[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nvshmem-select-remote-transport":3,"mdc-tkqrz6-key":36,"related-org-nvidia-nvshmem-select-remote-transport":1132,"related-repo-nvidia-nvshmem-select-remote-transport":1292},{"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":31,"sourceUrl":34,"mdContent":35},"nvshmem-select-remote-transport","select NVSHMEM remote transport configurations","Select an NVSHMEM remote transport from target system and kernel evidence. Use for inter-node selection, compatibility checks, or configuration.",{"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],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":17,"slug":18,"type":15},"Configuration","configuration",{"name":9,"slug":8,"type":15},579,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fnvshmem","2026-08-31T09:18:59.493722","Apache-2.0",104,[26,27,28,29,8,30],"communciations","cpp","cuda","deep-learning","python",{"repoUrl":21,"stars":20,"forks":24,"topics":32,"description":33},[26,27,28,29,8,30],"NVIDIA NVSHMEM is a parallel programming interface for NVIDIA GPUs based on OpenSHMEM. NVSHMEM can significantly reduce multi-process communication and coordination overheads by allowing programmers to perform one-sided communication from within CUDA kernels and on CUDA streams.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fnvshmem\u002Ftree\u002FHEAD\u002Fskills\u002Fnvshmem-select-remote-transport","---\nname: nvshmem-select-remote-transport\ndescription: Select an NVSHMEM remote transport from target system and kernel evidence. Use for inter-node selection, compatibility checks, or configuration.\nlicense: Apache-2.0\nmetadata:\n  version: \"1.0.0\"\n  author: NVIDIA NVSHMEM Team \u003Cnvshmem@nvidia.com>\n  tags:\n    - nvshmem\n    - transport\n    - tuning\n---\n\n# Select an NVSHMEM Remote Transport\n\n## Purpose\n\nRecommend one NVSHMEM transport configuration from static evidence. Separate hard eligibility from expected kernel fit, distinguish documented facts from inference.\n\nFocus on read-only system probes and source inspection. Do not run NVSHMEM applications, correctness tests, or performance benchmarks.\n\n## Prerequisites\n\n- Keep [transport-matrix.md](references\u002Ftransport-matrix.md) and [kernel-fit.md](references\u002Fkernel-fit.md) readable as part of the skill package. Read both before ranking transports.\n- Require an exact NVSHMEM version and an inventory of installed transport plugins for an unconditional recommendation.\n- Require the selected CUDA kernel\u002Fapplication source or an equivalent communication profile.\n- Collect target-node evidence through an already available unprivileged shell or obtain it from the user. Never request root access or inspect a login node as if it were the target compute node.\n- Use the sibling `$nvshmem-docs` workflow, or its readable `..\u002Fnvshmem-docs\u002FSKILL.md`, only when the user requests current\u002Fexact official verification or the bundled matrix cannot establish release applicability.\n- Require no API key, privileged command, network access for inspection, or benchmark execution.\n\nIf direct target-node access is unavailable, provide the bundled collector and request its complete labeled output.\n\n## Inputs\n\n### Required for an Unconditional Recommendation\n\n- Exact NVSHMEM version and installed transport-plugin inventory.\n- Target fabric\u002Fprovider: InfiniBand, RoCE, EFA, Slingshot\u002FCXI, UCX-managed, or NVLink-only.\n- Selected CUDA kernel\u002Fapplication source, or a profile covering:\n  - host, on-stream, or device API use;\n  - RMA, AMO, signaling, synchronization, or collective operations;\n  - typical and range of message sizes;\n  - issuing threads, warps, and CTAs;\n  - peer fan-out and expected simultaneous message rate.\n- Job scale: total PEs and local PEs per node.\n- Intended outcome: general default, latency-sensitive, throughput-oriented, or compatibility-first.\n\n### Optional\n\n- `NVSHMEM_PREFIX`, CMake cache, module\u002Fcontainer details, launcher command, environment, and NVSHMEM configuration files.\n- GPU, NIC, driver, kernel, OFED\u002Frdma-core, UCX, libfabric, GPUNetIO, or DOCA details not available through inspection.\n- GPU-to-NIC topology and local PE binding.\n- Host\u002Fdevice traffic mix, CPU availability, persistent-kernel behavior, batching opportunities, and completion frequency.\n- Required atomics, registered user buffers, VMM, DCI\u002FDCT, multi-NIC, Spectrum-X, or custom-QP features.\n- Preferred or prohibited transports.\n\nInspect before asking. Ask only for missing facts that can change eligibility, ranking, or the emitted configuration. When required evidence remains unavailable, return conditional branches instead of inventing it.\n\n## Instructions\n\n### 1. Resolve the Target\n\n- Identify the application, exactly selected kernel or host call path, deployment, NVSHMEM version, job scale, and optimization goal.\n- Treat assessment and recommendation as read-only. Do not patch application code or change launcher\u002Fsystem configuration.\n- Limit source inspection to the selected kernel, its device call graph, and directly corresponding host\u002Fon-stream call sites. Ignore unrelated NVSHMEM kernels.\n- Determine whether the relevant communication can cross nodes. Local P2P\u002FNVLink traffic does not use the selected remote transport.\n\n### 2. Collect System Evidence\n\nWhen already on the target compute node, announce the read-only probe and invoke the bundled collector:\n\n```text\nrun_script(\"scripts\u002Fcollect-transport-facts.sh\")\n```\n\nWhen the installation prefix is known, pass it explicitly rather than relying on shell expansion:\n\n```text\nrun_script(\"scripts\u002Fcollect-transport-facts.sh\", \"--prefix\", \"\u002Fopt\u002Fnvshmem\")\n```\n\nReplace the example prefix with the resolved absolute prefix. Resolve the script relative to this `SKILL.md`. Preserve partial output when it exits with status 2.\n\nWhen not on the target node, ask the user to run `bash scripts\u002Fcollect-transport-facts.sh [--prefix PATH]` there and return the complete output. Do not infer cluster-wide uniformity from one node without stating the assumption.\n\n### 3. Classify the Communication\n\nRead [kernel-fit.md](references\u002Fkernel-fit.md) and record:\n\n- API surface and exact operations;\n- remote versus local paths;\n- message-size distribution and whether operations are packed;\n- concurrent issuers, peers, and expected physical message rate;\n- ordering\u002Fcompletion frequency and QP pressure;\n- host\u002Fdevice mix, CPU availability, and persistent-kernel behavior;\n- scale-sensitive resource needs such as RC versus DCI\u002FDCT.\n\nDo not infer a high message rate merely from many logical elements. Account for application aggregation, one-thread issuance, warp coalescing, and batching.\n\n### 4. Apply Hard Eligibility Gates\n\nRead [transport-matrix.md](references\u002Ftransport-matrix.md) completely. Exclude a candidate when any required condition is disproven:\n\n- plugin unavailable in the selected NVSHMEM installation;\n- exact release or operation unsupported;\n- fabric\u002Fprovider mismatch;\n- missing driver, peer-memory\u002FDMA-BUF, GDRCopy, DevX, UCX, libfabric, GPUNetIO, or DOCA prerequisite;\n- incompatible memory mode, registered-buffer use, scale, or required feature.\n\nTreat a missing fact as `conditional`, not `eligible`. Keep build support, runtime loadability, and hardware suitability separate.\n\nFor IBGDA, evaluate both the GPU data path and a separate host remote transport. For GPUNetIO GDAKI, require `gpunetio` as the remote transport. Never enable IBGDA and GPUNetIO GDAKI together.\n\n### 5. Rank Eligible Candidates\n\nApply these rules in order without assigning artificial scores:\n\n1. **Fabric-native gate:** Prefer libfabric with `efa` on EFA and `cxi` on Slingshot\u002FCXI when the exact release supports the required operations and memory mode.\n2. **No remote network:** Select `none` only when all relevant PEs communicate within a peer-reachable NVLink domain and no remote path is required.\n3. **Fine-grained parallel device traffic:** Favor an eligible GDAKI path when many GPU issuers submit small independent remote operations and a CPU proxy would serialize them.\n4. **GDAKI choice:**\n   - Favor IBGDA when DCI\u002FDCT connection scaling is required, GPUNetIO is absent, or the exact release has stronger required operation coverage.\n   - Favor GPUNetIO GDAKI when its RC connection scale is acceptable and a unified GPUNetIO CPU\u002FGPU path or verified DOCA feature is useful.\n   - Do not treat GPUNetIO design goals as proof that it outperforms IBGDA for this workload.\n5. **Proxy-friendly traffic:** Favor a proxy path for host\u002Fon-stream-dominant work, packed larger transfers, low message concurrency, or latency-sensitive isolated operations.\n6. **Mellanox proxy choice:** Use IBRC as the conservative verbs default. Recommend IBDevX only when mlx5 DevX is available, its exact operation coverage is sufficient, and a concrete requirement such as avoiding a GDRCopy-dependent path justifies it. Do not claim general IBDevX superiority.\n7. **UCX choice:** Favor UCX for a UCX-controlled deployment or a verified release-specific capability unavailable from the native alternatives. Preserve its experimental\u002Fstatus caveats from the exact documentation.\n\nIf two candidates remain close, choose the compatibility-first option as the recommendation, report the other as runner-up, and state that static analysis cannot establish the performance winner.\n\n### 6. Emit Minimal Selection Configuration\n\nEmit only the variables required to select the resolved configuration:\n\n| Configuration | Minimal exports |\n| --- | --- |\n| IBRC | `NVSHMEM_REMOTE_TRANSPORT=ibrc` |\n| IBDevX | `NVSHMEM_REMOTE_TRANSPORT=ibdevx` |\n| IBGDA | `NVSHMEM_REMOTE_TRANSPORT=\u003Cresolved-host-transport>` and `NVSHMEM_IB_ENABLE_IBGDA=1` |\n| GPUNetIO CPU | `NVSHMEM_REMOTE_TRANSPORT=gpunetio` and `NVSHMEM_GPUNETIO_ENABLE_GDAKI=0` |\n| GPUNetIO GDAKI | `NVSHMEM_REMOTE_TRANSPORT=gpunetio` and `NVSHMEM_GPUNETIO_ENABLE_GDAKI=1` |\n| UCX | `NVSHMEM_REMOTE_TRANSPORT=ucx` |\n| libfabric | `NVSHMEM_REMOTE_TRANSPORT=libfabric` and `NVSHMEM_LIBFABRIC_PROVIDER=\u003Ccxi|efa|verbs>` |\n| No remote | `NVSHMEM_REMOTE_TRANSPORT=none` |\n\nDo not add HCA mapping, QP counts, batching, NIC handlers, memory-mode changes, or provider tuning to this block. Put mandatory compatibility preconditions outside the block and route mapping\u002Ftuning to the appropriate sibling skill.\n\nWithhold an unconditional export block when version, plugin, provider, operation coverage, or kernel profile is unresolved.\n\n### 7. Return the Recommendation\n\nUse this exact section order:\n\n```text\nRecommendation: \u003Ctransport or combined host\u002FGPU configuration; conditional if necessary>\n\nSelection configuration\n\u003Cminimal export block, or why it is withheld>\n\nSystem eligibility evidence\n\u003Cversion, plugins, fabric\u002Fprovider, driver\u002Fmodule\u002Fdependency evidence>\n\nKernel-fit evidence\n\u003CAPI surface, operation shape, sizes, concurrency, fan-out, completion, scale>\n\nRunner-up\n\u003Ccandidate and the fact or tradeoff that kept it second>\n\nExcluded candidates\n\u003Ccandidate: hard gate or unresolved prerequisite>\n\nConfidence and unknowns\n\u003Chigh|medium|low; documented facts versus inference; missing evidence>\n\nStatic follow-up checks\n\u003Cread-only checks or exact facts still needed; never benchmarks>\n```\n\nState the documentation version consulted when live verification was used. Describe performance as expected fit, not an assured result.\n\n## Routing\n\n- Route exact HCA selection and PE-to-NIC mapping to `$nvshmem-configure-nic-pe-mapping` when available.\n- Route missing packages, plugins, or builds to `$nvshmem-install` when available.\n- Route runtime initialization failures to `$nvshmem-troubleshoot-and-report-bugs` when available.\n- If a sibling skill is unavailable, provide the boundary and required next evidence without inventing its behavior.\n\n## Available Script\n\n| Script | Purpose | Arguments |\n| --- | --- | --- |\n| [`scripts\u002Fcollect-transport-facts.sh`](scripts\u002Fcollect-transport-facts.sh) | Collect read-only NVSHMEM, GPU, kernel-module, RDMA, and provider evidence from a target compute node. | Optional `--prefix PATH` |\n\n\n## Examples\n\n- `Use $nvshmem-select-remote-transport. Inspect this target compute node and src\u002Fexchange.cu, then recommend a throughput-oriented transport for 256 PEs with 8 local PEs per node.`\n- `Use $nvshmem-select-remote-transport with the attached collector output and communication profile. Recommend a compatibility-first transport for 32 PEs with 8 local PEs per node, and withhold exports if required evidence is missing.`\n\n\n## Guardrails\n\n- Do not run NVSHMEM binaries other than the read-only `nvshmem-info -n` query.\n- Do not launch distributed jobs, allocate nodes, benchmark transports, or mutate system\u002Fapplication state.\n- Do not recommend a plugin merely because its source exists; require evidence that the selected installation contains it.\n- Do not convert absent evidence into compatibility.\n- Do not claim that Ethernet link-layer evidence alone proves working RoCE.\n- Do not let kernel heuristics override an incompatible fabric\u002Fprovider or missing prerequisite.\n- Do not emit both IBGDA and GPUNetIO GDAKI enables.\n\n## Troubleshooting\n\n- If the collector exits with status 2, preserve its output and resolve only the missing facts listed under `diagnostics`.\n- If a plugin is missing or fails to initialize, verify the selected prefix and dependencies, then route runtime failures to `$nvshmem-troubleshoot-and-report-bugs`.\n- If the recommendation remains conditional, obtain the exact version, provider, operation coverage, or kernel profile before emitting exports.\n\n## Limitations\n\n- Static analysis estimates architectural fit; it does not validate runtime correctness, measure performance, or guarantee the fastest transport.\n- The bundled compatibility matrix is release-sensitive; verify the exact target release against its official documentation when the matrix cannot establish applicability.\n- The collector describes one node and one selected installation. It does not prove cluster-wide uniformity, working RoCE, or complete operation coverage.\n- Incomplete source, topology, plugin, or provider evidence requires a conditional recommendation with no unconditional exports.\n",{"data":37,"body":45},{"name":4,"description":6,"license":23,"metadata":38},{"version":39,"author":40,"tags":41},"1.0.0","NVIDIA NVSHMEM Team \u003Cnvshmem@nvidia.com>",[42,43,44],"nvshmem","transport","tuning",{"type":46,"children":47},"root",[48,57,64,70,75,81,150,155,161,168,224,230,269,274,280,286,309,315,320,332,337,346,359,372,378,389,427,432,438,448,476,497,510,516,521,636,641,647,652,837,842,847,853,858,867,872,878,923,929,986,992,1013,1019,1065,1071,1103,1109],{"type":49,"tag":50,"props":51,"children":53},"element","h1",{"id":52},"select-an-nvshmem-remote-transport",[54],{"type":55,"value":56},"text","Select an NVSHMEM Remote Transport",{"type":49,"tag":58,"props":59,"children":61},"h2",{"id":60},"purpose",[62],{"type":55,"value":63},"Purpose",{"type":49,"tag":65,"props":66,"children":67},"p",{},[68],{"type":55,"value":69},"Recommend one NVSHMEM transport configuration from static evidence. Separate hard eligibility from expected kernel fit, distinguish documented facts from inference.",{"type":49,"tag":65,"props":71,"children":72},{},[73],{"type":55,"value":74},"Focus on read-only system probes and source inspection. Do not run NVSHMEM applications, correctness tests, or performance benchmarks.",{"type":49,"tag":58,"props":76,"children":78},{"id":77},"prerequisites",[79],{"type":55,"value":80},"Prerequisites",{"type":49,"tag":82,"props":83,"children":84},"ul",{},[85,108,113,118,123,145],{"type":49,"tag":86,"props":87,"children":88},"li",{},[89,91,98,100,106],{"type":55,"value":90},"Keep ",{"type":49,"tag":92,"props":93,"children":95},"a",{"href":94},"references\u002Ftransport-matrix.md",[96],{"type":55,"value":97},"transport-matrix.md",{"type":55,"value":99}," and ",{"type":49,"tag":92,"props":101,"children":103},{"href":102},"references\u002Fkernel-fit.md",[104],{"type":55,"value":105},"kernel-fit.md",{"type":55,"value":107}," readable as part of the skill package. Read both before ranking transports.",{"type":49,"tag":86,"props":109,"children":110},{},[111],{"type":55,"value":112},"Require an exact NVSHMEM version and an inventory of installed transport plugins for an unconditional recommendation.",{"type":49,"tag":86,"props":114,"children":115},{},[116],{"type":55,"value":117},"Require the selected CUDA kernel\u002Fapplication source or an equivalent communication profile.",{"type":49,"tag":86,"props":119,"children":120},{},[121],{"type":55,"value":122},"Collect target-node evidence through an already available unprivileged shell or obtain it from the user. Never request root access or inspect a login node as if it were the target compute node.",{"type":49,"tag":86,"props":124,"children":125},{},[126,128,135,137,143],{"type":55,"value":127},"Use the sibling ",{"type":49,"tag":129,"props":130,"children":132},"code",{"className":131},[],[133],{"type":55,"value":134},"$nvshmem-docs",{"type":55,"value":136}," workflow, or its readable ",{"type":49,"tag":129,"props":138,"children":140},{"className":139},[],[141],{"type":55,"value":142},"..\u002Fnvshmem-docs\u002FSKILL.md",{"type":55,"value":144},", only when the user requests current\u002Fexact official verification or the bundled matrix cannot establish release applicability.",{"type":49,"tag":86,"props":146,"children":147},{},[148],{"type":55,"value":149},"Require no API key, privileged command, network access for inspection, or benchmark execution.",{"type":49,"tag":65,"props":151,"children":152},{},[153],{"type":55,"value":154},"If direct target-node access is unavailable, provide the bundled collector and request its complete labeled output.",{"type":49,"tag":58,"props":156,"children":158},{"id":157},"inputs",[159],{"type":55,"value":160},"Inputs",{"type":49,"tag":162,"props":163,"children":165},"h3",{"id":164},"required-for-an-unconditional-recommendation",[166],{"type":55,"value":167},"Required for an Unconditional Recommendation",{"type":49,"tag":82,"props":169,"children":170},{},[171,176,181,214,219],{"type":49,"tag":86,"props":172,"children":173},{},[174],{"type":55,"value":175},"Exact NVSHMEM version and installed transport-plugin inventory.",{"type":49,"tag":86,"props":177,"children":178},{},[179],{"type":55,"value":180},"Target fabric\u002Fprovider: InfiniBand, RoCE, EFA, Slingshot\u002FCXI, UCX-managed, or NVLink-only.",{"type":49,"tag":86,"props":182,"children":183},{},[184,186],{"type":55,"value":185},"Selected CUDA kernel\u002Fapplication source, or a profile covering:\n",{"type":49,"tag":82,"props":187,"children":188},{},[189,194,199,204,209],{"type":49,"tag":86,"props":190,"children":191},{},[192],{"type":55,"value":193},"host, on-stream, or device API use;",{"type":49,"tag":86,"props":195,"children":196},{},[197],{"type":55,"value":198},"RMA, AMO, signaling, synchronization, or collective operations;",{"type":49,"tag":86,"props":200,"children":201},{},[202],{"type":55,"value":203},"typical and range of message sizes;",{"type":49,"tag":86,"props":205,"children":206},{},[207],{"type":55,"value":208},"issuing threads, warps, and CTAs;",{"type":49,"tag":86,"props":210,"children":211},{},[212],{"type":55,"value":213},"peer fan-out and expected simultaneous message rate.",{"type":49,"tag":86,"props":215,"children":216},{},[217],{"type":55,"value":218},"Job scale: total PEs and local PEs per node.",{"type":49,"tag":86,"props":220,"children":221},{},[222],{"type":55,"value":223},"Intended outcome: general default, latency-sensitive, throughput-oriented, or compatibility-first.",{"type":49,"tag":162,"props":225,"children":227},{"id":226},"optional",[228],{"type":55,"value":229},"Optional",{"type":49,"tag":82,"props":231,"children":232},{},[233,244,249,254,259,264],{"type":49,"tag":86,"props":234,"children":235},{},[236,242],{"type":49,"tag":129,"props":237,"children":239},{"className":238},[],[240],{"type":55,"value":241},"NVSHMEM_PREFIX",{"type":55,"value":243},", CMake cache, module\u002Fcontainer details, launcher command, environment, and NVSHMEM configuration files.",{"type":49,"tag":86,"props":245,"children":246},{},[247],{"type":55,"value":248},"GPU, NIC, driver, kernel, OFED\u002Frdma-core, UCX, libfabric, GPUNetIO, or DOCA details not available through inspection.",{"type":49,"tag":86,"props":250,"children":251},{},[252],{"type":55,"value":253},"GPU-to-NIC topology and local PE binding.",{"type":49,"tag":86,"props":255,"children":256},{},[257],{"type":55,"value":258},"Host\u002Fdevice traffic mix, CPU availability, persistent-kernel behavior, batching opportunities, and completion frequency.",{"type":49,"tag":86,"props":260,"children":261},{},[262],{"type":55,"value":263},"Required atomics, registered user buffers, VMM, DCI\u002FDCT, multi-NIC, Spectrum-X, or custom-QP features.",{"type":49,"tag":86,"props":265,"children":266},{},[267],{"type":55,"value":268},"Preferred or prohibited transports.",{"type":49,"tag":65,"props":270,"children":271},{},[272],{"type":55,"value":273},"Inspect before asking. Ask only for missing facts that can change eligibility, ranking, or the emitted configuration. When required evidence remains unavailable, return conditional branches instead of inventing it.",{"type":49,"tag":58,"props":275,"children":277},{"id":276},"instructions",[278],{"type":55,"value":279},"Instructions",{"type":49,"tag":162,"props":281,"children":283},{"id":282},"_1-resolve-the-target",[284],{"type":55,"value":285},"1. Resolve the Target",{"type":49,"tag":82,"props":287,"children":288},{},[289,294,299,304],{"type":49,"tag":86,"props":290,"children":291},{},[292],{"type":55,"value":293},"Identify the application, exactly selected kernel or host call path, deployment, NVSHMEM version, job scale, and optimization goal.",{"type":49,"tag":86,"props":295,"children":296},{},[297],{"type":55,"value":298},"Treat assessment and recommendation as read-only. Do not patch application code or change launcher\u002Fsystem configuration.",{"type":49,"tag":86,"props":300,"children":301},{},[302],{"type":55,"value":303},"Limit source inspection to the selected kernel, its device call graph, and directly corresponding host\u002Fon-stream call sites. Ignore unrelated NVSHMEM kernels.",{"type":49,"tag":86,"props":305,"children":306},{},[307],{"type":55,"value":308},"Determine whether the relevant communication can cross nodes. Local P2P\u002FNVLink traffic does not use the selected remote transport.",{"type":49,"tag":162,"props":310,"children":312},{"id":311},"_2-collect-system-evidence",[313],{"type":55,"value":314},"2. Collect System Evidence",{"type":49,"tag":65,"props":316,"children":317},{},[318],{"type":55,"value":319},"When already on the target compute node, announce the read-only probe and invoke the bundled collector:",{"type":49,"tag":321,"props":322,"children":327},"pre",{"className":323,"code":325,"language":55,"meta":326},[324],"language-text","run_script(\"scripts\u002Fcollect-transport-facts.sh\")\n","",[328],{"type":49,"tag":129,"props":329,"children":330},{"__ignoreMap":326},[331],{"type":55,"value":325},{"type":49,"tag":65,"props":333,"children":334},{},[335],{"type":55,"value":336},"When the installation prefix is known, pass it explicitly rather than relying on shell expansion:",{"type":49,"tag":321,"props":338,"children":341},{"className":339,"code":340,"language":55,"meta":326},[324],"run_script(\"scripts\u002Fcollect-transport-facts.sh\", \"--prefix\", \"\u002Fopt\u002Fnvshmem\")\n",[342],{"type":49,"tag":129,"props":343,"children":344},{"__ignoreMap":326},[345],{"type":55,"value":340},{"type":49,"tag":65,"props":347,"children":348},{},[349,351,357],{"type":55,"value":350},"Replace the example prefix with the resolved absolute prefix. Resolve the script relative to this ",{"type":49,"tag":129,"props":352,"children":354},{"className":353},[],[355],{"type":55,"value":356},"SKILL.md",{"type":55,"value":358},". Preserve partial output when it exits with status 2.",{"type":49,"tag":65,"props":360,"children":361},{},[362,364,370],{"type":55,"value":363},"When not on the target node, ask the user to run ",{"type":49,"tag":129,"props":365,"children":367},{"className":366},[],[368],{"type":55,"value":369},"bash scripts\u002Fcollect-transport-facts.sh [--prefix PATH]",{"type":55,"value":371}," there and return the complete output. Do not infer cluster-wide uniformity from one node without stating the assumption.",{"type":49,"tag":162,"props":373,"children":375},{"id":374},"_3-classify-the-communication",[376],{"type":55,"value":377},"3. Classify the Communication",{"type":49,"tag":65,"props":379,"children":380},{},[381,383,387],{"type":55,"value":382},"Read ",{"type":49,"tag":92,"props":384,"children":385},{"href":102},[386],{"type":55,"value":105},{"type":55,"value":388}," and record:",{"type":49,"tag":82,"props":390,"children":391},{},[392,397,402,407,412,417,422],{"type":49,"tag":86,"props":393,"children":394},{},[395],{"type":55,"value":396},"API surface and exact operations;",{"type":49,"tag":86,"props":398,"children":399},{},[400],{"type":55,"value":401},"remote versus local paths;",{"type":49,"tag":86,"props":403,"children":404},{},[405],{"type":55,"value":406},"message-size distribution and whether operations are packed;",{"type":49,"tag":86,"props":408,"children":409},{},[410],{"type":55,"value":411},"concurrent issuers, peers, and expected physical message rate;",{"type":49,"tag":86,"props":413,"children":414},{},[415],{"type":55,"value":416},"ordering\u002Fcompletion