Image on HeadNode vs WorkerNode

Ray Libraries (Data, Train, Tune, Serve) Ray Serve
1

I am trying to run an LLM inference. The docker image is quite big, like 15 GB. It seems OK for the worker nodes, but should the head node have a smaller docker image, since it doesn’t need most of the components? Also, I am setting lower cpu/memory requirements for the head node.

Thanks.

2

I would recommend using the same image for all nodes for consistency sake. The image should realistically only get pulled once as the service is starting up, so the size shouldn’t impact the service during runtime.