Core concepts¶
Narwhal assigns prefill and decode roles across dual-capability engines serving one model. A role move changes placement for new requests while model weights, KV paths, and resident requests stay on their engines.
Concepts¶
Related reference material¶
- Backend continuation contract: producer ownership, local decode, descriptor validation, and timing boundaries.
- Configuration: placement policy, role guards, serving limits, and engine health.
- Measure a fleet: performance profiles, transfer checks, deployment load, and acceptance evidence.
- HTTP API: state-document definitions.
- Operate Narwhal: lifecycle and failover procedures.
- Source responsibilities: package ownership and import constraints.