skills.rest
atomicrajat
atomicrajatCOMMUNITY
Explore all skills by this author →
GitHub

rtvi-vlm-customize-model

Reconfigures the VLM endpoint across all three VLM consumers in the NVIDIA VSS Alerts Blueprint.

2|Updated Aug 20, 2026
Search Tags:#health checks#docker compose#openai-compatible#vllm#vlm#vss blueprint#rtvi-vlm
One-click install
npx skills add https://github.com/atomicrajat/industry_safety_monitoring_system --skill rtvi-vlm-customize-model-atomicrajat
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: rtvi-vlm-customize-model
Source: https://github.com/atomicrajat/industry_safety_monitoring_system/tree/main/.claude/skills/rtvi-vlm-customize-model
Command: npx skills add https://github.com/atomicrajat/industry_safety_monitoring_system --skill rtvi-vlm-customize-model-atomicrajat

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Swapping the vision-language model in the NVIDIA VSS Alerts Blueprint is error-prone because three separate services (rtvi-vlm, vlm-as-verifier, and vss-agent) each hold their own VLM configuration, so changing one environment variable silently leaves the other consumers pointing at the old endpoint. ## Core Features & Use Cases - Two deployment methods: Configure RTVI-VLM against an external OpenAI-compatible endpoint (NIM, vLLM, OpenAI) or run vLLM inside the RTVI container with NGC or Hugging Face model paths. - Three-consumer coverage: Guides updates to RTVI_VLM_* variables, the vlm-as-verifier config.yml, and vss-agent VLM_MODEL_TYPE/VLM_NAME/VLM_BASE_URL settings so no consumer is left stale. - Rigorous health checks: Verifies container status, absence of warm-up error lines, readiness, advertised model identity per method, and a real /v1/chat/completions inference rather than trusting misleading log lines. - Use Case: An engineer repoints the Alerts Blueprint to a host-side NIM at http://host.docker.internal:30082 and needs every VLM consumer, credential mapping, and verification step handled correctly. ## Quick Start Ask the agent to point the VSS Alerts Blueprint at your OpenAI-compatible VLM endpoint and verify all three consumers with health checks.

Frequently Asked Questions about rtvi-vlm-customize-model

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I change the VLM endpoint in the NVIDIA VSS Alerts Blueprint?▼

Set RTVI_VLM_MODEL_TO_USE=openai-compat and RTVI_VLM_ENDPOINT in the Alerts profile .env, then separately update vlm-as-verifier's config.yml and the vss-agent VLM_MODEL_TYPE/VLM_NAME/VLM_BASE_URL settings. Recreate the affected services and run the health checks.

Does changing RTVI_VLM_ENDPOINT update vlm-as-verifier and vss-agent automatically?▼

No. rtvi-vlm, vlm-as-verifier (alert-bridge), and vss-agent are three separate VLM consumers with independent configuration. Each must be updated and force-recreated individually, or the other services keep calling the old endpoint.

How do I run rtvi-vlm standalone with vLLM inside the container?▼

Set VLM_MODEL_TO_USE=vllm-compatible and point MODEL_PATH at the weights source, such as git:https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct or an ngc: path. Add HF_TOKEN only if the model requires authentication, then verify readiness and one real inference.

Why does the VLM health check pass even when warm-up failed?▼

VSS v3.2.1 logs 'Warmup VlmProcess-0 done' even when warm-up raised an exception, and readiness reports healthy for live processes whose model never loaded. Check for the absence of 'Error during warmup' lines and require a non-empty /v1/chat/completions response.

Why does /v1/models show a different model ID than VLM_NAME with vllm-compatible?▼

With Method B the advertised ID is derived from the resolved model directory basename, so ngc:nim/nvidia/cosmos-reason2-8b:hf-1208 becomes nim_nvidia_cosmos-reason2-8b_hf-1208. This is expected; point downstream consumers at the advertised ID, not the NGC path.

Can I reuse my NGC API key when switching from vLLM to an OpenAI endpoint?▼

No. Stock Compose maps RTVI_VLM_API_KEY (falling back to NGC_CLI_API_KEY) into VIA_VLM_API_KEY, which the client prefers over OPENAI_API_KEY. Replace the stale NGC credential and set both OPENAI_API_KEY and RTVI_VLM_API_KEY to the new endpoint's key.

Related Skills

Explore More Software Engineering Skills→
NVIDIANVIDIA
3.2k

vss-deploy-profile

Deploy, verify, debug, and tear down NVIDIA VSS profiles via Docker Compose.

Official
Advanced
rblake2320rblake2320

vss-manage-alerts

Route VSS camera alerts to Slack and dashboard backends.

Community
Advanced
NVIDIANVIDIA
3.2k

rtvi-cv-scaffold-vss-service

Scaffold a DeepStream YOLO26 microservice that publishes detection metadata to VSS via Kafka mdx-raw.

Official
Advanced
rish2jainrish2jain

nvidia-nim

Deploy and operate NVIDIA Inference Microservices for GPU-accelerated AI inference.

Community
Advanced