Narwhal CLI reference

narwhal-inference installs the commands below. Each executable accepts -h, --help and --version, and resolves relative paths from the process working directory. narwhal --help lists all six installed commands.

Use the installed narwhal-* commands in deployment scripts because internal Python module paths can change between releases. python -m narwhal.cli also starts a router.

In default text mode, the deployment and lifecycle commands exit 0 on success, 2 for invalid arguments or configuration inputs, and 1 when an operation fails, such as an engine HTTP request, runtime inspection, listener bind, verification gate or teardown. A degraded development instance returns 1 from narwhal dev status in that mode. Expected failures identify the command, operation and affected path or value on stderr; Python tracebacks identify unexpected failures. Engine inspections retain subprocess output in the run's diagnostic logs.

Command	Purpose
narwhal config	Validate fleet files and inspect resolved defaults and paths offline
narwhal diagnostics	Collect private router snapshots and selected incident artifacts
narwhal dev	Initialize, launch, verify and stop a local shared-GPU development fleet
narwhal-engine	Prepare and launch checked engine processes
narwhal-attest	Serve engine identity and attestation data for one vLLM engine
narwhal-serve	Run a Narwhal router
narwhal-profile	Measure engine behaviour and write the profile store used by the router
narwhal-check	Run deployment preflight gates

Finite commands accept --format json for versioned command results, including failures, artifact references and a documented status-to-exit-code mapping.