Transformers kernel and model addition proposals

Hugging Face Transformers saw several standalone proposals for kernels and model support. The highest-profile item seeks contributors for low-precision MoE training work.

MXFP4 MoE and attention backward kernels

A proposal in the Hugging Face Transformers repository calls for contributors to add MXFP4 MoE and attention backward kernels. The stated goal is to enable native low-precision training of GPT-OSS. Developers tracking efficient mixture-of-experts training have a concrete contribution target.

BART-LS model support

An opening issue requests addition of BART-LS model support through the modular guide. No discussion or checkpoints are present yet. The request remains at the earliest stage with no implementation artifacts.

Tokenizer compatibility with Auto classes

A feature request asks for tokenizer.Tokenizer compatibility with the Inference API and Auto classes. The proposal stands alone without further traffic. Users of the custom tokenizers library may watch for any follow-through.

Google Trillsson audio classification model

A single opening message requests addition of Google's Trillsson audio classification model via the modular guide. No further discussion is recorded. Audio model maintainers following Transformers coverage of Google releases have a new open request to track.