	var out []ggml.Tensor
+	var out []llm.Tensor
 	for n, e := range experts {
 		// TODO(mxyng): sanity check experts
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     n,
 			Kind:     e[0].Kind(),
 			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),

convert/convert_phi3.go

@@ -8,7 +8,7 @@ import (
 	"strings"
 	"sync"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type phi3Model struct {
@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
+func (p *phi3Model) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings
@@ -68,19 +68,19 @@ func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p *phi3Model) Tensors(ts []Tensor) []ggml.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
 	var addRopeFactors sync.Once
 
-	out := make([]ggml.Tensor, 0, len(ts)+2)
+	out := make([]llm.Tensor, 0, len(ts)+2)
 	for _, t := range ts {
 		if strings.HasPrefix(t.Name(), "blk.0.") {
 			addRopeFactors.Do(func() {
-				out = append(out, ggml.Tensor{
+				out = append(out, llm.Tensor{
 					Name:     "rope_factors_long.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
 					WriterTo: p.RopeScaling.LongFactor,
-				}, ggml.Tensor{
+				}, llm.Tensor{
 					Name:     "rope_factors_short.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
@@ -89,7 +89,7 @@ func (p *phi3Model) Tensors(ts []Tensor) []ggml.Tensor {
 			})
 		}
 
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_qwen2.go

@@ -1,6 +1,6 @@
 package convert
 
-import "github.com/ollama/ollama/fs/ggml"
+import "github.com/ollama/ollama/llm"
 
 type qwen2Model struct {
 	ModelParameters
@@ -21,7 +21,7 @@ type qwen2Model struct {
 
 var _ ModelConverter = (*qwen2Model)(nil)
 
-func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
+func (q *qwen2Model) KV(t *Tokenizer) llm.KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen2"
 	kv["qwen2.block_count"] = q.HiddenLayers
@@ -45,10 +45,10 @@ func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (q *qwen2Model) Tensors(ts []Tensor) []ggml.Tensor {
-	var out []ggml.Tensor
+func (q *qwen2Model) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
-		out = append(out, ggml.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_test.go

@@ -20,7 +20,7 @@ import (
 
 	"golang.org/x/exp/maps"
 
-	"github.com/ollama/ollama/fs/ggml"
+	"github.com/ollama/ollama/llm"
 )
 
 type tensorData struct {
@@ -29,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, *llm.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -48,7 +48,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	}
 	t.Cleanup(func() { r.Close() })
 
-	m, _, err := ggml.Decode(r, math.MaxInt)
+	m, _, err := llm.DecodeGGML(r, math.MaxInt)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -60,7 +60,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv ggml.KV, tensors ggml.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv llm.KV, tensors *llm.Tensors) map[string]string {
 	actual := make(map[string]string)
 	for k, v := range kv {
 		if s, ok := v.(json.Marshaler); !ok {
@@ -75,7 +75,7 @@ func generateResultsJSON(t *testing.T, f *os.File, kv ggml.KV, tensors ggml.Tens
 		}
 	}
 
-	for _, tensor := range tensors.Items() {
+	for _, tensor := range tensors.Items {
 		sha256sum := sha256.New()
 		sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
 		if _, err := io.Copy(sha256sum, sr); err != nil {
@@ -332,7 +332,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := ggml.Decode(r, math.MaxInt)
+			m, _, err := llm.DecodeGGML(r, math.MaxInt)
 			if err != nil {
 				t.Fatal(err)
 			}

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/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 +715,28 @@ 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
+	paths := []string{}
+	appExe, err := os.Executable()
+	if err != nil {
+		slog.Warn("failed to lookup executable path", "error", err)
+	} else {
+		appRelative := filepath.Join(filepath.Dir(appExe), envconfig.LibRelativeToExe(), "lib", "ollama")
+		if _, err := os.Stat(appRelative); err == nil {
+			paths = append(paths, appRelative)
+		}
+	}
+	rDir := runners.Locate()
+	if err != nil {
+		slog.Warn("unable to locate gpu dependency libraries", "error", err)
+	} else {
+		paths = append(paths, filepath.Dir(rDir))
+	}
+	return paths
+}
+
 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/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/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
 ```
 
@@ -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):
 
@@ -795,7 +795,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ##### Request
 
-```shell
+```
 curl http://localhost:11434/api/chat -d '{
   "model": "llama3.2",
   "messages": [
@@ -870,7 +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": []
@@ -878,7 +878,6 @@ curl http://localhost:11434/api/chat -d '{
 ```
 
 ##### Response
-
 ```json
 {
   "model": "llama3.2",
@@ -898,7 +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": [],
@@ -925,7 +924,7 @@ A single JSON object is returned:
 
 ## Create a Model
 
-```
+```shell
 POST /api/create
 ```
 
@@ -1021,7 +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"}
@@ -1052,7 +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"}
@@ -1119,7 +1118,7 @@ Return 200 OK if the blob exists, 404 Not Found if it does not.
 
 ## Push a Blob
 
-```
+```shell
 POST /api/blobs/:digest
 ```
 
@@ -1143,7 +1142,7 @@ Return 201 Created if the blob was successfully created, 400 Bad Request if the
 
 ## List Local Models
 
-```
+```shell
 GET /api/tags
 ```
 
@@ -1196,7 +1195,7 @@ A single JSON object will be returned.
 
 ## Show Model Information
 
-```
+```shell
 POST /api/show
 ```
 
@@ -1262,7 +1261,7 @@ curl http://localhost:11434/api/show -d '{
 
 ## Copy a Model
 
-```
+```shell
 POST /api/copy
 ```
 
@@ -1285,7 +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
 ```
 
@@ -1311,7 +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
 ```
 
@@ -1383,7 +1382,7 @@ if `stream` is set to false, then the response is a single JSON object:
 
 ## Push a Model
 
-```
+```shell
 POST /api/push
 ```
 
@@ -1448,7 +1447,7 @@ If `stream` is set to `false`, then the response is a single JSON object:
 
 ## Generate Embeddings
 
-```
+```shell
 POST /api/embed
 ```
 
@@ -1516,7 +1515,7 @@ curl http://localhost:11434/api/embed -d '{
 ```
 
 ## List Running Models
-```
+```shell
 GET /api/ps
 ```
 
@@ -1563,7 +1562,7 @@ A single JSON object will be returned.
 
 > Note: this endpoint has been superseded by `/api/embed`
 
-```
+```shell
 POST /api/embeddings
 ```
 
@@ -1603,7 +1602,7 @@ curl http://localhost:11434/api/embeddings -d '{
 
 ## Version
 
-```
+```shell
 GET /api/version
 ```
 

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

@@ -20,11 +20,11 @@ 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 2048 tokens. This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context length to 8K, use: `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 4096
 ```
 
@@ -46,15 +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
@@ -71,7 +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.
@@ -86,14 +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
    ```
@@ -226,19 +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 ""
 ```
@@ -258,13 +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`   |

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
@@ -69,30 +69,24 @@ To use this:
 
 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 +94,13 @@ 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 +111,7 @@ Additional models can be found at:
 
 #### Build from a Safetensors model
 
-```
+```modelfile
 FROM <model directory>
 ```
 
@@ -131,7 +125,7 @@ Currently supported model architectures:
 
 #### Build from a GGUF file
 
-```
+```modelfile
 FROM ./ollama-model.gguf
 ```
 
@@ -142,7 +136,7 @@ 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>
 ```
 
@@ -189,7 +183,7 @@ 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 +193,7 @@ The `ADAPTER` instruction specifies a fine tuned LoRA adapter that should apply
 
 #### Safetensor adapter
 
-```
+```modelfile
 ADAPTER <path to safetensor adapter>
 ```
 
@@ -210,7 +204,7 @@ Currently supported Safetensor adapters:
 
 #### GGUF adapter
 
-```
+```modelfile
 ADAPTER ./ollama-lora.gguf
 ```
 
@@ -218,7 +212,7 @@ 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 +222,7 @@ 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 +237,7 @@ 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 '{
@@ -207,6 +204,45 @@ curl http://localhost:11434/v1/embeddings \
     }'
 ```
 
+## Extra arguments
+
+### Setting context length
+- `context_length` parameter can be used to set the context length for the model
+
+#### OpenAI python library 
+- OpenAI python library does not support setting context length, however this can be set for Ollama through the `extra_body` parameter
+
+```py
+completion = client.chat.completions.create(
+    model="llama3.1:8b",
+    messages=[{"role": "user", "content": "Say this is a test"}],
+    extra_body={"context_length": 4096},
+)
+```
+
+#### OpenAI JavaScript library
+- OpenAI JavaScript library does not support setting context length, however this can be set for Ollama by passing `context_length` directly with a `@ts-expect-error` as an undocumented parameter in the OpenAI JavaScript library. [See documentation here](https://github.com/openai/openai-node?tab=readme-ov-file#making-customundocumented-requests)
+
+```ts
+const chatCompletion = await openai.chat.completions.create({
+    messages: [{ role: 'user', content: 'Say this is a test' }],
+    model: 'llama3.2',
+    // @ts-expect-error context_length is an additional parameter 
+    context_length: 4096,
+})
+```
+
+#### `curl`
+```shell
+curl http://localhost:11434/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "llama3.2",
+        "messages": [{"role": "user", "content": "Say this is a test"}],
+        "context_length": 4096
+    }'
+```
+
 ## Endpoints
 
 ### `/v1/chat/completions`
@@ -216,6 +252,7 @@ curl http://localhost:11434/v1/embeddings \
 - [x] Chat completions
 - [x] Streaming
 - [x] JSON mode
+- [x] Structured outputs
 - [x] Reproducible outputs
 - [x] Vision
 - [x] Tools
@@ -322,7 +359,7 @@ 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
 ```
 
@@ -342,27 +379,3 @@ curl http://localhost:11434/v1/chat/completions \
     }'
 ```
 
-### Setting the context size
-
-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:
-
-```
-FROM <some model>
-PARAMETER num_ctx <context size>
-```
-
-Use the `ollama create mymodel` command to create a new model with the updated context size. Call the API with the updated model name:
-
-```shell
-curl http://localhost:11434/v1/chat/completions \
-    -H "Content-Type: application/json" \
-    -d '{
-        "model": "mymodel",
-        "messages": [
-            {
-                "role": "user",
-                "content": "Hello!"
-            }
-        ]
-    }'
-```

docs/troubleshooting.md

@@ -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.
@@ -29,7 +28,6 @@ When you run Ollama on **Windows**, there are a few different locations. You can
 - `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"
@@ -51,13 +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
 ```
 
@@ -65,18 +62,14 @@ 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 tmp noexec 
 
 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/
 
-## Linux docker
-
-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.
-
 ## NVIDIA GPU Discovery
 
 When Ollama starts up, it takes inventory of the GPUs present in the system to determine compatibility and how much VRAM is available.  Sometimes this discovery can fail to find your GPUs.  In general, running the latest driver will yield the best results.
@@ -104,6 +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
 ```
@@ -55,7 +54,7 @@ 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
@@ -81,11 +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 = Uint("OLLAMA_CONTEXT_LENGTH", 2048)
 )
 
 func String(s string) func() string {
@@ -252,11 +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: 2048)"},
-		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 
 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
@@ -295,3 +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]uint{
-		"":     2048,
-		"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"},

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)
-			}
-		})
-	}
-}

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
@@ -11,20 +11,18 @@ 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 (
 	github.com/agnivade/levenshtein v1.1.1
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
-	github.com/dlclark/regexp2 v1.11.4
 	github.com/emirpasic/gods/v2 v2.0.0-alpha
 	github.com/google/go-cmp v0.6.0
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	golang.org/x/image v0.22.0
-	golang.org/x/tools v0.30.0
 )
 
 require (
@@ -70,12 +68,12 @@ 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 v0.25.0 // 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

@@ -42,8 +42,6 @@ github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/dgryski/trifles v0.0.0-20200323201526-dd97f9abfb48 h1:fRzb/w+pyskVMQ+UbP35JkH8yB7MYb4q/qhBarqZE6g=
 github.com/dgryski/trifles v0.0.0-20200323201526-dd97f9abfb48/go.mod h1:if7Fbed8SFyPtHLHbg49SI7NAdJiC5WIA09pe59rfAA=
-github.com/dlclark/regexp2 v1.11.4 h1:rPYF9/LECdNymJufQKmri9gV604RvvABwgOA8un7yAo=
-github.com/dlclark/regexp2 v1.11.4/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/emirpasic/gods/v2 v2.0.0-alpha h1:dwFlh8pBg1VMOXWGipNMRt8v96dKAIvBehtCt6OtunU=
 github.com/emirpasic/gods/v2 v2.0.0-alpha/go.mod h1:W0y4M2dtBB9U5z3YlghmpuUhiaZT2h6yoeE+C1sCp6A=
 github.com/envoyproxy/go-control-plane v0.9.0/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
@@ -214,16 +212,16 @@ 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 +255,8 @@ 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 v0.25.0 h1:d/OCCoBEUq33pjydKrGQhw7IlUPI2Oylr+8qLx49kac=
+golang.org/x/net v0.25.0/go.mod h1:JkAGAh7GEvH74S6FOH42FLoXpXbE/aqXSrIQjXgsiwM=
 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 +266,8 @@ 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 +283,17 @@ 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 +307,6 @@ 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/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=

kvcache/cache.go

@@ -1,66 +0,0 @@
-package kvcache
-
-import (
-	"errors"
-
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/model/input"
-)
-
-var (
-	ErrKvCacheFull  = errors.New("could not find a kv cache slot")
-	ErrNotSupported = errors.New("model does not support operation")
-)
-
-type Cache interface {
-	// ** used by model implementations **
-
-	// SetLayer sets the active layer of the cache
-	SetLayer(layer int)
-
-	// Get returns the history of key and value tensors plus a mask
-	//
-	// The shape of the tensors is documented in the specific
-	// cache implementation used.
-	Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor)
-
-	// Put stores a batch of key and value in the cache
-	//
-	// The shape of the tensors is documented in the specific
-	// cache implementation used.
-	Put(ctx ml.Context, key, value ml.Tensor)
-
-	// SetConfig controls optimizations (mostly backend-specific) that may transform
-	// the output of the cache to work better with specific kernels. If not called,
-	// the backend settings will be used. This works well when calling Attention.
-	//
-	// The config can be overridden by models, especially if they require vanilla
-	// output when implementing their own version of attention. To do this, pass
-	// an empty ml.CacheConfig.
-	//
-	// Most models will not need to use this.
-	SetConfig(ml.CacheConfig)
-
-	// ** cache management **
-
-	// Init sets up runtime parameters
-	Init(backend ml.Backend, dtype ml.DType, capacity int32)
-
-	// Close closes the cache and frees resources associated with it
-	Close()
-
-	// StartForward is called before the start of the model's forward pass.
-	// For each token in the coming batch, there must be a corresponding
-	// entry in positions and seqs.
-	StartForward(ctx ml.Context, opts input.Options) error
-
-	// CopyPrefix copies tokens in the range [0, len) from srcSeq to dstSeq
-	CopyPrefix(srcSeq, dstSeq int, len int32)
-
-	// Remove deletes tokens in the range [beginIndex, endIndex) from seq. Set
-	// endIndex to math.MaxInt32 to remove everything starting at beginIndex.
-	//
-	// If an error occurs, the entire context for the sequence should be
-	// removed by calling Remove(seq, 0, math.MaxInt32)
-	Remove(seq int, beginIndex, endIndex int32) error
-}

