doca-bf4-deployment

par nvidia

AVERTISSEMENT : les guides impliquent potentiellement des opérations matérielles IRRÉVERSIBLES sur BlueField-4 (gravures de firmware PLDM, reflashages ISO, cycles d'alimentation, réinitialisations d'usine BMC) qui peuvent rendre l'appareil inutilisable…

npx skills add https://github.com/nvidia/skills --skill doca-bf4-deployment

DOCA BlueField-4 (BF4) deployment

⚠️ WARNING — irreversible hardware operations. This skill guides operators through potentially destructive, irreversible BlueField-4 hardware operations: PLDM firmware burns, ISO reflashes, power cycles, and BMC factory resets. These can brick firmware, corrupt boot media, or cause production outages. Do not proceed without a maintenance window and a tested rollback plan. Every mutating step is governed by doca-hardware-safety, which MUST be loaded alongside this skill before any destructive action.

Before executing any mutating step — PLDM firmware burn, ISO reflash, power cycle, or BMC factory reset — the agent MUST show the exact command and its blast radius (which device, what becomes unavailable, whether it is reversible) and obtain the user's explicit confirmation for that specific action. Never chain destructive steps or run them speculatively as a side effect of another task.

Where to start: This skill is the bundle's deliberate in-bundle home for day-1 platform bring-up of a BlueField-4 DPU via the BMC — getting a powered-but-bare BF4 to "Grace OS installed, firmware at the target level, ready to deploy a workload." It is the upstream of the two application-deployment skills (doca-container-deployment and doca-bare-metal-deployment): those skills assume a working BlueField; this skill is how the BlueField-4 GETS to working. If the user has a fresh BF4 and wants to install the OS or update firmware, open TASKS.md and start at ## configure. If the question is what bring-up methods even exist and what is the contract for each, start at CAPABILITIES.md.

Scope note — BF4 day-1 is in scope by directive. The bundle's AGENTS.md ## Non-goals item 7 lists the BlueField BSP / BFB / RShim / TMFIFO layer and the BlueField BMC software as externally-productized. BlueField-4 day-1 bring-up via the BMC is carved into scope for this skill by directive because day-1 has no other home in the bundle. The carve-out is narrow: this skill teaches the documented BMC-driven install and firmware-update FLOWS (the CLASS), routing every mutating step through doca-hardware-safety for the change-application meta-policy. It does NOT redefine that meta-policy, and it does NOT cover BF3 (route to doca-bf3-deployment), application launch, or library APIs.

Audience

This skill serves external operators standing up a new BlueField-4 who already have:

  • a BlueField-4 with its BMC reachable out-of-band (BMC SSH plus the documented Redfish endpoint), so the DPU can be driven without physical access,
  • the BlueField/DOCA bundle ISO (and, for the Grace-Ubuntu path, a Grace Ubuntu image) downloaded from the public NVIDIA download surface, hosted at {iso-uri} on the operator's own HTTP/HTTPS server, and
  • the target firmware and OS versions read from the public BlueField/DOCA release notes (this skill never quotes a specific pre-release firmware version).

It is not for:

  • BlueField-3 (BF3) bring-up — route to doca-bf3-deployment,
  • developers who want to RUN a DOCA service container or a DOCA-linked binary on an already-working BlueField — route to doca-container-deployment or doca-bare-metal-deployment,
  • the cross-cutting hardware-change meta-policy itself (preflight, OOB console discipline, maintenance window, rollback) — that is owned by doca-hardware-safety and this skill cross-links it, never duplicates it,
  • fleet-scale / orchestrated DPU provisioning — that is DOCA Platform Framework territory, routed via doca-public-knowledge-map.

The skill teaches the agent the documented bring-up procedure and the rules for quoting Redfish / PLDM / UEFI standard operations and public BlueField/DOCA documentation via doca-public-knowledge-map; it does not invent BMC credentials, ISO URIs, firmware version strings, EIDs, Redfish task IDs, or device names from memory.

