Tutorials & Tips

Hugging Face runs asynchronous GRPO on HF Jobs with LoRA and Storage Bucket, eliminating NCCL and speeding up 3.9x

Original title:Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL

AI summary

Hugging Face's AsyncGRPOTrainer in TRL v1.14 supports training and synchronizing only LoRA adapters; combined with Storage Bucket mounts and proxy routing, training jobs and vLLM inference jobs can run on different machines without NCCL.

Key Takeaways

  • Hugging Face's AsyncGRPOTrainer in TRL v1.
  • 14 supports training and synchronizing only LoRA adapters
  • combined with Storage Bucket mounts and proxy routing,training jobs and vLLM inference jobs can…

Quick Facts

Published
Updated
Source
Hugging Face 社区博客(混合发现)
Category
Tutorials & Tips
AI score
61
Read the original Read the full article at the original source
Why it matters

Hugging Face 官方用五个实验展示了如何在 Jobs 上分离训练与推理跑异步 GRPO,并通过逐项定位瓶颈把运行提速 3.9 倍,方法可迁移。

Related entities

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: Hugging Face runs asynchronous GRPO on HF Jobs with LoRA and Storage Bucket, eliminating NCCL and speeding up 3.9x. September 10, 2026. https://new.nat6.net/en/news/cmu10mxhc08m6ro2nafd8smam/

View all →