vLLM adds day-0 LoRA reinforcement learning support for Muse Spark Open

vLLM announces day-0 LoRA reinforcement learning support for Muse Spark Open, with vLLM handling rollouts and verl FSDP2 handling training.

Rollouts run on the NVFP4 base model with LoRA, while the trainer uses a BF16 dequantization of the same weights. Only the LoRA adapter is updated and synced back to the rollout engine, with R3 replaying rollout routing.

In a GRPO run on DAPO-Math-17K over 150 steps using 16 GB300 GPUs, vLLM reports AIME-2024 mean@8 rising from 0.56 to 0.85, with train-rollout k3 KL near 2 × 10⁻⁴. These results come from vLLM's own announcement on X and have not been independently verified.