kvcache/causal.go

@@ -1,630 +0,0 @@
-package kvcache
-
-import (
-	"errors"
-	"fmt"
-	"log/slog"
-	"math"
-	"slices"
-
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/model/input"
-)
-
-type shiftFn func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error)
-
-// Causal cache stores K and V tensors according to their position in the
-// sequence. Returns the history and a mask for attending to past tokens
-//
-// The tensors are of shape embed dim, kv heads, batch size
-// The mask is of shape history size, batch size
-type Causal struct {
-	DType      ml.DType
-	Capacity   int32
-	windowSize int32
-
-	opts CausalOptions
-
-	// config controls mostly backend-specific optimizations
-	config *ml.CacheConfig
-
-	// ** current forward pass **
-
-	// the active layer for Get and Put
-	curLayer int
-
-	// starting location for data storage for this batch
-	curLoc int
-
-	// size of the current batch
-	curBatchSize int
-
-	// mask of the cache as used by this batch
-	curMask ml.Tensor
-
-	// locations in the cache that are needed for this batch
-	curCellRange cellRange
-
-	// curSequences is the sequences corresponding to this pass's entries in the cache
-	curSequences []int
-
-	// curPositions is the positions corresponding to this pass's entries in the cache
-	curPositions []int32
-
-	// ** cache metadata **
-
-	// for each possible location in the cache, stores the position and set of sequences
-	// that reference the data there
-	cells []cacheCell
-
-	// maps from sequence to the range of locations where it is stored in the cache
-	cellRanges map[int]cellRange
-
-	// ** cache data storage **
-
-	shiftFn      shiftFn
-	backend      ml.Backend
-	ctxs         map[int]ml.Context
-	keys, values map[int]ml.Tensor
-}
-
-type cacheCell struct {
-	pos       int32
-	sequences []int
-}
-
-type cellRange struct {
-	min int
-	max int
-}
-
-func NewCausalCache(shift shiftFn) *Causal {
-	return &Causal{
-		windowSize: math.MaxInt32,
-		shiftFn:    shift,
-		ctxs:       make(map[int]ml.Context),
-		keys:       make(map[int]ml.Tensor),
-		values:     make(map[int]ml.Tensor),
-	}
-}
-
-func NewSWACache(windowSize int32, shift shiftFn) *Causal {
-	return &Causal{
-		windowSize: windowSize,
-		shiftFn:    shift,
-		ctxs:       make(map[int]ml.Context),
-		keys:       make(map[int]ml.Tensor),
-		values:     make(map[int]ml.Tensor),
-	}
-}
-
-func (c *Causal) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
-	if c.config == nil {
-		var config ml.CacheConfig
-		if cc, ok := backend.(ml.BackendCacheConfig); ok {
-			config = cc.CacheConfig()
-		}
-		c.config = &config
-	}
-
-	if c.config.CachePadding == 0 {
-		c.config.CachePadding = 1
-	}
-
-	if c.config.MaskBatchPadding == 0 {
-		c.config.MaskBatchPadding = 1
-	}
-
-	if c.config.MaskDType == ml.DTypeOther {
-		c.config.MaskDType = ml.DTypeF32
-	}
-
-	c.DType = dtype
-	c.Capacity = int32(roundUp(int(capacity), c.config.CachePadding))
-	c.cells = make([]cacheCell, c.Capacity)
-	c.cellRanges = make(map[int]cellRange)
-	c.backend = backend
-}
-
-func (c *Causal) SetConfig(config ml.CacheConfig) {
-	if c.config != nil {
-		panic("config cannot be changed after being previously set, either by the model or backend")
-	}
-
-	c.config = &config
-}
-
-func (c *Causal) Close() {
-	for _, ctx := range c.ctxs {
-		ctx.Close()
-	}
-}
-
-func (c *Causal) StartForward(ctx ml.Context, opts input.Options) error {
-	c.curBatchSize = len(opts.Positions)
-	c.curSequences = opts.Sequences
-	c.curPositions = opts.Positions
-	c.opts.Except = nil
-
-	var err error
-	c.curLoc, err = c.findStartLoc()
-	if errors.Is(err, ErrKvCacheFull) {
-		c.defrag()
-		c.curLoc, err = c.findStartLoc()
-	}
-	if err != nil {
-		return err
-	}
-
-	c.curCellRange = newRange()
-	for i, pos := range opts.Positions {
-		seq := opts.Sequences[i]
-
-		c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
-
-		seqRange, ok := c.cellRanges[seq]
-		if !ok {
-			seqRange = newRange()
-		}
-
-		if c.curLoc+i > seqRange.max {
-			seqRange.max = c.curLoc + i
-		}
-		if seqRange.max > c.curCellRange.max {
-			c.curCellRange.max = seqRange.max
-		}
-
-		if c.curLoc+i < seqRange.min {
-			seqRange.min = c.curLoc + i
-		}
-		if seqRange.min < c.curCellRange.min {
-			c.curCellRange.min = seqRange.min
-		}
-		c.cellRanges[seq] = seqRange
-	}
-
-	c.curMask, err = c.buildMask(ctx)
-
-	return err
-}
-
-func newRange() cellRange {
-	return cellRange{
-		min: math.MaxInt,
-		max: 0,
-	}
-}
-
-// Find the first contiguous block of at least curBatchSize
-func (c *Causal) findStartLoc() (int, error) {
-	var start, count int
-	for i := range c.cells {
-		if len(c.cells[i].sequences) == 0 {
-			count++
-			if count >= c.curBatchSize {
-				return start, nil
-			}
-		} else {
-			start = i + 1
-			count = 0
-		}
-	}
-
-	return 0, fmt.Errorf("%w (length: %v)", ErrKvCacheFull, c.Capacity)
-}
-
-func roundDown(length, pad int) int {
-	return (length / pad) * pad
-}
-
-func roundUp(length, pad int) int {
-	return ((length + pad - 1) / pad) * pad
-}
-
-// Builds a mask of history x batch indicating whether for each token in the batch the
-// token in the history should apply. This is based on both the sequence and causality (the
-// position of the history is not ahead of the token in the batch).
-func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
-	// Align and pad the two dimensions as required by the backend
-	batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)
-
-	c.curCellRange.min = roundDown(c.curCellRange.min, c.config.CachePadding)
-	c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1
-
-	length := c.curCellRange.max - c.curCellRange.min + 1
-	mask := make([]float32, batchSize*length)
-
-	for i := range c.curBatchSize {
-		enabled := !slices.Contains(c.opts.Except, i)
-		for j := c.curCellRange.min; j <= c.curCellRange.max; j++ {
-			if !slices.Contains(c.cells[j].sequences, c.curSequences[i]) ||
-				(enabled && c.cells[j].pos > c.curPositions[i]) ||
-				c.cells[j].pos < c.curPositions[i]-c.windowSize {
-				mask[i*length+(j-c.curCellRange.min)] = float32(math.Inf(-1))
-			}
-		}
-	}
-
-	// Mask out any padding tokens we added. For padding that we added to the cache history, this
-	// has already been masked out because the sequence doesn't match.
-	for i := c.curBatchSize * length; i < len(mask); i++ {
-		mask[i] = float32(math.Inf(-1))
-	}
-
-	maskTensor, err := ctx.Input().FromFloatSlice(mask, length, batchSize)
-	if err != nil {
-		return nil, err
-	}
-
-	if c.config.MaskDType != ml.DTypeF32 {
-		out := ctx.Input().Empty(c.config.MaskDType, maskTensor.Shape()...)
-		ctx.Forward(maskTensor.Copy(ctx, out))
-		maskTensor = out
-	}
-
-	return maskTensor, nil
-}
-
-func (c *Causal) moveCells(ctx ml.Context, src, dst, len int) {
-	for i, key := range c.keys {
-		if key == nil {
-			continue
-		}
-
-		kHeadDim := key.Dim(0)
-		numKVHeads := key.Dim(1)
-		rowSize := key.Stride(2)
-
-		kSrcView := key.View(ctx, rowSize*src, kHeadDim*numKVHeads*len)
-		kDstView := key.View(ctx, rowSize*dst, kHeadDim*numKVHeads*len)
-
-		value := c.values[i]
-		var vSrcView, vDstView ml.Tensor
-		if c.config.PermutedV {
-			vHeadDim := value.Dim(1)
-			elemSize := value.Stride(0)
-
-			vSrcView = value.View(ctx, elemSize*src, len, int(c.Capacity)*elemSize, vHeadDim*numKVHeads)
-			vDstView = value.View(ctx, elemSize*dst, len, int(c.Capacity)*elemSize, vHeadDim*numKVHeads)
-		} else {
-			vHeadDim := value.Dim(0)
-			rowSize := value.Stride(2)
-
-			vSrcView = value.View(ctx, rowSize*src, vHeadDim*numKVHeads*len)
-			vDstView = value.View(ctx, rowSize*dst, vHeadDim*numKVHeads*len)
-		}
-
-		ctx.Forward(
-			kSrcView.Copy(ctx, kDstView),
-			vSrcView.Copy(ctx, vDstView),
-		)
-	}
-}
-
-func (c *Causal) defrag() {
-	slog.Debug("defragmenting kv cache")
-
-	// Defrag strategy:
-	// - Search for empty holes at the beginning of the cache,
-	//   filling them with active data starting at the end
-	// - If there are contiguous elements that need to be moved,
-	//   combine them into a single operation by holding new moves
-	//   until we see that the next one is non-contiguous
-	// - Fill up the context with the maximum number of operations it
-	//   can hold then compute that and continue with a new context
-	//
-	// We could try to optimize placement by grouping blocks from
-	// the same sequences together but most likely the next forward
-	// pass will disrupt this anyways, so the real world benefit
-	// seems limited as this time.
-
-	ctx := c.backend.NewContext()
-
-	// For every move, 6 tensors are required per layer (2 views and a
-	// copy for each of k and v).
-	layers := 0
-	for _, key := range c.keys {
-		if key == nil {
-			continue
-		}
-		layers++
-	}
-
-	maxMoves := ctx.MaxGraphNodes() / (6 * layers)
-	moves := 0
-
-	var pendingSrc, pendingDst, pendingLen int
-	src := len(c.cells) - 1
-
-	for dst := 0; dst < src; dst++ {
-		if len(c.cells[dst].sequences) == 0 {
-			for ; src > dst; src-- {
-				if len(c.cells[src].sequences) != 0 {
-					c.cells[dst] = c.cells[src]
-					c.cells[src] = cacheCell{}
-
-					if pendingLen > 0 {
-						if src == pendingSrc-pendingLen && dst == pendingDst+pendingLen {
-							pendingSrc = src
-							pendingLen++
-							break
-						} else {
-							c.moveCells(ctx, pendingSrc, pendingDst, pendingLen)
-							moves++
-						}
-					}
-
-					pendingSrc = src
-					pendingDst = dst
-					pendingLen = 1
-
-					break
-				}
-			}
-		}
-
-		if moves >= maxMoves {
-			ctx.Compute()
-			ctx.Close()
-			ctx = c.backend.NewContext()
-
-			moves = 0
-		}
-	}
-
-	if pendingLen > 0 {
-		c.moveCells(ctx, pendingSrc, pendingDst, pendingLen)
-		moves++
-	}
-
-	if moves > 0 {
-		ctx.Compute()
-	}
-	ctx.Close()
-
-	// Reset range metadata
-	for seq := range c.cellRanges {
-		seqRange := newRange()
-
-		for i, cell := range c.cells {
-			if slices.Contains(cell.sequences, seq) {
-				if i < seqRange.min {
-					seqRange.min = i
-				}
-				if i > seqRange.max {
-					seqRange.max = i
-				}
-			}
-		}
-
-		c.cellRanges[seq] = seqRange
-	}
-}
-
-func (c *Causal) SetLayer(layer int) {
-	c.curLayer = layer
-}
-
-type CausalOptions struct {
-	// Enabled controls whether the causal mask is generated for a particular index in a batch
-	Except []int
-}
-
-// SetCausal disables causal mask generation for a particular range of indicies in
-// the current batch for subsequent calls to Get. The state resets for the next forward pass.
-func (c *Causal) SetCausal(ctx ml.Context, opts CausalOptions) {
-	if !slices.Equal(c.opts.Except, opts.Except) {
-		c.opts = opts
-		if ctx != nil {
-			var err error
-			c.curMask, err = c.buildMask(ctx)
-			if err != nil {
-				// This error should never occur because we have previously built a mask with the same shape
-				panic(fmt.Errorf("SetCausal: %w", err))
-			}
-		}
-	}
-}
-
-func (c *Causal) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
-	key := c.keys[c.curLayer]
-	value := c.values[c.curLayer]
-
-	kHeadDim := key.Dim(0)
-	numKVHeads := key.Dim(1)
-	rowSize := key.Stride(2)
-	cachedSize := c.curMask.Dim(0)
-
-	key = key.View(ctx, rowSize*c.curCellRange.min,
-		kHeadDim, key.Stride(1),
-		numKVHeads, key.Stride(2),
-		cachedSize,
-	)
-
-	if c.config.PermutedV {
-		vHeadDim := value.Dim(1)
-		elemSize := value.Stride(0)
-
-		value = value.View(ctx, elemSize*c.curCellRange.min,
-			cachedSize, value.Stride(1),
-			vHeadDim, value.Stride(2),
-			numKVHeads,
-		)
-	} else {
-		vHeadDim := value.Dim(0)
-		rowSize := value.Stride(2)
-
-		value = value.View(ctx, rowSize*c.curCellRange.min,
-			vHeadDim, value.Stride(1),
-			numKVHeads, value.Stride(2),
-			cachedSize,
-		)
-	}
-
-	return key, value, c.curMask
-}
-
-func (c *Causal) Put(ctx ml.Context, key, value ml.Tensor) {
-	kHeadDim := key.Dim(0)
-	vHeadDim := value.Dim(0)
-	numKVHeads := key.Dim(1)
-	batchSize := key.Dim(2)
-
-	if c.curBatchSize != batchSize {
-		panic(fmt.Errorf("inconsistent batch sizes (layer: %v, batch size: %v layer batch size: %v)", c.curLayer, c.curBatchSize, batchSize))
-	}
-
-	if _, ok := c.ctxs[c.curLayer]; !ok {
-		c.ctxs[c.curLayer] = c.backend.NewContextSize(2).Layer(c.curLayer)
-	}
-
-	if _, ok := c.keys[c.curLayer]; !ok {
-		c.keys[c.curLayer] = c.ctxs[c.curLayer].Zeros(c.DType, kHeadDim, numKVHeads, int(c.Capacity))
-	}
-
-	if _, ok := c.values[c.curLayer]; !ok {
-		if c.config.PermutedV {
-			c.values[c.curLayer] = c.ctxs[c.curLayer].Zeros(c.DType, int(c.Capacity), vHeadDim, numKVHeads)
-		} else {
-			c.values[c.curLayer] = c.ctxs[c.curLayer].Zeros(c.DType, vHeadDim, numKVHeads, int(c.Capacity))
-		}
-	}
-
-	rowSize := c.keys[c.curLayer].Stride(2)
-	ctx.Forward(key.Copy(ctx, c.keys[c.curLayer].View(ctx, rowSize*c.curLoc, kHeadDim*numKVHeads*batchSize)))
-
-	if c.config.PermutedV {
-		elemSize := c.values[c.curLayer].Stride(0)
-
-		value = value.Permute(ctx, 1, 2, 0, 3)
-		ctx.Forward(value.Copy(ctx, c.values[c.curLayer].View(ctx, elemSize*c.curLoc, batchSize, int(c.Capacity)*elemSize, vHeadDim*numKVHeads)))
-	} else {
-		rowSize := c.values[c.curLayer].Stride(2)
-
-		ctx.Forward(value.Copy(ctx, c.values[c.curLayer].View(ctx, rowSize*c.curLoc, vHeadDim*numKVHeads*batchSize)))
-	}
-}
-
-func (c *Causal) CopyPrefix(srcSeq, dstSeq int, len int32) {
-	seqRange := newRange()
-
-	for i := range c.cells {
-		// Remove the contents of dstSeq so that we only have the copied prefix, metadata will be reset at the end
-		if slices.Contains(c.cells[i].sequences, dstSeq) {
-			c.cells[i].sequences = slices.DeleteFunc(c.cells[i].sequences, func(s int) bool { return s == dstSeq })
-		}
-
-		if slices.Contains(c.cells[i].sequences, srcSeq) && c.cells[i].pos < len {
-			c.cells[i].sequences = append(c.cells[i].sequences, dstSeq)
-			if i < seqRange.min {
-				seqRange.min = i
-			}
-			if i > seqRange.max {
-				seqRange.max = i
-			}
-		}
-	}
-
-	c.cellRanges[dstSeq] = seqRange
-}
-
-func (c *Causal) shift(seq int, beginIndex, offset int32) error {
-	if c.shiftFn == nil {
-		return ErrNotSupported
-	}
-
-	ctx := c.backend.NewContext()
-	defer ctx.Close()
-
-	seqRange := c.cellRanges[seq]
-	size := seqRange.max - seqRange.min + 1
-
-	offsets := make([]int32, size)
-	for i := range offsets {
-		cell := c.cells[seqRange.min+i]
-
-		if slices.Contains(cell.sequences, seq) && cell.pos >= beginIndex {
-			offsets[i] = offset
-		}
-	}
-
-	kShift, err := ctx.Input().FromIntSlice(offsets, len(offsets))
-	if err != nil {
-		return err
-	}
-
-	for i, key := range c.keys {
-		if key == nil {
-			continue
-		}
-
-		kHeadDim := key.Dim(0)
-		numKVHeads := key.Dim(1)
-		rowSize := key.Stride(2)
-
-		key = key.View(ctx, rowSize*seqRange.min,
-			kHeadDim, key.Stride(1),
-			numKVHeads, key.Stride(2),
-			size,
-		)
-
-		roped, err := c.shiftFn(ctx, i, key, kShift)
-		if err != nil {
-			return err
-		}
-
-		ctx.Forward(roped.Copy(ctx, key))
-	}
-
-	ctx.Compute()
-
-	return nil
-}
-
-func (c *Causal) Remove(seq int, beginIndex, endIndex int32) error {
-	var offset int32
-	if endIndex != math.MaxInt32 {
-		offset = beginIndex - endIndex
-	}
-
-	seqRange := newRange()
-
-	for i := range c.cells {
-		if slices.Contains(c.cells[i].sequences, seq) {
-			if c.cells[i].pos >= beginIndex && c.cells[i].pos < endIndex {
-				c.cells[i].sequences = slices.DeleteFunc(c.cells[i].sequences, func(s int) bool { return s == seq })
-			} else {
-				if c.cells[i].pos >= endIndex {
-					if slices.ContainsFunc(c.cells[i].sequences, func(s int) bool { return s != seq }) {
-						// TODO(jessegross): Need to be careful about data shared between sequences
-						return errors.New("shifting on cells shared by multiple sequences not yet implemented")
-					}
-
-					c.cells[i].pos += offset
-				}
-				if i < seqRange.min {
-					seqRange.min = i
-				}
-				if i > seqRange.max {
-					seqRange.max = i
-				}
-			}
-		}
-	}
-
-	if seqRange == newRange() {
-		delete(c.cellRanges, seq)
-		return nil
-	}
-
-	c.cellRanges[seq] = seqRange
-
-	if endIndex != math.MaxInt32 {
-		err := c.shift(seq, endIndex+offset, offset)
-		if err != nil {
-			return err
-		}
-	}
-
-	return nil
-}

