Hugging Face enables Async GRPO with LoRA using HF Jobs
September 9, 2026
Hugging Face implements asynchronous Group Relative Policy Optimization (GRPO) with LoRA by utilizing a bucket and proxy architecture. This approach avoids NCCL dependencies, allowing for distributed training across decoupled HF Jobs instances.
HOW THIS AFFECTS YOU
●
builderYou can scale reinforcement learning training more easily across distributed jobs without complex NCCL configurations.
●
researcherThis provides a more accessible framework for implementing GRPO-based alignment techniques.