TAO NVIDIA GPU Host Setup
Checks and standardizes NVIDIA drivers, CUDA, and container runtime prerequisites for TAO GPU hosts.
The documentation clearly separates read-only checks from user-approved installation and documents --yes, Docker opt-outs, and post-failure rechecks. However, installation requires sudo/root, adds repositories, installs system packages, may restart Docker, and by default adds the user to the docker group. There is no automatic rollback, complete change preview, or comprehensive dependency-source verification, so 7 points are deducted.
The script uses strict error handling, supports several distribution families, and provides multiple failure messages; the key workflow is broadly consistent statically. Execution was not performed, some prerequisite steps use || true, RHEL/SUSE driver selection is not strictly pinned to branch 580, and no complete test suite or recovery path is provided, so 2 points are deducted and the static ceiling applies.
The audience, Docker/Kubernetes scenarios, supported distributions, architectures, and unsupported cases are relatively clear, and the check path is broadly Linux-portable. However, core installation depends on NVIDIA/Docker overseas repositories and online documentation, mainland-China reachability is not addressed, and detailed non-fit boundaries and Chinese-language support are limited, so 5 points are deducted.
The skill includes structured metadata, Apache-2.0 licensing, version 0.1.0, ownership, a benchmark report, workflow guidance, failure modes, and references. It lacks the recommended Instructions and Examples sections, has no clear changelog or maintenance/update path, uses incomplete author formatting, and contains minor documentation/script inconsistencies, so 7 points are deducted.
The skill offers directly usable checks, installation commands, Docker runtime configuration, and Kubernetes capacity warnings. Static review cannot confirm package names, repository mappings, or successful end-state operation; the benchmark also contains only one task and no negative-trigger cases. The static ceiling therefore limits this to 7, with deductions for incomplete outcome evidence.
The committed script, evaluation task, benchmark report, and static findings provide a partially auditable trail. Coverage is thin: one task, one attempt, no committed suite covering key installation paths, and no independent corroboration, so 1 point is deducted.
- Installation modifies the host, may restart Docker, and grants docker-group privileges; use only with explicit approval on a controlled host and with a rollback plan.
- The RHEL/SUSE paths describe drivers as at least 580 rather than strictly pinning branch 580; verify compatibility on the target distribution.
- Installation depends on online NVIDIA and Docker repositories; mainland-China reachability, mirrors, and offline procedures are not addressed.
- Repository resolution, package versions, and successful installation were not reproduced in this static review.
What does this skill do, and when should you use it?
This skill prepares Linux hosts used by TAO Docker, local-Docker, and Kubernetes backends. After explicit user authorization, it can install NVIDIA driver branch 580, CUDA Toolkit 13.0, NVIDIA Container Toolkit 1.19.0, and Docker when required. Its default check-only path is read-only and works across Linux distributions, while automated installation targets documented Debian, RHEL, and SUSE families. It is a good fit for teams that need repeatable GPU worker prerequisites with an approval gate before system changes.
The skill runs a Bash setup script that detects the distribution family and checks nvidia-smi, the CUDA Toolkit path, the installed NVIDIA Container Toolkit version, and Docker's NVIDIA runtime. --check-only performs read-only probes and reports missing components. An authorized --install adds NVIDIA CUDA and Container Toolkit repositories, installs kernel headers or development packages, NVIDIA driver packages, cuda-toolkit-13-0, and NVIDIA Container Toolkit 1.19.0; for Docker backends it can install Docker, configure the NVIDIA runtime, enable or start the daemon, and add the invoking user to the docker group. It also attempts modprobe nvidia and prints verification commands, actionable failures, or manual-install guidance for unsupported distributions.
- A TAO user is preparing an Ubuntu, Debian, Fedora, RHEL, Rocky, AlmaLinux, openSUSE, or SLES GPU host and needs the required driver, CUDA, and container runtime installed.
- A platform engineer wants to validate a Docker GPU host before submitting TAO work without changing the system.
- A Kubernetes administrator is preparing each self-managed GPU worker before installing the NVIDIA GPU Operator or device plugin.
- An operator needs structured diagnostics for a failed nvidia-smi check, missing Docker NVIDIA runtime, or absent Kubernetes GPU capacity.
What are this skill's strengths and limitations?
- The default check path is safe and read-only and is portable across Linux distributions.
- Automated installation covers documented Debian, RHEL, and SUSE families with explicit package and version targets.
- It handles Docker runtime configuration, docker-group setup, kernel-module loading, and Kubernetes GPU-capacity warnings.
- Installation requires explicit user authorization, which supports controlled operations workflows.
- Installation requires sudo/root, internet access, and NVIDIA package repositories; Docker backends may also need Docker repositories.
- Automated installation does not cover every distribution; Arch, Alpine, Gentoo, NixOS, FreeBSD, and others require manual installation.
- On RHEL and SUSE families, the script uses the latest open driver supplied by the CUDA 13.0 repository rather than the same strict 580 package pin used on Debian-family systems.
- New docker-group membership does not affect the current shell until the user logs in again or runs newgrp docker.
- The documented quick-start script path differs in naming from the repository's stated top-level skill path, so the installed directory should be verified.
How do you install this skill?
Install the individual skill with the repository's documented skills CLI command:
npx skills add nvidia/skills --skill tao-setup-nvidia-gpu-host --yes
Omit --yes to choose the installation destination interactively. The README says the skill becomes available when the agent next loads skills and encounters a relevant task.
How do you use this skill?
Ask the agent to check or prepare the host, for example: “Check and prepare this host for the TAO Docker GPU backend” or “Check the TAO Kubernetes GPU worker runtime.” The documented check command is:
bash skills/platform/tao-setup-nvidia-gpu-host/scripts/setup-nvidia-gpu-host.sh --backend docker --check-only
Use kubernetes instead of docker for a Kubernetes worker. If the check fails, the workflow requires explaining the missing items and requesting authorization. After approval, use --install; for a non-interactive agent run, append --yes:
bash skills/platform/tao-setup-nvidia-gpu-host/scripts/setup-nvidia-gpu-host.sh --backend docker --install --yes
Rerun --check-only after installation and perform the documented nvidia-smi, CUDA 13.0, Docker runtime, and container GPU checks.