Create Dockerfiles for Accelerate

[muellerzr]

This PR adds two Dockerfiles for building on the CPU and the GPU. I chose to avoid writing one for Deepspeed for now.

Eventually these will be integrated into test runners and built nightly, similar to how transformers is setup.

This uses a staged process to reduce the size of the images a bit.

Uncompressed sizes of each docker container:

CPU: ~871mb
GPU: ~13gb

For perspective, the uncompressed size of the transformers docker image for torch is 11.2gb

The biggest difference is the inclusion of conda, to make it easier for us to switch between python versions when needed.

It's also recommended to use buildkit to build the images, as it reduces the time by a healthy chunk. E.g.:
sudo DOCKER_BUILDKIT=1 docker build .-t accelerate-cpu -f docker/accelerate-cpu/Dockerfile
(Note: times were ran cache-less)
GPU image:
5m56s -> 4m16s
CPU Image:
3m20s -> 2m44s

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

I know nothing about Docker, so asking Lysandre for a review :-)

[pacman100]

Great to have docker support! Helpful for production scenarios. Few comments:

1. The Python versions in CPU and GPU are different - 3.6 and 3.8, respectively. Would it be better to have same Python versions? Also, transformers will be discontinuing 3.6 support, so CPU version can be bumped up to match the GPU version?
2. Shouldn't the Dockerfile's be in the top-level directory to be able to copy accelerate codebase and install it? Maybe have Dockerfile_cpu and Dockerfile_gpu in top-level directory?
3. Getting below error while building GPU image :

W: GPG error: https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64  InRelease: The following signatures couldn't be verified because the public key is not available: NO_PUBKEY A4B469963BF863CC
E: The repository 'https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64  InRelease' is not signed.

[muellerzr]

Great to have docker support! Helpful for production scenarios. Few comments:

These will also be used for having GPU and multi GPU tests, these mimic the transformers docker images for this exact purpose. (This is actually the true purpose behind this PR)

1. The Python versions in CPU and GPU are different - 3.6 and 3.8, respectively. Would it be better to have same Python versions? Also, transformers will be discontinuing 3.6 support, so CPU version can be bumped up to match the GPU version?

Yes they are, it's a partial limitation of the GPU image. I could go through the effort of having it use conda instead, but Sylvain and I discussed that 3.6 support will be dropped on the next release. Since it's so soon, it's easier to just have it be this way.

2. Shouldn't the Dockerfile's be in the top-level directory to be able to copy accelerate codebase and install it? Maybe have `Dockerfile_cpu` and `Dockerfile_gpu` in top-level directory?

Nope, these assume the git repo is the top level directory and then copy it. So when building the dockerfiles it looks something like (assuming from a fresh accelerate clone and cd'd to it) docker build . -f docker/accelerate-cpu/Dockerfile. See the transformers repo for a similar behavior

3. Getting below error while building GPU image :

W: GPG error: https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64  InRelease: The following signatures couldn't be verified because the public key is not available: NO_PUBKEY A4B469963BF863CC
E: The repository 'https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64  InRelease' is not signed.

Unsure with this one. Do you have docker cuda properly setup? IIRC you might need to find the right keys to use as well. I was able to build it just fine yesterday.

[muellerzr]

Thinking on it more, having the python version be configurable in the cuda image would be nice. Will change this

Also would probably be better to do similar to this Dockerfile and clone the repo instead: https://github.com/huggingface/transformers/blob/main/docker/transformers-pytorch-gpu/Dockerfile#L11

[LysandreJik]

Looks good to me!