TAO Local Docker Runner
Run TAO jobs on a local or remote single-node Docker host with NVIDIA GPU support.
The documentation discloses Docker, remote DOCKER_HOST, GPU, S3, NGC_KEY, HF_TOKEN, and SSH-key data flows, and requires approval for some installations; however, allowed-tools includes broad Bash, Docker-daemon access is highly privileged, same-job replacement stops and removes an existing container, and there is no detailed launch confirmation, least-mount, secret-redaction, or rollback guidance. No malware, credential theft, or covert-exfiltration red line is evident.
The main workflow, preflight, GPU allocation, state mapping, logs, and failure modes are described, and one external evaluation report is included; however, static files do not establish reproducible key paths. TAO_SKILL_BANK_ROOT, TAO_SKILL_BANK_PATH, and helper default paths are inconsistent, while dependency installation, remote-path, and image-compatibility edge cases remain thin. The score is also constrained by static calibration.
The audience, local/remote Docker scenarios, and non-fit cases such as multi-node and SLURM scheduling are clear, with reasonably precise trigger phrases; however, input/output contracts, non-fit boundaries, and mount conventions are incomplete. The documentation is English-only and does not address mainland-China reachability for NGC, PyPI, SSH, or cloud storage dependencies.
The skill has frontmatter, version, compatibility, credential, monitoring, cancellation, and failure-mode notes, while the skill card identifies the owner and Apache 2.0 license; however, the benchmark records missing recommended Instructions and Examples sections, the author format is nonstandard, and no changelog, named maintenance responsibility, or explicit update path is shown. Repository license metadata is NOASSERTION, so points are deducted.
It provides a directly useful single-node TAO Docker workflow from preflight through launch, monitoring, and cancellation, with both SDK and docker-run options; however, only one positive evaluation task is present, reported correctness is about 47%/50% and effectiveness about 60%, and there are no representative verified outputs or broad failure-case results. User configuration and review remain substantial.
Revision-pinned source material, structured configuration, an evaluation report, and one eval definition provide some auditability; however, the evaluation has only one task and no negative cases, lacks independently reproducible execution artifacts, and contains no skill-specific tests covering the core Docker/GPU/S3 paths. Static evidence therefore supports only a limited score.
- Docker-daemon access, remote DOCKER_HOST, GPUs, S3 credentials, and HOST_SSH_PATH can widen host or data exposure; confirm the target host, mounts, image, credential scope, and GPU selection before execution.
- Replacement by the same job_id stops and removes an existing container; confirm that no other work is affected and preserve recoverable configuration, logs, and outputs.
- TAO_SKILL_BANK_ROOT and TAO_SKILL_BANK_PATH conventions are inconsistent, and key dependencies and image compatibility were not verified by this static review.
- Mainland-China users should separately verify reachability and compliance for PyPI, NGC, remote SSH, and the selected S3 endpoint.
What does this skill do, and when should you use it?
This skill guides an agent through running TAO SDK jobs in Docker containers backed by an NVIDIA GPU runtime. The Docker daemon may be on the agent host or on a remote GPU server exposed through DOCKER_HOST. It covers preflight checks, data and credential validation, GPU selection, container monitoring, logs, and cancellation. It is intended for development, debugging, and small single-node runs, not cluster scheduling, multi-node training, or SLURM queueing.
Checks Docker reachability, NVIDIA drivers, CUDA, NVIDIA Container Toolkit, GPU visibility and occupancy, dataset paths, model credentials, and host/container architecture compatibility; runs a GPU smoke container; creates the configured Docker network; starts detached containers named tao-job-<job_id> through docker run or the Docker Python client; mounts /dev/shm; maps container states; retrieves logs; and stops containers for cancellation. The optional TAO SDK wrapper adds Job handles, S3 input/output wrapping, and ActionWorkflow durability.
- A TAO developer running training or inference containers on a local NVIDIA GPU workstation and needing runtime validation first.
- A laptop or workstation submitting a job to a single remote GPU server through DOCKER_HOST or a Docker context.
- A debugging workflow requiring direct container status and log inspection.
- A shared single-host setup assigning multiple GPUs with Docker --gpus or the SDK gpu_count parameter.
- A TAO SDK workflow that needs Job handles or automatic S3 input downloads and output uploads.
What are this skill's strengths and limitations?
- Supports both local Docker and remote Docker daemons exposed through DOCKER_HOST.
- Includes checks for GPU runtime readiness, data paths, credentials, GPU occupancy, and architecture compatibility.
- Supports multiple GPUs on one host and documents container status, logs, and cancellation.
- The optional TAO SDK provides Job handles, S3 I/O wrapping, and cross-session durability.
- Explicitly does not support multi-node training, remote cluster scheduling, or SLURM queueing.
- Depends on specific NVIDIA driver, CUDA, Docker, and NVIDIA Container Toolkit versions.
- In remote mode, filesystem paths in job specifications are interpreted on the remote Docker host.
- The supplied material does not document a test suite or standalone test results for this individual skill.
How do you install this skill?
Install the individual skill with the repository README's skills CLI command: npx skills add nvidia/skills --skill tao-run-on-local-docker --yes. The README does not document additional manual installation steps for this skill; after installation, it becomes available when the agent loads skills and encounters a relevant task.
How do you use this skill?
Ask the agent to “run locally,” “use local Docker,” “run remote Docker,” or “use my GPU.” Before launching, run the host preflight with bash "$SETUP_SCRIPT" --backend docker --check-only and confirm that docker info and docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi succeed. For remote Docker, set DOCKER_HOST=ssh://user@host or use a Docker context, and pass paths that exist on the remote Docker host to the remote-docker preflight. For the SDK mode, install nvidia-tao-sdk[docker], create a job with DockerSDK().create_job(...), then use get_job_status and get_job_logs.
How does this skill compare with similar options?
The skill documents two execution modes: direct docker run, which does not require the TAO SDK, and the optional TAO SDK Docker wrapper, which adds Job handles, S3 I/O, and ActionWorkflow durability. It also directs users toward the platform that owns a remote filesystem, such as SLURM for Lustre paths.