Who contributed to Hugging Face on July 23, 2026?

9 developers shipped this update, including dacorvo, tarekziade, vbayanag, eustlb, xhg, Yuanyuan Chen, Eric Bezzam, and qgallouedec, and 1 more.

What were the notable Hugging Face updates?

fix: guard DTensor import in sharding_utils.py for PyTorch < 2.5 (#47481), CPU can incur a slow path on non-contiguous magnitudes (#47351), and [fix] fix requirements audio feature and proc.

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

TRANSFORMERS HARDENS AGAINST PYTORCH FRAGMENTATION WHILE TRL SIMPLIFIES DISTILLATION

Q: What shipped in Hugging Face on July 23, 2026?

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. In total, 41 commits and 29 pull requests landed.

By RepoJournal · Filed 07:00 UTC on July 23, 2026 · About Hugging Face

9 people shipped this

qgallouedec @qgallouedec 2 cited

dacorvo @dacorvo 1 cited

tarekziade @tarekziade 1 cited

vbayanag @vbayanag 1 cited

eustlb @eustlb 1 cited

xhg @gygdh-001 1 cited

Yuanyuan Chen @cyyever

Eric Bezzam @ebezzam

Albert Villanova del Moral @albertvillanova

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.

In brief

transformers guards DTensor imports for PyTorch < 2.5. one more version gate in a codebase that's getting fragile across the PyTorch matrix.
transformers patches a slow path on CPU for non-contiguous tensors. small fix, but the kind that kills latency in production without anyone noticing.
trl's DistillationTrainer ditches messages-format and prompt-length config. cleaner API, fewer footguns. distillation pipelines just got less gnarly.
funes hardens remote memory setup. infrastructure churn nobody sees until it breaks. this ship probably stops a few outages.

The DTensor import guard [1] lands as a critical fix: sharding_utils.py was importing DTensor unconditionally, causing ImportError cascades through core_model_loading.py and downstream projects like Accelerate when PyTorch drops below 2.5. This one unblocks CI across the ecosystem. On the performance front, Whisper feature extraction got 8x faster by forcing mel magnitudes contiguous [2], sidestepping a pathological strided-GEMM slowdown on CPU backends that was dominating 30-second audio windows. The audio feature requirements PR [3] tightens decorator usage across Voxtral and other models where missing `@requires` decorators masked import errors. Meanwhile, TRL's DistillationTrainer refactor [4] [5] removes the messages-format branch and prompt-truncation config entirely, simplifying the collator API since on-policy distillation only needs prompts, the student generates completions. This mirrors GRPO's cleaner surface and unblocks after the `num_items` fix stopped prompt-only batches from producing NaN. Funes hardened remote memory setup [6] by requiring TruffleHog validation before hub integration and marking newly created datasets as private in prompts, addressing the privacy and secret-scanning surface. Across all desks, 41 commits and 29 PRs shipped focused on robustness, performance, and API coherence.

Action items

→ Upgrade PyTorch to 2.5+ or apply DTensor guard patch before deploying models with sharding huggingface/transformers [immediate]
→ Review and test Whisper inference pipelines to confirm 8x speedup on CPU backends huggingface/transformers [plan]
→ Update DistillationTrainer integrations to use prompt-only datasets; messages-format is deprecated huggingface/trl [plan]
→ If using funes remote memory, validate TruffleHog installation before next `funes add` huggingface/funes [monitor]

References

[1] fix: guard DTensor import in sharding_utils.py for PyTorch < 2.5 (#47481) huggingface/transformers
[2] CPU can incur a slow path on non-contiguous magnitudes (#47351) huggingface/transformers
[3] [fix] fix requirements audio feature and proc ↗ huggingface/transformers
[4] [DistillationTrainer refactor] Remove messages-format support and prompt-length config ↗ huggingface/trl
[5] [DistillationTrainer refactor] Switch tests, docs, and example to prompt-only datasets ↗ huggingface/trl
[6] Harden remote memory setup in funes add ↗ huggingface/funes
[7] add a reproduce step ↗ huggingface/transformers-ci

Quick answers

What shipped in Hugging Face on July 23, 2026?: 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. In total, 41 commits and 29 pull requests landed.
Who contributed to Hugging Face on July 23, 2026?: 9 developers shipped this update, including dacorvo, tarekziade, vbayanag, eustlb, xhg, Yuanyuan Chen, Eric Bezzam, and qgallouedec, and 1 more.
What were the notable Hugging Face updates?: fix: guard DTensor import in sharding_utils.py for PyTorch < 2.5 (#47481), CPU can incur a slow path on non-contiguous magnitudes (#47351), and [fix] fix requirements audio feature and proc.

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

@huggingface 2 days ago

TRANSFORMERS FIXES RECURRENT MASKING BUG; MOON-IDE KILLS PAIRING WINDOW

A critical padding mask bug in transformers is now fixed for chunked prefill and cache continuation, while moon-ide removes dead startup ceremony and hardens the remote bridge client.

python 97 shipped 2-min read

@huggingface 3 days ago

TAU FIXES SESSION CORRUPTION BUG, SERGE ADDS GPU VERIFICATION LOOP

Tau's Unicode line separator bug was silently bricking user sessions. The fix is retroactive and ships today.

python 36 shipped 2-min read

@huggingface 4 days ago

MLCLAW TIGHTENS AUTHENTICATION, TAU SHIPS REASONING PERSISTENCE

MLClaw fixed a critical authentication gap where private Space edge cookies weren't reaching approval requests, forcing a shift to protected Space defaults across all hosted gateways.

python 69 shipped 2-min read

Elsewhere on the wire

AI Agents about 8 hours ago

LANGCHAIN REFRESHES MODEL PROFILES, OPENROUTER 0.2.7 SHIPS

OpenRouter integration lands with updated model profiles including Google's Gemini 3.5 Flash Lite, while core token counting gets a performance fix for multi-tool agents.

ai-agents 29 shipped 1-min read

Local LLMs about 8 hours ago

OLLAMA AGENTS NOW IMPORT SKILLS FROM CODEX AND CLAUDE

Ollama shipped a skills import system that lets coding agents pull validated skill directories from external sources, while simultaneous fixes to GLM tool-call parsing and Laguna model alignment prevent agents from silently failing.

+10

llms 212 shipped 2-min read

@CachyOS about 8 hours ago

BRAVE BROWSER SOURCES SHIFT TO APT, GAME RULES SURGE, GERMAN WIKI TRANSLATIONS ACCELERATE

CachyOS is consolidating Brave packaging around apt sources while ananicy-rules absorbs a massive gaming ruleset expansion and the wiki team pushes German localization across multiple docs.

infra 81 shipped 1-min read

Elixir & Phoenix about 8 hours ago

ELIXIR FIXES CALENDAR AND COMPILER EDGE CASES ACROSS FIVE PATCHES

Alexander Gubarev landed five targeted fixes to Calendar.ISO and ParallelCompiler that close gaps in time parsing, validation guards, and documentation.

elixir 15 shipped 1-min read

Want every project, not just this one?

Follow @huggingface