Verifies, configures, and troubleshoots the NVIDIA DOCA install and runtime environment on BlueField hosts.
npx skills add https://github.com/NVIDIA/skills --skill doca-setup
Please help me install this Agent Skill. Skill: doca-setup Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-setup Command: npx skills add https://github.com/NVIDIA/skills --skill doca-setup
Setting up a working NVIDIA DOCA environment on a host or BlueField DPU involves many fragile steps — install profiles, pkg-config paths, hugepages, devlink, representors, and kernel modules — and failures surface as confusing build or runtime errors. This Skill guides an agent through verifying install health, preparing the build and runtime environment, diagnosing env-class failures, and routing deployment-shaped questions to the correct container or bare-metal path.
Ask your agent: "I just got a BlueField and my DOCA code is built — walk me through verifying the install and running it."
High-intent search queries and answers about installing and using this skill.
Run the install-health snapshot: check /opt/mellanox/doca exists, confirm pkg-config --modversion doca-common returns a version, list *.pc files under the infrastructure pkgconfig directory, and run doca_caps. The skill walks these checks in install, capability, then runtime order.
Use the public NGC DOCA container at nvcr.io/nvidia/doca/doca, which runs on any OS with Docker. It ships the full doca-all profile with samples and headers for building and learning, but real hardware runtime requires graduating to a host with a NIC or DPU.
The most common causes are a missing PKG_CONFIG_PATH entry for /opt/mellanox/doca/infrastructure/lib/pkgconfig or a wrong install profile that never shipped doca-flow.pc. Check the pkgconfig directory first, then reinstall with the doca-all profile if the file is absent.
Containers via kubelet-standalone with a YAML pod-spec are the canonical path for packaged NVIDIA DOCA services on BlueField. Bare-metal launch via CLI, tmux, or systemd is the canonical path for a DOCA-linked application you wrote yourself. The skill's recognize workflow routes based on your system shape and workload kind.
Hugepages must be mounted and reserved, the BlueField device must appear in devlink dev show, representors must be visible under /sys/class/net, mlx5 kernel modules must be loaded, and the device mode must match what the program expects. All five checks must pass before debugging the program itself.
Do not use it for library API questions like Flow pipe construction or RDMA queue setup, for deriving a first application from a sample, or for DOCA_ERROR_* program-side debugging. Those belong to the doca-flow, doca-programming-guide, or doca-debug skills respectively.
freire19Create audit documentation and map codebase structure for pre-audit context.
toplyr-narfurDeploy Urbit planets, comets, and fake ships on Ubuntu/Debian systems.
AntonBespalovDetect No-Go safety issues in firmware changes affecting BKIN/BKIN2, STO, timers, and watchdogs.
air-gappedAutomate NVIDIA datacenter GPU bring-up on Ubuntu 24.04 with health validation.
