How to use data-parallel in vLLM

I want know how to use data-parallel not for deployment, but for offline inference.