How should I enable vllm distributed inference on multiple single GPU nodes?

[Jie2GG]

I have multiple xinference worker nodes deployed in a server cluster, and when I run Qwen1.5-14B-Chat, I want to enable distributed inference on multiple nodes to better leverage multiple Gpus in multiple clusters (like the ray cluster), But xinference doesn't seem to allow me to do that.

How should I configure xinference so that the vllm framework can leverage Gpus on multiple nodes for distributed inference?