frequency and QP pressure;",{"type":49,"tag":86,"props":418,"children":419},{},[420],{"type":55,"value":421},"host\u002Fdevice mix, CPU availability, and persistent-kernel behavior;",{"type":49,"tag":86,"props":423,"children":424},{},[425],{"type":55,"value":426},"scale-sensitive resource needs such as RC versus DCI\u002FDCT.",{"type":49,"tag":65,"props":428,"children":429},{},[430],{"type":55,"value":431},"Do not infer a high message rate merely from many logical elements. Account for application aggregation, one-thread issuance, warp coalescing, and batching.",{"type":49,"tag":162,"props":433,"children":435},{"id":434},"_4-apply-hard-eligibility-gates",[436],{"type":55,"value":437},"4. Apply Hard Eligibility Gates",{"type":49,"tag":65,"props":439,"children":440},{},[441,442,446],{"type":55,"value":382},{"type":49,"tag":92,"props":443,"children":444},{"href":94},[445],{"type":55,"value":97},{"type":55,"value":447}," completely. Exclude a candidate when any required condition is disproven:",{"type":49,"tag":82,"props":449,"children":450},{},[451,456,461,466,471],{"type":49,"tag":86,"props":452,"children":453},{},[454],{"type":55,"value":455},"plugin unavailable in the selected NVSHMEM installation;",{"type":49,"tag":86,"props":457,"children":458},{},[459],{"type":55,"value":460},"exact release or operation unsupported;",{"type":49,"tag":86,"props":462,"children":463},{},[464],{"type":55,"value":465},"fabric\u002Fprovider mismatch;",{"type":49,"tag":86,"props":467,"children":468},{},[469],{"type":55,"value":470},"missing driver, peer-memory\u002FDMA-BUF, GDRCopy, DevX, UCX, libfabric, GPUNetIO, or DOCA prerequisite;",{"type":49,"tag":86,"props":472,"children":473},{},[474],{"type":55,"value":475},"incompatible memory mode, registered-buffer use, scale, or required feature.",{"type":49,"tag":65,"props":477,"children":478},{},[479,481,487,489,495],{"type":55,"value":480},"Treat a missing fact as ",{"type":49,"tag":129,"props":482,"children":484},{"className":483},[],[485],{"type":55,"value":486},"conditional",{"type":55,"value":488},", not ",{"type":49,"tag":129,"props":490,"children":492},{"className":491},[],[493],{"type":55,"value":494},"eligible",{"type":55,"value":496},". Keep build support, runtime loadability, and hardware suitability separate.",{"type":49,"tag":65,"props":498,"children":499},{},[500,502,508],{"type":55,"value":501},"For IBGDA, evaluate both the GPU data path and a separate host remote transport. For GPUNetIO GDAKI, require ",{"type":49,"tag":129,"props":503,"children":505},{"className":504},[],[506],{"type":55,"value":507},"gpunetio",{"type":55,"value":509}," as the remote transport. Never enable IBGDA and GPUNetIO GDAKI together.",{"type":49,"tag":162,"props":511,"children":513},{"id":512},"_5-rank-eligible-candidates",[514],{"type":55,"value":515},"5. Rank Eligible Candidates",{"type":49,"tag":65,"props":517,"children":518},{},[519],{"type":55,"value":520},"Apply these rules in order without assigning artificial scores:",{"type":49,"tag":522,"props":523,"children":524},"ol",{},[525,552,570,580,606,616,626],{"type":49,"tag":86,"props":526,"children":527},{},[528,534,536,542,544,550],{"type":49,"tag":529,"props":530,"children":531},"strong",{},[532],{"type":55,"value":533},"Fabric-native gate:",{"type":55,"value":535}," Prefer libfabric with ",{"type":49,"tag":129,"props":537,"children":539},{"className":538},[],[540],{"type":55,"value":541},"efa",{"type":55,"value":543}," on EFA and ",{"type":49,"tag":129,"props":545,"children":547},{"className":546},[],[548],{"type":55,"value":549},"cxi",{"type":55,"value":551}," on Slingshot\u002FCXI when the exact release supports the required operations and memory mode.",{"type":49,"tag":86,"props":553,"children":554},{},[555,560,562,568],{"type":49,"tag":529,"props":556,"children":557},{},[558],{"type":55,"value":559},"No remote network:",{"type":55,"value":561}," Select ",{"type":49,"tag":129,"props":563,"children":565},{"className":564},[],[566],{"type":55,"value":567},"none",{"type":55,"value":569}," only when all relevant PEs communicate within a peer-reachable NVLink domain and no remote path is required.",{"type":49,"tag":86,"props":571,"children":572},{},[573,578],{"type":49,"tag":529,"props":574,"children":575},{},[576],{"type":55,"value":577},"Fine-grained parallel device traffic:",{"type":55,"value":579}," Favor an eligible GDAKI path when many GPU issuers submit small independent remote operations and a CPU proxy would serialize them.",{"type":49,"tag":86,"props":581,"children":582},{},[583,588],{"type":49,"tag":529,"props":584,"children":585},{},[586],{"type":55,"value":587},"GDAKI choice:",{"type":49,"tag":82,"props":589,"children":590},{},[591,596,601],{"type":49,"tag":86,"props":592,"children":593},{},[594],{"type":55,"value":595},"Favor IBGDA when DCI\u002FDCT connection scaling is required, GPUNetIO is absent, or the exact release has stronger required operation coverage.",{"type":49,"tag":86,"props":597,"children":598},{},[599],{"type":55,"value":600},"Favor GPUNetIO GDAKI when its RC connection scale is acceptable and a unified GPUNetIO CPU\u002FGPU path or verified DOCA feature is useful.",{"type":49,"tag":86,"props":602,"children":603},{},[604],{"type":55,"value":605},"Do not treat GPUNetIO design goals as proof that it outperforms IBGDA for this workload.",{"type":49,"tag":86,"props":607,"children":608},{},[609,614],{"type":49,"tag":529,"props":610,"children":611},{},[612],{"type":55,"value":613},"Proxy-friendly traffic:",{"type":55,"value":615}," Favor a proxy path for host\u002Fon-stream-dominant work, packed larger transfers, low message concurrency, or latency-sensitive isolated operations.",{"type":49,"tag":86,"props":617,"children":618},{},[619,624],{"type":49,"tag":529,"props":620,"children":621},{},[622],{"type":55,"value":623},"Mellanox proxy choice:",{"type":55,"value":625}," Use IBRC as the conservative verbs default. Recommend IBDevX only when mlx5 DevX is available, its exact operation coverage is sufficient, and a concrete requirement such as avoiding a GDRCopy-dependent path justifies it. Do not claim general IBDevX superiority.",{"type":49,"tag":86,"props":627,"children":628},{},[629,634],{"type":49,"tag":529,"props":630,"children":631},{},[632],{"type":55,"value":633},"UCX choice:",{"type":55,"value":635}," Favor UCX for a UCX-controlled deployment or a verified release-specific capability unavailable from the native alternatives. Preserve its experimental\u002Fstatus