[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-doca-gpunetio":3,"mdc-ikifon-key":31,"related-repo-nvidia-doca-gpunetio":1380,"related-org-nvidia-doca-gpunetio":1486},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":20,"repoUrl":21,"updatedAt":22,"license":23,"forks":24,"topics":25,"repo":26,"sourceUrl":29,"mdContent":30},"doca-gpunetio","program NVIDIA GPUNetIO for CUDA kernels","Use this skill when the user is doing hands-on DOCA GPUNetIO programming — wiring a CUDA kernel on an NVIDIA GPU to a doca-eth queue via doca_gpu_eth_rxq \u002F doca_gpu_eth_txq, standing up the per-CUDA-device doca_gpu context, designing the persistent CUDA kernel that drains the GPU-visible queue, running the dual capability check (DOCA cap-query plus cudaGetDeviceProperties), registering cudaMalloc pools via doca_buf_arr_create_*, or debugging DOCA_ERROR_* returns from the GPUNetIO API. Trigger even when the user does not explicitly mention \"DOCA GPUNetIO\" or \"persistent kernel\" — typical implicit phrasings include \"CUDA kernel reading packets directly from the NIC\", \"GPU-initiated networking on BlueField\", \"DOCA_ERROR_DRIVER on doca_gpu_create\", \"nvidia_peermem not loaded\", \"kernel-per-packet is too slow\", or \"which GPU supports GPU-side packet I\u002FO\". Refuse and route elsewhere for general CUDA programming, DOCA Ethernet queue bring-up, DOCA DPA, or DOCA install — 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",[12,14,17],{"name":9,"slug":8,"type":13},"tag",{"name":15,"slug":16,"type":13},"Engineering","engineering",{"name":18,"slug":19,"type":13},"CUDA","cuda",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-30T05:28:26.417995","Apache-2.0",281,[],{"repoUrl":21,"stars":20,"forks":24,"topics":27,"description":28},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fdoca-gpunetio","---\nlicense: Apache-2.0\nname: doca-gpunetio\ndescription: >\n  Use this skill when the user is doing hands-on DOCA GPUNetIO\n  programming — wiring a CUDA kernel on an NVIDIA GPU to a doca-eth\n  queue via doca_gpu_eth_rxq \u002F doca_gpu_eth_txq, standing up the\n  per-CUDA-device doca_gpu context, designing the persistent CUDA\n  kernel that drains the GPU-visible queue, running the dual\n  capability check (DOCA cap-query plus cudaGetDeviceProperties),\n  registering cudaMalloc pools via doca_buf_arr_create_*, or\n  debugging DOCA_ERROR_* returns from the GPUNetIO API. Trigger\n  even when the user does not explicitly mention \"DOCA GPUNetIO\"\n  or \"persistent kernel\" — typical implicit phrasings include\n  \"CUDA kernel reading packets directly from the NIC\",\n  \"GPU-initiated networking on BlueField\", \"DOCA_ERROR_DRIVER on\n  doca_gpu_create\", \"nvidia_peermem not loaded\",\n  \"kernel-per-packet is too slow\", or \"which GPU supports GPU-side\n  packet I\u002FO\". Refuse and route elsewhere for general CUDA\n  programming, DOCA Ethernet queue bring-up, DOCA DPA, or\n  DOCA install — those belong to other skills.\nmetadata:\n  kind: library\ncompatibility: >\n  Requires DOCA SDK at \u002Fopt\u002Fmellanox\u002Fdoca on Linux (Ubuntu 22.04\u002F24.04 or\n  RHEL\u002FSLES) with a BlueField DPU or ConnectX NIC. Reads the local install\n  via `pkg-config doca-gpunetio`. Requires an NVIDIA GPU with CUDA toolkit\n  (matched to DOCA per the DOCA Compatibility Policy) and the\n  nvidia_peermem kernel module loaded for GPUDirect RDMA; some samples\n  need an InfiniBand-capable RNIC.\n\n---\n\n# DOCA GPUNetIO\n\n**Where to start:** This skill assumes DOCA is already installed,\nthe CUDA toolkit is installed and matched to the DOCA install, and\nthe user is doing **hands-on GPUNetIO work** — i.e. wiring a DOCA\nnetwork queue into a CUDA kernel on an NVIDIA GPU. Open\n[`TASKS.md`](TASKS.md) if the user wants to *do* something\n(configure \u002F build \u002F modify \u002F run \u002F test \u002F debug); open\n[`CAPABILITIES.md`](CAPABILITIES.md) when the question is *what\ncan GPUNetIO express* on this version + this GPU. If the user has\nnot installed DOCA yet, route to\n[`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) first; if the user has\nnot set up the underlying Ethernet RX\u002FTX queues yet, that is a\nDOCA Ethernet question — route to\n[`doca-eth`](..\u002Fdoca-eth\u002FSKILL.md).\n\n## Example questions this skill answers well\n\nThe CLASSES of GPUNetIO questions this skill is built to answer,\neach with one worked example. The agent should treat the *class*\nas the load-bearing piece — the worked example is a single\ninstance.\n\n- **\"How do I get a CUDA kernel to receive packets directly from\n  the NIC?\"** — worked example: *\"persistent kernel on one GPU\n  reads packets from a `doca_gpu_eth_rxq` built on top of a\n  representor `doca_eth_rxq` and counts them per-flow\"*. Answered\n  by the persistent-kernel pattern in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the GPU-side bring-up workflow in\n  [`TASKS.md ## configure`](TASKS.md#configure).\n- **\"Can I run GPUNetIO on this GPU?\"** — worked example: *\"my\n  host has one Ampere card and one Turing card; which one\n  supports GPU-initiated networking?\"*. Answered by the dual\n  capability-discovery rule (DOCA cap-query AND\n  `cudaGetDeviceProperties` against the CUDA device ordinal) in\n  [`CAPABILITIES.md ## Capabilities and modes`](CAPABILITIES.md#capabilities-and-modes)\n  + the device-enumeration step in\n  [`TASKS.md ## configure`](TASKS.md#configure).\n- **\"Why does my GPUNetIO setup fail with\n  `DOCA_ERROR_NOT_SUPPORTED` even though doca-eth came up\n  fine?\"** — worked example: *\"`nvidia_peermem` is not loaded so\n  GPUDirect RDMA is unavailable\"*. Answered by the env preconditions\n  in [`CAPABILITIES.md ## Safety policy`](CAPABILITIES.md#safety-policy)\n  + the env checklist in\n  [`TASKS.md ## configure`](TASKS.md#configure) step 1.\n- **\"How do I move data between CUDA-allocated buffers and a DOCA\n  queue?\"** — worked example: *\"use `cudaMalloc` for the receive\n  buffer pool and register it with DOCA via `doca_buf_arr_create_*`\n  before starting the context\"*. Answered by the CUDA-allocator\n  + DOCA-registration overlay in\n  [`CAPABILITIES.md ## Safety policy`](CAPABILITIES.md#safety-policy)\n  + the buffer-prep step in\n  [`TASKS.md ## configure`](TASKS.md#configure) step 4.\n- **\"Is the GPUNetIO API I'm reading about on my installed DOCA +\n  CUDA combination?\"** — worked example: *\"is the persistent-kernel\n  helper available with the CUDA toolkit version I have?\"*.\n  Answered by the version-compatibility overlay in\n  [`CAPABILITIES.md ## Version compatibility`](CAPABILITIES.md#version-compatibility)\n  which cross-links the canonical detection chain in\n  [`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) and adds the\n  GPUNetIO-specific *DOCA must match CUDA* overlay.\n- **\"What does this `DOCA_ERROR_*` from a GPUNetIO call mean and\n  which layer caused it?\"** — worked example: *\"`DOCA_ERROR_DRIVER`\n  on `doca_gpu_*_create` — is it DOCA, CUDA, or the underlying\n  doca-eth queue?\"*. Answered by the GPUNetIO overlay on the\n  cross-library 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 to\n  [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md).\n\n## Audience\n\nThis skill serves **external developers building applications\nthat consume the DOCA GPUNetIO library** — i.e., users whose code\ncalls `doca_gpu_*` from host C\u002FC++ to stand up the per-GPU\ncontext and the GPU-visible queue handles, and whose CUDA kernel\n(`.cu` translation unit) uses those handles from device code to\nsubmit \u002F receive packets. The canonical target shape is the GPU\nPacket Processing reference application: a CUDA persistent\nkernel on an NVIDIA GPU that polls a GPU-visible RX queue and\nprocesses packets in-place on the GPU. It is *not* for NVIDIA\ndevelopers contributing to DOCA GPUNetIO itself.\n\n**Language scope.** DOCA GPUNetIO ships as a C \u002F CUDA library\nwith `pkg-config` module name `doca-gpunetio`. The host-side API\nis C; the device-side API is CUDA C++ used inside a `.cu`\nkernel. The shipped samples and the GPU Packet Processing\nreference application are written in C + CUDA C++ (NVIDIA's\nchoice). Other-language consumers are limited in practice — the\ndevice-side API has no FFI escape hatch because the kernel must\nbe a CUDA translation unit — but a Rust \u002F Go \u002F Python host-side\nwrapper that drives the host-side `doca_gpu_*` setup and\nlaunches a CUDA kernel built separately is still useful, and the\nskill keeps the lifecycle, capability-discovery, env-precondition,\nand error-taxonomy guidance language-neutral.\n\n## When to load this skill\n\nLoad this skill when the user is doing hands-on DOCA GPUNetIO\nwork, in any host language plus CUDA. Concretely:\n\n- Initializing a `doca_gpu` against a specific CUDA device\n  ordinal on a host with one or more NVIDIA GPUs.\n- Creating a GPU-visible queue handle (`doca_gpu_eth_rxq`,\n  `doca_gpu_eth_txq`) on top of an existing `doca_eth_rxq` \u002F\n  `doca_eth_txq` from DOCA Ethernet, and passing the handle into\n  a CUDA kernel for device-side use.\n- Writing or modifying the persistent CUDA kernel that drains\n  the GPU-visible RX queue in a long-running loop (the canonical\n  GPU Packet Processing shape).\n- Allocating GPU buffers via `cudaMalloc` and registering them\n  with DOCA via the `doca_buf_arr_create_*` family before\n  `doca_ctx_start()`.\n- Checking which GPUNetIO features are supported on the active\n  `doca_devinfo` (DOCA cap-query family) AND on the candidate\n  CUDA device (`cudaGetDeviceProperties` and CUDA-driver-version\n  checks).\n- Debugging a `DOCA_ERROR_*` returned from a GPUNetIO call — in\n  particular disambiguating *DOCA capability missing* from *CUDA\n  device too old* from *`nvidia_peermem` not loaded* from *CUDA\n  driver + DOCA version skew*.\n- Designing host-side bindings for non-C languages that drive a\n  CUDA kernel they built separately — the env-precondition and\n  capability-discovery rules in this skill still apply.\n\nDo **not** load this skill for general DOCA orientation, install\nof DOCA or the CUDA toolkit, the underlying DOCA Ethernet queue\nsetup, or non-GPUNetIO library questions. For those, route\nthrough [`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md)\nto the matching upstream guide.\n\n## What this skill provides\n\nThis is a **thin loader**. The body keeps only the orientation\nneeded to pick the right next file. The substantive\nGPUNetIO-specific material lives in two companion files:\n\n- `CAPABILITIES.md` — what GPUNetIO can express on this version\n  + this GPU: the `doca_gpu` per-device context, the GPU-visible\n  RX \u002F TX queue handles layered on doca-eth, the persistent\n  CUDA-kernel pattern as the default usage shape, the\n  capability-query surface (the doca-eth\n  `doca_eth_rxq_cap_is_type_supported` \u002F `doca_eth_rxq_cap_get_*`\n  family in `doca_eth_rxq.h`, plus the matching\n  `doca_eth_txq_cap_*` family, on the DOCA side, plus\n  `cudaGetDeviceProperties` on the CUDA side), the GPUNetIO error taxonomy mapped onto the cross-library\n  `DOCA_ERROR_*` set, the observability surface (CUDA-side\n  counters + DOCA-side per-task completion), and the safety\n  policy that gates env preconditions (CUDA + DOCA version\n  match, `nvidia_peermem`, CUDA buffer registration).\n- `TASKS.md` — step-by-step workflows for the six in-scope\n  GPUNetIO verbs: `configure`, `build`, `modify`, `run`, `test`,\n  `debug`. Plus a `## rollback` overlay (GPUNetIO-specific\n  five-step teardown that signals the persistent kernel to\n  drain, unregisters GPU buffers in reverse-register order, and\n  leaves the parent doca-eth queue intact) and the 5-phase\n  universal debug-loop instantiation appended to `## debug`.\n  Plus a `Deferred task verbs` block that points out-of-scope\n  questions at the right next skill.\n\nThe skill assumes a host where DOCA is already installed at the\nstandard location, an NVIDIA GPU is physically present, the CUDA\ntoolkit is installed and its version is matched to the DOCA\ninstall per the DOCA Compatibility Policy, and the underlying\nDOCA Ethernet RX\u002FTX queues are *already* configured (this skill\nsits on top of doca-eth, not below it). It does not cover\ninstalling DOCA or the CUDA toolkit — that path goes through\n[`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md).\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- **Pre-written DOCA GPUNetIO application source code or CUDA\n  kernel source, in any language.** The verified GPUNetIO\n  source is the shipped C + CUDA samples at\n  `\u002Fopt\u002Fmellanox\u002Fdoca\u002Fsamples\u002Fdoca_gpunetio\u002F` and the GPU Packet\n  Processing reference application. The agent's job is to route\n  the user to those files and prescribe a minimum-diff\n  modification on them via the universal modify-a-sample\n  workflow in\n  [`doca-programming-guide`](..\u002F..\u002Fdoca-programming-guide\u002FSKILL.md),\n  layered with the GPUNetIO-specific overrides in\n  [`TASKS.md ## modify`](TASKS.md#modify).\n- **Standalone build manifests** (`meson.build`,\n  `CMakeLists.txt`, …) parked inside the skill. The agent\n  constructs the build manifest *in the user's project\n  directory* against the user's installed DOCA + CUDA toolkit,\n  where `pkg-config --modversion doca-gpunetio`,\n  `pkg-config --modversion doca-common`, and `nvcc --version`\n  form the version gate.\n- **A `samples\u002F`, `bindings\u002F`, or `reference\u002F` subtree** of any\n  kind. A mock or incomplete artifact in this skill's tree, even\n  one labeled \"reference\", is misleading: users will read it as\n  buildable.\n\n## Loading order\n\n1. Read this `SKILL.md` first to confirm the user's question is\n   in scope.\n2. **For the GPUNetIO capability matrix, the `doca_gpu` per-device\n   context, the persistent-kernel pattern, the dual capability\n   query, the env-precondition policy, the error taxonomy, the\n   observability surface, and the safety policy, see\n   [CAPABILITIES.md](CAPABILITIES.md).**\n3. **For step-by-step workflows — configure, build, modify, run,\n   test, debug — see [TASKS.md](TASKS.md).**\n\nBoth companion files cross-link to each other,\n[`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) for the canonical\nDOCA version-handling rules (with the GPUNetIO overlay that DOCA\nmust match CUDA), 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\nGPUNetIO guide, the DOCA Compatibility Policy, the CUDA toolkit\ndocs, or the on-disk install layout\" rather than\n\"GPUNetIO-specific guidance\".\n\n## Related skills\n\n- [`doca-public-knowledge-map`](..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md) —\n  the routing table for every public DOCA documentation source\n  and the on-disk layout of an installed DOCA package. The\n  GPUNetIO public guide is at\n  \u003Chttps:\u002F\u002Fdocs.nvidia.com\u002Fdoca\u002Fsdk\u002FDOCA-GPUNetIO\u002Findex.html>;\n  the GPU Packet Processing reference application is reachable\n  from there. The CUDA toolkit and DOCA Compatibility Policy\n  links live in the same routing table.\n- [`doca-setup`](..\u002F..\u002Fdoca-setup\u002FSKILL.md) — env preparation,\n  install verification, CUDA toolkit install \u002F verification, and\n  the *I have no install yet* path with the public NGC DOCA\n  container. This skill assumes its preconditions are satisfied\n  AND that CUDA is installed at a version that matches DOCA.\n- [`doca-version`](..\u002F..\u002Fdoca-version\u002FSKILL.md) — canonical DOCA\n  version-handling rules. This skill's `## Version\n  compatibility` cross-links the four-way match rule and adds\n  the GPUNetIO-specific *DOCA-and-CUDA must match* overlay per\n  the DOCA Compatibility Policy.\n- [`doca-structured-tools-contract`](..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md) —\n  the bundle's structured-tools precedence rule (detect \u002F prefer\n  \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 shared by every library:\n  the canonical `pkg-config` + meson build pattern, the\n  universal modify-a-shipped-sample first-app workflow, the\n  universal lifecycle, the cross-library `DOCA_ERROR_*`\n  taxonomy, and the program-side debug order. This skill\n  layers GPUNetIO specifics on top.\n- [`doca-debug`](..\u002F..\u002Fdoca-debug\u002FSKILL.md) — the cross-cutting\n  debug ladder (install \u002F version \u002F build \u002F link \u002F runtime \u002F\n  program \u002F driver). GPUNetIO-specific debug (CUDA + DOCA\n  version skew, `nvidia_peermem` missing, persistent-kernel\n  silent hangs, CUDA-allocator + DOCA-registration mismatches)\n  overlays on top of that ladder.\n\nDOCA Ethernet is GPUNetIO's mandatory companion library: GPU-visible\nRX \u002F TX queue handles are layered on top of `doca_eth_rxq` \u002F\n`doca_eth_txq` from DOCA Ethernet. For the underlying queue setup,\nroute to [`doca-eth`](..\u002Fdoca-eth\u002FSKILL.md).\n\n",{"data":32,"body":36},{"license":23,"name":4,"description":6,"metadata":33,"compatibility":35},{"kind":34},"library","Requires DOCA SDK at \u002Fopt\u002Fmellanox\u002Fdoca on Linux (Ubuntu 22.04\u002F24.04 or RHEL\u002FSLES) with a BlueField DPU or ConnectX NIC. Reads the local install via `pkg-config doca-gpunetio`. Requires an NVIDIA GPU with CUDA toolkit (matched to DOCA per the DOCA Compatibility Policy) and the nvidia_peermem kernel module loaded for GPUDirect RDMA; some samples need an InfiniBand-capable RNIC.\n",{"type":37,"children":38},"root",[39,47,129,136,148,514,520,555,594,600,605,754,777,783,795,954,975,981,993,1125,1131,1181,1206,1212,1353],{"type":40,"tag":41,"props":42,"children":43},"element","h1",{"id":4},[44],{"type":45,"value":46},"text","DOCA GPUNetIO",{"type":40,"tag":48,"props":49,"children":50},"p",{},[51,57,59,64,66,77,79,85,87,96,98,103,105,115,117,127],{"type":40,"tag":52,"props":53,"children":54},"strong",{},[55],{"type":45,"value":56},"Where to start:",{"type":45,"value":58}," This skill assumes DOCA is already installed,\nthe CUDA toolkit is installed and matched to the DOCA install, and\nthe user is doing ",{"type":40,"tag":52,"props":60,"children":61},{},[62],{"type":45,"value":63},"hands-on GPUNetIO work",{"type":45,"value":65}," — i.e. wiring a DOCA\nnetwork queue into a CUDA kernel on an NVIDIA GPU. Open\n",{"type":40,"tag":67,"props":68,"children":70},"a",{"href":69},"TASKS.md",[71],{"type":40,"tag":72,"props":73,"children":75},"code",{"className":74},[],[76],{"type":45,"value":69},{"type":45,"value":78}," if the user wants to ",{"type":40,"tag":80,"props":81,"children":82},"em",{},[83],{"type":45,"value":84},"do",{"type":45,"value":86}," something\n(configure \u002F build \u002F modify \u002F run \u002F test \u002F debug); open\n",{"type":40,"tag":67,"props":88,"children":90},{"href":89},"CAPABILITIES.md",[91],{"type":40,"tag":72,"props":92,"children":94},{"className":93},[],[95],{"type":45,"value":89},{"type":45,"value":97}," when the question is ",{"type":40,"tag":80,"props":99,"children":100},{},[101],{"type":45,"value":102},"what\ncan GPUNetIO express",{"type":45,"value":104}," on this version + this GPU. If the user has\nnot installed DOCA yet, route to\n",{"type":40,"tag":67,"props":106,"children":108},{"href":107},"..\u002F..\u002Fdoca-setup\u002FSKILL.md",[109],{"type":40,"tag":72,"props":110,"children":112},{"className":111},[],[113],{"type":45,"value":114},"doca-setup",{"type":45,"value":116}," first; if the user has\nnot set up the underlying Ethernet RX\u002FTX queues yet, that is a\nDOCA Ethernet question — route to\n",{"type":40,"tag":67,"props":118,"children":120},{"href":119},"..\u002Fdoca-eth\u002FSKILL.md",[121],{"type":40,"tag":72,"props":122,"children":124},{"className":123},[],[125],{"type":45,"value":126},"doca-eth",{"type":45,"value":128},".",{"type":40,"tag":130,"props":131,"children":133},"h2",{"id":132},"example-questions-this-skill-answers-well",[134],{"type":45,"value":135},"Example questions this skill answers well",{"type":40,"tag":48,"props":137,"children":138},{},[139,141,146],{"type":45,"value":140},"The CLASSES of GPUNetIO questions this skill is built to answer,\neach with one worked example. The agent should treat the ",{"type":40,"tag":80,"props":142,"children":143},{},[144],{"type":45,"value":145},"class",{"type":45,"value":147},"\nas the load-bearing piece — the worked example is a single\ninstance.",{"type":40,"tag":149,"props":150,"children":151},"ul",{},[152,215,264,324,387,434],{"type":40,"tag":153,"props":154,"children":155},"li",{},[156,161,163,184,186,196],{"type":40,"tag":52,"props":157,"children":158},{},[159],{"type":45,"value":160},"\"How do I get a CUDA kernel to receive packets directly from\nthe NIC?\"",{"type":45,"value":162}," — worked example: ",{"type":40,"tag":80,"props":164,"children":165},{},[166,168,174,176,182],{"type":45,"value":167},"\"persistent kernel on one GPU\nreads packets from a ",{"type":40,"tag":72,"props":169,"children":171},{"className":170},[],[172],{"type":45,"value":173},"doca_gpu_eth_rxq",{"type":45,"value":175}," built on top of a\nrepresentor ",{"type":40,"tag":72,"props":177,"children":179},{"className":178},[],[180],{"type":45,"value":181},"doca_eth_rxq",{"type":45,"value":183}," and counts them per-flow\"",{"type":45,"value":185},". Answered\nby the persistent-kernel pattern in\n",{"type":40,"tag":67,"props":187,"children":189},{"href":188},"CAPABILITIES.md#capabilities-and-modes",[190],{"type":40,"tag":72,"props":191,"children":193},{"className":192},[],[194],{"type":45,"value":195},"CAPABILITIES.md ## Capabilities and modes",{"type":40,"tag":149,"props":197,"children":198},{},[199],{"type":40,"tag":153,"props":200,"children":201},{},[202,204,214],{"type":45,"value":203},"the GPU-side bring-up workflow in\n",{"type":40,"tag":67,"props":205,"children":207},{"href":206},"TASKS.md#configure",[208],{"type":40,"tag":72,"props":209,"children":211},{"className":210},[],[212],{"type":45,"value":213},"TASKS.md ## configure",{"type":45,"value":128},{"type":40,"tag":153,"props":216,"children":217},{},[218,223,224,229,231,237,239,247],{"type":40,"tag":52,"props":219,"children":220},{},[221],{"type":45,"value":222},"\"Can I run GPUNetIO on this GPU?\"",{"type":45,"value":162},{"type":40,"tag":80,"props":225,"children":226},{},[227],{"type":45,"value":228},"\"my\nhost has one Ampere card and one Turing card; which one\nsupports GPU-initiated networking?