[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-doca-flow-perf":3,"mdc--n5slsw-key":34,"related-repo-nvidia-doca-flow-perf":1392,"related-org-nvidia-doca-flow-perf":1496},{"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-flow-perf","measure DOCA Flow pipeline performance","Use this skill when the user is measuring the host or DPU-CPU control-plane rate of a DOCA Flow pipeline with doca_flow_perf — picking a JSON policy from configs\u002F, choosing the DPDK or DOCA backend, running the single-iteration smoke then the iterative eval loop, interpreting per-iteration CPU cycles and num_pushed \u002F num_failed, or capturing the four-tuple (DOCA version, BlueField\u002Ffirmware, JSON policy, worker\u002Fqueue\u002Fburst config) that makes a Kops\u002Fsec number defensible. Trigger even when the user does not explicitly mention \"doca-flow-perf\" — typical implicit phrasings include \"how many rules per second can my BlueField insert\", \"5-tuple hairpin rule rate\", \"Kops\u002Fsec for steering\", \"flow-perf number does not match release notes\", \"DPDK vs DOCA benchmark\", or \"rule-install variance too high\". Refuse and route elsewhere for optimizing a live Flow app (doca-flow-tune), the DPA-offloaded path (doca-flow-dpa-perf), dataplane throughput or latency, or library-internal pipe semantics — 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},"Performance","performance","tag",{"name":17,"slug":18,"type":15},"Monitoring","monitoring",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"Engineering","engineering",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-20T06:24:24.419227","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-flow-perf","---\nlicense: Apache-2.0\nname: doca-flow-perf\ndescription: >\n  Use this skill when the user is measuring the host or DPU-CPU\n  control-plane rate of a DOCA Flow pipeline with doca_flow_perf —\n  picking a JSON policy from configs\u002F, choosing the DPDK or DOCA\n  backend, running the single-iteration smoke then the iterative eval\n  loop, interpreting per-iteration CPU cycles and num_pushed \u002F\n  num_failed, or capturing the four-tuple (DOCA version,\n  BlueField\u002Ffirmware, JSON policy, worker\u002Fqueue\u002Fburst config) that\n  makes a Kops\u002Fsec number defensible. Trigger even when the user does\n  not explicitly mention \"doca-flow-perf\" — typical implicit phrasings\n  include \"how many rules per second can my BlueField insert\",\n  \"5-tuple hairpin rule rate\", \"Kops\u002Fsec for steering\", \"flow-perf\n  number does not match release notes\", \"DPDK vs DOCA benchmark\", or\n  \"rule-install variance too high\". Refuse and route elsewhere for\n  optimizing a live Flow app (doca-flow-tune), the DPA-offloaded path\n  (doca-flow-dpa-perf), dataplane throughput or latency, or\n  library-internal pipe semantics — those belong to other skills.\nmetadata:\n  kind: tool\ncompatibility: >\n  Requires DOCA SDK installed at \u002Fopt\u002Fmellanox\u002Fdoca on Linux (Ubuntu\n  22.04\u002F24.04 or RHEL\u002FSLES) with a BlueField DPU or ConnectX NIC\n  attached. The doca_flow_perf binary plus its configs\u002F JSON exemplars\n  must be present (the DOCA Flow Perf install component), with the\n  underlying doca-flow library healthy. Reads `pkg-config doca-flow`\n  and inspects \u002Fopt\u002Fmellanox\u002Fdoca\u002F{lib,include,samples,applications}.\n---\n\n# DOCA Flow Perf (`doca_flow_perf`)\n\n**Where to start:** This is a tool skill for invoking\n`doca_flow_perf`, the host-side \u002F DPU-CPU-side DOCA Flow\nperformance measurement tool. Open [`TASKS.md`](TASKS.md) and\nstart at [`## configure`](TASKS.md#configure) to commit to\nthe three-axis decision (target Flow pipeline shape × traffic\nclass × measurement axis) and pick the JSON policy file that\nexpresses the workload, then [`## run`](TASKS.md#run) for the\nsingle-iteration smoke, then [`## test`](TASKS.md#test) for\nthe iterative eval loop that produces a defensible\nKops\u002Fsec-class number. Open [`CAPABILITIES.md`](CAPABILITIES.md)\nwhen the question is *what `doca_flow_perf` measures and what\nit deliberately does not measure*, *how its DPDK and DOCA\nbackends differ behind the same JSON contract*, *how to\ninterpret the per-iteration CPU-cycle output*, or *how it\ndiffers from `doca-flow-tune` (measurement vs. optimization)\nand `doca-flow-dpa-perf` (host \u002F DPU-CPU vs. DPA-offloaded\npath)*. If DOCA is not installed, route to\n[`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) first; if the\ntarget measurement is the DPA-offloaded path, route to\n[`doca-flow-dpa-perf`](..\u002Fdoca-flow-dpa-perf\u002FSKILL.md)\ninstead; if the goal is to optimize an already-deployed Flow\npipeline rather than measure a synthetic one, route to\n[`doca-flow-tune`](..\u002Fdoca-flow-tune\u002FSKILL.md) — `flow-perf`\nis a synthetic-driver microbenchmark, not a tuner of a live\nFlow application.\n\n## Example questions this skill answers well\n\n- *\"I want a defensible host-side baseline number for how\n  many `doca-flow` rules per second a single BlueField-3 can\n  insert for a 5-tuple match-and-hairpin workload. Which\n  policy JSON do I start from, how do I make the result\n  reproducible, and what do I have to capture alongside the\n  number for it to be defensible?\"* — class-shaped flow-perf\n  baseline question; the agent walks the `configs\u002F` library,\n  the JSON contract, and the four-tuple capture rule.\n- *\"What is the difference between `doca-flow-perf`,\n  `doca-flow-dpa-perf`, and `doca-flow-tune`? They all\n  mention `doca-flow` and `perf` in their names — when do I\n  reach for each?\"* — measurement-vs-optimization plus\n  host-vs-DPA-path; the agent surfaces the boundaries.\n- *\"My policy JSON looks like the example, but the reported\n  Kops\u002Fsec is dramatically lower than the published numbers I\n  see in NVIDIA's release notes. What variables do I have to\n  control before I can trust the comparison?\"* — methodology\n  question; the agent walks the controllable axes (number of\n  workers, queue depth, burst size, fixed-vs-incremented match\n  fields, DPDK vs DOCA backend, BlueField mode, driver \u002F\n  firmware).\n- *\"I have a workload that does not match any of the shipped\n  policy JSONs in `configs\u002F`. How do I author a new policy\n  JSON, what is the JSON schema in broad strokes, and what\n  changes when I switch a match field from `mode: fixed` to\n  `mode: increase`?\"* — JSON authoring question; the agent\n  walks the shipped configs as exemplars and refuses to\n  invent schema fields not present in the source tree.\n- *\"What does the tool actually NOT measure? I am trying to\n  understand whether a flow-perf number tells me anything\n  about end-to-end traffic latency or just about the\n  rule-programming control-plane rate.\"* — methodology\n  perimeter question; the agent draws a hard line: this tool\n  measures rule install \u002F delete (control-plane) rate plus\n  optional query rate, NOT dataplane latency, NOT dataplane\n  throughput, NOT end-to-end application performance.\n- *\"I see two backends — DPDK and DOCA — behind the same\n  JSON. When do I pick which, and what does the choice mean\n  for the result I report?\"* — backend choice question; the\n  agent walks the DPDK-backend vs. DOCA-backend trade-off and\n  insists the operator REPORT which one they used.\n\n## Audience\n\nExperienced AI agents and platform \u002F network engineers who\nare comfortable with the `doca-flow` programming model and\nthe DPDK control-plane, who want a *defensible* number for\nthe host-side \u002F DPU-CPU-side Flow rule-install \u002F rule-delete\nrate. Readers are expected to know that the published\nnumbers in NVIDIA release notes are run with very specific\npreconditions (specific DOCA version, specific BlueField\nfirmware, specific traffic class) and that any number they\nproduce locally must explicitly state those preconditions.\n\nThis skill is NOT for:\n\n- operators who want to optimize an already-deployed\n  `doca-flow` application — that is\n  [`doca-flow-tune`](..\u002Fdoca-flow-tune\u002FSKILL.md);\n- operators measuring the DPA-offloaded Flow path — that is\n  [`doca-flow-dpa-perf`](..\u002Fdoca-flow-dpa-perf\u002FSKILL.md);\n- operators measuring end-to-end dataplane throughput or\n  latency — that is the application's responsibility,\n  layered on\n  [`doca-flow`](..\u002F..\u002Flibs\u002Fdoca-flow\u002FSKILL.md);\n- contributors authoring or modifying the tool itself.\n\n## Language scope\n\nUser interaction with `doca_flow_perf` is via:\n\n1. The shipped binary's command-line flags (documented by\n   `doca_flow_perf --help` and the public DOCA Flow Perf\n   guide on `docs.nvidia.com`).\n2. A JSON policy file describing the pipeline (ports, pipes,\n   matchers, actions, forwarding). The shipped `configs\u002F`\n   directory contains canned policies for the most common\n   traffic classes; new policies are authored by copying and\n   editing one of those.\n3. The tool's per-iteration output (CPU cycles per iteration,\n   number-processed, number-failed; reported via the tool's\n   stdout — the exact format is the public guide and the\n   binary's runtime output, NOT this skill's invention).\n\nThe skill itself is Markdown. There is no programmatic API on\ntop of `doca_flow_perf`; consumers of its results read its\nstdout \u002F captured logs.\n\n## When to load this skill\n\nLoad `doca-flow-perf` when ANY of the following is true:\n\n- the user mentions `doca_flow_perf`, `doca-flow-perf`, the\n  `configs\u002F` JSON library, or asks for a \"host-side flow\n  rules per second\" number;\n- the user wants to baseline an underlying Flow path (not\n  optimize a live application);\n- the user is comparing host-side \u002F DPU-CPU-side Flow\n  performance across DOCA releases, BlueField generations,\n  or firmware versions;\n- the user wants to design a new traffic class JSON and\n  needs to know which canned `configs\u002F` JSON to start from\n  and which fields they can change.\n\nCo-load this skill with:\n\n- [`doca-flow`](..\u002F..\u002Flibs\u002Fdoca-flow\u002FSKILL.md) (the\n  underlying library; flow-perf programs the same\n  matchers \u002F actions \u002F pipes the library exposes);\n- [`doca-flow-tune`](..\u002Fdoca-flow-tune\u002FSKILL.md) (the\n  measurement-vs-optimization distinction is the most\n  common confusion);\n- [`doca-flow-dpa-perf`](..\u002Fdoca-flow-dpa-perf\u002FSKILL.md)\n  (the host-vs-DPA-path distinction is the second most\n  common confusion);\n- [`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) (the\n  four-way version match every reported flow-perf number\n  must carry);\n- [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md) and\n  [`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) for the\n  env-side debug ladder.\n\nDo NOT load this skill when the user wants to optimize a live\nFlow application (route to\n[`doca-flow-tune`](..\u002Fdoca-flow-tune\u002FSKILL.md)) or measure\nthe DPA-offloaded path (route to\n[`doca-flow-dpa-perf`](..\u002Fdoca-flow-dpa-perf\u002FSKILL.md)).\n\n## What this skill provides\n\nThree companion files in this directory, each owning a\ndifferent question shape:\n\n- [`SKILL.md`](SKILL.md) — this file. Audience, scope,\n  loading order, related skills. Routes everything else.\n- [`CAPABILITIES.md`](CAPABILITIES.md) — *what\n  `doca_flow_perf` is*, what it measures, what it\n  deliberately doesn't measure, the DPDK-vs-DOCA backend\n  duality, the JSON contract surface, the per-iteration\n  output interpretation, version compatibility (versioned\n  with `doca-flow` and `doca-version`), the layered error\n  taxonomy, observability, and the safety policy overlay.\n- [`TASKS.md`](TASKS.md) — the procedural verbs (`configure`,\n  `run`, `test`, `debug`, etc.) plus a `doca_flow_perf`-\n  specific command appendix and the agent-side `use`\n  workflow that consumes the captured per-iteration output.\n\nThe combined skill teaches an AI agent to drive the\n*measurement-class* of `doca_flow_perf` questions: pick a\nshipped or author-new policy JSON, run the single-iteration\nsmoke, run the iterative eval loop, capture the four-tuple\nthat makes the resulting number defensible, interpret the\noutput, and route every adjacent question (tune the live\napp, measure the DPA path, optimize the firmware) to the\nright neighbouring skill.\n\n## What this skill deliberately does not ship\n\n- **End-to-end dataplane throughput or latency\n  measurement.** `doca_flow_perf` measures the\n  *control-plane* rate of programming rules, plus optional\n  per-entry query timing. It does NOT measure how fast\n  packets traverse the resulting rules in the dataplane.\n  That is the application's responsibility, layered on\n  [`doca-flow`](..\u002F..\u002Flibs\u002Fdoca-flow\u002FSKILL.md). The agent\n  must say this explicitly when the operator asks for \"Flow\n  throughput\".\n- **DPA-offloaded Flow path measurement.** Route to\n  [`doca-flow-dpa-perf`](..\u002Fdoca-flow-dpa-perf\u002FSKILL.md).\n- **Optimization of a deployed Flow application.** Route to\n  [`doca-flow-tune`](..\u002Fdoca-flow-tune\u002FSKILL.md). flow-perf\n  is a synthetic driver of a JSON-described pipeline, not a\n  tuner of a live one.\n- **A canonical \"right answer\" Kops\u002Fsec number.** The agent\n  refuses to quote published numbers from memory as\n  authoritative; the published numbers live in NVIDIA's\n  release notes per the DOCA version and BlueField\n  generation, and the operator must reproduce on their own\n  exact preconditions before comparing.\n- **Invented JSON schema fields.** The agent does NOT invent\n  policy JSON keys that are not present in the shipped\n  `configs\u002F` exemplars. If a key the operator wants is not\n  in any shipped exemplar, the agent says so and routes to\n  the public DOCA Flow Perf guide.\n- **Library-internal `doca-flow` API explanations.** The\n  underlying matchers and actions belong to\n  [`doca-flow`](..\u002F..\u002Flibs\u002Fdoca-flow\u002FSKILL.md); this skill\n  references them but does not duplicate the library's API\n  documentation.\n- **Cross-tool benchmarking apples-to-apples claims** when\n  preconditions differ. Two flow-perf numbers from different\n  DOCA versions \u002F BlueField generations \u002F firmware versions\n  are NOT directly comparable; the agent insists on the\n  four-tuple capture so consumers can judge.\n\n## Loading order\n\nWhen a `doca_flow_perf` question arrives:\n\n1. Confirm DOCA is installed and the binary plus the\n   `configs\u002F` JSON library are reachable — if not, route to\n   [`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md);\n2. Confirm the underlying `doca-flow` library is healthy on\n   the device — if not, route to\n   [`doca-flow TASKS.md ## test`](..\u002F..\u002Flibs\u002Fdoca-flow\u002FTASKS.md#test);\n3. Confirm the user wants to *measure*, not *optimize* —\n   if optimize, route to\n   [`doca-flow-tune`](..\u002Fdoca-flow-tune\u002FSKILL.md);\n4. Confirm the target path is host \u002F DPU-CPU, not DPA — if\n   DPA, route to\n   [`doca-flow-dpa-perf`](..\u002Fdoca-flow-dpa-perf\u002FSKILL.md);\n5. Read [`CAPABILITIES.md`](CAPABILITIES.md) to commit to\n   the three-axis decision (pipeline shape × traffic class ×\n   measurement axis);\n6. Read [`TASKS.md`](TASKS.md) and walk\n   `## configure → ## run → ## test → ## debug` in that\n   order; do NOT start with `## run` without the\n   `## configure` precondition step.\n\n## Related skills\n\nCross-link conventions follow the bundle's relative path\ncontract from `tools\u002F\u003CX>\u002F`:\n\n- [`doca-flow`](..\u002F..\u002Flibs\u002Fdoca-flow\u002FSKILL.md) — the\n  underlying library. flow-perf programs Flow pipes, entries,\n  matchers, and actions; the library is the source of truth\n  for the API surface flow-perf exercises.\n- [`doca-flow-tune`](..\u002Fdoca-flow-tune\u002FSKILL.md) — the\n  unified Flow tuning tool. **Measurement vs. optimization\n  boundary** lives here. Ask: \"do I want a number, or do I\n  want to change the deployed pipeline?\"\n- [`doca-flow-dpa-perf`](..\u002Fdoca-flow-dpa-perf\u002FSKILL.md) —\n  the DPA-offloaded Flow performance tool. **Host \u002F\n  DPU-CPU vs. DPA path boundary** lives here. Ask: \"am I\n  measuring the path that executes on the CPU, or the path\n  that executes on the DPA processor?\"\n- [`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) — every\n  reported flow-perf number must come with the four-way\n  match (host package, kernel module, firmware, target\n  application's linked `doca-flow` version) and the BlueField\n  \u002F ConnectX generation. flow-perf overlays this rule, not\n  contradicts it.\n- [`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) — DOCA install\n  posture; routing for \"is the binary even here?\" questions.\n- [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md) — the\n  cross-cutting debug ladder for env-side issues (driver,\n  firmware, BlueField mode, kernel module).\n- [`doca-bench`](..\u002Fdoca-bench\u002FSKILL.md) — a peer\n  benchmarking tool with a broader scope (multiple DOCA\n  primitives, not just Flow). flow-perf is the Flow-specific\n  microbenchmark; doca-bench is the broader workload\n  benchmark.\n- [`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md)\n  — routing to the public `docs.nvidia.com` DOCA Flow Perf\n  page, release notes, and forums for release-specific\n  published numbers and reproducibility notes.\n- [`doca-structured-tools-contract`](..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md)\n  — the agent's detect → prefer → fall back → report contract\n  for the structured helpers (`doca-env --json`,\n  `doca-capability-snapshot`, `version-matrix.json`)\n  flow-perf preconditions rely on.\n- [`doca-hardware-safety`](..\u002F..\u002Fdoca-hardware-safety\u002FSKILL.md)\n  — the canonical hardware-safety meta-policy that\n  [`CAPABILITIES.md ## Safety policy`](CAPABILITIES.md#safety-policy)\n  overlays.\n\nThis skill assumes the surrounding doca-flow application is\nthe operator's existing source artifact; flow-perf does not\nship a sample doca-flow application of its own.\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 DPU or ConnectX NIC attached. The doca_flow_perf binary plus its configs\u002F JSON exemplars must be present (the DOCA Flow Perf install component), with the underlying doca-flow library healthy. Reads `pkg-config doca-flow` and inspects \u002Fopt\u002Fmellanox\u002Fdoca\u002F{lib,include,samples,applications}.\n",{"type":40,"children":41},"root",[42,60,231,238,378,384,403,408,467,473,485,527,539,545,557,607,612,694,718,724,729,842,861,867,1007,1013,1025,1165,1171,1184,1387],{"type":43,"tag":44,"props":45,"children":47},"element","h1",{"id":46},"doca-flow-perf-doca_flow_perf",[48,51,58],{"type":49,"value":50},"text","DOCA Flow Perf (",{"type":43,"tag":52,"props":53,"children":55},"code",{"className":54},[],[56],{"type":49,"value":57},"doca_flow_perf",{"type":49,"value":59},")",{"type":43,"tag":61,"props":62,"children":63},"p",{},[64,70,72,77,79,89,91,101,103,113,115,125,127,136,138,151,153,158,159,164,166,187,189,199,201,210,212,221,223,229],{"type":43,"tag":65,"props":66,"children":67},"strong",{},[68],{"type":49,"value":69},"Where to start:",{"type":49,"value":71}," This is a tool skill for invoking\n",{"type":43,"tag":52,"props":73,"children":75},{"className":74},[],[76],{"type":49,"value":57},{"type":49,"value":78},", the host-side \u002F DPU-CPU-side DOCA Flow\nperformance measurement tool. Open ",{"type":43,"tag":80,"props":81,"children":83},"a",{"href":82},"TASKS.md",[84],{"type":43,"tag":52,"props":85,"children":87},{"className":86},[],[88],{"type":49,"value":82},{"type":49,"value":90}," and\nstart at ",{"type":43,"tag":80,"props":92,"children":94},{"href":93},"TASKS.md#configure",[95],{"type":43,"tag":52,"props":96,"children":98},{"className":97},[],[99],{"type":49,"value":100},"## configure",{"type":49,"value":102}," to commit to\nthe three-axis decision (target Flow pipeline shape × traffic\nclass × measurement axis) and pick the JSON policy file that\nexpresses the workload, then ",{"type":43,"tag":80,"props":104,"children":106},{"href":105},"TASKS.md#run",[107],{"type":43,"tag":52,"props":108,"children":110},{"className":109},[],[111],{"type":49,"value":112},"## run",{"type":49,"value":114}," for the\nsingle-iteration smoke, then ",{"type":43,"tag":80,"props":116,"children":118},{"href":117},"TASKS.md#test",[119],{"type":43,"tag":52,"props":120,"children":122},{"className":121},[],[123],{"type":49,"value":124},"## test",{"type":49,"value":126}," for\nthe iterative eval loop that produces a defensible\nKops\u002Fsec-class number. Open ",{"type":43,"tag":80,"props":128,"children":130},{"href":129},"CAPABILITIES.md",[131],{"type":43,"tag":52,"props":132,"children":134},{"className":133},[],[135],{"type":49,"value":129},{"type":49,"value":137},"\nwhen the question is ",{"type":43,"tag":139,"props":140,"children":141},"em",{},[142,144,149],{"type":49,"value":143},"what ",{"type":43,"tag":52,"props":145,"children":147},{"className":146},[],[148],{"type":49,"value":57},{"type":49,"value":150}," measures and what\nit deliberately does not measure",{"type":49,"value":152},", ",{"type":43,"tag":139,"props":154,"children":155},{},[156],{"type":49,"value":157},"how its DPDK and DOCA\nbackends differ behind the same JSON contract",{"type":49,"value":152},{"type":43,"tag":139,"props":160,"children":161},{},[162],{"type":49,"value":163},"how to\ninterpret the per-iteration CPU-cycle output",{"type":49,"value":165},", or ",{"type":43,"tag":139,"props":167,"children":168},{},[169,171,177,179,185],{"type":49,"value":170},"how it\ndiffers from ",{"type":43,"tag":52,"props":172,"children":174},{"className":173},[],[175],{"type":49,"value":176},"doca-flow-tune",{"type":49,"value":178}," (measurement vs. optimization)\nand ",{"type":43,"tag":52,"props":180,"children":182},{"className":181},[],[183],{"type":49,"value":184},"doca-flow-dpa-perf",{"type":49,"value":186}," (host \u002F DPU-CPU vs. DPA-offloaded\npath)",{"type":49,"value":188},". If DOCA is not installed, route to\n",{"type":43,"tag":80,"props":190,"children":192},{"href":191},"..\u002F..\u002Fdoca-setup\u002FSKILL.md",[193],{"type":43,"tag":52,"props":194,"children":196},{"className":195},[],[197],{"type":49,"value":198},"doca-setup",{"type":49,"value":200}," first; if the\ntarget measurement is the DPA-offloaded path, route to\n",{"type":43,"tag":80,"props":202,"children":204},{"href":203},"..\u002Fdoca-flow-dpa-perf\u002FSKILL.md",[205],{"type":43,"tag":52,"props":206,"children":208},{"className":207},[],[209],{"type":49,"value":184},{"type":49,"value":211},"\ninstead; if the goal is to optimize an already-deployed Flow\npipeline rather than measure a synthetic one, route to\n",{"type":43,"tag":80,"props":213,"children":215},{"href":214},"..\u002Fdoca-flow-tune\u002FSKILL.md",[216],{"type":43,"tag":52,"props":217,"children":219},{"className":218},[],[220],{"type":49,"value":176},{"type":49,"value":222}," — ",{"type":43,"tag":52,"props":224,"children":226},{"className":225},[],[227],{"type":49,"value":228},"flow-perf",{"type":49,"value":230},"\nis a synthetic-driver microbenchmark, not a tuner of a live\nFlow application.",{"type":43,"tag":232,"props":233,"children":235},"h2",{"id":234},"example-questions-this-skill-answers-well",[236],{"type":49,"value":237},"Example questions this skill answers well",{"type":43,"tag":239,"props":240,"children":241},"ul",{},[242,269,315,325,358,368],{"type":43,"tag":243,"props":244,"children":245},"li",{},[246,259,261,267],{"type":43,"tag":139,"props":247,"children":248},{},[249,251,257],{"type":49,"value":250},"\"I want a defensible host-side baseline number for how\nmany ",{"type":43,"tag":52,"props":252,"children":254},{"className":253},[],[255],{"type":49,"value":256},"doca-flow",{"type":49,"value":258}," rules per second a single BlueField-3 can\ninsert for a 5-tuple match-and-hairpin workload. Which\npolicy JSON do I start from, how do I make the result\nreproducible, and what do I have to capture alongside the\nnumber for it to be defensible?\"",{"type":49,"value":260}," — class-shaped flow-perf\nbaseline question; the agent walks the ",{"type":43,"tag":52,"props":262,"children":264},{"className":263},[],[265],{"type":49,"value":266},"configs\u002F",{"type":49,"value":268}," library,\nthe JSON contract, and the four-tuple capture rule.",{"type":43,"tag":243,"props":270,"children":271},{},[272,313],{"type":43,"tag":139,"props":273,"children":274},{},[275,277,282,284,289,291,296,298,303,305,311],{"type":49,"value":276},"\"What is the difference between ",{"type":43,"tag":52,"props":278,"children":280},{"className":279},[],[281],{"type":49,"value":4},{"type":49,"value":283},",\n",{"type":43,"tag":52,"props":285,"children":287},{"className":286},[],[288],{"type":49,"value":184},{"type":49,"value":290},", and ",{"type":43,"tag":52,"props":292,"children":294},{"className":293},[],[295],{"type":49,"value":176},{"type":49,"value":297},"? They all\nmention ",{"type":43,"tag":52,"props":299,"children":301},{"className":300},[],[302],{"type":49,"value":256},{"type":49,"value":304}," and ",{"type":43,"tag":52,"props":306,"children":308},{"className":307},[],[309],{"type":49,"value":310},"perf",{"type":49,"value":312}," in their names — when do I\nreach for each?\"",{"type":49,"value":314}," — measurement-vs-optimization plus\nhost-vs-DPA-path; the agent surfaces the boundaries.",{"type":43,"tag":243,"props":316,"children":317},{},[318,323],{"type":43,"tag":139,"props":319,"children":320},{},[321],{"type":49,"value":322},"\"My policy JSON looks like the example, but the reported\nKops\u002Fsec is dramatically lower than the published numbers I\nsee in NVIDIA's release notes. What variables do I have to\ncontrol before I can trust the comparison?\"",{"type":49,"value":324}," — methodology\nquestion; the agent walks the controllable axes (number of\nworkers, queue depth, burst size, fixed-vs-incremented match\nfields, DPDK vs DOCA backend, BlueField mode, driver \u002F\nfirmware).",{"type":43,"tag":243,"props":326,"children":327},{},[328,356],{"type":43,"tag":139,"props":329,"children":330},{},[331,333,338,340,346,348,354],{"type":49,"value":332},"\"I have a workload that does not match any of the shipped\npolicy JSONs in ",{"type":43,"tag":52,"props":334,"children":336},{"className":335},[],[337],{"type":49,"value":266},{"type":49,"value":339},". How do I author a new policy\nJSON, what is the JSON schema in broad strokes, and what\nchanges when I switch a match field from ",{"type":43,"tag":52,"props":341,"children":343},{"className":342},[],[344],{"type":49,"value":345},"mode: fixed",{"type":49,"value":347}," to\n",{"type":43,"tag":52,"props":349,"children":351},{"className":350},[],[352],{"type":49,"value":353},"mode: increase",{"type":49,"value":355},"?\"",{"type":49,"value":357}," — JSON authoring question; the agent\nwalks the shipped configs as exemplars and refuses to\ninvent schema fields not present in the source tree.",{"type":43,"tag":243,"props":359,"children":360},{},[361,366],{"type":43,"tag":139,"props":362,"children":363},{},[364],{"type":49,"value":365},"\"What does the tool actually NOT measure? I am trying to\nunderstand whether a flow-perf number tells me anything\nabout end-to-end traffic latency or just about the\nrule-programming control-plane rate.\"",{"type":49,"value":367}," — methodology\nperimeter question; the agent draws a hard line: this tool\nmeasures rule install \u002F delete (control-plane) rate plus\noptional query rate, NOT dataplane latency, NOT dataplane\nthroughput, NOT end-to-end application performance.",{"type":43,"tag":243,"props":369,"children":370},{},[371,376],{"type":43,"tag":139,"props":372,"children":373},{},[374],{"type":49,"value":375},"\"I see two backends — DPDK and DOCA — behind the same\nJSON. When do I pick which, and what does the choice mean\nfor the result I report?\"",{"type":49,"value":377}," — backend choice question; the\nagent walks the DPDK-backend vs. DOCA-backend trade-off and\ninsists the operator REPORT which one they used.",{"type":43,"tag":232,"props":379,"children":381},{"id":380},"audience",[382],{"type":49,"value":383},"Audience",{"type":43,"tag":61,"props":385,"children":386},{},[387,389,394,396,401],{"type":49,"value":388},"Experienced AI agents and platform \u002F network engineers who\nare comfortable with the ",{"type":43,"tag":52,"props":390,"children":392},{"className":391},[],[393],{"type":49,"value":256},{"type":49,"value":395}," programming model and\nthe DPDK control-plane, who want a ",{"type":43,"tag":139,"props":397,"children":398},{},[399],{"type":49,"value":400},"defensible",{"type":49,"value":402}," number for\nthe host-side \u002F DPU-CPU-side Flow rule-install \u002F rule-delete\nrate. Readers are expected to know that the published\nnumbers in NVIDIA release notes are run with very specific\npreconditions (specific DOCA version, specific BlueField\nfirmware, specific traffic class) and that any number they\nproduce locally must explicitly state those preconditions.",{"type":43,"tag":61,"props":404,"children":405},{},[406],{"type":49,"value":407},"This skill is NOT for:",{"type":43,"tag":239,"props":409,"children":410},{},[411,433,447,462],{"type":43,"tag":243,"props":412,"children":413},{},[414,416,421,423,431],{"type":49,"value":415},"operators who want to optimize an already-deployed\n",{"type":43,"tag":52,"props":417,"children":419},{"className":418},[],[420],{"type":49,"value":256},{"type":49,"value":422}," application — that is\n",{"type":43,"tag":80,"props":424,"children":425},{"href":214},[426],{"type":43,"tag":52,"props":427,"children":429},{"className":428},[],[430],{"type":49,"value":176},{"type":49,"value":432},";",{"type":43,"tag":243,"props":434,"children":435},{},[436,438,446],{"type":49,"value":437},"operators measuring the DPA-offloaded Flow path — that is\n",{"type":43,"tag":80,"props":439,"children":440},{"href":203},[441],{"type":43,"tag":52,"props":442,"children":444},{"className":443},[],[445],{"type":49,"value":184},{"type":49,"value":432},{"type":43,"tag":243,"props":448,"children":449},{},[450,452,461],{"type":49,"value":451},"operators measuring end-to-end dataplane throughput or\nlatency — that is the application's responsibility,\nlayered on\n",{"type":43,"tag":80,"props":453,"children":455},{"href":454},"..\u002F..\u002Flibs\u002Fdoca-flow\u002FSKILL.md",[456],{"type":43,"tag":52,"props":457,"children":459},{"className":458},[],[460],{"type":49,"value":256},{"type":49,"value":432},{"type":43,"tag":243,"props":463,"children":464},{},[465],{"type":49,"value":466},"contributors authoring or modifying the tool itself.",{"type":43,"tag":232,"props":468,"children":470},{"id":469},"language-scope",[471],{"type":49,"value":472},"Language scope",{"type":43,"tag":61,"props":474,"children":475},{},[476,478,483],{"type":49,"value":477},"User interaction with ",{"type":43,"tag":52,"props":479,"children":481},{"className":480},[],[482],{"type":49,"value":57},{"type":49,"value":484}," is via:",{"type":43,"tag":486,"props":487,"children":488},"ol",{},[489,510,522],{"type":43,"tag":243,"props":490,"children":491},{},[492,494,500,502,508],{"type":49,"value":493},"The shipped binary's command-line flags (documented by\n",{"type":43,"tag":52,"props":495,"children":497},{"className":496},[],[498],{"type":49,"value":499},"doca_flow_perf --help",{"type":49,"value":501}," and the public DOCA Flow Perf\nguide on ",{"type":43,"tag":52,"props":503,"children":505},{"className":504},[],[506],{"type":49,"value":507},"docs.nvidia.com",{"type":49,"value":509},").",{"type":43,"tag":243,"props":511,"children":512},{},[513,515,520],{"type":49,"value":514},"A JSON policy file describing the pipeline (ports, pipes,\nmatchers, actions, forwarding). The shipped ",{"type":43,"tag":52,"props":516,"children":518},{"className":517},[],[519],{"type":49,"value":266},{"type":49,"value":521},"\ndirectory contains canned policies for the most common\ntraffic classes; new policies are authored by copying and\nediting one of those.",{"type":43,"tag":243,"props":523,"children":524},{},[525],{"type":49,"value":526},"The tool's per-iteration output (CPU cycles per iteration,\nnumber-processed, number-failed; reported via the tool's\nstdout — the exact format is the public guide and the\nbinary's runtime output, NOT this skill's invention).",{"type":43,"tag":61,"props":528,"children":529},{},[530,532,537],{"type":49,"value":531},"The skill itself is Markdown. There is no programmatic API on\ntop of ",{"type":43,"tag":52,"props":533,"children":535},{"className":534},[],[536],{"type":49,"value":57},{"type":49,"value":538},"; consumers of its results read its\nstdout \u002F captured logs.",{"type":43,"tag":232,"props":540,"children":542},{"id":541},"when-to-load-this-skill",[543],{"type":49,"value":544},"When to load this skill",{"type":43,"tag":61,"props":546,"children":547},{},[548,550,555],{"type":49,"value":549},"Load ",{"type":43,"tag":52,"props":551,"children":553},{"className":552},[],[554],{"type":49,"value":4},{"type":49,"value":556}," when ANY of the following is true:",{"type":43,"tag":239,"props":558,"children":559},{},[560,585,590,595],{"type":43,"tag":243,"props":561,"children":562},{},[563,565,570,571,576,578,583],{"type":49,"value":564},"the user mentions ",{"type":43,"tag":52,"props":566,"children":568},{"className":567},[],[569],{"type":49,"value":57},{"type":49,"value":152},{"type":43,"tag":52,"props":572,"children":574},{"className":573},[],[575],{"type":49,"value":4},{"type":49,"value":577},", the\n",{"type":43,"tag":52,"props":579,"children":581},{"className":580},[],[582],{"type":49,"value":266},{"type":49,"value":584}," JSON library, or asks for a \"host-side flow\nrules per second\" number;",{"type":43,"tag":243,"props":586,"children":587},{},[588],{"type":49,"value":589},"the user wants to baseline an underlying Flow path (not\noptimize a live application);",{"type":43,"tag":243,"props":591,"children":592},{},[593],{"type":49,"value":594},"the user is comparing host-side \u002F DPU-CPU-side Flow\nperformance across DOCA releases, BlueField generations,\nor firmware versions;",{"type":43,"tag":243,"props":596,"children":597},{},[598,600,605],{"type":49,"value":599},"the user wants to design a new traffic class JSON and\nneeds to know which canned ",{"type":43,"tag":52,"props":601,"children":603},{"className":602},[],[604],{"type":49,"value":266},{"type":49,"value":606}," JSON to start from\nand which fields they can change.",{"type":43,"tag":61,"props":608,"children":609},{},[610],{"type":49,"value":611},"Co-load this skill with:",{"type":43,"tag":239,"props":613,"children":614},{},[615,628,641,654,669],{"type":43,"tag":243,"props":616,"children":617},{},[618,626],{"type":43,"tag":80,"props":619,"children":620},{"href":454},[621],{"type":43,"tag":52,"props":622,"children":624},{"className":623},[],[625],{"type":49,"value":256},{"type":49,"value":627}," (the\nunderlying library; flow-perf programs the same\nmatchers \u002F actions \u002F pipes the library exposes);",{"type":43,"tag":243,"props":629,"children":630},{},[631,639],{"type":43,"tag":80,"props":632,"children":633},{"href":214},[634],{"type":43,"tag":52,"props":635,"children":637},{"className":636},[],[638],{"type":49,"value":176},{"type":49,"value":640}," (the\nmeasurement-vs-optimization distinction is the most\ncommon confusion);",{"type":43,"tag":243,"props":642,"children":643},{},[644,652],{"type":43,"tag":80,"props":645,"children":646},{"href":203},[647],{"type":43,"tag":52,"props":648,"children":650},{"className":649},[],[651],{"type":49,"value":184},{"type":49,"value":653},"\n(the host-vs-DPA-path distinction is the second most\ncommon confusion);",{"type":43,"tag":243,"props":655,"children":656},{},[657,667],{"type":43,"tag":80,"props":658,"children":660},{"href":659},"..\u002F..\u002Fdoca-version\u002FSKILL.md",[661],{"type":43,"tag":52,"props":662,"children":664},{"className":663},[],[665],{"type":49,"value":666},"doca-version",{"type":49,"value":668}," (the\nfour-way version match every reported flow-perf number\nmust carry);",{"type":43,"tag":243,"props":670,"children":671},{},[672,682,684,692],{"type":43,"tag":80,"props":673,"children":675},{"href":674},"..\u002F..\u002Fdoca-debug\u002FSKILL.md",[676],{"type":43,"tag":52,"props":677,"children":679},{"className":678},[],[680],{"type":49,"value":681},"doca-debug",{"type":49,"value":683}," and\n",{"type":43,"tag":80,"props":685,"children":686},{"href":191},[687],{"type":43,"tag":52,"props":688,"children":690},{"className":689},[],[691],{"type":49,"value":198},{"type":49,"value":693}," for the\nenv-side debug ladder.",{"type":43,"tag":61,"props":695,"children":696},{},[697,699,707,709,717],{"type":49,"value":698},"Do NOT load this skill when the user wants to optimize a live\nFlow application (route to\n",{"type":43,"tag":80,"props":700,"children":701},{"href":214},[702],{"type":43,"tag":52,"props":703,"children":705},{"className":704},[],[706],{"type":49,"value":176},{"type":49,"value":708},") or measure\nthe DPA-offloaded path (route to\n",{"type":43,"tag":80,"props":710,"children":711},{"href":203},[712],{"type":43,"tag":52,"props":713,"children":715},{"className":714},[],[716],{"type":49,"value":184},{"type":49,"value":509},{"type":43,"tag":232,"props":719,"children":721},{"id":720},"what-this-skill-provides",[722],{"type":49,"value":723},"What this skill provides",{"type":43,"tag":61,"props":725,"children":726},{},[727],{"type":49,"value":728},"Three companion files in this directory, each owning a\ndifferent question shape:",{"type":43,"tag":239,"props":730,"children":731},{},[732,746,785],{"type":43,"tag":243,"props":733,"children":734},{},[735,744],{"type":43,"tag":80,"props":736,"children":738},{"href":737},"SKILL.md",[739],{"type":43,"tag":52,"props":740,"children":742},{"className":741},[],[743],{"type":49,"value":737},{"type":49,"value":745}," — this file. Audience, scope,\nloading order, related skills. Routes everything else.",{"type":43,"tag":243,"props":747,"children":748},{},[749,757,758,770,772,777,778,783],{"type":43,"tag":80,"props":750,"children":751},{"href":129},[752],{"type":43,"tag":52,"props":753,"children":755},{"className":754},[],[756],{"type":49,"value":129},{"type":49,"value":222},{"type":43,"tag":139,"props":759,"children":760},{},[761,763,768],{"type":49,"value":762},"what\n",{"type":43,"tag":52,"props":764,"children":766},{"className":765},[],[767],{"type":49,"value":57},{"type":49,"value":769}," is",{"type":49,"value":771},", what it measures, what it\ndeliberately doesn't measure, the DPDK-vs-DOCA backend\nduality, the JSON contract surface, the per-iteration\noutput interpretation, version compatibility (versioned\nwith ",{"type":43,"tag":52,"props":773,"children":775},{"className":774},[],[776],{"type":49,"value":256},{"type":49,"value":304},{"type":43,"tag":52,"props":779,"children":781},{"className":780},[],[782],{"type":49,"value":666},{"type":49,"value":784},"), the layered error\ntaxonomy, observability, and the safety policy overlay.",{"type":43,"tag":243,"props":786,"children":787},{},[788,796,798,804,805,811,812,818,819,825,827,832,834,840],{"type":43,"tag":80,"props":789,"children":790},{"href":82},[791],{"type":43,"tag":52,"props":792,"children":794},{"className":793},[],[795],{"type":49,"value":82},{"type":49,"value":797}," — the procedural verbs (",{"type":43,"tag":52,"props":799,"children":801},{"className":800},[],[802],{"type":49,"value":803},"configure",{"type":49,"value":283},{"type":43,"tag":52,"props":806,"children":808},{"className":807},[],[809],{"type":49,"value":810},"run",{"type":49,"value":152},{"type":43,"tag":52,"props":813,"children":815},{"className":814},[],[816],{"type":49,"value":817},"test",{"type":49,"value":152},{"type":43,"tag":52,"props":820,"children":822},{"className":821},[],[823],{"type":49,"value":824},"debug",{"type":49,"value":826},", etc.) plus a ",{"type":43,"tag":52,"props":828,"children":830},{"className":829},[],[831],{"type":49,"value":57},{"type":49,"value":833},"-\nspecific command appendix and the agent-side ",{"type":43,"tag":52,"props":835,"children":837},{"className":836},[],[838],{"type":49,"value":839},"use",{"type":49,"value":841},"\nworkflow that consumes the captured per-iteration output.",{"type":43,"tag":61,"props":843,"children":844},{},[845,847,852,854,859],{"type":49,"value":846},"The combined skill teaches an AI agent to drive the\n",{"type":43,"tag":139,"props":848,"children":849},{},[850],{"type":49,"value":851},"measurement-class",{"type":49,"value":853}," of ",{"type":43,"tag":52,"props":855,"children":857},{"className":856},[],[858],{"type":49,"value":57},{"type":49,"value":860}," questions: pick a\nshipped or author-new policy JSON, run the single-iteration\nsmoke, run the iterative eval loop, capture the four-tuple\nthat makes the resulting number defensible, interpret the\noutput, and route every adjacent question (tune the live\napp, measure the DPA path, optimize the firmware) to the\nright neighbouring skill.",{"type":43,"tag":232,"props":862,"children":864},{"id":863},"what-this-skill-deliberately-does-not-ship",[865],{"type":49,"value":866},"What this skill deliberately does not ship",{"type":43,"tag":239,"props":868,"children":869},{},[870,904,924,943,953,970,997],{"type":43,"tag":243,"props":871,"children":872},{},[873,878,880,885,887,892,894,902],{"type":43,"tag":65,"props":874,"children":875},{},[876],{"type":49,"value":877},"End-to-end dataplane throughput or latency\nmeasurement.",{"type":49,"value":879}," ",{"type":43,"tag":52,"props":881,"children":883},{"className":882},[],[884],{"type":49,"value":57},{"type":49,"value":886}," measures the\n",{"type":43,"tag":139,"props":888,"children":889},{},[890],{"type":49,"value":891},"control-plane",{"type":49,"value":893}," rate of programming rules, plus optional\nper-entry query timing. It does NOT measure how fast\npackets traverse the resulting rules in the dataplane.\nThat is the application's responsibility, layered on\n",{"type":43,"tag":80,"props":895,"children":896},{"href":454},[897],{"type":43,"tag":52,"props":898,"children":900},{"className":899},[],[901],{"type":49,"value":256},{"type":49,"value":903},". The agent\nmust say this explicitly when the operator asks for \"Flow\nthroughput\".",{"type":43,"tag":243,"props":905,"children":906},{},[907,912,914,922],{"type":43,"tag":65,"props":908,"children":909},{},[910],{"type":49,"value":911},"DPA-offloaded Flow path measurement.",{"type":49,"value":913}," Route to\n",{"type":43,"tag":80,"props":915,"children":916},{"href":203},[917],{"type":43,"tag":52,"props":918,"children":920},{"className":919},[],[921],{"type":49,"value":184},{"type":49,"value":923},".",{"type":43,"tag":243,"props":925,"children":926},{},[927,932,933,941],{"type":43,"tag":65,"props":928,"children":929},{},[930],{"type":49,"value":931},"Optimization of a deployed Flow application.",{"type":49,"value":913},{"type":43,"tag":80,"props":934,"children":935},{"href":214},[936],{"type":43,"tag":52,"props":937,"children":939},{"className":938},[],[940],{"type":49,"value":176},{"type":49,"value":942},". flow-perf\nis a synthetic driver of a JSON-described pipeline, not a\ntuner of a live one.",{"type":43,"tag":243,"props":944,"children":945},{},[946,951],{"type":43,"tag":65,"props":947,"children":948},{},[949],{"type":49,"value":950},"A canonical \"right answer\" Kops\u002Fsec number.",{"type":49,"value":952}," The agent\nrefuses to quote published numbers from memory as\nauthoritative; the published numbers live in NVIDIA's\nrelease notes per the DOCA version and BlueField\ngeneration, and the operator must reproduce on their own\nexact preconditions before comparing.",{"type":43,"tag":243,"props":954,"children":955},{},[956,961,963,968],{"type":43,"tag":65,"props":957,"children":958},{},[959],{"type":49,"value":960},"Invented JSON schema fields.",{"type":49,"value":962}," The agent does NOT invent\npolicy JSON keys that are not present in the shipped\n",{"type":43,"tag":52,"props":964,"children":966},{"className":965},[],[967],{"type":49,"value":266},{"type":49,"value":969}," exemplars. If a key the operator wants is not\nin any shipped exemplar, the agent says so and routes to\nthe public DOCA Flow Perf guide.",{"type":43,"tag":243,"props":971,"children":972},{},[973,985,987,995],{"type":43,"tag":65,"props":974,"children":975},{},[976,978,983],{"type":49,"value":977},"Library-internal ",{"type":43,"tag":52,"props":979,"children":981},{"className":980},[],[982],{"type":49,"value":256},{"type":49,"value":984}," API explanations.",{"type":49,"value":986}," The\nunderlying matchers and actions belong to\n",{"type":43,"tag":80,"props":988,"children":989},{"href":454},[990],{"type":43,"tag":52,"props":991,"children":993},{"className":992},[],[994],{"type":49,"value":256},{"type":49,"value":996},"; this skill\nreferences them but does not duplicate the library's API\ndocumentation.",{"type":43,"tag":243,"props":998,"children":999},{},[1000,1005],{"type":43,"tag":65,"props":1001,"children":1002},{},[1003],{"type":49,"value":1004},"Cross-tool benchmarking apples-to-apples claims",{"type":49,"value":1006}," when\npreconditions differ. Two flow-perf numbers from different\nDOCA versions \u002F BlueField generations \u002F firmware versions\nare NOT directly comparable; the agent insists on the\nfour-tuple capture so consumers can judge.",{"type":43,"tag":232,"props":1008,"children":1010},{"id":1009},"loading-order",[1011],{"type":49,"value":1012},"Loading order",{"type":43,"tag":61,"props":1014,"children":1015},{},[1016,1018,1023],{"type":49,"value":1017},"When a ",{"type":43,"tag":52,"props":1019,"children":1021},{"className":1020},[],[1022],{"type":49,"value":57},{"type":49,"value":1024}," question arrives:",{"type":43,"tag":486,"props":1026,"children":1027},{},[1028,1049,1072,1100,1114,1129],{"type":43,"tag":243,"props":1029,"children":1030},{},[1031,1033,1038,1040,1048],{"type":49,"value":1032},"Confirm DOCA is installed and the binary plus the\n",{"type":43,"tag":52,"props":1034,"children":1036},{"className":1035},[],[1037],{"type":49,"value":266},{"type":49,"value":1039}," JSON library are reachable — if not, route to\n",{"type":43,"tag":80,"props":1041,"children":1042},{"href":191},[1043],{"type":43,"tag":52,"props":1044,"children":1046},{"className":1045},[],[1047],{"type":49,"value":198},{"type":49,"value":432},{"type":43,"tag":243,"props":1050,"children":1051},{},[1052,1054,1059,1061,1071],{"type":49,"value":1053},"Confirm the underlying ",{"type":43,"tag":52,"props":1055,"children":1057},{"className":1056},[],[1058],{"type":49,"value":256},{"type":49,"value":1060}," library is healthy on\nthe device — if not, route to\n",{"type":43,"tag":80,"props":1062,"children":1064},{"href":1063},"..\u002F..\u002Flibs\u002Fdoca-flow\u002FTASKS.md#test",[1065],{"type":43,"tag":52,"props":1066,"children":1068},{"className":1067},[],[1069],{"type":49,"value":1070},"doca-flow TASKS.md ## test",{"type":49,"value":432},{"type":43,"tag":243,"props":1073,"children":1074},{},[1075,1077,1082,1084,1089,1091,1099],{"type":49,"value":1076},"Confirm the user wants to ",{"type":43,"tag":139,"props":1078,"children":1079},{},[1080],{"type":49,"value":1081},"measure",{"type":49,"value":1083},", not ",{"type":43,"tag":139,"props":1085,"children":1086},{},[1087],{"type":49,"value":1088},"optimize",{"type":49,"value":1090}," —\nif optimize, route to\n",{"type":43,"tag":80,"props":1092,"children":1093},{"href":214},[1094],{"type":43,"tag":52,"props":1095,"children":1097},{"className":1096},[],[1098],{"type":49,"value":176},{"type":49,"value":432},{"type":43,"tag":243,"props":1101,"children":1102},{},[1103,1105,1113],{"type":49,"value":1104},"Confirm the target path is host \u002F DPU-CPU, not DPA — if\nDPA, route to\n",{"type":43,"tag":80,"props":1106,"children":1107},{"href":203},[1108],{"type":43,"tag":52,"props":1109,"children":1111},{"className":1110},[],[1112],{"type":49,"value":184},{"type":49,"value":432},{"type":43,"tag":243,"props":1115,"children":1116},{},[1117,1119,1127],{"type":49,"value":1118},"Read ",{"type":43,"tag":80,"props":1120,"children":1121},{"href":129},[1122],{"type":43,"tag":52,"props":1123,"children":1125},{"className":1124},[],[1126],{"type":49,"value":129},{"type":49,"value":1128}," to commit to\nthe three-axis decision (pipeline shape × traffic class ×\nmeasurement axis);",{"type":43,"tag":243,"props":1130,"children":1131},{},[1132,1133,1141,1143,1149,1151,1156,1158,1163],{"type":49,"value":1118},{"type":43,"tag":80,"props":1134,"children":1135},{"href":82},[1136],{"type":43,"tag":52,"props":1137,"children":1139},{"className":1138},[],[1140],{"type":49,"value":82},{"type":49,"value":1142}," and walk\n",{"type":43,"tag":52,"props":1144,"children":1146},{"className":1145},[],[1147],{"type":49,"value":1148},"## configure → ## run → ## test → ## debug",{"type":49,"value":1150}," in that\norder; do NOT start with ",{"type":43,"tag":52,"props":1152,"children":1154},{"className":1153},[],[1155],{"type":49,"value":112},{"type":49,"value":1157}," without the\n",{"type":43,"tag":52,"props":1159,"children":1161},{"className":1160},[],[1162],{"type":49,"value":100},{"type":49,"value":1164}," precondition step.",{"type":43,"tag":232,"props":1166,"children":1168},{"id":1167},"related-skills",[1169],{"type":49,"value":1170},"Related skills",{"type":43,"tag":61,"props":1172,"children":1173},{},[1174,1176,1182],{"type":49,"value":1175},"Cross-link conventions follow the bundle's relative path\ncontract from ",{"type":43,"tag":52,"props":1177,"children":1179},{"className":1178},[],[1180],{"type":49,"value":1181},"tools\u002F\u003CX>\u002F",{"type":49,"value":1183},":",{"type":43,"tag":239,"props":1185,"children":1186},{},[1187,1200,1220,1240,1260,1273,1286,1301,1323,1360],{"type":43,"tag":243,"props":1188,"children":1189},{},[1190,1198],{"type":43,"tag":80,"props":1191,"children":1192},{"href":454},[1193],{"type":43,"tag":52,"props":1194,"children":1196},{"className":1195},[],[1197],{"type":49,"value":256},{"type":49,"value":1199}," — the\nunderlying library. flow-perf programs Flow pipes, entries,\nmatchers, and actions; the library is the source of truth\nfor the API surface flow-perf exercises.",{"type":43,"tag":243,"props":1201,"children":1202},{},[1203,1211,1213,1218],{"type":43,"tag":80,"props":1204,"children":1205},{"href":214},[1206],{"type":43,"tag":52,"props":1207,"children":1209},{"className":1208},[],[1210],{"type":49,"value":176},{"type":49,"value":1212}," — the\nunified Flow tuning tool. ",{"type":43,"tag":65,"props":1214,"children":1215},{},[1216],{"type":49,"value":1217},"Measurement vs. optimization\nboundary",{"type":49,"value":1219}," lives here. Ask: \"do I want a number, or do I\nwant to change the deployed pipeline?\"",{"type":43,"tag":243,"props":1221,"children":1222},{},[1223,1231,1233,1238],{"type":43,"tag":80,"props":1224,"children":1225},{"href":203},[1226],{"type":43,"tag":52,"props":1227,"children":1229},{"className":1228},[],[1230],{"type":49,"value":184},{"type":49,"value":1232}," —\nthe DPA-offloaded Flow performance tool. ",{"type":43,"tag":65,"props":1234,"children":1235},{},[1236],{"type":49,"value":1237},"Host \u002F\nDPU-CPU vs. DPA path boundary",{"type":49,"value":1239}," lives here. Ask: \"am I\nmeasuring the path that executes on the CPU, or the path\nthat executes on the DPA processor?\"",{"type":43,"tag":243,"props":1241,"children":1242},{},[1243,1251,1253,1258],{"type":43,"tag":80,"props":1244,"children":1245},{"href":659},[1246],{"type":43,"tag":52,"props":1247,"children":1249},{"className":1248},[],[1250],{"type":49,"value":666},{"type":49,"value":1252}," — every\nreported flow-perf number must come with the four-way\nmatch (host package, kernel module, firmware, target\napplication's linked ",{"type":43,"tag":52,"props":1254,"children":1256},{"className":1255},[],[1257],{"type":49,"value":256},{"type":49,"value":1259}," version) and the BlueField\n\u002F ConnectX generation. flow-perf overlays this rule, not\ncontradicts it.",{"type":43,"tag":243,"props":1261,"children":1262},{},[1263,1271],{"type":43,"tag":80,"props":1264,"children":1265},{"href":191},[1266],{"type":43,"tag":52,"props":1267,"children":1269},{"className":1268},[],[1270],{"type":49,"value":198},{"type":49,"value":1272}," — DOCA install\nposture; routing for \"is the binary even here?\" questions.",{"type":43,"tag":243,"props":1274,"children":1275},{},[1276,1284],{"type":43,"tag":80,"props":1277,"children":1278},{"href":674},[1279],{"type":43,"tag":52,"props":1280,"children":1282},{"className":1281},[],[1283],{"type":49,"value":681},{"type":49,"value":1285}," — the\ncross-cutting debug ladder for env-side issues (driver,\nfirmware, BlueField mode, kernel module).",{"type":43,"tag":243,"props":1287,"children":1288},{},[1289,1299],{"type":43,"tag":80,"props":1290,"children":1292},{"href":1291},"..\u002Fdoca-bench\u002FSKILL.md",[1293],{"type":43,"tag":52,"props":1294,"children":1296},{"className":1295},[],[1297],{"type":49,"value":1298},"doca-bench",{"type":49,"value":1300}," — a peer\nbenchmarking tool with a broader scope (multiple DOCA\nprimitives, not just Flow). flow-perf is the Flow-specific\nmicrobenchmark; doca-bench is the broader workload\nbenchmark.",{"type":43,"tag":243,"props":1302,"children":1303},{},[1304,1314,1316,1321],{"type":43,"tag":80,"props":1305,"children":1307},{"href":1306},"..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md",[1308],{"type":43,"tag":52,"props":1309,"children":1311},{"className":1310},[],[1312],{"type":49,"value":1313},"doca-public-knowledge-map",{"type":49,"value":1315},"\n— routing to the public ",{"type":43,"tag":52,"props":1317,"children":1319},{"className":1318},[],[1320],{"type":49,"value":507},{"type":49,"value":1322}," DOCA Flow Perf\npage, release notes, and forums for release-specific\npublished numbers and reproducibility notes.",{"type":43,"tag":243,"props":1324,"children":1325},{},[1326,1336,1338,1344,1345,1351,1352,1358],{"type":43,"tag":80,"props":1327,"children":1329},{"href":1328},"..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md",[1330],{"type":43,"tag":52,"props":1331,"children":1333},{"className":1332},[],[1334],{"type":49,"value":1335},"doca-structured-tools-contract",{"type":49,"value":1337},"\n— the agent's detect → prefer → fall back → report contract\nfor the structured helpers (",{"type":43,"tag":52,"props":1339,"children":1341},{"className":1340},[],[1342],{"type":49,"value":1343},"doca-env --json",{"type":49,"value":283},{"type":43,"tag":52,"props":1346,"children":1348},{"className":1347},[],[1349],{"type":49,"value":1350},"doca-capability-snapshot",{"type":49,"value":152},{"type":43,"tag":52,"props":1353,"children":1355},{"className":1354},[],[1356],{"type":49,"value":1357},"version-matrix.json",{"type":49,"value":1359},")\nflow-perf preconditions rely on.",{"type":43,"tag":243,"props":1361,"children":1362},{},[1363,1373,1375,1385],{"type":43,"tag":80,"props":1364,"children":1366},{"href":1365},"..