cuda-docker-generator

cuda-docker-generator is a tool to generate Dockerfile that installs CUDA / cuDNN on arbitrary base Docker image.

Installation steps are automatically downloaded from assets in nvidia/cuda repository. Images can be run using nvidia-docker.

Usage

The following is an example to generate Dockerfile (and cuda.repo) that installs development components of CUDA 9.1 for CentOS 7 and cuDNN 7 on Fedora 27 Docker image.

$ ./generate.py --os centos7 --cuda 9.1 --cudnn 7 --variant devel --base fedora:27

By default, privilege used to run the image is reset to root. You can override this behavior by specifying --user option.

$ ./generate.py --os ubuntu16.04 --cuda 9.0 --cudnn 7 --base jupyter/datascience-notebook --user jovyan

See ./generate.py --help for the detailed usage.

Notes

• If you specify --base image containing different operating system than one specified in --os, CUDA may not work properly.
• If you specify invalid combination (e.g., CUDA 9.1 with cuDNN 6), you may see HTTP Error 404: Not Found error. --verbose option may help diagnosing such issue.