What does this skill do, and when should you use it?
This skill is for external operators bringing up a new BlueField-4 DPU through its BMC. It covers installing the BlueField/DOCA bundle ISO onto Grace, updating BMC, NIC firmware, SBIOS, and ERoT through PLDM, and installing a Grace Ubuntu image with optional cloud-init. The documented installation methods are UEFI HTTP Boot, PXE Boot, and Redfish Virtual Media. Because the workflows can include irreversible firmware burns, ISO reflashes, power cycles, and BMC factory resets, the skill requires doca-hardware-safety, a maintenance window, a tested rollback plan, and action-specific confirmation.
Guides agents through BlueField/DOCA bundle ISO installation using UEFI HTTP Boot, PXE Boot, or Redfish Virtual Media; uploads .fwpkg bundles through the Redfish UpdateService multipart endpoint and monitors or verifies PLDM updates with Redfish Task, FirmwareInventory, and pldmtool; installs Grace Ubuntu and optional CIDATA cloud-init configuration through Redfish Virtual Media; reaches the out-of-band serial console using BMC SSH and obmc-console-client; checks /etc/mlnx-release; and diagnoses boot-source, virtual-media, firmware-task, activation, cloud-init, and boot-loop failures.
- An operator needs to install Grace OS on a fresh BlueField-4 through the BMC.
- An administrator needs to update BMC, NIC firmware, SBIOS, or ERoT through Redfish UpdateService during a maintenance window.
- An engineer needs to choose among UEFI HTTP Boot, PXE Boot, and Redfish Virtual Media.
- An operator is investigating a firmware Task stuck at Running, reporting an Exception, or failing to activate a pending image.
- An administrator needs to install Grace Ubuntu with an optional cloud-init configuration ISO.
- A field engineer needs the out-of-band serial console to investigate ISO boot, virtual-media, or boot-loop problems.
How do you install this skill?
- Firmware, reflashing, power-cycle, and BMC-reset actions can make the device unavailable or brick it; load doca-hardware-safety first, verify target versions, OOB access, maintenance window, and a testable rollback plan, and obtain confirmation for each action.
- Do not treat placeholders, example URIs, EIDs, task IDs, or command parameters as site values; verify Redfish behavior, dpu-bmc support, firmware layout, and release targets against current documentation for the target device.
- Chinese usability and mainland-China network reachability are not demonstrated; when NVIDIA download or documentation services are required, validate reachability and prepare a compliant local alternative.
- Shell / CLI
- Network access
- Local filesystem
BlueField-4 DPUout-of-band reachable BMCBMC SSHRedfish endpointobmc-console-clientpldmtoolHTTP/HTTPS server
Install the specific skill with the skills CLI command documented by the repository README:
npx skills add nvidia/skills --skill doca-bf4-deployment --yes
The source does not document another dedicated installation method for this skill. After installation, the skill becomes available when an agent loads skills and encounters a relevant task; manual repository cloning or folder copying is not required.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/doca-bf4-deployment" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- bring up my new BlueField-4,
- the BlueField ISO will not boot over HTTP from the BMC,
- attach BF4 virtual media via Redfish,
- BF4 firmware Task stuck at Running.
After installation, trigger it with requests such as “bring up my new BlueField-4,” “the BlueField ISO will not boot over HTTP from the BMC,” “attach BF4 virtual media via Redfish,” or “BF4 firmware Task stuck at Running.” Before any PLDM burn, ISO reflash, power cycle, or BMC factory reset, load doca-hardware-safety alongside it, show the exact command and blast radius, and obtain explicit confirmation for that specific action. Use CAPABILITIES.md for the bring-up contract and TASKS.md for procedural workflows.
What are this skill's strengths and limitations?
- Covers the principal BlueField-4 day-one bring-up paths.
- Combines OS installation, PLDM firmware updates, console observability, and layered troubleshooting.
- Clearly separates BF4 bring-up from BF3, application deployment, environment preparation, and fleet orchestration.
- Defines explicit safety controls for operations that may damage hardware or cause outages.
- Limited to BlueField-4; BlueField-3 is out of scope.
- Does not cover application launch, DOCA tooling installation, library APIs, or post-install environment preparation.
- Assumes an accessible BMC, operator-hosted image files, an HTTP/HTTPS server, and target version information.
- SKILL.md is a thin loader whose substantive guidance depends on companion files such as CAPABILITIES.md, TASKS.md, and references/details.md.
- The supplied source provides no specific hardware test matrix or standalone test-suite evidence.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| DOCA BlueField-4 Deployment this page ✓ NVIDIA · Official | 57 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Hardware Safety ✓ NVIDIA · Official | 57 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Socket Relay ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA BlueField-3 Bring-Up ✓ NVIDIA · Official | 56 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA BlueField Bare-Metal Deployment ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
The skill explicitly covers three installation alternatives: recommended UEFI HTTP Boot, PXE Boot, and Redfish Virtual Media. It provides no comparison with non-NVIDIA vendors or unrelated tools.
How did FollowSkills review this skill?
The document explicitly identifies the irreversible risks of firmware burns, ISO reflashes, power cycles, and BMC resets; requires a maintenance window, rollback plan, per-action confirmation, credential redaction, placeholders, and verified OOB access. However, rollback details are delegated to the external doca-hardware-safety skill, while least privilege, credential lifecycle, data-flow disclosure, and dependency security remain incomplete, so points are deducted.
The workflow is internally organized with prerequisites, six diagnostic layers, task-state checks, version checks, and media-state checks. However, key Redfish endpoints, compatibility details, and environment assumptions depend on external documentation; there are no executable tests covering the skill itself and failure guidance remains partly procedural. Under the static cap, reliability cannot exceed 10, so 8 is awarded conservatively.
The audience, BF4 day-1 bring-up scenarios, trigger phrases, and non-fit boundaries for BF3, applications, and fleet provisioning are clearly stated. However, Chinese-language support and mainland-China reachability are not addressed, and core use depends on public NVIDIA download and documentation surfaces, so points are deducted.
The skill has a useful information architecture, loading order, task-verb organization, related-skill routing, risk notes, dependency notes, Apache-2.0 metadata, and NVIDIA ownership signals. It lacks an author field, a formal changelog, and a clear maintenance/update path; BENCHMARK also reports missing recommended Instructions and Examples sections, and the skill-card's CC-BY-4.0 statement is not fully aligned with SKILL.md metadata. Points are therefore deducted.
The files cover three installation methods, PLDM updates, cloud-init, verification, and layered debugging, with a clear operational goal and potentially useful outputs. However, the skill is explicitly a thin loader, depends on companion files, live target versions, and external public documentation, and provides no statically verifiable representative result or comparison with alternatives. The static ceiling is 7, so 6 is awarded.
The committed benchmark report, static-validation findings, evaluation metrics, and traceable command/state checks provide limited auditability. The evaluation dataset is unavailable, there is no real CI plus committed key-path test suite, and standards/documentation are referenced only generally, preventing independent reproduction; 4 points are awarded.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →