HuggingFace Jobs GRPO experiment: one code change makes each training step 40% faster

FerdinandMom · x · 2026-11-18

Testing AsyncGRPO + LoRA on Qwen3-30B-A3B with HuggingFace Jobs (3 vLLM Jobs + 2 trainer Jobs), the author found GPUs talk over NVLink inside a Job but only TCP between Jobs. Same GPUs, same model — one simple code change cut 40% off each training step. The thread breaks down the cause and the fix.

Related event: HSDP Beats FSDP by 40% in Async GRPO Training on HuggingFace Jobs(3 posts)→

Original post →

More from Infra

Infra channel →