What does this skill do, and when should you use it?
This skill is for developers using the DOCA DPA library from a BlueField host program. It covers the host-side lifecycle for a doca_dpa context, dpacc-produced doca_dpa_app image, DPA threads, kernel launches, arguments, and completions. It also covers capability discovery, DOCA/DPACC compatibility, and doca_dpa error diagnosis. It does not replace DOCA installation guidance or the dedicated guides for writing DPA-side kernels, DPA-Comms, or DPA-Verbs code.
Guides the agent to inspect DOCA files under /opt/mellanox/doca, read pkg-config --modversion doca-dpa and the installed dpacc version, create a doca_dpa Core context, load a doca_dpa_app image, create doca_dpa_thread contexts, launch kernels through the doca_dpa_kernel_launch_update_* family, drain doca_dpa_completion, query doca_dpa_cap_* capabilities, and classify DOCA_ERROR_* failures. It routes out-of-scope questions to related DOCA skills and public guides.
- A BlueField application developer needs to load a dpacc-produced DPA image and launch a parameterized DPA kernel from a host program.
- A developer needs to determine whether the active BlueField generation and DOCA installation expose a required DPA capability.
- A host launch call returns DOCA_ERROR_NOT_SUPPORTED, DOCA_ERROR_DRIVER, or another doca_dpa error and the failure layer must be isolated.
- A developer needs to coordinate DPA threads through the device-side DPA-Comms component.
- A developer needs to assess direct remote RDMA from a DPA kernel through the device-side DPA-Verbs component.
How do you install this skill?
- Do not treat this skill as a verified runnable implementation; independently confirm the DOCA, DPACC, hardware capability, and API semantics on the target BlueField.
- SKILL.md is contradictory about whether DPA-side comms and verbs programming is in scope; clarify routing to the appropriate public guides before acting.
- BENCHMARK.md, SKILLCARD.yaml, and SKILL.md disagree on repository paths, licenses, and evaluation metadata; the PASS verdict is not independent proof for this revision.
- Execution touches real hardware and asynchronous kernels; begin with a low-risk smoke test and define stop, cleanup, and recovery steps.
- Shell / CLI
- Local filesystem
DOCA SDK at /opt/mellanox/docaBlueField DPU with a host-visible DPA processorMatching DPACC compilerLinux: Ubuntu 22.04/24.04, RHEL, or SLES
Install the skill with the command documented by the repository README:
npx skills add nvidia/skills --skill doca-dpa --yes
For Codex, the README also supports:
npx skills add nvidia/skills --skill doca-dpa --agent codex --yes
The source does not document the exact local destination path; the CLI prompts for or applies the installation target.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/doca-dpa" ~/.claude/skills/
rm -rf "$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:
- Load a dpacc-produced DPA image from my host program, launch a kernel with a buffer and length argument, and debug why no completion arrives.
Give the agent a concrete host-side task, for example: "Load a dpacc-produced DPA image from my host program, launch a kernel with a buffer and length argument, and debug why no completion arrives."
Before using the skill, confirm that DOCA is installed at /opt/mellanox/doca, the BlueField DPA is visible to the host, and the DOCA and DPACC versions satisfy the compatibility policy. Consult TASKS.md for configure, build, modify, run, test, and debug workflows, and CAPABILITIES.md for capability, version, error, and safety guidance.
What are this skill's strengths and limitations?
- Covers the host-side DPA lifecycle and six concrete engineering workflows.
- Separates hardware capability, installed DOCA capability, and DOCA/DPACC version compatibility.
- Clearly distinguishes DPA-Comms and DPA-Verbs from host-side doca-comch and doca-rdma.
- Keeps lifecycle and diagnostic guidance language-neutral for host-side bindings beyond C and C++.
- Requires an existing DOCA SDK, a matching DPACC compiler, and a BlueField environment with a host-visible DPA.
- Ships no buildable application source, build manifest, bindings, samples, or reference subtree.
- Does not cover DOCA or DPACC installation, compiler internals, or the complete DPA-side kernel programming model.
- The supplied material gives no independent test-platform matrix or skill-specific benchmark results.
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 DPA Host-Side Development this page ✓ NVIDIA · Official | 48 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA PCC Custom Congestion Control ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA RDMA Initiator ✓ NVIDIA · Official | 57 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Flow DPA Provider ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Compress Hardware Offload ✓ NVIDIA · Official | 58 · Recommended | ★ 3.8k | today | Apache-2.0 |
This is not a replacement for doca-setup, which handles DOCA and DPACC preparation. It is also not the complete DPA-side programming documentation: DPA kernel, DPA-Comms, and DPA-Verbs questions are routed through doca-public-knowledge-map to the relevant public DOCA guides.
How did FollowSkills review this skill?
The documentation clearly scopes the skill to host-side work, requires version and capability checks, distinguishes error layers, and warns against partial rebuilds; it requests no credentials or secrets. However, it does not systematically define permission boundaries, hardware-operation confirmation, sensitive-data handling, rollback, or recovery procedures, so points are deducted.
TASKS.md and CAPABILITIES.md provide substantial configure, build, run, test, and debug guidance, including PE progress, completions, and version-skew failure causes. Static review cannot reproduce key paths, and the documents contain scope contradictions and some over-absolute error attributions, so the score remains conservative and below the static ceiling.
Audience, environment assumptions, trigger phrases, and out-of-scope routing are clearly described, including BlueField, DOCA, and DPACC dependencies. However, the skill is constrained to Linux, a standard install path, specific hardware, and an English API ecosystem; Chinese support is not demonstrated, and DPA-side coverage is contradictory, so points are deducted.
The documentation is well layered across SKILL.md, CAPABILITIES.md, TASKS.md, evaluations, and related skills, and the skill declares Apache-2.0. However, versioning, changelog, maintenance ownership, and publication metadata are incomplete or inconsistent; the benchmark also reports path and license discrepancies, so points are deducted.
The skill can directly guide the host-side DPA lifecycle, capability checks, error diagnosis, and routing, with broad core-task coverage. It ships no runnable samples or templates, depends on local DOCA/DPACC installations and companion skills, and provides no independently verifiable representative output in the static material, so the score stays below the static ceiling.
The files provide structured workflows, named symbols, an evaluation summary, and external reference pointers, giving limited auditability. The evaluation dataset is unavailable, nothing was executed, the report contains path and metadata inconsistencies, and there is no independent reproduction evidence, so only a low score is justified.
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 →