Update PyTorch Inference + Training & Text Embeddings Inference (TEI) DLCs

[alvarobartt]

Description

This PR updates the Deep Learning Containers (DLCs) for PyTorch Training, PyTorch Inference and Text Embeddings Inference (TEI), as part of the renewed partnership between Hugging Face and Google Cloud.

It mainly upgrades dependencies in these containers to leverage the latest libraries, for example:

• PyTorch Training now uses the latest versions of transformers, accelerate, kernels, etc., including the newest trl for training and fine-tuning LLMs and VLMs.
• PyTorch Inference runs with a new engine (still shipped as a private wheel for now; we plan to publish it on PyPI soon) and the latest released versions of transformers, sentence-transformers, and diffusers, benefiting from new model architectures and inference optimizations.
• Text Embeddings Inference (TEI) now supports google/embeddinggemma-300m, along with many other architectures, improvements, and bug fixes.

These are just some of the improvements in the new DLCs; additional details will follow in upcoming documentation and content on how to best leverage them.

Warning

Note that the DLCs might not be released yet on Google Cloud (as of late Nov 2025), but once those are released we'll update the README.md and the examples/ to point and benefit from the latest DLCs!