NVIDIA VSS Quickstart

Why it matters

A video agent is an LLM, a vision-language model and a media pipeline deployed together. This quickstart stands the whole stack up, so you can see where each model sits.

What to practise

Deploy the base profile with dev-profile.sh up -p base and the -H flag for your GPU, choosing shared GPUs, dedicated device IDs, or --use-remote-llm and --use-remote-vlm. Upload warehouse_safety_0001.mp4 in the VSS UI, ask questions about it, and generate a report, editing the VLM prompt when the agent asks you to review it. Follow the agent’s workflow telemetry in the Phoenix UI. Note which requests the Nemotron LLM routes and which the Cosmos Reasoner VLM answers.

How it fits

The deployed services are 7.5.1’s video route: ingestion and storage in VIOS, understanding in the VLM. Compare the default settings with 7.5.2’s frame-sampling and input-size budgets before sizing a deployment of your own.