cuda-toolkit

This action installs the NVIDIA® CUDA® Toolkit on the system. It adds the cuda install location as CUDA_PATH to GITHUB_ENV so you can access the CUDA install location in subsequent steps. CUDA_PATH/bin is added to GITHUB_PATH so you can use commands such as nvcc directly in subsequent steps. Right now both windows-2019 and ubuntu-20.04 runners have been tested to work successfully.

Inputs

cuda

Optional The CUDA version to install. View src/link/windows-links.ts and src/link/linux-links.ts for available versions.

Default: '11.8.0'.

cudnn, cudnn_url and cudnn_archive_dir

Optional The cudnn version, the URL to the cudnn archive file (.zip file for Windows and .tar.xz file for Linux) and the name of the unarchived directory.

A user provided URL is needed as the official cudnn archive would require a valid login to download. Therefore, we need to follow these steps:

1. Download the offical archive file from NVIDIA, cudnn-download.
2. Serve the archive file where (only) you can access.
3. Set cudnn, cudnn_url and cudnn_archive_dir.

For example, to use cudnn 8.7.0 on Windows, you'll need to download https://developer.nvidia.com/downloads/c118-cudnn-windows-8664-87084cuda11-archivezip from NVIDIA website. Note that this requires you to have an NVIDIA developer account.

Since we are using cudnn 8.7.0, we need to set cudnn to "8.7.0".

Once you suceesfully downloaded the file, and if you unarchive that file, there will be a directory with the same name as the zip file (excluding the .zip file extension):

.
├── cudnn-windows-x86_64-8.7.0.84_cuda11-archive.zip
└── cudnn-windows-x86_64-8.7.0.84_cuda11-archive
    ├── LICENSE
    ├── bin
    │   └── *.dll
    ├── include
    │   └── *.h
    └── lib
        └── x64
            └── *.lib

The dir name, cudnn-windows-x86_64-8.7.0.84_cuda11-archive, should be the value of cudnn_archive_dir. Note that we only need to serve the .zip file; the unarchiving operation here is to show the expected structure and the name of the unarchived directory.

Lastly, once you uploaded this .zip file to your server, and let' suppose it can be accessed at https://example.com/cudnn-windows-x86_64-8.7.0.84_cuda11-archive.zip, then cudnn_url should be set to "https://example.com/cudnn-windows-x86_64-8.7.0.84_cuda11-archive.zip".

Of course, you can store the URL as a secret in GitHub Actions, and use the corresponding secret in your workflow file.

sub-packages

NOTE: On Linux this only works with the 'network' method view details

Optional If set, only the specified CUDA subpackages will be installed. Only installs specified subpackages, must be in the form of a JSON array. For example, if you only want to install nvcc and visual studio integration: '["nvcc", "visual_studio_integration"]' (double quotes required)

Default: '[]'.

method

Optional Installation method, can be either 'local' or 'network'.

• 'local' downloads the entire installer with all packages and runs that (you can still only install certain packages with sub-packages on Windows).
• 'network' downloads a smaller executable which only downloads necessary packages which you can define in sub-packages.

Default: 'local'.

linux-local-args

Optional (For Linux and 'local' method only) override arguments for the Linux .run installer. For example if you don't want samples use '["--toolkit"]' (double quotes required) See the Nvidia Docs for available options. Note that the --silent option is already always added by the action itself.

Default: '["--toolkit", "--samples"]'.

Outputs

cuda

The cuda version installed (same as cuda from input).

CUDA_PATH

The path where cuda is installed (same as CUDA_PATH in GITHUB_ENV).

Example usage

steps:
- uses: Jimver/[email protected]
  id: cuda-toolkit
  with:
    cuda: '11.7.0'

- run: echo "Installed cuda version is: ${{steps.cuda-toolkit.outputs.cuda}}"

- run: echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"

- run: nvcc -V