[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-doca-pcc-ztr-rttcc-algo":3,"mdc-qkc8uc-key":34,"related-org-nvidia-doca-pcc-ztr-rttcc-algo":1519,"related-repo-nvidia-doca-pcc-ztr-rttcc-algo":1680},{"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-pcc-ztr-rttcc-algo","tune ZTR RTTCC congestion control algorithms","Use this skill when the user is doing hands-on deployment, tuning, or evaluation of the DOCA-shipped Zero-Touch RoCE RTT-based Congestion Control (ZTR RTTCC) reference algorithm on a BlueField-3 DPA — wiring `doca_pcc_dev_ztr_rttcc_algo` into the shipped DOCA PCC sample, picking a variant (vanilla \u002F PM \u002F RX-rate \u002F multipath \u002F window-probeless) at DPACC build time, tuning host-set parameters, or diagnosing `DOCA_PCC_DEV_STATUS_FAIL` from the algorithm. Trigger even when the user does not say 'DOCA PCC' or 'ZTR RTTCC' — typical implicit phrasings: 'my RoCE-v2 flows aren't being throttled', 'PCC sample isn't dispatching to my algo', 'how do I pick the multipath PCC variant', 'set-params returns fail', 'algorithm loaded but counters are flat', or 'do I need a custom CC algorithm on BF3'. Refuse and route elsewhere for writing a custom PCC algorithm from scratch, read-only PCC counter inspection, the host-side `doca-pcc` lifecycle, or firmware-only pre-Programmable PCC — 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},"Hardware","hardware",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"Engineering","engineering",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-20T06:24:29.379683","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-pcc-ztr-rttcc-algo","---\nlicense: Apache-2.0\nname: doca-pcc-ztr-rttcc-algo\ndescription: >\n  Use this skill when the user is doing hands-on deployment, tuning,\n  or evaluation of the DOCA-shipped Zero-Touch RoCE RTT-based\n  Congestion Control (ZTR RTTCC) reference algorithm on a BlueField-3\n  DPA — wiring `doca_pcc_dev_ztr_rttcc_algo` into the shipped DOCA\n  PCC sample, picking a variant (vanilla \u002F PM \u002F RX-rate \u002F multipath \u002F\n  window-probeless) at DPACC build time, tuning host-set parameters,\n  or diagnosing `DOCA_PCC_DEV_STATUS_FAIL` from the algorithm.\n  Trigger even when the user does not say 'DOCA PCC' or 'ZTR RTTCC' —\n  typical implicit phrasings: 'my RoCE-v2 flows aren't being\n  throttled', 'PCC sample isn't dispatching to my algo', 'how do I\n  pick the multipath PCC variant', 'set-params returns fail',\n  'algorithm loaded but counters are flat', or 'do I need a custom\n  CC algorithm on BF3'. Refuse and route elsewhere for writing a\n  custom PCC algorithm from scratch, read-only PCC counter\n  inspection, the host-side `doca-pcc` lifecycle, or firmware-only\n  pre-Programmable PCC — those belong to other skills.\nmetadata:\n  kind: library\ncompatibility: >\n  Requires DOCA SDK installed at \u002Fopt\u002Fmellanox\u002Fdoca on Linux (Ubuntu\n  22.04\u002F24.04 or RHEL\u002FSLES) with a BlueField-3 DPU exposing the DPA\n  processor, the firmware custom-PCC slot enabled, a matched-version\n  DPACC compiler, and live RoCE-v2 traffic on the attached port.\n  Reads `pkg-config doca-pcc-ztr-rttcc-algo` and inspects\n  \u002Fopt\u002Fmellanox\u002Fdoca\u002F{lib,include,applications\u002Fpcc}.\n---\n\n# DOCA PCC ZTR RTTCC Algorithm\n\n**Where to start:** This skill assumes DOCA is already\ninstalled, the user's BlueField has a DPA processor that\nthe host can see through DOCA (a BlueField-3-generation\ndevice per the README), the BlueField firmware has the\ncustom-PCC slot enabled, the DPACC compiler is installed at\na matched version per the DOCA Compatibility Policy, and the\nuser is doing **hands-on deployment of the DOCA-shipped ZTR\nRTTCC reference algorithm** on a BlueField port that\nalready carries RoCE-v2 traffic — i.e. either deploying it\nas the no-config-required baseline, tuning its documented\nparameters, or evaluating it against a custom algorithm the\nuser intends to write. Open [`TASKS.md`](TASKS.md) if the\nuser wants to *do* something (install \u002F configure \u002F build \u002F\nmodify \u002F run \u002F test \u002F debug \u002F use); open\n[`CAPABILITIES.md`](CAPABILITIES.md) when the question is\n*what does the algorithm express, what are its variants and\nparameters, what does it ship vs not ship*. If the user has\nnot installed DOCA yet, route to\n[`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) first; if the user\nhas not stood up the host-side `doca-pcc` framework yet,\nroute to [`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md) first (this\nalgorithm is a *library consumed by* the PCC framework, not\na standalone program); if the user only wants to *inspect*\nPCC counters at runtime without changing the running\nalgorithm, route to\n[`doca-pcc-counters`](..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md);\nif the user wants to *write their own algorithm from\nscratch*, that is the `doca-pcc` library plus the public\nPCC programming guide — this skill is for the shipped\nreference algorithm specifically.\n\n## Example questions this skill answers well\n\nThe CLASSES of ZTR RTTCC questions this skill is built to\nanswer, each with one worked example. The agent should\ntreat the *class* as the load-bearing piece — the worked\nexample is a single instance.\n\n- **\"Is the ZTR RTTCC reference algorithm the right\n  baseline for my deployment, or should I write a custom\n  algorithm?\"** — worked example: *\"I have a BlueField-3\n  carrying production RoCE-v2 traffic from a GPU cluster;\n  is the shipped algorithm a fine default or do I need\n  custom logic?\"*. Answered by the decision rule in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  (\"when to use the reference vs custom\") + the env\n  preconditions in\n  [`TASKS.md ## install`](TASKS.md#install).\n- **\"How do I wire the shipped algorithm into the DOCA\n  PCC application that's already running on my host?\"** —\n  worked example: *\"`\u002Fopt\u002Fmellanox\u002Fdoca\u002Fapplications\u002Fpcc`\n  is already building from sample sources; what do I\n  change so the user algo callback dispatches to\n  `doca_pcc_dev_ztr_rttcc_algo` under a chosen algo\n  slot?\"*. Answered by the integration sequence in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the in-place edits in\n  [`TASKS.md ## modify`](TASKS.md#modify).\n- **\"Which variant of the algorithm am I getting — vanilla\n  RTT-CC, path-migration mode, RX-rate mode, multipath,\n  multipath with credits, window-probeless?\"** — worked\n  example: *\"the shipped library exposes one public\n  symbol `doca_pcc_dev_ztr_rttcc_algo` but the device-\n  side source ships several variants; how do I know\n  which one I get and how do I pick another?\"*.\n  Answered by the variants table in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes).\n- **\"How do I confirm the algorithm is actually\n  modulating my RDMA \u002F RoCE traffic, and not just\n  loading?\"** — worked example: *\"I followed the\n  integration steps; the application starts; how do I\n  know the algorithm is shaping flows under load?\"*.\n  Answered by the observability surface in\n  [`CAPABILITIES.md ## Observability`](CAPABILITIES.md#observability)\n  + the counter-watch loop in\n  [`TASKS.md ## test`](TASKS.md#test) which routes to\n  [`doca-pcc-counters`](..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md).\n- **\"Which tunables does the algorithm expose, and how\n  do I change them from the host without rebuilding the\n  DPA-side image?\"** — worked example: *\"my workload is\n  more latency-sensitive than the default profile assumes\n  — which parameter knob do I adjust?\"*. Answered by\n  the parameter surface in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the `doca_pcc_dev_set_ztr_rttcc_params` workflow in\n  [`TASKS.md ## use`](TASKS.md#use).\n- **\"What does this `DOCA_PCC_DEV_STATUS_FAIL` or\n  `DOCA_ERROR_*` from a `doca_pcc_dev_ztr_rttcc_*` call\n  mean and which layer caused it?\"** — worked example:\n  *\"my init callback returns `DOCA_PCC_DEV_STATUS_FAIL`\n  on first launch\"*. Answered by the algorithm overlay\n  on the host-side PCC taxonomy in\n  [`CAPABILITIES.md ## Error taxonomy`](CAPABILITIES.md#error-taxonomy)\n  + the layered ladder in\n  [`TASKS.md ## debug`](TASKS.md#debug) that escalates\n  through\n  [`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md) and\n  [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md).\n\n## Audience\n\nThis skill serves **external developers operating a\nBlueField-3-class DPU who want to deploy NVIDIA's shipped\nreference PCC algorithm on RoCE-v2 traffic, OR who are\nevaluating it against a custom algorithm they intend to\nwrite**. The reference algorithm is *zero-touch* by design\n— the no-config-required baseline — and the canonical use\ncase is dropping it onto a port and confirming it shapes\nflows correctly under congestion. It is *not* for NVIDIA\ndevelopers contributing to the algorithm itself, nor for\nusers who want general PCC programming theory (route via\nthe public DOCA PCC programming guide), nor for users who\nonly want to *inspect* PCC counters (route to\n[`doca-pcc-counters`](..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md)).\n\n**Language scope.** The algorithm ships as a DPA-side\nlibrary (`pkg-config` module `doca-pcc-ztr-rttcc-algo`)\nplus a public header `doca_pcc_dev_ztr_rttcc_algo.h` that\nDPA-side translation units include. The shipped algorithm\nbinary is the static library\n`libdoca_pcc_ztr_rttcc_algo_dev.a` per the README; the\ndevice-side translation unit that consumes it is C and is\ncompiled by DPACC. The host-side that drives the PCC\ncontext comes from [`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md);\nthis library does NOT add a host-side surface beyond the\nhost-side helpers (also shipped as\n`libdoca_pcc_ztr_rttcc_algo.{a,so}` per the README) that\nthe `doca-pcc` framework links. Other-language host-side\nwrappers around `doca-pcc` can drive this algorithm through\nthe same lifecycle described in `doca-pcc`; the DPA-side\nintegration always stays C-via-DPACC.\n\n## When to load this skill\n\nLoad this skill when the user is doing hands-on deployment,\ntuning, or evaluation of the DOCA-shipped ZTR RTTCC\nreference algorithm on a BlueField port carrying RoCE-v2\ntraffic, in any host language plus the DPA-side translation\nunit built by `dpacc`. Concretely:\n\n- Deciding whether the shipped algorithm is the right\n  baseline for the user's RoCE-v2 workload, or whether the\n  user needs a custom algorithm.\n- Wiring the algorithm into the DOCA PCC sample\n  application by patching the user-algo \u002F user-init \u002F\n  user-set-algo-params callbacks per the steps the\n  shipped README documents.\n- Picking which of the algorithm's documented variants\n  (vanilla, path-migration, RX-rate, RX-rate + PM,\n  multipath, multipath + credits, window-probeless)\n  matches the user's intent — and surfacing that the\n  *public* surface only ships one symbol\n  (`doca_pcc_dev_ztr_rttcc_algo`); the variants live in\n  the DPA-side source the user compiles against.\n- Tuning the host-set parameters the algorithm exposes\n  (per the parameter list in\n  `doca_pcc_dev_ztr_rttcc_algo.h` and the shipped\n  `doca_pcc_dev_set_ztr_rttcc_params`).\n- Observing whether the running algorithm is actually\n  modulating RoCE-v2 flows under load (route to\n  [`doca-pcc-counters`](..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md)\n  for the read-only inspection side).\n- Deciding when to *replace* the shipped algorithm with a\n  custom one because latency target, fairness policy, or\n  convergence behavior requirements diverge from what the\n  reference provides.\n\nDo **not** load this skill for general DOCA orientation;\nfor the host-side `doca-pcc` lifecycle (route to\n[`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md)); for writing a custom\nalgorithm from scratch (route to\n[`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md) and the public PCC\nprogramming guide via\n[`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md));\nfor read-only PCC counter inspection (route to\n[`doca-pcc-counters`](..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md));\nor for the default firmware-shipped PCC algorithms that\npredate Programmable Congestion Control entirely (no\nhost-side code, no DPACC compile — that is a firmware-only\npath routed via\n[`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md)).\n\n## What this skill provides\n\nThis is a **thin loader**. The body keeps only the\norientation needed to pick the right next file. The\nsubstantive algorithm-specific material lives in two\ncompanion files:\n\n- `CAPABILITIES.md` — what the shipped ZTR RTTCC\n  algorithm expresses on this version + this BlueField\n  generation + this firmware: the public DPA-side API\n  surface (`doca_pcc_dev_ztr_rttcc_init`,\n  `doca_pcc_dev_ztr_rttcc_algo`,\n  `doca_pcc_dev_set_ztr_rttcc_params`,\n  `doca_pcc_dev_ztr_rttcc_get_param_num`,\n  `doca_pcc_dev_ztr_rttcc_get_counter_num`,\n  `doca_pcc_dev_ztr_rttcc_get_num_of_histograms`), the\n  documented variants (vanilla \u002F path-migration \u002F RX-rate\n  \u002F multipath \u002F window-probeless — pick one at\n  DPA-side compile time), the relationship to the\n  host-side `doca-pcc` framework (this is an algorithm\n  body the framework loads), the relationship to the\n  `doca-pcc-counters` tool (which is the canonical\n  inspection surface), the algorithm's parameter and\n  counter surface (RTT-based congestion signal,\n  per-feature parameter blocks), the error taxonomy in\n  `DOCA_PCC_DEV_STATUS_OK` \u002F `_FAIL`, and the safety\n  policy.\n- `TASKS.md` — step-by-step workflows for the in-scope\n  algorithm verbs: `install`, `configure`, `build`,\n  `modify`, `run`, `test`, `debug`, `use`. Plus a\n  `Deferred task verbs` block that points out-of-scope\n  questions at the right next skill.\n\nThe skill assumes DOCA + the DPACC compiler + the\n[`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md) host-side framework are\nalready installed; the BlueField is a generation that\nexposes the DPA processor (the algorithm runs on the DPA);\nthe BlueField firmware has the custom-PCC slot enabled\n(inherited from\n[`doca-pcc CAPABILITIES.md ## Safety policy`](..\u002Fdoca-pcc\u002FCAPABILITIES.md#safety-policy));\nand the BlueField port the algorithm will modulate has\nRoCE-v2 traffic actually flowing on it (the algorithm\nmodulates *existing* RDMA \u002F RoCE traffic — without traffic,\nthere is nothing for it to do).\n\n## What this skill deliberately does not ship\n\nThis skill is **agent guidance**, not a samples or templates\nbundle. To keep the boundary clean, it deliberately does not\ncontain — and pull requests should not add:\n\n- **The algorithm body itself, in any language.** The\n  shipped algorithm is the static library\n  `libdoca_pcc_ztr_rttcc_algo_dev.a` (plus the host-side\n  helpers) installed by the matching DOCA host package\n  per the README. The agent's job is to route the user to\n  the installed library and header\n  (`doca_pcc_dev_ztr_rttcc_algo.h`) and to prescribe the\n  in-place edits documented in the README on the\n  shipped DOCA PCC application source, not to author the\n  algorithm.\n- **A standalone PCC application.** The DOCA PCC\n  application that hosts this algorithm is the shipped\n  C sample under\n  `\u002Fopt\u002Fmellanox\u002Fdoca\u002Fapplications\u002Fpcc\u002F` (per the\n  README). The agent's job is to prescribe the\n  minimum-diff modifications the README documents and to\n  walk the user through the rebuild — not to author a\n  parallel application.\n- **A description of every internal variant's\n  behavior.** The DPA-side source ships several\n  variants (vanilla, path-migration, RX-rate, multipath,\n  multipath + credits, window-probeless). The agent\n  names the variant set, says which one is reachable via\n  the public symbol on this install, and routes\n  algorithm-design questions to the public PCC\n  programming guide via\n  [`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md).\n  It does not redefine each variant's mathematical\n  behavior.\n- **A specific congestion-control algorithm tutorial.**\n  Congestion-control theory, RoCE-v2 fairness analysis,\n  workload-specific tuning — out of scope. Route to the\n  user's own domain expertise and to the public DOCA PCC\n  guide.\n\n## Loading order\n\n1. Read this `SKILL.md` first to confirm the user's\n   question is in scope (deployment \u002F tuning \u002F evaluation\n   of the shipped reference algorithm, not algorithm\n   design from scratch and not read-only counter\n   inspection).\n2. **For the algorithm capability matrix, the public\n   DPA-side API surface, the variant set, the parameter\n   and counter surface, the host-side `doca-pcc`\n   framework relationship, the\n   `doca-pcc-counters` inspection-side relationship, the\n   error taxonomy, the observability surface, and the\n   safety policy, see [CAPABILITIES.md](CAPABILITIES.md).**\n3. **For step-by-step workflows — install, configure,\n   build, modify, run, test, debug, use — see\n   [TASKS.md](TASKS.md).**\n\nBoth companion files cross-link to each other,\n[`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md) for the host-side PCC\nlifecycle that loads this algorithm,\n[`doca-pcc-counters`](..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md)\nfor the read-only counter-inspection side of validating that\nthe algorithm is modulating traffic,\n[`doca-dpa`](..\u002Fdoca-dpa\u002FSKILL.md) for the DPA-side\ntwo-side-program model and the DPACC compiler discipline,\n[`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) for the\ncanonical DOCA version-handling rules (with the DPACC\noverlay inherited from\n[`doca-dpa`](..\u002Fdoca-dpa\u002FSKILL.md) and\n[`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md)), and\n[`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md)\nwhenever the right answer is \"look it up in the public DOCA\nPCC programming guide or the on-disk install layout\".\n\n## Related skills\n\n- [`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md) — the host-side PCC\n  control library. This algorithm is loaded INTO a\n  `doca_pcc` context that `doca-pcc` stands up; the\n  host-side lifecycle (`doca_pcc` create \u002F configure \u002F\n  start \u002F stop \u002F destroy, the algorithm image\n  `doca_pcc_app`, the attach-to-port semantics) is owned\n  by `doca-pcc`. This skill prescribes only the DPA-side\n  algorithm integration on top.\n- [`doca-pcc-counters`](..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md) —\n  the read-only diagnostic CLI for PCC counters at the\n  port. The canonical *\"is the algorithm actually\n  modulating traffic\"* check goes through the counter\n  tool; this skill names what counters the algorithm\n  emits (CNP \u002F NACK \u002F AI \u002F HAI \u002F decrement \u002F RTT-band\n  counters per the public header) and routes the\n  inspection workflow to the tool skill.\n- [`doca-dpa`](..\u002Fdoca-dpa\u002FSKILL.md) — the host-side\n  DPA control library. The algorithm runs on the DPA,\n  compiled by DPACC; the two-side-program rule and the\n  DOCA-and-DPACC version-match overlay inherited from\n  here apply.\n- [`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md) —\n  the routing table for every public DOCA documentation\n  source (the DOCA PCC programming guide at\n  \u003Chttps:\u002F\u002Fdocs.nvidia.com\u002Fdoca\u002Fsdk\u002Fdoca-pcc\u002Findex.html>;\n  the DOCA PCC application guide; the DOCA Compatibility\n  Policy) and the on-disk layout of an installed DOCA\n  package.\n- [`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) — env\n  preparation, install verification, DPACC compiler\n  install \u002F verification, BlueField firmware\n  configuration (including the custom-PCC slot enable),\n  and the *I have no install yet* path with the public\n  NGC DOCA container. This skill assumes its\n  preconditions are satisfied AND that DPACC is installed\n  at a version that matches DOCA AND that the firmware-\n  level custom-PCC slot is enabled.\n- [`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) —\n  canonical DOCA version-handling rules. This skill's\n  `## Version compatibility` cross-links the four-way\n  match rule plus the DOCA-and-DPACC overlay inherited\n  from [`doca-pcc`](..\u002Fdoca-pcc\u002FSKILL.md).\n- [`doca-structured-tools-contract`](..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md) —\n  the bundle's structured-tools precedence rule (detect \u002F\n  prefer \u002F fall back \u002F report). The Command appendix in\n  [TASKS.md](TASKS.md) honors this contract.\n- [`doca-programming-guide`](..\u002F..\u002Fdoca-programming-guide\u002FSKILL.md) —\n  general DOCA programming patterns. This skill layers\n  algorithm-specific overlays on top of the universal\n  build, modify-a-shipped-sample, and Core lifecycle\n  patterns.\n- [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md) — cross-cutting\n  debug ladder. Algorithm-specific debug (the algorithm\n  loaded but counters do not move; the algorithm fails to\n  initialize; the algorithm modulates traffic too\n  aggressively \u002F too gently for the workload) overlays on\n  top of that ladder.\n- [`doca-hardware-safety`](..\u002F..\u002Fdoca-hardware-safety\u002FSKILL.md) —\n  cross-cutting hardware-safety meta-policy. Because the\n  algorithm modulates production RoCE-v2 flows on a\n  BlueField port, the meta-policy's pre-flight inventory,\n  replica-first, and rollback rules apply via this\n  skill's `## Safety policy` overlay.\n\n",{"data":35,"body":39},{"license":26,"name":4,"description":6,"metadata":36,"compatibility":38},{"kind":37},"library","Requires DOCA SDK installed at \u002Fopt\u002Fmellanox\u002Fdoca on Linux (Ubuntu 22.04\u002F24.04 or RHEL\u002FSLES) with a BlueField-3 DPU exposing the DPA processor, the firmware custom-PCC slot enabled, a matched-version DPACC compiler, and live RoCE-v2 traffic on the attached port. Reads `pkg-config doca-pcc-ztr-rttcc-algo` and inspects \u002Fopt\u002Fmellanox\u002Fdoca\u002F{lib,include,applications\u002Fpcc}.\n",{"type":40,"children":41},"root",[42,51,180,187,199,543,549,591,671,677,690,767,836,842,854,1015,1049,1055,1067,1142,1148,1205,1283,1289],{"type":43,"tag":44,"props":45,"children":47},"element","h1",{"id":46},"doca-pcc-ztr-rttcc-algorithm",[48],{"type":49,"value":50},"text","DOCA PCC ZTR RTTCC Algorithm",{"type":43,"tag":52,"props":53,"children":54},"p",{},[55,61,63,68,70,81,83,89,91,100,102,107,109,119,121,127,129,138,140,145,147,152,154,164,166,171,173,178],{"type":43,"tag":56,"props":57,"children":58},"strong",{},[59],{"type":49,"value":60},"Where to start:",{"type":49,"value":62}," This skill assumes DOCA is already\ninstalled, the user's BlueField has a DPA processor that\nthe host can see through DOCA (a BlueField-3-generation\ndevice per the README), the BlueField firmware has the\ncustom-PCC slot enabled, the DPACC compiler is installed at\na matched version per the DOCA Compatibility Policy, and the\nuser is doing ",{"type":43,"tag":56,"props":64,"children":65},{},[66],{"type":49,"value":67},"hands-on deployment of the DOCA-shipped ZTR\nRTTCC reference algorithm",{"type":49,"value":69}," on a BlueField port that\nalready carries RoCE-v2 traffic — i.e. either deploying it\nas the no-config-required baseline, tuning its documented\nparameters, or evaluating it against a custom algorithm the\nuser intends to write. Open ",{"type":43,"tag":71,"props":72,"children":74},"a",{"href":73},"TASKS.md",[75],{"type":43,"tag":76,"props":77,"children":79},"code",{"className":78},[],[80],{"type":49,"value":73},{"type":49,"value":82}," if the\nuser wants to ",{"type":43,"tag":84,"props":85,"children":86},"em",{},[87],{"type":49,"value":88},"do",{"type":49,"value":90}," something (install \u002F configure \u002F build \u002F\nmodify \u002F run \u002F test \u002F debug \u002F use); open\n",{"type":43,"tag":71,"props":92,"children":94},{"href":93},"CAPABILITIES.md",[95],{"type":43,"tag":76,"props":96,"children":98},{"className":97},[],[99],{"type":49,"value":93},{"type":49,"value":101}," when the question is\n",{"type":43,"tag":84,"props":103,"children":104},{},[105],{"type":49,"value":106},"what does the algorithm express, what are its variants and\nparameters, what does it ship vs not ship",{"type":49,"value":108},". If the user has\nnot installed DOCA yet, route to\n",{"type":43,"tag":71,"props":110,"children":112},{"href":111},"..\u002F..\u002Fdoca-setup\u002FSKILL.md",[113],{"type":43,"tag":76,"props":114,"children":116},{"className":115},[],[117],{"type":49,"value":118},"doca-setup",{"type":49,"value":120}," first; if the user\nhas not stood up the host-side ",{"type":43,"tag":76,"props":122,"children":124},{"className":123},[],[125],{"type":49,"value":126},"doca-pcc",{"type":49,"value":128}," framework yet,\nroute to ",{"type":43,"tag":71,"props":130,"children":132},{"href":131},"..\u002Fdoca-pcc\u002FSKILL.md",[133],{"type":43,"tag":76,"props":134,"children":136},{"className":135},[],[137],{"type":49,"value":126},{"type":49,"value":139}," first (this\nalgorithm is a ",{"type":43,"tag":84,"props":141,"children":142},{},[143],{"type":49,"value":144},"library consumed by",{"type":49,"value":146}," the PCC framework, not\na standalone program); if the user only wants to ",{"type":43,"tag":84,"props":148,"children":149},{},[150],{"type":49,"value":151},"inspect",{"type":49,"value":153},"\nPCC counters at runtime without changing the running\nalgorithm, route to\n",{"type":43,"tag":71,"props":155,"children":157},{"href":156},"..\u002F..\u002Ftools\u002Fdoca-pcc-counters\u002FSKILL.md",[158],{"type":43,"tag":76,"props":159,"children":161},{"className":160},[],[162],{"type":49,"value":163},"doca-pcc-counters",{"type":49,"value":165},";\nif the user wants to ",{"type":43,"tag":84,"props":167,"children":168},{},[169],{"type":49,"value":170},"write their own algorithm from\nscratch",{"type":49,"value":172},", that is the ",{"type":43,"tag":76,"props":174,"children":176},{"className":175},[],[177],{"type":49,"value":126},{"type":49,"value":179}," library plus the public\nPCC programming guide — this skill is for the shipped\nreference algorithm specifically.",{"type":43,"tag":181,"props":182,"children":184},"h2",{"id":183},"example-questions-this-skill-answers-well",[185],{"type":49,"value":186},"Example questions this skill answers well",{"type":43,"tag":52,"props":188,"children":189},{},[190,192,197],{"type":49,"value":191},"The CLASSES of ZTR RTTCC questions this skill is built to\nanswer, each with one worked example. The agent should\ntreat the ",{"type":43,"tag":84,"props":193,"children":194},{},[195],{"type":49,"value":196},"class",{"type":49,"value":198}," as the load-bearing piece — the worked\nexample is a single instance.",{"type":43,"tag":200,"props":201,"children":202},"ul",{},[203,245,305,338,393,444],{"type":43,"tag":204,"props":205,"children":206},"li",{},[207,212,214,219,221,231,233,243],{"type":43,"tag":56,"props":208,"children":209},{},[210],{"type":49,"value":211},"\"Is the ZTR RTTCC reference algorithm the right\nbaseline for my deployment, or should I write a custom\nalgorithm?\"",{"type":49,"value":213}," — worked example: ",{"type":43,"tag":84,"props":215,"children":216},{},[217],{"type":49,"value":218},"\"I have a BlueField-3\ncarrying production RoCE-v2 traffic from a GPU cluster;\nis the shipped algorithm a fine default or do I need\ncustom logic?\"",{"type":49,"value":220},". Answered by the decision rule in\n",{"type":43,"tag":71,"props":222,"children":224},{"href":223},"CAPABILITIES.md#capabilities-and-modes",[225],{"type":43,"tag":76,"props":226,"children":228},{"className":227},[],[229],{"type":49,"value":230},"CAPABILITIES.md ## Capabilities and modes",{"type":49,"value":232},"\n(\"when to use the reference vs custom\") + the env\npreconditions in\n",{"type":43,"tag":71,"props":234,"children":236},{"href":235},"TASKS.md#install",[237],{"type":43,"tag":76,"props":238,"children":240},{"className":239},[],[241],{"type":49,"value":242},"TASKS.md ## install",{"type":49,"value":244},".",{"type":43,"tag":204,"props":246,"children":247},{},[248,253,255,276,278,286],{"type":43,"tag":56,"props":249,"children":250},{},[251],{"type":49,"value":252},"\"How do I wire the shipped algorithm into the DOCA\nPCC application that's already running on my host?\"",{"type":49,"value":254}," —\nworked example: ",{"type":43,"tag":84,"props":256,"children":257},{},[258,260,266,268,274],{"type":49,"value":259},"\"",{"type":43,"tag":76,"props":261,"children":263},{"className":262},[],[264],{"type":49,"value":265},"\u002Fopt\u002Fmellanox\u002Fdoca\u002Fapplications\u002Fpcc",{"type":49,"value":267},"\nis already building from sample sources; what do I\nchange so the user algo callback dispatches to\n",{"type":43,"tag":76,"props":269,"children":271},{"className":270},[],[272],{"type":49,"value":273},"doca_pcc_dev_ztr_rttcc_algo",{"type":49,"value":275}," under a chosen algo\nslot?\"",{"type":49,"value":277},". Answered by the integration sequence in\n",{"type":43,"tag":71,"props":279,"children":280},{"href":223},[281],{"type":43,"tag":76,"props":282,"children":284},{"className":283},[],[285],{"type":49,"value":230},{"type":43,"tag":200,"props":287,"children":288},{},[289],{"type":43,"tag":204,"props":290,"children":291},{},[292,294,304],{"type":49,"value":293},"the in-place edits in\n",{"type":43,"tag":71,"props":295,"children":297},{"href":296},"TASKS.md#modify",[298],{"type":43,"tag":76,"props":299,"children":301},{"className":300},[],[302],{"type":49,"value":303},"TASKS.md ## modify",{"type":49,"value":244},{"type":43,"tag":204,"props":306,"children":307},{},[308,313,315,327,329,337],{"type":43,"tag":56,"props":309,"children":310},{},[311],{"type":49,"value":312},"\"Which variant of the algorithm am I getting — vanilla\nRTT-CC, path-migration mode, RX-rate mode, multipath,\nmultipath with credits, window-probeless?\"",{"type":49,"value":314}," — worked\nexample: ",{"type":43,"tag":84,"props":316,"children":317},{},[318,320,325],{"type":49,"value":319},"\"the shipped library exposes one public\nsymbol ",{"type":43,"tag":76,"props":321,"children":323},{"className":322},[],[324],{"type":49,"value":273},{"type":49,"value":326}," but the device-\nside source ships several variants; how do I know\nwhich one I get and how do I pick another?\"",{"type":49,"value":328},".\nAnswered by the variants table in\n",{"type":43,"tag":71,"props":330,"children":331},{"href":223},[332],{"type":43,"tag":76,"props":333,"children":335},{"className":334},[],[336],{"type":49,"value":230},{"type":49,"value":244},{"type":43,"tag":204,"props":339,"children":340},{},[341,346,347,352,354,364],{"type":43,"tag":56,"props":342,"children":343},{},[344],{"type":49,"value":345},"\"How do I confirm the algorithm is actually\nmodulating my RDMA \u002F RoCE traffic, and not just\nloading?\"",{"type":49,"value":213},{"type":43,"tag":84,"props":348,"children":349},{},[350],{"type":49,"value":351},"\"I followed the\nintegration steps; the application starts; how do I\nknow the algorithm is shaping flows under load?\"",{"type":49,"value":353},".\nAnswered by the observability surface in\n",{"type":43,"tag":71,"props":355,"children":357},{"href":356},"CAPABILITIES.md#observability",[358],{"type":43,"tag":76,"props":359,"children":361},{"className":360},[],[362],{"type":49,"value":363},"CAPABILITIES.md ## Observability",{"type":43,"tag":200,"props":365,"children":366},{},[367],{"type":43,"tag":204,"props":368,"children":369},{},[370,372,382,384,392],{"type":49,"value":371},"the counter-watch loop in\n",{"type":43,"tag":71,"props":373,"children":375},{"href":374},"TASKS.md#test",[376],{"type":43,"tag":76,"props":377,"children":379},{"className":378},[],[380],{"type":49,"value":381},"TASKS.md ## test",{"type":49,"value":383}," which routes to\n",{"type":43,"tag":71,"props":385,"children":386},{"href":156},[387],{"type":43,"tag":76,"props":388,"children":390},{"className":389},[],[391],{"type":49,"value":163},{"type":49,"value":244},{"type":43,"tag":204,"props":394,"children":395},{},[396,401,402,407,409,417],{"type":43,"tag":56,"props":397,"children":398},{},[399],{"type":49,"value":400},"\"Which tunables does the algorithm expose, and how\ndo I change them from the host without rebuilding the\nDPA-side image?\"",{"type":49,"value":213},{"type":43,"tag":84,"props":403,"children":404},{},[405],{"type":49,"value":406},"\"my workload is\nmore latency-sensitive than the default profile assumes\n— which parameter knob do I adjust?\"",{"type":49,"value":408},". Answered by\nthe parameter surface in\n",{"type":43,"tag":71,"props":410,"children":411},{"href":223},[412],{"type":43,"tag":76,"props":413,"children":415},{"className":414},[],[416],{"type":49,"value":230},{"type":43,"tag":200,"props":418,"children":419},{},[420],{"type":43,"tag":204,"props":421,"children":422},{},[423,425,431,433,443],{"type":49,"value":424},"the ",{"type":43,"tag":76,"props":426,"children":428},{"className":427},[],[429],{"type":49,"value":430},"doca_pcc_dev_set_ztr_rttcc_params",{"type":49,"value":432}," workflow in\n",{"type":43,"tag":71,"props":434,"children":436},{"href":435},"TASKS.md#use",[437],{"type":43,"tag":76,"props":438,"children":440},{"className":439},[],[441],{"type":49,"value":442},"TASKS.md ## use",{"type":49,"value":244},{"type":43,"tag":204,"props":445,"children":446},{},[447,476,478,490,492,502],{"type":43,"tag":56,"props":448,"children":449},{},[450,452,458,460,466,468,474],{"type":49,"value":451},"\"What does this ",{"type":43,"tag":76,"props":453,"children":455},{"className":454},[],[456],{"type":49,"value":457},"DOCA_PCC_DEV_STATUS_FAIL",{"type":49,"value":459}," or\n",{"type":43,"tag":76,"props":461,"children":463},{"className":462},[],[464],{"type":49,"value":465},"DOCA_ERROR_*",{"type":49,"value":467}," from a ",{"type":43,"tag":76,"props":469,"children":471},{"className":470},[],[472],{"type":49,"value":473},"doca_pcc_dev_ztr_rttcc_*",{"type":49,"value":475}," call\nmean and which layer caused it?\"",{"type":49,"value":477}," — worked example:\n",{"type":43,"tag":84,"props":479,"children":480},{},[481,483,488],{"type":49,"value":482},"\"my init callback returns ",{"type":43,"tag":76,"props":484,"children":486},{"className":485},[],[487],{"type":49,"value":457},{"type":49,"value":489},"\non first launch\"",{"type":49,"value":491},". Answered by the algorithm overlay\non the host-side PCC taxonomy in\n",{"type":43,"tag":71,"props":493,"children":495},{"href":494},"CAPABILITIES.md#error-taxonomy",[496],{"type":43,"tag":76,"props":497,"children":499},{"className":498},[],[500],{"type":49,"value":501},"CAPABILITIES.md ## Error taxonomy",{"type":43,"tag":200,"props":503,"children":504},{},[505],{"type":43,"tag":204,"props":506,"children":507},{},[508,510,520,522,530,532,542],{"type":49,"value":509},"the layered ladder in\n",{"type":43,"tag":71,"props":511,"children":513},{"href":512},"TASKS.md#debug",[514],{"type":43,"tag":76,"props":515,"children":517},{"className":516},[],[518],{"type":49,"value":519},"TASKS.md ## debug",{"type":49,"value":521}," that escalates\nthrough\n",{"type":43,"tag":71,"props":523,"children":524},{"href":131},[525],{"type":43,"tag":76,"props":526,"children":528},{"className":527},[],[529],{"type":49,"value":126},{"type":49,"value":531}," and\n",{"type":43,"tag":71,"props":533,"children":535},{"href":534},"..\u002F..