kvcache/causal_test.go

@@ -1,533 +0,0 @@
-package kvcache
-
-import (
-	"math"
-	"slices"
-	"testing"
-
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/model/input"
-)
-
-type testCase struct {
-	name          string
-	in            []float32
-	inShape       []int
-	seqs          []int
-	pos           []int32
-	expected      []float32
-	expectedShape []int
-	expectedMask  []float32
-}
-
-func TestStore(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{111, 211, 121, 221, 131, 231, 112, 212, 122, 222, 132, 232, 113, 213, 123, 223, 133, 233, 114, 214, 124, 224, 134, 234},
-			inShape:       []int{2, 3, 4},
-			seqs:          []int{0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3},
-			expected:      []float32{111, 211, 121, 221, 131, 231, 112, 212, 122, 222, 132, 232, 113, 213, 123, 223, 133, 233, 114, 214, 124, 224, 134, 234},
-			expectedShape: []int{2, 3, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0},
-		},
-		{
-			name:          "SecondBatch",
-			in:            []float32{115, 215, 125, 225, 135, 235},
-			inShape:       []int{2, 3, 1},
-			seqs:          []int{0},
-			pos:           []int32{4},
-			expected:      []float32{111, 211, 121, 221, 131, 231, 112, 212, 122, 222, 132, 232, 113, 213, 123, 223, 133, 233, 114, 214, 124, 224, 134, 234, 115, 215, 125, 225, 135, 235},
-			expectedShape: []int{2, 3, 5},
-			expectedMask:  []float32{0, 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestSWA(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewSWACache(1, nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF32, 16)
-
-	tests := []testCase{
-		{
-			name:          "SlidingWindow",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestSequences(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(nil)
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 1, 1},
-			pos:           []int32{0, 1, 0, 1},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-		{
-			name:          "SecondBatch",
-			in:            []float32{5, 6},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{0, 1},
-			pos:           []int32{2, 2},
-			expected:      []float32{1, 2, 3, 4, 5, 6},
-			expectedShape: []int{1, 1, 6},
-			expectedMask:  []float32{0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestRemove(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-		return key.Add(ctx, shift), nil
-	})
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 1, 1},
-			pos:           []int32{0, 1, 0, 1},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	err := cache.Remove(0, 1, math.MaxInt32)
-	if err != nil {
-		panic(err)
-	}
-
-	tests = []testCase{
-		{
-			name:          "RemoveEnd",
-			in:            []float32{5, 6},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{0, 1},
-			pos:           []int32{1, 2},
-			expected:      []float32{1, 2, 3, 4, 5, 6},
-			expectedShape: []int{1, 1, 6},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	err = cache.Remove(0, 0, 1)
-	if err != nil {
-		panic(err)
-	}
-
-	tests = []testCase{
-		{
-			name:          "RemoveMiddle",
-			in:            []float32{7, 8},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{0, 0},
-			pos:           []int32{1, 2},
-			expected:      []float32{7, 8, 3, 4, 4},
-			expectedShape: []int{1, 1, 5},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestDefrag(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-		return key.Add(ctx, shift), nil
-	})
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
-			inShape:       []int{1, 1, 16},
-			seqs:          []int{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
-			expected:      []float32{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
-			expectedShape: []int{1, 1, 16},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	err := cache.Remove(0, 2, 4)
-	if err != nil {
-		panic(err)
-	}
-
-	err = cache.Remove(0, 13, math.MaxInt32)
-	if err != nil {
-		panic(err)
-	}
-
-	tests = []testCase{
-		{
-			name:          "Defrag",
-			in:            []float32{17, 18, 19},
-			inShape:       []int{1, 1, 3},
-			seqs:          []int{0, 0, 0},
-			pos:           []int32{16, 17, 18},
-			expected:      []float32{1, 2, 12, 13, 3, 4, 5, 6, 7, 8, 9, 10, 11, 17, 18, 19},
-			expectedShape: []int{1, 1, 16},
-			expectedMask:  []float32{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func TestCopy(t *testing.T) {
-	backend := &testBackend{}
-	cache := NewCausalCache(func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) { return key, nil })
-	defer cache.Close()
-
-	cache.Init(backend, ml.DTypeF16, 16)
-
-	tests := []testCase{
-		{
-			name:          "FirstBatch",
-			in:            []float32{1, 2, 3, 4},
-			inShape:       []int{1, 1, 4},
-			seqs:          []int{0, 0, 0, 0},
-			pos:           []int32{0, 1, 2, 3},
-			expected:      []float32{1, 2, 3, 4},
-			expectedShape: []int{1, 1, 4},
-			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), 0, 0, 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-
-	cache.CopyPrefix(0, 1, 2)
-
-	tests = []testCase{
-		{
-			name:          "Copy",
-			in:            []float32{5, 6},
-			inShape:       []int{1, 1, 2},
-			seqs:          []int{1, 1},
-			pos:           []int32{3, 4},
-			expected:      []float32{1, 2, 3, 4, 5, 6},
-			expectedShape: []int{1, 1, 6},
-			expectedMask:  []float32{0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
-		},
-	}
-
-	testCache(t, backend, cache, tests)
-}
-
-func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase) {
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			context := backend.NewContext()
-			defer context.Close()
-
-			err := cache.StartForward(context, input.Options{Positions: test.pos, Sequences: test.seqs})
-			if err != nil {
-				panic(err)
-			}
-
-			cache.SetLayer(0)
-			tensor, _ := context.FromFloatSlice(test.in, test.inShape...)
-			cache.Put(context, tensor, tensor)
-
-			out, _, mask := cache.Get(context)
-
-			context.Forward(out, mask).Compute(out, mask)
-
-			if !slices.Equal(out.Floats(), test.expected) || !slices.Equal(out.Shape(), test.expectedShape) || !slices.Equal(mask.Floats(), test.expectedMask) {
-				t.Errorf("TestCache: have %v (shape %v); want %v (shape %v); mask: have %v (shape %v) want %v", out.Floats(), out.Shape(), test.expected, test.expectedShape, mask.Floats(), mask.Shape(), test.expectedMask)
-			}
-		})
-	}
-}
-
-type testBackend struct{}
-
-func (b *testBackend) Config() ml.Config {
-	panic("not implemented")
-}
-
-func (b *testBackend) Get(name string) ml.Tensor {
-	panic("not implemented")
-}
-
-func (b *testBackend) NewContext() ml.Context {
-	return &testContext{}
-}
-
-func (b *testBackend) NewContextSize(int) ml.Context {
-	return &testContext{}
-}
-
-func (b *testBackend) SystemInfo() string {
-	return "not implemented"
-}
-
-type testContext struct{}
-
-func (c *testContext) Empty(dtype ml.DType, shape ...int) ml.Tensor {
-	total := 0
-
-	if len(shape) > 0 {
-		total = 1
-		for _, s := range shape {
-			total *= s
-		}
-	}
-
-	return &testTensor{dtype: dtype, elementSize: 4, data: make([]float32, total), shape: shape}
-}
-
-func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
-	return c.Empty(dtype, shape...)
-}
-
-func (c *testContext) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
-	t := c.Empty(ml.DTypeF32, shape...).(*testTensor)
-
-	copy(t.data, s)
-
-	return t, nil
-}
-
-func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
-	f := make([]float32, len(s))
-	for i := range f {
-		f[i] = float32(s[i])
-	}
-
-	out, _ := c.FromFloatSlice(f, shape...)
-	out.(*testTensor).dtype = ml.DTypeI32
-
-	return out, nil
-}
-
-func (c *testContext) Input() ml.Context    { return c }
-func (c *testContext) Output() ml.Context   { return c }
-func (c *testContext) Layer(int) ml.Context { return c }
-
-func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }
-
-func (c *testContext) Compute(...ml.Tensor) {}
-
-func (c *testContext) MaxGraphNodes() int {
-	return 10
-}
-
-func (c *testContext) Close() {}
-
-type testTensor struct {
-	dtype       ml.DType
-	elementSize int
-	data        []float32
-	shape       []int
-}
-
-func (t *testTensor) Dim(n int) int {
-	return t.shape[n]
-}
-
-func (t *testTensor) Stride(n int) int {
-	stride := t.elementSize
-	for i := range n {
-		stride *= t.shape[i]
-	}
-
-	return stride
-}
-
-func (t *testTensor) Shape() []int {
-	return t.shape
-}
-
-func (t *testTensor) DType() ml.DType {
-	return t.dtype
-}
-
-func (t *testTensor) Bytes() []byte {
-	panic("not implemented")
-}
-
-func (t *testTensor) Floats() []float32 {
-	out := make([]float32, len(t.data))
-	copy(out, t.data)
-	return out
-}
-
-func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	out := ctx.Empty(t.DType(), t.Shape()...).(*testTensor)
-
-	for i := range out.data {
-		out.data[i] = t.data[i] + t2.(*testTensor).data[i]
-	}
-
-	return out
-}
-
-func (t *testTensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Softmax(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) LayerNorm(ctx ml.Context, weight, bias ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RMSNorm(ctx ml.Context, weight ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Scale(ctx ml.Context, s float64) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) AvgPool1D(ctx ml.Context, k, s, p int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim, ropeType uint32, base, scale float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Tanh(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) GELU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) SILU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
-	offset /= t.elementSize
-
-	var s []int
-
-	switch len(shape) {
-	case 1:
-		s = []int{shape[0]}
-	case 5:
-		s = []int{shape[0], shape[2], shape[4]}
-	default:
-		panic("unsupported number of dimensions")
-	}
-
-	context := &testContext{}
-
-	view := context.Empty(t.dtype, s...).(*testTensor)
-	view.data = t.data[offset : offset+len(view.data)]
-
-	return view
-}
-
-func (t *testTensor) Permute(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Contiguous(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Set(ctx ml.Context, t2 ml.Tensor, offset int, strides ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Stack(ctx ml.Context, dim int, s ...ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Concat(ctx ml.Context, t2 ml.Tensor, dim int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Rows(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Copy(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	copy(t2.(*testTensor).data, t.data)
-	return nil
-}

kvcache/encoder.go

@@ -1,139 +0,0 @@
-package kvcache
-
-import (
-	"fmt"
-
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/model/input"
-)
-
-// Encoder cache stores K and V tensors that are position independent
-//
-// The tensors can be of any shape and will be returned as they were stored
-// The mask is currently always nil
-//
-// Not currently safe for multiple sequences
-type EncoderCache struct {
-	// config controls mostly backend-specific optimizations
-	config *ml.CacheConfig
-
-	// ** current forward pass **
-
-	// the active layer for Get and Put
-	curLayer int
-
-	// if something is stored during this pass, this
-	// will be the position (but there is no guarantee
-	// anything will be stored)
-	curPos int32
-
-	// ** cache metadata **
-
-	// was something stored in the cache?
-	encoderCached bool
-
-	// position of the cached data
-	encoderPos int32
-
-	// ** cache data storage **
-	backend      ml.Backend
-	ctxs         map[int]ml.Context
-	keys, values map[int]ml.Tensor
-}
-
-func NewEncoderCache() *EncoderCache {
-	return &EncoderCache{
-		ctxs:   make(map[int]ml.Context),
-		keys:   make(map[int]ml.Tensor),
-		values: make(map[int]ml.Tensor),
-	}
-}
-
-func (c *EncoderCache) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
-	if c.config == nil {
-		var config ml.CacheConfig
-		if cc, ok := backend.(ml.BackendCacheConfig); ok {
-			config = cc.CacheConfig()
-		}
-		c.config = &config
-	}
-
-	if c.config.CachePadding != 0 && c.config.CachePadding != 1 {
-		panic(fmt.Errorf("encoder cache is unable to enforce requested CachePadding (%v)", c.config.CachePadding))
-	}
-
-	c.backend = backend
-}
-
-func (c *EncoderCache) SetConfig(config ml.CacheConfig) {
-	if c.config != nil {
-		panic("config cannot be changed after being previously set, either by the model or backend")
-	}
-
-	c.config = &config
-}
-
-func (c *EncoderCache) Close() {
-	for _, ctx := range c.ctxs {
-		ctx.Close()
-	}
-}
-
-func (c *EncoderCache) StartForward(ctx ml.Context, opts input.Options) error {
-	// We work with the most recent image
-	if len(opts.Multimodal) > 0 {
-		c.curPos = opts.Positions[opts.Multimodal[len(opts.Multimodal)-1].Index]
-	}
-
-	return nil
-}
-
-func (c *EncoderCache) SetLayer(layer int) {
-	c.curLayer = layer
-}
-
-func (c *EncoderCache) EncoderCached() bool {
-	return c.encoderCached
-}
-
-func (c *EncoderCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
-	return c.keys[c.curLayer], c.values[c.curLayer], nil
-}
-
-func (c *EncoderCache) Put(ctx ml.Context, key, value ml.Tensor) {
-	c.encoderPos = c.curPos
-	c.encoderCached = true
-
-	if c.config.PermutedV {
-		value = value.Permute(ctx, 1, 2, 0, 3)
-	}
-
-	if _, ok := c.ctxs[c.curLayer]; !ok {
-		c.ctxs[c.curLayer] = c.backend.NewContextSize(2).Layer(c.curLayer)
-	}
-
-	if _, ok := c.keys[c.curLayer]; !ok {
-		c.keys[c.curLayer] = c.ctxs[c.curLayer].Empty(key.DType(), key.Shape()...)
-	}
-
-	if _, ok := c.values[c.curLayer]; !ok {
-		c.values[c.curLayer] = c.ctxs[c.curLayer].Empty(value.DType(), value.Shape()...)
-	}
-
-	ctx.Forward(
-		key.Copy(ctx, c.keys[c.curLayer]),
-		value.Copy(ctx, c.values[c.curLayer]),
-	)
-}
-
-func (c *EncoderCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
-	panic("encoder cache does not support multiple sequences")
-}
-
-func (c *EncoderCache) Remove(seq int, beginIndex, endIndex int32) error {
-	if c.encoderPos >= beginIndex && c.encoderPos < endIndex {
-		c.encoderCached = false
-	}
-
-	return nil
-}

