# Rtvi Vlm Perf Testing

Plan, run, and diagnose reproducible RT-VLM GPU performance canaries and benchmarks.

- Repository: https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization (1959 stars)
- Category: Engineering
- License: Apache-2.0
- Last updated: 2026-11-01
- Runs scripts: yes, review before use

## Install

Claude Code: `npx skills add NVIDIA-AI-Blueprints/video-search-and-summarization --skill rtvi-vlm-perf-testing -g -a claude-code -y`
Codex: `npx skills add NVIDIA-AI-Blueprints/video-search-and-summarization --skill rtvi-vlm-perf-testing -g -a codex -y`
Cursor: `npx skills add NVIDIA-AI-Blueprints/video-search-and-summarization --skill rtvi-vlm-perf-testing -g -a cursor -y`
Gemini CLI: `npx skills add NVIDIA-AI-Blueprints/video-search-and-summarization --skill rtvi-vlm-perf-testing -g -a gemini-cli -y`

## Description

Plan, run, and diagnose reproducible RT-VLM GPU performance canaries and benchmarks. Use this skill when running fresh-container stream-capacity, semantic-isolation, latency, throughput, or regression experiments against an RTVI microservices checkout.

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# RTVI VLM Perf Testing

## Operating Rules

This repository ships the benchmark implementation under `services/rtvi/rt-vlm/perf/`. Set the manifest `repo` to that service directory and confirm that it contains `perf/benchmark/` before using repository-specific benchmark commands. Read the repository contributor instructions before changing scripts, configs, docs, or benchmark behavior. Check `git status --short` before editing and preserve unrelated dirty worktree changes.

The committed planning and canary helpers require Python 3.9+. Executed canaries additionally require SSH, `tmux`, Docker with NVIDIA Container Toolkit, FFmpeg/FFprobe, and an NVIDIA GPU on the named remote host. Model images and protected artifacts may require the user's existing NGC or artifact-registry access; never request that credentials be stored in a manifest.

Do not start or stop services when the user only asks to inspect existing reports. Treat setup, teardown, and benchmark runs as side-effecting operations and state that you are about to run them.

When editing shell scripts, validate with `bash -n`. When changing commands, environment variables, Docker behavior, benchmark scenarios, or report workflows, update the matching guidance in the repository `AGENTS.md`, service README, and `perf/benchmark/README.RTVI_VLM.md` or `PERF_GUIDE.RTVI_VLM.md`.

Before launching a new experiment, freeze its identity, workload, metrics, scenarios, and isolated
paths using `references/performance-contract.md`. Validate and render the plan without executing it:

```bash
python3 skills/benchmarking/rtvi-vlm-perf-testing/scripts/perf_plan.py validate plan.json
python3 skills/benchmarking/rtvi-vlm-perf-testing/scripts/perf_plan.py render plan.json
```

The rest: https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization/blob/HEAD/skills/benchmarking/rtvi-vlm-perf-testing/SKILL.md