$ the-wire · showcase
LEROBOT CUTS FASTWAM OVERHEAD WITH COMPILED INFERENCE
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
![moon-bot-app[bot]](https://avatars.githubusercontent.com/in/3225802?v=4)
12 people shipped this
darthjaja6
@darthjaja6
1 cited
hq-fang
@hq-fang
1 cited
pkooij
@pkooij
1 cited
moon-bot-app[bot]
@moon-bot-app[bot]
1 cited
Pierrci
@Pierrci
1 cited
sergiopaniego
@sergiopaniego
1 cited
albertvillanova
@albertvillanova
1 cited
sayakpaul
@sayakpaul
akshan-main
@akshan-main
dg845
@dg845
moon-bot-app
@moon-bot-app
HuggingFaceInfra
@HuggingFaceInfra
LeRobot's FastWAM policy gets an opt-in compiled path that slashes Python and CUDA kernel-launch overhead on cached action inference.
LeRobot shipped an opt-in compiled inference path for FastWAM that cuts Python and CUDA kernel-launch overhead on cached action inference, a win for anyone running repeated denoising steps on the same video features [1]. The same repo tightened MolmoAct2 continuous training to match the official recipe, adopting a BF16 VLM / FP32 action-expert precision policy and the official 30K scheduler clock [2]. On the policy side, Pi05 now supports training-time Real-Time Chunking behind a flag that defaults to 0, so existing pipelines stay untouched [3]. Outside LeRobot, a stale Pi install command on model pages has been fixed: huggingface.js now points to the new @earendil-works/pi-coding-agent package, which sits at 0.84.2 versus the frozen 0.73.1 release [4]. Hub docs clarify that PRO subscription credits are now general-purpose compute credits, applied automatically before pay-as-you-go, usable across Inference Endpoints, Spaces, and Jobs [5]. TRL fixed a CLI parsing bug where --scale_rewards group failed because SDPOConfig typed it as bool; it now matches GRPOConfig's string type [6]. The async trainer tests in TRL also upgraded to a Flash Attention-compatible model so head_size is a multiple of 8, un-xfailing a set of tests that were expected to fail [7].
Action items
- → Test FastWAM with the new compiled action inference path to measure latency gains huggingface/lerobot [plan]
- → Update any stale Pi install commands to @earendil-works/pi-coding-agent huggingface/huggingface.js [immediate]
- → Review MolmoAct2 fine-tuning scripts to match the aligned official recipe huggingface/lerobot [monitor]
References
- [1] perf(fastwam): add compiled cached action inference ↗ huggingface/lerobot
- [2] fix(molmoact2): align continuous training with official recipe ↗ huggingface/lerobot
- [3] feat(pi05): add optional training-time RTC ↗ huggingface/lerobot
- [4] Update Pi local-app snippet to the new package name and repo URL (#2390) ↗ huggingface/huggingface.js
- [5] included sub credits are now compute credits (#2719) ↗ huggingface/hub-docs
- [6] Fix SDPO scale_rewards CLI parsing and warn on ignored feedback ↗ huggingface/trl
- [7] Use a Flash Attention compatible model in the async trainer tests ↗ huggingface/trl
