CI/CD for a CUDA/C++ Project with GitHub Actions

Compile CUDA kernels on a CPU runner and gate device tests on GPUs.

A CUDA/C++ project compiles kernels with nvcc but only needs a GPU to run them. This recipe installs the toolkit, builds with CMake on a CPU runner, and runs device tests on a self-hosted GPU job.

What the pipeline does

install the CUDA toolkit
configure with CMake and CUDA architectures
compile kernels with nvcc on a CPU runner
run host-side tests
run device tests on a self-hosted GPU runner

The workflow

Compilation does not need a GPU, so the build runs on a standard runner. Set CMAKE_CUDA_ARCHITECTURES to your target SM versions; device tests run on a labeled self-hosted GPU runner.

.github/workflows/ci.yml

name: CI
on:
  push:
    branches: [main]
  pull_request:
jobs:
  build:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: Jimver/cuda-toolkit@v0.2.19
        with:
          cuda: '12.4.0'
      - run: cmake -B build -DCMAKE_CUDA_ARCHITECTURES='80;86'
      - run: cmake --build build -j
      - run: ctest --test-dir build -L host
      - uses: actions/upload-artifact@v4
        with:
          name: cuda-build
          path: build

Caching and speed

Cache the CUDA toolkit install and use ccache (cache ~/.ccache) to skip recompiling unchanged translation units across multi-arch builds. nvcc multi-arch compiles are slow; cheaper managed runners such as Latchkey (around 69% cheaper than GitHub-hosted) keep these long builds affordable and auto-retry transient toolkit download failures.

Deploying

CPU runners verify the build and host-side tests; add a self-hosted GPU runner job for device tests. Package the compiled binaries or a CUDA-enabled container image and deploy to a GPU host or Kubernetes GPU node pool.

Key takeaways

nvcc compiles on a CPU runner; only device tests need a GPU.
Set CMAKE_CUDA_ARCHITECTURES to your target SM versions.
Use ccache to speed multi-arch kernel compiles.