\"",{"type":45,"value":230},". Answered by the dual\ncapability-discovery rule (DOCA cap-query AND\n",{"type":40,"tag":72,"props":232,"children":234},{"className":233},[],[235],{"type":45,"value":236},"cudaGetDeviceProperties",{"type":45,"value":238}," against the CUDA device ordinal) in\n",{"type":40,"tag":67,"props":240,"children":241},{"href":188},[242],{"type":40,"tag":72,"props":243,"children":245},{"className":244},[],[246],{"type":45,"value":195},{"type":40,"tag":149,"props":248,"children":249},{},[250],{"type":40,"tag":153,"props":251,"children":252},{},[253,255,263],{"type":45,"value":254},"the device-enumeration step in\n",{"type":40,"tag":67,"props":256,"children":257},{"href":206},[258],{"type":40,"tag":72,"props":259,"children":261},{"className":260},[],[262],{"type":45,"value":213},{"type":45,"value":128},{"type":40,"tag":153,"props":265,"children":266},{},[267,280,281,294,296,306],{"type":40,"tag":52,"props":268,"children":269},{},[270,272,278],{"type":45,"value":271},"\"Why does my GPUNetIO setup fail with\n",{"type":40,"tag":72,"props":273,"children":275},{"className":274},[],[276],{"type":45,"value":277},"DOCA_ERROR_NOT_SUPPORTED",{"type":45,"value":279}," even though doca-eth came up\nfine?\"",{"type":45,"value":162},{"type":40,"tag":80,"props":282,"children":283},{},[284,286,292],{"type":45,"value":285},"\"",{"type":40,"tag":72,"props":287,"children":289},{"className":288},[],[290],{"type":45,"value":291},"nvidia_peermem",{"type":45,"value":293}," is not loaded so\nGPUDirect RDMA is unavailable\"",{"type":45,"value":295},". Answered by the env preconditions\nin ",{"type":40,"tag":67,"props":297,"children":299},{"href":298},"CAPABILITIES.md#safety-policy",[300],{"type":40,"tag":72,"props":301,"children":303},{"className":302},[],[304],{"type":45,"value":305},"CAPABILITIES.md ## Safety policy",{"type":40,"tag":149,"props":307,"children":308},{},[309],{"type":40,"tag":153,"props":310,"children":311},{},[312,314,322],{"type":45,"value":313},"the env checklist in\n",{"type":40,"tag":67,"props":315,"children":316},{"href":206},[317],{"type":40,"tag":72,"props":318,"children":320},{"className":319},[],[321],{"type":45,"value":213},{"type":45,"value":323}," step 1.",{"type":40,"tag":153,"props":325,"children":326},{},[327,332,333,354,356],{"type":40,"tag":52,"props":328,"children":329},{},[330],{"type":45,"value":331},"\"How do I move data between CUDA-allocated buffers and a DOCA\nqueue?\"",{"type":45,"value":162},{"type":40,"tag":80,"props":334,"children":335},{},[336,338,344,346,352],{"type":45,"value":337},"\"use ",{"type":40,"tag":72,"props":339,"children":341},{"className":340},[],[342],{"type":45,"value":343},"cudaMalloc",{"type":45,"value":345}," for the receive\nbuffer pool and register it with DOCA via ",{"type":40,"tag":72,"props":347,"children":349},{"className":348},[],[350],{"type":45,"value":351},"doca_buf_arr_create_*",{"type":45,"value":353},"\nbefore starting the context\"",{"type":45,"value":355},". Answered by the CUDA-allocator\n",{"type":40,"tag":149,"props":357,"children":358},{},[359,372],{"type":40,"tag":153,"props":360,"children":361},{},[362,364],{"type":45,"value":363},"DOCA-registration overlay in\n",{"type":40,"tag":67,"props":365,"children":366},{"href":298},[367],{"type":40,"tag":72,"props":368,"children":370},{"className":369},[],[371],{"type":45,"value":305},{"type":40,"tag":153,"props":373,"children":374},{},[375,377,385],{"type":45,"value":376},"the buffer-prep step in\n",{"type":40,"tag":67,"props":378,"children":379},{"href":206},[380],{"type":40,"tag":72,"props":381,"children":383},{"className":382},[],[384],{"type":45,"value":213},{"type":45,"value":386}," step 4.",{"type":40,"tag":153,"props":388,"children":389},{},[390,395,396,401,403,413,415,425,427,432],{"type":40,"tag":52,"props":391,"children":392},{},[393],{"type":45,"value":394},"\"Is the GPUNetIO API I'm reading about on my installed DOCA +\nCUDA combination?\"",{"type":45,"value":162},{"type":40,"tag":80,"props":397,"children":398},{},[399],{"type":45,"value":400},"\"is the persistent-kernel\nhelper available with the CUDA toolkit version I have?\"",{"type":45,"value":402},".\nAnswered by the version-compatibility overlay in\n",{"type":40,"tag":67,"props":404,"children":406},{"href":405},"CAPABILITIES.md#version-compatibility",[407],{"type":40,"tag":72,"props":408,"children":410},{"className":409},[],[411],{"type":45,"value":412},"CAPABILITIES.md ## Version compatibility",{"type":45,"value":414},"\nwhich cross-links the canonical detection chain in\n",{"type":40,"tag":67,"props":416,"children":418},{"href":417},"..\u002F..\u002Fdoca-version\u002FSKILL.md",[419],{"type":40,"tag":72,"props":420,"children":422},{"className":421},[],[423],{"type":45,"value":424},"doca-version",{"type":45,"value":426}," and adds the\nGPUNetIO-specific ",{"type":40,"tag":80,"props":428,"children":429},{},[430],{"type":45,"value":431},"DOCA must match CUDA",{"type":45,"value":433}," overlay.",{"type":40,"tag":153,"props":435,"children":436},{},[437,450,451,471,473,483],{"type":40,"tag":52,"props":438,"children":439},{},[440,442,448],{"type":45,"value":441},"\"What does this ",{"type":40,"tag":72,"props":443,"children":445},{"className":444},[],[446],{"type":45,"value":447},"DOCA_ERROR_*",{"type":45,"value":449}," from a GPUNetIO call mean and\nwhich layer caused it?\"",{"type":45,"value":162},{"type":40,"tag":80,"props":452,"children":453},{},[454,455,461,463,469],{"type":45,"value":285},{"type":40,"tag":72,"props":456,"children":458},{"className":457},[],[459],{"type":45,"value":460},"DOCA_ERROR_DRIVER",{"type":45,"value":462},"\non ",{"type":40,"tag":72,"props":464,"children":466},{"className":465},[],[467],{"type":45,"value":468},"doca_gpu_*_create",{"type":45,"value":470}," — is it DOCA, CUDA, or the underlying\ndoca-eth queue?\"",{"type":45,"value":472},". Answered by the GPUNetIO overlay on the\ncross-library taxonomy in\n",{"type":40,"tag":67,"props":474,"children":476},{"href":475},"CAPABILITIES.md#error-taxonomy",[477],{"type":40,"tag":72,"props":478,"children":480},{"className":479},[],[481],{"type":45,"value":482},"CAPABILITIES.md ## Error taxonomy",{"type":40,"tag":149,"props":484,"children":485},{},[486],{"type":40,"tag":153,"props":487,"children":488},{},[489,491,501,503,513],{"type":45,"value":490},"the layered ladder in\n",{"type":40,"tag":67,"props":492,"children":494},{"href":493},"TASKS.md#debug",[495],{"type":40,"tag":72,"props":496,"children":498},{"className":497},[],[499],{"type":45,"value":500},"TASKS.md ## debug",{"type":45,"value":502}," that escalates to\n",{"type":40,"tag":67,"props":504,"children":506},{"href":505},"..\u002F..\u002Fdoca-debug\u002FSKILL.md",[507],{"type":40,"tag":72,"props":508,"children":510},{"className":509},[],[511],{"type":45,"value":512},"doca-debug",{"type":45,"value":128},{"type":40,"tag":130,"props":515,"children":517},{"id":516},"audience",[518],{"type":45,"value":519},"Audience",{"type":40,"tag":48,"props":521,"children":522},{},[523,525,530,532,538,540,546,548,553],{"type":45,"value":524},"This skill serves ",{"type":40,"tag":52,"props":526,"children":527},{},[528],{"type":45,"value":529},"external developers building applications\nthat consume the DOCA GPUNetIO library",{"type":45,"value":531}," — i.e., users whose code\ncalls ",{"type":40,"tag":72,"props":533,"children":535},{"className":534},[],[536],{"type":45,"value":537},"doca_gpu_*",{"type":45,"value":539}," from host C\u002FC++ to stand up the per-GPU\ncontext and the GPU-visible queue handles, and whose CUDA kernel\n(",{"type":40,"tag":72,"props":541,"children":543},{"className":542},[],[544],{"type":45,"value":545},".cu",{"type":45,"value":547}," translation unit) uses those handles from device code to\nsubmit \u002F receive packets. The canonical target shape is the GPU\nPacket Processing reference application: a CUDA persistent\nkernel on an NVIDIA GPU that polls a GPU-visible RX queue and\nprocesses packets in-place on the GPU. It is ",{"type":40,"tag":80,"props":549,"children":550},{},[551],{"type":45,"value":552},"not",{"type":45,"value":554}," for NVIDIA\ndevelopers contributing to DOCA GPUNetIO