Who contributed to Hugging Face on July 17, 2026?

8 developers shipped this update, including tomaarsen, Quentin Gallouédec, albertvillanova, Andrey Churkin, Sayak Paul, atharvajoshi10, yiyixuxu, and PreethamNoelP.

What were the notable Hugging Face updates?

[`losses`] Consolidate GradCache into one shared engine, fix silently wrong gradients, add `mini_batch_num_tokens`, Check causal left padding against the attention mask, and Improve the error message when `video_metadata` is passed as a sibling modality key (#3866).

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

SENTENCE-TRANSFORMERS FIXES SILENT GRADIENT CORRUPTION ON GPU

By RepoJournal · Filed 06:59 UTC on July 17, 2026 · About Hugging Face

8 people shipped this

albertvillanova @albertvillanova 4 cited

tomaarsen @tomaarsen 3 cited

Andrey Churkin @andrey-churkin 1 cited

atharvajoshi10 @atharvajoshi10 1 cited

PreethamNoelP @PreethamNoelP 1 cited

Quentin Gallouédec @qgallouedec

Sayak Paul @sayakpaul

yiyixuxu @yiyixuxu

A critical bug in cached losses was silently producing wrong gradients on GPU, and it's fixed now [ref:1].

Sentence-transformers consolidated its fragmented GradCache machinery into a single shared engine and patched a silent gradient corruption bug that affected cross-encoder cached losses on GPU and prompt-training with `Pooling(include_prompt=False)` across all cached losses [1]. This is the kind of bug that ships bad models without warning. In related fixes, the library corrected causal padding validation that was reading the wrong attribute on multimodal processors, causing VLM padding checks to never fire [2], and improved error messages for video metadata passed as sibling keys instead of nested under the video modality [3]. TRL is systematically hardening its experimental trainers with comprehensive eval_dataset initialization tests across GRPO-family [4], Online-DPO-family [5], SFT and distillation [6], and preference trainers [7], mirroring coverage already added to core trainers. Diffusers added Cosmos 3 Edge generator support [9] and began shipping Anima img2img pipeline blocks [10], extending modular generation capabilities. Optimum-Intel patched a missing compression parameter that should have been exposed [8].

Action items

→ Update sentence-transformers immediately if using cached losses, cross-encoders, or prompt training on GPU huggingface/sentence-transformers [immediate]
→ Review TRL trainer eval_dataset handling against your use case now that comprehensive tests are shipping huggingface/trl [plan]
→ Monitor sentence-transformers releases for any downstream effects from the GradCache consolidation huggingface/sentence-transformers [monitor]

References

[1] [`losses`] Consolidate GradCache into one shared engine, fix silently wrong gradients, add `mini_batch_num_tokens` ↗ huggingface/sentence-transformers
[2] Check causal left padding against the attention mask huggingface/sentence-transformers
[3] Improve the error message when `video_metadata` is passed as a sibling modality key (#3866) huggingface/sentence-transformers
[4] Add eval_dataset init tests for experimental GRPO-family trainers ↗ huggingface/trl
[5] Add eval_dataset init tests for experimental Online-DPO-family trainers ↗ huggingface/trl
[6] Add eval_dataset init tests for experimental SFT and distillation trainers ↗ huggingface/trl
[7] Add eval_dataset init tests for experimental preference trainers ↗ huggingface/trl
[8] Add missing compression parameter ↗ huggingface/optimum-intel
[9] Cosmos3 edge support ↗ huggingface/diffusers
[10] [Anima] Add img2img pipeline blocks ↗ huggingface/diffusers

Quick answers

What shipped in Hugging Face on July 17, 2026?: A critical bug in cached losses was silently producing wrong gradients on GPU, and it's fixed now [ref:1]. In total, 31 commits and 29 pull requests landed.
Who contributed to Hugging Face on July 17, 2026?: 8 developers shipped this update, including tomaarsen, Quentin Gallouédec, albertvillanova, Andrey Churkin, Sayak Paul, atharvajoshi10, yiyixuxu, and PreethamNoelP.
What were the notable Hugging Face updates?: [`losses`] Consolidate GradCache into one shared engine, fix silently wrong gradients, add `mini_batch_num_tokens`, Check causal left padding against the attention mask, and Improve the error message when `video_metadata` is passed as a sibling modality key (#3866).

TRANSFORMERS 5.14 SHIPS WITH INKLING MULTIMODAL MODEL

Hugging Face shipped transformers v5.14.0 with native support for Inkling, a 975B parameter multimodal model that processes text, image, and audio in a single forward pass.

python 53 shipped 1-min read

@huggingface 2 days ago

TRANSFORMERS PATCHES TORCH.COMPILE CRASH; KERNELS SHIP LAYER REGISTRY UPDATES

Transformers just fixed a half-precision torch.compile crash that broke DETR models in v5.9.0, while the kernel ecosystem shipped three layer-repo updates that unify the custom kernel interface across rotary, liger, and rope implementations.

python 39 shipped 2-min read

@huggingface 3 days ago

TRANSFORMERS ADDS MTP DECODING, TRL FIXES DEEPSEEP ZERO-3 CRASH

Transformers finally ships proper multi-token prediction support [ref:3], while TRL patches a critical DeepSpeed ZeRO-3 incompatibility that breaks Liger kernel fusion in production training.

python 46 shipped 1-min read

@huggingface 4 days ago

TRANSFORMERS CUTS DISTRIBUTED LOAD TIME IN HALF WITH SHARD-ON-READ

Transformers shipped native DTensor shard-on-read support, letting each rank load its own checkpoint slice instead of materializing the full model on every device first.