DeepStream Import Vision Model

Automated end-to-end pipeline: HuggingFace model → TensorRT engine → DeepStream multi-stream benchmark → PDF report.

Overview

This self-contained skill uses four phase-specific reference documents. Together they automate the full model bringup workflow for NVIDIA DeepStream, from downloading a model on HuggingFace or NVIDIA NGC to a publication-ready benchmark report.

Supported input formats: ONNX (direct), SafeTensors (auto-exported via torch.onnx.export).

Current scope: object detection models only. Classification, segmentation, pose estimation, and other vision tasks are not yet supported — the pipeline fails fast if a non-detection architecture is detected in config.json.

Prerequisites

Host: only Docker + the NVIDIA driver. Everything else runs inside the DeepStream container — DeepStream, TensorRT/trtexec, the Python export venv (torch/onnx/onnxruntime), wkhtmltopdf, deepstream-app/gst-launch-1.0 — and is bootstrapped by setup.sh. Nothing is installed on the host. Runs identically on Linux and Windows (Docker Desktop + WSL2 backend, required for GPU).

Docker — Docker Desktop with the WSL2 backend on Windows; Docker Engine + NVIDIA Container Toolkit on Linux.
NVIDIA GPU + driver (on Windows, the WSL2 GPU driver — no host CUDA/TensorRT/DeepStream needed).
docker pull nvcr.io/nvidia/deepstream:9.1-triton-multiarch, then run setup.sh through the container.

See references/windows.md for the cross-platform run model and the per-shell docker run mount token.

Installation

bash <path-to-deepstream-import-vision-model>/install.sh --target <your-project-path>

Preview what will be installed first with --dry-run:

bash <path-to-deepstream-import-vision-model>/install.sh --target <your-project-path> --dry-run

Where <path-to-deepstream-import-vision-model> is the location of this skill in your repo, e.g.:

In team-mind-hub: team-skills/deepstream-sdk/deepstream-import-vision-model
In ds-copilot: team-skills/deepstream-sdk/deepstream-import-vision-model (same path)

The script copies the complete skill into the target project for Claude Code, Codex, and Cursor. Re-running it safely refreshes an existing installation.

Usage

Claude Code:

Use deepstream-import-vision-model to run this model: https://huggingface.co/onnx-community/yolov8n

Codex:

Use $deepstream-import-vision-model to deploy and benchmark https://huggingface.co/onnx-community/yolov8n

Cursor:

@deepstream-import-vision-model run this model: https://huggingface.co/onnx-community/yolov8n

The skill runs the full pipeline autonomously — no manual steps required.

Pipeline Steps

Step	Phase reference	Action
1–3	`references/model-acquire.md`	Browse HF repo, download ONNX or export SafeTensors
4–5	`references/engine-build.md`	Build dynamic TRT engine, run trtexec benchmarks
6–7	`references/pipeline-run.md`	Custom bbox parser, DeepStream single + multi-stream
8	`references/report-generation.md`	5 charts, HTML report, PDF

Output Structure

Per-model outputs are written to models/<model_name>/ in your project:

models/<model_name>/
  model/          ONNX file(s)
  parser/         Custom nvinfer bbox parser (.cpp, .so)
  config/         nvinfer config, DS app config, labels.txt
  scripts/        Model-specific run helpers
  benchmarks/     TRT engines, trtexec logs
  reports/        benchmark_report.md / .html / .pdf + charts/
  samples/        Output videos, test frames, KITTI detections

Files in this package

deepstream-import-vision-model/
├── SKILL.md                    Top-level skill definition
├── README.md                   This file
├── references/                 Phase-specific runbooks
├── scripts/                    Utility scripts by pipeline phase
└── tests/                      Installer and script regression tests

/deepstream-import-vision-model

README.md