runner: Remove CGO engines, use llama-server exclusively for GGML models

Remove the vendored GGML and llama.cpp backend, CGO runner, Go model implementations, and sample. llama-server (built from upstream llama.cpp via FetchContent) is now the sole inference engine for GGUF-based models. (Safetensor based models continue to run on the new MLX engine.) This allows us to more rapidly pick up new capabilities and fixes from llama.cpp as they come out. On windows this now requires recent AMD driver versions to support ROCm v7 as llama.cpp currently does not support building against v6.
2026-04-23 09:15:44 +02:00 · 2026-03-25 16:59:18 -07:00
parent ff23dd343f
commit 56c735d871
902 changed files with 5888 additions and 423367 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -24,7 +24,7 @@ jobs:
        run: |
          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${GITHUB_REF_NAME#v}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" | tee -a $GITHUB_OUTPUT
          echo VERSION="${GITHUB_REF_NAME#v}" | tee -a $GITHUB_OUTPUT
-          echo vendorsha=$(make -f Makefile.sync print-base) | tee -a $GITHUB_OUTPUT
+          echo vendorsha=$(cat LLAMA_CPP_VERSION) | tee -a $GITHUB_OUTPUT

  darwin-build:
    runs-on: macos-26-xlarge
@@ -57,7 +57,7 @@ jobs:
          go-version-file: go.mod
          cache-dependency-path: |
            go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
      - run: |
          ./scripts/build_darwin.sh
      - name: Log build results
@@ -106,10 +106,10 @@ jobs:
            flags: ''
          - os: windows
            arch: amd64
-            preset: 'ROCm 6'
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
-            rocm-version: '6.2'
-            flags: '-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
+            preset: 'ROCm 7'
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-26.Q1-Win11-For-HIP.exe
+            rocm-version: '7.1'
+            flags: '-DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma"'
            runner_dir: 'rocm'
          - os: windows
            arch: amd64
@@ -232,14 +232,66 @@ jobs:
        with:
          path: ${{ github.workspace }}\.ccache
          key: ccache-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.preset }}-${{ needs.setup-environment.outputs.vendorsha }}
-      - name: Build target "${{ matrix.preset }}"
+      - name: Build MLX
+        if: startsWith(matrix.preset, 'MLX ')
        run: |
          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
          cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
-          cmake --build --parallel ([Environment]::ProcessorCount) --preset "${{ matrix.preset }}"
-          cmake --install build --component "${{ startsWith(matrix.preset, 'MLX ') && 'MLX' || startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || startsWith(matrix.preset, 'Vulkan') && 'Vulkan' || 'CPU' }}" --strip
-          Remove-Item -Path dist\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
+          cmake --build --preset "${{ matrix.preset }}" -- -l $([Environment]::ProcessorCount)
+          cmake --install build --component MLX --strip
+      - name: Build llama-server
+        if: matrix.preset == 'CPU'
+        run: |
+          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          cmake -S llama\server --preset cpu --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
+          cmake --build build\llama-server-cpu -- -l $([Environment]::ProcessorCount)
+          cmake --install build\llama-server-cpu --component llama-server --strip
+        env:
+          CMAKE_GENERATOR: Ninja
+      - name: Build llama-server (CUDA)
+        if: startsWith(matrix.preset, 'CUDA ')
+        run: |
+          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          $cudaVer = "${{ matrix.preset }}".Replace('CUDA ', '').Replace(' ', '')
+          $preset = "cuda-v$cudaVer"
+          if ($cudaVer -eq "13") { $preset = "cuda-v13-windows" }
+          cmake -S llama\server --preset $preset ${{ matrix.flags }} --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
+          cmake --build "build\llama-server-$preset" -- -l $([Environment]::ProcessorCount)
+          cmake --install "build\llama-server-$preset" --component llama-server --strip
+        env:
+          CMAKE_GENERATOR: Ninja
+      - name: Build llama-server (ROCm)
+        if: startsWith(matrix.preset, 'ROCm ')
+        run: |
+          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          $hipPath = "${{ env.HIP_PATH }}"
+          $env:HIPCXX = "$hipPath\bin\clang++.exe"
+          $env:HIP_PLATFORM = "amd"
+          $env:CMAKE_PREFIX_PATH = "$hipPath"
+          $env:CC = "$hipPath\bin\clang.exe"
+          $env:CXX = "$hipPath\bin\clang++.exe"
+          cmake -S llama\server --preset rocm -G Ninja `
+              -DCMAKE_HIP_FLAGS="-parallel-jobs=4" `
+              -DCMAKE_C_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" `
+              -DCMAKE_CXX_FLAGS="-parallel-jobs=4 -Wno-ignored-attributes -Wno-deprecated-pragma" `
+              -DAMDGPU_TARGETS="gfx942;gfx950;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1200;gfx1201;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-" `
+              --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
+          cmake --build build\llama-server-rocm -- -l $([Environment]::ProcessorCount)
+          cmake --install build\llama-server-rocm --component llama-server --strip
+        env:
+          CMAKE_GENERATOR: Ninja
+      - name: Build llama-server (Vulkan)
+        if: matrix.preset == 'Vulkan'
+        run: |
+          Import-Module 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
+          cmake -S llama\server --preset vulkan --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
+          cmake --build build\llama-server-vulkan -- -l $([Environment]::ProcessorCount)
+          cmake --install build\llama-server-vulkan --component llama-server --strip
        env:
          CMAKE_GENERATOR: Ninja
      - name: Log build results
@@ -297,7 +349,7 @@ jobs:
          go-version-file: go.mod
          cache-dependency-path: |
            go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
      - name: Verify gcc is actually clang
        run: |
          $ErrorActionPreference='Continue'
@@ -353,7 +405,7 @@ jobs:
          go-version-file: go.mod
          cache-dependency-path: |
            go.sum
-            Makefile.sync
+            LLAMA_CPP_VERSION
      - uses: actions/download-artifact@v4
        with:
          pattern: depends-windows*
@@ -421,6 +473,7 @@ jobs:
            case "$COMPONENT" in
              bin/ollama*)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/llama-server*|lib/ollama/llama-quantize*)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/vulkan*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
              lib/ollama/mlx*)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
@@ -439,7 +492,7 @@ jobs:
          done
      - run: |
          for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd --ultra -22 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst);
+            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd -9 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst);
          done
      - uses: actions/upload-artifact@v4
        with: