[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-doca-dpa-hl-tracer":3,"mdc-8h83d5-key":34,"related-org-nvidia-doca-dpa-hl-tracer":1616,"related-repo-nvidia-doca-dpa-hl-tracer":1777},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":29,"sourceUrl":32,"mdContent":33},"doca-dpa-hl-tracer","capture and decode DPA programming event traces","Use this skill when the user runs doca_dpa_hl_tracer to capture\u002Fdecode DPA-side traces at the programming-events layer (kernel entry\u002Fexit, sync points, comm primitive calls, RDMA WR submission, completion drain) — picking TRACE vs CRIT, tuning the JSON config (file-size limits + file_size_limit_policy, thread priorities\u002Fcores), decoding against the matching DPA-side ELF, or diagnosing empty\u002Fnoisy captures. Trigger even when the user does not explicitly mention \"DOCA DPA tracer\" or \"high-level tracer\" — typical implicit phrasings include \"DPA kernel returns wrong result but host completions look clean\", \"kernel-entry to first-comm latency is huge\", \"RDMA WR to drain gap on the DPA\", \"trace file truncated mid-run\", \"TRACE doubled my DPA latency\", or \"tracer wrote a file but parser shows zero events\". Refuse and route elsewhere for writing DPA kernels, DPA-Comms\u002FDPA-Verbs programming, raw per-cycle DPA profiling, host-side doca-dpa debugging, or production DPA telemetry — those belong to other skills.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,20],{"name":13,"slug":14,"type":15},"Tracing","tracing","tag",{"name":17,"slug":18,"type":15},"Observability","observability",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"Engineering","engineering",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-20T06:24:36.959685","Apache-2.0",281,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fdoca-dpa-hl-tracer","---\nlicense: Apache-2.0\nname: doca-dpa-hl-tracer\ndescription: >\n  Use this skill when the user runs doca_dpa_hl_tracer to\n  capture\u002Fdecode DPA-side traces at the programming-events\n  layer (kernel entry\u002Fexit, sync points, comm primitive\n  calls, RDMA WR submission, completion drain) — picking\n  TRACE vs CRIT, tuning the JSON config (file-size limits\n  + file_size_limit_policy, thread priorities\u002Fcores),\n  decoding against the matching DPA-side ELF, or\n  diagnosing empty\u002Fnoisy captures. Trigger even when the\n  user does not explicitly mention \"DOCA DPA tracer\" or\n  \"high-level tracer\" — typical implicit phrasings include\n  \"DPA kernel returns wrong result but host completions\n  look clean\", \"kernel-entry to first-comm latency is\n  huge\", \"RDMA WR to drain gap on the DPA\", \"trace file\n  truncated mid-run\", \"TRACE doubled my DPA latency\", or\n  \"tracer wrote a file but parser shows zero events\".\n  Refuse and route elsewhere for writing DPA kernels,\n  DPA-Comms\u002FDPA-Verbs programming, raw per-cycle DPA\n  profiling, host-side doca-dpa debugging, or production\n  DPA telemetry — those belong to other skills.\nmetadata:\n  kind: tool\ncompatibility: >\n  Requires DOCA SDK installed at \u002Fopt\u002Fmellanox\u002Fdoca on\n  Linux (Ubuntu 22.04\u002F24.04 or RHEL\u002FSLES) with a BlueField\n  device whose DPA processor is exposed to the host, plus\n  the DOCA DPA Tools optional component (binary at\n  \u002Fopt\u002Fmellanox\u002Fdoca\u002Ftools\u002Fdoca_dpa_hl_tracer). Requires a\n  DPACC-built DPA-side ELF and a live doca-dpa-launched\n  workload for events to fire.\n---\n\n# DOCA DPA High-Level Tracer\n\n**Where to start:** This is a tool skill for invoking\n`doca_dpa_hl_tracer` — the documented host-side CLI that\ncaptures DPA-side execution traces in higher-level terms\n(DPA programming events: kernel entry \u002F exit, sync points,\ncomm primitive calls, RDMA WR submission, completions) rather\nthan raw cycle counts. Open [`TASKS.md`](TASKS.md) and start at\n[`## configure`](TASKS.md#configure) for the\nmode-vs-overhead decision and the JSON config layout, then\n[`## run`](TASKS.md#run) for the\ncapture → decode → render pipeline. Open\n[`CAPABILITIES.md`](CAPABILITIES.md) when the question is\n*what does this tool actually trace*, *which DPA programming\nevents does it expose*, *what is the trace-overhead vs\nfidelity tradeoff*, or *how does it slot into a DPA debug\nloop alongside [`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md)\nand [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md)*. If DPA is not\nthe right surface for the user's question (e.g. the bug is\nhost-side, the bug is in the DPACC-produced image, the user\nwants raw cycle counts), the path-selection rule in\n[`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\nroutes the agent before any capture is attempted.\n\n## Example questions this skill answers well\n\nThe CLASSES of `doca_dpa_hl_tracer` questions this skill is\nbuilt to answer, each with one worked example. The class is\nthe load-bearing piece; the worked example is one instance.\n\n- **\"My DPA kernel is doing the wrong thing — where do I\n  look?\"** — worked example: *\"my host-side\n  `doca_dpa_kernel_launch_update_*` completes, but the\n  kernel's reported result is wrong; no host-side\n  `DOCA_ERROR_*`\"*. Answered by the *when DPA-side\n  high-level tracing is the right surface* gate in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the\n  capture → decode → render flow in\n  [`TASKS.md ## run`](TASKS.md#run) + the\n  *which DPA programming events to focus on* rule in\n  [`TASKS.md ## debug`](TASKS.md#debug).\n- **\"My DPA kernel is slow at a granularity that doesn't\n  show up in cycle profiles — how do I see kernel-entry to\n  first-comm-call latency?\"** — worked example: *\"my DPA\n  kernel runs but the time between launch and the first\n  RDMA WR submission is bigger than I expected\"*. Answered\n  by the event-taxonomy table in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the iterative loop in\n  [`TASKS.md ## test`](TASKS.md#test) which treats trace\n  overhead, mode (`TRACE` vs `CRIT`), and capture window\n  as axes to tune.\n- **\"How do I capture a trace without burying the DPA in\n  observation overhead?\"** — worked example: *\"`TRACE` mode\n  is producing too much data and my measured DPA latency\n  went up by 2x compared to without the tracer\"*. Answered\n  by the mode-vs-overhead tradeoff in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the `CRIT`-first guidance in\n  [`TASKS.md ## configure`](TASKS.md#configure) (start\n  with critical-events-only; widen to `TRACE` only when the\n  bug demands per-event detail).\n- **\"My trace file got truncated mid-run — how should I\n  configure the file-size limits?\"** — worked example:\n  *\"binary trace file hit 5 GB and the capture stopped\"*.\n  Answered by the `log_file_max_size_in_bytes` \u002F\n  `bin_file_max_size_in_bytes` \u002F `file_size_limit_policy`\n  triple in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the JSON config layout in\n  [`TASKS.md ## configure`](TASKS.md#configure).\n- **\"Is the tracer on my install, and is it paired with the\n  matching `doca-dpa` library and DPACC compiler\n  version?\"** — worked example: *\"is the tracer ABI on my\n  install compatible with the DPA image my DPACC just\n  produced?\"*. Answered by the version-overlay in\n  [`CAPABILITIES.md ## Version compatibility`](CAPABILITIES.md#version-compatibility),\n  which redirects to the canonical\n  [`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) chain and\n  adds the *tracer ↔ `doca-dpa` library ↔ DPACC compiler*\n  match rule.\n- **\"The capture file looks empty \u002F decode failed — is the\n  install broken, no events fired, or am I tracing the\n  wrong thing?\"** — worked example: *\"`doca_dpa_hl_tracer`\n  ran, wrote a file, but the parser shows zero events\"*.\n  Answered by the layered error taxonomy in\n  [`CAPABILITIES.md ## Error taxonomy`](CAPABILITIES.md#error-taxonomy)\n  (install \u002F device-binding \u002F DPA-image-instrumented \u002F\n  capture-window \u002F decode-vs-elf \u002F overhead-saturated \u002F\n  version \u002F cross-cutting) + the layered walk in\n  [`TASKS.md ## debug`](TASKS.md#debug).\n\n## Audience\n\nThis skill serves **external developers, platform operators,\nand AI agents who have already brought up a DPA-side workload\nthrough [`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md) and now\nneed higher-level visibility into what the DPA kernel is\nactually doing on the wire** — DPA programming events\nordering, sync gaps, comm-call latencies, RDMA-WR \u002F completion\ntiming — without dropping all the way down to raw cycle\ncounters. Concretely:\n\n- A DPA developer who can launch their kernel cleanly from\n  the host side but whose kernel's *result* is wrong or\n  whose *DPA-side performance* is below expectation, and\n  who needs a DPA-side ground truth before triaging.\n- A platform operator running a DPA-using workload (RDMA\n  offload from accelerator, custom CC algorithm via\n  `doca-pcc`) and needs to localize a regression to the\n  DPA side without instrumenting the application.\n- An AI agent producing a *DPA-side trace report* as\n  evidence for the host-side\n  [`doca-dpa TASKS.md ## debug`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FTASKS.md#debug)\n  ladder when the host side reports clean completions but\n  the DPA-side behaviour is wrong.\n\nIt is **not** for users debugging the tracer binary itself,\n**not** a substitute for the live public DOCA DPA Tools\nguide, **not** the right place for users learning how to\nwrite a DPA kernel (that audience belongs in\n[`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md) plus the public\nDOCA DPA \u002F DPACC \u002F DPA-Comms \u002F DPA-Verbs guides), and **not**\nthe right place for raw per-instruction cycle profiling\n(different surface, different tool — route via\n[`doca-public-knowledge-map ## DOCA tools`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md#doca-tools)).\n\nThe tracer is shipped as a **CLI binary** under\n`\u002Fopt\u002Fmellanox\u002Fdoca\u002Ftools\u002F`, not a library you link against.\nThe skill uses the same `kind: tool` three-file shape as\nthe rest of the bundle so the agent's task-verb contract is\nuniform across libraries, services, and tools.\n\n## Language scope\n\n`doca_dpa_hl_tracer` is a C++ host-side CLI. Its inputs are\nits JSON config file, the DPA-side ELF (the\n`doca_dpa_app`-class image produced by DPACC), and a running\nDPA-side workload that the host-side `doca-dpa` lifecycle\nalready started. Its outputs are a binary trace file\n(`bin_file`) and a human-readable log file (`log_file`).\nThe skill keeps the workflow guidance language-neutral —\nthe DPA-side workload it traces can be C compiled by DPACC\nor any other DPA translation unit DPACC accepts — and\nroutes per-language questions to the public DPA \u002F DPACC\nguides via\n[`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md).\n\n## When to load this skill\n\nLoad this skill when the user is — or the agent needs to —\ninvoke `doca_dpa_hl_tracer` on a real host with DOCA\ninstalled against a BlueField with a DPA processor visible to\nthe host, and the host-side\n[`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md) lifecycle has\nalready brought a DPA workload up at least once. Concretely:\n\n- Capturing a DPA-side trace to localize a DPA kernel's\n  wrong-result or wrong-ordering behaviour when the\n  host-side `doca-dpa` lifecycle reports clean completions.\n- Capturing a DPA-side trace to localize a DPA-side\n  performance gap (kernel-entry to first-comm latency,\n  RDMA-WR-issue to completion gap, sync-point dwell time)\n  at a granularity above raw cycle counts.\n- Choosing between `TRACE` and `CRIT` capture modes based\n  on the bug-vs-overhead tradeoff and the available\n  capture window.\n- Tuning the JSON config (thread priorities, core\n  affinities, file size limits, file-size-limit policy) so\n  the capture itself does not perturb the workload more\n  than the bug it is investigating.\n- Decoding a captured `bin_file` against the matching\n  DPA-side ELF to render the human-readable event stream.\n- Capturing a side-effect-bounded trace as prerequisite\n  evidence for a host-side\n  [`doca-dpa TASKS.md ## debug`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FTASKS.md#debug)\n  ladder step.\n\nDo **not** load this skill for general DOCA orientation,\nDPA-side programming model questions, raw cycle profiling,\nor DOCA \u002F DPACC install. For those, route to\n[`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md),\n[`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md), or\n[`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md).\n\n## What this skill provides\n\nThis is a **thin loader**. Substantive material lives in two\ncompanion files:\n\n- `CAPABILITIES.md` — what `doca_dpa_hl_tracer` captures: the\n  DPA programming event taxonomy (kernel entry \u002F exit, sync\n  points, comm primitive calls, RDMA WR submission and\n  completion drain), the two documented capture modes\n  (`TRACE` for full per-event, `CRIT` for critical-events\n  only), the trace-overhead-vs-fidelity tradeoff, the\n  config-file shape (receiver \u002F binary-writer \u002F file-writer\n  \u002F printer threads with priority + core affinity, file\n  size limits, `file_size_limit_policy`), the\n  capture-window + workload-must-be-running invariant, the\n  ELF-must-match-image rule for decode, the\n  version-availability overlay (tracer ↔\n  [`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md) library ↔\n  DPACC compiler), the layered error taxonomy\n  (install \u002F device-binding \u002F image-instrumented \u002F\n  capture-window \u002F decode \u002F overhead-saturated \u002F version \u002F\n  cross-cutting), the observability surface (binary trace\n  file + log file + tool's own stderr), and the safety\n  policy (capture is bounded; tracing is not a production\n  observability surface).\n- `TASKS.md` — step-by-step workflows for the in-scope\n  task verbs: `install` (route to host-side DOCA install +\n  DPA prerequisites), `configure` (mode + JSON config\n  layout + capture window), `build` (route to install —\n  the binary is shipped, the DPA-side application is\n  user-built by DPACC), `modify` (refuse — do not patch\n  the binary; modify the JSON config and the invocation\n  instead), `run` (the capture flow with `--mode`,\n  `--config-file`, `--output-file`), `test` (iterative\n  loop tuning mode, window, and overhead), `debug` (walk\n  the error taxonomy), `use` (consume the decoded trace\n  in a `doca-dpa` debug session), plus a `Deferred task\n  verbs` block.\n\nThe skill assumes a host where DOCA is already installed at\nthe standard location, a BlueField with a DPA processor is\npresent and visible to the host, the DPACC compiler is\ninstalled at a version matched to the host-side DOCA, the\nDPA-side application image (the ELF the tracer decodes\nagainst) is on disk and matches what the\n[`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md) lifecycle loaded,\nand the operator has the privileges the public DOCA DPA\nTools guide requires.\n\n## What this skill deliberately does not ship\n\nThis skill is **agent guidance**, not a samples or scripts\nbundle. To keep the boundary clean, it deliberately does not\ncontain — and pull requests should not add:\n\n- **Specific flag strings, event names, or mode tokens\n  beyond what the public DOCA DPA Tools page and `--help`\n  document.** The DPA programming events surface evolves\n  release to release; `--help` on the installed binary is\n  the authoritative inventory.\n- **Pre-baked example traces or expected event timings.**\n  Trace output is workload-, DPA-image-, BlueField-, and\n  firmware-specific; a captured example pinned to one\n  setup misleads operators elsewhere.\n- **Wrappers, parsers, or rendering scripts** in any\n  language that consume the binary trace format. The\n  format is documented; users who want to script against\n  it should read the live guide and write the parser\n  against their installed version.\n- **A specific tuning recommendation derived from a single\n  trace.** A DPA-side perf decision (move a sync, batch a\n  comm call, change a launch argument) is a workload\n  question and the skill prescribes how to *capture and\n  read* traces — it refuses to translate a captured gap\n  into a kernel-rewrite recommendation without the user's\n  own analysis.\n- **A `samples\u002F` or `reference\u002F` subtree.** This is a thin\n  loader for a shipped CLI; substantive material lives on\n  the public page, in `--help`, and in\n  [`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md).\n\n## Loading order\n\n1. Read this `SKILL.md` first to confirm the user's question\n   is in scope (DPA-side high-level tracing, not DPA-side\n   programming and not raw cycle profiling).\n2. **For the event taxonomy, capture modes, overhead\n   tradeoff, JSON config layout, version overlay, error\n   taxonomy, observability, and safety policy, see\n   [CAPABILITIES.md](CAPABILITIES.md).**\n3. **For the documented invocations and the\n   capture → decode → render workflow — `install`,\n   `configure`, `build`, `modify`, `run`, `test`, `debug`,\n   `use` — see [TASKS.md](TASKS.md).**\n\n## Related skills\n\n- [`doca-dpa`](..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md) — the host-side\n  DPA control library whose loaded application image the\n  tracer captures. Pair them in every DPA debug session:\n  `doca-dpa` brings the workload up; the tracer captures\n  what the workload does at the DPA programming event\n  layer. Conflating the library with the tracer is the\n  most common DPA-debug first-touch error.\n- [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md) — the\n  cross-cutting debug ladder. The tracer slots in at the\n  *runtime* layer as the DPA-side ground truth before any\n  DPA-side perf or correctness conclusion is made.\n- [`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md)\n  — routing to the public DOCA DPA Tools page on\n  `docs.nvidia.com` and the rest of the public DOCA\n  documentation set.\n- [`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) — canonical\n  DOCA version-handling rules. The `## Version\n  compatibility` section in [`CAPABILITIES.md`](CAPABILITIES.md)\n  is a concise overlay that redirects here for the body\n  and adds the *tracer ↔ `doca-dpa` library ↔ DPACC\n  compiler* matching rule.\n- [`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) — env\n  preparation, install verification, DPACC compiler\n  install \u002F verification, BlueField mode (the DPA\n  processor must be exposed before any tracing is\n  meaningful), and the *I have no install yet* path with\n  the public NGC DOCA container.\n- [`doca-structured-tools-contract`](..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md)\n  — the bundle's detect → prefer → fall back → report\n  contract for structured helper tools. The command\n  appendix in [`TASKS.md`](TASKS.md) honors this contract.\n- [`doca-programming-guide`](..\u002F..\u002Fdoca-programming-guide\u002FSKILL.md)\n  — general DOCA programming patterns shared by every\n  library \u002F tool surface, including the cross-library\n  `DOCA_ERROR_*` taxonomy this tool's host-side error\n  layer overlays on top of when host-side `doca-dpa`\n  calls fail in tandem.\n\nThe DPA-side companion libraries `doca-dpa-comms` (comm\nprimitives the DPA kernel itself calls) and\n`doca-dpa-verbs` (RDMA verbs the DPA kernel itself calls)\nare **different artifacts** that the tracer's *DPA\nprogramming events* surface visibly names; for the\nDPA-side programming model itself, route through\n[`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md)\nto the public DOCA DPA-Comms and DPA-Verbs guides and to\nthe shipped `\u002Fopt\u002Fmellanox\u002Fdoca\u002Fsamples\u002Fdoca_dpa\u002F` samples.\nThis tool *traces* their use; it does not redefine them.\n",{"data":35,"body":39},{"license":26,"name":4,"description":6,"metadata":36,"compatibility":38},{"kind":37},"tool","Requires DOCA SDK installed at \u002Fopt\u002Fmellanox\u002Fdoca on Linux (Ubuntu 22.04\u002F24.04 or RHEL\u002FSLES) with a BlueField device whose DPA processor is exposed to the host, plus the DOCA DPA Tools optional component (binary at \u002Fopt\u002Fmellanox\u002Fdoca\u002Ftools\u002Fdoca_dpa_hl_tracer). Requires a DPACC-built DPA-side ELF and a live doca-dpa-launched workload for events to fire.\n",{"type":40,"children":41},"root",[42,51,181,188,200,584,590,612,671,723,751,757,809,815,837,908,950,956,968,1130,1145,1151,1163,1270,1276,1368,1374,1556],{"type":43,"tag":44,"props":45,"children":47},"element","h1",{"id":46},"doca-dpa-high-level-tracer",[48],{"type":49,"value":50},"text","DOCA DPA High-Level Tracer",{"type":43,"tag":52,"props":53,"children":54},"p",{},[55,61,63,70,72,82,84,94,96,106,108,117,119,125,127,132,133,138,140,167,169,179],{"type":43,"tag":56,"props":57,"children":58},"strong",{},[59],{"type":49,"value":60},"Where to start:",{"type":49,"value":62}," This is a tool skill for invoking\n",{"type":43,"tag":64,"props":65,"children":67},"code",{"className":66},[],[68],{"type":49,"value":69},"doca_dpa_hl_tracer",{"type":49,"value":71}," — the documented host-side CLI that\ncaptures DPA-side execution traces in higher-level terms\n(DPA programming events: kernel entry \u002F exit, sync points,\ncomm primitive calls, RDMA WR submission, completions) rather\nthan raw cycle counts. Open ",{"type":43,"tag":73,"props":74,"children":76},"a",{"href":75},"TASKS.md",[77],{"type":43,"tag":64,"props":78,"children":80},{"className":79},[],[81],{"type":49,"value":75},{"type":49,"value":83}," and start at\n",{"type":43,"tag":73,"props":85,"children":87},{"href":86},"TASKS.md#configure",[88],{"type":43,"tag":64,"props":89,"children":91},{"className":90},[],[92],{"type":49,"value":93},"## configure",{"type":49,"value":95}," for the\nmode-vs-overhead decision and the JSON config layout, then\n",{"type":43,"tag":73,"props":97,"children":99},{"href":98},"TASKS.md#run",[100],{"type":43,"tag":64,"props":101,"children":103},{"className":102},[],[104],{"type":49,"value":105},"## run",{"type":49,"value":107}," for the\ncapture → decode → render pipeline. Open\n",{"type":43,"tag":73,"props":109,"children":111},{"href":110},"CAPABILITIES.md",[112],{"type":43,"tag":64,"props":113,"children":115},{"className":114},[],[116],{"type":49,"value":110},{"type":49,"value":118}," when the question is\n",{"type":43,"tag":120,"props":121,"children":122},"em",{},[123],{"type":49,"value":124},"what does this tool actually trace",{"type":49,"value":126},", ",{"type":43,"tag":120,"props":128,"children":129},{},[130],{"type":49,"value":131},"which DPA programming\nevents does it expose",{"type":49,"value":126},{"type":43,"tag":120,"props":134,"children":135},{},[136],{"type":49,"value":137},"what is the trace-overhead vs\nfidelity tradeoff",{"type":49,"value":139},", or ",{"type":43,"tag":120,"props":141,"children":142},{},[143,145,155,157],{"type":49,"value":144},"how does it slot into a DPA debug\nloop alongside ",{"type":43,"tag":73,"props":146,"children":148},{"href":147},"..\u002F..\u002Flibs\u002Fdoca-dpa\u002FSKILL.md",[149],{"type":43,"tag":64,"props":150,"children":152},{"className":151},[],[153],{"type":49,"value":154},"doca-dpa",{"type":49,"value":156},"\nand ",{"type":43,"tag":73,"props":158,"children":160},{"href":159},"..\u002F..\u002Fdoca-debug\u002FSKILL.md",[161],{"type":43,"tag":64,"props":162,"children":164},{"className":163},[],[165],{"type":49,"value":166},"doca-debug",{"type":49,"value":168},". If DPA is not\nthe right surface for the user's question (e.g. the bug is\nhost-side, the bug is in the DPACC-produced image, the user\nwants raw cycle counts), the path-selection rule in\n",{"type":43,"tag":73,"props":170,"children":172},{"href":171},"CAPABILITIES.md#capabilities-and-modes",[173],{"type":43,"tag":64,"props":174,"children":176},{"className":175},[],[177],{"type":49,"value":178},"CAPABILITIES.md ## Capabilities and modes",{"type":49,"value":180},"\nroutes the agent before any capture is attempted.",{"type":43,"tag":182,"props":183,"children":185},"h2",{"id":184},"example-questions-this-skill-answers-well",[186],{"type":49,"value":187},"Example questions this skill answers well",{"type":43,"tag":52,"props":189,"children":190},{},[191,193,198],{"type":49,"value":192},"The CLASSES of ",{"type":43,"tag":64,"props":194,"children":196},{"className":195},[],[197],{"type":49,"value":69},{"type":49,"value":199}," questions this skill is\nbuilt to answer, each with one worked example. The class is\nthe load-bearing piece; the worked example is one instance.",{"type":43,"tag":201,"props":202,"children":203},"ul",{},[204,291,351,414,480,541],{"type":43,"tag":205,"props":206,"children":207},"li",{},[208,213,215,236,238,243,245,253],{"type":43,"tag":56,"props":209,"children":210},{},[211],{"type":49,"value":212},"\"My DPA kernel is doing the wrong thing — where do I\nlook?\"",{"type":49,"value":214}," — worked example: ",{"type":43,"tag":120,"props":216,"children":217},{},[218,220,226,228,234],{"type":49,"value":219},"\"my host-side\n",{"type":43,"tag":64,"props":221,"children":223},{"className":222},[],[224],{"type":49,"value":225},"doca_dpa_kernel_launch_update_*",{"type":49,"value":227}," completes, but the\nkernel's reported result is wrong; no host-side\n",{"type":43,"tag":64,"props":229,"children":231},{"className":230},[],[232],{"type":49,"value":233},"DOCA_ERROR_*",{"type":49,"value":235},"\"",{"type":49,"value":237},". Answered by the ",{"type":43,"tag":120,"props":239,"children":240},{},[241],{"type":49,"value":242},"when DPA-side\nhigh-level tracing is the right surface",{"type":49,"value":244}," gate in\n",{"type":43,"tag":73,"props":246,"children":247},{"href":171},[248],{"type":43,"tag":64,"props":249,"children":251},{"className":250},[],[252],{"type":49,"value":178},{"type":43,"tag":201,"props":254,"children":255},{},[256],{"type":43,"tag":205,"props":257,"children":258},{},[259,261,270,272,277,279,289],{"type":49,"value":260},"the\ncapture → decode → render flow in\n",{"type":43,"tag":73,"props":262,"children":263},{"href":98},[264],{"type":43,"tag":64,"props":265,"children":267},{"className":266},[],[268],{"type":49,"value":269},"TASKS.md ## run",{"type":49,"value":271}," + the\n",{"type":43,"tag":120,"props":273,"children":274},{},[275],{"type":49,"value":276},"which DPA programming events to focus on",{"type":49,"value":278}," rule in\n",{"type":43,"tag":73,"props":280,"children":282},{"href":281},"TASKS.md#debug",[283],{"type":43,"tag":64,"props":284,"children":286},{"className":285},[],[287],{"type":49,"value":288},"TASKS.md ## debug",{"type":49,"value":290},".",{"type":43,"tag":205,"props":292,"children":293},{},[294,299,300,305,307,315],{"type":43,"tag":56,"props":295,"children":296},{},[297],{"type":49,"value":298},"\"My DPA kernel is slow at a granularity that doesn't\nshow up in cycle profiles — how do I see kernel-entry to\nfirst-comm-call latency?\"",{"type":49,"value":214},{"type":43,"tag":120,"props":301,"children":302},{},[303],{"type":49,"value":304},"\"my DPA\nkernel runs but the time between launch and the first\nRDMA WR submission is bigger than I expected\"",{"type":49,"value":306},". Answered\nby the event-taxonomy table in\n",{"type":43,"tag":73,"props":308,"children":309},{"href":171},[310],{"type":43,"tag":64,"props":311,"children":313},{"className":312},[],[314],{"type":49,"value":178},{"type":43,"tag":201,"props":316,"children":317},{},[318],{"type":43,"tag":205,"props":319,"children":320},{},[321,323,333,335,341,343,349],{"type":49,"value":322},"the iterative loop in\n",{"type":43,"tag":73,"props":324,"children":326},{"href":325},"TASKS.md#test",[327],{"type":43,"tag":64,"props":328,"children":330},{"className":329},[],[331],{"type":49,"value":332},"TASKS.md ## test",{"type":49,"value":334}," which treats trace\noverhead, mode (",{"type":43,"tag":64,"props":336,"children":338},{"className":337},[],[339],{"type":49,"value":340},"TRACE",{"type":49,"value":342}," vs ",{"type":43,"tag":64,"props":344,"children":346},{"className":345},[],[347],{"type":49,"value":348},"CRIT",{"type":49,"value":350},"), and capture window\nas axes to tune.",{"type":43,"tag":205,"props":352,"children":353},{},[354,359,360,371,373,381],{"type":43,"tag":56,"props":355,"children":356},{},[357],{"type":49,"value":358},"\"How do I capture a trace without burying the DPA in\nobservation overhead?\"",{"type":49,"value":214},{"type":43,"tag":120,"props":361,"children":362},{},[363,364,369],{"type":49,"value":235},{"type":43,"tag":64,"props":365,"children":367},{"className":366},[],[368],{"type":49,"value":340},{"type":49,"value":370}," mode\nis producing too much data and my measured DPA latency\nwent up by 2x compared to without the tracer\"",{"type":49,"value":372},". Answered\nby the mode-vs-overhead tradeoff in\n",{"type":43,"tag":73,"props":374,"children":375},{"href":171},[376],{"type":43,"tag":64,"props":377,"children":379},{"className":378},[],[380],{"type":49,"value":178},{"type":43,"tag":201,"props":382,"children":383},{},[384],{"type":43,"tag":205,"props":385,"children":386},{},[387,389,394,396,405,407,412],{"type":49,"value":388},"the ",{"type":43,"tag":64,"props":390,"children":392},{"className":391},[],[393],{"type":49,"value":348},{"type":49,"value":395},"-first guidance in\n",{"type":43,"tag":73,"props":397,"children":398},{"href":86},[399],{"type":43,"tag":64,"props":400,"children":402},{"className":401},[],[403],{"type":49,"value":404},"TASKS.md ## configure",{"type":49,"value":406}," (start\nwith critical-events-only; widen to ",{"type":43,"tag":64,"props":408,"children":410},{"className":409},[],[411],{"type":49,"value":340},{"type":49,"value":413}," only when the\nbug demands per-event detail).",{"type":43,"tag":205,"props":415,"children":416},{},[417,422,424,429,431,437,439,445,447,453,455,463],{"type":43,"tag":56,"props":418,"children":419},{},[420],{"type":49,"value":421},"\"My trace file got truncated mid-run — how should I\nconfigure the file-size limits?\"",{"type":49,"value":423}," — worked example:\n",{"type":43,"tag":120,"props":425,"children":426},{},[427],{"type":49,"value":428},"\"binary trace file hit 5 GB and the capture stopped\"",{"type":49,"value":430},".\nAnswered by the ",{"type":43,"tag":64,"props":432,"children":434},{"className":433},[],[435],{"type":49,"value":436},"log_file_max_size_in_bytes",{"type":49,"value":438}," \u002F\n",{"type":43,"tag":64,"props":440,"children":442},{"className":441},[],[443],{"type":49,"value":444},"bin_file_max_size_in_bytes",{"type":49,"value":446}," \u002F ",{"type":43,"tag":64,"props":448,"children":450},{"className":449},[],[451],{"type":49,"value":452},"file_size_limit_policy",{"type":49,"value":454},"\ntriple in\n",{"type":43,"tag":73,"props":456,"children":457},{"href":171},[458],{"type":43,"tag":64,"props":459,"children":461},{"className":460},[],[462],{"type":49,"value":178},{"type":43,"tag":201,"props":464,"children":465},{},[466],{"type":43,"tag":205,"props":467,"children":468},{},[469,471,479],{"type":49,"value":470},"the JSON config layout in\n",{"type":43,"tag":73,"props":472,"children":473},{"href":86},[474],{"type":43,"tag":64,"props":475,"children":477},{"className":476},[],[478],{"type":49,"value":404},{"type":49,"value":290},{"type":43,"tag":205,"props":481,"children":482},{},[483,495,496,501,503,513,515,525,527,539],{"type":43,"tag":56,"props":484,"children":485},{},[486,488,493],{"type":49,"value":487},"\"Is the tracer on my install, and is it paired with the\nmatching ",{"type":43,"tag":64,"props":489,"children":491},{"className":490},[],[492],{"type":49,"value":154},{"type":49,"value":494}," library and DPACC compiler\nversion?\"",{"type":49,"value":214},{"type":43,"tag":120,"props":497,"children":498},{},[499],{"type":49,"value":500},"\"is the tracer ABI on my\ninstall compatible with the DPA image my DPACC just\nproduced?\"",{"type":49,"value":502},". Answered by the version-overlay in\n",{"type":43,"tag":73,"props":504,"children":506},{"href":505},"CAPABILITIES.md#version-compatibility",[507],{"type":43,"tag":64,"props":508,"children":510},{"className":509},[],[511],{"type":49,"value":512},"CAPABILITIES.md ## Version compatibility",{"type":49,"value":514},",\nwhich redirects to the canonical\n",{"type":43,"tag":73,"props":516,"children":518},{"href":517},"..\u002F..\u002Fdoca-version\u002FSKILL.md",[519],{"type":43,"tag":64,"props":520,"children":522},{"className":521},[],[523],{"type":49,"value":524},"doca-version",{"type":49,"value":526}," chain and\nadds the ",{"type":43,"tag":120,"props":528,"children":529},{},[530,532,537],{"type":49,"value":531},"tracer ↔ ",{"type":43,"tag":64,"props":533,"children":535},{"className":534},[],[536],{"type":49,"value":154},{"type":49,"value":538}," library ↔ DPACC compiler",{"type":49,"value":540},"\nmatch rule.",{"type":43,"tag":205,"props":542,"children":543},{},[544,549,550,561,563,573,575,583],{"type":43,"tag":56,"props":545,"children":546},{},[547],{"type":49,"value":548},"\"The capture file looks empty \u002F decode failed — is the\ninstall broken, no events fired, or am I tracing the\nwrong thing?\"",{"type":49,"value":214},{"type":43,"tag":120,"props":551,"children":552},{},[553,554,559],{"type":49,"value":235},{"type":43,"tag":64,"props":555,"children":557},{"className":556},[],[558],{"type":49,"value":69},{"type":49,"value":560},"\nran, wrote a file, but the parser shows zero events\"",{"type":49,"value":562},".\nAnswered by the layered error taxonomy in\n",{"type":43,"tag":73,"props":564,"children":566},{"href":565},"CAPABILITIES.md#error-taxonomy",[567],{"type":43,"tag":64,"props":568,"children":570},{"className":569},[],[571],{"type":49,"value":572},"CAPABILITIES.md ## Error taxonomy",{"type":49,"value":574},"\n(install \u002F device-binding \u002F DPA-image-instrumented \u002F\ncapture-window \u002F decode-vs-elf \u002F overhead-saturated \u002F\nversion \u002F cross-cutting) + the layered walk in\n",{"type":43,"tag":73,"props":576,"children":577},{"href":281},[578],{"type":43,"tag":64,"props":579,"children":581},{"className":580},[],[582],{"type":49,"value":288},{"type":49,"value":290},{"type":43,"tag":182,"props":585,"children":587},{"id":586},"audience",[588],{"type":49,"value":589},"Audience",{"type":43,"tag":52,"props":591,"children":592},{},[593,595,610],{"type":49,"value":594},"This skill serves ",{"type":43,"tag":56,"props":596,"children":597},{},[598,600,608],{"type":49,"value":599},"external developers, platform operators,\nand AI agents who have already brought up a DPA-side workload\nthrough ",{"type":43,"tag":73,"props":601,"children":602},{"href":147},[603],{"type":43,"tag":64,"props":604,"children":606},{"className":605},[],[607],{"type":49,"value":154},{"type":49,"value":609}," and now\nneed higher-level visibility into what the DPA kernel is\nactually doing on the wire",{"type":49,"value":611}," — DPA programming events\nordering, sync gaps, comm-call latencies, RDMA-WR \u002F completion\ntiming — without dropping all the way down to raw cycle\ncounters. Concretely:",{"type":43,"tag":201,"props":613,"children":614},{},[615,634,647],{"type":43,"tag":205,"props":616,"children":617},{},[618,620,625,627,632],{"type":49,"value":619},"A DPA developer who can launch their kernel cleanly from\nthe host side but whose kernel's ",{"type":43,"tag":120,"props":621,"children":622},{},[623],{"type":49,"value":624},"result",{"type":49,"value":626}," is wrong or\nwhose ",{"type":43,"tag":120,"props":628,"children":629},{},[630],{"type":49,"value":631},"DPA-side performance",{"type":49,"value":633}," is below expectation, and\nwho needs a DPA-side ground truth before triaging.",{"type":43,"tag":205,"props":635,"children":636},{},[637,639,645],{"type":49,"value":638},"A platform operator running a DPA-using workload (RDMA\noffload from accelerator, custom CC algorithm via\n",{"type":43,"tag":64,"props":640,"children":642},{"className":641},[],[643],{"type":49,"value":644},"doca-pcc",{"type":49,"value":646},") and needs to localize a regression to the\nDPA side without instrumenting the application.",{"type":43,"tag":205,"props":648,"children":649},{},[650,652,657,659,669],{"type":49,"value":651},"An AI agent producing a ",{"type":43,"tag":120,"props":653,"children":654},{},[655],{"type":49,"value":656},"DPA-side trace report",{"type":49,"value":658}," as\nevidence for the host-side\n",{"type":43,"tag":73,"props":660,"children":662},{"href":661},"..\u002F..\u002Flibs\u002Fdoca-dpa\u002FTASKS.md#debug",[663],{"type":43,"tag":64,"props":664,"children":666},{"className":665},[],[667],{"type":49,"value":668},"doca-dpa TASKS.md ## debug",{"type":49,"value":670},"\nladder when the host side reports clean completions but\nthe DPA-side behaviour is wrong.",{"type":43,"tag":52,"props":672,"children":673},{},[674,676,681,683,687,689,693,695,703,705,709,711,721],{"type":49,"value":675},"It is ",{"type":43,"tag":56,"props":677,"children":678},{},[679],{"type":49,"value":680},"not",{"type":49,"value":682}," for users debugging the tracer binary itself,\n",{"type":43,"tag":56,"props":684,"children":685},{},[686],{"type":49,"value":680},{"type":49,"value":688}," a substitute for the live public DOCA DPA Tools\nguide, ",{"type":43,"tag":56,"props":690,"children":691},{},[692],{"type":49,"value":680},{"type":49,"value":694}," the right place for users learning how to\nwrite a DPA kernel (that audience belongs in\n",{"type":43,"tag":73,"props":696,"children":697},{"href":147},[698],{"type":43,"tag":64,"props":699,"children":701},{"className":700},[],[702],{"type":49,"value":154},{"type":49,"value":704}," plus the public\nDOCA DPA \u002F DPACC \u002F DPA-Comms \u002F DPA-Verbs guides), and ",{"type":43,"tag":56,"props":706,"children":707},{},[708],{"type":49,"value":680},{"type":49,"value":710},"\nthe right place for raw per-instruction cycle profiling\n(different surface, different tool — route via\n",{"type":43,"tag":73,"props":712,"children":714},{"href":713},"..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md#doca-tools",[715],{"type":43,"tag":64,"props":716,"children":718},{"className":717},[],[719],{"type":49,"value":720},"doca-public-knowledge-map ## DOCA tools",{"type":49,"value":722},").",{"type":43,"tag":52,"props":724,"children":725},{},[726,728,733,735,741,743,749],{"type":49,"value":727},"The tracer is shipped as a ",{"type":43,"tag":56,"props":729,"children":730},{},[731],{"type":49,"value":732},"CLI binary",{"type":49,"value":734}," under\n",{"type":43,"tag":64,"props":736,"children":738},{"className":737},[],[739],{"type":49,"value":740},"\u002Fopt\u002Fmellanox\u002Fdoca\u002Ftools\u002F",{"type":49,"value":742},", not a library you link against.\nThe skill uses the same ",{"type":43,"tag":64,"props":744,"children":746},{"className":745},[],[747],{"type":49,"value":748},"kind: tool",{"type":49,"value":750}," three-file shape as\nthe rest of the bundle so the agent's task-verb contract is\nuniform across libraries, services, and tools.",{"type":43,"tag":182,"props":752,"children":754},{"id":753},"language-scope",[755],{"type":49,"value":756},"Language scope",{"type":43,"tag":52,"props":758,"children":759},{},[760,765,767,773,775,780,782,788,790,796,798,808],{"type":43,"tag":64,"props":761,"children":763},{"className":762},[],[764],{"type":49,"value":69},{"type":49,"value":766}," is a C++ host-side CLI. Its inputs are\nits JSON config file, the DPA-side ELF (the\n",{"type":43,"tag":64,"props":768,"children":770},{"className":769},[],[771],{"type":49,"value":772},"doca_dpa_app",{"type":49,"value":774},"-class image produced by DPACC), and a running\nDPA-side workload that the host-side ",{"type":43,"tag":64,"props":776,"children":778},{"className":777},[],[779],{"type":49,"value":154},{"type":49,"value":781}," lifecycle\nalready started. Its outputs are a binary trace file\n(",{"type":43,"tag":64,"props":783,"children":785},{"className":784},[],[786],{"type":49,"value":787},"bin_file",{"type":49,"value":789},") and a human-readable log file (",{"type":43,"tag":64,"props":791,"children":793},{"className":792},[],[794],{"type":49,"value":795},"log_file",{"type":49,"value":797},").\nThe skill keeps the workflow guidance language-neutral —\nthe DPA-side workload it traces can be C compiled by DPACC\nor any other DPA translation unit DPACC accepts — and\nroutes per-language questions to the public DPA \u002F DPACC\nguides via\n",{"type":43,"tag":73,"props":799,"children":801},{"href":800},"..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md",[802],{"type":43,"tag":64,"props":803,"children":805},{"className":804},[],[806],{"type":49,"value":807},"doca-public-knowledge-map",{"type":49,"value":290},{"type":43,"tag":182,"props":810,"children":812},{"id":811},"when-to-load-this-skill",[813],{"type":49,"value":814},"When to load this skill",{"type":43,"tag":52,"props":816,"children":817},{},[818,820,825,827,835],{"type":49,"value":819},"Load this skill when the user is — or the agent needs to —\ninvoke ",{"type":43,"tag":64,"props":821,"children":823},{"className":822},[],[824],{"type":49,"value":69},{"type":49,"value":826}," on a real host with DOCA\ninstalled against a BlueField with a DPA processor visible to\nthe host, and the host-side\n",{"type":43,"tag":73,"props":828,"children":829},{"href":147},[830],{"type":43,"tag":64,"props":831,"children":833},{"className":832},[],[834],{"type":49,"value":154},{"type":49,"value":836}," lifecycle has\nalready brought a DPA workload up at least once. Concretely:",{"type":43,"tag":201,"props":838,"children":839},{},[840,852,857,876,881,893],{"type":43,"tag":205,"props":841,"children":842},{},[843,845,850],{"type":49,"value":844},"Capturing a DPA-side trace to localize a DPA kernel's\nwrong-result or wrong-ordering behaviour when the\nhost-side ",{"type":43,"tag":64,"props":846,"children":848},{"className":847},[],[849],{"type":49,"value":154},{"type":49,"value":851}," lifecycle reports clean completions.",{"type":43,"tag":205,"props":853,"children":854},{},[855],{"type":49,"value":856},"Capturing a DPA-side trace to localize a DPA-side\nperformance gap (kernel-entry to first-comm latency,\nRDMA-WR-issue to completion gap, sync-point dwell time)\nat a granularity above raw cycle counts.",{"type":43,"tag":205,"props":858,"children":859},{},[860,862,867,869,874],{"type":49,"value":861},"Choosing between ",{"type":43,"tag":64,"props":863,"children":865},{"className":864},[],[866],{"type":49,"value":340},{"type":49,"value":868}," and ",{"type":43,"tag":64,"props":870,"children":872},{"className":871},[],[873],{"type":49,"value":348},{"type":49,"value":875}," capture modes based\non the bug-vs-overhead tradeoff and the available\ncapture window.",{"type":43,"tag":205,"props":877,"children":878},{},[879],{"type":49,"value":880},"Tuning the JSON config (thread priorities, core\naffinities, file size limits, file-size-limit policy) so\nthe capture itself does not perturb the workload more\nthan the bug it is investigating.",{"type":43,"tag":205,"props":882,"children":883},{},[884,886,891],{"type":49,"value":885},"Decoding a captured ",{"type":43,"tag":64,"props":887,"children":889},{"className":888},[],[890],{"type":49,"value":787},{"type":49,"value":892}," against the matching\nDPA-side ELF to render the human-readable event stream.",{"type":43,"tag":205,"props":894,"children":895},{},[896,898,906],{"type":49,"value":897},"Capturing a side-effect-bounded trace as prerequisite\nevidence for a host-side\n",{"type":43,"tag":73,"props":899,"children":900},{"href":661},[901],{"type":43,"tag":64,"props":902,"children":904},{"className":903},[],[905],{"type":49,"value":668},{"type":49,"value":907},"\nladder step.",{"type":43,"tag":52,"props":909,"children":910},{},[911,913,917,919,927,929,937,939,949],{"type":49,"value":912},"Do ",{"type":43,"tag":56,"props":914,"children":915},{},[916],{"type":49,"value":680},{"type":49,"value":918}," load this skill for general DOCA orientation,\nDPA-side programming model questions, raw cycle profiling,\nor DOCA \u002F DPACC install. For those, route to\n",{"type":43,"tag":73,"props":920,"children":921},{"href":800},[922],{"type":43,"tag":64,"props":923,"children":925},{"className":924},[],[926],{"type":49,"value":807},{"type":49,"value":928},",\n",{"type":43,"tag":73,"props":930,"children":931},{"href":147},[932],{"type":43,"tag":64,"props":933,"children":935},{"className":934},[],[936],{"type":49,"value":154},{"type":49,"value":938},", or\n",{"type":43,"tag":73,"props":940,"children":942},{"href":941},"..\u002F..\u002Fdoca-setup\u002FSKILL.md",[943],{"type":43,"tag":64,"props":944,"children":946},{"className":945},[],[947],{"type":49,"value":948},"doca-setup",{"type":49,"value":290},{"type":43,"tag":182,"props":951,"children":953},{"id":952},"what-this-skill-provides",[954],{"type":49,"value":955},"What this skill provides",{"type":43,"tag":52,"props":957,"children":958},{},[959,961,966],{"type":49,"value":960},"This is a ",{"type":43,"tag":56,"props":962,"children":963},{},[964],{"type":49,"value":965},"thin loader",{"type":49,"value":967},". Substantive material lives in two\ncompanion files:",{"type":43,"tag":201,"props":969,"children":970},{},[971,1019],{"type":43,"tag":205,"props":972,"children":973},{},[974,979,981,986,988,993,995,1000,1002,1007,1009,1017],{"type":43,"tag":64,"props":975,"children":977},{"className":976},[],[978],{"type":49,"value":110},{"type":49,"value":980}," — what ",{"type":43,"tag":64,"props":982,"children":984},{"className":983},[],[985],{"type":49,"value":69},{"type":49,"value":987}," captures: the\nDPA programming event taxonomy (kernel entry \u002F exit, sync\npoints, comm primitive calls, RDMA WR submission and\ncompletion drain), the two documented capture modes\n(",{"type":43,"tag":64,"props":989,"children":991},{"className":990},[],[992],{"type":49,"value":340},{"type":49,"value":994}," for full per-event, ",{"type":43,"tag":64,"props":996,"children":998},{"className":997},[],[999],{"type":49,"value":348},{"type":49,"value":1001}," for critical-events\nonly), the trace-overhead-vs-fidelity tradeoff, the\nconfig-file shape (receiver \u002F binary-writer \u002F file-writer\n\u002F printer threads with priority + core affinity, file\nsize limits, ",{"type":43,"tag":64,"props":1003,"children":1005},{"className":1004},[],[1006],{"type":49,"value":452},{"type":49,"value":1008},"), the\ncapture-window + workload-must-be-running invariant, the\nELF-must-match-image rule for decode, the\nversion-availability overlay (tracer ↔\n",{"type":43,"tag":73,"props":1010,"children":1011},{"href":147},[1012],{"type":43,"tag":64,"props":1013,"children":1015},{"className":1014},[],[1016],{"type":49,"value":154},{"type":49,"value":1018}," library ↔\nDPACC compiler), the layered error taxonomy\n(install \u002F device-binding \u002F image-instrumented \u002F\ncapture-window \u002F decode \u002F overhead-saturated \u002F version \u002F\ncross-cutting), the observability surface (binary trace\nfile + log file + tool's own stderr), and the safety\npolicy (capture is bounded; tracing is not a production\nobservability surface).",{"type":43,"tag":205,"props":1020,"children":1021},{},[1022,1027,1029,1035,1037,1043,1045,1051,1053,1059,1061,1067,1069,1075,1076,1082,1083,1089,1091,1097,1099,1105,1107,1113,1115,1120,1122,1128],{"type":43,"tag":64,"props":1023,"children":1025},{"className":1024},[],[1026],{"type":49,"value":75},{"type":49,"value":1028}," — step-by-step workflows for the in-scope\ntask verbs: ",{"type":43,"tag":64,"props":1030,"children":1032},{"className":1031},[],[1033],{"type":49,"value":1034},"install",{"type":49,"value":1036}," (route to host-side DOCA install +\nDPA prerequisites), ",{"type":43,"tag":64,"props":1038,"children":1040},{"className":1039},[],[1041],{"type":49,"value":1042},"configure",{"type":49,"value":1044}," (mode + JSON config\nlayout + capture window), ",{"type":43,"tag":64,"props":1046,"children":1048},{"className":1047},[],[1049],{"type":49,"value":1050},"build",{"type":49,"value":1052}," (route to install —\nthe binary is shipped, the DPA-side application is\nuser-built by DPACC), ",{"type":43,"tag":64,"props":1054,"children":1056},{"className":1055},[],[1057],{"type":49,"value":1058},"modify",{"type":49,"value":1060}," (refuse — do not patch\nthe binary; modify the JSON config and the invocation\ninstead), ",{"type":43,"tag":64,"props":1062,"children":1064},{"className":1063},[],[1065],{"type":49,"value":1066},"run",{"type":49,"value":1068}," (the capture flow with ",{"type":43,"tag":64,"props":1070,"children":1072},{"className":1071},[],[1073],{"type":49,"value":1074},"--mode",{"type":49,"value":928},{"type":43,"tag":64,"props":1077,"children":1079},{"className":1078},[],[1080],{"type":49,"value":1081},"--config-file",{"type":49,"value":126},{"type":43,"tag":64,"props":1084,"children":1086},{"className":1085},[],[1087],{"type":49,"value":1088},"--output-file",{"type":49,"value":1090},"), ",{"type":43,"tag":64,"props":1092,"children":1094},{"className":1093},[],[1095],{"type":49,"value":1096},"test",{"type":49,"value":1098}," (iterative\nloop tuning mode, window, and overhead), ",{"type":43,"tag":64,"props":1100,"children":1102},{"className":1101},[],[1103],{"type":49,"value":1104},"debug",{"type":49,"value":1106}," (walk\nthe error taxonomy), ",{"type":43,"tag":64,"props":1108,"children":1110},{"className":1109},[],[1111],{"type":49,"value":1112},"use",{"type":49,"value":1114}," (consume the decoded trace\nin a ",{"type":43,"tag":64,"props":1116,"children":1118},{"className":1117},[],[1119],{"type":49,"value":154},{"type":49,"value":1121}," debug session), plus a ",{"type":43,"tag":64,"props":1123,"children":1125},{"className":1124},[],[1126],{"type":49,"value":1127},"Deferred task verbs",{"type":49,"value":1129}," block.",{"type":43,"tag":52,"props":1131,"children":1132},{},[1133,1135,1143],{"type":49,"value":1134},"The skill assumes a host where DOCA is already installed at\nthe standard location, a BlueField with a DPA processor is\npresent and visible to the host, the DPACC compiler is\ninstalled at a version matched to the host-side DOCA, the\nDPA-side application image (the ELF the tracer decodes\nagainst) is on disk and matches what the\n",{"type":43,"tag":73,"props":1136,"children":1137},{"href":147},[1138],{"type":43,"tag":64,"props":1139,"children":1141},{"className":1140},[],[1142],{"type":49,"value":154},{"type":49,"value":1144}," lifecycle loaded,\nand the operator has the privileges the public DOCA DPA\nTools guide requires.",{"type":43,"tag":182,"props":1146,"children":1148},{"id":1147},"what-this-skill-deliberately-does-not-ship",[1149],{"type":49,"value":1150},"What this skill deliberately does not ship",{"type":43,"tag":52,"props":1152,"children":1153},{},[1154,1156,1161],{"type":49,"value":1155},"This skill is ",{"type":43,"tag":56,"props":1157,"children":1158},{},[1159],{"type":49,"value":1160},"agent guidance",{"type":49,"value":1162},", not a samples or scripts\nbundle. To keep the boundary clean, it deliberately does not\ncontain — and pull requests should not add:",{"type":43,"tag":201,"props":1164,"children":1165},{},[1166,1191,1201,1211,1228],{"type":43,"tag":205,"props":1167,"children":1168},{},[1169,1182,1184,1189],{"type":43,"tag":56,"props":1170,"children":1171},{},[1172,1174,1180],{"type":49,"value":1173},"Specific flag strings, event names, or mode tokens\nbeyond what the public DOCA DPA Tools page and ",{"type":43,"tag":64,"props":1175,"children":1177},{"className":1176},[],[1178],{"type":49,"value":1179},"--help",{"type":49,"value":1181},"\ndocument.",{"type":49,"value":1183}," The DPA programming events surface evolves\nrelease to release; ",{"type":43,"tag":64,"props":1185,"children":1187},{"className":1186},[],[1188],{"type":49,"value":1179},{"type":49,"value":1190}," on the installed binary is\nthe authoritative inventory.",{"type":43,"tag":205,"props":1192,"children":1193},{},[1194,1199],{"type":43,"tag":56,"props":1195,"children":1196},{},[1197],{"type":49,"value":1198},"Pre-baked example traces or expected event timings.",{"type":49,"value":1200},"\nTrace output is workload-, DPA-image-, BlueField-, and\nfirmware-specific; a captured example pinned to one\nsetup misleads operators elsewhere.",{"type":43,"tag":205,"props":1202,"children":1203},{},[1204,1209],{"type":43,"tag":56,"props":1205,"children":1206},{},[1207],{"type":49,"value":1208},"Wrappers, parsers, or rendering scripts",{"type":49,"value":1210}," in any\nlanguage that consume the binary trace format. The\nformat is documented; users who want to script against\nit should read the live guide and write the parser\nagainst their installed version.",{"type":43,"tag":205,"props":1212,"children":1213},{},[1214,1219,1221,1226],{"type":43,"tag":56,"props":1215,"children":1216},{},[1217],{"type":49,"value":1218},"A specific tuning recommendation derived from a single\ntrace.",{"type":49,"value":1220}," A DPA-side perf decision (move a sync, batch a\ncomm call, change a launch argument) is a workload\nquestion and the skill prescribes how to ",{"type":43,"tag":120,"props":1222,"children":1223},{},[1224],{"type":49,"value":1225},"capture and\nread",{"type":49,"value":1227}," traces — it refuses to translate a captured gap\ninto a kernel-rewrite recommendation without the user's\nown analysis.",{"type":43,"tag":205,"props":1229,"children":1230},{},[1231,1252,1254,1259,1261,1269],{"type":43,"tag":56,"props":1232,"children":1233},{},[1234,1236,1242,1244,1250],{"type":49,"value":1235},"A ",{"type":43,"tag":64,"props":1237,"children":1239},{"className":1238},[],[1240],{"type":49,"value":1241},"samples\u002F",{"type":49,"value":1243}," or ",{"type":43,"tag":64,"props":1245,"children":1247},{"className":1246},[],[1248],{"type":49,"value":1249},"reference\u002F",{"type":49,"value":1251}," subtree.",{"type":49,"value":1253}," This is a thin\nloader for a shipped CLI; substantive material lives on\nthe public page, in ",{"type":43,"tag":64,"props":1255,"children":1257},{"className":1256},[],[1258],{"type":49,"value":1179},{"type":49,"value":1260},", and in\n",{"type":43,"tag":73,"props":1262,"children":1263},{"href":147},[1264],{"type":43,"tag":64,"props":1265,"children":1267},{"className":1266},[],[1268],{"type":49,"value":154},{"type":49,"value":290},{"type":43,"tag":182,"props":1271,"children":1273},{"id":1272},"loading-order",[1274],{"type":49,"value":1275},"Loading order",{"type":43,"tag":1277,"props":1278,"children":1279},"ol",{},[1280,1293,1306],{"type":43,"tag":205,"props":1281,"children":1282},{},[1283,1285,1291],{"type":49,"value":1284},"Read this ",{"type":43,"tag":64,"props":1286,"children":1288},{"className":1287},[],[1289],{"type":49,"value":1290},"SKILL.md",{"type":49,"value":1292}," first to confirm the user's question\nis in scope (DPA-side high-level tracing, not DPA-side\nprogramming and not raw cycle profiling).",{"type":43,"tag":205,"props":1294,"children":1295},{},[1296],{"type":43,"tag":56,"props":1297,"children":1298},{},[1299,1301,1305],{"type":49,"value":1300},"For the event taxonomy, capture modes, overhead\ntradeoff, JSON config layout, version overlay, error\ntaxonomy, observability, and safety policy, see\n",{"type":43,"tag":73,"props":1302,"children":1303},{"href":110},[1304],{"type":49,"value":110},{"type":49,"value":290},{"type":43,"tag":205,"props":1307,"children":1308},{},[1309],{"type":43,"tag":56,"props":1310,"children":1311},{},[1312,1314,1319,1320,1325,1326,1331,1332,1337,1338,1343,1344,1349,1350,1355,1356,1361,1363,1367],{"type":49,"value":1313},"For the documented invocations and the\ncapture → decode → render workflow — ",{"type":43,"tag":64,"props":1315,"children":1317},{"className":1316},[],[1318],{"type":49,"value":1034},{"type":49,"value":928},{"type":43,"tag":64,"props":1321,"children":1323},{"className":1322},[],[1324],{"type":49,"value":1042},{"type":49,"value":126},{"type":43,"tag":64,"props":1327,"children":1329},{"className":1328},[],[1330],{"type":49,"value":1050},{"type":49,"value":126},{"type":43,"tag":64,"props":1333,"children":1335},{"className":1334},[],[1336],{"type":49,"value":1058},{"type":49,"value":126},{"type":43,"tag":64,"props":1339,"children":1341},{"className":1340},[],[1342],{"type":49,"value":1066},{"type":49,"value":126},{"type":43,"tag":64,"props":1345,"children":1347},{"className":1346},[],[1348],{"type":49,"value":1096},{"type":49,"value":126},{"type":43,"tag":64,"props":1351,"children":1353},{"className":1352},[],[1354],{"type":49,"value":1104},{"type":49,"value":928},{"type":43,"tag":64,"props":1357,"children":1359},{"className":1358},[],[1360],{"type":49,"value":1112},{"type":49,"value":1362}," — see ",{"type":43,"tag":73,"props":1364,"children":1365},{"href":75},[1366],{"type":49,"value":75},{"type":49,"value":290},{"type":43,"tag":182,"props":1369,"children":1371},{"id":1370},"related-skills",[1372],{"type":49,"value":1373},"Related skills",{"type":43,"tag":201,"props":1375,"children":1376},{},[1377,1397,1417,1438,1482,1502,1527],{"type":43,"tag":205,"props":1378,"children":1379},{},[1380,1388,1390,1395],{"type":43,"tag":73,"props":1381,"children":1382},{"href":147},[1383],{"type":43,"tag":64,"props":1384,"children":1386},{"className":1385},[],[1387],{"type":49,"value":154},{"type":49,"value":1389}," — the host-side\nDPA control library whose loaded application image the\ntracer captures. Pair them in every DPA debug session:\n",{"type":43,"tag":64,"props":1391,"children":1393},{"className":1392},[],[1394],{"type":49,"value":154},{"type":49,"value":1396}," brings the workload up; the tracer captures\nwhat the workload does at the DPA programming event\nlayer. Conflating the library with the tracer is the\nmost common DPA-debug first-touch error.",{"type":43,"tag":205,"props":1398,"children":1399},{},[1400,1408,1410,1415],{"type":43,"tag":73,"props":1401,"children":1402},{"href":159},[1403],{"type":43,"tag":64,"props":1404,"children":1406},{"className":1405},[],[1407],{"type":49,"value":166},{"type":49,"value":1409}," — the\ncross-cutting debug ladder. The tracer slots in at the\n",{"type":43,"tag":120,"props":1411,"children":1412},{},[1413],{"type":49,"value":1414},"runtime",{"type":49,"value":1416}," layer as the DPA-side ground truth before any\nDPA-side perf or correctness conclusion is made.",{"type":43,"tag":205,"props":1418,"children":1419},{},[1420,1428,1430,1436],{"type":43,"tag":73,"props":1421,"children":1422},{"href":800},[1423],{"type":43,"tag":64,"props":1424,"children":1426},{"className":1425},[],[1427],{"type":49,"value":807},{"type":49,"value":1429},"\n— routing to the public DOCA DPA Tools page on\n",{"type":43,"tag":64,"props":1431,"children":1433},{"className":1432},[],[1434],{"type":49,"value":1435},"docs.nvidia.com",{"type":49,"value":1437}," and the rest of the public DOCA\ndocumentation set.",{"type":43,"tag":205,"props":1439,"children":1440},{},[1441,1449,1451,1457,1459,1467,1469,1480],{"type":43,"tag":73,"props":1442,"children":1443},{"href":517},[1444],{"type":43,"tag":64,"props":1445,"children":1447},{"className":1446},[],[1448],{"type":49,"value":524},{"type":49,"value":1450}," — canonical\nDOCA version-handling rules. The ",{"type":43,"tag":64,"props":1452,"children":1454},{"className":1453},[],[1455],{"type":49,"value":1456},"## Version compatibility",{"type":49,"value":1458}," section in ",{"type":43,"tag":73,"props":1460,"children":1461},{"href":110},[1462],{"type":43,"tag":64,"props":1463,"children":1465},{"className":1464},[],[1466],{"type":49,"value":110},{"type":49,"value":1468},"\nis a concise overlay that redirects here for the body\nand adds the ",{"type":43,"tag":120,"props":1470,"children":1471},{},[1472,1473,1478],{"type":49,"value":531},{"type":43,"tag":64,"props":1474,"children":1476},{"className":1475},[],[1477],{"type":49,"value":154},{"type":49,"value":1479}," library ↔ DPACC\ncompiler",{"type":49,"value":1481}," matching rule.",{"type":43,"tag":205,"props":1483,"children":1484},{},[1485,1493,1495,1500],{"type":43,"tag":73,"props":1486,"children":1487},{"href":941},[1488],{"type":43,"tag":64,"props":1489,"children":1491},{"className":1490},[],[1492],{"type":49,"value":948},{"type":49,"value":1494}," — env\npreparation, install verification, DPACC compiler\ninstall \u002F verification, BlueField mode (the DPA\nprocessor must be exposed before any tracing is\nmeaningful), and the ",{"type":43,"tag":120,"props":1496,"children":1497},{},[1498],{"type":49,"value":1499},"I have no install yet",{"type":49,"value":1501}," path with\nthe public NGC DOCA container.",{"type":43,"tag":205,"props":1503,"children":1504},{},[1505,1515,1517,1525],{"type":43,"tag":73,"props":1506,"children":1508},{"href":1507},"..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md",[1509],{"type":43,"tag":64,"props":1510,"children":1512},{"className":1511},[],[1513],{"type":49,"value":1514},"doca-structured-tools-contract",{"type":49,"value":1516},"\n— the bundle's detect → prefer → fall back → report\ncontract for structured helper tools. The command\nappendix in ",{"type":43,"tag":73,"props":1518,"children":1519},{"href":75},[1520],{"type":43,"tag":64,"props":1521,"children":1523},{"className":1522},[],[1524],{"type":49,"value":75},{"type":49,"value":1526}," honors this contract.",{"type":43,"tag":205,"props":1528,"children":1529},{},[1530,1540,1542,1547,1549,1554],{"type":43,"tag":73,"props":1531,"children":1533},{"href":1532},"..\u002F..\u002Fdoca-programming-guide\u002FSKILL.md",[1534],{"type":43,"tag":64,"props":1535,"children":1537},{"className":1536},[],[1538],{"type":49,"value":1539},"doca-programming-guide",{"type":49,"value":1541},"\n— general DOCA programming patterns shared by every\nlibrary \u002F tool surface, including the cross-library\n",{"type":43,"tag":64,"props":1543,"children":1545},{"className":1544},[],[1546],{"type":49,"value":233},{"type":49,"value":1548}," taxonomy this tool's host-side error\nlayer overlays on top of when host-side ",{"type":43,"tag":64,"props":1550,"children":1552},{"className":1551},[],[1553],{"type":49,"value":154},{"type":49,"value":1555},"\ncalls fail in tandem.",{"type":43,"tag":52,"props":1557,"children":1558},{},[1559,1561,1567,1569,1575,1577,1582,1584,1589,1591,1599,1601,1607,1609,1614],{"type":49,"value":1560},"The DPA-side companion libraries ",{"type":43,"tag":64,"props":1562,"children":1564},{"className":1563},[],[1565],{"type":49,"value":1566},"doca-dpa-comms",{"type":49,"value":1568}," (comm\nprimitives the DPA kernel itself calls) and\n",{"type":43,"tag":64,"props":1570,"children":1572},{"className":1571},[],[1573],{"type":49,"value":1574},"doca-dpa-verbs",{"type":49,"value":1576}," (RDMA verbs the DPA kernel itself calls)\nare ",{"type":43,"tag":56,"props":1578,"children":1579},{},[1580],{"type":49,"value":1581},"different artifacts",{"type":49,"value":1583}," that the tracer's ",{"type":43,"tag":120,"props":1585,"children":1586},{},[1587],{"type":49,"value":1588},"DPA\nprogramming events",{"type":49,"value":1590}," surface visibly names; for the\nDPA-side programming model itself, route through\n",{"type":43,"tag":73,"props":1592,"children":1593},{"href":800},[1594],{"type":43,"tag":64,"props":1595,"children":1597},{"className":1596},[],[1598],{"type":49,"value":807},{"type":49,"value":1600},"\nto the public DOCA DPA-Comms and DPA-Verbs guides and to\nthe shipped ",{"type":43,"tag":64,"props":1602,"children":1604},{"className":1603},[],[1605],{"type":49,"value":1606},"\u002Fopt\u002Fmellanox\u002Fdoca\u002Fsamples\u002Fdoca_dpa\u002F",{"type":49,"value":1608}," samples.