Official Library Image in Docker Hub

[ruffsl]

It'd be great to get started on the review process to get an official cuda image up in Docker Hub for researchers and developers to build off from. If this is inline with what the maintainers here have in mind, I'd suggest forks of the docker-library/official-images and docker-library/docs be made on the Nvidia's github org so that further PRs can be made.

Truth be told, I'd like to see some official image of a few other projects, like Tensorflow, however official images can only derive from other official images, and I think it makes sense that cuda warrants it's own official image, rather that roping in these dependencies in other project Dockerfiles elsewhere.

[flx42]

We are preparing some improvements to our solution, but we can start preparing for an official Docker Hub image, and apparently you have the knowledge to help us :)
I would still like to have more feedback from the community before we actually commit to an official image, did you test our images with complex CUDA applications?

[ruffsl]

Hey @flx42 ,
I've been testing the setup some more using the cuda samples,
and so checkout this libGL issue I've made here: ruffsl#1

Also, pinging @tleyden and @Kaixhin 🙏,
I imagine you both would like to provide some feedback.

[3XX0]

OpenGL is not supported for now, we have ongoing discussion internally to enable it though.

[elezar]

I was able to use the CUDA images to build a GPU-enabled Caffe (direct from Berkley). Will still test with something like DIGITS ...

[flx42]

That's great to hear. I've been compiling and using Caffe with our solution too.
We are planning on doing a DIGITS image soon.

[flx42]

Update: we pushed the images to the Docker Hub under the NVIDIA organization:
https://hub.docker.com/r/nvidia/cuda/
I'm keeping this bug open to track the progress of our work towards the submission of an official (reviewed by Docker maintainers) image for CUDA.

[Yangqing]

May I ask why cudnn3 is not on the official dockerhub list yet? It seems that it is already present in the repo, but not synced to dockerhub.

[flx42]

We don't have automated builds yet (see #18). The images for cudnn3 are now on the official dockerhub.

[ruffsl]

Just noticing that because of #18 the images up at nvidia/cuda on Docker Hub have become quite stale (over a month old), yet most of the cuda docker files have become at the same time stable, aside from the acasonal version bump for cudnn here or there.

I'd recommend we start a PR to docker-library/official-images to just get some palemanary feedback from the library devs. We already know official repos are not subject to nesting limitation of automated builds, as such should help us make sure cuda images stay fresh and triggered user build's up to date for security as well.

[flx42]

Thank you @ruffsl! I haven't forgotten, I actually wanted to regenerate the images today but I realized it's not the right day to do so. I'm waiting for the new Ubuntu images, after the glibc security patch (CVE-2015-7547):
tianon/docker-brew-ubuntu-core#41
docker-library/official-images#1448

tianon said there was a glitch when regenerating the new Ubuntu images, so I should be able to push the new images tomorrow.

Regarding the official image, we still have this in mind and we would love to start the process soon. Unfortunately we are very busy on another project right now and we had to postpone this process, temporarily.

[ruffsl]

👌 Any idea when might be better? A week or a month, or two?

[flx42]

New images are online! Fresh and up-to-date!

It's difficult to give you a timeline for the official image. I hope we will be able to start the process in ~1 month.

[ruffsl]

@flx42 , would before or after the nvidia-docker beta concludes be a good time to resume this endeavor?

[3XX0]

We have ongoing discussions with Docker regarding this matter. nvidia-docker is really peculiar, therefore we are considering another approach than the official docker-library. We can't really disclose anything yet but hopefully it will change soon.

[3XX0]

Closing since we now have automated builds on Gitlab.
We will only use this moving forward as it gives us more flexibility.