Run nested docker containers (containers that run inside a docker-in-docker container) with access to the NVIDIA GPUs of the host system.
This docker image is based on the NVIDIA CUDA image, but installs docker and its required runtime deps too. This can be run easilily with docker, assuming you already have the container toolkit installed and setup on the docker engine host.
To run the image, the following command can be run:
docker run --rm -it --privileged --gpus all \
--name nvdind ghcr.io/willfantom/nv-dind:<TAG>
Tags can be found in the relevant GitHub Package for this project and a named by
the Docker version and CUDA version used for the image. For example, if using
Docker v26.0 and CUDA 12.5.0, the tag would be v26.0-12.5.0.