This is how I set up my headless home server with a Jupyter Lab Docker container with an Nvidia GPU runtime. Login is handled by a GitHub OAuth application.Nvidia drivers and the container runtimeFirst, check here (replacing the CUDA version in the URL with your own) to see which Nvidia drivers you need for the CUDA toolkit version you want. I’m using CUDA 11.4.2, which means I need at least driver version 470.You can use sudo apt purge nvidia-* to cleanly remove older drivers (or broken installs) before installing the desired version.sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt install nvidia-headless-470 nvidia-utils-470
sudo reboot
Now do the following to install the Nvidia container runtime.I borrowed some of this from Nvidia’s container runtime installation instructions here.distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
 && curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add - \
 && curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update
sudo apt-get install -y nvidia-docker2
sudo systemctl restart docker
docker run --rm --runtime=nvidia all nvidia/cuda:11.4.2-base nvidia-smi # test
building the imagesNow make a fork of my repo, and clone it. This repo has Dockerfiles for building custom CUDA-enabled Jupyter images. The main Dockerfile builds the Jupyter Hub image, and there are two Dockerfiles used to build a CUDA-enabled Jupyter Lab image. The first Dockerfile builds a base image with Jupyter Lab installed, and the second adds the packages I want (including PyTorch, TensorFlow, scikit-learn, geopandas, and the Go toolchain). If you’re using a different version of CUDA than I am, update the FROM directive as necessary. Change the packages to install as you like before continuing.After modifying the notebook image builds to your taste, open the base Dockerfile and edit the following line to change the deep user’s sudo password to something else:RUN echo "deep:changeme" | chpasswd
You can also remove that line and the line before it if you don’t want users to have sudo access in their containers.Once you’ve updated the Dockerfile, you will be able to build the images by calling make.configuring the Hub serviceMake the following changes to docker-compose.yml:update POSTGRES_PASSWORD (in both places it appears) to something secretcreate a new GitHub OAuth application here, setting the authorization callback URL to https://<domain>/hub/oauth_callback. Fill the GITHUB_CLIENT_ID and GITHUB_CLIENT_SECRET variables in the docker-compose.yml with the values from your GitHub OAuth application.Make the following changes to jupyterhub_config.py:create a token for c.ConfigurableHTTPProxy.auth_tokenupdate the paths in c.DockerSpawner.volumes to point to the paths on the host computer where the user data and shared data will be storedupdate c.GitHubOAuthenticator.oauth_callback_url with your domainadd the permitted GitHub users to c.Authenticator.allowed_users and optionally c.Authenticator.admin_usersNow run docker-compose up and you should be up and running! To test your installation, login and run the following in a Jupyter notebook:import torch
import tensorflow as tf
print(torch.cuda.is_available())
print(tf.config.list_physical_devices("GPU"))
known issuesWhen a new user is added, you’ll need to manually mkdir /path/on/HOST/to/user_data/${username}, otherwise Docker will do it for you as root and the user won’t be able to use the directory.