The training loop
Moving the policy
What a filesystem broadcast and a CUDA IPC refresh actually transfer.
After an optimizer update, the trainer and vLLM hold different versions of the policy. Refreshing inference is a correctness boundary before it is a speed optimization.