caveats from the exact documentation.",{"type":49,"tag":65,"props":637,"children":638},{},[639],{"type":55,"value":640},"If two candidates remain close, choose the compatibility-first option as the recommendation, report the other as runner-up, and state that static analysis cannot establish the performance winner.",{"type":49,"tag":162,"props":642,"children":644},{"id":643},"_6-emit-minimal-selection-configuration",[645],{"type":55,"value":646},"6. Emit Minimal Selection Configuration",{"type":49,"tag":65,"props":648,"children":649},{},[650],{"type":55,"value":651},"Emit only the variables required to select the resolved configuration:",{"type":49,"tag":653,"props":654,"children":655},"table",{},[656,674],{"type":49,"tag":657,"props":658,"children":659},"thead",{},[660],{"type":49,"tag":661,"props":662,"children":663},"tr",{},[664,669],{"type":49,"tag":665,"props":666,"children":667},"th",{},[668],{"type":55,"value":17},{"type":49,"tag":665,"props":670,"children":671},{},[672],{"type":55,"value":673},"Minimal exports",{"type":49,"tag":675,"props":676,"children":677},"tbody",{},[678,696,713,737,761,784,801,820],{"type":49,"tag":661,"props":679,"children":680},{},[681,687],{"type":49,"tag":682,"props":683,"children":684},"td",{},[685],{"type":55,"value":686},"IBRC",{"type":49,"tag":682,"props":688,"children":689},{},[690],{"type":49,"tag":129,"props":691,"children":693},{"className":692},[],[694],{"type":55,"value":695},"NVSHMEM_REMOTE_TRANSPORT=ibrc",{"type":49,"tag":661,"props":697,"children":698},{},[699,704],{"type":49,"tag":682,"props":700,"children":701},{},[702],{"type":55,"value":703},"IBDevX",{"type":49,"tag":682,"props":705,"children":706},{},[707],{"type":49,"tag":129,"props":708,"children":710},{"className":709},[],[711],{"type":55,"value":712},"NVSHMEM_REMOTE_TRANSPORT=ibdevx",{"type":49,"tag":661,"props":714,"children":715},{},[716,721],{"type":49,"tag":682,"props":717,"children":718},{},[719],{"type":55,"value":720},"IBGDA",{"type":49,"tag":682,"props":722,"children":723},{},[724,730,731],{"type":49,"tag":129,"props":725,"children":727},{"className":726},[],[728],{"type":55,"value":729},"NVSHMEM_REMOTE_TRANSPORT=\u003Cresolved-host-transport>",{"type":55,"value":99},{"type":49,"tag":129,"props":732,"children":734},{"className":733},[],[735],{"type":55,"value":736},"NVSHMEM_IB_ENABLE_IBGDA=1",{"type":49,"tag":661,"props":738,"children":739},{},[740,745],{"type":49,"tag":682,"props":741,"children":742},{},[743],{"type":55,"value":744},"GPUNetIO CPU",{"type":49,"tag":682,"props":746,"children":747},{},[748,754,755],{"type":49,"tag":129,"props":749,"children":751},{"className":750},[],[752],{"type":55,"value":753},"NVSHMEM_REMOTE_TRANSPORT=gpunetio",{"type":55,"value":99},{"type":49,"tag":129,"props":756,"children":758},{"className":757},[],[759],{"type":55,"value":760},"NVSHMEM_GPUNETIO_ENABLE_GDAKI=0",{"type":49,"tag":661,"props":762,"children":763},{},[764,769],{"type":49,"tag":682,"props":765,"children":766},{},[767],{"type":55,"value":768},"GPUNetIO GDAKI",{"type":49,"tag":682,"props":770,"children":771},{},[772,777,778],{"type":49,"tag":129,"props":773,"children":775},{"className":774},[],[776],{"type":55,"value":753},{"type":55,"value":99},{"type":49,"tag":129,"props":779,"children":781},{"className":780},[],[782],{"type":55,"value":783},"NVSHMEM_GPUNETIO_ENABLE_GDAKI=1",{"type":49,"tag":661,"props":785,"children":786},{},[787,792],{"type":49,"tag":682,"props":788,"children":789},{},[790],{"type":55,"value":791},"UCX",{"type":49,"tag":682,"props":793,"children":794},{},[795],{"type":49,"tag":129,"props":796,"children":798},{"className":797},[],[799],{"type":55,"value":800},"NVSHMEM_REMOTE_TRANSPORT=ucx",{"type":49,"tag":661,"props":802,"children":803},{},[804,809],{"type":49,"tag":682,"props":805,"children":806},{},[807],{"type":55,"value":808},"libfabric",{"type":49,"tag":682,"props":810,"children":811},{},[812,818],{"type":49,"tag":129,"props":813,"children":815},{"className":814},[],[816],{"type":55,"value":817},"NVSHMEM_REMOTE_TRANSPORT=libfabric",{"type":55,"value":819}," and `NVSHMEM_LIBFABRIC_PROVIDER=\u003Ccxi",{"type":49,"tag":661,"props":821,"children":822},{},[823,828],{"type":49,"tag":682,"props":824,"children":825},{},[826],{"type":55,"value":827},"No remote",{"type":49,"tag":682,"props":829,"children":830},{},[831],{"type":49,"tag":129,"props":832,"children":834},{"className":833},[],[835],{"type":55,"value":836},"NVSHMEM_REMOTE_TRANSPORT=none",{"type":49,"tag":65,"props":838,"children":839},{},[840],{"type":55,"value":841},"Do not add HCA mapping, QP counts, batching, NIC handlers, memory-mode changes, or provider tuning to this block. Put mandatory compatibility preconditions outside the block and route mapping\u002Ftuning to the appropriate sibling skill.",{"type":49,"tag":65,"props":843,"children":844},{},[845],{"type":55,"value":846},"Withhold an unconditional export block when version, plugin, provider, operation coverage, or kernel profile is unresolved.",{"type":49,"tag":162,"props":848,"children":850},{"id":849},"_7-return-the-recommendation",[851],{"type":55,"value":852},"7. Return the Recommendation",{"type":49,"tag":65,"props":854,"children":855},{},[856],{"type":55,"value":857},"Use this exact section order:",{"type":49,"tag":321,"props":859,"children":862},{"className":860,"code":861,"language":55,"meta":326},[324],"Recommendation: \u003Ctransport or combined host\u002FGPU configuration; conditional if necessary>\n\nSelection configuration\n\u003Cminimal export block, or why it is withheld>\n\nSystem eligibility evidence\n\u003Cversion, plugins, fabric\u002Fprovider, driver\u002Fmodule\u002Fdependency evidence>\n\nKernel-fit evidence\n\u003CAPI surface, operation shape, sizes, concurrency, fan-out, completion, scale>\n\nRunner-up\n\u003Ccandidate and the fact or tradeoff that kept it second>\n\nExcluded candidates\n\u003Ccandidate: hard gate or unresolved prerequisite>\n\nConfidence and unknowns\n\u003Chigh|medium|low; documented facts versus inference; missing evidence>\n\nStatic follow-up checks\n\u003Cread-only checks or exact facts still needed; never benchmarks>\n",[863],{"type":49,"tag":129,"props":864,"children":865},{"__ignoreMap":326},[866],{"type":55,"value":861},{"type":49,"tag":65,"props":868,"children":869},{},[870],{"type":55,"value":871},"State the documentation version consulted when live verification was used. Describe performance as expected fit, not an assured