RepoJournal
Hugging Face @huggingface
139 wires and counting

$ follow Hugging Face

Keep up with Hugging Face in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

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.

$ status

wire 2026-08-26
stories 82
GitHub RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

TRL UNBLOCKS MOE LORA AS PEFT PATCHES AUTOCAST

By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology

Albert Villanova del Moral hanouticelina github-actions[bot] jiqing-feng akshan-main moon-bot-app[bot] Yih-Dar
13 people shipped this
Albert Villanova del Moral @albertvillanova 2 cited hanouticelina @hanouticelina 1 cited github-actions[bot] @github-actions[bot] 1 cited jiqing-feng @jiqing-feng 1 cited akshan-main @akshan-main 1 cited moon-bot-app[bot] @moon-bot-app[bot] Yih-Dar @ydshieh Ferdinand Mom @3outeille stevhliu @stevhliu eustlb @eustlb tarekziade @tarekziade dg845 @dg845 dependabot @dependabot

The MoE LoRA CI failures that have been plaguing TRL for days are gone, and the fix came from upstream.

TRL reverted its xfail hotfix for the seven MoE LoRA training tests now that peft#3602 fixed the autocast dtype bug that was crashing CI with "ValueError: Registering a parametrization may not change the dtype of the tensor" [1]. The temporary exclusion comment is also gone, signaling the team trusts the upstream resolution [2]. Meanwhile, huggingface_hub shipped v1.26.1, a patch release on the heels of the new bucket visibility update that adds a `--private`/`--public` CLI switch for changing bucket settings after creation [3][4]. That bucket change is client-side plumbing only, so this is a feature to adopt, not a security emergency. Transformers is quieter but notable: a fix for CpmAnt ensures the model's lm_head is sized to the actual vocab_size, preventing a mismatch with the released checkpoint's 30720-row head, a subtle loading bug that could corrupt inference [5]. The pipeline parallel inference PR is still cooking, and the DeepFloyd IF test refactors across four pipelines are a sign of housekeeping discipline [6]. The headline verdict: patch TRL's CI, and update huggingface_hub for the new CLI capability.

Action items

References

  1. [1] Revert xfail for MoE LoRA tests now that peft#3602 fixed the autocast dtype bug ↗ huggingface/trl
  2. [2] Remove the stale detector exclusion comment (#6926) ↗ huggingface/trl
  3. [3] v1.26.1 ↗ huggingface/huggingface_hub
  4. [4] [Buckets] Add support for updating bucket visibility ↗ huggingface/huggingface_hub
  5. [5] Cpmant fix use cache (#48013) ↗ huggingface/transformers
  6. [6] refactor deepfloyd if img2img superresolution pipeline tests to the new mixin structure ↗ huggingface/diffusers

$ ls huggingface/reviews/ # the step-back reads

the week in review Diffusers 0.41.0 ships Qwen-Image 2.1, TRL 1.15 goes breaking the month in review TRL ships 1.13 and 1.14, removing trl.losses
Full archive → Weekly reviews →

Quick answers

What shipped in Hugging Face on August 26, 2026?
The MoE LoRA CI failures that have been plaguing TRL for days are gone, and the fix came from upstream. In total, 40 commits, 40 pull requests, and 2 releases landed.
Who contributed to Hugging Face on August 26, 2026?
13 developers shipped this update, including hanouticelina, moon-bot-app[bot], github-actions[bot], Yih-Dar, Ferdinand Mom, stevhliu, eustlb, and jiqing-feng, and 5 more.
What were the notable Hugging Face updates?
Revert xfail for MoE LoRA tests now that peft#3602 fixed the autocast dtype bug, Remove the stale detector exclusion comment (#6926), and v1.26.1.
♥

$ log @huggingface --recent

MoE expert sharding gets a test, lerobot swaps vendor SDKs for rustypot 59 shipped Transformers 5.20 ships Muse Spark Open, TRL adds AWS Trainium 97 shipped TRL defaults padding to "auto" on Neuron, drops vLLM 0.21.0 164 shipped AsyncGRPO packs rollouts as prefix trees, TRL types tool arguments by schema 80 shipped

$ log --all # elsewhere on the wire

Alignment & Post-training verl rejects conflicting vLLM engine_kwargs, open-instruct adds MILES OPD behind a Megatron learner Interpretability nnsight 0.8.0 cuts import time from 7.6s to 1.6s

$ ls showcases/

Safety Evals AI Red Teaming Agent Security Interpretability Alignment & Post-training

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.