Add pytorch/training/gpu/2.3.1/transformers/4.48.0/py311/Dockerfile

[alvarobartt]

Description

This PR bumps the dependencies to release a new PyTorch DLC for training with improvements, support for newer model architectures, bug fixes and much more.

Additionally, besides the version bumps, this PR also includes the gcloud CLI and installs huggingface_hub with the hf-transfer utility for improvements on download/upload speed to the Hugging Face Hub.

Note

This PR will enable the example on how to fine-tune PaliGemma 2 with TRL to be shipped within #133

[philschmid]

Whats the time benefit of uv? Not sure if we should add another dependency for a few seconds faster build time of containers. Especially with the alias, e.g. adding custom dependencies in training jobs, what if they are not supported etc.

[alvarobartt]

Whats the time benefit of uv? Not sure if we should add another dependency for a few seconds faster build time of containers. Especially with the alias, e.g. adding custom dependencies in training jobs, what if they are not supported etc.

So it's mainly for both the docker build time and also for installing dependencies on e.g. Kubeflow Pipelines as it will be faster, I'm expecting there's not much impact besides faster pip-installs and bigger image size; so we can roll back those changes if not desired at the current stage; whatever you feel like is better, I'm happy with either! 🤗

[alvarobartt]

Ok @philschmid after checking, apparently kfp is automatically installing the Python dependencies as python3 -m pip install ... as per https://github.com/kubeflow/pipelines/blob/2686e017ceca21671d47a6f8d5703ad94b7f0615/sdk/python/kfp/dsl/component_factory.py#L126, meaning that the current alias won't work on those scenarios i.e. when installing the packages via a Kubeflow Component for Vertex Pipelines; the current Dockerfile can be updated to work even in that case, but as you mentioned, not sure that's worth it, WDYT?

[philschmid]

Ok @philschmid after checking, apparently kfp is automatically installing the Python dependencies as python3 -m pip install ... as per https://github.com/kubeflow/pipelines/blob/2686e017ceca21671d47a6f8d5703ad94b7f0615/sdk/python/kfp/dsl/component_factory.py#L126, meaning that the current alias won't work on those scenarios i.e. when installing the packages via a Kubeflow Component for Vertex Pipelines; the current Dockerfile can be updated to work even in that case, but as you mentioned, not sure that's worth it, WDYT?

Lets remove it and maybe revisit in a few months.