itself.",{"type":40,"tag":48,"props":556,"children":557},{},[558,563,565,571,573,578,580,585,587,592],{"type":40,"tag":52,"props":559,"children":560},{},[561],{"type":45,"value":562},"Language scope.",{"type":45,"value":564}," DOCA GPUNetIO ships as a C \u002F CUDA library\nwith ",{"type":40,"tag":72,"props":566,"children":568},{"className":567},[],[569],{"type":45,"value":570},"pkg-config",{"type":45,"value":572}," module name ",{"type":40,"tag":72,"props":574,"children":576},{"className":575},[],[577],{"type":45,"value":4},{"type":45,"value":579},". The host-side API\nis C; the device-side API is CUDA C++ used inside a ",{"type":40,"tag":72,"props":581,"children":583},{"className":582},[],[584],{"type":45,"value":545},{"type":45,"value":586},"\nkernel. The shipped samples and the GPU Packet Processing\nreference application are written in C + CUDA C++ (NVIDIA's\nchoice). Other-language consumers are limited in practice — the\ndevice-side API has no FFI escape hatch because the kernel must\nbe a CUDA translation unit — but a Rust \u002F Go \u002F Python host-side\nwrapper that drives the host-side ",{"type":40,"tag":72,"props":588,"children":590},{"className":589},[],[591],{"type":45,"value":537},{"type":45,"value":593}," setup and\nlaunches a CUDA kernel built separately is still useful, and the\nskill keeps the lifecycle, capability-discovery, env-precondition,\nand error-taxonomy guidance language-neutral.",{"type":40,"tag":130,"props":595,"children":597},{"id":596},"when-to-load-this-skill",[598],{"type":45,"value":599},"When to load this skill",{"type":40,"tag":48,"props":601,"children":602},{},[603],{"type":45,"value":604},"Load this skill when the user is doing hands-on DOCA GPUNetIO\nwork, in any host language plus CUDA. Concretely:",{"type":40,"tag":149,"props":606,"children":607},{},[608,621,656,661,687,707,749],{"type":40,"tag":153,"props":609,"children":610},{},[611,613,619],{"type":45,"value":612},"Initializing a ",{"type":40,"tag":72,"props":614,"children":616},{"className":615},[],[617],{"type":45,"value":618},"doca_gpu",{"type":45,"value":620}," against a specific CUDA device\nordinal on a host with one or more NVIDIA GPUs.",{"type":40,"tag":153,"props":622,"children":623},{},[624,626,631,633,639,641,646,648,654],{"type":45,"value":625},"Creating a GPU-visible queue handle (",{"type":40,"tag":72,"props":627,"children":629},{"className":628},[],[630],{"type":45,"value":173},{"type":45,"value":632},",\n",{"type":40,"tag":72,"props":634,"children":636},{"className":635},[],[637],{"type":45,"value":638},"doca_gpu_eth_txq",{"type":45,"value":640},") on top of an existing ",{"type":40,"tag":72,"props":642,"children":644},{"className":643},[],[645],{"type":45,"value":181},{"type":45,"value":647}," \u002F\n",{"type":40,"tag":72,"props":649,"children":651},{"className":650},[],[652],{"type":45,"value":653},"doca_eth_txq",{"type":45,"value":655}," from DOCA Ethernet, and passing the handle into\na CUDA kernel for device-side use.",{"type":40,"tag":153,"props":657,"children":658},{},[659],{"type":45,"value":660},"Writing or modifying the persistent CUDA kernel that drains\nthe GPU-visible RX queue in a long-running loop (the canonical\nGPU Packet Processing shape).",{"type":40,"tag":153,"props":662,"children":663},{},[664,666,671,673,678,680,686],{"type":45,"value":665},"Allocating GPU buffers via ",{"type":40,"tag":72,"props":667,"children":669},{"className":668},[],[670],{"type":45,"value":343},{"type":45,"value":672}," and registering them\nwith DOCA via the ",{"type":40,"tag":72,"props":674,"children":676},{"className":675},[],[677],{"type":45,"value":351},{"type":45,"value":679}," family before\n",{"type":40,"tag":72,"props":681,"children":683},{"className":682},[],[684],{"type":45,"value":685},"doca_ctx_start()",{"type":45,"value":128},{"type":40,"tag":153,"props":688,"children":689},{},[690,692,698,700,705],{"type":45,"value":691},"Checking which GPUNetIO features are supported on the active\n",{"type":40,"tag":72,"props":693,"children":695},{"className":694},[],[696],{"type":45,"value":697},"doca_devinfo",{"type":45,"value":699}," (DOCA cap-query family) AND on the candidate\nCUDA device (",{"type":40,"tag":72,"props":701,"children":703},{"className":702},[],[704],{"type":45,"value":236},{"type":45,"value":706}," and CUDA-driver-version\nchecks).",{"type":40,"tag":153,"props":708,"children":709},{},[710,712,717,719,724,726,731,732,742,743,748],{"type":45,"value":711},"Debugging a ",{"type":40,"tag":72,"props":713,"children":715},{"className":714},[],[716],{"type":45,"value":447},{"type":45,"value":718}," returned from a GPUNetIO call — in\nparticular disambiguating ",{"type":40,"tag":80,"props":720,"children":721},{},[722],{"type":45,"value":723},"DOCA capability missing",{"type":45,"value":725}," from ",{"type":40,"tag":80,"props":727,"children":728},{},[729],{"type":45,"value":730},"CUDA\ndevice too old",{"type":45,"value":725},{"type":40,"tag":80,"props":733,"children":734},{},[735,740],{"type":40,"tag":72,"props":736,"children":738},{"className":737},[],[739],{"type":45,"value":291},{"type":45,"value":741}," not loaded",{"type":45,"value":725},{"type":40,"tag":80,"props":744,"children":745},{},[746],{"type":45,"value":747},"CUDA\ndriver + DOCA version skew",{"type":45,"value":128},{"type":40,"tag":153,"props":750,"children":751},{},[752],{"type":45,"value":753},"Designing host-side bindings for non-C languages that drive a\nCUDA kernel they built separately — the env-precondition and\ncapability-discovery rules in this skill still apply.",{"type":40,"tag":48,"props":755,"children":756},{},[757,759,763,765,775],{"type":45,"value":758},"Do ",{"type":40,"tag":52,"props":760,"children":761},{},[762],{"type":45,"value":552},{"type":45,"value":764}," load this skill for general DOCA orientation, install\nof DOCA or the CUDA toolkit, the underlying DOCA Ethernet queue\nsetup, or non-GPUNetIO library questions. For those, route\nthrough ",{"type":40,"tag":67,"props":766,"children":768},{"href":767},"..\u002F..\u002Fdoca-public-knowledge-map\u002FSKILL.md",[769],{"type":40,"tag":72,"props":770,"children":772},{"className":771},[],[773],{"type":45,"value":774},"doca-public-knowledge-map",{"type":45,"value":776},"\nto the matching upstream guide.",{"type":40,"tag":130,"props":778,"children":780},{"id":779},"what-this-skill-provides",[781],{"type":45,"value":782},"What this skill provides",{"type":40,"tag":48,"props":784,"children":785},{},[786,788,793],{"type":45,"value":787},"This is a ",{"type":40,"tag":52,"props":789,"children":790},{},[791],{"type":45,"value":792},"thin loader",{"type":45,"value":794},". The body keeps only the orientation\nneeded to pick the right next file. The substantive\nGPUNetIO-specific material lives in two companion files:",{"type":40,"tag":149,"props":796,"children":797},{},[798,876],{"type":40,"tag":153,"props":799,"children":800},{},[801,806,808],{"type":40,"tag":72,"props":802,"children":804},{"className":803},[],[805],{"type":45,"value":89},{"type":45,"value":807}," — what GPUNetIO can express on this version\n",{"type":40,"tag":149,"props":809,"children":810},{},[811],{"type":40,"tag":153,"props":812,"children":813},{},[814,816,821,823,829,831,837,839,845,847,853,855,860,862,867,869,874],{"type":45,"value":815},"this GPU: the ",{"type":40,"tag":72,"props":817,"children":819},{"className":818},[],[820],{"type":45,"value":618},{"type":45,"value":822}," per-device context, the GPU-visible\nRX \u002F TX queue handles layered on doca-eth, the persistent\nCUDA-kernel pattern as the default usage shape, the\ncapability-query surface (the