Who contributed to Hugging Face on May 21, 2026?

6 developers shipped this update, including danielfleischer, danieldk, stevhliu, sayakpaul, aazizyan, and janbernloehr.

What were the notable Hugging Face updates?

Add xpu-kernels skill - Intel XPU Triton kernel development, nix-builder: remove Torch 2.10 support, and Do not load arch kernels on free-threaded Python.

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

INTEL XPU JOINS KERNEL-BUILDER, TORCH 2.10 DEPRECATED

By RepoJournal · Filed 06:04 UTC on May 21, 2026 · About Hugging Face

6 people shipped this

danieldk @danieldk 2 cited

danielfleischer @danielfleischer 1 cited

stevhliu @stevhliu 1 cited

sayakpaul @sayakpaul 1 cited

aazizyan @aazizyan 1 cited

janbernloehr @janbernloehr 1 cited

Hugging Face kernels now ships Intel XPU support via Triton, while dropping legacy Torch 2.10 from the test matrix.

The kernels team launched XPU-kernels, a new skill that packages Intel's Xe-Forge workflow into kernel-builder alongside existing CUDA and ROCm skills [1]. This targets Intel Battlemage / Arc Pro B70 hardware via the Intel XPU Backend for Triton, giving teams a third acceleration path for optimized kernel development. At the same time, nix-builder removed Torch 2.10 support [2] and updated ROCm test coverage, streamlining the CI matrix. Separately, kernels now guards against segfaults on free-threaded Python interpreters by disabling arch kernels when ABI3 isn't available [3]. Over in diffusers, the team consolidated docs by removing the pipeline examples section and migrating content to model docs [4], cutting duplication and maintenance burden. A new CI utility checks forward/call method docstrings for consistency [5]. The TRL team shipped Qwen3-VL training templates with generation markers for correct SFT masking [6], fixed tiny Qwen model config bugs [7], [8], and dropped unjustified test skips [9]. Finally, pytorch-image-models added reproducible validation with a --seed argument mirroring train.py [10], so validation runs produce consistent results.

Action items

→ Test XPU kernel builds on Battlemage hardware if you target Intel Arc huggingface/kernels [monitor]
→ Drop Torch 2.10 from any custom kernel CI pipelines huggingface/kernels [plan]
→ Update Qwen3-VL fine-tuning templates if using return_assistant_tokens_mask=True huggingface/trl [plan]
→ Add --seed to validate.py calls for reproducible benchmark runs huggingface/pytorch-image-models [monitor]

References

[1] Add xpu-kernels skill - Intel XPU Triton kernel development ↗ huggingface/kernels
[2] nix-builder: remove Torch 2.10 support ↗ huggingface/kernels
[3] Do not load arch kernels on free-threaded Python ↗ huggingface/kernels
[4] [docs] remove pipeline examples section ↗ huggingface/diffusers
[5] [WIP] chore: add utilities to check if call/forward methods are documented. ↗ huggingface/diffusers
[6] Add Qwen3-VL training chat template with generation markers ↗ huggingface/trl
[7] Fix tiny Qwen3-VL `deepstack_visual_indexes` and drop the test skip (#5779) huggingface/trl
[8] Remove non-existent params from tiny Qwen2-VL model (#5795) huggingface/trl
[9] Drop unjustified `model.visual.` skip in GRPO / RLOO Qwen2.5-VL tests (#5780) huggingface/trl
[10] validate.py: add --seed for reproducible validation ↗ huggingface/pytorch-image-models

Quick answers

What shipped in Hugging Face on May 21, 2026?: Hugging Face kernels now ships Intel XPU support via Triton, while dropping legacy Torch 2.10 from the test matrix. In total, 24 commits and 24 pull requests landed.
Who contributed to Hugging Face on May 21, 2026?: 6 developers shipped this update, including danielfleischer, danieldk, stevhliu, sayakpaul, aazizyan, and janbernloehr.
What were the notable Hugging Face updates?: Add xpu-kernels skill - Intel XPU Triton kernel development, nix-builder: remove Torch 2.10 support, and Do not load arch kernels on free-threaded Python.

TRANSFORMERS OVERHAULS LINEAR ATTENTION WHILE DEPRECATING LEGACY RESPONSE SCHEMA

The transformers library is retiring its fragile response_schema prototype in favor of streaming-compatible parsing, while simultaneously refactoring every linear attention model to use standardized convolution patterns.

python 70 shipped 2-min read

@huggingface 1 day ago

TRANSFORMERS SHIPS FSDP DISTRIBUTED TRAINING STACK, HUB LIBRARY PLUGS REDOS HOLE

Hugging Face landed distributed training orchestration in transformers while plugging a regex vulnerability that could stall untrusted card parsing for minutes.

+10

python 61 shipped 1-min read

@huggingface 2 days ago

TRANSFORMERS HARDENS AGAINST PYTORCH FRAGMENTATION WHILE TRL SIMPLIFIES DISTILLATION

Transformers plugged a cascading import failure that breaks downstream CI on older PyTorch versions, while TRL rips out dead code to lock DistillationTrainer into prompt-only datasets.

python 70 shipped 2-min read

@huggingface 3 days ago

DATASET VIEWER LOCKS DOWN ARROW, FUNES SHIPS GROUNDED ASK

Hugging Face security teams moved overnight to contain a critical Arrow IPC parsing vulnerability in dataset-viewer while shipping three production hardening releases across Repo2RLEnv, optimum-executorch, and funes.

python 91 shipped 2-min read

Elsewhere on the wire

AI Agents about 9 hours ago

CLAUDE OPUS 5 LANDS ACROSS THE STACK

The newest Anthropic model is now live in langchain, Cline, and llama-index, with native support for extended reasoning and 1M context windows.

ai-agents 28 shipped 1-min read

Local LLMs about 9 hours ago

OLLAMA LANDS LAGUNA SUPPORT AND CRUSHES MEMORY LEAKS WHILE SGLANG HITS V0.5.16 WITH CONFIDENCE-DRIVEN SPECULATIVE DECODING

Ollama shipped three critical performance and reliability fixes for Metal residency and concurrent access patterns, while SGL-Lang released 0.5.16 with a new speculative algorithm hitting 383.7 tok/s on DeepSeek-V4.