Narwhal fleet configuration and deployment reference¶
A fleet file configures one router process and one engine fleet for one model. A deployment with several models runs one router and one fleet per model, and ingress selects the model.
Print an annotated example fleet file:
narwhal-serve, narwhal-profile and narwhal-check read the fleet file named by --fleet. create_app() reads the path from the NARWHAL_FLEET environment variable.
Every fleet file declares this header:
The declared schema and version must match a supported interface version.
Ingress handles client identity and content capture. The Narwhal router holds one global admission budget, and request records hold timings, identifiers, placements and outcomes.
Fleet and deployment contracts¶
-
Fleet schema and engine contract
Top-level keys, JSON types, paths, environment variables, minimal fleet definition, shared-device allocation, hardware block,
engine_contractfields, and attestation. -
Admission, decode admission checks, queueing, retries, placement, prefix-cache pricing, deadlines, token counting, and role-controller settings.
-
Recovery, authentication, and profile validation
Breaker and drift settings, restart policy, resume, engine credentials, protocol adapters, profile store path, and decode-fit error limits.
-
Deployment inputs and SSH trust
Generated deployment files, role environments, host inventory, and known-hosts handling.
-
Engine endpoints and launch records
Engine URLs, launch-record fields, image verification, prefix caching, and KV cache events.
-
Fabric, CLI, and configuration operations
Fabric workload budgets, CLI precedence, request journal, provenance and publication, and deployment sequence.