doca-eth\n",{"type":40,"tag":72,"props":824,"children":826},{"className":825},[],[827],{"type":45,"value":828},"doca_eth_rxq_cap_is_type_supported",{"type":45,"value":830}," \u002F ",{"type":40,"tag":72,"props":832,"children":834},{"className":833},[],[835],{"type":45,"value":836},"doca_eth_rxq_cap_get_*",{"type":45,"value":838},"\nfamily in ",{"type":40,"tag":72,"props":840,"children":842},{"className":841},[],[843],{"type":45,"value":844},"doca_eth_rxq.h",{"type":45,"value":846},", plus the matching\n",{"type":40,"tag":72,"props":848,"children":850},{"className":849},[],[851],{"type":45,"value":852},"doca_eth_txq_cap_*",{"type":45,"value":854}," family, on the DOCA side, plus\n",{"type":40,"tag":72,"props":856,"children":858},{"className":857},[],[859],{"type":45,"value":236},{"type":45,"value":861}," on the CUDA side), the GPUNetIO error taxonomy mapped onto the cross-library\n",{"type":40,"tag":72,"props":863,"children":865},{"className":864},[],[866],{"type":45,"value":447},{"type":45,"value":868}," set, the observability surface (CUDA-side\ncounters + DOCA-side per-task completion), and the safety\npolicy that gates env preconditions (CUDA + DOCA version\nmatch, ",{"type":40,"tag":72,"props":870,"children":872},{"className":871},[],[873],{"type":45,"value":291},{"type":45,"value":875},", CUDA buffer registration).",{"type":40,"tag":153,"props":877,"children":878},{},[879,884,886,892,894,900,901,907,908,914,915,921,922,928,930,936,938,944,946,952],{"type":40,"tag":72,"props":880,"children":882},{"className":881},[],[883],{"type":45,"value":69},{"type":45,"value":885}," — step-by-step workflows for the six in-scope\nGPUNetIO verbs: ",{"type":40,"tag":72,"props":887,"children":889},{"className":888},[],[890],{"type":45,"value":891},"configure",{"type":45,"value":893},", ",{"type":40,"tag":72,"props":895,"children":897},{"className":896},[],[898],{"type":45,"value":899},"build",{"type":45,"value":893},{"type":40,"tag":72,"props":902,"children":904},{"className":903},[],[905],{"type":45,"value":906},"modify",{"type":45,"value":893},{"type":40,"tag":72,"props":909,"children":911},{"className":910},[],[912],{"type":45,"value":913},"run",{"type":45,"value":893},{"type":40,"tag":72,"props":916,"children":918},{"className":917},[],[919],{"type":45,"value":920},"test",{"type":45,"value":632},{"type":40,"tag":72,"props":923,"children":925},{"className":924},[],[926],{"type":45,"value":927},"debug",{"type":45,"value":929},". Plus a ",{"type":40,"tag":72,"props":931,"children":933},{"className":932},[],[934],{"type":45,"value":935},"## rollback",{"type":45,"value":937}," overlay (GPUNetIO-specific\nfive-step teardown that signals the persistent kernel to\ndrain, unregisters GPU buffers in reverse-register order, and\nleaves the parent doca-eth queue intact) and the 5-phase\nuniversal debug-loop instantiation appended to ",{"type":40,"tag":72,"props":939,"children":941},{"className":940},[],[942],{"type":45,"value":943},"## debug",{"type":45,"value":945},".\nPlus a ",{"type":40,"tag":72,"props":947,"children":949},{"className":948},[],[950],{"type":45,"value":951},"Deferred task verbs",{"type":45,"value":953}," block that points out-of-scope\nquestions at the right next skill.",{"type":40,"tag":48,"props":955,"children":956},{},[957,959,964,966,974],{"type":45,"value":958},"The skill assumes a host where DOCA is already installed at the\nstandard location, an NVIDIA GPU is physically present, the CUDA\ntoolkit is installed and its version is matched to the DOCA\ninstall per the DOCA Compatibility Policy, and the underlying\nDOCA Ethernet RX\u002FTX queues are ",{"type":40,"tag":80,"props":960,"children":961},{},[962],{"type":45,"value":963},"already",{"type":45,"value":965}," configured (this skill\nsits on top of doca-eth, not below it). It does not cover\ninstalling DOCA or the CUDA toolkit — that path goes through\n",{"type":40,"tag":67,"props":967,"children":968},{"href":107},[969],{"type":40,"tag":72,"props":970,"children":972},{"className":971},[],[973],{"type":45,"value":114},{"type":45,"value":128},{"type":40,"tag":130,"props":976,"children":978},{"id":977},"what-this-skill-deliberately-does-not-ship",[979],{"type":45,"value":980},"What this skill deliberately does not ship",{"type":40,"tag":48,"props":982,"children":983},{},[984,986,991],{"type":45,"value":985},"This skill is ",{"type":40,"tag":52,"props":987,"children":988},{},[989],{"type":45,"value":990},"agent guidance",{"type":45,"value":992},", not a samples or templates\nbundle. To keep the boundary clean, it deliberately does not\ncontain — and pull requests should not add:",{"type":40,"tag":149,"props":994,"children":995},{},[996,1037,1092],{"type":40,"tag":153,"props":997,"children":998},{},[999,1004,1006,1012,1014,1024,1026,1036],{"type":40,"tag":52,"props":1000,"children":1001},{},[1002],{"type":45,"value":1003},"Pre-written DOCA GPUNetIO application source code or CUDA\nkernel source, in any language.",{"type":45,"value":1005}," The verified GPUNetIO\nsource is the shipped C + CUDA samples at\n",{"type":40,"tag":72,"props":1007,"children":1009},{"className":1008},[],[1010],{"type":45,"value":1011},"\u002Fopt\u002Fmellanox\u002Fdoca\u002Fsamples\u002Fdoca_gpunetio\u002F",{"type":45,"value":1013}," and the GPU Packet\nProcessing reference application. The agent's job is to route\nthe user to those files and prescribe a minimum-diff\nmodification on them via the universal modify-a-sample\nworkflow in\n",{"type":40,"tag":67,"props":1015,"children":1017},{"href":1016},"..\u002F..\u002Fdoca-programming-guide\u002FSKILL.md",[1018],{"type":40,"tag":72,"props":1019,"children":1021},{"className":1020},[],[1022],{"type":45,"value":1023},"doca-programming-guide",{"type":45,"value":1025},",\nlayered with the GPUNetIO-specific overrides in\n",{"type":40,"tag":67,"props":1027,"children":1029},{"href":1028},"TASKS.md#modify",[1030],{"type":40,"tag":72,"props":1031,"children":1033},{"className":1032},[],[1034],{"type":45,"value":1035},"TASKS.md ## modify",{"type":45,"value":128},{"type":40,"tag":153,"props":1038,"children":1039},{},[1040,1045,1047,1053,1054,1060,1062,1067,1069,1075,1076,1082,1084,1090],{"type":40,"tag":52,"props":1041,"children":1042},{},[1043],{"type":45,"value":1044},"Standalone build manifests",{"type":45,"value":1046}," (",{"type":40,"tag":72,"props":1048,"children":1050},{"className":1049},[],[1051],{"type":45,"value":1052},"meson.build",{"type":45,"value":632},{"type":40,"tag":72,"props":1055,"children":1057},{"className":1056},[],[1058],{"type":45,"value":1059},"CMakeLists.txt",{"type":45,"value":1061},", …) parked inside the skill. The agent\nconstructs the build manifest ",{"type":40,"tag":80,"props":1063,"children":1064},{},[1065],{"type":45,"value":1066},"in the user's project\ndirectory",{"type":45,"value":1068}," against the user's installed DOCA + CUDA toolkit,\nwhere ",{"type":40,"tag":72,"props":1070,"children":1072},{"className":1071},[],[1073],{"type":45,"value":1074},"pkg-config --modversion doca-gpunetio",{"type":45,"value":632},{"type":40,"tag":72,"props":1077,"children":1079},{"className":1078},[],[1080],{"type":45,"value":1081},"pkg-config --modversion doca-common",{"type":45,"value":1083},", and ",{"type":40,"tag":72,"props":1085,"children":1087},{"className":1086},[],[1088],{"type":45,"value":1089},"nvcc --version",{"type":45,"value":1091},"\nform the version gate.",{"type":40,"tag":153,"props":1093,"children":1094},{},[1095,1123],{"type":40,"tag":52,"props":1096,"children":1097},{},[1098,1100,1106,1107,1113,1115,1121],{"type":45,"value":1099},"A ",{"type":40,"tag":72,"props":1101,"children":1103},{"className":1102},[],[1104],{"type":45,"value":1105},"samples\u002F",{"type":45,"value":893},{"type":40,"tag":72,"props":1108,"children":1110},{"className":1109},[],[1111],{"type":45,"value":1112},"bindings\u002F",{"type":45,"value":1114},", or ",{"type":40,"tag":72,"props":1116,"children":1118},{"className":1117},[],[1119],{"type":45,"value":1120},"reference\u002F",{"type":45,"value":1122}," subtree",{"type":45,"value":1124}," of any\nkind. A mock or incomplete artifact in this skill's tree, even\none labeled \"reference\", is misleading: users will read it as\nbuildable.",{"type":40,"tag":130,"props":1126,"children":1128},{"id":1127},"loading-order",[1129],{"type":45,"value":1130},"Loading order",{"type":40,"tag":1132,"props":1133,"children":1134},"ol",{},[1135,1148,1168],{"type":40,"tag":153,"props":1136,"children":1137},{},[1138,1140,1146],{"type":45,"value":1139},"Read this ",{"type":40,"tag":72,"props":1141,"children":1143},{"className":1142},[],[1144],{"type":45,"value":1145},"SKILL.md",{"type":45,"value":1147}," first to confirm the user's question is\nin scope.",{"type":40,"tag":153,"props":1149,"children":1150},{},[1151],{"type":40,"tag":52,"props":1152,"children":1153},{},[1154,1156,1161,1163,1167],{"type":45,"value":1155},"For the GPUNetIO capability matrix, the ",{"type":40,"tag":72,"props":1157,"children":1159},{"className":1158},[],[1160],{"type":45,"value":618},{"type":45,"value":1162}," per-device\ncontext, the persistent-kernel pattern, the dual capability\nquery, the env-precondition policy, the error taxonomy, the\nobservability surface, and the safety policy, see\n",{"type":40,"tag":67,"props":1164,"children":1165},{"href":89},[1166],{"type":45,"value":89},{"type":45,"value":128},{"type":40,"tag":153,"props":1169,"children":1170},{},[1171],{"type":40,"tag":52,"props":1172,"children":1173},{},[1174,1176,1180],{"type":45,"value":1175},"For step-by-step workflows — configure, build, modify, run,\ntest, debug — see ",{"type":40,"tag":67,"props":1177,"children":1178},{"href":69},[1179],{"type":45,"value":69},{"type":45,"value":128},{"type":40,"tag":48,"props":1182,"children":1183},{},[1184,1186,1194,1196,1204],{"type":45,"value":1185},"Both companion files cross-link to each other,\n",{"type":40,"tag":67,"props":1187,"children":1188},{"href":417},[1189],{"type":40,"tag":72,"props":1190,"children":1192},{"className":1191},[],[1193],{"type":45,"value":424},{"type":45,"value":1195}," for the canonical\nDOCA version-handling rules (with the GPUNetIO overlay that DOCA\nmust match CUDA), and\n",{"type":40,"tag":67,"props":1197,"children":1198},{"href":767},[1199],{"type":40,"tag":72,"props":1200,"children":1202},{"className":1201},[],[1203],{"type":45,"value":774},{"type":45,"value":1205},"\nwhenever the right answer is \"look it up in the public DOCA\nGPUNetIO guide, the DOCA Compatibility Policy, the CUDA toolkit\ndocs, or the on-disk install layout\" rather than\n\"GPUNetIO-specific guidance\".",{"type":40,"tag":130,"props":1207,"children":1209},{"id":1208},"related-skills",[1210],{"type":45,"value":1211},"Related skills",{"type":40,"tag":149,"props":1213,"children":1214},{},[1215,1237,1257,1285,1306,1333],{"type":40,"tag":153,"props":1216,"children":1217},{},[1218,1226,1228,1235],{"type":40,"tag":67,"props":1219,"children":1220},{"href":767},[1221],{"type":40,"tag":72,"props":1222,"children":1224},{"className":1223},[],[1225],{"type":45,"value":774},{"type":45,"value":1227}," —\nthe routing table for every public DOCA documentation source\nand the on-disk layout of an installed DOCA package. The\nGPUNetIO public guide is at\n",{"type":40,"tag":67,"props":1229,"children":1233},{"href":1230,"rel":1231},"https:\u002F\u002Fdocs.nvidia.com\u002Fdoca\u002Fsdk\u002FDOCA-GPUNetIO\u002Findex.html",[1232],"nofollow",[1234],{"type":45,"value":1230},{"type":45,"value":1236},";\nthe GPU Packet Processing reference application is reachable\nfrom there. The CUDA toolkit and DOCA Compatibility Policy\nlinks live in the same routing table.",{"type":40,"tag":153,"props":1238,"children":1239},{},[1240,1248,1250,1255],{"type":40,"tag":67,"props":1241,"children":1242},{"href":107},[1243],{"type":40,"tag":72,"props":1244,"children":1246},{"className":1245},[],[1247],{"type":45,"value":114},{"type":45,"value":1249}," — env preparation,\ninstall verification, CUDA toolkit install \u002F verification, and\nthe ",{"type":40,"tag":80,"props":1251,"children":1252},{},[1253],{"type":45,"value":1254},"I have no install yet",{"type":45,"value":1256}," path with the public NGC DOCA\ncontainer. This skill assumes its preconditions are satisfied\nAND that CUDA is installed at a version that matches DOCA.",{"type":40,"tag":153,"props":1258,"children":1259},{},[1260,1268,1270,1276,1278,1283],{"type":40,"tag":67,"props":1261,"children":1262},{"href":417},[1263],{"type":40,"tag":72,"props":1264,"children":1266},{"className":1265},[],[1267],{"type":45,"value":424},{"type":45,"value":1269}," — canonical DOCA\nversion-handling rules. This skill's ",{"type":40,"tag":72,"props":1271,"children":1273},{"className":1272},[],[1274],{"type":45,"value":1275},"## Version compatibility",{"type":45,"value":1277}," cross-links the four-way match rule and adds\nthe GPUNetIO-specific ",{"type":40,"tag":80,"props":1279,"children":1280},{},[1281],{"type":45,"value":1282},"DOCA-and-CUDA must match",{"type":45,"value":1284}," overlay per\nthe DOCA Compatibility Policy.",{"type":40,"tag":153,"props":1286,"children":1287},{},[1288,1298,1300,1304],{"type":40,"tag":67,"props":1289,"children":1291},{"href":1290},"..\u002F..\u002Fdoca-structured-tools-contract\u002FSKILL.md",[1292],{"type":40,"tag":72,"props":1293,"children":1295},{"className":1294},[],[1296],{"type":45,"value":1297},"doca-structured-tools-contract",{"type":45,"value":1299}," —\nthe bundle's structured-tools precedence rule (detect \u002F prefer\n\u002F fall back \u002F report). The Command appendix in\n",{"type":40,"tag":67,"props":1301,"children":1302},{"href":69},[1303],{"type":45,"value":69},{"type":45,"value":1305}," honors this contract.",{"type":40,"tag":153,"props":1307,"children":1308},{},[1309,1317,1319,1324,1326,1331],{"type":40,"tag":67,"props":1310,"children":1311},{"href":1016},[1312],{"type":40,"tag":72,"props":1313,"children":1315},{"className":1314},[],[1316],{"type":45,"value":1023},{"type":45,"value":1318}," —\ngeneral DOCA programming patterns shared by every library:\nthe canonical ",{"type":40,"tag":72,"props":1320,"children":1322},{"className":1321},[],[1323],{"type":45,"value":570},{"type":45,"value":1325}," + meson build pattern, the\nuniversal modify-a-shipped-sample first-app workflow, the\nuniversal lifecycle, the cross-library ",{"type":40,"tag":72,"props":1327,"children":1329},{"className":1328},[],[1330],{"type":45,"value":447},{"type":45,"value":1332},"\ntaxonomy, and the program-side debug order. This skill\nlayers GPUNetIO specifics on top.",{"type":40,"tag":153,"props":1334,"children":1335},{},[1336,1344,1346,1351],{"type":40,"tag":67,"props":1337,"children":1338},{"href":505},[1339],{"type":40,"tag":72,"props":1340,"children":1342},{"className":1341},[],[1343],{"type":45,"value":512},{"type":45,"value":1345}," — the cross-cutting\ndebug ladder (install \u002F version \u002F build \u002F link \u002F runtime \u002F\nprogram \u002F driver). GPUNetIO-specific debug (CUDA + DOCA\nversion skew, ",{"type":40,"tag":72,"props":1347,"children":1349},{"className":1348},[],[1350],{"type":45,"value":291},{"type":45,"value":1352}," missing, persistent-kernel\nsilent hangs, CUDA-allocator + DOCA-registration mismatches)\noverlays on top of that ladder.",{"type":40,"tag":48,"props":1354,"children":1355},{},[1356,1358,1363,1364,1369,1371,1379],{"type":45,"value":1357},"DOCA