Docker recipes for Julia builds with JuliaGPU packages.
Deprecation notice: With Julia 0.7/1.0, installability of the GPU stack has improved, and this image isn't worth maintaining anymore.
For a list of included packaged, refer to the REQUIRE file.
- Docker
- NVIDIA drivers on your host system
- nvidia-docker
Pull and start the container as follows:
$ docker pull maleadt/juliagpu
$ docker run --runtime=nvidia -v $PKG:/pkg -it maleadt/juliagpu
Replace $PKG with a named volume (see docker volume create), or the path to
a local directory. When the container starts, and /pkg does not contain an
initialized package installation, a pre-configured installation of the packages
as listed in the REQUIRE file will be copied to the package directory
and configured for your system. This will take a while, but the results will be
persistent across invocations (on the condition you keep your package volume or
directory intact).
The container starts in /data, which you can mount in order to access files
from your host.
Note that the container has Julia as entry point, and thus can be used as if it were a
regular binary (especially powerful in combination with $PWD mounted as /data):
$ alias juliagpu='docker run --runtime=nvidia -v $PKG:/pkg -v $PWD:/data -it maleadt/juliagpu'
$ juliagpu -e 'println("Hello, World!")'
Hello, World!
$ echo 'println("Hello, World!")' > test.jl
$ juliagpu test.jl
Hello, World!
- Did you run the container with
--runtime=nvidia? - Some packages might support a debugging mode; run again with
DEBUG=trueand file an issue.
Members of the JuliaGPU organization can force a rebuild of the image at the JuliaGPU CI. Apart from that, the image is rebuilt weekly, including new versions of Julia and any included packages.
If you want to include a new package, create a PR modifying the REQUIRE file. Note that all included packages need to pass tests, or the image won't be pushed.
Manual build instructions:
$ docker build -t maleadt/juliagpu .
$ docker push maleadt/juliagpu