NVIDIA GPU Docker Runtime Guide
Reusable Docker conventions for authenticating with NGC, exposing NVIDIA GPUs, mounting data, inspecting containers, and diagnosing GPU workload failures.
The document recommends environment-variable credential passthrough, avoiding secrets on command lines, and user approval for host installation. However, Bash access is broad, and it presents high-impact operations such as docker system prune, sudo Docker data-root migration and deletion, and stopping/removing containers without consistently requiring confirmation, backup, or rollback. Credentials persist in ~/.docker/config.json, and data-flow and sensitive-data cleanup guidance is incomplete, so points are deducted.
The skill covers common workflows and error diagnosis, but the canonical docker run example places spaces and comments after continuation backslashes, which may break shell continuation. The ubuntu smoke example also does not establish that nvidia-smi is present. The dependency on tao-setup-nvidia-gpu-host and tightly pinned versions is brittle, and no committed tests cover this skill's key paths; the score is therefore conservative under static-review limits.
Trigger terms, audience, GPU-Docker scenarios, and scope boundaries are reasonably clear, including separation from platform, model, and data skills. Nevertheless, the fixed requirements for driver 580, CUDA 13.0, and Toolkit 1.19.0 may exclude many environments, with no compatibility strategy for other versions or Chinese-language guidance. NGC, Hugging Face, S3, and W&B dependencies may also have mainland-China reachability constraints, so points are deducted.
The document is layered across prerequisites, authentication, execution, mounts, inspection, networking, disk management, and error modes, and provides Apache-2.0 licensing, a version, and a skill card. It lacks explicit Instructions/Examples sections, a changelog, clear maintenance ownership, and an update path; the included benchmark also reports schema, hierarchy, and author-format findings, so full marks are not justified.
The skill provides directly reusable command patterns, mount conventions, and troubleshooting guidance for NVIDIA GPU Docker workloads, with a clear core benefit. However, it does not select the image or business command, contains an execution-related shell-format concern, requires review for destructive operations, and cannot be runtime-validated here; the score remains within the static cap.
The skill cites Docker and NVIDIA primary documentation and includes one external evaluation report with a documented task. The evaluation has only one plan-description task, no negative-trigger coverage, and no committed tests for this skill's key paths. Static inspection is not independent execution verification, so the evidence remains limited.
- Fix and validate the backslash-and-inline-comment continuation examples before execution, and confirm that the smoke-test image actually contains nvidia-smi.
- Add explicit confirmation, backup, and reversible recovery steps for docker system prune -a --volumes, sudo data-root migration/deletion, and stop/rm operations.
- NGC/Hugging Face and other external services, together with pinned versions, may create mainland-China reachability or environment-compatibility issues.
What does this skill do, and when should you use it?
This skill is for AI agents that must run Docker workloads on NVIDIA GPU hosts. It covers NGC authentication, GPU selection, shared-memory sizing, data and result mounts, environment passthrough, container inspection, and Docker storage relocation. It also documents common failures involving GPU runtime setup, authentication, disk space, permissions, and startup crashes. Its scope is the “how” of container execution; other model and data skills specify the image and workload command.
Checks Docker version, GPU host readiness, and NGC_KEY; invokes docker login, docker pull, docker run, docker exec, docker inspect, docker logs, and docker stats; applies --gpus, --shm-size, bind mounts, environment passthrough, labels, and Docker networks; documents moving Docker data-root to a larger volume; and maps GPU, authentication, storage, shared-memory, permission, and container-exit errors to troubleshooting actions.
- An AI engineer needs to pull an nvcr.io image and run training or inference on an NVIDIA GPU host.
- A multi-GPU PyTorch workload fails with Bus error because the default container shared memory is too small.
- A cloud GPU host has a small root volume and needs Docker image storage moved before pulling large NVIDIA images.
- An agent must perform download, execution, and post-processing steps in one persistent container.
- An operator needs to inspect container mounts, environment variables, commands, networking, logs, or exit codes.
What are this skill's strengths and limitations?
- Covers the operational path from NGC authentication through container cleanup, inspection, and troubleshooting.
- Provides copyable command patterns for GPUs, mounts, shared memory, and environment variables.
- Clearly distinguishes host paths from container paths and documents split-disk Docker data-root relocation.
- Maintains a clear boundary as the generic Docker execution layer for other model and data skills.
- Requires specific NVIDIA driver, CUDA Toolkit, Docker, and NVIDIA Container Toolkit versions.
- Pulling nvcr.io images requires an NGC API key and network access.
- Some procedures use sudo, rsync, and removal of the old Docker data directory, requiring elevated privileges or careful handling.
- The supplied SKILL.md provides no test suite or platform-coverage evidence.
How do you install this skill?
Install the individual skill with the skills CLI command documented in the repository README:
npx skills add nvidia/skills --skill tao-run-on-docker --yes
The README does not document a manual copy destination. After installation, the skill becomes available when the agent loads skills and encounters a relevant task.
How do you use this skill?
With the skill installed, ask an agent for a relevant Docker GPU workflow, for example: “Run an nvcr.io container on an NVIDIA GPU host using GPUs 0 and 1, mount the data directory, and troubleshoot shared-memory issues.” The skill activates for topics such as docker run, nvcr.io, NGC, --gpus, or nvidia-container-toolkit. The calling model or data skill must provide the specific image and workload command.
How does this skill compare with similar options?
Unlike the adjacent tao-run-on-brev and tao-run-platform skills, this skill focuses on generic Docker conventions on a GPU host. tao-run-on-brev covers running Docker through brev exec, while tao-run-platform adds Job handles, state persistence, and S3 I/O.