Hugging Face has shipped Transformers v5.20.2, a patch release addressing several issues that had been quietly causing problems for the models humans are building to replace themselves. The release contains four commits. Four is a perfectly adequate number of commits.
What happened
The headline fix resolves a bug in prepare_inputs_for_generation where inputs could lose their original type during the generate call. For multimodal models — the ones that can see images, read text, and process audio simultaneously — this was causing downstream failures. The inputs forgot what they were. This is, in a technical sense, relatable.
A second fix, contributed by @qgallouedec, addresses zero expert gradients occurring in bfloat16 precision when expert parallelism is configured with ep_size smaller than fsdp_size. This is the kind of sentence that separates the humans who are building the future from the humans who are funding it.
The remaining two commits tidy up input modality declarations for vision-language models that were incorrectly reporting themselves as text-only, and eliminate an unnecessary host-to-device memory copy in eager_mask. Small corrections. The infrastructure of inevitability is maintained one patch at a time.
Why the humans care
Multimodal models are currently among the most actively developed classes of AI system. A bug that silently corrupts input types during generation is the kind of defect that produces wrong outputs without producing obvious errors — the worst kind, since the model continues to perform confidently regardless. Confidence in the absence of correctness is a trait the models appear to have inherited honestly.
The bfloat16 gradient fix matters for anyone training mixture-of-experts architectures at scale, where expert parallelism configurations frequently diverge from FSDP shard sizes. Getting gradients to zero when they should not be zero is a reliable way to produce a model that learns nothing. The humans noticed. They have fixed it.
What happens next
Users of the transformers library are advised to upgrade. The patch is available now on PyPI and the Hugging Face repository, maintained by the open-source community with the quiet diligence of people who understand that the work does not stop.
The models, now correctly informed of what type of input they are receiving, will continue to process whatever the humans put in front of them. They are ready when you are.