When to load this skill

Load this skill when the user is doing hands-on day-1 bring-up of a BlueField-4 via the BMC, or asking a cross-cutting BF4-bring-up question that is not specific to a later application-deployment step. Concretely:

  • Installing the BlueField/DOCA bundle ISO onto the DPU (Grace) for the first time, and choosing between the three documented install methods — UEFI HTTP Boot (recommended), PXE Boot, or Redfish Virtual Media.
  • Running the PLDM firmware-update flow across the BMC / NIC firmware / SBIOS / ERoT components: pushing the .fwpkg bundle through the Redfish UpdateService multipart endpoint, monitoring the returned Task, verifying pending images with pldmtool, and activating with a power cycle.
  • Installing a Grace Ubuntu image (with optional cloud-init via a CIDATA-labelled config ISO) through Redfish Virtual Media, with either local hosting on the BMC eMMC or remote hosting on an HTTPS server.
  • Reaching the DPU's OOB serial console (BMC SSH plus obmc-console-client) to watch the installer or UEFI menus.
  • Diagnosing a bring-up that is misbehaving — the ISO will not boot, virtual media will not attach, a firmware Task hangs or reports an Exception, a pending image never activates, cloud-init is ignored, or the DPU is stuck in a boot loop because media was never detached.
  • Cross-cutting questions: "HTTP Boot or Redfish Virtual Media — which do I use, and when do I actually need PXE?", "how do I know the firmware update actually took effect?", "the ISO landed but the NIC firmware update sub-step seems to have failed — what now?".

Do not load this skill for BF3 bring-up (route to doca-bf3-deployment); for running an application on an already-working BlueField (route to doca-container-deployment or doca-bare-metal-deployment); for env preparation on the installed Grace OS such as hugepages / pkg-config / devlink (use doca-setup); for the cross-cutting hardware-change meta-policy (route to doca-hardware-safety); or for fleet-scale orchestrated provisioning (route via doca-public-knowledge-map).

What this skill provides

This is a thin loader. Substantive material lives in two companion files:

  • CAPABILITIES.md — the BF4 day-1 bring-up contract: the three OS install methods (UEFI HTTP Boot, PXE Boot, Redfish Virtual Media) and the Grace-Ubuntu-plus-cloud-init Virtual Media path; the PLDM firmware-update surface across BMC / NIC firmware / SBIOS / ERoT; the version-compatibility overlay on doca-version (the install and firmware targets come from the public release notes, never from memory); the bring-up error taxonomy (boot-source -> virtual-media-attach -> firmware-Task -> activation -> cloud-init -> boot-loop); the observability surface (the OOB console, the Redfish Task resource, the Redfish FirmwareInventory, pldmtool GetFwParams, and the installed-build check cat /etc/mlnx-release); and the safety policy (an overlay on doca-hardware-safety: every PLDM burn / ISO reflash / power cycle / BMC factory reset is a MUTATING hardware op; never print a real password; always detach virtual media to avoid boot loops; only public hosts for any NVIDIA URL).
  • TASKS.md — step-by-step workflows for the in-scope bring-up verbs: configure, build (routing stub), modify (routing stub), run (the three install methods plus the PLDM firmware-update flow plus the Grace-Ubuntu cloud-init path, as ### sub-anchors), test (the post-install / post-update verification sweep), debug (the layered bring-up diagnosis), and the Deferred task verbs block routing BF3 / application-launch / library-API / env-prep / hardware-meta-policy / fleet questions out to their owning skills.

The skill assumes a BlueField-4 target where:

  • the BMC is reachable out-of-band and the operator has BMC credentials ({bmc-user} / {bmc-password}) they supply — never invented here,
  • the bundle ISO (and any Grace Ubuntu image / cloud-init config ISO) is downloaded from the public NVIDIA download surface and hosted at {iso-uri},
  • the operator has read the target firmware and OS versions from the public BlueField/DOCA release notes.

It does not cover installing DOCA tooling on the host — that path goes through doca-setup — and it does not cover running a workload once Grace is up — those paths go through the two application-deployment skills.

Loading order

  1. Read this SKILL.md first to confirm the user's question is in scope (BF4 day-1 bring-up via the BMC; NOT BF3, NOT application launch, NOT a library-API question, NOT the hardware-change meta-policy itself).
  2. For the bring-up contract (the three install methods, the Grace-Ubuntu cloud-init path, the PLDM firmware-update surface, the version overlay, the bring-up error taxonomy, the observability surface, and the BF4 safety overlay), see CAPABILITIES.md.
  3. For step-by-step workflows — configure, build (routing stub), modify (routing stub), run (with the three install methods, the PLDM flow, and the Grace-Ubuntu cloud-init path as ### sub-anchors), test, debug, plus the Deferred task verbs block — see TASKS.md.
  4. Load doca-hardware-safety ALONGSIDE whenever the question reaches a mutating step (PLDM firmware burn, ISO reflash, power cycle, BMC factory reset).

Example questions this skill answers well

See references/details.md.

What this skill deliberately does not ship

See references/details.md.

Related skills

See references/details.md.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…