RepoJournal
Hugging Face

@huggingface

Transformers, Datasets, and the open AI-model layer

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

One email a day. Unsubscribe in one click.

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

TRAINING GOES PADDING-FREE WHILE DIFFUSERS EXPANDS TO SOUND AND PIXEL SPACE

By RepoJournal · Filed · About Hugging Face

AsyncGRPO just ditched padding overhead entirely, diffusers shipped pixel-space PRX and audio synthesis to Cosmos3, and LeRobot fixed a critical parquet row-group bug that was tanking episode fetch performance.

The training stack got leaner overnight. TRL's AsyncGRPOTrainer [1] now runs padding-free by default, unpacking each rank's padded slice into concatenated sequences and letting FlashAttention derive `cu_seq_lens` from position resets. This eliminates the memory tax of padded forward passes without changing the collator, so your existing dataloaders stay compatible. On the diffusers front, PRXPixelPipeline [2] adds pixel-space denoising (raw RGB, no VAE) conditioned on Qwen3-VL, while Cosmos3 [3] gained sound encoder support, expanding into multimodal generation territory. LeRobot fixed a nasty data layer bug [4] where aggregate and annotation re-write paths were collapsing multiple episodes into single row groups, tanking performance when fetching individual episodes. The fix enforces one row group per episode so reads hit a single byte range instead of scanning the whole shard. Security-wise, TRL [5] deleted the pr_style_bot workflow entirely after it proved repeatedly exploitable, consolidating style enforcement into the existing check_code_quality pipeline. transformers.js [6] refactored to the new singular COS API, switching from the deprecated plural `requestFileHandles()` to `requestFileHandle()`, and huggingface.js extended download metrics [7] to include `.zst` JAX weights in IntelliFold queries.

One email a day. Unsubscribe in one click.

Action items

References

  1. [1] Padding-free training in AsyncGRPO ↗ huggingface/trl
  2. [2] Add PRXPixelPipeline: pixel-space PRX text-to-image pipeline (#13928) huggingface/diffusers
  3. [3] Add Sound Encoder to Cosmos3 (#13911) huggingface/diffusers
  4. [4] fix(datasets): enforce one parquet row group per episode in v3 data writes (#3807) huggingface/lerobot
  5. [5] Delete CI pr_style_bot workflow ↗ huggingface/trl
  6. [6] refactor(cos): use singular requestFileHandle() instead of requestFileHandles() (#1709) huggingface/transformers.js
  7. [7] Update countDownloads filter for intellifold to include 'zst' extension (#2229) huggingface/huggingface.js

Quick answers

What shipped in Hugging Face on June 17, 2026?
AsyncGRPO just ditched padding overhead entirely, diffusers shipped pixel-space PRX and audio synthesis to Cosmos3, and LeRobot fixed a critical parquet row-group bug that was tanking episode fetch performance. In total, 23 commits, 20 pull requests, and 1 releases landed.
Who contributed to Hugging Face on June 17, 2026?
2 developers shipped this update, including qgallouedec and albertvillanova.
What were the notable Hugging Face updates?
Padding-free training in AsyncGRPO, Add PRXPixelPipeline: pixel-space PRX text-to-image pipeline (#13928), and Add Sound Encoder to Cosmos3 (#13911).

More from @huggingface

Daily updates, in your inbox

Follow Hugging Face

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

or

One email a day. Unsubscribe in one click. Read a past issue →

Elsewhere on the wire

Want every project, not just this one?

We use privacy-friendly analytics (Google Analytics, IP-anonymized) to see which pages help readers. No ads, and we never sell your data. See our Privacy Policy.