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

Summary

Hugging Face published a method for running asynchronous GRPO reinforcement learning with LoRA adapters across its Jobs infrastructure, using a bucket and a proxy instead of NCCL.

Key facts

• Hugging Face published the method on 2026-09-10.
• The approach runs async GRPO with LoRA across HF Jobs.
• It relies on a bucket and a proxy rather than NCCL.

Why it matters

Teams training reinforcement learning models on Hugging Face Jobs can distribute GRPO with LoRA without depending on NCCL, which may simplify setup for builders using that infrastructure.