Hybrid rollout switching is merged in verl-omni (#575): the v1 separate-async diffusion trainer now lends its co-located rollout GPUs to generation when the replay buffer runs short, raising Wan2.2 video training throughput by 22% at a fixed GPU budget.