components.d: add Dynamo

[dagil-nvidia]

Overview

Add Dynamo to the components catalog.

NVIDIA Dynamo is a distributed LLM inference framework. Seven agent skills covering the full lifecycle of Dynamo are maintained at ai-dynamo/dynamo:.agents/skills/ and ready to mirror into this catalog via the standard daily sync pipeline.

This PR adds the one-file registry entry that registers Dynamo. No skill content lands in this repo directly — the sync workflow pulls .agents/skills/ from ai-dynamo/dynamo and renders skills/Dynamo/ on its next run after the upstream PR merges.

Details

What's in this PR

A single file: components.d/dynamo.yml.

name: Dynamo
repo: ai-dynamo/dynamo
description: Distributed LLM inference framework — pre-deployment sizing, model optimization, local serving, Kubernetes deployment, request-path configuration (Frontend + DynamoModel + gateway), day-2 troubleshooting, and benchmarking.
skills:
 - path: .agents/skills/
 catalog_dir: Dynamo

The schema matches the components.d/README.md spec; the fields follow the pattern of existing entries (compare cuopt.yml, tensorrt-llm.yml, nemoclaw.yml).

What the catalog will publish

When the sync runs after the upstream merges, skills/Dynamo/ will contain seven skill directories:

Skill	Lifecycle stage	Owns
dynamo-plan	Plan	AIConfigurator workflow, DGDR searchStrategy, SLA framing, recipe selection
dynamo-optimize	Optimize	Model Optimizer (modelopt) quantization — FP8 / NVFP4 / INT8 / AWQ
dynamo-serve	Local run	python3 -m dynamo.<backend> workstation workflow
dynamo-deploy	Deploy	dynamo-platform Helm, DGD + DGDR authoring, recipes, conversion webhooks, day-2 ops
dynamo-frontend	Request path	Frontend service (OpenAI endpoints), DynamoModel CR, multi-model, GAIE / kgateway / Istio
dynamo-troubleshoot	Day-2	Worker crashloops, inference 5xx, Planner stuck states, KV transfer fallback
dynamo-benchmark	Benchmark	AIPerf workflow, in-tree benchmarks/ suites, recipe-attached benchmarks

Gating

This catalog PR depends on the upstream Dynamo PR landing the skills at ai-dynamo/dynamo:.agents/skills/. Submitted as draft until that upstream PR merges.

• Upstream PR: ai-dynamo/dynamo#9847 — also draft for Computex visibility; ready for review when set out of draft.
• Once the upstream lands, this PR can be moved out of draft; the next daily sync will populate skills/Dynamo/.

Methodology note

The seven Dynamo skills follow conventions developed in NVIDIA's internal ai-infra-agent project — the 4-phase workflow shape, the DESTRUCTIVE / MUTATING / SAFE command-tier rubric, the Human-in-the-Loop behavioral contract, the pass / fail / warn and check() script helper patterns, the references/ + scripts/ subdirectory layout. Full attribution and the per-release update model live in the upstream PR's .agents/skills/README.md (visible in the upstream PR's diff; will be at the path .agents/skills/README.md once the upstream merges).

Quality

NV-ACES Tier 1 deterministic scoring on the seven skills (2026-05-21): average 92.1 / 100, lowest 90. All grades A- or A. Zero errors. All scripts shellcheck-clean. All YAML frontmatters parse with required fields present.

Where should the reviewer start?

• components.d/dynamo.yml (the only file in this PR).
• Optionally compare against the upstream .agents/skills/ directory in ai-dynamo/dynamo#9847 for the actual skill content.

Related Issues

• Companion upstream PR: ai-dynamo/dynamo#9847.