\u002F..\u002Fdoca-hardware-safety\u002FSKILL.md",[1367],{"type":43,"tag":52,"props":1368,"children":1370},{"className":1369},[],[1371],{"type":49,"value":1372},"doca-hardware-safety",{"type":49,"value":1374},"\n— the canonical hardware-safety meta-policy that\n",{"type":43,"tag":80,"props":1376,"children":1378},{"href":1377},"CAPABILITIES.md#safety-policy",[1379],{"type":43,"tag":52,"props":1380,"children":1382},{"className":1381},[],[1383],{"type":49,"value":1384},"CAPABILITIES.md ## Safety policy",{"type":49,"value":1386},"\noverlays.",{"type":43,"tag":61,"props":1388,"children":1389},{},[1390],{"type":49,"value":1391},"This skill assumes the surrounding doca-flow application is\nthe operator's existing source artifact; flow-perf does not\nship a sample doca-flow application of its own.",{"items":1393,"total":1495},[1394,1409,1423,1437,1449,1466,1481],{"slug":1395,"name":1395,"fn":1396,"description":1397,"org":1398,"tags":1399,"stars":23,"repoUrl":24,"updatedAt":1408},"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},[1400,1403,1406,1407],{"name":1401,"slug":1402,"type":15},"Data Analysis","data-analysis",{"name":1404,"slug":1405,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:28:43.176466",{"slug":1410,"name":1410,"fn":1411,"description":1412,"org":1413,"tags":1414,"stars":23,"repoUrl":24,"updatedAt":1422},"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},[1415,1418,1421],{"name":1416,"slug":1417,"type":15},"Deployment","deployment",{"name":1419,"slug":1420,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1424,"name":1424,"fn":1425,"description":1426,"org":1427,"tags":1428,"stars":23,"repoUrl":24,"updatedAt":1436},"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},[1429,1432,1433],{"name":1430,"slug":1431,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1434,"slug":1435,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1438,"name":1438,"fn":1439,"description":1440,"org":1441,"tags":1442,"stars":23,"repoUrl":24,"updatedAt":1448},"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},[1443,1444,1445],{"name":1401,"slug":1402,"type":15},{"name":9,"slug":8,"type":15},{"name":1446,"slug":1447,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":1450,"name":1450,"fn":1451,"description":1452,"org":1453,"tags":1454,"stars":23,"repoUrl":24,"updatedAt":1465},"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},[1455,1458,1461,1462],{"name":1456,"slug":1457,"type":15},"Automation","automation",{"name":1459,"slug":1460,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":1463,"slug":1464,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":1467,"name":1467,"fn":1468,"description":1469,"org":1470,"tags":1471,"stars":23,"repoUrl":24,"updatedAt":1480},"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},[1472,1473,1476,1477],{"name":1416,"slug":1417,"type":15},{"name":1474,"slug":1475,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1478,"slug":1479,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1482,"name":1482,"fn":1483,"description":1484,"org":1485,"tags":1486,"stars":23,"repoUrl":24,"updatedAt":1494},"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},[1487,1488,1491],{"name":9,"slug":8,"type":15},{"name":1489,"slug":1490,"type":15},"Quantum Computing","quantum-computing",{"name":1492,"slug":1493,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":1497,"total":1648},[1498,1516,1532,1543,1555,1569,1582,1596,1607,1616,1630,1639],{"slug":1499,"name":1499,"fn":1500,"description":1501,"org":1502,"tags":1503,"stars":1513,"repoUrl":1514,"updatedAt":1515},"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},[1504,1507,1510],{"name":1505,"slug":1506,"type":15},"Documentation","documentation",{"name":1508,"slug":1509,"type":15},"MCP","mcp",{"name":1511,"slug":1512,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1517,"name":1517,"fn":1518,"description":1519,"org":1520,"tags":1521,"stars":1529,"repoUrl":1530,"updatedAt":1531},"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},[1522,1525,1526],{"name":1523,"slug":1524,"type":15},"Containers","containers",{"name":1416,"slug":1417,"type":15},{"name":1527,"slug":1528,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1533,"name":1533,"fn":1534,"description":1535,"org":1536,"tags":1537,"stars":1529,"repoUrl":1530,"updatedAt":1542},"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},[1538,1541],{"name":1539,"slug":1540,"type":15},"CI\u002FCD","ci-cd",{"name":1416,"slug":1417,"type":15},"2026-07-14T05:25:59.97109",{"slug":1544,"name":1544,"fn":1545,"description":1546,"org":1547,"tags":1548,"stars":1529,"repoUrl":1530,"updatedAt":1554},"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},[1549,1550,1551],{"name":1539,"slug":1540,"type":15},{"name":1416,"slug":1417,"type":15},{"name":1552,"slug":1553,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1556,"name":1556,"fn":1557,"description":1558,"org":1559,"tags":1560,"stars":1529,"repoUrl":1530,"updatedAt":1568},"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},[1561,1564,1565],{"name":1562,"slug":1563,"type":15},"Debugging","debugging",{"name":1552,"slug":1553,"type":15},{"name":1566,"slug":1567,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1570,"name":1570,"fn":1571,"description":1572,"org":1573,"tags":1574,"stars":1529,"repoUrl":1530,"updatedAt":1581},"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},[1575,1578],{"name":1576,"slug":1577,"type":15},"Best Practices","best-practices",{"name":1579,"slug":1580,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1583,"name":1583,"fn":1584,"description":1585,"org":1586,"tags":1587,"stars":1529,"repoUrl":1530,"updatedAt":1595},"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},[1588,1591,1594],{"name":1589,"slug":1590,"type":15},"Machine Learning","machine-learning",{"name":1592,"slug":1593,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1597,"name":1597,"fn":1598,"description":1599,"org":1600,"tags":1601,"stars":1529,"repoUrl":1530,"updatedAt":1606},"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},[1602,1605],{"name":1603,"slug":1604,"type":15},"QA","qa",{"name":1446,"slug":1447,"type":15},"2026-07-14T05:25:53.673039",{"slug":1608,"name":1608,"fn":1609,"description":1610,"org":1611,"tags":1612,"stars":1529,"repoUrl":1530,"updatedAt":1615},"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},[1613,1614],{"name":1416,"slug":1417,"type":15},{"name":1419,"slug":1420,"type":15},"2026-07-14T05:25:49.362534",{"slug":1617,"name":1617,"fn":1618,"description":1619,"org":1620,"tags":1621,"stars":1529,"repoUrl":1530,"updatedAt":1629},"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},[1622,1625,1626],{"name":1623,"slug":1624,"type":15},"Code Review","code-review",{"name":1552,"slug":1553,"type":15},{"name":1627,"slug":1628,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1631,"name":1631,"fn":1632,"description":1633,"org":1634,"tags":1635,"stars":1529,"repoUrl":1530,"updatedAt":1638},"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},[1636,1637],{"name":1603,"slug":1604,"type":15},{"name":1446,"slug":1447,"type":15},"2026-07-14T05:25:54.928983",{"slug":1640,"name":1640,"fn":1641,"description":1642,"org":1643,"tags":1644,"stars":1529,"repoUrl":1530,"updatedAt":1647},"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},[1645,1646],{"name":1456,"slug":1457,"type":15},{"name":1539,"slug":1540,"type":15},"2026-07-30T05:29:03.275638",496]