This repository provides a Docker recipe suitable for running the latest version of Julia and its CUDA GPU stack. It has been developed for use on NGC, NVIDIA's catalog of Docker containers, but can be used with plain Docker as well.
The container recipe is based on NVIDIA's Ubuntu images with CUDA pre-installed, and is currently hard-coded for use on x86_64 hosts.
$ docker build -t ngc .
$ docker run --gpus=all -it --rm ngc
Note that the --gpus=all argument requires a sufficiently recent version of
Docker, as well as the NVIDIA Container Runtime to be installed and configured.
To test CUDA.jl, it is recommended to launch Julia with multiple threads. This
can be done by passing -tauto to the container invocation, e.g., docker run --gpus=all -it --rm ngc -tauto.
To update the software used by this container, note that both Julia and CUDA.jl are pinned to specific versions.
Edit the Dockerfile and change the JULIA_RELEASE and JULIA_VERSION
arguments at the top. Then edit the Project.toml to reflect these changes.
The repository contains a Project.toml listing the packages that will be
installed, currently only CUDA.jl, and a Manifest.toml locking the specific
versions that will be used. To update this environment, use Julia on your host
system:
$ julia --project
$ pkg> update
Make sure you use the same version of Julia as will be used in the container, or incompatible package versions may be selected.