kvcache/wrapper.go

@@ -1,100 +0,0 @@
-package kvcache
-
-import (
-	"math"
-
-	"github.com/ollama/ollama/ml"
-	"github.com/ollama/ollama/model/input"
-)
-
-// Wrapper cache is a container for multiple types of caches,
-// such as for the encoding and decoding portions of a model.
-type WrapperCache struct {
-	// caches we are wrapping
-	caches []Cache
-
-	// cache to be used for this layer
-	curType int
-}
-
-func NewWrapperCache(caches ...Cache) *WrapperCache {
-	return &WrapperCache{
-		caches: caches,
-	}
-}
-
-func (c *WrapperCache) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
-	for _, cache := range c.caches {
-		cache.Init(backend, dtype, capacity)
-	}
-}
-
-func (c *WrapperCache) SetConfig(config ml.CacheConfig) {
-	for _, cache := range c.caches {
-		cache.SetConfig(config)
-	}
-}
-
-func (c *WrapperCache) Close() {
-	for _, cache := range c.caches {
-		cache.Close()
-	}
-}
-
-func (c *WrapperCache) StartForward(ctx ml.Context, opts input.Options) error {
-	for i, cache := range c.caches {
-		err := cache.StartForward(ctx, opts)
-		if err != nil {
-			// unwind on error - Remove with endIndex set to math.MaxInt32 does not fail
-			for j := i - 1; j >= 0; j-- {
-				for k := range opts.Positions {
-					_ = c.caches[j].Remove(opts.Sequences[k], opts.Positions[k], math.MaxInt32)
-				}
-			}
-			return err
-		}
-	}
-
-	c.curType = 0
-	return nil
-}
-
-func (c *WrapperCache) SetLayer(layer int) {
-	for _, cache := range c.caches {
-		cache.SetLayer(layer)
-	}
-}
-
-func (c *WrapperCache) SetLayerType(layerType int) {
-	c.curType = layerType
-}
-
-func (c *WrapperCache) UnderlyingCache() Cache {
-	return c.caches[c.curType]
-}
-
-func (c *WrapperCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
-	return c.caches[c.curType].Get(ctx)
-}
-
-func (c *WrapperCache) Put(ctx ml.Context, key, value ml.Tensor) {
-	c.caches[c.curType].Put(ctx, key, value)
-}
-
-func (c *WrapperCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
-	for _, cache := range c.caches {
-		cache.CopyPrefix(srcSeq, dstSeq, len)
-	}
-}
-
-func (c *WrapperCache) Remove(seq int, beginIndex, endIndex int32) error {
-	// If the one of these fails, the caller is supposed to retry with endIndex set to math.MaxInt32, which should not fail
-	for _, cache := range c.caches {
-		err := cache.Remove(seq, beginIndex, endIndex)
-		if err != nil {
-			return err
-		}
-	}
-
-	return nil
-}

llama/README.md

@@ -1,52 +1,157 @@
 # `llama`
 
-This package provides Go bindings to [llama.cpp](https://github.com/ggerganov/llama.cpp).
+This package integrates the [llama.cpp](https://github.com/ggerganov/llama.cpp) library as a Go package and makes it easy to build it with tags for different CPU and GPU processors.
+
+Supported:
+
+- [x] CPU
+- [x] avx, avx2
+- [x] macOS Metal
+- [x] Windows CUDA
+- [x] Windows ROCm
+- [x] Linux CUDA
+- [x] Linux ROCm
+- [x] Llava
+
+Extra build steps are required for CUDA and ROCm on Windows since `nvcc` and `hipcc` both require using msvc as the host compiler. For these shared libraries are created:
+
+- `ggml_cuda.dll` on Windows or `ggml_cuda.so` on Linux
+- `ggml_hipblas.dll` on Windows or `ggml_hipblas.so` on Linux
+
+> Note: it's important that memory is allocated and freed by the same compiler (e.g. entirely by code compiled with msvc or mingw). Issues from this should be rare, but there are some places where pointers are returned by the CUDA or HIP runtimes and freed elsewhere, causing a a crash. In a future change the same runtime should be used in both cases to avoid crashes.
+
+## Building
+
+```
+go build .
+```
+
+### AVX
+
+```shell
+go build -tags avx .
+```
+
+### AVX2
+
+```shell
+# go doesn't recognize `-mfma` as a valid compiler flag
+# see https://github.com/golang/go/issues/17895
+go env -w "CGO_CFLAGS_ALLOW=-mfma|-mf16c"
+go env -w "CGO_CXXFLAGS_ALLOW=-mfma|-mf16c"
+go build -tags=avx,avx2 .
+```
+
+## Linux
+
+### CUDA
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive):
+
+```shell
+make ggml_cuda.so
+go build -tags avx,cuda .
+```
+
+### ROCm
+
+Install [ROCm](https://rocm.docs.amd.com/en/latest/).
+
+```shell
+make ggml_hipblas.so
+go build -tags avx,rocm .
+```
+
+## Windows
+
+Download [w64devkit](https://github.com/skeeto/w64devkit/releases/latest) for a simple MinGW development environment.
+
+### CUDA
+
+Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive) then build the cuda code:
+
+```shell
+make ggml_cuda.dll
+go build -tags avx,cuda .
+```
+
+### ROCm
+
+Install [ROCm](https://rocm.docs.amd.com/en/latest/).
+
+```shell
+make ggml_hipblas.dll
+go build -tags avx,rocm .
+```
+
+## Building runners
+
+```shell
+# build all runners for this platform
+make -j
+```
 
 ## Vendoring
 
-Ollama vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/llama.cpp/tree/master/ggml/src). While we generally strive to contribute changes back upstream to avoid drift, we carry a small set of patches which are applied to the tracking commit.
+Ollama currently vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/ggml) through a vendoring model. While we generally strive to contribute changes back upstream to avoid drift, we cary a small set of patches which are applied to the tracking commit. A set of make targets are available to aid developers in updating to a newer tracking commit, or to work on changes.
 
 If you update the vendoring code, start by running the following command to establish the tracking llama.cpp repo in the `./vendor/` directory.
 
-```shell
-make -f Makefile.sync apply-patches
+```
+make apply-patches
 ```
 
 ### Updating Base Commit
 
 **Pin to new base commit**
 
-To change the base commit, update `FETCH_HEAD` in Makefile.sync.
+To update to a newer base commit, select the upstream git tag or commit and update `llama/vendoring`
+
+#### Applying patches
 
 When updating to a newer base commit, the existing patches may not apply cleanly and require manual merge resolution.
 
 Start by applying the patches. If any of the patches have conflicts, the `git am` will stop at the first failure.
 
-```shell
-make -f Makefile.sync apply-patches
+```
+make apply-patches
 ```
 
-If there are conflicts, you will see an error message. Resolve the conflicts in `./vendor/`, and continue the patch series with `git am --continue` and rerun `make -f Makefile.sync apply-patches`. Repeat until all patches are successfully applied.
+If you see an error message about a conflict, go into the `./vendor/` directory, and perform merge resolution using your preferred tool to the patch commit which failed. Save the file(s) and continue the patch series with `git am --continue` . If any additional patches fail, follow the same pattern until the full patch series is applied. Once finished, run a final `create-patches` and `sync` target to ensure everything is updated.
 
-Once all patches are applied, commit the changes to the tracking repository.
-
-```shell
-make -f Makefile.sync format-patches sync
 ```
+make create-patches sync
+```
+
+Build and test Ollama, and make any necessary changes to the Go code based on the new base commit. Submit your PR to the Ollama repo.
 
 ### Generating Patches
 
 When working on new fixes or features that impact vendored code, use the following model. First get a clean tracking repo with all current patches applied:
 
-```shell
-make -f Makefile.sync clean apply-patches
 ```
+make apply-patches
+```
+
+Now edit the upstream native code in the `./vendor/` directory. You do not need to commit every change in order to build, a dirty working tree in the tracking repo is OK while developing. Simply save in your editor, and run the following to refresh the vendored code with your changes, build the backend(s) and build ollama:
+
+```
+make sync
+make -j 8
+go build .
+```
+
+> [!IMPORTANT]
+> Do **NOT** run `apply-patches` while you're iterating as that will reset the tracking repo. It will detect a dirty tree and abort, but if your tree is clean and you accidentally ran this target, use `git reflog` to recover your commit(s).
 
 Iterate until you're ready to submit PRs. Once your code is ready, commit a change in the `./vendor/` directory, then generate the patches for ollama with
 
