[RFC]: Provide recipes and a skill/documentation for multi-node deployment on Kubernetes clusters

Motivation

Current recipes don't support multi-node deployment:

• AFD rendezvous host is 127.0.0.1 which is only reachable from processes on the same machine
• GPU selection via CUDA_VISIBLE_DEVICES only makes sense if all ranks' GPUs live in one host's device list.
• One bash script executes every role (attention, ffn).

Proposed change

Goal:

1. Create examples of templated recipes for multi-node deployment that do not hard-code host, list of devices, etc., and split vllm processes per role.
2. Create a deployment skill that uses the templated multi-node recipes for the deployment on a kubernetes cluster.

Initial scope:

• collocation recipes
• CUDA device backend

Plugin boundary

No patches required.

Risks and alternatives

No changes to existing recipes/scripts.

Feedback period

No response

CC list

No response

Start by reviewing the existing recipes and bash scripts to understand the current collocation and CUDA device backend setup. Define templated multi-node recipes with configurable hosts and devices, split processes by role, and document a Kubernetes deployment skill. Done means the initial collocation and CUDA backend scope is covered without changing existing recipes or scripts.