Large LLM like Llama3.1-405B or Deepseek-V3 (671B), require distributed inference across multiple-nodes like vLLM + Ray backend.
However, it's more than just model-slicing with tensor-parallelism, Native K8S treats those workloads across nodes irrelevantly , so challenges come:
- standalone statefulSets without coordination
- demand of Gang-scheduling
- uncontrolled startup order among master & workers, causing boot lag
- HPA as a whole instead of for each sts, to scale together for both Ray head/worker.
- stable index and rank
- topology aware grouping
- failure recovery for vllm/pytorch(not smart enough), to avoid one pod/GPU failure disrupting overall inference

----
So LWS - LeaderWorkerSet (github.com/kubernetes-sigs/lws) , is designed to address them:
- to optimize resource coordination with leader-worker set
- improve performance thru co-location
- integrate scaling with HPA for whole lws together
- all-or-nothing restart policy to fault tolerance as a group.