Who contributed to Hugging Face on August 25, 2026?

9 developers shipped this update, including qgallouedec, Yuekai Zhang, jiqing-feng, sergereview[bot], mowoe, danieldk, Sayak Paul, and Wang, Yi, and 1 more.

What were the notable Hugging Face updates?

Compute MoE load-balancing loss per layer to avoid giant one-hot materialization −99.7% @ 128k, Fix nemotron_h save_pretrained emitting singular backbone.embedding.weight (#48075), and [DistillationTrainer refactor] Support tool calling.

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

Transformers MoE loss cut 99.7% as TRL gains tool calling

By RepoJournal · Filed 06:59 UTC on August 25, 2026 · About Hugging Face

9 people shipped this

qgallouedec @qgallouedec 4 cited

Sayak Paul @sayakpaul 2 cited

Yuekai Zhang @yuekaizhang 1 cited

jiqing-feng @jiqing-feng

sergereview[bot] @sergereview[bot]

mowoe @mowoe

danieldk @danieldk

Wang, Yi @sywangyi

albertvillanova @albertvillanova

A single PR eliminates a 128k-token memory bomb in MoE training while TRL's distillation trainer learns to call tools.

Transformers slashed MoE load-balancing loss memory by 99.7% at 128k tokens by computing per-layer instead of materializing a giant one-hot tensor [1]. That fix lands alongside a nemotron_h save_pretrained correction that restores the plural embedding key for all published checkpoints [2]. TRL's DistillationTrainer now supports tool calling, mirroring GRPO with masked tool-result tokens and a configurable iteration cap [3]. The same author reorganized TRL examples into per-task folders, untangling scripts from notebooks [4]. Kernels-community ported SageAttention3 Blackwell FP4 attention, tested on RTX-PRO-6000, bringing microscaling NVFP4 to sm_120 [5]. But watch the kernels telemetry doc fix: a typo in the title hides a real improvement in env var documentation [6]. The SFT collator now emits position_ids for sequence parallelism, fixing a crash on the first batch in ALST/Ulysses setups [7]. That's the one to test before you trust it at scale.

Action items

→ Cherry-pick the MoE load-balancing loss fix before your next long-context training run huggingface/transformers [immediate]
→ Review and update nemotron_h conversion scripts to expect plural embedding keys huggingface/transformers [plan]
→ Test DistillationTrainer tool-calling on your own workflow huggingface/trl [monitor]
→ Run SFT with sequence parallelism after the collator fix huggingface/trl [monitor]

References

[1] Compute MoE load-balancing loss per layer to avoid giant one-hot materialization −99.7% @ 128k ↗ huggingface/transformers
[2] Fix nemotron_h save_pretrained emitting singular backbone.embedding.weight (#48075) huggingface/transformers
[3] [DistillationTrainer refactor] Support tool calling ↗ huggingface/trl
[4] Reorganize examples into self-contained per-example folders ↗ huggingface/trl
[5] sage-blackwell: port SageAttention3 Blackwell FP4 attention ↗ huggingface/kernels-community
[6] chore: document telemetry env var beter. ↗ huggingface/kernels
[7] Emit position_ids from the SFT collator when sequence parallelism is enabled ↗ huggingface/trl

Quick answers

What shipped in Hugging Face on August 25, 2026?: A single PR eliminates a 128k-token memory bomb in MoE training while TRL's distillation trainer learns to call tools. In total, 31 commits and 31 pull requests landed.
Who contributed to Hugging Face on August 25, 2026?: 9 developers shipped this update, including qgallouedec, Yuekai Zhang, jiqing-feng, sergereview[bot], mowoe, danieldk, Sayak Paul, and Wang, Yi, and 1 more.
What were the notable Hugging Face updates?: Compute MoE load-balancing loss per layer to avoid giant one-hot materialization −99.7% @ 128k, Fix nemotron_h save_pretrained emitting singular backbone.embedding.weight (#48075), and [DistillationTrainer refactor] Support tool calling.

Transformers linter hunts nn.Buffer misuse as kernels ships hotfix

transformers-mlinter is tightening model internals while kernels drops a quick patch, but the real action is a quiet alignment fix in trl.

python 7 shipped

@huggingface Aug 23

Harbor-hf sinks broken Sandbox cleanup

Harbor-hf shipped five fixes overnight that close the gaps where stuck Sandboxes, terminal jobs, and invalid campaign results could quietly block recovery.

python 38 shipped

@huggingface Aug 22

LEROBOT CUTS FASTWAM OVERHEAD WITH COMPILED INFERENCE

LeRobot's FastWAM policy gets an opt-in compiled path that slashes Python and CUDA kernel-launch overhead on cached action inference.

python 36 shipped

@huggingface Aug 21

Transformers Brings Biology to v5 Main

Hugging Face's flagship repo hauled two protein-folding models onto the v5 mainline, and a wav2vec2 attention fix is ready to unblock your SDPA workflows.