Nvidia CUDA Linux Container Image Sources

Usage of the container images require the nvidia-docker v2 runtime.

Container images are available from:

• https://ngc.nvidia.com/catalog/containers/nvidia:cuda

• https://hub.docker.com/r/nvidia/cuda

The python script is used to generate container sources and tests from templates.

1. Install poetry. See https://poetry.eustace.io/docs/#installation

2. Install the dependencies with poetry install

3. Enter the virtual environment with poetry shell

Once in the virtual environment, run the script with:

python manager.py

Pipelines can be triggered automatically or explicitly via commit message.

The pipeline mechanism used for this project within gitlab is a workround until multiple pipelines per gitlab-ci.yaml file is implemented in gitlab.

By default on the master branch, the manifest.yaml file is compared against the previous version from the last commit. The changed keys within the manifest are then used to trigger builds via the gitlab api.

Pipeline triggering can be overridden via the git commit message with:

ci.trigger = <pipeline>[,...]

Where pipeline can be:

• all: All of the pipelines are built.
• <distro><distro_version>_cuda<major>_<minor>: A specific cuda version for a specific cuda version.
• TODO: <distro><distro_version>: all cuda versions for a distro version.
• TODO: cuda<major>_<minor>: A cuda version for all distros.
• TODO: cudnn<version>: A cudnn version for all distros.

Configuration of Dockerfile and Test output for the numerous platforms and architectures supported by CUDA Docker is defined in manifest.yaml.

To define a new image:

ubi7:
  template_path: redhat
  cuda:
    repo_url: "http://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64"
    v10.0:
      build_version: 130
      cuda_requires: "brand=tesla,driver>=384,driver<385 brand=tesla,driver>=410,driver<411"
      cudnn7:
        version: "7.6.0.64"
        sha256sum: "c4e1ee4168f4cadabaa989487a47bed09f34d34e35398b6084a2699d11bd2560"

Notable keys:

• template_path: The path to search in for templates for this image. The path should be in the same directory as manager.py.

• cuda: Controls the versions of cuda supported by this image definition. It is possible to define multiple cuda versions as well as multiple cudnn versions.

• repo_url: A variable used in the container source templates.

And to generate the image:

python manager.py generate --os ubi --os-version 7 --cuda-version 10.1

The container repos that manager.py can push are defined on a global level and can be excluded on a image level.

If only_if is defined on the repo object in manifest.yaml, manager.py will only use that repo if the defined value is also present in the global environment. For example:

docker_repos:
  docker.io:
    only_if: REGISTRY_TOKEN

will result in manager.py pushing to docker.io only if the REGISTRY_TOKEN global variable is defined in the environment.

If an image root object in manifest.yaml contains the exclude_repos key, then manager.py will not push to that repo for that image. For example,

ubi7:
  base_image: registry.access.redhat.com/ubi7/ubi:latest
  exclude_repos:
    - docker.io

manager.py will not push the UBI-7 images to docker.io.