Basic

The basic guide runs a disaggregated prefill/decode deployment where each role is its own child LeaderWorkerSet. Prompt processing (prefill) and token generation (decode) scale and roll out on their own.

• vLLM streams the KV cache from prefill to decode via --kv-transfer-config. The connector (NIXL, LMCache, etc.) and the prefill<->decode wiring are deployment-specific; the config here is a starting point, not a turnkey value.
• SGLang runs PD disaggregation via --disaggregation-mode. It needs a router/load balancer (for example sglang-router in mini-lb mode) to connect prefill and decode. That wiring is deployment-specific and omitted here. See the llm-d P/D disaggregation guide.

The other guides build on this one: autoscaling adds a per-role scaler and HPA, multi-slice fans the role set out into independent slices, and topology-aware scheduling co-locates a slice’s roles in one topology domain.

Deploy

• vLLM
• SGLang

export HF_TOKEN=<your-hf-token>
curl https://raw.githubusercontent.com/kubernetes-sigs/lws/refs/heads/main/docs/examples/disaggregatedset/basic/vllm.yaml -s | envsubst | kubectl apply -f -

export HF_TOKEN=<your-hf-token>
curl https://raw.githubusercontent.com/kubernetes-sigs/lws/refs/heads/main/docs/examples/disaggregatedset/basic/sglang.yaml -s | envsubst | kubectl apply -f -

Verify the child LeaderWorkerSets and pods:

kubectl get leaderworkersets
kubectl get pods

Roll out a role by editing its container spec; the per-role rolloutStrategy upgrades that role independently of the others. See the DisaggregatedSet concepts for the API details.