GDS diagnostics with gds-diag

NVIDIA’s guide introduces gds-diag for diagnosing GPUDirect Storage, the data path between storage and GPUs. The command-line tool identifies configuration problems and suggests mitigations; it does not change the system itself. An optional agent skill selects diagnostic commands and interprets their results. The distinction matters for administrators: identifying a problem and applying its remedy are separate operations. Findings distinguish hard blockers from potential blockers and explain the proposed mitigation. An agent can help implement changes, with confirmation governed by the coding assistant’s local settings and site policies.

The repository requires Linux and Python 3.8 or newer, without third-party Python packages. Most commands need CUDA and GDS packages, while deeper diagnostics need sudo. The pre-install command works before GDS installation; support-matrix and config-audit can run without a GPU. Start the general diagnostic sequence with all, optionally supplying a filesystem path; otherwise it uses the current directory. For narrower questions, mount-check investigates a path, post-install validates the runtime, and config-audit examines configuration. JSON output supports processing reports elsewhere. Inside the checkout, Claude Code and Codex discover the shared skill without a separate installation.