TAO Toolkit - Data Services

TAO Data Services contains the source for data annotation, augmentation, auto-labeling, analytics, mining, image validation, and the lightweight API service used by NVIDIA TAO workflows.

This repository builds the nvidia-tao-ds Python package and the TAO Data Services containers. The developer launcher is tao_ds; the package console scripts run inside that container after the source tree is mounted at /workspace.

Contents

• Start Here
• Source Layout
• Requirements
• Supported Commands
• Container Builds
• Contributing
• License

Start Here

source scripts/envsetup.sh
tao_ds --gpus all --volume /data/tao:/data
tao_ds -- make build

Use docs/index.md as the detailed source map. It links to architecture, workflow, testing, container, and extension guides for maintainers, coding agents, and container power users.

Need	Document
Repository map and first-pass commands	docs/agent_onboarding.md
Command, config, API, and package flow	docs/architecture.md
Common source-change recipes	docs/development_workflows.md
Test selection and debugging notes	docs/testing_and_debugging.md
Launcher, mounts, GPU, and image details	docs/container_power_users.md
Adding a new data-service command	docs/new_data_service_command.md

Source Layout

Path	Purpose
nvidia_tao_ds/	Main package for CLI subtasks, configs, shared utilities, API service, and data-service domains.
runner/tao_ds.py	Host-side Docker launcher used by the tao_ds shell function from scripts/envsetup.sh.
docker/	Base development image Dockerfile, requirements, build script, and digest manifest.
release/	Python package metadata plus release-container build scripts.
ci/ and .gitlab-ci.yml	Static-test helpers and merge-request pipeline wiring.
tests/	Unit and integration-style tests for conversion, analytics, auto-label, mining, and config behavior.
tao-core/, tao-pytorch/	In-repo submodules used by local development and container builds.

Requirements

Software	Version
Ubuntu LTS	>=18.04
Python	>=3.8.x
Docker CE	>19.03.5
Docker API	1.40
nvidia-container-toolkit	>1.3.0-1
NVIDIA container runtime	3.4.0-1
nvidia-docker2	2.5.0-1
NVIDIA driver	>525.85
python-pip	>21.06

Minimum hardware is 8 GB system RAM, 4 GB GPU RAM, 8 CPU cores, one NVIDIA GPU, and 100 GB SSD space. Recommended hardware is 32 GB system RAM and 32 GB GPU RAM.

Supported Commands

Generated by python tools/update_readme_supported_commands.py from setup.py, runner/tao_ds.py, and docker/manifest.json.

tao_ds Container Launcher

scripts/envsetup.sh exports tao_ds as a shell function that runs runner/tao_ds.py. Commands after -- execute inside the container.

Option	Default	Purpose
--gpus	all	Comma separated GPU indices to be exposed to the docker.
--volume	[]	Volumes to bind.
--env	[]	Environment variables to bind.
--mounts_file	empty string	Path to the mounts file.
--shm_size	16G	Shared memory size for docker
--run_as_user	False	Flag to run as user
--tag	None	The tag value for the local dev docker.
--ulimit	None	Docker ulimits for the host machine.
--run_as_service	False	Flag to run as a microservice
--ip	0.0.0.0	Microservice ip address (e.g. 0.0.0.0).
--port	8000	Microservice port (e.g. 8000).
--port_mapping	8000	Port mapping for the micorservices port (e.g. 8000).
--no-tty	True	Set TTY

In-Container Console Scripts

These entrypoints are installed by the nvidia-tao-ds package. Script subtasks are discovered from each command package's scripts/ directory.

Command	Python entry point	Script subtasks
analytics	nvidia_tao_ds.data_analytics.entrypoint.analytics:main	analyze kpi_analyze validate
annotations	nvidia_tao_ds.annotations.entrypoint.annotations:main	convert merge qa_to_llava_annotation slice
augmentation	nvidia_tao_ds.augmentation.entrypoint.augment:main	generate
auto_label	nvidia_tao_ds.auto_label.entrypoint.auto_label:main	generate
embedding	nvidia_tao_ds.mining.embedding.entrypoint.embedding:main	image_embeddings text_embeddings
gap_analysis	nvidia_tao_ds.rcca.gap_analysis.entrypoint.gap_analysis:main	object_detection vcn_aoi vlm_bcq
image	nvidia_tao_ds.image.entrypoint.image:main	validate
tmm	nvidia_tao_ds.mining.tmm.entrypoint.tmm:main	nearest_neighbors unique_neighbor_matching

Base Image Source

runner/tao_ds.py and ci/utils.py resolve the immutable base image (nvstaging/tao/data_services_base_image) from docker/manifest.json, choosing the architecture-specific digest for the host. The pinned digests are intentionally not duplicated here — they live in docker/manifest.json (and the CI / Jenkins / release files), and a static CI check (ci/run_static_tests.py) verifies those digest references stay in sync.

Container Builds

The base development image is defined by docker/Dockerfile, Python requirements live in docker/requirements-pip.txt, and immutable image digests live in docker/manifest.json.

source scripts/envsetup.sh
cd "$NV_TAO_DS_TOP/docker"
./build.sh --build --x86
./build.sh --build --arm
./build.sh --build --multiplatform --push

The release image builds a wheel for nvidia_tao_ds and installs it through release/docker/Dockerfile.release.

source scripts/envsetup.sh
cd "$NV_TAO_DS_TOP/release/docker"
./deploy.sh --build --wheel

Contributing

TAO Toolkit Data Services is not accepting external contributions for this release stream. Internal changes should keep generated README content in sync by running:

python tools/update_readme_supported_commands.py --check

License

This project is licensed under the Apache-2.0 License.