Skip to content

Docker Containers with GPU access in Fedora

This guide will let you use a NVIDIA GPU in a Docker container (e.g. for Tensorflow/Pytorch).

Prerequisites:

  • Docker must be installed.
  • You have a NVIDIA GPU.

1. Install NVIDIA drivers from the RPM Fusion repository1:

sudo dnf install xorg-x11-drv-nvidia-cuda

It is necessary to reboot the system, as this package includes a script which blacklists the default Nouveau drivers.

2. Add the NVIDIA Container Runtime Repository2:

curl -s -L https://nvidia.github.io/libnvidia-container/centos7/libnvidia-container.repo | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo

3. Install the NVIDIA Container Toolkit

This includes the NVIDIA Container Runtime, which adds a custom pre-start hook to all containers to allow GPU access):

sudo dnf install nvidia-container-toolkit

4. Register the runtime with Docker:

sudo tee /etc/docker/daemon.json <<EOF
{
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}
EOF

Tell dockerd to reload configuration:

sudo pkill -SIGHUP dockerd

(Optional) To make all containers use the NVIDIA runtime by default, add the following to /etc/docker/daemon.json:

"default-runtime": "nvidia"

Alternatively, you can use nvidia-ctk:

sudo nvidia-ctk runtime configure --runtime=docker

5. Finally, test GPU access in Docker with either:

docker run --rm nvidia/cuda:10.2-base nvidia-smi

or

docker run --gpus all --rm nvcr.io/nvidia/k8s/cuda-sample:nbody nbody -benchmark -numbodies=512000

And that's it!

Note on VSCode

If you are using VSCode Remote Containers, you will need to add

"runArgs": ["--gpus","all"]

to your devcontainer.json. For some reason, --runtime=nvidia does not work.

  1. Enable the RPM Fusion repositories in Fedora here. ↩

  2. While there is no official Fedora support for nvidia-container-runtime, the Centos 7 repository seems to work. ↩

Comments