Auto-Labeling

August 31, 2026 · View on GitHub

How to run Auto-Labeling: install it, pick a cookbook, wire endpoints and checkpoints, and read the outputs.

This guide is for external customers and for NVIDIA operators using the product. If you are changing the code, follow this guide once so you can run a cookbook, then switch to Developer Docs.

Typical commands:

  • make sync — install the workspace
  • make run SCRIPT=workflow-runner:main ARGS='...' — run a cookbook
  • make build IMAGE=<service>:<target> — build a stage image when needed

Read the pages in this order the first time. After that, use the table as a lookup.

  1. Installation — clone, make sync, confirm --help
  2. Model Provisioning — download the checkpoints your cookbook needs
  3. VLM and LLM Endpoints — stand up or point at a model server and export an API key
  4. Getting Started — copy a cookbook, dry-run, run
  5. Experiment Output Layout — find the results on disk

Open a service page only when you need one stage on its own. Use Remote Storage only for S3/GCS/Azure/HTTP paths. If something fails, use Troubleshooting.

Start Here By Goal

GoalPage
First end-to-end runGetting Started
Clone, sync, host requirements, and limitationsInstallation
Pick a sample cookbookSamples and Cookbooks
Run, debug, or understand one serviceServices
Compile and launch a cookbookOperations: Workflow Runner
Wire VLM and LLM endpointsVLM and LLM Endpoints
Download checkpointsModel Provisioning
Use S3/GCS/Azure/HTTP scene or media pathsRemote Storage
Understand the output tree, or run one service by handExperiment Output Layout
Fix a failed runTroubleshooting

What you can run from this guide

  • Video auto-labeling (video_data_augmentation)
  • Visual Attribute Search cookbooks
  • Image spatial grounding (2D grounding plus referring expressions)
  • Smart Spaces warehouse event-reasoning
  • Training export as a last stage or a standalone batch job

The workflow runner runs those stages locally, in order, on your machine.