nvidia-docker in nvidia linux repositories

[jendap]

Can you please add nvidia-docker to linux repositories?

Ubuntu ppa would be nice. Or just add it to nvidia rpm and deb repos? For example here: http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64

[3XX0]

We plan on doing it when we get out of beta (relatively soon)

[jendap]

Cool. Thanks.

[jendap]

@3XX0 are you getting out of beta?

[3XX0]

A lot of things are moving with CUDA 8 and Pascal out and we might tag RC3 before the final release.
Also, it may be wise to wait for Dockercon since it's most likely going to bring more users to the mix and help us identify any blocker before 1.0.

[jendap]

That's the beauty of having it in repo - once can run apt-get upgrade - and get your changes.

Should we recommend it in tensorflow? The hack using variables we recomend is not good. But if you change instructions every other week? That is also not great.

BTW: Have you at least talked with Docker folks if this is the right way to do it? Is this what Docker folks expect to be the main way special hardware should be wired into docker?

[Kaixhin]

I will also have to submit PRs to various repos (Theano, Lasagne, mxnet etc.) to update their docs when I move my files to use the Docker plugin. Currently I would recommend the following: docker run -itcurl -s localhost:3476/docker/clikaixhin/cuda-caffe. I can point users to this repo with respect to installing nvidia-docker, but the command to run the containers needs to be stable.

[3XX0]

The commands are stable but I would recommend using nvidia-docker rather than docker as it will perform additional version checks (see today's issue #106 for example).
The only instructions that will change is how to install it. @jendap I will try to do my best to have our packages distributed through our repo asap.

Yes we had multiple discussion with the Docker folks.
It's not perfect but there are no other ways to do what's needed at the moment.
I will probably discuss a proposal I have at Dockercon see if we can work something out. The idea would be to extend Docker to support hooks. I can already see several other applications (e.g. Security policy, Infiniband).

[Kaixhin]

@3XX0 What's the use case for the plugin then? According to the wiki, run and create are modified, but the rest of the commands are passed through. It seems that nvidia-docker only mounts the required files with the right metadata (regardless of version checking)? My plan is to move over to nvidia/cuda for base images, but for now I can work on adding them to kaixhin/cuda to ease the transition.

[3XX0]

Everything is explained in the documentation. nvidia-docker and nvidia-docker-plugin work in pair, the plugin handles the creation and mounts of the driver files.

nvidia-docker is similar to docker run -itcurl -s localhost:3476/docker/cli`` except that it checks the image/driver compatibility and enables remote deployment.

[Kaixhin]

@3XX0 Thanks for clarifying. I could work this out from the nvidia docker page, but it might also help to have the summary you just gave on the Why NVIDIA Docker page?

[jendap]

Thank you @3XX0! I will update all of our documentation to send people to nvidia-docker. Hence the repo will be nice.

[flx42]

Sorry, we're late, but we're working on it...

[jeffstokes72]

Still working on it @flx42 ?

[flx42]

@jeffstokes72: we're working on a new approach for enabling GPU support in Docker and other container technologies. We should be able to have this new package on a NVIDIA package repository.
However, for nvidia-docker, I don't think it will happen, sorry.

[3XX0]

Closing, our repositories are live!