TRL async GRPO adds LoRA sync via storage bucket and proxy, cutting 500-step training to 53 min

SergioPaniego · x · 2026-09-22

Hugging Face engineers detail LoRA support in TRL's AsyncGRPOTrainer (v1.14): only a few-MB adapter syncs to vLLM via a shared Storage Bucket instead of NCCL. Trainer and two vLLM replicas run as separate HF Jobs; a proxy routes each rollout to the replica holding its KV prefix. Iterating on AsyncGRPO metrics cut a 500-step run from 3h27m to 53m. Rationale: rank-1 LoRA suffices for policy-gradient RL per Thinking Machines' findings.

Original post →

More from Infra

Infra channel →