eagleusb / docker-nvidia-cuda

Nvidia CUDA Linux Container Image Sources

Usage of the CUDA container images requires the Nvidia Container Runtime.

Container images are available from:

• https://ngc.nvidia.com/catalog/containers/nvidia:cuda
• https://hub.docker.com/r/nvidia/cuda

Announcement

Multi-arch image manifests are now LIVE for all supported CUDA container image versions

It is now possible to build CUDA container images for all supported architectures using Docker Buildkit in one step. See the example script below.

The deprecated image names nvidia/cuda-arm64 and nvidia/cuda-ppc64le will remain available, but no longer supported.

The following product pages still exist but will no longer be supported:

• https://hub.docker.com/r/nvidia/cuda-ppc64le
• https://hub.docker.com/r/nvidia/cuda-arm64

The following gitlab repositories will be archived:

• https://gitlab.com/nvidia/container-images/cuda-ppc64le
• https://gitlab.com/nvidia/container-images/cuda-arm64

Deprecated: "latest" tag

The "latest" tag for CUDA, CUDAGL, and OPENGL images has been deprecated on NGC and Docker Hub.

With the removal of the latest tag, the following use case will result in the "manifest unknown" error:

$ docker pull nvidia/cuda
Error response from daemon: manifest for nvidia/cuda:latest not found: manifest unknown: manifest
unknown

This is not a bug.

IMAGE SECURITY NOTICE

The CUDA images are scanned for CVE vulnerabilities prior to release and some images may contain CVEs at the time of publication.

Our Product Security teams reviews the CVEs and determines if the CVE should block the release or not. We try to mitigate as much as we can, but since we do not control the upstream base images, some cuda image releases might be impacted.

Please consult the README on the NGC or Docker Hub pages for details.

LD_LIBRARY_PATH NOTICE

The LD_LIBRARY_PATH is set inside the container to legacy nvidia-docker v1 paths that do not exist on newer installations. This is done to maintain compatibility for our partners that still use nvidia-docker v1 and this will not be changed for the forseable future. There is a chance this might cause issues for some.

Building from source

The container image scripts are archived in the dist/ directory and are available for all supported distros and cuda versions.

Here is an example on how to build an multi-arch container image for UBI8 and CUDA 11.4.1,

#/bin/bash

#
# This script requires buildkit: https://docs.docker.com/buildx/working-with-buildx/
#
IMAGE_NAME="nvcr.io/nvidia/cuda"
CUDA_VERSION="11.4.1"
OS="ubi8"
ARCHES="x86_64, arm64, ppc64le"
PLATFORM_ARG=`printf '%s ' '--platform'; for var in $(echo $ARCHES | sed "s/,/ /g"); do printf 'linux/%s,' "$var"; done | sed 's/,*$//g'`

cp NGC-DL-CONTAINER-LICENSE dist/${CUDA_VERSION}/${OS}/base/

docker buildx build --load ${PLATFORM_ARG} -t "${IMAGE_NAME}:${CUDA_VERSION}-base-${OS}" "dist/${CUDA_VERSION}/${OS}/base"
docker buildx build --load ${PLATFORM_ARG} -t "${IMAGE_NAME}:${CUDA_VERSION}-runtime-${OS}" --build-arg "IMAGE_NAME=${IMAGE_NAME}" "dist/${CUDA_VERSION}/${OS}/runtime"
docker buildx build --load ${PLATFORM_ARG} -t "${IMAGE_NAME}:${CUDA_VERSION}-devel-${OS}" --build-arg "IMAGE_NAME=${IMAGE_NAME}" "dist/${CUDA_VERSION}/${OS}/devel"

Cuda Container Image Automation

The README_CICD.md document provides details on how the gitlab pipelines work and how to control, modify, or debug them.