-```shell
-make -f Makefile.sync format-patches
 ```
+make create-patches
+```
+
+> [!IMPORTANT]
+> Once you have completed this step, it is safe to run `apply-patches` since your change is preserved in the patches.
 
 In your `./vendor/` directory, create a branch, and cherry-pick the new commit to that branch, then submit a PR upstream to llama.cpp.
 

llama/amx.cpp

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #include "amx.h"
 #include "common.h"
 #include "mmq.h"

llama/amx.h

@@ -0,0 +1,34 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ggml-backend.h"
+#include "ggml-cpu-impl.h"
+
+// GGML internal header
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+#endif

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "d7cfe1ffe0f435d0048a6058d529daf76e072d9c";
+char const *LLAMA_COMMIT = "ba1cb19cdd0d92e012e0f6e009e0620f854b6afd";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/build-info.cpp.in

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

llama/clip.cpp

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 // NOTE: This is modified from clip.cpp only for LLaVA,
 // so there might be still unnecessary artifacts hanging around
 // I'll gradually clean and extend it
@@ -7,7 +33,6 @@
 #include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
-#include "gguf.h"
 
 #ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
@@ -40,7 +65,6 @@
 #include <map>
 #include <regex>
 #include <stdexcept>
-#include <unordered_set>
 #include <vector>
 #include <sstream>
 #include <cinttypes>
@@ -116,7 +140,6 @@ static std::string format(const char * fmt, ...) {
 #define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
 #define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
 #define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
-#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
 #define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
 #define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
 #define KEY_USE_GELU            "clip.use_gelu"
@@ -134,7 +157,6 @@ static std::string format(const char * fmt, ...) {
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"
 #define KEY_IMAGE_STD           "clip.vision.image_std"
 #define KEY_PROJ_TYPE           "clip.projector_type"
-#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -176,15 +198,6 @@ static std::string format(const char * fmt, ...) {
 #define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
 #define TN_MINICPMV_LN "resampler.ln_%s.%s"
 
-#define TN_GLM_ADAPER_CONV "adapter.conv.%s"
-#define TN_GLM_ADAPTER_LINEAR "adapter.linear.linear.%s"
-#define TN_GLM_ADAPTER_NORM_1 "adapter.linear.norm1.%s"
-#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
-#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s"
-#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
-#define TN_GLM_BOI_W "adapter.boi"
-#define TN_GLM_EOI_W "adapter.eoi"
-
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
@@ -192,7 +205,6 @@ enum projector_type {
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
     PROJECTOR_TYPE_RESAMPLER,
-    PROJECTOR_TYPE_GLM_EDGE,
     PROJECTOR_TYPE_MERGER,
     PROJECTOR_TYPE_UNKNOWN,
 };
@@ -202,7 +214,6 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_LDP, "ldp" },
     { PROJECTOR_TYPE_LDPV2, "ldpv2"},
     { PROJECTOR_TYPE_RESAMPLER, "resampler"},
-    { PROJECTOR_TYPE_GLM_EDGE, "adapter"},
     { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
 };
 
@@ -290,7 +301,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
             {
                 const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
                 int arr_n = gguf_get_arr_n(ctx_gguf, i);
-                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
+                const void * data = gguf_get_arr_data(ctx_gguf, i);
                 std::stringstream ss;
                 ss << "[";
                 for (int j = 0; j < arr_n; j++) {
@@ -459,9 +470,8 @@ struct clip_hparams {
 
     char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default)
 
-    std::vector<int32_t> image_grid_pinpoints;
+    int32_t image_grid_pinpoints[32];
     int32_t image_crop_resolution;
-    std::unordered_set<int32_t> vision_feature_layer;
 };
 
 struct clip_layer {
@@ -528,12 +538,6 @@ struct clip_vision_model {
     struct ggml_tensor * mm_4_w = NULL;
     struct ggml_tensor * mm_4_b = NULL;
 
-    //GLMV-Edge projection
-    struct ggml_tensor * mm_model_adapter_conv_w;
-    struct ggml_tensor * mm_model_adapter_conv_b;
-    struct ggml_tensor * boi_w;
-    struct ggml_tensor * eoi_w;
-
     // MobileVLM projection
     struct ggml_tensor * mm_model_mlp_1_w;
     struct ggml_tensor * mm_model_mlp_1_b;
@@ -594,14 +598,12 @@ struct clip_ctx {
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
     bool has_minicpmv_projector = false;
-    bool has_glm_projector = false;
     bool has_qwen2vl_merger = false;
     int minicpmv_version = 2;
 
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
 
-    int32_t max_feature_layer;
     float image_mean[3];
     float image_std[3];
     bool use_gelu = false;
@@ -668,12 +670,13 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
+    int n_layer                    = hparams.n_layer;
     const float eps                = hparams.eps;
     int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
 
     const int batch_size = imgs->size;
 
-    if (ctx->has_llava_projector || ctx->has_minicpmv_projector || ctx->has_glm_projector) {
+    if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
         GGML_ASSERT(batch_size == 1);
     }
 
@@ -753,9 +756,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         else if (ctx->minicpmv_version == 3) {
             pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
         }
-        else if (ctx->minicpmv_version == 4) {
-            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
-        }
         ggml_set_name(pos_embed, "pos_embed");
         ggml_set_input(pos_embed);
     }
@@ -768,19 +768,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
     }
 
-    std::vector<struct ggml_tensor *> embedding_stack;
-    const auto & vision_feature_layer = hparams.vision_feature_layer;
-
     // loop over layers
-    for (int il = 0; il < ctx->max_feature_layer; il++) {
+    if (ctx->has_minicpmv_projector || ctx->has_qwen2vl_merger) {
+        // TODO: figure out why we doing thing in this way ???
+        n_layer += 1;
+    }
+    for (int il = 0; il < n_layer - 1; il++) {
         struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
 
-        // If this is an embedding feature layer, save the output.
-        // NOTE: 0 index here refers to the input to the encoder.
-        if (vision_feature_layer.find(il) != vision_feature_layer.end()) {
-            embedding_stack.push_back(embeddings);
-        }
-
         //const size_t nb_q_w = model.layers[il].q_w->nb[0];
 
         // layernorm1
@@ -868,6 +863,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         cur = ggml_add(ctx0, embeddings, cur);
 
         embeddings = cur;
+
     }
 
     // post-layernorm
@@ -878,19 +874,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
     }
 
-    // final layer is a vision feature layer
-    if (vision_feature_layer.find(ctx->max_feature_layer) != vision_feature_layer.end()) {
-        embedding_stack.push_back(embeddings);
-    }
-
-    // If feature layers are explicitly set, stack them (if we have multiple)
-    if (!embedding_stack.empty()) {
-        embeddings = embedding_stack[0];
-        for (size_t i = 1; i < embedding_stack.size(); i++) {
-            embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0);
-        }
-    }
-
     // llava projector
     if (ctx->has_llava_projector) {
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
@@ -1108,11 +1091,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
                     n_head = hidden_size/d_head;
                     num_query = 64;
                 }
-                else if (ctx->minicpmv_version == 4) {
-                    hidden_size = 3584;
-                    n_head = hidden_size/d_head;
-                    num_query = 64;
-                }
 
                 struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
                 Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
@@ -1147,33 +1125,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             GGML_ASSERT(false);
         }
     }
-    // glm projector
-    else if (ctx->has_glm_projector) {
-        if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) {
-            size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
-            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
-            embeddings = ggml_reshape_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
-            embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
-            embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
-            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3));
-            embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b);
-            //GLU
-            {
-                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
-                embeddings = ggml_norm(ctx0, embeddings, eps);
-                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
-                embeddings = ggml_gelu_inplace(ctx0, embeddings);
-                struct ggml_tensor * x = embeddings;
-                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
-                x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x);
-                embeddings = ggml_silu_inplace(ctx0, embeddings);
-                embeddings = ggml_mul(ctx0, embeddings,x);
-                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
-            }
-        } else {
-            GGML_ABORT("fatel error");
-        }
-    } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+    else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size);
 
         embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
@@ -1309,15 +1261,35 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-    ggml_backend_t backend = ggml_backend_init_best();
-    if (backend == nullptr) {
-        LOG_ERR("%s: failed to initialize backend\n", __func__);
-        clip_free(new_clip);
-        gguf_free(ctx);
-        return nullptr;
+#ifdef GGML_USE_CUDA
+   new_clip->backend = ggml_backend_cuda_init(0);
+   LOG_INF("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+   new_clip->backend = ggml_backend_metal_init();
+   LOG_INF("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_CANN
+   new_clip->backend = ggml_backend_cann_init(0);
+   LOG_INF("%s: CLIP using CANN backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_VULKAN
+   new_clip->backend = ggml_backend_vk_init(0);
+   LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_SYCL
+   new_clip->backend = ggml_backend_sycl_init(0);
+   LOG_INF("%s: CLIP using SYCL backend\n", __func__);
+#endif
+
+    if (!new_clip->backend) {
+        new_clip->backend = ggml_backend_cpu_init();
+        LOG_INF("%s: CLIP using CPU backend\n", __func__);
     }
-    LOG_INF("%s: using %s backend\n", __func__, ggml_backend_name(backend));
-    new_clip->backend = backend;
 
     // model size and capabilities
     {
@@ -1342,11 +1314,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
         }
 
-        idx = gguf_find_key(ctx, KEY_HAS_GLM_PROJ);
-        if (idx != -1) {
-            new_clip->has_glm_projector = gguf_get_val_bool(ctx, idx);
-        }
-
         idx = gguf_find_key(ctx, KEY_HAS_QWEN2VL_MERGER);
         if (idx != -1) {
             new_clip->has_qwen2vl_merger = gguf_get_val_bool(ctx, idx);
@@ -1371,7 +1338,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             LOG_INF("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
             LOG_INF("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
-            LOG_INF("%s: glm_projector:  %d\n", __func__, new_clip->has_glm_projector);
             LOG_INF("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
             LOG_INF("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
@@ -1482,26 +1448,14 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS);
             int n = gguf_get_arr_n(ctx, idx);
             const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx);
-            for (int i = 0; i < n; ++i) {
-                hparams.image_grid_pinpoints.push_back(pinpoints[i]);
+            for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) {
+                hparams.image_grid_pinpoints[i] = pinpoints[i];
             }
-        } catch (std::runtime_error & /*e*/) { }
-
-        // Load the vision feature layer indices if they are explicitly provided;
-        // if multiple vision feature layers are present, the values will be concatenated
-        // to form the final visual features.
-        // NOTE: gguf conversions should standardize the values of the vision feature layer to
-        // be non-negative, since we use -1 to mark values as unset here.
-        try {
-            int idx = get_key_idx(ctx, KEY_FEATURE_LAYER);
-            int n = gguf_get_arr_n(ctx, idx);
-
-            const int32_t * vision_feature_layer = (const int32_t *)gguf_get_arr_data(ctx, idx);
-
-            for (int i = 0; i < n; ++i) {
-                hparams.vision_feature_layer.insert(vision_feature_layer[i]);
-            }
-        } catch (std::runtime_error & /*e*/) { }
+            if (n < 32)
+                hparams.image_grid_pinpoints[n] = 0;
+        } catch (std::runtime_error & /*e*/) {
+            hparams.image_grid_pinpoints[0]=0;
+        }
 
         try {
             int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE);
@@ -1527,9 +1481,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->image_std[i]  = std_data[i];
         }
 
-        // Calculate the deepest feature layer based on hparams and projector type
-        new_clip->max_feature_layer = get_deepest_feature_layer(new_clip);
-
         if (verbosity >= 2) {
             LOG_INF("\n%s: vision model hparams\n", __func__);
             LOG_INF("image_size         %d\n", hparams.image_size);
@@ -1543,13 +1494,8 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             LOG_INF("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
             LOG_INF("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
             LOG_INF("v_image_grid_pinpoints: ");
-            for (const auto & pp : hparams.image_grid_pinpoints) {
-                LOG_INF("%d ", pp);
-            }
-            LOG_INF("\n");
-            LOG_INF("v_vision_feature_layer: ");
-            for (const auto & feature_layer: hparams.vision_feature_layer) {
-                LOG_INF("%d ", feature_layer);
+            for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
+                LOG_INF("%d ", hparams.image_grid_pinpoints[i]);
             }
             LOG_INF("\n");
             LOG_INF("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
@@ -1684,18 +1630,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
             vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
         }
-        else if (new_clip->proj_type == PROJECTOR_TYPE_GLM_EDGE) {
-            vision_model.mm_model_adapter_conv_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "weight"));
-            vision_model.mm_model_adapter_conv_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "bias"));
-            vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_LINEAR,"weight"));
-            vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"weight"));
-            vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"bias"));
-            vision_model.mm_model_mlp_1_w =  get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_H_2_4H,"weight"));
-            vision_model.mm_model_mlp_2_w =  get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_GATE,"weight"));
-            vision_model.mm_model_mlp_3_w =  get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_4H_2_H,"weight"));
-            vision_model.boi_w = get_tensor(new_clip->ctx_data, TN_GLM_BOI_W);
-            vision_model.eoi_w = get_tensor(new_clip->ctx_data, TN_GLM_EOI_W);
-        }
         else if (new_clip->proj_type == PROJECTOR_TYPE_MERGER) {
             vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
             vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
@@ -1788,11 +1722,11 @@ void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
     }
 }
 
-void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img) {
+static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
     img->nx = nx;
     img->ny = ny;
     img->buf.resize(3 * nx * ny);
-    memcpy(img->buf.data(), rgb_pixels, img->buf.size());
+    memcpy(img->buf.data(), data, img->buf.size());
 }
 
 bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
@@ -1802,7 +1736,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
         LOG_ERR("%s: failed to load image '%s'\n", __func__, fname);
         return false;
     }
-    clip_build_img_from_pixels(data, nx, ny, img);
+    build_clip_img_from_data(data, nx, ny, img);
     stbi_image_free(data);
     return true;
 }
@@ -1814,7 +1748,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
         LOG_ERR("%s: failed to decode image bytes\n", __func__);
         return false;
     }
-    clip_build_img_from_pixels(data, nx, ny, img);
+    build_clip_img_from_data(data, nx, ny, img);
     stbi_image_free(data);
     return true;
 }
@@ -2170,7 +2104,6 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
                 images[images.size()-1].push_back(patch);
             }
         }
-        clip_image_u8_free(refine_image);
     }
     return images;
 }
@@ -2209,13 +2142,6 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
                 clip_image_f32_free(res);
             }
         }
-        for (size_t i = 0; i < imgs.size(); ++i) {
-            for (size_t j = 0; j < imgs[i].size(); ++j) {
-                if (imgs[i][j] != nullptr) {
-                    clip_image_u8_free(imgs[i][j]);
-                }
-            }
-        }
         return true;
     }
     else if (ctx->has_qwen2vl_merger) {
@@ -2236,20 +2162,6 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
         return true;
     }
 
-    if (ctx->has_glm_projector) {
-        res_imgs->size = 1;
-        res_imgs->data = new clip_image_f32[res_imgs->size];
-        clip_image_u8 resized_image;
-        int32_t sz=ctx->vision_model.hparams.image_size;
-        bicubic_resize(*img, resized_image,sz,sz);
-        clip_image_f32 * res = clip_image_f32_init();
-        //clip_image_save_to_bmp(resized_image, "resized.bmp");
-        normalize_image_u8_to_f32(&resized_image, res, ctx->image_mean, ctx->image_std);
-        res_imgs->data[0] = *res;
-        clip_image_f32_free(res);
-        return true;
-    }
-
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
         LOG_ERR("This gguf file seems to have no vision encoder\n");
@@ -2294,10 +2206,10 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
             }
         }
     } else {
-        if (!params.image_grid_pinpoints.empty()) {
+        if (params.image_grid_pinpoints[0] != 0) {
             // "spatial_unpad" with "anyres" processing for llava-1.6
             std::vector<std::pair<int, int>> possible_resolutions;
-            for (size_t i = 0; i < params.image_grid_pinpoints.size(); i+=2) {
+            for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
                 possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
             }
             std::pair<int, int> best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions);
@@ -2435,8 +2347,7 @@ void clip_free(clip_ctx * ctx) {
 }
 
 size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
-    int extra_tokens = ctx->has_glm_projector ? 2 : 0;
-    return (clip_n_patches(ctx) + extra_tokens) * clip_n_mmproj_embd(ctx) * sizeof(float);
+    return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }
 
 size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) {
@@ -2463,14 +2374,7 @@ const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
 }
 
 const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
-    if (ctx->vision_model.hparams.image_grid_pinpoints.size()) {
-        return &ctx->vision_model.hparams.image_grid_pinpoints.front();
-    }
-    return nullptr;
-}
-
-size_t get_clip_image_grid_size(const struct clip_ctx * ctx) {
-    return ctx->vision_model.hparams.image_grid_pinpoints.size();
+    return ctx->vision_model.hparams.image_grid_pinpoints;
 }
 
 int clip_n_patches(const struct clip_ctx * ctx) {
@@ -2485,7 +2389,7 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i
 
     int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
 
-    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2 || ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
     } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
         if (ctx->minicpmv_version == 2) {
@@ -2494,9 +2398,6 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i
         else if (ctx->minicpmv_version == 3) {
             n_patches = 64;
         }
-        else if (ctx->minicpmv_version == 4) {
-            n_patches = 64;
-        }
     } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         int patch_size = params.patch_size * 2;
         int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
@@ -2618,12 +2519,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     if (ctx->has_minicpmv_projector) {
