Start or patch Dynamo router modes and run router endpoint smoke checks. Use for round-robin, KV-aware, least-loaded, or device-aware routing setup; use recipe-runner for recipe deployment and troubleshoot for failure diagnosis.
Dynamo Router Starter makes NVIDIA Dynamo request routing easy to bring up and verify. It solves the problem of getting a baseline router mode running, enabling KV-aware routing when appropriate, and proving the endpoint works, all while keeping the user focused on exact commands and success signals rather than router internals. It complements dynamo-recipe-runner (which deploys recipes) and dynamo-troubleshoot (which diagnoses failures).
The workflow establishes a baseline router mode (for example round-robin) either locally via python3 -m dynamo.frontend or on Kubernetes by patching the frontend service env, then enables KV routing (with an approximate-mode fallback when workers do not publish KV cache events), and finally runs a bundled check_router_health.py smoke test against /v1/models and one /v1/chat/completions request. It supports modes including round-robin, kv, least-loaded, device-aware-weighted, direct, and random, and it is careful about methodology: when comparing modes it insists on holding model, workers, prompts, concurrency, and sampling constant, labeling single-request results as smoke comparisons rather than benchmarks, and never claiming throughput gains from one chat request.
Target users are developers and engineers configuring Dynamo routing and verifying endpoint health. The skill's changes are limited to starting a local frontend or patching a small set of frontend environment variables, and the smoke test is read-oriented, exercising standard OpenAI-compatible endpoints. Its pre-publication benchmark returned an overall PASS verdict. For throughput and latency numbers it explicitly defers to a separate dynamo-benchmark skill, and for unhealthy endpoints it routes to dynamo-troubleshoot.
round-robin, kv (KV-aware), least-loaded, device-aware-weighted, direct, and random. It also supports an approximate KV mode when workers do not publish KV cache events.
Python 3.10+ with the dynamo package importable, and for Kubernetes runs kubectl access to the target namespace with a deployed Dynamo recipe, network reachability to the frontend service, and at least one model loaded into a worker.
It runs the bundled check_router_health.py against the frontend, verifying /v1/models and, when a model is discoverable, one /v1/chat/completions request.
No. The smoke test is a single chat completion, not a benchmark. The skill explicitly defers throughput and latency measurement to a separate dynamo-benchmark skill and warns against claiming improvement from one request.
It starts a local frontend or patches only the frontend service environment variables (for example DYN_ROUTER_MODE and DYN_ROUTER_USE_KV_EVENTS); it does not deploy full recipes.
Quick Setup:
.claude/skills/Repository
nvidia/skills