Hugging Face runs async GRPO with LoRA and Storage Bucket on HF Jobs, drops NCCL and speeds up 3.9×

Hugging Face 用 LoRA 与 Storage Bucket 在 HF Jobs 上跑异步 GRPO，免 NCCL 并提速 3.9 倍

TRL v1.14's AsyncGRPOTrainer now supports training only LoRA adapters and syncing them via a Storage Bucket, so the training job and vLLM inference job run on separate machines without NCCL. Hugging Face reports a 3.9× speedup over the synchronous setup. The post doesn't disclose test configs or latency numbers, so I'd hold for those details.