DockerFile for PyTorch?

[jzf2101]

@parente : @koustuvsinha are looking at jupyterhub/zero-to-jupyterhub-k8s#994 and @consideRatio 's dockerfile and we need to tweak both to get a pytorch image perhaps we can also add this when it's done?

[parente]

Hi @jzf2101 . We've been trying to build, test, and maintain fewer image in this one uber-repo over time and encouraging more links to images based on the ones here managed elsewhere. Ways to go about it (ordered by my preference FWIW) would be:

1. Start a new repo / image based on the info here: https://jupyter-docker-stacks.readthedocs.io/en/latest/contributing/stacks.html
2. Repurpose the jupyter/tensorflow-notebook to be jupyter/deeplearning-notebook and add it there.
3. Update the automation here and on Docker Cloud to build and maintain a new jupyter/pytorch-notebook image in this repo.

[jzf2101]

I'm fine with option 1 as long as we presume this doesn't imply we're providing preferential support for TF over Pytorch. Both seem quite popular

[consideRatio]

@jzf2101 I added two bullet points to the issue i posted previously:

• Read this to understand why we pin cudatoolkit, and what NVIDIA drivers are required for various versions of CUDA: https://docs.anaconda.com/anaconda/user-guide/tasks/gpu-packages/
• The following project to build deeplearning images also contains a lot of suggestions of deeplearning packages to include, but by using this you don't get a conda installed package that will have better performance, nor do you get the docker-stacks specific things for jupyterhub like fix-permissions etc: https://github.com/ufoym/deepo#comparison-to-alternatives

@parente thanks for the input on this!

---

@jzf2101 I could help out on option 1!

[parente]

I'm fine with option 1 as long as we presume this doesn't imply we're providing preferential support for TF over Pytorch. Both seem quite popular

Agreed. The tensorflow-notebook image predates our push to include fewer images in this single repo. Perhaps we should consider moving it to its own repo too. It's of limited use because it is a CPU-only image.

On that topic: A handful of users have requested that we update it (tensorflow-notebook) to include the CUDA drivers. After consulting with the Jupyter Steering Council, we decided that we did not want to introduce commercial software into the images we maintain. See #516 for a brief history. I don't know how that affects plans for PyTorch.

[consideRatio]

When you write conda install tensorflow-gpu, CUDA is installed automatically as a dependency through cudatoolkit I think. I manually pinned it while installing tensorflow-gpu though to avoid getting the latest version as that is coupled with a too new CUDA version that requires more modern graphics cards than I were utilizing on google cloud (K80, P100).

Does this mean the CUDA licence has changed in some way, as it is available simple enough with a conda install command? If not, I'm wondering if we are keeping track of all kinds of things installed as dependencies through conda install?

I supressed confirmations etc during install with --quiet --yes during the conda install though, so I may have accepted something without knowing it. Hmmm...

[parente]

Does this mean the CUDA licence has changed in some way, as it is available simple enough with a conda install command?

It does look like the EULA has changed since the last time I looked: https://docs.nvidia.com/cuda/eula/index.html#distribution-requirements

IANAL so I won't attempt to interpret the changes. That said, the original decision was about not including non-open source software in the images here, not the details of the CUDA license in particular.

If not, I'm wondering if we are keeping track of all kinds of things installed as dependencies through conda install?

We purposely installed tensorflow but not tensorflow-gpu to avoid pulling the CUDA lib when we originally set up the image definition.

https://github.com/jupyter/docker-stacks/blob/master/tensorflow-notebook/Dockerfile#L10

But that begs the valid question of what else might we be pulling in.

[ericdill]

@consideRatio @parente -- Here is my understanding from our open source team at Anaconda regarding CUDA, nvidia and redistribution.

• EULA lists basically all the shared libraries as redistributable, but not compiler, debugger, or profiler
• cuDNN is not part of CUDA toolkit and is not technically redistributable
• NVIDIA has an exception clause for Docker images based on their images / Dockerfiles. It's not totally unambiguous what "based on" means, but the rough idea seems to be that you can start from their docker images, layer in new things and still be able to redistribute them.

There's also a nice discussion over on conda-forge here

[jzf2101]

@alexbw followed up on this issue looking at the underlying conda build for PyTorch conda-forge/conda-forge.github.io#63 (comment) @scopatz are you saying that PyTorch likely has a license that allows people to use these conda packages to install CUDA with PyTorch wherever?

[scopatz]

Nope, users are not allowed to install cudatoolkit anywhere that is redistributed

[beniz]

See here for a Pytorch GPU recipe: #706 (comment)

The Tensorflow + Pytorch recipe is here: https://github.com/jolibrain/docker-stacks/tree/master/jupyter-dd-notebook-gpu

[parente]

This issue has been idle for some time now so I'm going to close it. If anyone would like to submit a PR to the recipes or community images doc pages linking to information about GPU-enabled images, I'll be happy to review it. For now, the decision stands that we do not wish to deal with the distribution of non-open source software in these images.

I'm not sure if it's been said yet, so I'll point out that it might be much easier to direct users to images like https://hub.docker.com/r/pytorch/pytorch and https://hub.docker.com/r/tensorflow/tensorflow/ and help users add Jupyter components to them instead of working the other way around (i.e., starting with Jupyter images and adding deep learning frameworks with GPU acceleration to them). I'd be fine with the recipes page including such instructions if someone wants to figure out that approach and send a PR.