flash-attention-prebuild-wheels-rocm/.github/workflows/test-windows-build.yml

name: "Test Windows build"
on:
  workflow_dispatch:

jobs:
  build_wheels:
    name: Build wheels and Upload
    runs-on: windows-latest
    timeout-minutes: 1000
    strategy:
      fail-fast: false
      matrix:
        flash-attn-version: ["2.6.3"]
        python-version: ["3.11"]
        torch-version: ["2.6.0"]
        # https://developer.nvidia.com/cuda-toolkit-archive
        cuda-version: ["12.6.3"]
        exclude:
          # torch < 2.2 does not support Python 3.12
          - python-version: "3.12"
            torch-version: "2.0.1"
          - python-version: "3.12"
            torch-version: "2.1.2"
          # torch 2.0.1 does not support CUDA 12.x
          - torch-version: "2.0.1"
            cuda-version: "12.1.1"
          - torch-version: "2.0.1"
            cuda-version: "12.4.1"
          - torch-version: "2.0.1"
            cuda-version: "12.6.3"
          - torch-version: "2.0.1"
            cuda-version: "12.8.1"
          # torch 2.7.0 does not support CUDA 12.4
          - torch-version: "2.7.0"
            cuda-version: "12.4.1"
    steps:
      - uses: actions/checkout@v4

      - name: Enable Git long paths
        run: git config --system core.longpaths true
        shell: pwsh

      - name: Install VS2022 BuildTools 17.9.7
        run: choco install -y visualstudio2022buildtools --version=117.9.7.0 --params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64 --installChannelUri https://aka.ms/vs/17/release/180911598_-255012421/channel"
        shell: pwsh

      - uses: actions/setup-python@v5
        with:
          python-version: ${{ matrix.python-version }}

      - uses: Jimver/cuda-toolkit@master
        with:
          cuda: ${{ matrix.cuda-version }}
          method: "network"
          sub-packages: '["nvcc", "cudart", "cublas", "cufft", "curand", "cusolver", "cusparse", "visual_studio_integration"]'

      - name: Set CUDA and PyTorch versions
        run: |
          $cudaVersion = "${{ matrix.cuda-version }}" -replace '\.', ''
          $cudaVersion = $cudaVersion.Substring(0, [Math]::Min(3, $cudaVersion.Length))
          echo "MATRIX_CUDA_VERSION=$cudaVersion" >> $env:GITHUB_ENV
          $torchVersion = "${{ matrix.torch-version }}" -replace '^(\d+\.\d+).*', '$1'
          echo "MATRIX_TORCH_VERSION=$torchVersion" >> $env:GITHUB_ENV
        shell: pwsh

      - name: Install build dependencies
        run: |
          choco install -y ninja
          pip install -U pip setuptools==75.8.0 wheel setuptools packaging psutil
        shell: pwsh

      - name: Install PyTorch ${{ matrix.torch-version }}+cu${{ matrix.cuda-version }}
        run: |
          $env:TORCH_CUDA_VERSION = python -c "from os import environ as env; support_cuda_versions = { '2.0': [117, 118], '2.1': [118, 121], '2.2': [118, 121], '2.3': [118, 121], '2.4': [118, 121, 124], '2.5': [118, 121, 124], '2.6': [118, 124, 126], '2.7': [118, 126, 128], '2.8': [128], }; target_cuda_versions = support_cuda_versions[env['MATRIX_TORCH_VERSION']]; cuda_version = int(env['MATRIX_CUDA_VERSION']); closest_version = min(target_cuda_versions, key=lambda x: abs(x - cuda_version)); print(closest_version)"

          if ("${{ matrix.torch-version }}" -like "*dev*") {
            pip install --pre torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu$env:TORCH_CUDA_VERSION
          } else {
            pip install --no-cache-dir torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/cu$env:TORCH_CUDA_VERSION
          }

          nvcc --version
          python -V
          python -c "import torch; print('PyTorch:', torch.__version__)"
          python -c "import torch; print('CUDA:', torch.version.cuda)"
          python -c "from torch.utils import cpp_extension; print(cpp_extension.CUDA_HOME)"
        shell: pwsh

      - name: Checkout flash-attn
        run: |
          git clone https://github.com/Dao-AILab/flash-attention.git -b "v${{ matrix.flash-attn-version }}"
        shell: pwsh

      - name: Build wheels
        timeout-minutes: 800
        run: |
          Import-Module 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -VsInstallPath 'C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools' -DevCmdArguments '-arch=x64 -host_arch=x64'
          $env:DISTUTILS_USE_SDK = 1
          $env:BUILD_TARGET = "cuda"

          $env:CUDA_HOME = $env:CUDA_PATH
          $env:CUDA_ROOT = $env:CUDA_PATH
          $env:PATH = "$env:CUDA_PATH\bin;$env:PATH"
          $env:MAX_JOBS = "2"
          $env:NVCC_THREADS = "2"
          $env:FLASH_ATTENTION_FORCE_BUILD = "TRUE"

          cd flash-attention
          python setup.py bdist_wheel --dist-dir=dist
          $baseWheelName = Get-ChildItem -Path "dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
          $wheelName = $baseWheelName -replace "${{ matrix.flash-attn-version }}", "${{ matrix.flash-attn-version }}+cu$env:MATRIX_CUDA_VERSION" + "torch$env:MATRIX_TORCH_VERSION" + "-win"
          Move-Item "dist\$baseWheelName" "dist\$wheelName"
          echo "wheel_name=$wheelName" >> $env:GITHUB_ENV
        shell: pwsh

      - name: Install Test
        run: |
          pip install flash-attention/dist/$env:wheel_name
          python -c "import flash_attn; print(flash_attn.__version__)"
        shell: pwsh