DinD with nvidia-docker

Use nvidia-docker inside a container.

Requirements

nvidia-docker is required on the host machine.

Building image

You can either pull the image henderake/dind:nvidia-docker or build the image by yourself with the following command:

  $ nvidia-docker build -t dind:nvidia-docker .

CUDA version and Docker version and be specified with --build-arg:

  $ nvidia-docker build -t dind:nvidia-docker --build-arg CUDA_IMAGE=nvidia/cuda:9.0-runtime --build-arg DOCKER_CE_VERSION=5:18.09.1~3-0~ubuntu-xenial .

Running the dind:nvidia-docker container

The usage of the container is the same as the official dind image, except that you have to run it in a nvidia runtime.

Here is an example. First, run the container on the host machine.

  $ DIND=$(docker run --privileged -d --runtime=nvidia dind:nvidia-docker)
  $ docker exec -it $DIND /bin/bash

Now we have a shell in the dind container. Inside this container, you can run any container that requires nvidia runtime.

You can also connect a second container to dind:nvidia-docker. Refer to official dind image to read the steps.

Acknowledgement

The dind part of this Dockerfile is copied from https://github.com/docker-library/docker/tree/master/18.09/dind