DisaggregatedSet

DisaggregatedSet runs a multi-role inference deployment where each role is its own child LeaderWorkerSet. The common pattern splits inference into a prefill role for prompt processing and a decode role for token generation, so each one scales and rolls out on its own.

These examples show four things. The prefill and decode roles are separate. spec.slices fans the role set out into several independent slices. A per-role rolloutStrategy upgrades each role on its own. spec.placementPolicy co-locates a slice’s roles in one topology domain and spreads slices across domains.

Each guide isolates one feature and ships both a vllm.yaml and a sglang.yaml.

• Basic: minimal prefill/decode disaggregation.
• Autoscaling: autoscale a role with a DisaggregatedSetRoleScaler and HPA.
• Multi-slice: fan the role set out into independent slices with spec.slices.
• Topology-aware scheduling: co-locate a slice’s roles with spec.placementPolicy.

See the DisaggregatedSet concepts for the API details.