setup-cuda-deps

GitHub Action to install GPU Python packages using gpkg.

Resolves prebuilt CUDA wheels from the gpkg registry and community sources. Optionally builds from source when no prebuilt wheel exists.

Usage

- uses: mapika/setup-cuda-deps@v1
  with:
    packages: flash-attn causal-conv1d mamba-ssm
    torch: '2.11.0'
    cuda: '130'

Inputs

Input	Required	Default	Description
packages	Yes		Space-separated list of packages
torch	No	auto-detect	PyTorch version
cuda	No	auto-detect	CUDA version
python	No	auto-detect	Python version
build-missing	No	false	Build from source when no prebuilt wheel exists
gpkg-version	No	latest	Specific gpkg version to install

Examples

Basic — specify versions

jobs:
  train:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: '3.12'
      - uses: mapika/setup-cuda-deps@v1
        with:
          packages: flash-attn causal-conv1d mamba-ssm
          torch: '2.11.0'
          cuda: '130'
      - run: uv sync

Auto-detect from environment

      - uses: mapika/setup-cuda-deps@v1
        with:
          packages: flash-attn causal-conv1d

Build missing wheels from source

      - uses: mapika/setup-cuda-deps@v1
        with:
          packages: flash-attn causal-conv1d
          torch: '2.11.0'
          cuda: '128'
          build-missing: 'true'

What it does

1. Installs gpkg via pip
2. Resolves prebuilt CUDA wheels for your packages
3. Writes pyproject.toml with direct wheel URLs
4. Writes gpkg.lock.toml for reproducibility
5. Optionally builds from source if build-missing: true

License

MIT