         GGML_ASSERT(batch_size == 1);
     }
-    if (ctx->has_glm_projector) {
-        GGML_ASSERT(batch_size == 1);
-        ggml_tensor * boi = ctx->vision_model.boi_w;
-        ggml_backend_tensor_get(boi,vec,0,ggml_nbytes(boi));
-        vec = (float*)(vec+ggml_nelements(boi)); //offset for boi
-    }
 
     // build the inference graph
     ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
@@ -2682,8 +2577,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
             int* positions_data = (int*)malloc(ggml_nbytes(positions));
-            int bucket_coords_h[1024];
-            int bucket_coords_w[1024];
+            int bucket_coords_h[70];
+            int bucket_coords_w[70];
             for (int i = 0; i < pos_h; i++){
                 bucket_coords_h[i] = std::floor(70.0*i/pos_h);
             }
@@ -2711,9 +2606,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             else if (ctx->minicpmv_version == 3) {
                 embed_dim = 3584;
             }
-            else if (ctx->minicpmv_version == 4) {
-                embed_dim = 3584;
-            }
             auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
 
             float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
@@ -2776,15 +2668,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
             free(positions_data);
 
-            if (!ctx->has_glm_projector) {
+            {
                 struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
-                // The patches vector is used to get rows to index into the embeds with;
-                // we should skip dim 0 only if we have CLS to avoid going out of bounds
-                // when retrieving the rows.
-                int patch_offset = ctx->has_class_embedding ? 1 : 0;
                 int* patches_data = (int*)malloc(ggml_nbytes(patches));
                 for (int i = 0; i < num_patches; i++) {
-                    patches_data[i] = i + patch_offset;
+                    patches_data[i] = i + 1;
                 }
                 ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
                 free(patches_data);
@@ -2804,19 +2692,14 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     // copy the embeddings to the location passed by the user
     ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
 
-    if (ctx->has_glm_projector) {
-        //eoi
-        ggml_tensor * eoi = ctx->vision_model.eoi_w;
-        int offset = ggml_nelements(embeddings);
-        ggml_backend_tensor_get(eoi, vec+offset, 0, ggml_nbytes(eoi));
-    }
-
     return true;
 }
 
 bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
     assert(itype < GGML_TYPE_COUNT);
-    ggml_type type = static_cast<ggml_type>(itype);
+    type = static_cast<ggml_type>(itype);
 
     auto * ctx_clip = clip_model_load(fname_inp, 2);
 
@@ -2869,8 +2752,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             }
         }
 
-        // quantize only 2D tensors and bigger than block size
-        quantize &= (ggml_n_dims(cur) == 2) && cur->ne[0] > ggml_blck_size(type);
+        // quantize only 2D tensors
+        quantize &= (ggml_n_dims(cur) == 2);
 
         if (quantize) {
             new_type = type;
@@ -2915,8 +2798,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         total_size_org += orig_size;
         total_size_new += new_size;
         gguf_set_tensor_type(ctx_out, name.c_str(), new_type);
-        GGML_ASSERT(gguf_get_tensor_size(ctx_out, gguf_find_tensor(ctx_out, name.c_str())) == new_size);
-        gguf_set_tensor_data(ctx_out, name.c_str(), new_data);
+        gguf_set_tensor_data(ctx_out, name.c_str(), new_data, new_size);
         fout.write((const char *)new_data, new_size);
         size_t pad = GGML_PAD(new_size, gguf_get_alignment(ctx_out)) - new_size;
         for (size_t j = 0; j < pad; ++j) {
@@ -2966,12 +2848,6 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         else if (ctx->minicpmv_version == 3) {
             return 3584;
         }
-        else if (ctx->minicpmv_version == 4) {
-            return 3584;
-        }
-    }
-    if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE){
-        return ctx->vision_model.mm_model_mlp_3_w->ne[1];
     }
     if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
         return ctx->vision_model.mm_1_b->ne[0];
@@ -2988,35 +2864,10 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) {
     return 0;
 }
 
-bool clip_is_glm(const struct clip_ctx * ctx) {
-    return ctx->has_glm_projector;
-}
 bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
     return ctx->has_qwen2vl_merger;
 }
 
-// Determine the number of encoder layers to iterate over
-int get_deepest_feature_layer(const struct clip_ctx * ctx) {
-    // Get the index of the second to last layer; this is the
-    // default for models that have a llava projector
-    const auto & hparams = ctx->vision_model.hparams;
-    int n_layer = hparams.n_layer - 1;
-    int deepest_feature_layer = -1;
-
-    // Handle other projectors; incrementing here indicates that we
-    // should use the last encoder layer for the vision features.
-    if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) {
-        n_layer += 1;
-    }
-
-    // If we set explicit vision feature layers, only go up to the deepest one
-    for (const auto & feature_layer : hparams.vision_feature_layer) {
-        if (feature_layer > deepest_feature_layer) {
-            deepest_feature_layer = feature_layer;
-        }
-    }
-    return deepest_feature_layer < 0 ? n_layer : deepest_feature_layer;
-}
 
 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {
     clip_image_f32 clip_img;

llama/clip.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #ifndef CLIP_H
 #define CLIP_H
 
@@ -55,7 +81,6 @@ CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 
 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
-CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
 
 CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
 CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
@@ -74,9 +99,6 @@ CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
 CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
 
-/** build image from pixels decoded by other libraries instead of stb_image.h for better performance. The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes */
-CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
-
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
 
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
@@ -93,14 +115,10 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
-CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
 
-CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
-
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 
-
 #ifdef __cplusplus
 }
 #endif

llama/common.cpp

@@ -1,10 +1,33 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #if defined(_MSC_VER)
 #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
 #endif
 
-#include "ggml.h"
-#include "gguf.h"
-
 #include "common.h"
 #include "log.h"
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
@@ -73,22 +96,6 @@
 #include <sys/syslimits.h>
 #endif
 #define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-
-//
-// CURL utils
-//
-
-using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
-
-// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
-struct curl_slist_ptr {
-    struct curl_slist * ptr = nullptr;
-    ~curl_slist_ptr() {
-        if (ptr) {
-            curl_slist_free_all(ptr);
-        }
-    }
-};
 #endif // LLAMA_USE_CURL
 
 using json = nlohmann::ordered_json;
@@ -483,48 +490,6 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
-std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
-    std::ostringstream result;
-    for (size_t i = 0; i < values.size(); ++i) {
-        if (i > 0) {
-            result << separator;
-        }
-        result << values[i];
-    }
-    return result.str();
-}
-
-std::vector<std::string> string_split(const std::string & str, const std::string & delimiter) {
-    std::vector<std::string> parts;
-    size_t start = 0;
-    size_t end = str.find(delimiter);
-
-    while (end != std::string::npos) {
-        parts.push_back(str.substr(start, end - start));
-        start = end + delimiter.length();
-        end = str.find(delimiter, start);
-    }
-
-    parts.push_back(str.substr(start));
-
-    return parts;
-}
-
-std::string string_repeat(const std::string & str, size_t n) {
-    if (n == 0) {
-        return "";
-    }
-
-    std::string result;
-    result.reserve(str.length() * n);
-
-    for (size_t i = 0; i < n; ++i) {
-        result += str;
-    }
-
-    return result;
-}
-
 std::string string_from(bool value) {
     return value ? "true" : "false";
 }
@@ -907,7 +872,7 @@ struct common_init_result common_init_from_params(common_params & params) {
     } else if (!params.model_url.empty()) {
         model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
     } else {
-        model = llama_model_load_from_file(params.model.c_str(), mparams);
+        model = llama_load_model_from_file(params.model.c_str(), mparams);
     }
 
     if (model == NULL) {
@@ -915,28 +880,26 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
     if (params.reranking) {
         bool ok = true;
 
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+        if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  BOS token, reranking will not work\n", __func__);
             ok = false;
         }
 
-        if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
+        if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
             ok = false;
         }
 
-        if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  SEP token, reranking will not work\n", __func__);
+        if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  SEP token, reranking will not work\n", __func__);
             ok = false;
         }
 
         if (!ok) {
-            llama_model_free(model);
+            llama_free_model(model);
 
             return iparams;
         }
@@ -944,10 +907,10 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     auto cparams = common_context_params_to_llama(params);
 
-    llama_context * lctx = llama_init_from_model(model, cparams);
+    llama_context * lctx = llama_new_context_with_model(model, cparams);
     if (lctx == NULL) {
         LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
-        llama_model_free(model);
+        llama_free_model(model);
         return iparams;
     }
 
@@ -958,26 +921,25 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     if (!params.control_vectors.empty()) {
         if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
-        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_model_n_layer(model);
+        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
             llama_free(lctx);
-            llama_model_free(model);
+            llama_free_model(model);
 
             return iparams;
         }
 
-        int err = llama_apply_adapter_cvec(
-                lctx,
-                cvec.data.data(),
-                cvec.data.size(),
-                cvec.n_embd,
-                params.control_vector_layer_start,
-                params.control_vector_layer_end);
+        int err = llama_control_vector_apply(lctx,
+                                             cvec.data.data(),
+                                             cvec.data.size(),
+                                             cvec.n_embd,
+                                             params.control_vector_layer_start,
+                                             params.control_vector_layer_end);
         if (err) {
             llama_free(lctx);
-            llama_model_free(model);
+            llama_free_model(model);
 
             return iparams;
         }
@@ -985,12 +947,12 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
-        llama_adapter_lora_ptr lora;
-        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
+        llama_lora_adapter_ptr lora;
+        lora.reset(llama_lora_adapter_init(model, la.path.c_str()));
         if (lora == nullptr) {
             LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
             llama_free(lctx);
-            llama_model_free(model);
+            llama_free_model(model);
             return iparams;
         }
 
@@ -999,17 +961,17 @@ struct common_init_result common_init_from_params(common_params & params) {
     }
 
     if (!params.lora_init_without_apply) {
-        common_set_adapter_lora(lctx, params.lora_adapters);
+        common_lora_adapters_apply(lctx, params.lora_adapters);
     }
 
-    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
+    if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sampling.ignore_eos = false;
     }
 
     if (params.sampling.ignore_eos) {
-        for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-            if (llama_vocab_is_eog(vocab, i)) {
+        for (llama_token i = 0; i < llama_n_vocab(model); i++) {
+            if (llama_token_is_eog(model, i)) {
                 LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
                 params.sampling.logit_bias.push_back({i, -INFINITY});
             }
@@ -1030,9 +992,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
         std::vector<llama_token> tmp;
-        llama_token bos = llama_vocab_bos(vocab);
-        llama_token eos = llama_vocab_eos(vocab);
-
+        llama_token bos = llama_token_bos(model);
+        llama_token eos = llama_token_eos(model);
         // some models (e.g. T5) don't have a BOS token
         if (bos != LLAMA_TOKEN_NULL) {
             tmp.push_back(bos);
@@ -1047,7 +1008,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_encoder(model)) {
             llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
             llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
-            if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
+            if (decoder_start_token_id == -1) {
                 decoder_start_token_id = bos;
             }
             tmp.clear();
@@ -1067,11 +1028,11 @@ struct common_init_result common_init_from_params(common_params & params) {
     return iparams;
 }
 
-void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
-    llama_clear_adapter_lora(ctx);
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora) {
+    llama_lora_adapter_clear(ctx);
     for (auto & la : lora) {
         if (la.scale != 0.0f) {
-            llama_set_adapter_lora(ctx, la.ptr, la.scale);
+            llama_lora_adapter_set(ctx, la.ptr, la.scale);
         }
     }
 }
@@ -1085,6 +1046,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
+    mparams.rpc_servers     = params.rpc_servers.c_str();
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
@@ -1187,8 +1149,7 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
 
 static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
     // Initialize libcurl
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
+    std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
     if (!curl) {
         LOG_ERR("%s: error initializing libcurl\n", __func__);
         return false;
@@ -1202,9 +1163,11 @@ static bool common_download_file(const std::string & url, const std::string & pa
 
     // Check if hf-token or bearer-token was specified
     if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-        curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+      std::string auth_header = "Authorization: Bearer ";
+      auth_header += hf_token.c_str();
+      struct curl_slist *http_headers = NULL;
+      http_headers = curl_slist_append(http_headers, auth_header.c_str());
+      curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers);
     }
 
 #if defined(_WIN32)
@@ -1474,7 +1437,7 @@ struct llama_model * common_load_model_from_url(
         }
     }
 
-    return llama_model_load_from_file(local_path.c_str(), params);
+    return llama_load_model_from_file(local_path.c_str(), params);
 }
 
 struct llama_model * common_load_model_from_hf(
@@ -1500,80 +1463,6 @@ struct llama_model * common_load_model_from_hf(
     return common_load_model_from_url(model_url, local_path, hf_token, params);
 }
 
-/**
- * Allow getting the HF file from the HF repo with tag (like ollama), for example:
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
- * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
- * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
- *
- * Return pair of <repo, file> (with "repo" already having tag removed)
- *
- * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
- */
-std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
-    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
-    std::string tag = parts.size() > 1 ? parts.back() : "latest";
-    std::string hf_repo = parts[0];
-    if (string_split<std::string>(hf_repo, '/').size() != 2) {
-        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
-    }
-
-    // fetch model info from Hugging Face Hub API
-    json model_info;
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
-    std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-    if (!hf_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + hf_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-    }
-    // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-
-    CURLcode res = curl_easy_perform(curl.get());
-
-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
-    }
-
-    long res_code;
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-    if (res_code == 200) {
-        model_info = json::parse(res_str);
-    } else if (res_code == 401) {
-        throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
-    } else {
-        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
-    }
-
-    // check response
-    if (!model_info.contains("ggufFile")) {
-        throw std::runtime_error("error: model does not have ggufFile");
-    }
-    json & gguf_file = model_info.at("ggufFile");
-    if (!gguf_file.contains("rfilename")) {
-        throw std::runtime_error("error: ggufFile does not have rfilename");
-    }
-
-    return std::make_pair(hf_repo, gguf_file.at("rfilename"));
-}
-
 #else
 
 struct llama_model * common_load_model_from_url(
@@ -1595,11 +1484,6 @@ struct llama_model * common_load_model_from_hf(
     return nullptr;
 }
 
-std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return std::make_pair("", "");
-}
-
 #endif // LLAMA_USE_CURL
 
 //
@@ -1698,23 +1582,21 @@ std::vector<llama_token> common_tokenize(
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-    return common_tokenize(vocab, text, add_special, parse_special);
+    return common_tokenize(llama_get_model(ctx), text, add_special, parse_special);
 }
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_vocab * vocab,
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -1723,18 +1605,12 @@ std::vector<llama_token> common_tokenize(
 }
 
 std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-    return common_token_to_piece(vocab, token, special);
-}
-
-std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token token, bool special) {
     std::string piece;
     piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
-    const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+    const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
     if (n_chars < 0) {
         piece.resize(-n_chars);
-        int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
         GGML_ASSERT(check == -n_chars);
     }
     else {
@@ -1744,19 +1620,13 @@ std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token
     return piece;
 }
 
-std::string common_detokenize(const struct llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-    return common_detokenize(vocab, tokens, special);
-}
-
-std::string common_detokenize(const struct llama_vocab * vocab, const std::vector<llama_token> & tokens, bool special) {
+std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
     std::string text;
     text.resize(std::max(text.capacity(), tokens.size()));
-    int32_t n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
     if (n_chars < 0) {
         text.resize(-n_chars);
-        n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
         GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
     }
 
@@ -1766,6 +1636,103 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto
     return text;
 }
 
+//
+// Chat template utils
+//
+
+std::string common_get_builtin_chat_template(const struct llama_model * model) {
+    static const char * template_key = "tokenizer.chat_template";
+    // call with NULL buffer to get the total size of the string
+    int32_t res = llama_model_meta_val_str(model, template_key, NULL, 0);
+    if (res > 0) {
+        std::vector<char> model_template(res + 1, 0);
+        llama_model_meta_val_str(model, template_key, model_template.data(), model_template.size());
+        return std::string(model_template.data(), model_template.size() - 1);
+    }
+    return "";
+}
+
+bool common_chat_verify_template(const std::string & tmpl) {
+    llama_chat_message chat[] = {{"user", "test"}};
+    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
+    return res >= 0;
+}
+
+std::string common_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<common_chat_msg> & msgs,
+        bool add_ass) {
+    int alloc_size = 0;
+    bool fallback = false; // indicate if we must fallback to default chatml
+    std::vector<llama_chat_message> chat;