\u002Fdoca-debug\u002FSKILL.md",[536],{"type":43,"tag":76,"props":537,"children":539},{"className":538},[],[540],{"type":49,"value":541},"doca-debug",{"type":49,"value":244},{"type":43,"tag":181,"props":544,"children":546},{"id":545},"audience",[547],{"type":49,"value":548},"Audience",{"type":43,"tag":52,"props":550,"children":551},{},[552,554,559,561,566,568,573,575,579,581,589],{"type":49,"value":553},"This skill serves ",{"type":43,"tag":56,"props":555,"children":556},{},[557],{"type":49,"value":558},"external developers operating a\nBlueField-3-class DPU who want to deploy NVIDIA's shipped\nreference PCC algorithm on RoCE-v2 traffic, OR who are\nevaluating it against a custom algorithm they intend to\nwrite",{"type":49,"value":560},". The reference algorithm is ",{"type":43,"tag":84,"props":562,"children":563},{},[564],{"type":49,"value":565},"zero-touch",{"type":49,"value":567}," by design\n— the no-config-required baseline — and the canonical use\ncase is dropping it onto a port and confirming it shapes\nflows correctly under congestion. It is ",{"type":43,"tag":84,"props":569,"children":570},{},[571],{"type":49,"value":572},"not",{"type":49,"value":574}," for NVIDIA\ndevelopers contributing to the algorithm itself, nor for\nusers who want general PCC programming theory (route via\nthe public DOCA PCC programming guide), nor for users who\nonly want to ",{"type":43,"tag":84,"props":576,"children":577},{},[578],{"type":49,"value":151},{"type":49,"value":580}," PCC counters (route to\n",{"type":43,"tag":71,"props":582,"children":583},{"href":156},[584],{"type":43,"tag":76,"props":585,"children":587},{"className":586},[],[588],{"type":49,"value":163},{"type":49,"value":590},").",{"type":43,"tag":52,"props":592,"children":593},{},[594,599,601,607,609,614,616,622,624,630,632,640,642,648,650,655,657,662,664,669],{"type":43,"tag":56,"props":595,"children":596},{},[597],{"type":49,"value":598},"Language scope.",{"type":49,"value":600}," The algorithm ships as a DPA-side\nlibrary (",{"type":43,"tag":76,"props":602,"children":604},{"className":603},[],[605],{"type":49,"value":606},"pkg-config",{"type":49,"value":608}," module ",{"type":43,"tag":76,"props":610,"children":612},{"className":611},[],[613],{"type":49,"value":4},{"type":49,"value":615},")\nplus a public header ",{"type":43,"tag":76,"props":617,"children":619},{"className":618},[],[620],{"type":49,"value":621},"doca_pcc_dev_ztr_rttcc_algo.h",{"type":49,"value":623}," that\nDPA-side translation units include. The shipped algorithm\nbinary is the static library\n",{"type":43,"tag":76,"props":625,"children":627},{"className":626},[],[628],{"type":49,"value":629},"libdoca_pcc_ztr_rttcc_algo_dev.a",{"type":49,"value":631}," per the README; the\ndevice-side translation unit that consumes it is C and is\ncompiled by DPACC. The host-side that drives the PCC\ncontext comes from ",{"type":43,"tag":71,"props":633,"children":634},{"href":131},[635],{"type":43,"tag":76,"props":636,"children":638},{"className":637},[],[639],{"type":49,"value":126},{"type":49,"value":641},";\nthis library does NOT add a host-side surface beyond the\nhost-side helpers (also shipped as\n",{"type":43,"tag":76,"props":643,"children":645},{"className":644},[],[646],{"type":49,"value":647},"libdoca_pcc_ztr_rttcc_algo.{a,so}",{"type":49,"value":649}," per the README) that\nthe ",{"type":43,"tag":76,"props":651,"children":653},{"className":652},[],[654],{"type":49,"value":126},{"type":49,"value":656}," framework links. Other-language host-side\nwrappers around ",{"type":43,"tag":76,"props":658,"children":660},{"className":659},[],[661],{"type":49,"value":126},{"type":49,"value":663}," can drive this algorithm through\nthe same lifecycle described in ",{"type":43,"tag":76,"props":665,"children":667},{"className":666},[],[668],{"type":49,"value":126},{"type":49,"value":670},"; the DPA-side\nintegration always stays C-via-DPACC.",{"type":43,"tag":181,"props":672,"children":674},{"id":673},"when-to-load-this-skill",[675],{"type":49,"value":676},"When to load this skill",{"type":43,"tag":52,"props":678,"children":679},{},[680,682,688],{"type":49,"value":681},"Load this skill when the user is doing hands-on deployment,\ntuning, or evaluation of the DOCA-shipped ZTR RTTCC\nreference algorithm on a BlueField port carrying RoCE-v2\ntraffic, in any host language plus the DPA-side translation\nunit built by ",{"type":43,"tag":76,"props":683,"children":685},{"className":684},[],[686],{"type":49,"value":687},"dpacc",{"type":49,"value":689},". Concretely:",{"type":43,"tag":200,"props":691,"children":692},{},[693,698,703,722,740,755],{"type":43,"tag":204,"props":694,"children":695},{},[696],{"type":49,"value":697},"Deciding whether the shipped algorithm is the right\nbaseline for the user's RoCE-v2 workload, or whether the\nuser needs a custom algorithm.",{"type":43,"tag":204,"props":699,"children":700},{},[701],{"type":49,"value":702},"Wiring the algorithm into the DOCA PCC sample\napplication by patching the user-algo \u002F user-init \u002F\nuser-set-algo-params callbacks per the steps the\nshipped README documents.",{"type":43,"tag":204,"props":704,"children":705},{},[706,708,713,715,720],{"type":49,"value":707},"Picking which of the algorithm's documented variants\n(vanilla, path-migration, RX-rate, RX-rate + PM,\nmultipath, multipath + credits, window-probeless)\nmatches the user's intent — and surfacing that the\n",{"type":43,"tag":84,"props":709,"children":710},{},[711],{"type":49,"value":712},"public",{"type":49,"value":714}," surface only ships one symbol\n(",{"type":43,"tag":76,"props":716,"children":718},{"className":717},[],[719],{"type":49,"value":273},{"type":49,"value":721},"); the variants live in\nthe DPA-side source the user compiles against.",{"type":43,"tag":204,"props":723,"children":724},{},[725,727,732,734,739],{"type":49,"value":726},"Tuning the host-set parameters the algorithm exposes\n(per the parameter list in\n",{"type":43,"tag":76,"props":728,"children":730},{"className":729},[],[731],{"type":49,"value":621},{"type":49,"value":733}," and the shipped\n",{"type":43,"tag":76,"props":735,"children":737},{"className":736},[],[738],{"type":49,"value":430},{"type":49,"value":590},{"type":43,"tag":204,"props":741,"children":742},{},[743,745,753],{"type":49,"value":744},"Observing whether the running algorithm is actually\nmodulating RoCE-v2 flows under load (route to\n",{"type":43,"tag":71,"props":746,"children":747},{"href":156},[748],{"type":43,"tag":76,"props":749,"children":751},{"className":750},[],[752],{"type":49,"value":163},{"type":49,"value":754},"\nfor the read-only inspection side).",{"type":43,"tag":204,"props":756,"children":757},{},[758,760,765],{"type":49,"value":759},"Deciding when to ",{"type":43,"tag":84,"props":761,"children":762},{},[763],{"type":49,"value":764},"replace",{"type":49,"value":766}," the shipped algorithm with a\ncustom one because latency target, fairness policy, or\nconvergence behavior requirements diverge from what the\nreference provides.",{"type":43,"tag":52,"props":768,"children":769},{},[770,772,776,778,783,785,793,795,803,805,815,817,825,827,835],{"type":49,"value":771},"Do ",{"type":43,"tag":56,"props":773,"children":774},{},[775],{"type":49,"value":572},{"type":49,"value":777}," load this skill for general DOCA orientation;\nfor the host-side ",{"type":43,"tag":76,"props":779,"children":781},{"className":780},[],[782],{"type":49,"value":126},{"type":49,"value":784}," lifecycle (route to\n",{"type":43,"tag":71,"props":786,"children":787},{"href":131},[788],{"type":43,"tag":76,"props":789,"children":791},{"className":790},[],[792],{"type":49,"value":126},{"type":49,"value":794},"); for writing a custom\nalgorithm from scratch (route to\n",{"type":43,"tag":71,"props":796,"children":797},{"href":131},[798],{"type":43,"tag":76,"props":799,"children":801},{"className":800},[],[802],{"type":49,"value":126},{"type":49,"value":804}," and the public PCC\nprogramming guide via\n",{"type":43,"tag":71,"props":806,"children":808},{"href":807},"..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md",[809],{"type":43,"tag":76,"props":810,"children":812},{"className":811},[],[813],{"type":49,"value":814},"doca-public-knowledge-map",{"type":49,"value":816},");\nfor read-only PCC counter inspection (route to\n",{"type":43,"tag":71,"props":818,"children":819},{"href":156},[820],{"type":43,"tag":76,"props":821,"children":823},{"className":822},[],[824],{"type":49,"value":163},{"type":49,"value":826},");\nor for the default firmware-shipped PCC algorithms that\npredate Programmable Congestion Control entirely (no\nhost-side code, no DPACC compile — that is a firmware-only\npath routed via\n",{"type":43,"tag":71,"props":828,"children":829},{"href":807},[830],{"type":43,"tag":76,"props":831,"children":833},{"className":832},[],[834],{"type":49,"value":814},{"type":49,"value":590},{"type":43,"tag":181,"props":837,"children":839},{"id":838},"what-this-skill-provides",[840],{"type":49,"value":841},"What this skill provides",{"type":43,"tag":52,"props":843,"children":844},{},[845,847,852],{"type":49,"value":846},"This is a ",{"type":43,"tag":56,"props":848,"children":849},{},[850],{"type":49,"value":851},"thin loader",{"type":49,"value":853},". The body keeps only the\norientation needed to pick the right next file. The\nsubstantive algorithm-specific material lives in two\ncompanion files:",{"type":43,"tag":200,"props":855,"children":856},{},[857,939],{"type":43,"tag":204,"props":858,"children":859},{},[860,865,867,873,875,880,881,886,887,893,894,900,901,907,909,914,916,921,923,929,931,937],{"type":43,"tag":76,"props":861,"children":863},{"className":862},[],[864],{"type":49,"value":93},{"type":49,"value":866}," — what the shipped ZTR RTTCC\nalgorithm expresses on this version + this BlueField\ngeneration + this firmware: the public DPA-side API\nsurface (",{"type":43,"tag":76,"props":868,"children":870},{"className":869},[],[871],{"type":49,"value":872},"doca_pcc_dev_ztr_rttcc_init",{"type":49,"value":874},",\n",{"type":43,"tag":76,"props":876,"children":878},{"className":877},[],[879],{"type":49,"value":273},{"type":49,"value":874},{"type":43,"tag":76,"props":882,"children":884},{"className":883},[],[885],{"type":49,"value":430},{"type":49,"value":874},{"type":43,"tag":76,"props":888,"children":890},{"className":889},[],[891],{"type":49,"value":892},"doca_pcc_dev_ztr_rttcc_get_param_num",{"type":49,"value":874},{"type":43,"tag":76,"props":895,"children":897},{"className":896},[],[898],{"type":49,"value":899},"doca_pcc_dev_ztr_rttcc_get_counter_num",{"type":49,"value":874},{"type":43,"tag":76,"props":902,"children":904},{"className":903},[],[905],{"type":49,"value":906},"doca_pcc_dev_ztr_rttcc_get_num_of_histograms",{"type":49,"value":908},"), the\ndocumented variants (vanilla \u002F path-migration \u002F RX-rate\n\u002F multipath \u002F window-probeless — pick one at\nDPA-side compile time), the relationship to the\nhost-side ",{"type":43,"tag":76,"props":910,"children":912},{"className":911},[],[913],{"type":49,"value":126},{"type":49,"value":915}," framework (this is an algorithm\nbody the framework loads), the relationship to the\n",{"type":43,"tag":76,"props":917,"children":919},{"className":918},[],[920],{"type":49,"value":163},{"type":49,"value":922}," tool (which is the canonical\ninspection surface), the algorithm's parameter and\ncounter surface (RTT-based congestion signal,\nper-feature parameter blocks), the error taxonomy in\n",{"type":43,"tag":76,"props":924,"children":926},{"className":925},[],[927],{"type":49,"value":928},"DOCA_PCC_DEV_STATUS_OK",{"type":49,"value":930}," \u002F ",{"type":43,"tag":76,"props":932,"children":934},{"className":933},[],[935],{"type":49,"value":936},"_FAIL",{"type":49,"value":938},", and the safety\npolicy.",{"type":43,"tag":204,"props":940,"children":941},{},[942,947,949,955,957,963,964,970,971,977,978,984,985,991,992,998,999,1005,1007,1013],{"type":43,"tag":76,"props":943,"children":945},{"className":944},[],[946],{"type":49,"value":73},{"type":49,"value":948}," — step-by-step workflows for the in-scope\nalgorithm verbs: ",{"type":43,"tag":76,"props":950,"children":952},{"className":951},[],[953],{"type":49,"value":954},"install",{"type":49,"value":956},", ",{"type":43,"tag":76,"props":958,"children":960},{"className":959},[],[961],{"type":49,"value":962},"configure",{"type":49,"value":956},{"type":43,"tag":76,"props":965,"children":967},{"className":966},[],[968],{"type":49,"value":969},"build",{"type":49,"value":874},{"type":43,"tag":76,"props":972,"children":974},{"className":973},[],[975],{"type":49,"value":976},"modify",{"type":49,"value":956},{"type":43,"tag":76,"props":979,"children":981},{"className":980},[],[982],{"type":49,"value":983},"run",{"type":49,"value":956},{"type":43,"tag":76,"props":986,"children":988},{"className":987},[],[989],{"type":49,"value":990},"test",{"type":49,"value":956},{"type":43,"tag":76,"props":993,"children":995},{"className":994},[],[996],{"type":49,"value":997},"debug",{"type":49,"value":956},{"type":43,"tag":76,"props":1000,"children":1002},{"className":1001},[],[1003],{"type":49,"value":1004},"use",{"type":49,"value":1006},". Plus a\n",{"type":43,"tag":76,"props":1008,"children":1010},{"className":1009},[],[1011],{"type":49,"value":1012},"Deferred task verbs",{"type":49,"value":1014}," block that points out-of-scope\nquestions at the right next skill.",{"type":43,"tag":52,"props":1016,"children":1017},{},[1018,1020,1028,1030,1040,1042,1047],{"type":49,"value":1019},"The skill assumes DOCA + the DPACC compiler + the\n",{"type":43,"tag":71,"props":1021,"children":1022},{"href":131},[1023],{"type":43,"tag":76,"props":1024,"children":1026},{"className":1025},[],[1027],{"type":49,"value":126},{"type":49,"value":1029}," host-side framework are\nalready installed; the BlueField is a generation that\nexposes the DPA processor (the algorithm runs on the DPA);\nthe BlueField firmware has the custom-PCC slot enabled\n(inherited from\n",{"type":43,"tag":71,"props":1031,"children":1033},{"href":1032},"..\u002Fdoca-pcc\u002FCAPABILITIES.md#safety-policy",[1034],{"type":43,"tag":76,"props":1035,"children":1037},{"className":1036},[],[1038],{"type":49,"value":1039},"doca-pcc CAPABILITIES.md ## Safety policy",{"type":49,"value":1041},");\nand the BlueField port the algorithm will modulate has\nRoCE-v2 traffic actually flowing on it (the algorithm\nmodulates ",{"type":43,"tag":84,"props":1043,"children":1044},{},[1045],{"type":49,"value":1046},"existing",{"type":49,"value":1048}," RDMA \u002F RoCE traffic — without traffic,\nthere is nothing for it to do).",{"type":43,"tag":181,"props":1050,"children":1052},{"id":1051},"what-this-skill-deliberately-does-not-ship",[1053],{"type":49,"value":1054},"What this skill deliberately does not ship",{"type":43,"tag":52,"props":1056,"children":1057},{},[1058,1060,1065],{"type":49,"value":1059},"This skill is ",{"type":43,"tag":56,"props":1061,"children":1062},{},[1063],{"type":49,"value":1064},"agent guidance",{"type":49,"value":1066},", not a samples or templates\nbundle. To keep the boundary clean, it deliberately does not\ncontain — and pull requests should not add:",{"type":43,"tag":200,"props":1068,"children":1069},{},[1070,1094,1112,1132],{"type":43,"tag":204,"props":1071,"children":1072},{},[1073,1078,1080,1085,1087,1092],{"type":43,"tag":56,"props":1074,"children":1075},{},[1076],{"type":49,"value":1077},"The algorithm body itself, in any language.",{"type":49,"value":1079}," The\nshipped algorithm is the static library\n",{"type":43,"tag":76,"props":1081,"children":1083},{"className":1082},[],[1084],{"type":49,"value":629},{"type":49,"value":1086}," (plus the host-side\nhelpers) installed by the matching DOCA host package\nper the README. The agent's job is to route the user to\nthe installed library and header\n(",{"type":43,"tag":76,"props":1088,"children":1090},{"className":1089},[],[1091],{"type":49,"value":621},{"type":49,"value":1093},") and to prescribe the\nin-place edits documented in the README on the\nshipped DOCA PCC application source, not to author the\nalgorithm.",{"type":43,"tag":204,"props":1095,"children":1096},{},[1097,1102,1104,1110],{"type":43,"tag":56,"props":1098,"children":1099},{},[1100],{"type":49,"value":1101},"A standalone PCC application.",{"type":49,"value":1103}," The DOCA PCC\napplication that hosts this algorithm is the shipped\nC sample under\n",{"type":43,"tag":76,"props":1105,"children":1107},{"className":1106},[],[1108],{"type":49,"value":1109},"\u002Fopt\u002Fmellanox\u002Fdoca\u002Fapplications\u002Fpcc\u002F",{"type":49,"value":1111}," (per the\nREADME). The agent's job is to prescribe the\nminimum-diff modifications the README documents and to\nwalk the user through the rebuild — not to author a\nparallel application.",{"type":43,"tag":204,"props":1113,"children":1114},{},[1115,1120,1122,1130],{"type":43,"tag":56,"props":1116,"children":1117},{},[1118],{"type":49,"value":1119},"A description of every internal variant's\nbehavior.",{"type":49,"value":1121}," The DPA-side source ships several\nvariants (vanilla, path-migration, RX-rate, multipath,\nmultipath + credits, window-probeless). The agent\nnames the variant set, says which one is reachable via\nthe public symbol on this install, and routes\nalgorithm-design questions to the public PCC\nprogramming guide via\n",{"type":43,"tag":71,"props":1123,"children":1124},{"href":807},[1125],{"type":43,"tag":76,"props":1126,"children":1128},{"className":1127},[],[1129],{"type":49,"value":814},{"type":49,"value":1131},".