Multi-slice

This guide is the basic deployment plus slice fan-out. spec.slices: 2 replicates the whole prefill/decode role set into two independent slices, so each slice is a self-contained prefill+decode unit. Combine this with topology-aware scheduling to co-locate each slice’s roles and spread slices across domains.

Slice fan-out and external role autoscaling are mutually exclusive: external scaling requires spec.slices: 1.

Deploy

• vLLM
• SGLang

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

Verify the slices:

kubectl get leaderworkersets

See the slices concepts for details.