+    for (auto & msg : msgs) {
+        chat.push_back({msg.role.c_str(), msg.content.c_str()});
+        alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
+    }
+
+    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
+    std::vector<char> buf(alloc_size);
+
+    // run the first time to get the total output length
+    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+
+    // error: chat template is not supported
+    if (res < 0) {
+        if (ptr_tmpl != nullptr) {
+            // if the custom "tmpl" is not supported, we throw an error
+            // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
+            throw std::runtime_error("this custom template is not supported");
+        } else {
+            // If the built-in template is not supported, we default to chatml
+            res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+            fallback = true;
+        }
+    }
+
+    // if it turns out that our buffer is too small, we resize it
+    if ((size_t) res > buf.size()) {
+        buf.resize(res);
+        res = llama_chat_apply_template(
+            fallback ? nullptr : model,
+            fallback ? "chatml" : ptr_tmpl,
+            chat.data(), chat.size(), add_ass, buf.data(), buf.size());
+    }
+
+    std::string formatted_chat(buf.data(), res);
+    return formatted_chat;
+}
+
+std::string common_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
+        bool add_ass) {
+    std::ostringstream ss;
+    auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(model, tmpl, past_msg, false);
+    std::vector<common_chat_msg> chat_new(past_msg);
+    // if the past_msg ends with a newline, we must preserve it in the formatted version
+    if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
+        ss << "\n";
+    };
+    // format chat with new_msg
+    chat_new.push_back(new_msg);
+    auto fmt_new_msg = common_chat_apply_template(model, tmpl, chat_new, add_ass);
+    // get the diff part
+    ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
+    return ss.str();
+}
+
+std::string common_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl) {
+    std::vector<common_chat_msg> msgs = {
+        {"system",    "You are a helpful assistant"},
+        {"user",      "Hello"},
+        {"assistant", "Hi there"},
+        {"user",      "How are you?"},
+    };
+    return common_chat_apply_template(model, tmpl, msgs, true);
+}
+
 //
 // KV cache utils
 //

llama/common.h

@@ -1,10 +1,35 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 // Various helper functions and utilities
 
 #pragma once
 
 #include "llama-cpp.h"
 
-#include <set>
 #include <string>
 #include <vector>
 #include <sstream>
@@ -25,11 +50,11 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
-struct common_adapter_lora_info {
+struct common_lora_adapter_info {
     std::string path;
     float scale;
 
-    struct llama_adapter_lora * ptr;
+    struct llama_lora_adapter * ptr;
 };
 
 using llama_tokens = std::vector<llama_token>;
@@ -104,17 +129,6 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
-enum common_conversation_mode {
-    COMMON_CONVERSATION_MODE_DISABLED = 0,
-    COMMON_CONVERSATION_MODE_ENABLED  = 1,
-    COMMON_CONVERSATION_MODE_AUTO     = 2,
-};
-
-struct common_grammar_trigger {
-    std::string word;
-    bool at_start;
-};
-
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -140,7 +154,6 @@ struct common_params_sampling {
     int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
     int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
     int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   top_n_sigma        = -1.00f;// -1.0 = disabled
     float   mirostat_tau       = 5.00f; // target entropy
     float   mirostat_eta       = 0.10f; // learning rate
     bool    ignore_eos         = false;
@@ -161,11 +174,7 @@ struct common_params_sampling {
         COMMON_SAMPLER_TYPE_TEMPERATURE,
     };
 
-    std::string                         grammar; // optional BNF-like grammar to constrain sampling
-    bool                                grammar_lazy = false;
-    std::vector<common_grammar_trigger> grammar_trigger_words;  // optional trigger words to trigger lazy grammar
-    std::vector<llama_token>            grammar_trigger_tokens; // optional trigger tokens to trigger lazy grammar and print trigger special tokens.
-    std::set<llama_token>               preserved_tokens;
+    std::string grammar; // optional BNF-like grammar to constrain sampling
 
     std::vector<llama_logit_bias> logit_bias; // logit biases to apply
 
@@ -178,19 +187,15 @@ struct common_params_speculative {
 
     int32_t n_ctx        =     0; // draft context size
     int32_t n_max        =    16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min        =     0; // minimum number of draft tokens to use for speculative decoding
+    int32_t n_min        =     5; // minimum number of draft tokens to use for speculative decoding
     int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     float   p_split      =  0.1f; // speculative decoding split probability
-    float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
+    float   p_min        =  0.9f; // minimum speculative decoding probability (greedy)
 
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
-
-    std::string model = "";     // draft model for speculative decoding                      // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    std::string model = ""; // draft model for speculative decoding                          // NOLINT
 };
 
 struct common_params_vocoder {
@@ -199,13 +204,6 @@ struct common_params_vocoder {
 
     std::string model     = ""; // model path                                                // NOLINT
     std::string model_url = ""; // model url to download                                     // NOLINT
-
-    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
-};
-
-enum common_reasoning_format {
-    COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
 };
 
 struct common_params {
@@ -268,13 +266,14 @@ struct common_params {
     std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
+    std::string rpc_servers          = ""; // comma separated list of RPC servers                           // NOLINT
 
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
-    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
-    std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
+    std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -298,11 +297,11 @@ struct common_params {
     bool   kl_divergence    = false; // compute KL divergence
 
     bool usage             = false; // print usage
-    bool completion        = false; // print source-able completion script
     bool use_color         = false; // use color to distinguish generations and inputs
     bool special           = false; // enable special token output
     bool interactive       = false; // interactive mode
     bool interactive_first = false; // wait for user input immediately
+    bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
     bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
@@ -328,8 +327,6 @@ struct common_params {
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
-    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
-
     // multimodal models (see examples/llava)
     std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
     std::vector<std::string> image; // path to image file(s)
@@ -351,9 +348,7 @@ struct common_params {
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
 
     std::vector<std::string> api_keys;
 
@@ -432,13 +427,13 @@ bool set_process_priority(enum ggml_sched_priority prio);
 //
 
 #ifdef __GNUC__
-#    if defined(__MINGW32__) && !defined(__clang__)
-#        define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
-#    else
-#        define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
-#    endif
+#ifdef __MINGW32__
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
 #else
-#    define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
 #endif
 
 LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
@@ -447,10 +442,6 @@ std::string string_format(const char * fmt, ...);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
-std::string string_join(const std::vector<std::string> & values, const std::string & separator);
-std::vector<std::string> string_split(const std::string & str, const std::string & delimiter);
-std::string string_repeat(const std::string & str, size_t n);
-
 void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
 
 template<class T>
@@ -489,11 +480,6 @@ static bool string_starts_with(const std::string & str,
     return str.rfind(prefix, 0) == 0;
 }
 
-static bool string_ends_with(const std::string & str,
-                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
-
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 
@@ -521,7 +507,7 @@ struct common_init_result {
     llama_model_ptr   model;
     llama_context_ptr context;
 
-    std::vector<llama_adapter_lora_ptr> lora;
+    std::vector<llama_lora_adapter_ptr> lora;
 };
 
 struct common_init_result     common_init_from_params(common_params & params);
@@ -535,7 +521,6 @@ struct llama_model * common_load_model_from_url(
     const std::string & local_path,
     const std::string & hf_token,
     const struct llama_model_params & params);
-
 struct llama_model * common_load_model_from_hf(
     const std::string & repo,
     const std::string & remote_path,
@@ -543,12 +528,8 @@ struct llama_model * common_load_model_from_hf(
     const std::string & hf_token,
     const struct llama_model_params & params);
 
-std::pair<std::string, std::string> common_get_hf_file(
-    const std::string & hf_repo_with_tag,
-    const std::string & hf_token);
-
 // clear LoRA adapters from context, then apply new list of adapters
-void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora);
 
 //
 // Batch utils
@@ -586,7 +567,7 @@ std::vector<llama_token> common_tokenize(
                         bool   parse_special = false);
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_vocab * vocab,
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special = false);
@@ -598,23 +579,48 @@ std::string common_token_to_piece(
                        llama_token   token,
                        bool          special = true);
 
-std::string common_token_to_piece(
-          const struct llama_vocab * vocab,
-                       llama_token   token,
-                       bool          special = true);
-
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
 // optionally renders special/control tokens
 std::string common_detokenize(
-            const struct llama_context * ctx,
+                         llama_context * ctx,
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-std::string common_detokenize(
-              const struct llama_vocab * vocab,
-        const std::vector<llama_token> & tokens,
-                                  bool   special = true);
+//
+// Chat template utils
+//
+
+// same with llama_chat_message, but uses std::string
+struct common_chat_msg {
+    std::string role;
+    std::string content;
+};
+
+// Get the built-in chat template for the model. Return empty string if not present.
+std::string common_get_builtin_chat_template(const struct llama_model * model);
+
+// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
+bool common_chat_verify_template(const std::string & tmpl);
+
+// CPP wrapper for llama_chat_apply_template
+// If the built-in template is not supported, we default to chatml
+// If the custom "tmpl" is not supported, we throw an error
+std::string common_chat_apply_template(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<common_chat_msg> & chat,
+        bool add_ass);
+
+// Format single message, while taking into account the position of that message in chat history
+std::string common_chat_format_single(const struct llama_model * model,
+        const std::string & tmpl,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
+        bool add_ass);
+
+// Returns an example of formatted chat
+std::string common_chat_format_example(const struct llama_model * model,
+        const std::string & tmpl);
 
 //
 // KV cache utils

llama/ggml-alloc.c

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #include "ggml-alloc.h"
 #include "ggml-backend-impl.h"
 #include "ggml.h"
@@ -37,7 +63,6 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
     return true;
 }
 
-// ops that return true for this function must not use restrict pointers for their backend implementations
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
         case GGML_OP_SCALE:
@@ -53,12 +78,8 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_LOG:
         case GGML_OP_UNARY:
         case GGML_OP_ROPE:
-        case GGML_OP_ROPE_BACK:
-        case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_SOFT_MAX_BACK:
             return true;
 
         default:
@@ -989,7 +1010,19 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
             this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
         }
 
-        if (cur_buf_size > 0 && (cur_buf_size + this_size) > max_size) {
+        if (this_size > max_size) {
+            GGML_LOG_ERROR("%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
             // allocate tensors in the current buffer
             if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
                 return NULL;

llama/ggml-alloc.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 #include "ggml.h"

llama/ggml-backend-impl.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 // ggml-backend internal header
@@ -208,6 +234,7 @@ extern "C" {
 
     // Internal backend registry API
     GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
+    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
 
     // Add backend dynamic loading support to the backend
 

llama/ggml-backend-reg.cpp

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
 #include "ggml-impl.h"
@@ -66,6 +92,26 @@
 #include "ggml-kompute.h"
 #endif
 
+// disable C++17 deprecation warning for std::codecvt_utf8
+#if defined(__clang__)
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
+
+static std::wstring utf8_to_utf16(const std::string & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.from_bytes(str);
+}
+
+static std::string utf16_to_utf8(const std::wstring & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.to_bytes(str);
+}
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#endif
+
 #ifdef _WIN32
 
 using dl_handle = std::remove_pointer_t<HMODULE>;
@@ -76,7 +122,7 @@ struct dl_handle_deleter {
     }
 };
 
-static dl_handle * dl_load_library(const std::filesystem::path & path) {
+static dl_handle * dl_load_library(const std::wstring & path) {
     // suppress error dialogs for missing DLLs
     DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
     SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
@@ -109,8 +155,8 @@ struct dl_handle_deleter {
     }
 };
 
-static void * dl_load_library(const std::filesystem::path & path) {
-    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
+static void * dl_load_library(const std::wstring & path) {
+    dl_handle * handle = dlopen(utf16_to_utf8(path).c_str(), RTLD_NOW | RTLD_LOCAL);
 
     return handle;
 }
@@ -121,25 +167,6 @@ static void * dl_get_sym(dl_handle * handle, const char * name) {
 
 #endif
 
-static std::string path_to_string(const std::filesystem::path & path)
-{
-#ifdef _WIN32
-    const std::wstring wstr = path.wstring();
-    const int size_needed = WideCharToMultiByte(CP_UTF8, 0, wstr.c_str(), -1, nullptr, 0, nullptr, nullptr);
-    if (size_needed <= 0) {
-        return std::string();
-    }
-
-    // size_needed includes the null terminator
-    std::string str(size_needed - 1, '\0');
-    WideCharToMultiByte(CP_UTF8, 0, wstr.c_str(), -1, str.data(), size_needed, nullptr, nullptr);
-    return str;
-#else
-    return path.string();
-#endif
-}
-
-
 using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
 
 struct ggml_backend_reg_entry {
@@ -149,7 +176,7 @@ struct ggml_backend_reg_entry {
 
 struct ggml_backend_registry {
     std::vector<ggml_backend_reg_entry> backends;
-    std::vector<std::pair<ggml_backend_dev_t, int>> devices;
+    std::vector<ggml_backend_dev_t> devices;
 
     ggml_backend_registry() {
 #ifdef GGML_USE_CUDA
@@ -194,7 +221,7 @@ struct ggml_backend_registry {
         }
     }
 
-    void register_backend(ggml_backend_reg_t reg, int score = -1, dl_handle_ptr handle = nullptr) {
+    void register_backend(ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
         if (!reg) {
             return;
         }
@@ -205,27 +232,22 @@ struct ggml_backend_registry {
 #endif
         backends.push_back({ reg, std::move(handle) });
         for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
-            register_device(ggml_backend_reg_dev_get(reg, i), score);
+            register_device(ggml_backend_reg_dev_get(reg, i));
         }
     }
 
-    void register_device(ggml_backend_dev_t device, int score = -1) {
+    void register_device(ggml_backend_dev_t device) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
 #endif
-        devices.push_back({device, score});
-        std::stable_sort(devices.begin(), devices.end(),
-            [](const auto & a, const auto & b) {
-                return a.second > b.second;
-            }
-        );
+        devices.push_back(device);
     }
 
-    ggml_backend_reg_t load_backend(const std::filesystem::path & path, bool silent) {
+    ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) {
         dl_handle_ptr handle { dl_load_library(path) };
         if (!handle) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_to_string(path).c_str());
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(path).c_str());
             }
             return nullptr;
         }
@@ -233,7 +255,7 @@ struct ggml_backend_registry {
         auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
         if (score_fn && score_fn() == 0) {
             if (!silent) {
-                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path_to_string(path).c_str());
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, utf16_to_utf8(path).c_str());
             }
             return nullptr;
         }
@@ -241,7 +263,7 @@ struct ggml_backend_registry {
         auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
         if (!backend_init_fn) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path_to_string(path).c_str());
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, utf16_to_utf8(path).c_str());
             }
             return nullptr;
         }
@@ -250,18 +272,18 @@ struct ggml_backend_registry {
         if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
             if (!silent) {
                 if (!reg) {
-                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path_to_string(path).c_str());
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, utf16_to_utf8(path).c_str());
                 } else {
                     GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
-                        __func__, path_to_string(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
+                        __func__, utf16_to_utf8(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
                 }
             }
             return nullptr;
         }
 
-        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path_to_string(path).c_str());
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str());
 
-        register_backend(reg, score_fn ? score_fn() : -1, std::move(handle));
+        register_backend(reg, std::move(handle));
 
         return reg;
     }
@@ -284,7 +306,7 @@ struct ggml_backend_registry {
         // remove devices
         devices.erase(
             std::remove_if(devices.begin(), devices.end(),
-                            [reg](std::pair<ggml_backend_dev_t, int> dev) { return ggml_backend_dev_backend_reg(dev.first) == reg; }),
+                            [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
             devices.end());
 
         // remove backend
@@ -342,7 +364,7 @@ size_t ggml_backend_dev_count() {
 
 ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
     GGML_ASSERT(index < ggml_backend_dev_count());
-    return get_reg().devices[index].first;
+    return get_reg().devices[index];
 }
 
 ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
@@ -395,14 +417,14 @@ ggml_backend_t ggml_backend_init_best(void) {
 
 // Dynamic loading
 ggml_backend_reg_t ggml_backend_load(const char * path) {
-    return get_reg().load_backend(path, false);
+    return get_reg().load_backend(utf8_to_utf16(path), false);
 }
 
 void ggml_backend_unload(ggml_backend_reg_t reg) {
     get_reg().unload_backend(reg, true);
 }
 
-static std::filesystem::path get_executable_path() {
+static std::wstring get_executable_path() {
 #if defined(__APPLE__)
     // get executable path
     std::vector<char> path;
@@ -414,9 +436,15 @@ static std::filesystem::path get_executable_path() {
         }
         path.resize(size);
     }
-
-    return std::filesystem::path(path.data()).parent_path();
+    std::string base_path(path.data(), size);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('/');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return utf8_to_utf16(base_path + "/");
 #elif defined(__linux__) || defined(__FreeBSD__)
+    std::string base_path = ".";
     std::vector<char> path(1024);