\nIt does not redefine each variant's mathematical\nbehavior.",{"type":43,"tag":204,"props":1133,"children":1134},{},[1135,1140],{"type":43,"tag":56,"props":1136,"children":1137},{},[1138],{"type":49,"value":1139},"A specific congestion-control algorithm tutorial.",{"type":49,"value":1141},"\nCongestion-control theory, RoCE-v2 fairness analysis,\nworkload-specific tuning — out of scope. Route to the\nuser's own domain expertise and to the public DOCA PCC\nguide.",{"type":43,"tag":181,"props":1143,"children":1145},{"id":1144},"loading-order",[1146],{"type":49,"value":1147},"Loading order",{"type":43,"tag":1149,"props":1150,"children":1151},"ol",{},[1152,1165,1192],{"type":43,"tag":204,"props":1153,"children":1154},{},[1155,1157,1163],{"type":49,"value":1156},"Read this ",{"type":43,"tag":76,"props":1158,"children":1160},{"className":1159},[],[1161],{"type":49,"value":1162},"SKILL.md",{"type":49,"value":1164}," first to confirm the user's\nquestion is in scope (deployment \u002F tuning \u002F evaluation\nof the shipped reference algorithm, not algorithm\ndesign from scratch and not read-only counter\ninspection).",{"type":43,"tag":204,"props":1166,"children":1167},{},[1168],{"type":43,"tag":56,"props":1169,"children":1170},{},[1171,1173,1178,1180,1185,1187,1191],{"type":49,"value":1172},"For the algorithm capability matrix, the public\nDPA-side API surface, the variant set, the parameter\nand counter surface, the host-side ",{"type":43,"tag":76,"props":1174,"children":1176},{"className":1175},[],[1177],{"type":49,"value":126},{"type":49,"value":1179},"\nframework relationship, the\n",{"type":43,"tag":76,"props":1181,"children":1183},{"className":1182},[],[1184],{"type":49,"value":163},{"type":49,"value":1186}," inspection-side relationship, the\nerror taxonomy, the observability surface, and the\nsafety policy, see ",{"type":43,"tag":71,"props":1188,"children":1189},{"href":93},[1190],{"type":49,"value":93},{"type":49,"value":244},{"type":43,"tag":204,"props":1193,"children":1194},{},[1195],{"type":43,"tag":56,"props":1196,"children":1197},{},[1198,1200,1204],{"type":49,"value":1199},"For step-by-step workflows — install, configure,\nbuild, modify, run, test, debug, use — see\n",{"type":43,"tag":71,"props":1201,"children":1202},{"href":73},[1203],{"type":49,"value":73},{"type":49,"value":244},{"type":43,"tag":52,"props":1206,"children":1207},{},[1208,1210,1218,1220,1228,1230,1240,1242,1252,1254,1262,1263,1271,1273,1281],{"type":49,"value":1209},"Both companion files cross-link to each other,\n",{"type":43,"tag":71,"props":1211,"children":1212},{"href":131},[1213],{"type":43,"tag":76,"props":1214,"children":1216},{"className":1215},[],[1217],{"type":49,"value":126},{"type":49,"value":1219}," for the host-side PCC\nlifecycle that loads this algorithm,\n",{"type":43,"tag":71,"props":1221,"children":1222},{"href":156},[1223],{"type":43,"tag":76,"props":1224,"children":1226},{"className":1225},[],[1227],{"type":49,"value":163},{"type":49,"value":1229},"\nfor the read-only counter-inspection side of validating that\nthe algorithm is modulating traffic,\n",{"type":43,"tag":71,"props":1231,"children":1233},{"href":1232},"..\u002Fdoca-dpa\u002FSKILL.md",[1234],{"type":43,"tag":76,"props":1235,"children":1237},{"className":1236},[],[1238],{"type":49,"value":1239},"doca-dpa",{"type":49,"value":1241}," for the DPA-side\ntwo-side-program model and the DPACC compiler discipline,\n",{"type":43,"tag":71,"props":1243,"children":1245},{"href":1244},"..\u002F..\u002Fdoca-version\u002FSKILL.md",[1246],{"type":43,"tag":76,"props":1247,"children":1249},{"className":1248},[],[1250],{"type":49,"value":1251},"doca-version",{"type":49,"value":1253}," for the\ncanonical DOCA version-handling rules (with the DPACC\noverlay inherited from\n",{"type":43,"tag":71,"props":1255,"children":1256},{"href":1232},[1257],{"type":43,"tag":76,"props":1258,"children":1260},{"className":1259},[],[1261],{"type":49,"value":1239},{"type":49,"value":531},{"type":43,"tag":71,"props":1264,"children":1265},{"href":131},[1266],{"type":43,"tag":76,"props":1267,"children":1269},{"className":1268},[],[1270],{"type":49,"value":126},{"type":49,"value":1272},"), and\n",{"type":43,"tag":71,"props":1274,"children":1275},{"href":807},[1276],{"type":43,"tag":76,"props":1277,"children":1279},{"className":1278},[],[1280],{"type":49,"value":814},{"type":49,"value":1282},"\nwhenever the right answer is \"look it up in the public DOCA\nPCC programming guide or the on-disk install layout\".",{"type":43,"tag":181,"props":1284,"children":1286},{"id":1285},"related-skills",[1287],{"type":49,"value":1288},"Related skills",{"type":43,"tag":200,"props":1290,"children":1291},{},[1292,1342,1362,1375,1397,1417,1447,1468,1483,1496],{"type":43,"tag":204,"props":1293,"children":1294},{},[1295,1303,1305,1311,1313,1318,1320,1325,1327,1333,1335,1340],{"type":43,"tag":71,"props":1296,"children":1297},{"href":131},[1298],{"type":43,"tag":76,"props":1299,"children":1301},{"className":1300},[],[1302],{"type":49,"value":126},{"type":49,"value":1304}," — the host-side PCC\ncontrol library. This algorithm is loaded INTO a\n",{"type":43,"tag":76,"props":1306,"children":1308},{"className":1307},[],[1309],{"type":49,"value":1310},"doca_pcc",{"type":49,"value":1312}," context that ",{"type":43,"tag":76,"props":1314,"children":1316},{"className":1315},[],[1317],{"type":49,"value":126},{"type":49,"value":1319}," stands up; the\nhost-side lifecycle (",{"type":43,"tag":76,"props":1321,"children":1323},{"className":1322},[],[1324],{"type":49,"value":1310},{"type":49,"value":1326}," create \u002F configure \u002F\nstart \u002F stop \u002F destroy, the algorithm image\n",{"type":43,"tag":76,"props":1328,"children":1330},{"className":1329},[],[1331],{"type":49,"value":1332},"doca_pcc_app",{"type":49,"value":1334},", the attach-to-port semantics) is owned\nby ",{"type":43,"tag":76,"props":1336,"children":1338},{"className":1337},[],[1339],{"type":49,"value":126},{"type":49,"value":1341},". This skill prescribes only the DPA-side\nalgorithm integration on top.",{"type":43,"tag":204,"props":1343,"children":1344},{},[1345,1353,1355,1360],{"type":43,"tag":71,"props":1346,"children":1347},{"href":156},[1348],{"type":43,"tag":76,"props":1349,"children":1351},{"className":1350},[],[1352],{"type":49,"value":163},{"type":49,"value":1354}," —\nthe read-only diagnostic CLI for PCC counters at the\nport. The canonical ",{"type":43,"tag":84,"props":1356,"children":1357},{},[1358],{"type":49,"value":1359},"\"is the algorithm actually\nmodulating traffic\"",{"type":49,"value":1361}," check goes through the counter\ntool; this skill names what counters the algorithm\nemits (CNP \u002F NACK \u002F AI \u002F HAI \u002F decrement \u002F RTT-band\ncounters per the public header) and routes the\ninspection workflow to the tool skill.",{"type":43,"tag":204,"props":1363,"children":1364},{},[1365,1373],{"type":43,"tag":71,"props":1366,"children":1367},{"href":1232},[1368],{"type":43,"tag":76,"props":1369,"children":1371},{"className":1370},[],[1372],{"type":49,"value":1239},{"type":49,"value":1374}," — the host-side\nDPA control library. The algorithm runs on the DPA,\ncompiled by DPACC; the two-side-program rule and the\nDOCA-and-DPACC version-match overlay inherited from\nhere apply.",{"type":43,"tag":204,"props":1376,"children":1377},{},[1378,1386,1388,1395],{"type":43,"tag":71,"props":1379,"children":1380},{"href":807},[1381],{"type":43,"tag":76,"props":1382,"children":1384},{"className":1383},[],[1385],{"type":49,"value":814},{"type":49,"value":1387}," —\nthe routing table for every public DOCA documentation\nsource (the DOCA PCC programming guide at\n",{"type":43,"tag":71,"props":1389,"children":1393},{"href":1390,"rel":1391},"https:\u002F\u002Fdocs.nvidia.com\u002Fdoca\u002Fsdk\u002Fdoca-pcc\u002Findex.html",[1392],"nofollow",[1394],{"type":49,"value":1390},{"type":49,"value":1396},";\nthe DOCA PCC application guide; the DOCA Compatibility\nPolicy) and the on-disk layout of an installed DOCA\npackage.",{"type":43,"tag":204,"props":1398,"children":1399},{},[1400,1408,1410,1415],{"type":43,"tag":71,"props":1401,"children":1402},{"href":111},[1403],{"type":43,"tag":76,"props":1404,"children":1406},{"className":1405},[],[1407],{"type":49,"value":118},{"type":49,"value":1409}," — env\npreparation, install verification, DPACC compiler\ninstall \u002F verification, BlueField firmware\nconfiguration (including the custom-PCC slot enable),\nand the ",{"type":43,"tag":84,"props":1411,"children":1412},{},[1413],{"type":49,"value":1414},"I have no install yet",{"type":49,"value":1416}," path with the public\nNGC DOCA container. This skill assumes its\npreconditions are satisfied AND that DPACC is installed\nat a version that matches DOCA AND that the firmware-\nlevel custom-PCC slot is enabled.",{"type":43,"tag":204,"props":1418,"children":1419},{},[1420,1428,1430,1436,1438,1446],{"type":43,"tag":71,"props":1421,"children":1422},{"href":1244},[1423],{"type":43,"tag":76,"props":1424,"children":1426},{"className":1425},[],[1427],{"type":49,"value":1251},{"type":49,"value":1429}," —\ncanonical DOCA version-handling rules. This skill's\n",{"type":43,"tag":76,"props":1431,"children":1433},{"className":1432},[],[1434],{"type":49,"value":1435},"## Version compatibility",{"type":49,"value":1437}," cross-links the four-way\nmatch rule plus the DOCA-and-DPACC overlay inherited\nfrom ",{"type":43,"tag":71,"props":1439,"children":1440},{"href":131},[1441],{"type":43,"tag":76,"props":1442,"children":1444},{"className":1443},[],[1445],{"type":49,"value":126},{"type":49,"value":244},{"type":43,"tag":204,"props":1448,"children":1449},{},[1450,1460,1462,1466],{"type":43,"tag":71,"props":1451,"children":1453},{"href":1452},"..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md",[1454],{"type":43,"tag":76,"props":1455,"children":1457},{"className":1456},[],[1458],{"type":49,"value":1459},"doca-structured-tools-contract",{"type":49,"value":1461}," —\nthe bundle's structured-tools precedence rule (detect \u002F\nprefer \u002F fall back \u002F report). The Command appendix in\n",{"type":43,"tag":71,"props":1463,"children":1464},{"href":73},[1465],{"type":49,"value":73},{"type":49,"value":1467}," honors this contract.",{"type":43,"tag":204,"props":1469,"children":1470},{},[1471,1481],{"type":43,"tag":71,"props":1472,"children":1474},{"href":1473},"..\u002F..\u002Fdoca-programming-guide\u002FSKILL.md",[1475],{"type":43,"tag":76,"props":1476,"children":1478},{"className":1477},[],[1479],{"type":49,"value":1480},"doca-programming-guide",{"type":49,"value":1482}," —\ngeneral DOCA programming patterns. This skill layers\nalgorithm-specific overlays on top of the universal\nbuild, modify-a-shipped-sample, and Core lifecycle\npatterns.",{"type":43,"tag":204,"props":1484,"children":1485},{},[1486,1494],{"type":43,"tag":71,"props":1487,"children":1488},{"href":534},[1489],{"type":43,"tag":76,"props":1490,"children":1492},{"className":1491},[],[1493],{"type":49,"value":541},{"type":49,"value":1495}," — cross-cutting\ndebug ladder. Algorithm-specific debug (the algorithm\nloaded but counters do not move; the algorithm fails to\ninitialize; the algorithm modulates traffic too\naggressively \u002F too gently for the workload) overlays on\ntop of that ladder.",{"type":43,"tag":204,"props":1497,"children":1498},{},[1499,1509,1511,1517],{"type":43,"tag":71,"props":1500,"children":1502},{"href":1501},"..\u002F..\u002Fdoca-hardware-safety\u002FSKILL.md",[1503],{"type":43,"tag":76,"props":1504,"children":1506},{"className":1505},[],[1507],{"type":49,"value":1508},"doca-hardware-safety",{"type":49,"value":1510}," —\ncross-cutting hardware-safety meta-policy. Because the\nalgorithm modulates production RoCE-v2 flows on a\nBlueField port, the meta-policy's pre-flight inventory,\nreplica-first, and rollback rules apply via this\nskill's ",{"type":43,"tag":76,"props":1512,"children":1514},{"className":1513},[],[1515],{"type":49,"value":1516},"## Safety policy",{"type":49,"value":1518}," overlay.",{"items":1520,"total":1679},[1521,1539,1557,1568,1580,1594,1607,1621,1634,1645,1659,1668],{"slug":1522,"name":1522,"fn":1523,"description":1524,"org":1525,"tags":1526,"stars":1536,"repoUrl":1537,"updatedAt":1538},"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},[1527,1530,1533],{"name":1528,"slug":1529,"type":15},"Documentation","documentation",{"name":1531,"slug":1532,"type":15},"MCP","mcp",{"name":1534,"slug":1535,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1540,"name":1540,"fn":1541,"description":1542,"org":1543,"tags":1544,"stars":1554,"repoUrl":1555,"updatedAt":1556},"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},[1545,1548,1551],{"name":1546,"slug":1547,"type":15},"Containers","containers",{"name":1549,"slug":1550,"type":15},"Deployment","deployment",{"name":1552,"slug":1553,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1558,"name":1558,"fn":1559,"description":1560,"org":1561,"tags":1562,"stars":1554,"repoUrl":1555,"updatedAt":1567},"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},[1563,1566],{"name":1564,"slug":1565,"type":15},"CI\u002FCD","ci-cd",{"name":1549,"slug":1550,"type":15},"2026-07-14T05:25:59.97109",{"slug":1569,"name":1569,"fn":1570,"description":1571,"org":1572,"tags":1573,"stars":1554,"repoUrl":1555,"updatedAt":1579},"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},[1574,1575,1576],{"name":1564,"slug":1565,"type":15},{"name":1549,"slug":1550,"type":15},{"name":1577,"slug":1578,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1581,"name":1581,"fn":1582,"description":1583,"org":1584,"tags":1585,"stars":1554,"repoUrl":1555,"updatedAt":1593},"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},[1586,1589,1590],{"name":1587,"slug":1588,"type":15},"Debugging","debugging",{"name":1577,"slug":1578,"type":15},{"name":1591,"slug":1592,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1595,"name":1595,"fn":1596,"description":1597,"org":1598,"tags":1599,"stars":1554,"repoUrl":1555,"updatedAt":1606},"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},[1600,1603],{"name":1601,"slug":1602,"type":15},"Best Practices","best-practices",{"name":1604,"slug":1605,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1608,"name":1608,"fn":1609,"description":1610,"org":1611,"tags":1612,"stars":1554,"repoUrl":1555,"updatedAt":1620},"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},[1613,1616,1619],{"name":1614,"slug":1615,"type":15},"Machine Learning","machine-learning",{"name":1617,"slug":1618,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1622,"name":1622,"fn":1623,"description":1624,"org":1625,"tags":1626,"stars":1554,"repoUrl":1555,"updatedAt":1633},"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},[1627,1630],{"name":1628,"slug":1629,"type":15},"QA","qa",{"name":1631,"slug":1632,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1635,"name":1635,"fn":1636,"description":1637,"org":1638,"tags":1639,"stars":1554,"repoUrl":1555,"updatedAt":1644},"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},[1640,1641],{"name":1549,"slug":1550,"type":15},{"name":1642,"slug":1643,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1646,"name":1646,"fn":1647,"description":1648,"org":1649,"tags":1650,"stars":1554,"repoUrl":1555,"updatedAt":1658},"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},[1651,1654,1655],{"name":1652,"slug":1653,"type":15},"Code Review","code-review",{"name":1577,"slug":1578,"type":15},{"name":1656,"slug":1657,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1660,"name":1660,"fn":1661,"description":1662,"org":1663,"tags":1664,"stars":1554,"repoUrl":1555,"updatedAt":1667},"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},[1665,1666],{"name":1628,"slug":1629,"type":15},{"name":1631,"slug":1632,"type":15},"2026-07-14T05:25:54.928983",{"slug":1669,"name":1669,"fn":1670,"description":1671,"org":1672,"tags":1673,"stars":1554,"repoUrl":1555,"updatedAt":1678},"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},[1674,1677],{"name":1675,"slug":1676,"type":15},"Automation","automation",{"name":1564,"slug":1565,"type":15},"2026-07-30T05:29:03.275638",496,{"items":1681,"total":1775},[1682,1697,1707,1721,1731,1746,1761],{"slug":1683,"name":1683,"fn":1684,"description":1685,"org":1686,"tags":1687,"stars":23,"repoUrl":24,"updatedAt":1696},"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},[1688,1691,1694,1695],{"name":1689,"slug":1690,"type":15},"Data Analysis","data-analysis",{"name":1692,"slug":1693,"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":1698,"name":1698,"fn":1699,"description":1700,"org":1701,"tags":1702,"stars":23,"repoUrl":24,"updatedAt":1706},"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},[1703,1704,1705],{"name":1549,"slug":1550,"type":15},{"name":1642,"slug":1643,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1708,"name":1708,"fn":1709,"description":1710,"org":1711,"tags":1712,"stars":23,"repoUrl":24,"updatedAt":1720},"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},[1713,1716,1717],{"name":1714,"slug":1715,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1718,"slug":1719,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1722,"name":1722,"fn":1723,"description":1724,"org":1725,"tags":1726,"stars":23,"repoUrl":24,"updatedAt":1730},"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},[1727,1728,1729],{"name":1689,"slug":1690,"type":15},{"name":9,"slug":8,"type":15},{"name":1631,"slug":1632,"type":15},"2026-07-17T05:29:03.913266",{"slug":1732,"name":1732,"fn":1733,"description":1734,"org":1735,"tags":1736,"stars":23,"repoUrl":24,"updatedAt":1745},"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},[1737,1738,1741,1742],{"name":1675,"slug":1676,"type":15},{"name":1739,"slug":1740,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":1743,"slug":1744,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":1747,"name":1747,"fn":1748,"description":1749,"org":1750,"tags":1751,"stars":23,"repoUrl":24,"updatedAt":1760},"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},[1752,1753,1756,1757],{"name":1549,"slug":1550,"type":15},{"name":1754,"slug":1755,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1758,"slug":1759,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1762,"name":1762,"fn":1763,"description":1764,"org":1765,"tags":1766,"stars":23,"repoUrl":24,"updatedAt":1774},"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},[1767,1768,1771],{"name":9,"slug":8,"type":15},{"name":1769,"slug":1770,"type":15},"Quantum Computing","quantum-computing",{"name":1772,"slug":1773,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]