tao-setup-nvidia-gpu-host

This NVIDIA skill prepares and standardizes a GPU host runtime for TAO workflows on Docker, local-Docker, or Kubernetes backends. It checks and — only after explicit user approval — installs a pinned stack: NVIDIA driver branch 580, CUDA Toolkit 13.0, and NVIDIA Container Toolkit 1.19.0, plus a Docker engine when missing on Docker backends. It solves the common friction of getting a machine's GPU driver, CUDA, container toolkit, and container runtime aligned before TAO training or inference runs.

The design cleanly separates a safe default from a privileged action. The `--check-only` path is read-only and works on any Linux distribution: it only probes `nvidia-smi`, the CUDA toolkit path, the installed container-toolkit package version, and the Docker daemon's NVIDIA runtime. The `--install` path must be explicitly authorized and automates debian-family (apt), rhel-family (dnf/yum, adding download.docker.com for Docker), and suse-family (zypper) hosts; unsupported distributions get a clear error naming the version targets and the official NVIDIA install-guide URLs. A driving shell script (`scripts/setup-nvidia-gpu-host.sh`) implements the check/install logic, and the skill ships an NVSkills-Eval BENCHMARK.md recording a PASS verdict with 100% security scores.

It targets developers and platform engineers standing up NVIDIA GPU hosts for TAO Docker or Kubernetes work. The install step requires sudo/root and internet access to NVIDIA (and, on RHEL, Docker) package repositories, and it modifies the host — so the skill gates it behind a preview via `--check-only` and explicit approval, and warns that the non-interactive `--yes` flag should only be used on a host you control. There is no credential exfiltration, no destructive wipe, and no remote code execution beyond installing official vendor packages from official repositories.