NVIDIA outlines GDS diagnostics and setup tools for GPU storage paths

NVIDIA said its GPUDirect Storage tooling is designed to help users check, configure and troubleshoot GPU storage paths across on-premises and cloud environments. In a blog post published Oct. 30, 2026, the company introduced gds-diag, a command-line utility and optional agent skill intended to assess deployments, surface configuration issues and suggest improvements.

Evidence and context

According to NVIDIA Technical Blog on Oct. 30, 2026, GPUDirect Storage (GDS) provides a data path between NVIDIA GPUs and storage targets to increase bandwidth, reduce latency and lower CPU and memory utilization. The company said gds-diag was developed by the GPUDirect Storage team as an open-source tool that can identify problems and recommend remediation steps, while leaving the actual system changes to the user or administrator.

The post says gds-diag requires a Linux system with Python 3.8 or newer. NVIDIA also noted that GDS itself requires an NVIDIA datacenter-class GPU and CUDA Toolkit 12.2 or newer, although those are not dependencies of gds-diag. The utility is distributed in the NVIDIA Magnum IO GitHub repository and can be used either as a command-line tool or through an agent skill embedded in an AI coding assistant.

NVIDIA said the command-line interface includes subcommands for different stages of deployment. The all command runs a full diagnostic sequence, while pre-install, support-matrix, post-install and config-audit cover installation planning, support validation and configuration review. The post also highlighted deeper checks such as container-check for container environments including containerd, Docker, Enroot and Kubernetes, and mount-check for a specific mount point or directory.

The company described gds-diag findings as status-based outputs labeled FAIL, WARN, INFO and PASS, with PASS shown only when the verbose option is used. It said the tool can emit reports as structured JSON for ingestion or archiving, and that most subcommands follow a common format. An exception is support-matrix, which returns a table with footnotes and can run in static or live mode to show supported platform combinations.

For the agent skill, NVIDIA said the layer is meant to route requests to the right gds-diag subcommand and interpret the results rather than duplicate deterministic checks already available in the CLI. The company said the skill was validated with Claude and Codex, and gave examples such as checking whether a system is eligible for GDS, whether GDS is available for a given mount point, and which NVMe device has the best locality to a specified GPU.

NVIDIA also said community contributions are welcome and that both the source code and the agent skill are published under open-source licenses: Apache 2.0 and Creative Commons Attribution 4.0 International.

The release follows another NVIDIA technical blog post dated Oct. 22, 2026, on reading data into GPU memory in Python with GPUDirect Storage. The company did not provide a new product launch date in the Oct. 30 post beyond the publication date of the guide itself.