result.",{"type":49,"tag":58,"props":873,"children":875},{"id":874},"routing",[876],{"type":55,"value":877},"Routing",{"type":49,"tag":82,"props":879,"children":880},{},[881,894,906,918],{"type":49,"tag":86,"props":882,"children":883},{},[884,886,892],{"type":55,"value":885},"Route exact HCA selection and PE-to-NIC mapping to ",{"type":49,"tag":129,"props":887,"children":889},{"className":888},[],[890],{"type":55,"value":891},"$nvshmem-configure-nic-pe-mapping",{"type":55,"value":893}," when available.",{"type":49,"tag":86,"props":895,"children":896},{},[897,899,905],{"type":55,"value":898},"Route missing packages, plugins, or builds to ",{"type":49,"tag":129,"props":900,"children":902},{"className":901},[],[903],{"type":55,"value":904},"$nvshmem-install",{"type":55,"value":893},{"type":49,"tag":86,"props":907,"children":908},{},[909,911,917],{"type":55,"value":910},"Route runtime initialization failures to ",{"type":49,"tag":129,"props":912,"children":914},{"className":913},[],[915],{"type":55,"value":916},"$nvshmem-troubleshoot-and-report-bugs",{"type":55,"value":893},{"type":49,"tag":86,"props":919,"children":920},{},[921],{"type":55,"value":922},"If a sibling skill is unavailable, provide the boundary and required next evidence without inventing its behavior.",{"type":49,"tag":58,"props":924,"children":926},{"id":925},"available-script",[927],{"type":55,"value":928},"Available Script",{"type":49,"tag":653,"props":930,"children":931},{},[932,952],{"type":49,"tag":657,"props":933,"children":934},{},[935],{"type":49,"tag":661,"props":936,"children":937},{},[938,943,947],{"type":49,"tag":665,"props":939,"children":940},{},[941],{"type":55,"value":942},"Script",{"type":49,"tag":665,"props":944,"children":945},{},[946],{"type":55,"value":63},{"type":49,"tag":665,"props":948,"children":949},{},[950],{"type":55,"value":951},"Arguments",{"type":49,"tag":675,"props":953,"children":954},{},[955],{"type":49,"tag":661,"props":956,"children":957},{},[958,970,975],{"type":49,"tag":682,"props":959,"children":960},{},[961],{"type":49,"tag":92,"props":962,"children":964},{"href":963},"scripts\u002Fcollect-transport-facts.sh",[965],{"type":49,"tag":129,"props":966,"children":968},{"className":967},[],[969],{"type":55,"value":963},{"type":49,"tag":682,"props":971,"children":972},{},[973],{"type":55,"value":974},"Collect read-only NVSHMEM, GPU, kernel-module, RDMA, and provider evidence from a target compute node.",{"type":49,"tag":682,"props":976,"children":977},{},[978,980],{"type":55,"value":979},"Optional ",{"type":49,"tag":129,"props":981,"children":983},{"className":982},[],[984],{"type":55,"value":985},"--prefix PATH",{"type":49,"tag":58,"props":987,"children":989},{"id":988},"examples",[990],{"type":55,"value":991},"Examples",{"type":49,"tag":82,"props":993,"children":994},{},[995,1004],{"type":49,"tag":86,"props":996,"children":997},{},[998],{"type":49,"tag":129,"props":999,"children":1001},{"className":1000},[],[1002],{"type":55,"value":1003},"Use $nvshmem-select-remote-transport. Inspect this target compute node and src\u002Fexchange.cu, then recommend a throughput-oriented transport for 256 PEs with 8 local PEs per node.",{"type":49,"tag":86,"props":1005,"children":1006},{},[1007],{"type":49,"tag":129,"props":1008,"children":1010},{"className":1009},[],[1011],{"type":55,"value":1012},"Use $nvshmem-select-remote-transport with the attached collector output and communication profile. Recommend a compatibility-first transport for 32 PEs with 8 local PEs per node, and withhold exports if required evidence is missing.",{"type":49,"tag":58,"props":1014,"children":1016},{"id":1015},"guardrails",[1017],{"type":55,"value":1018},"Guardrails",{"type":49,"tag":82,"props":1020,"children":1021},{},[1022,1035,1040,1045,1050,1055,1060],{"type":49,"tag":86,"props":1023,"children":1024},{},[1025,1027,1033],{"type":55,"value":1026},"Do not run NVSHMEM binaries other than the read-only ",{"type":49,"tag":129,"props":1028,"children":1030},{"className":1029},[],[1031],{"type":55,"value":1032},"nvshmem-info -n",{"type":55,"value":1034}," query.",{"type":49,"tag":86,"props":1036,"children":1037},{},[1038],{"type":55,"value":1039},"Do not launch distributed jobs, allocate nodes, benchmark transports, or mutate system\u002Fapplication state.",{"type":49,"tag":86,"props":1041,"children":1042},{},[1043],{"type":55,"value":1044},"Do not recommend a plugin merely because its source exists; require evidence that the selected installation contains it.",{"type":49,"tag":86,"props":1046,"children":1047},{},[1048],{"type":55,"value":1049},"Do not convert absent evidence into compatibility.",{"type":49,"tag":86,"props":1051,"children":1052},{},[1053],{"type":55,"value":1054},"Do not claim that Ethernet link-layer evidence alone proves working RoCE.",{"type":49,"tag":86,"props":1056,"children":1057},{},[1058],{"type":55,"value":1059},"Do not let kernel heuristics override an incompatible fabric\u002Fprovider or missing prerequisite.",{"type":49,"tag":86,"props":1061,"children":1062},{},[1063],{"type":55,"value":1064},"Do not emit both IBGDA and GPUNetIO GDAKI enables.",{"type":49,"tag":58,"props":1066,"children":1068},{"id":1067},"troubleshooting",[1069],{"type":55,"value":1070},"Troubleshooting",{"type":49,"tag":82,"props":1072,"children":1073},{},[1074,1087,1098],{"type":49,"tag":86,"props":1075,"children":1076},{},[1077,1079,1085],{"type":55,"value":1078},"If the collector exits with status 2, preserve its output and resolve only the missing facts listed under ",{"type":49,"tag":129,"props":1080,"children":1082},{"className":1081},[],[1083],{"type":55,"value":1084},"diagnostics",{"type":55,"value":1086},".",{"type":49,"tag":86,"props":1088,"children":1089},{},[1090,1092,1097],{"type":55,"value":1091},"If a plugin is missing or fails to initialize, verify the selected prefix and dependencies, then route runtime failures to ",{"type":49,"tag":129,"props":1093,"children":1095},{"className":1094},[],[1096],{"type":55,"value":916},{"type":55,"value":1086},{"type":49,"tag":86,"props":1099,"children":1100},{},[1101],{"type":55,"value":1102},"If the recommendation remains conditional, obtain the exact version, provider, operation coverage, or kernel profile before emitting exports.",{"type":49,"tag":58,"props":1104,"children":1106},{"id":1105},"limitations",[1107],{"type":55,"value":1108},"Limitations",{"type":49,"tag":82,"props":1110,"children":1111},{},[1112,1117,1122,1127],{"type":49,"tag":86,"props":1113,"children":1114},{},[1115],{"type":55,"value":1116},"Static analysis