llama.cpp/.github/workflows/build-wheels-cuda.yaml

name: Build Wheels (CUDA)

on: workflow_dispatch

permissions:
  contents: write

jobs:
  define_matrix:
    name: Define Build Matrix
    runs-on: ubuntu-latest
    outputs:
      matrix: ${{ steps.set-matrix.outputs.matrix }}
    defaults:
      run:
        shell: pwsh

    steps:
      - name: Define Job Output
        id: set-matrix
        run: |
          $matrix = @{
              'os' = @('ubuntu-20.04', 'windows-latest')
              'pyver' = @("3.10", "3.11", "3.12")
              'cuda' = @("12.1.1", "12.2.2", "12.3.2")
              'releasetag' = @("basic")
          }

          $matrixOut = ConvertTo-Json $matrix -Compress
          Write-Output ('matrix=' + $matrixOut) >> $env:GITHUB_OUTPUT          

  build_wheels:
    name: Build Wheel ${{ matrix.os }} ${{ matrix.pyver }} ${{ matrix.cuda }} ${{ matrix.releasetag == 'wheels' && 'AVX2' || matrix.releasetag }}
    needs: define_matrix
    runs-on: ${{ matrix.os }}
    strategy:
      matrix: ${{ fromJSON(needs.define_matrix.outputs.matrix) }}
    defaults:
      run:
        shell: pwsh
    env:
      CUDAVER: ${{ matrix.cuda }}
      AVXVER: ${{ matrix.releasetag }}

    steps:
      - uses: actions/checkout@v4
        with:
          submodules: "recursive"

      - uses: actions/setup-python@v4
        with:
          python-version: ${{ matrix.pyver }}

      - name: Setup Mamba
        uses: conda-incubator/setup-miniconda@v3.0.4
        with:
          activate-environment: "build"
          python-version: ${{ matrix.pyver }}
          miniforge-variant: Mambaforge
          miniforge-version: latest
          use-mamba: true
          add-pip-as-python-dependency: true
          auto-activate-base: false

      - name: VS Integration Cache
        id: vs-integration-cache
        if: runner.os == 'Windows'
        uses: actions/cache@v4.0.2
        with:
          path: ./MSBuildExtensions
          key: cuda-${{ matrix.cuda }}-vs-integration

      - name: Get Visual Studio Integration
        if: runner.os == 'Windows' && steps.vs-integration-cache.outputs.cache-hit != 'true'
        run: |
          if ($env:CUDAVER -eq '12.1.1') {$x = '12.1.0'} else {$x = $env:CUDAVER}
          $links = (Invoke-RestMethod 'https://github.com/Jimver/cuda-toolkit/raw/dc0ca7bb29c5a92f7a963d3d5c93f8d59765136a/src/links/windows-links.ts').Trim().split().where({$_ -ne ''})
          for ($i=$q=0;$i -lt $links.count -and $q -lt 2;$i++) {if ($links[$i] -eq "'$x',") {$q++}}
          Invoke-RestMethod $links[$i].Trim("'") -OutFile 'cudainstaller.zip'
          & 'C:\Program Files\7-Zip\7z.exe' e cudainstaller.zip -oMSBuildExtensions -r *\MSBuildExtensions\* > $null
          Remove-Item 'cudainstaller.zip'          

      - name: Install Visual Studio Integration
        if: runner.os == 'Windows'
        run: |
          $y = (gi '.\MSBuildExtensions').fullname + '\*'
          (gi 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise\MSBuild\Microsoft\VC\*\BuildCustomizations').fullname.foreach({cp $y $_})
          $cupath = 'CUDA_PATH_V' + $env:CUDAVER.Remove($env:CUDAVER.LastIndexOf('.')).Replace('.','_')
          echo "$cupath=$env:CONDA_PREFIX" >> $env:GITHUB_ENV          

      - name: Install Dependencies
        env:
          MAMBA_DOWNLOAD_FAILFAST: "0"
          MAMBA_NO_LOW_SPEED_LIMIT: "1"
        run: |
          $cudaVersion = $env:CUDAVER
          mamba install -y 'cuda' -c nvidia/label/cuda-$cudaVersion
          python -m pip install build wheel          

      - name: Build Wheel
        run: |
          $cudaVersion = $env:CUDAVER.Remove($env:CUDAVER.LastIndexOf('.')).Replace('.','')
          $env:CUDA_PATH = $env:CONDA_PREFIX
          $env:CUDA_HOME = $env:CONDA_PREFIX
          $env:CUDA_TOOLKIT_ROOT_DIR = $env:CONDA_PREFIX
          if ($IsLinux) {
            $env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH
          }
          $env:VERBOSE = '1'
          $env:CMAKE_ARGS = '-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all'
          $env:CMAKE_ARGS = "-DLLAMA_CUDA_FORCE_MMQ=ON $env:CMAKE_ARGS"
          if ($env:AVXVER -eq 'AVX') {
            $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off'
          }
          if ($env:AVXVER -eq 'AVX512') {
            $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX512=on'
          }
          if ($env:AVXVER -eq 'basic') {
            $env:CMAKE_ARGS = $env:CMAKE_ARGS + ' -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_FMA=off -DLLAMA_F16C=off'
          }
          python -m build --wheel
          # write the build tag to the output
          Write-Output "CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV          

      - uses: softprops/action-gh-release@v1
        with:
          files: dist/*
          # Set tag_name to <tag>-cu<cuda_version>
          tag_name: ${{ github.ref_name }}-cu${{ env.CUDA_VERSION }}
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}