\nThis tool ",{"type":43,"tag":120,"props":1610,"children":1611},{},[1612],{"type":49,"value":1613},"traces",{"type":49,"value":1615}," their use; it does not redefine them.",{"items":1617,"total":1776},[1618,1636,1654,1665,1677,1691,1704,1718,1731,1742,1756,1765],{"slug":1619,"name":1619,"fn":1620,"description":1621,"org":1622,"tags":1623,"stars":1633,"repoUrl":1634,"updatedAt":1635},"nemoclaw-user-guide","retrieve NemoClaw documentation and configuration","Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1624,1627,1630],{"name":1625,"slug":1626,"type":15},"Documentation","documentation",{"name":1628,"slug":1629,"type":15},"MCP","mcp",{"name":1631,"slug":1632,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1637,"name":1637,"fn":1638,"description":1639,"org":1640,"tags":1641,"stars":1651,"repoUrl":1652,"updatedAt":1653},"mcore-build-and-dependency","manage Megatron-LM development environments","Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1642,1645,1648],{"name":1643,"slug":1644,"type":15},"Containers","containers",{"name":1646,"slug":1647,"type":15},"Deployment","deployment",{"name":1649,"slug":1650,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1655,"name":1655,"fn":1656,"description":1657,"org":1658,"tags":1659,"stars":1651,"repoUrl":1652,"updatedAt":1664},"mcore-bump-base-image","update NVIDIA PyTorch base images","Bump the NVIDIA PyTorch base image (`nvcr.io\u002Fnvidia\u002Fpytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker\u002F.ngc_version.dev` and GitLab CI in `.gitlab\u002Fstages\u002F01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1660,1663],{"name":1661,"slug":1662,"type":15},"CI\u002FCD","ci-cd",{"name":1646,"slug":1647,"type":15},"2026-07-14T05:25:59.97109",{"slug":1666,"name":1666,"fn":1667,"description":1668,"org":1669,"tags":1670,"stars":1651,"repoUrl":1652,"updatedAt":1676},"mcore-cicd","manage CI\u002FCD pipelines for Megatron-LM","CI\u002FCD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request\u002FBRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1671,1672,1673],{"name":1661,"slug":1662,"type":15},{"name":1646,"slug":1647,"type":15},{"name":1674,"slug":1675,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1678,"name":1678,"fn":1679,"description":1680,"org":1681,"tags":1682,"stars":1651,"repoUrl":1652,"updatedAt":1690},"mcore-create-issue","investigate CI failures and create issues","Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1683,1686,1687],{"name":1684,"slug":1685,"type":15},"Debugging","debugging",{"name":1674,"slug":1675,"type":15},{"name":1688,"slug":1689,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1692,"name":1692,"fn":1693,"description":1694,"org":1695,"tags":1696,"stars":1651,"repoUrl":1652,"updatedAt":1703},"mcore-linting-and-formatting","lint and format Megatron-LM code","Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1697,1700],{"name":1698,"slug":1699,"type":15},"Best Practices","best-practices",{"name":1701,"slug":1702,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1705,"name":1705,"fn":1706,"description":1707,"org":1708,"tags":1709,"stars":1651,"repoUrl":1652,"updatedAt":1717},"mcore-migrate-gpt-to-hybrid","migrate Megatron-LM models to HybridModel","Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1710,1713,1716],{"name":1711,"slug":1712,"type":15},"Machine Learning","machine-learning",{"name":1714,"slug":1715,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1719,"name":1719,"fn":1720,"description":1721,"org":1722,"tags":1723,"stars":1651,"repoUrl":1652,"updatedAt":1730},"mcore-onboard-gb200-1node-tests","onboard functional tests for GB200","Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1724,1727],{"name":1725,"slug":1726,"type":15},"QA","qa",{"name":1728,"slug":1729,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1732,"name":1732,"fn":1733,"description":1734,"org":1735,"tags":1736,"stars":1651,"repoUrl":1652,"updatedAt":1741},"mcore-run-on-slurm","launch distributed training jobs on SLURM","How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1737,1738],{"name":1646,"slug":1647,"type":15},{"name":1739,"slug":1740,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1743,"name":1743,"fn":1744,"description":1745,"org":1746,"tags":1747,"stars":1651,"repoUrl":1652,"updatedAt":1755},"mcore-split-pr","split pull requests to reduce review load","Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1748,1751,1752],{"name":1749,"slug":1750,"type":15},"Code Review","code-review",{"name":1674,"slug":1675,"type":15},{"name":1753,"slug":1754,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1757,"name":1757,"fn":1758,"description":1759,"org":1760,"tags":1761,"stars":1651,"repoUrl":1652,"updatedAt":1764},"mcore-testing","run and manage Megatron-LM tests","Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1762,1763],{"name":1725,"slug":1726,"type":15},{"name":1728,"slug":1729,"type":15},"2026-07-14T05:25:54.928983",{"slug":1766,"name":1766,"fn":1767,"description":1768,"org":1769,"tags":1770,"stars":1651,"repoUrl":1652,"updatedAt":1775},"nightly-sync","manage nightly main-to-dev sync workflows","Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1771,1774],{"name":1772,"slug":1773,"type":15},"Automation","automation",{"name":1661,"slug":1662,"type":15},"2026-07-30T05:29:03.275638",496,{"items":1778,"total":1874},[1779,1796,1806,1820,1830,1845,1860],{"slug":1780,"name":1780,"fn":1781,"description":1782,"org":1783,"tags":1784,"stars":23,"repoUrl":24,"updatedAt":1795},"accelerated-computing-cudf","accelerate data processing with cuDF","Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV\u002FParquet I\u002FO, nullable semantics, and multi-GPU DataFrame workloads.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1785,1788,1791,1792],{"name":1786,"slug":1787,"type":15},"Data Analysis","data-analysis",{"name":1789,"slug":1790,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":1793,"slug":1794,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":1797,"name":1797,"fn":1798,"description":1799,"org":1800,"tags":1801,"stars":23,"repoUrl":24,"updatedAt":1805},"aiq-deploy","deploy and manage NVIDIA AI-Q infrastructure","Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1802,1803,1804],{"name":1646,"slug":1647,"type":15},{"name":1739,"slug":1740,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1807,"name":1807,"fn":1808,"description":1809,"org":1810,"tags":1811,"stars":23,"repoUrl":24,"updatedAt":1819},"aiq-research","conduct deep research with AI-Q","Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1812,1815,1816],{"name":1813,"slug":1814,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1817,"slug":1818,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1821,"name":1821,"fn":1822,"description":1823,"org":1824,"tags":1825,"stars":23,"repoUrl":24,"updatedAt":1829},"amc-run-sample-calibration","run AMC sample dataset calibration","Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1826,1827,1828],{"name":1786,"slug":1787,"type":15},{"name":9,"slug":8,"type":15},{"name":1728,"slug":1729,"type":15},"2026-07-17T05:29:03.913266",{"slug":1831,"name":1831,"fn":1832,"description":1833,"org":1834,"tags":1835,"stars":23,"repoUrl":24,"updatedAt":1844},"amc-run-video-calibration","calibrate video datasets with AutoMagicCalib","Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP\u002Flive streams, use amc-run-rtsp-calibration instead.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1836,1837,1840,1841],{"name":1772,"slug":1773,"type":15},{"name":1838,"slug":1839,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":1842,"slug":1843,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":1846,"name":1846,"fn":1847,"description":1848,"org":1849,"tags":1850,"stars":23,"repoUrl":24,"updatedAt":1859},"amc-setup-calibration-stack","deploy AutoMagicCalib microservice with Docker","Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1851,1852,1855,1856],{"name":1646,"slug":1647,"type":15},{"name":1853,"slug":1854,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1857,"slug":1858,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1861,"name":1861,"fn":1862,"description":1863,"org":1864,"tags":1865,"stars":23,"repoUrl":24,"updatedAt":1873},"cudaq-guide","develop quantum applications with CUDA-Q","CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1866,1867,1870],{"name":9,"slug":8,"type":15},{"name":1868,"slug":1869,"type":15},"Quantum Computing","quantum-computing",{"name":1871,"slug":1872,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]