What does this skill do, and when should you use it?
This skill targets developers who already have a host-side DOCA Flow pipe and a host-side DOCA DPA execution environment. It explains how to use doca-flow-dpa-provider to export pipes, index-selector resources, and memory resources into DPA address space, with coordinated queues and lifecycle handling. It covers the host C library and the DPACC-linked DPA device library, but ships no application source, kernel source, or standalone build manifests. It is one skill in NVIDIA/skills, a monorepo containing 324 separate skills.
Guides the agent to inspect the local DOCA installation and the availability and versions of doca-flow-dpa-provider, doca-flow, doca-dpa, and dpacc; configure a per-port doca_flow_dpa_ctx and general, resources-write, and resources-read queues; export an existing Flow pipe in the required prepare, add entries, export, and get-device-address order; export index-selector and memory resources; pass device addresses to a DPA kernel; use pipe, resource, and completion-queue APIs; and diagnose DOCA_ERROR_* results, lifecycle-ordering errors, missing queues, and undrained DPA-side queues.
- A developer has a Flow pipe with counters and wants a DPA kernel to read them inline and make a follow-up decision.
- A developer needs a DPA kernel to enable, disable, or replace entries in an exported hash pipe.
- A developer needs to read or update Flow-associated memory or index-selector resources from the DPA side.
- A developer receives BAD_STATE, sees no entries from the DPA kernel, or gets repeated AGAIN results and needs to isolate the host, DPA, or export-handshake failure.
- A developer is building non-C host bindings and needs to pass a device address to a separately DPACC-built kernel image.
How do you install this skill?
- This is a low-confidence static review; do not treat the PASS claims, percentages, or signature placeholders in BENCHMARK.md or the skill cards as independent validation.
- Before operating real BlueField hardware, verify compatibility among local DOCA Flow, DOCA DPA, provider, and DPACC versions, and validate export ordering, queue capacity, address ownership, and teardown/recovery in an isolated environment.
- License, repository-path, and version metadata are inconsistent; maintainers should reconcile them and add reproducible test evidence before publication.
- Core guidance depends on NVIDIA packages, hardware, and external documentation; mainland-China users should confirm reachability or arrange an offline mirror.
- Shell / CLI
- Local filesystem
LinuxDOCA SDK installed at /opt/mellanox/docaBlueField DPU with a host-visible DPA processorDPACC compiler matched to DOCApkg-config
Install it through the NVIDIA/skills collection with: npx skills add nvidia/skills --skill doca-flow-dpa-provider --yes. To target a client, for example: npx skills add nvidia/skills --skill doca-flow-dpa-provider --agent codex. The source does not document additional single-skill installation commands; the runtime still requires a DOCA SDK, a compatible BlueField DPU and DPA processor, and a DOCA-matched DPACC installation.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/doca-flow-dpa-provider" ~/.claude/skills/
rm -rf "$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:
- export my existing DOCA Flow pipe to a DPA kernel so it can read counters
- doca_flow_dpa_pipe_export returns DOCA_ERROR_BAD_STATE; help me debug it.
After installing it in an Agent Skills-compatible client, trigger it with a request such as “export my existing DOCA Flow pipe to a DPA kernel so it can read counters” or “doca_flow_dpa_pipe_export returns DOCA_ERROR_BAD_STATE; help me debug it.” For configure, build, run, test, or debug work, the skill directs the agent to TASKS.md; for capability questions, it directs the agent to CAPABILITIES.md. The skill itself does not provide the companion-file contents or a directly buildable sample.
What are this skill's strengths and limitations?
- Covers pipe and external-resource export, all three queue types, device-address handoff, and completion polling.
- Clearly separates host-side DOCA Flow, host-side DPA, and DPA-kernel responsibilities.
- Provides dedicated guidance for version matching, environmental prerequisites, and error classification.
- The skill declares Apache-2.0 licensing; the repository README declares the collection dual-licensed under Apache-2.0 and CC BY-4.0.
- It does not replace DOCA installation, host-side Flow setup, or general DPA kernel development.
- It requires a physical BlueField DPU with a host-visible DPA processor, Linux, DOCA, and a matching DPACC version.
- It contains no application source, DPA kernel source, samples, bindings, or standalone build system.
- The supplied source gives no per-skill test results, platform coverage matrix, or performance data.
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 Flow DPA Provider this page ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Flow DPA Performance ✓ NVIDIA · Official | 58 · Recommended | ★ 3.8k | today | Apache-2.0 |
| DOCA GPUNetIO Development Skill ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA DPDK Bridge ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Flow gRPC Remote Control ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
The skill positions doca-flow as the host-side Flow alternative, doca-dpa as the host-side DPA control library, and the public DOCA Flow, DOCA DPA, and DPACC guides as the reference for DPA kernel writing. It fits Flow-and-DPA integration work, not pure host-side Flow or unrelated generic DPA computation.
How did FollowSkills review this skill?
The documentation defines BlueField, same-device, queue, and export-lifecycle constraints, and asks for confirmation of ports, resources, queues, and kernel termination before proceeding. It also documents address handoff, version matching, and teardown ordering. Points are deducted because the skill affects real hardware and DPA/Flow state, while rollback, least-privilege controls, data-flow boundaries, and sensitive-data handling are not presented as a complete executable policy; some safety rules are only cross-referenced.
The workflows, error taxonomy, preconditions, and abnormal-path guidance are detailed enough to make BAD_STATE, AGAIN, and NOT_SUPPORTED diagnosable. Points are deducted because this is a static review with no runnable implementation, committed test suite, or verifiable CI coverage of key paths; behavior also depends on local DOCA versions, hardware, and installation headers not supplied here, so the static ceiling applies.
The audience, trigger phrases, prerequisites, task scenarios, and out-of-scope boundaries are unusually explicit. Points are deducted because there is no Chinese-language usage or localization guidance, and core operation depends on BlueField hardware, DOCA, DPACC, and NVIDIA documentation/installations; mainland-China network reachability is not addressed.
The package uses layered loader, capability, and task documentation and includes routing, version requirements, error taxonomy, examples, and maintenance/bug-intake references. Points are deducted for inconsistent license metadata (Apache-2.0 in SKILL.md versus Apache-2.0 AND CC-BY-4.0 elsewhere and NOASSERTION in the assessment metadata), TBD version/signature fields, and benchmark-reported missing sections plus path/structure inconsistencies.
The skill targets a narrow, well-defined Flow-to-DPA export task and supplies queue selection, export ordering, device-address handoff, smoke tests, and debugging guidance. Points are deducted because it explicitly ships no buildable samples, source, or standalone build manifests, so the user must rely on local samples and external documentation; no directly verifiable representative output or execution result is included here.
The files provide concrete API names, call ordering, error codes, version-detection commands, benchmark results, and references to installed headers and documentation, giving useful audit hooks. Points are deducted because much of the evidence is author assertion or cross-reference, the benchmark dataset is unavailable, and no CI/test suite or independently reproducible third-party execution evidence is provided in the supplied material.
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 →