Who contributed to Hugging Face on July 31, 2026?

14 developers shipped this update, including Cyril Vallez, Yoni Gozlan, Ferdinand Mom, delmalih, asomoza, albertvillanova, Strongich, and HUANG TZU-CHUN, and 6 more.

What were the notable Hugging Face updates?

Drop multimodal inputs natively in prepare_inputs_for_generation if not in prefill (#47622), Remove Rotary warning (#47642), and [Chat Parsing] Type inline tool-call arguments from the calling tool's JSON Schema (#47529).

@huggingface

Transformers, Datasets, and the open AI-model layer

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

TRANSFORMERS FIXES MULTIMODAL INFERENCE PIPELINE, DIFFUSERS ADDS SDNQ QUANTIZATION

By RepoJournal · Filed 06:59 UTC on July 31, 2026 · About Hugging Face

14 people shipped this

Cyril Vallez @Cyrilvallez 3 cited

Yoni Gozlan @yonigozlan 1 cited

Ferdinand Mom @3outeille 1 cited

asomoza @asomoza 1 cited

albertvillanova @albertvillanova 1 cited

Xingdong Zuo @zuoxingdong 1 cited

Steven Palma @imstevenpmwork 1 cited

Caroline Pascal @CarolinePascal 1 cited

delmalih @delmalih

Strongich @Strongich

HUANG TZU-CHUN @tc-huang

github-actions[bot] @github-actions[bot]

changdazhou @changdazhou

machineuser @machineuser

Transformers dropped native multimodal input handling in prepare_inputs_for_generation while diffusers shipped core support for SDNQ quantization, the first major quantization framework addition in months.

The transformers team merged three critical fixes across the inference stack. A patch [1] now drops multimodal inputs natively in prepare_inputs_for_generation if they're not in the prefill stage, fixing a class of edge cases where vision inputs would contaminate token generation. The same effort [2] removed the Rotary deprecation warning and [4] cleaned up Gemma 4 warnings that were firing on every model load. Separately, [5] improved torch distributed availability checking to handle binaries compiled with USE_DISTRIBUTED=0, unblocking users on constrained deployments. Chat parsing [3] got smarter about tool-call arguments, now typing inline arguments directly from the calling tool's JSON schema rather than applying generic value parsers downstream. Over in diffusers, [6] landed SDNQ quantization core loading, letting you load prequantized checkpoints with a normal from_pretrained call or quantize on the fly by passing SDNQConfig, with full save/reload support inside pipelines. TRL [7] added a regression test for KTO with pre-wrapped PEFT models using target_parameters, matching equivalent coverage in GRPO, DPO and RLOO. LeRobot fixed [8] SmolVLM layer freezing that was silently leaving the final VLM layer trainable, [9] corrected frame errors being treated as shard exhaustion in streaming datasets, and [10] removed stereo fallback on RealSense RGB-only camera color parameter setting.

Action items

→ Upgrade transformers to pick up multimodal inference fix and warning removal huggingface/transformers [plan]
→ Review SDNQ quantization workflow if shipping diffusion models to production huggingface/diffusers [monitor]
→ Verify SmolVLM fine-tuning layer freezing behavior if using train_expert_only=False huggingface/lerobot [plan]

References

[1] Drop multimodal inputs natively in prepare_inputs_for_generation if not in prefill (#47622) huggingface/transformers
[2] Remove Rotary warning (#47642) huggingface/transformers
[3] [Chat Parsing] Type inline tool-call arguments from the calling tool's JSON Schema (#47529) huggingface/transformers
[4] Remove gemma4 warnings (#47664) huggingface/transformers
[5] better guarding to handle torch compiled with USE_DISTRIBUTED=0 (#47619) huggingface/transformers
[6] [quantization] SDNQ core loading ↗ huggingface/diffusers
[7] Add `test_train_moe_peft_model` for KTO (#6589) huggingface/trl
[8] fix(smolvla): freeze the intended VLM layers when train_expert_only=False (#4019) huggingface/lerobot
[9] fix(datasets): stop frame errors being treated as shard exhaustion in StreamingLeRobotDataset (#4237) huggingface/lerobot
[10] fix(RGB only): remove the stereo module fallback when setting colors parameters on RealSense cameras (#4225) huggingface/lerobot

Quick answers

What shipped in Hugging Face on July 31, 2026?: Transformers dropped native multimodal input handling in prepare_inputs_for_generation while diffusers shipped core support for SDNQ quantization, the first major quantization framework addition in months. In total, 39 commits, 38 pull requests, and 1 releases landed.
Who contributed to Hugging Face on July 31, 2026?: 14 developers shipped this update, including Cyril Vallez, Yoni Gozlan, Ferdinand Mom, delmalih, asomoza, albertvillanova, Strongich, and HUANG TZU-CHUN, and 6 more.
What were the notable Hugging Face updates?: Drop multimodal inputs natively in prepare_inputs_for_generation if not in prefill (#47622), Remove Rotary warning (#47642), and [Chat Parsing] Type inline tool-call arguments from the calling tool's JSON Schema (#47529).

SPEECH-TO-SPEECH 0.2.12 SHIPS WITH SMARTER TURN-TAKING AND WEBRTC SUPPORT

The final 0.2.x release lands Smart Turn v3.2 endpointing, WebRTC support, and direct audio input for audio-capable LLMs, while diffusers onboards MiniMax-H3 for joint video and audio generation.

python 57 shipped

@huggingface Aug 5

KERNELS REFACTOR CLEANS UP LOADING PATH, TOKENIZERS KILLS DEPENDABOT NOISE

Hugging Face is untangling kernel loading internals across multiple repos while simultaneously eliminating 50% of its Dependabot burden through surgical dependency cleanup.

python 26 shipped

@huggingface Aug 4

LEROBOT BREAKS IMPORTS, FIXES TRAINING PIPELINE SILENTLY BREAKING

LeRobot 0.6.1 renames the types module and patches a critical bug where multi-stage training produced all-zero targets when episode annotations were malformed.

+11

python 69 shipped

@huggingface Aug 3

LEROBOT FIXES D405 REALSENSE STARTUP HANGS WITH AUTO-RECOVERY

LeRobot's camera stack now survives RealSense D405 hardware lockups that used to require a physical replug.

python 5 shipped

Elsewhere on the wire

AI Agents Aug 6

LANGCHAIN HARDENS ANTHROPIC TOOL CALLING AND GIT CREDENTIALS AS OPENHANDS SHIPS 1.10.0

LangChain fixed a critical credential leak in evaluation runs while Anthropic tool schemas get proper fallback handling, as three major agent frameworks push releases across the ecosystem.

ai-agents 54 shipped

Local LLMs Aug 6

OLLAMA FIXES KV CACHE CORRUPTION, LLAMA.CPP HARDENS FILE SEARCH, VLLM KILLS MOE LEGACY CODE

Ollama patched a critical cache corruption bug that corrupts lazy snapshots during multi-request prefills, while llama.cpp hardened Windows junction traversal and vLLM removed 14 hardcoded MoE shards across quantization methods.