Who contributed to Hugging Face on May 25, 2026?

1 developer shipped this update, including JoelNiklaus.

What were the notable Hugging Face updates?

Fix caching allocator warmup byte estimation for EP model loading (#46149), Add Pillow dependency, and Update Inference Providers documentation (automated) (#2498).

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

EXPERT PARALLEL MEMORY BUG FIXED, LIGHTEVAL GAINS IMAGE SUPPORT

By RepoJournal · Filed 06:03 UTC on May 25, 2026 · About Hugging Face

1 person shipped this

JoelNiklaus @JoelNiklaus 1 cited

Transformers shipped a critical fix for distributed model loading that was overestimating device memory in expert parallel runs, while lighteval patched a silent failure in task initialization.

The transformers team fixed a memory allocation bug [1] where the caching allocator's warmup phase was reading the wrong tensor parallelism plan and trying to preallocate as if expert weights weren't sharded. This could blow past device memory limits on multi-GPU setups running expert parallel inference. In a separate but equally quiet fix, lighteval addressed a ModuleNotFoundError [2] where the built-in MathVista task was importing PIL without declaring Pillow as a dependency. That meant running lighteval accelerate would fail during task registry initialization for any user without Pillow already installed. The hub-docs team also shipped automated Inference Providers documentation updates [3] reflecting the latest @huggingface/tasks and @huggingface/inference package versions. Dev version bumps landed across transformers as preparation for the next release.

Action items

→ Review distributed model loading if you run expert parallel inference - upgrade transformers for the warmup fix huggingface/transformers [plan]
→ Upgrade lighteval if you use task evaluation - Pillow dependency now declared huggingface/lighteval [plan]

References

[1] Fix caching allocator warmup byte estimation for EP model loading (#46149) huggingface/transformers
[2] Add Pillow dependency ↗ huggingface/lighteval
[3] Update Inference Providers documentation (automated) (#2498) huggingface/hub-docs

Quick answers

What shipped in Hugging Face on May 25, 2026?: Transformers shipped a critical fix for distributed model loading that was overestimating device memory in expert parallel runs, while lighteval patched a silent failure in task initialization. In total, 4 commits and 4 pull requests landed.
Who contributed to Hugging Face on May 25, 2026?: 1 developer shipped this update, including JoelNiklaus.
What were the notable Hugging Face updates?: Fix caching allocator warmup byte estimation for EP model loading (#46149), Add Pillow dependency, and Update Inference Providers documentation (automated) (#2498).

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.