tao-run-on-local-docker

This NVIDIA platform skill runs TAO SDK job containers on a Docker daemon with the NVIDIA GPU runtime, either locally or on a remote GPU box exposed through DOCKER_HOST. It solves the execution-target problem for TAO workflows: a coding agent needs a consistent way to submit training, evaluation, and inference jobs as named Docker containers, whether the daemon is on the same machine or on a remote single-GPU server, with the understanding that in remote mode all filesystem paths in specs are interpreted on the remote host.

The skill emphasizes preflight verification and safe defaults. Before starting jobs it checks the host GPU runtime (NVIDIA driver branch 580, CUDA 13.0, NVIDIA Container Toolkit 1.19.0) via a setup script in check-only mode, verifies the Docker daemon is reachable, and confirms the container toolkit works by running nvidia-smi in a throwaway container. If a check fails, it prompts the user to approve and run the install command, then reruns preflight. It documents two modes: a direct-docker mode needing only Docker and the GPU runtime, and a TAO SDK wrapper mode (pip install nvidia-tao-sdk[docker]) that adds Job handles, S3 I/O wrapping, and ActionWorkflow durability, plus idempotent creation of a Docker network for job containers. It requires no platform credentials beyond Docker daemon access, and documents optional environment variables such as DOCKER_HOST, DOCKER_NETWORK, and registry credentials. A bundled benchmark reports a passing NVSkills-Eval verdict including a 100% security score.

Target users are ML engineers running TAO jobs for development, debugging, and small runs, or submitting jobs from a laptop to a remote GPU server. It is explicitly not for multi-node training, SLURM queueing, or cluster scheduling.