$ cat huggingface/month/2026-09-01.log
the month in review · September 2026
TRL ships 1.13 and 1.14, removing trl.losses
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
Two breaking TRL releases landed this month, alongside Transformers' deprecated-code purge and diffusers' ONNX deprecation.
v1.14.0 huggingface/trl
The fused loss module introduced in v1.13 lasted one release: FusedLinearDPOLoss, FusedLinearKTOLoss, FusedLinearGRPOLoss and FusedLinearJSDLoss are gone, and DPO, KTO and GRPO now stream their own log-probs via a chunked path. If you imported from trl.losses, that import fails on upgrade.
v1.13.0 huggingface/trl
Alongside the long-context material, this release is the version v1.14 unwinds. Read the two together before pinning: the losses module you may have adopted here is exactly what the next release removes.
Deprecated stuff gone (#48367) huggingface/transformers
The deprecated surface is gone (see also the removal of the deprecated mask functions and the position-indexed token type lookup in RoPE encoders). Code that still called those helpers breaks on upgrade rather than warning.
[core] deprecate onnx huggingface/diffusers
Diffusers marked ONNX support deprecated this month. Export and inference paths built on it now have a finite window; there is no replacement named in the announcement yet.
v0.5.0 huggingface/OpenEnv
by github-actions[bot]
OpenEnv's public surface changed here: RFC 008 manifest and validation contracts were restored, and the environment later dropped implicit port 8000 and shipped 0.6.0.
feat: support and check minver huggingface/kernels
Kernels now carries minver in its metadata so the client library can check whether the installed version satisfies a kernel's requirements at load time, rather than failing later at exec_module.
fix vlm support asyncgrpo huggingface/trl
AsyncGRPOTrainer accepts VLMs by loading through create_model_from_path, with everything outside the text tower frozen. The tradeoff is explicit in the change: no images in the dataset means the vision tower is never trained, so this does not give you multimodal training.
[Fix] Sparse TikToken tokenizers silently fail (#48446) huggingface/transformers
Sparse TikToken tokenizers that failed silently now surface, via a size-estimation fix and removal of the sentencepiece warning path. Faster to diagnose than the silent failure.
$ ls huggingface/month/ # 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.