dynamo-recipe-runner

Dynamo Recipe Runner helps an operator go from user intent to a working NVIDIA Dynamo endpoint on Kubernetes with minimal back-and-forth. Rather than authoring new content, it operates on the existing recipes/ tree in the ai-dynamo/dynamo repository: it selects an appropriate recipe by model, framework (vllm, sglang, trtllm, tokenspeed), deployment mode, and GPU type/count, validates it, patches only the minimal set of manifest values needed for the run, deploys when the user has cluster access, and proves success with an OpenAI-compatible smoke request.

The documented workflow starts with read-only preflight checks (git status, a recipe-scanner listing, kubectl context, storage classes, nodes, namespace, and the Hugging Face token secret), then recipe selection via a bundled recipe_tool.py, inspection and validation, minimal patching (storage class, image tag, model path, GPU requests, router mode, namespace), deployment via kubectl apply of the model-cache and deploy manifests, and a port-forwarded smoke test against /v1/models and one chat completion. It defers router-only work to dynamo-router-starter and failure diagnosis to dynamo-troubleshoot. It handles secrets carefully, instructing that Hugging Face tokens never be written into files or logs and always be sourced from Kubernetes secrets.

Target users are ML infrastructure engineers and operators bringing up Dynamo inference recipes on a shared cluster. Its pre-publication benchmark returned an overall FAIL verdict, with a noted finding that the skill instructs applying manifests to a live Kubernetes cluster without an explicit user-confirmation gate, so this is state-changing infrastructure automation that should be reviewed before execution.