Installation

Install Auto-Labeling so you can run a cookbook. This is the product install path for customers and operators.

If you are changing the code, do this install first, then continue with Local Development.

Requirements

Install only what the selected workflow needs.

Item	Requirement
OS	Linux-like host with bind mounts and a container runtime
Python	3.12 or newer
Package manager	uv
Build/launcher glue	GNU Make
Container runtime	Docker or Podman for cookbook execution
GPU runtime	NVIDIA GPU, NVIDIA driver, and NVIDIA Container Toolkit

Workflow	Extra requirements
Workflow-runner dry run	Docker or Podman
Remote VLM/LLM-backed runs	Reachable endpoint URLs and matching credentials
GPU-backed stages	Local checkpoints plus NVIDIA Container Toolkit
Remote storage	Multi-Storage Client configuration

GPU stages need the NVIDIA Container Toolkit (or an equivalent GPU runtime) so stage containers can see the GPU.

ffmpeg / ffprobe on the host is required to chunk the NGC VSS sample clips and is useful for other local validation. SAM3, SeedVR2, and RF-DETR checkpoints are required only for the stages that use them. See Model Provisioning.

Clone And Sync

git clone <repo-url>
cd paidf-auto-labeling
make sync

make sync installs the workspace packages needed to run services and cookbooks.

Copy tracked cookbooks to *.local.yaml before you edit paths. Those *.local.yaml files are gitignored so machine-specific paths and endpoints do not get committed.

Verify The Install

Confirm the workflow-runner entrypoint resolves:

make run SCRIPT=workflow-runner:main ARGS='--help'

Then confirm the host can compile a cookbook plan without running containers:

make run SCRIPT=workflow-runner:main \
  ARGS='--cookbook-file cookbooks/video_data_augmentation/configs/pipeline_video.yaml --container-dry-run'

A successful dry-run only means the runner and container runtime can parse that cookbook. It is not a real labeling run. Tracked configs contain placeholders for media, checkpoints, and endpoints.

You still need, before Getting Started:

1. Checkpoints for the stages you will run — Model Provisioning
2. A VLM/LLM endpoint and an API key in your shell — VLM and LLM Endpoints
3. Staged NGC traffic/warehouse clips for a real sample run — Samples and Cookbooks
4. Remote storage config only if media or outputs are not on local disk — Remote Storage

Never put API keys or cloud credentials in cookbook YAML. Export them in the shell and pass the variable name with --container-env.

Limitations

• workflow-runner is a local sequential launcher, not a scheduler.
• grounding_2d and referring_expressions are image-only.
• Person Attribute Search is an assembly stage: it needs upstream sidecars or explicit attribute JSON.
• reasoning-service documents openai-compatible as its LLM provider surface.
• Secrets belong in the environment or the execution platform, not in tracked configs.
• The repo does not publish a qualified performance matrix. Time a short representative asset on your cookbook and serving stack before you scale.

Media Policy

• Supported input video codecs are H.264, VP9, and MPEG-4 Part 2 only.
• H.264 requires the NVIDIA CUVID hardware decoder.
• Generated videos are standardized on VP9 output.