[Pytorch][GPU][Training] Initial Release

[philschmid]

As part of our collaboration with Google Cloud, we want to create dedicated first-party Hugging Face Deep Learning Containers, which make it easy and straightforward to train and deploy models on GKE and Vertex AI.

We want to create a dedicated Pytorch Training container that includes all important Hugging Face Ecosystem libraries, including transformers, datasets, evaluate, diffusers, trl, peft, sentence-transformers and common ML libraries, like torch, scikit-learn, tensorboard, bitsandbytes, deepspeed etc. to allow Google users to train models easily.

For this we have the following requirements:

• [ ] #3
• [ ] #4
• [ ] maintain a container.yaml, which defines which versions should be built. (Read the version and build the correct container based on the information see https://github.com/huggingface/Google-Cloud-Containers/blob/main/containers/container.yaml)
  • [ ] have a "script" (can be python) to build containers.
    • [ ] Automatic generated markdown table based on which contaienrs are available
    • [ ] Documentation on how to add and release new containers

The implementation should be easily extendable. We aim to maintain up-to-date, secured, and maintained versions of the Hugging Face DLC.

[alvarobartt]

Closing this issue as the PyTorch Training DLC has already been released at us-docker.pkg.dev/deeplearning-platform-release/gcr.io/huggingface-pytorch-training-cu121.2-3.transformers.4-42.ubuntu2204.py310, even though some of the tasks defined within the bullet points are missing, but those will be tackled in separate issues / PRs.