$ cat huggingface/week/2026-10-26.log
the week in review · Oct 26 – Nov 1, 2026
Transformers 5.22.0 ships Mistral Large 4
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
TRL 1.17.0 can fine-tune the 1T-parameter model on a single node, and Transformers 5.22 also drops torch.jit.script.
v5.22.0: Mistral Large 4 (Le Chonk) huggingface/transformers
The release adds Mistral's largest model to date, a 1 trillion-parameter natively multimodal model, alongside the removal of torch.jit.script and a changed attention path on Mac serve. If you script models or serve on Apple hardware, test before bumping.
v1.17.0 huggingface/trl
TRL gains support for mistralai/Mistral-Large-4-1T-A52B, enough to fine-tune it on one node. The same release removes trl.experimental.tpo entirely: TPOTrainer, TPOConfig, its docs and its example are gone, and the release notes note nobody asked to keep it after the proposal on 2026-10-13.
Composite transparency in shared RGB conversion and image processors huggingface/transformers
PNG transparency now composites correctly through the shared RGB conversion path, so tRNS chunks survive preprocessing instead of flattening. Any pipeline that feeds transparent images into a processor gets the right pixel values.
Merge commit from fork huggingface/diffusers
Diffusers stops importing modules named by a model repository: a component's library name from model_index.json no longer reaches importlib.import_module, and requirement names from modular_config.json no longer go through find_spec. Loading an untrusted repo is a smaller attack surface.
Fix shared LoRA component weight broadcasting huggingface/diffusers
When set_adapters() receives one component-weight dictionary for several LoRA adapters, only the first adapter got the requested weight. This fix makes the rest honor it, which matters if you weight adapters per component rather than loading them one at a time.
Fix chunked loss precision huggingface/trl
The chunked log-probability and KL/JSD heads now handle the two precision issues described in the paper the fix cites. Confident targets keep their small gradients and matching teachers produce zero gradients, so losses stay correct at the edges of the distribution.
Fix/fused logprob entropy strided huggingface/trl
The fused log-prob entropy kernel stored strides during forward, which broke when SFT and DPO passed shifted views like logits[:, :-1] and input_ids[:, 1:]. Those trainers can now use selective_log_softmax again.
Asyncgrpo score centering huggingface/trl
AsyncGRPO's ratio mixed sampler staleness with the server-trainer numerical mismatch, and the clipped objective collapsed with an FP8 vLLM sampler. Score centering stabilizes it, and a separate change drops rollout groups whose advantages are all zero before they reach the trainer. Alongside: liger-kernel leaves the dev dependencies and Docker images with use_liger_kernel=True deprecated, PEFT ...
$ ls huggingface/week/ # the briefings behind this review
Keep up with Hugging Face in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.