Ethernet is GPUNetIO's mandatory companion library: GPU-visible\nRX \u002F TX queue handles are layered on top of ",{"type":40,"tag":72,"props":1359,"children":1361},{"className":1360},[],[1362],{"type":45,"value":181},{"type":45,"value":647},{"type":40,"tag":72,"props":1365,"children":1367},{"className":1366},[],[1368],{"type":45,"value":653},{"type":45,"value":1370}," from DOCA Ethernet. For the underlying queue setup,\nroute to ",{"type":40,"tag":67,"props":1372,"children":1373},{"href":119},[1374],{"type":40,"tag":72,"props":1375,"children":1377},{"className":1376},[],[1378],{"type":45,"value":126},{"type":45,"value":128},{"items":1381,"total":1485},[1382,1399,1413,1427,1439,1456,1471],{"slug":1383,"name":1383,"fn":1384,"description":1385,"org":1386,"tags":1387,"stars":20,"repoUrl":21,"updatedAt":1398},"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},[1388,1391,1394,1395],{"name":1389,"slug":1390,"type":13},"Data Analysis","data-analysis",{"name":1392,"slug":1393,"type":13},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":13},{"name":1396,"slug":1397,"type":13},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":1400,"name":1400,"fn":1401,"description":1402,"org":1403,"tags":1404,"stars":20,"repoUrl":21,"updatedAt":1412},"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},[1405,1408,1411],{"name":1406,"slug":1407,"type":13},"Deployment","deployment",{"name":1409,"slug":1410,"type":13},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":13},"2026-07-14T05:29:06.667109",{"slug":1414,"name":1414,"fn":1415,"description":1416,"org":1417,"tags":1418,"stars":20,"repoUrl":21,"updatedAt":1426},"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},[1419,1422,1423],{"name":1420,"slug":1421,"type":13},"Agents","agents",{"name":9,"slug":8,"type":13},{"name":1424,"slug":1425,"type":13},"Research","research","2026-07-14T05:28:06.816956",{"slug":1428,"name":1428,"fn":1429,"description":1430,"org":1431,"tags":1432,"stars":20,"repoUrl":21,"updatedAt":1438},"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},[1433,1434,1435],{"name":1389,"slug":1390,"type":13},{"name":9,"slug":8,"type":13},{"name":1436,"slug":1437,"type":13},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":1440,"name":1440,"fn":1441,"description":1442,"org":1443,"tags":1444,"stars":20,"repoUrl":21,"updatedAt":1455},"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},[1445,1448,1451,1452],{"name":1446,"slug":1447,"type":13},"Automation","automation",{"name":1449,"slug":1450,"type":13},"Imaging","imaging",{"name":9,"slug":8,"type":13},{"name":1453,"slug":1454,"type":13},"Video","video","2026-07-17T05:28:53.905004",{"slug":1457,"name":1457,"fn":1458,"description":1459,"org":1460,"tags":1461,"stars":20,"repoUrl":21,"updatedAt":1470},"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},[1462,1463,1466,1467],{"name":1406,"slug":1407,"type":13},{"name":1464,"slug":1465,"type":13},"Docker","docker",{"name":9,"slug":8,"type":13},{"name":1468,"slug":1469,"type":13},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1472,"name":1472,"fn":1473,"description":1474,"org":1475,"tags":1476,"stars":20,"repoUrl":21,"updatedAt":1484},"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},[1477,1478,1481],{"name":9,"slug":8,"type":13},{"name":1479,"slug":1480,"type":13},"Quantum Computing","quantum-computing",{"name":1482,"slug":1483,"type":13},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":1487,"total":1638},[1488,1506,1522,1533,1545,1559,1572,1586,1597,1606,1620,1629],{"slug":1489,"name":1489,"fn":1490,"description":1491,"org":1492,"tags":1493,"stars":1503,"repoUrl":1504,"updatedAt":1505},"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},[1494,1497,1500],{"name":1495,"slug":1496,"type":13},"Documentation","documentation",{"name":1498,"slug":1499,"type":13},"MCP","mcp",{"name":1501,"slug":1502,"type":13},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1507,"name":1507,"fn":1508,"description":1509,"org":1510,"tags":1511,"stars":1519,"repoUrl":1520,"updatedAt":1521},"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},[1512,1515,1516],{"name":1513,"slug":1514,"type":13},"Containers","containers",{"name":1406,"slug":1407,"type":13},{"name":1517,"slug":1518,"type":13},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1523,"name":1523,"fn":1524,"description":1525,"org":1526,"tags":1527,"stars":1519,"repoUrl":1520,"updatedAt":1532},"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},[1528,1531],{"name":1529,"slug":1530,"type":13},"CI\u002FCD","ci-cd",{"name":1406,"slug":1407,"type":13},"2026-07-14T05:25:59.97109",{"slug":1534,"name":1534,"fn":1535,"description":1536,"org":1537,"tags":1538,"stars":1519,"repoUrl":1520,"updatedAt":1544},"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},[1539,1540,1541],{"name":1529,"slug":1530,"type":13},{"name":1406,"slug":1407,"type":13},{"name":1542,"slug":1543,"type":13},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1546,"name":1546,"fn":1547,"description":1548,"org":1549,"tags":1550,"stars":1519,"repoUrl":1520,"updatedAt":1558},"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},[1551,1554,1555],{"name":1552,"slug":1553,"type":13},"Debugging","debugging",{"name":1542,"slug":1543,"type":13},{"name":1556,"slug":1557,"type":13},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1560,"name":1560,"fn":1561,"description":1562,"org":1563,"tags":1564,"stars":1519,"repoUrl":1520,"updatedAt":1571},"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},[1565,1568],{"name":1566,"slug":1567,"type":13},"Best Practices","best-practices",{"name":1569,"slug":1570,"type":13},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1573,"name":1573,"fn":1574,"description":1575,"org":1576,"tags":1577,"stars":1519,"repoUrl":1520,"updatedAt":1585},"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},[1578,1581,1584],{"name":1579,"slug":1580,"type":13},"Machine Learning","machine-learning",{"name":1582,"slug":1583,"type":13},"Migration","migration",{"name":9,"slug":8,"type":13},"2026-07-17T06:07:11.777011",{"slug":1587,"name":1587,"fn":1588,"description":1589,"org":1590,"tags":1591,"stars":1519,"repoUrl":1520,"updatedAt":1596},"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},[1592,1595],{"name":1593,"slug":1594,"type":13},"QA","qa",{"name":1436,"slug":1437,"type":13},"2026-07-14T05:25:53.673039",{"slug":1598,"name":1598,"fn":1599,"description":1600,"org":1601,"tags":1602,"stars":1519,"repoUrl":1520,"updatedAt":1605},"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},[1603,1604],{"name":1406,"slug":1407,"type":13},{"name":1409,"slug":1410,"type":13},"2026-07-14T05:25:49.362534",{"slug":1607,"name":1607,"fn":1608,"description":1609,"org":1610,"tags":1611,"stars":1519,"repoUrl":1520,"updatedAt":1619},"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},[1612,1615,1616],{"name":1613,"slug":1614,"type":13},"Code Review","code-review",{"name":1542,"slug":1543,"type":13},{"name":1617,"slug":1618,"type":13},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1621,"name":1621,"fn":1622,"description":1623,"org":1624,"tags":1625,"stars":1519,"repoUrl":1520,"updatedAt":1628},"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},[1626,1627],{"name":1593,"slug":1594,"type":13},{"name":1436,"slug":1437,"type":13},"2026-07-14T05:25:54.928983",{"slug":1630,"name":1630,"fn":1631,"description":1632,"org":1633,"tags":1634,"stars":1519,"repoUrl":1520,"updatedAt":1637},"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},[1635,1636],{"name":1446,"slug":1447,"type":13},{"name":1529,"slug":1530,"type":13},"2026-07-30T05:29:03.275638",496]