The Wire · Showcase
DIFFUSERS SHIPS COSMOS3 FEW-STEP INFERENCE, LEROBOT STRIPS 3K LINES OF VENDORED CODE
By RepoJournal · Filed · About Hugging Face
Cosmos3 distilled checkpoints hit diffusers with 4-step T2I and I2V support, while lerobot kills its Qwen2.5-VL fork and rebases onto native transformers.
The diffusers pipeline now supports nvidia's Cosmos3-Super distilled variants [1], both 4-step checkpoints shipping with fixed sigma schedules already baked in for immediate inference without tuning. Parallel to that, diffusers is gutting its test layer [2] and adding granular control over pipeline signatures [3], letting test classes declare which canonical parameters a pipeline deliberately drops (like guidance-distilled models skipping negative_prompt) so regressions fail hard. Over in lerobot, the Wall-X policy got a major haircut [7]: the team ripped out 2.8k lines of vendored Qwen2.5-VL code and subclassed the native transformers version instead, killing the hand-sync tax every time transformers updates. That refactor is part of a broader policy cleanup [8] [9] [10] extracting shared model components and standardizing how policies register themselves by convention rather than boilerplate. trl moved fast on housekeeping: KTO trainer got redundant tensor re-wrapping removed [4] [6], and both GRPO and RLOO trainers now document the max_steps requirement for iterable datasets [5], closing the docs gap that already existed in DPO and SFT.
Action items
- → Review lerobot Wall-X rebase for compatibility before next deployment huggingface/lerobot [plan]
- → Adopt Cosmos3 distilled checkpoints in downstream diffusers workflows huggingface/diffusers [monitor]
- → Update any custom KTO trainers to remove redundant torch.tensor wraps huggingface/trl [plan]
References
- [1] Cosmos3 Distilled support ↗ huggingface/diffusers
- [2] [tests] Pipeline test refactor ↗ huggingface/diffusers
- [3] Add not_params to ModularPipelineTesterMixin (#14207) huggingface/diffusers
- [4] Remove redundant torch.tensor re-wrapping of batch labels in KTO (#6431) huggingface/trl
- [5] Document max_steps requirement for iterable datasets in GRPO and RLOO ↗ huggingface/trl
- [6] Remove redundant torch.tensor re-wrapping of batch labels in KTO ↗ huggingface/trl
- [7] refactor(wall-x): subclass native Transformers Qwen2.5-VL instead of vendoring it ↗ huggingface/lerobot
- [8] refactor(policies): extract shared model components ↗ huggingface/lerobot
- [9] refactor(processors): share policy pipeline builders ↗ huggingface/lerobot
- [10] refactor(policies): resolve policy components by convention ↗ huggingface/lerobot