Dell Technologies Dell Deskside Agentic AI Solutions
Home / solutions / agentic-vss / it-admin-guide

Agentic VSS — IT Admin Deployment Guide

Who this is for: enterprise IT admins rolling out the VSS Blueprint across a set of Dell deskside systems.

Developer quick-start: Developer Deploy Guide.

Overview

VSS deploys as a stack of Docker Compose services managed by the upstream NVIDIA VSS Blueprint. This solution adds Dell-specific sizing, per-platform model overlays, and a fleet-install path.

Component Purpose Install
VSS Agent + UI Interactive UI and API for search / summarization / alerts Upstream Blueprint (Docker Compose)
VIOS / VST Video ingest, embedding, stream understanding Upstream Blueprint
Vector / graph DB Search indexes Upstream Blueprint
LLM (vLLM) Summarization, Q&A, alert reasoning Local vLLM container (per-tier model below)
VLM (vLLM) Vision-language reasoning over frames Local vLLM container (Nemotron 3 Nano Omni)
Phoenix Observability Upstream Blueprint

Prerequisites (per host)

  • Ubuntu 24.04 LTS or newer on Dell Pro Precision 9 towers and Dell Pro Precision with GB300; GB10 systems ship with DGX OS.
  • Docker CE 28.3.3 ≤ version < 29.5.0 — upstream VSS is sensitive to Docker versions; 29.5.0+ may fail pulling NGC images.
  • NVIDIA Container Toolkit.
  • Git LFS.
  • NGC API key (NGC_CLI_API_KEY).
  • Hugging Face token (HF_TOKEN) with access to the Nemotron models.
  • 50–80 GB free disk for VSS containers + LLM + VLM weights, plus space for the video corpus (specific to you).

Model selection per tier

Use the per-tier LLM default (the VLM is the same across all tiers):

Dell system LLM (default · alt) VLM
Dell Pro Precision with GB10 (128 GB unified) nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · alt nvidia/Qwen3.8-27B-NVFP4 nvidia/nemotron-3-nano-omni-30b-a3b
Dell Pro Precision 9 T2 (1× RTX PRO 6000, 96 GB) nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 nvidia/nemotron-3-nano-omni-30b-a3b
Dell Pro Precision 9 T4 (2× RTX PRO 6000) Nemotron 3.5 Super VL · alt nvidia/Qwen3.8-Flash-Next-NVFP4 nvidia/nemotron-3-nano-omni-30b-a3b
Dell Pro Precision with GB300 (252 GB HBM3e, aarch64) Nemotron 3.5 Super VL · alt RedHatAI/GLM-5.3-Flash-NVFP4 nvidia/nemotron-3-nano-omni-30b-a3b

Standalone LLM serving recipes per tier are in the recipe catalog; inside the VSS hybrid stack the LLM runs co-located with the VLM at reduced memory utilization (see below). VLM serving command: inline in the Developer Deploy Guide.

Install (per host)

Follow Developer Deploy Guide — the developer flow is the fleet flow, run per host. For multi-host rollouts, wrap the steps in your existing configuration-management tool or Ubuntu Landscape script execution (see Fleet Deploy via Landscape for the pattern). The upstream VSS Blueprint's Docker Compose is the deployment unit; your tooling handles inventory, secret distribution, and rollout ordering.

VLM + LLM co-location on S tier

VSS runs the VLM and the LLM on the same GPU in Standard hybrid mode. This is intentional and validated on the S tier:

  • Peak co-located GPU memory: 68.3 GB of 96 GB (71.4 %), measured on Dell Pro Precision 9 T2.
  • Both containers use --gpu-memory-utilization 0.35.
  • VLM must start before LLM — vLLM requires the target GPU to have free memory at startup.

For M / L tiers with additional GPUs, split VLM and LLM across GPUs and drop the 0.35 cap.

Validation checklist

Check Command Expected
LLM endpoint healthy curl -s :30081/v1/models Lists LLM model
VLM endpoint healthy curl -s :30082/v1/models Lists VLM model
VSS Agent UI reachable curl -I :7777 HTTP 200
VSS health endpoint curl -fsS :8000/health 200 OK
End-to-end smoke test Ingest → NL query → result 0 errors

Monitoring

  • vLLM exposes /metrics for both LLM and VLM containers — scrape with Prometheus. Cover request rate, tok/s, TTFT, KV-cache utilization.
  • VSS exposes ingestion metrics via Phoenix.
  • GPU telemetry via NVIDIA DCGM exporter.

Extend your Grafana dashboards with video-pipeline panels: VLM ingest throughput (chunks / sec), LLM summarization latency, storage growth rate.

Storage

Video corpus size + vector / graph DB indices are the dominant storage cost. Size against the retention window:

  • Local disk — fastest, suitable for S-tier evaluations.
  • NAS / NFS — recommended for M / L tiers with shared corpora.
  • S3-compatible object store — supported by upstream VSS via environment configuration; recommended for large / long-retention archives.

Actual GB / hour of video depends on your footage (depends on bitrate, resolution, container). Measure early; document per-site. Index size is typically 5–15 % of raw video size for typical embedding schemes.

Data handling

Video at rest may be sensitive (PII, security-camera footage, medical imaging). Document upfront:

  • Retention window.
  • Encryption at rest (Docker volume encryption, filesystem-level, or hardware-level).
  • Access-control (who can query VSS, who can retrieve raw clips).
  • Deletion / audit trail.

Lifecycle

Operation How
Update VSS Blueprint git fetch in the VSS repo, check out the new tag, restart via dev-profile.sh.
Update the LLM or VLM Change the model tag in the vLLM docker run, restart the container.
Rollback dev-profile.sh down + docker rm -f the vLLM containers; redeploy previous versions.
Re-index a corpus Restart VSS with the same corpus path — VSS re-processes as needed.

Common issues

Symptom Fix
NGC image pull fails Confirm Docker version is 28.3.3 ≤ version < 29.5.0.
VLM container OOM at startup Start VLM before LLM; vLLM needs free GPU at startup.
Deploy passes bad -H flag silently dev-profile.sh exits on unknown host profiles — verify the -H argument matches the platform (RTXPRO6000BW, DGX-SPARK, …).
Long TTFT at c8 on T2 Prefer c2–c4 for interactive workloads on the S tier.
NVFP4 MoE errors on public images Pin the Marlin backend (--moe-backend marlin + VLLM_NVFP4_GEMM_BACKEND=marlin) as the Lightning recipes do.
Camera capacity exceeded Move to a larger YOLO26 variant, add a GPU (M tier), or downsample the camera ingest rate.