     while (true) {
         // get executable path
@@ -429,55 +457,76 @@ static std::filesystem::path get_executable_path() {
             break;
         }
         if (len < (ssize_t) path.size()) {
-            return std::filesystem::path(path.data()).parent_path();
+            base_path = std::string(path.data(), len);
+            // remove executable name
+            auto last_slash = base_path.find_last_of('/');
+            if (last_slash != std::string::npos) {
+                base_path = base_path.substr(0, last_slash);
+            }
+            break;
         }
         path.resize(path.size() * 2);
     }
+
+    return utf8_to_utf16(base_path + "/");
 #elif defined(_WIN32)
     std::vector<wchar_t> path(MAX_PATH);
     DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
     if (len == 0) {
         return {};
     }
-
-    return std::filesystem::path(path.data()).parent_path();
-#endif
+    std::wstring base_path(path.data(), len);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('\\');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + L"\\";
+#else
     return {};
-}
-
-static std::string backend_filename_prefix() {
-#ifdef _WIN32
-    return "ggml-";
-#else
-    return "libggml-";
 #endif
 }
 
-static std::string backend_filename_suffix() {
+static std::wstring backend_filename_prefix() {
 #ifdef _WIN32
-    return ".dll";
+    return L"ggml-";
 #else
-    return ".so";
+    return L"libggml-";
+#endif
+}
+
+static std::wstring backend_filename_suffix() {
+#ifdef _WIN32
+    return L".dll";
+#else
+    return L".so";
+#endif
+}
+
+static std::wstring path_separator() {
+#ifdef _WIN32
+    return L"\\";
+#else
+    return L"/";
 #endif
 }
 
 static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
     // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
      // TODO: search system paths
-    namespace fs = std::filesystem;
-    std::string file_prefix = backend_filename_prefix() + name + "-";
-    std::vector<fs::path> search_paths;
-
+    std::wstring file_prefix = backend_filename_prefix() + utf8_to_utf16(name) + L"-";
+    std::vector<std::wstring> search_paths;
     if (user_search_path == nullptr) {
-        search_paths.push_back(fs::current_path());
+        search_paths.push_back(L"." + path_separator());
         search_paths.push_back(get_executable_path());
     } else {
-        search_paths.push_back(fs::u8path(user_search_path));
+        search_paths.push_back(utf8_to_utf16(user_search_path) + path_separator());
     }
 
     int best_score = 0;
-    fs::path best_path;
+    std::wstring best_path;
 
+    namespace fs = std::filesystem;
     for (const auto & search_path : search_paths) {
         if (!fs::exists(search_path)) {
             continue;
@@ -485,26 +534,29 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
         for (const auto & entry : dir_it) {
             if (entry.is_regular_file()) {
-                std::string filename = entry.path().filename().string();
-                std::string ext = entry.path().extension().string();
+                std::wstring filename = entry.path().filename().wstring();
+                std::wstring ext = entry.path().extension().wstring();
                 if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
-                    dl_handle_ptr handle { dl_load_library(entry.path()) };
-                    if (!handle) {
-                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_to_string(entry.path()).c_str());
-                        continue;
+                    dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
+                    if (!handle && !silent) {
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
                     }
-
-                    auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
-                    if (!score_fn) {
-                        GGML_LOG_DEBUG("%s: failed to find ggml_backend_score in %s\n", __func__, path_to_string(entry.path()).c_str());
-                        continue;
-                    }
-
-                    int s = score_fn();
-                    GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, path_to_string(entry.path()).c_str(), s);
-                    if (s > best_score) {
-                        best_score = s;
-                        best_path = entry.path();
+                    if (handle) {
+                        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+                        if (score_fn) {
+                            int s = score_fn();
+#ifndef NDEBUG
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
+#endif
+                            if (s > best_score) {
+                                best_score = s;
+                                best_path = entry.path().wstring();
+                            }
+                        } else {
+                            if (!silent) {
+                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
+                            }
+                        }
                     }
                 }
             }
@@ -514,7 +566,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
     if (best_score == 0) {
         // try to load the base backend
         for (const auto & search_path : search_paths) {
-            fs::path path = fs::path(search_path) / (backend_filename_prefix() + name + backend_filename_suffix());
+            std::wstring path = search_path + backend_filename_prefix() + utf8_to_utf16(name) + backend_filename_suffix();
             if (fs::exists(path)) {
                 return get_reg().load_backend(path, silent);
             }
@@ -529,14 +581,6 @@ void ggml_backend_load_all() {
     ggml_backend_load_all_from_path(nullptr);
 }
 
-static void ggml_backend_try_load_best(const char * name, bool silent, const char * user_search_path) {
-    try {
-        ggml_backend_load_best(name, silent, user_search_path);
-    } catch (const std::exception & e) {
-        GGML_LOG_DEBUG("%s: failed to load %s: %s\n", __func__, name, e.what());
-    }
-}
-
 void ggml_backend_load_all_from_path(const char * dir_path) {
 #ifdef NDEBUG
     bool silent = true;
@@ -544,21 +588,16 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     bool silent = false;
 #endif
 
-    ggml_backend_try_load_best("blas", silent, dir_path);
-    ggml_backend_try_load_best("cann", silent, dir_path);
-    ggml_backend_try_load_best("cuda", silent, dir_path);
-    ggml_backend_try_load_best("hip", silent, dir_path);
-    ggml_backend_try_load_best("kompute", silent, dir_path);
-    ggml_backend_try_load_best("metal", silent, dir_path);
-    ggml_backend_try_load_best("rpc", silent, dir_path);
-    ggml_backend_try_load_best("sycl", silent, dir_path);
-    ggml_backend_try_load_best("vulkan", silent, dir_path);
-    ggml_backend_try_load_best("opencl", silent, dir_path);
-    ggml_backend_try_load_best("musa", silent, dir_path);
-    ggml_backend_try_load_best("cpu", silent, dir_path);
-    // check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend
-    const char * backend_path = std::getenv("GGML_BACKEND_PATH");
-    if (backend_path) {
-        ggml_backend_load(backend_path);
-    }
+    ggml_backend_load_best("blas", silent, dir_path);
+    ggml_backend_load_best("cann", silent, dir_path);
+    ggml_backend_load_best("cuda", silent, dir_path);
+    ggml_backend_load_best("hip", silent, dir_path);
+    ggml_backend_load_best("kompute", silent, dir_path);
+    ggml_backend_load_best("metal", silent, dir_path);
+    ggml_backend_load_best("rpc", silent, dir_path);
+    ggml_backend_load_best("sycl", silent, dir_path);
+    ggml_backend_load_best("vulkan", silent, dir_path);
+    ggml_backend_load_best("opencl", silent, dir_path);
+    ggml_backend_load_best("musa", silent, dir_path);
+    ggml_backend_load_best("cpu", silent, dir_path);
 }

llama/ggml-backend.cpp

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 // Note: porting this file to C++ is a work in progress
 
 #ifdef _WIN32
@@ -106,6 +132,12 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
     if (buffer->iface.free_buffer != NULL) {
         buffer->iface.free_buffer(buffer);
     }
+
+// TODO: this needs to be freed in cuda and hip backends because
+// the cuda backend implementation compiled with msvc
+#if !defined(GGML_USE_CUDA) && !defined(GGML_USE_HIP)
+    delete buffer;
+#endif
 }
 
 size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
@@ -763,7 +795,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
+            if (src_backend_id == sched->n_backends - 1) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");

llama/ggml-backend.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 #include "ggml.h"
@@ -203,8 +229,6 @@ extern "C" {
     // Backend registry
     //
 
-    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
-
     // Backend (reg) enumeration
     GGML_API size_t             ggml_backend_reg_count(void);
     GGML_API ggml_backend_reg_t ggml_backend_reg_get(size_t index);

llama/ggml-blas.cpp

@@ -1,3 +1,31 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifdef GGML_USE_BLAS
+
 #include "ggml-impl.h"
 #include "ggml-blas.h"
 #include "ggml-backend-impl.h"
@@ -515,3 +543,5 @@ ggml_backend_reg_t ggml_backend_blas_reg(void) {
 }
 
 GGML_BACKEND_DL_IMPL(ggml_backend_blas_reg)
+
+#endif // GGML_USE_BLAS

llama/ggml-blas.h

@@ -0,0 +1,51 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
+
+GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
+
+// number of threads used for conversion to float
+// for openblas and blis, this will also set the number of threads used for blas operations
+GGML_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
+
+
+#ifdef  __cplusplus
+}
+#endif

llama/ggml-common.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #ifndef GGML_COMMON_DECL
 
 #if defined(GGML_COMMON_DECL_C)
@@ -473,6 +499,7 @@ GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
     240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
 GGML_TABLE_END()
 
+//#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A // lowest compute capability for integer intrinsics
 GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
     0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
@@ -507,6 +534,7 @@ GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
     0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
 GGML_TABLE_END()
+//#endif
 
 
 GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)

llama/ggml-cpp.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 #ifndef __cplusplus
@@ -7,7 +33,6 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
-#include "gguf.h"
 #include <memory>
 
 // Smart pointers for ggml types

llama/ggml-cpu-aarch64.cpp

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #define GGML_COMMON_IMPL_CPP
 #define GGML_COMMON_DECL_CPP
 #include "ggml-common.h"
@@ -4169,8 +4195,6 @@ static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(g
     buffer->buft              = buft;
     buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
     buffer->iface.set_tensor  = ggml_backend_cpu_aarch64_buffer_set_tensor;
-    buffer->iface.get_tensor  = nullptr;
-    buffer->iface.cpy_tensor  = nullptr;
     return buffer;
 }
 

llama/ggml-cpu-aarch64.h

@@ -0,0 +1,34 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml-cpu-traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

llama/ggml-cpu-impl.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 // GGML CPU internal header
@@ -59,15 +85,6 @@ struct ggml_compute_params {
 #endif
 #endif
 
-#if defined(__s390x__) && defined(__VEC__)
-#ifndef __VXE__
-#define __VXE__
-#endif
-#ifndef __VXE2__
-#define __VXE2__
-#endif
-#endif
-
 #if defined(__ARM_FEATURE_SVE)
 #include <arm_sve.h>
 #include <sys/prctl.h>
@@ -368,158 +385,22 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #endif
 #endif
 
-#if defined(__VXE__) || defined(__VXE2__)
-#include <vecintrin.h>
-
-#define vec_neg(a)    (-(a))                // Vector Negate
-#define vec_add(a, b) ((a) + (b))           // Vector Add
-#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
-#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
-#define vec_div(a, b) ((a) / (b))           // Vector Divide
-#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
-#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
-#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
-#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
-#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
-
-#ifndef vec_and
-#define vec_and(a, b) ((a) & (b)) // Vector AND
-#endif
-
-#ifndef vec_or
-#define vec_or(a, b)  ((a) | (b)) // Vector OR
-#endif
-
-#ifndef vec_xor
-#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
-#endif
-
-typedef signed char char8x16_t __attribute__((vector_size(16)));
-typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
-
-typedef int8_t  int8x16_t __attribute__((vector_size(16)));
-typedef int16_t int16x8_t __attribute__((vector_size(16)));
-typedef int32_t int32x4_t __attribute__((vector_size(16)));
-
-typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
-typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
-typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
-
-typedef float float32x4_t __attribute__((vector_size(16)));
-typedef double double64x2_t __attribute((vector_size(16)));
-
-typedef signed long long long64x2_t __attribute((vector_size(16)));
-typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vec_xl( 0, ptr);
-    res.val[1] = vec_xl(16, ptr);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vec_xl( 0, ptr);
-    res.val[1] = vec_xl(16, ptr);
-    res.val[2] = vec_xl(32, ptr);
-    res.val[3] = vec_xl(48, ptr);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vec_xl( 0, ptr);
-    res.val[1] = vec_xl(16, ptr);
-    res.val[2] = vec_xl(32, ptr);
-    res.val[3] = vec_xl(48, ptr);
-
-    return res;
-}
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vec_xl( 0, ptr);
-    res.val[1] = vec_xl(16, ptr);
-
-    return res;
-}
-
-/*
-    ! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
-    !          or iq4_nl for example implementation.
-*/
-inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
-    const uchar8x16_t v_maske = {  0,  1,  4,  5,  8,  9, 12, 13,
-                                  16, 17, 20, 21, 24, 25, 28, 29 };
-
-    const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
-    const int16x8_t v_abe = vec_perm(a, b, v_maske);
-    return v_abo + v_abe;
-}
-
-inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
-    return acc + (vec_unpackh(p) + vec_unpackl(p));
-}
-
-#endif
-
 #if defined(__loongarch_asx)
+
+typedef union {
+    int32_t i;
+    float f;
+} ft_union;
+
 /* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(const float val) {
-    v4f32 res = {val, val, val, val};
-    return (__m128)res;
+static __m128 __lsx_vreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
 }
 
-static __m256 __lasx_xvreplfr2vr_s(const float val) {
-    v8f32 res = {val, val, val, val, val, val, val, val};
-    return (__m256)res;
+static __m256 __lasx_xvreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
 }
 #endif
 

llama/ggml-cpu-quants.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 #define GGML_COMMON_DECL_C

llama/ggml-cpu-traits.cpp

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #include "ggml-cpu-traits.h"
 
 #include "ggml-backend-impl.h"

llama/ggml-cpu-traits.h

@@ -0,0 +1,64 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#pragma once
+#include "ggml-backend-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+#    include <vector>
+extern "C" {
+#endif
+
+// return true if op part of extra "accelerator"
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
+
+#ifdef __cplusplus
+}
+
+namespace ggml::cpu {
+// register in tensor->extra
+class tensor_traits {
+  public:
+    virtual ~tensor_traits();
+    virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size)        = 0;
+    virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
+};
+
+class extra_buffer_type {
+  public:
+    virtual ~extra_buffer_type();
+    virtual bool            supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
+    virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op)                   = 0;
+};
+}  // namespace ggml::cpu
+
+// implemented in ggml-cpu.cpp.
+std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
+
+#endif

llama/ggml-cpu.cpp

@@ -1,11 +1,36 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-aarch64.h"
 #include "ggml-cpu-traits.h"
 #include "ggml-impl.h"
-#include "amx/amx.h"
-
+#include "amx.h"
 #include <cctype>
 #include <string>
 #include <vector>
@@ -14,10 +39,6 @@
 #include "ggml-cpu-hbm.h"
 #endif
 
-#ifdef GGML_USE_CPU_KLEIDIAI
-#include "kleidiai/kleidiai.h"
-#endif
-
 #if defined(__APPLE__)
 #include <sys/types.h>
 #include <sys/sysctl.h>
@@ -43,12 +64,6 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_KLEIDIAI
-        if (ggml_backend_cpu_kleidiai_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());
-        }
-#endif
-
 #ifdef GGML_USE_CPU_AARCH64
         if (ggml_backend_cpu_aarch64_buffer_type()) {
             bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
@@ -294,14 +309,14 @@ struct ggml_backend_cpu_device_context {
                         &hKey) == ERROR_SUCCESS) {
             DWORD cpu_brand_size = 0;
             if (RegQueryValueExA(hKey,
-                                "ProcessorNameString",
+                                TEXT("ProcessorNameString"),
                                 NULL,
                                 NULL,
                                 NULL,
                                 &cpu_brand_size) == ERROR_SUCCESS) {
                 description.resize(cpu_brand_size);
                 if (RegQueryValueExA(hKey,
-                                    "ProcessorNameString",
+                                    TEXT("ProcessorNameString"),
                                     NULL,
                                     NULL,
                                     (LPBYTE)&description[0], // NOLINT
@@ -413,21 +428,12 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
                 op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
-        case GGML_OP_SOFT_MAX_BACK: {
-            if (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type != GGML_TYPE_F32) {
-                return false;
-            }
-            float max_bias = 0.0f;
-
-            memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
-
-            return max_bias == 0.0f;
-        }
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
         case GGML_OP_IM2COL_BACK:
             return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
         case GGML_OP_OUT_PROD:
-            return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
-                src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+            return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32;
         default:
             return true;
     }
@@ -544,22 +550,19 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_dotprod()) {
             features.push_back({ "DOTPROD", "1" });
         }
+        if (ggml_cpu_has_matmul_int8()) {
+            features.push_back({ "MATMUL_INT8", "1" });
+        }
         if (ggml_cpu_get_sve_cnt() > 0) {
             static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
             features.push_back({ "SVE_CNT", sve_cnt.c_str() });
         }
-        if (ggml_cpu_has_sme()) {
-            features.push_back({ "SME", "1" });
-        }
         if (ggml_cpu_has_riscv_v()) {
             features.push_back({ "RISCV_V", "1" });
         }
         if (ggml_cpu_has_vsx()) {
             features.push_back({ "VSX", "1" });
         }
-        if (ggml_cpu_has_vxe()) {
-            features.push_back({ "VXE", "1" });
-        }
         if (ggml_cpu_has_wasm_simd()) {
             features.push_back({ "WASM_SIMD", "1" });
         }
@@ -575,9 +578,6 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
     #ifdef GGML_USE_OPENMP
         features.push_back({ "OPENMP", "1" });
     #endif
-    #ifdef GGML_USE_CPU_KLEIDIAI
-        features.push_back({ "KLEIDIAI", "1" });
-    #endif
     #ifdef GGML_USE_CPU_AARCH64
         features.push_back({ "AARCH64_REPACK", "1" });
     #endif

llama/ggml-cpu.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 #include "ggml.h"
@@ -8,7 +34,7 @@ extern "C" {
 #endif
 
     // the compute plan that needs to be prepared for ggml_graph_compute()
-    // since https://github.com/ggml-org/ggml/issues/287
+    // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
         size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
         uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
@@ -95,11 +121,9 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
     GGML_BACKEND_API int ggml_cpu_has_sve        (void);
     GGML_BACKEND_API int ggml_cpu_get_sve_cnt    (void);  // sve vector length in bytes
-    GGML_BACKEND_API int ggml_cpu_has_sme        (void);
     // other
     GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
     GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
-    GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
     GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 

llama/ggml-cuda.h

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #pragma once
 
 #include "ggml.h"

llama/ggml-cuda/acc.cu

@@ -1,3 +1,29 @@
+/**
+ * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
 #include "acc.cuh"
 
 static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,