estimates architectural fit; it does not validate runtime correctness, measure performance, or guarantee the fastest transport.",{"type":49,"tag":86,"props":1118,"children":1119},{},[1120],{"type":55,"value":1121},"The bundled compatibility matrix is release-sensitive; verify the exact target release against its official documentation when the matrix cannot establish applicability.",{"type":49,"tag":86,"props":1123,"children":1124},{},[1125],{"type":55,"value":1126},"The collector describes one node and one selected installation. It does not prove cluster-wide uniformity, working RoCE, or complete operation coverage.",{"type":49,"tag":86,"props":1128,"children":1129},{},[1130],{"type":55,"value":1131},"Incomplete source, topology, plugin, or provider evidence requires a conditional recommendation with no unconditional exports.",{"items":1133,"total":1291},[1134,1152,1169,1180,1192,1206,1219,1233,1246,1257,1271,1280],{"slug":1135,"name":1135,"fn":1136,"description":1137,"org":1138,"tags":1139,"stars":1149,"repoUrl":1150,"updatedAt":1151},"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},[1140,1143,1146],{"name":1141,"slug":1142,"type":15},"Documentation","documentation",{"name":1144,"slug":1145,"type":15},"MCP","mcp",{"name":1147,"slug":1148,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-08-25T03:29:57.273192",{"slug":1153,"name":1153,"fn":1154,"description":1155,"org":1156,"tags":1157,"stars":1166,"repoUrl":1167,"updatedAt":1168},"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},[1158,1161,1164],{"name":1159,"slug":1160,"type":15},"Containers","containers",{"name":1162,"slug":1163,"type":15},"Deployment","deployment",{"name":1165,"slug":30,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1170,"name":1170,"fn":1171,"description":1172,"org":1173,"tags":1174,"stars":1166,"repoUrl":1167,"updatedAt":1179},"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},[1175,1178],{"name":1176,"slug":1177,"type":15},"CI\u002FCD","ci-cd",{"name":1162,"slug":1163,"type":15},"2026-07-14T05:25:59.97109",{"slug":1181,"name":1181,"fn":1182,"description":1183,"org":1184,"tags":1185,"stars":1166,"repoUrl":1167,"updatedAt":1191},"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},[1186,1187,1188],{"name":1176,"slug":1177,"type":15},{"name":1162,"slug":1163,"type":15},{"name":1189,"slug":1190,"type":15},"GitHub","github","2026-08-28T14:38:16.959248",{"slug":1193,"name":1193,"fn":1194,"description":1195,"org":1196,"tags":1197,"stars":1166,"repoUrl":1167,"updatedAt":1205},"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},[1198,1201,1202],{"name":1199,"slug":1200,"type":15},"Debugging","debugging",{"name":1189,"slug":1190,"type":15},{"name":1203,"slug":1204,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1207,"name":1207,"fn":1208,"description":1209,"org":1210,"tags":1211,"stars":1166,"repoUrl":1167,"updatedAt":1218},"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},[1212,1215],{"name":1213,"slug":1214,"type":15},"Best Practices","best-practices",{"name":1216,"slug":1217,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1220,"name":1220,"fn":1221,"description":1222,"org":1223,"tags":1224,"stars":1166,"repoUrl":1167,"updatedAt":1232},"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, including the mechanical steps for transferring an existing pretrain_gpt.py launch script.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1225,1228,1231],{"name":1226,"slug":1227,"type":15},"Machine Learning","machine-learning",{"name":1229,"slug":1230,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-08-21T03:36:57.470256",{"slug":1234,"name":1234,"fn":1235,"description":1236,"org":1237,"tags":1238,"stars":1166,"repoUrl":1167,"updatedAt":1245},"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},[1239,1242],{"name":1240,"slug":1241,"type":15},"QA","qa",{"name":1243,"slug":1244,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1247,"name":1247,"fn":1248,"description":1249,"org":1250,"tags":1251,"stars":1166,"repoUrl":1167,"updatedAt":1256},"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},[1252,1253],{"name":1162,"slug":1163,"type":15},{"name":1254,"slug":1255,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1258,"name":1258,"fn":1259,"description":1260,"org":1261,"tags":1262,"stars":1166,"repoUrl":1167,"updatedAt":1270},"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},[1263,1266,1267],{"name":1264,"slug":1265,"type":15},"Code Review","code-review",{"name":1189,"slug":1190,"type":15},{"name":1268,"slug":1269,"type":15},"Pull Requests","pull-requests","2026-08-25T03:29:16.211287",{"slug":1272,"name":1272,"fn":1273,"description":1274,"org":1275,"tags":1276,"stars":1166,"repoUrl":1167,"updatedAt":1279},"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},[1277,1278],{"name":1240,"slug":1241,"type":15},{"name":1243,"slug":1244,"type":15},"2026-07-14T05:25:54.928983",{"slug":1281,"name":1281,"fn":1282,"description":1283,"org":1284,"tags":1285,"stars":1166,"repoUrl":1167,"updatedAt":1290},"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},[1286,1289],{"name":1287,"slug":1288,"type":15},"Automation","automation",{"name":1176,"slug":1177,"type":15},"2026-07-30T05:29:03.275638",563,{"items":1293,"total":1370},[1294,1306,1318,1330,1342,1354,1364],{"slug":1295,"name":1295,"fn":1296,"description":1297,"org":1298,"tags":1299,"stars":20,"repoUrl":21,"updatedAt":1305},"nvshmem-collect-performance-data","collect NVSHMEM performance metrics","Collect and package NVSHMEM put\u002Fget bandwidth, latency, and other perftest results with system and topology evidence for performance sanity checks.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1300,1303,1304],{"name":1301,"slug":1302,"type":15},"Monitoring","monitoring",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-08-31T09:19:00.917887",{"slug":1307,"name":1307,"fn":1308,"description":1309,"org":1310,"tags":1311,"stars":20,"repoUrl":21,"updatedAt":1317},"nvshmem-configure-nic-pe-mapping","configure NVSHMEM NIC-to-PE mappings","Recommend NVSHMEM NIC-to-PE mappings and environment exports. Use for HCA selection, multi-NIC configuration, or topology-based mapping diagnostics.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1312,1313,1316],{"name":17,"slug":18,"type":15},{"name":1314,"slug":1315,"type":15},"Networking","networking",{"name":9,"slug":8,"type":15},"2026-08-31T09:19:06.10294",{"slug":1319,"name":1319,"fn":1320,"description":1321,"org":1322,"tags":1323,"stars":20,"repoUrl":21,"updatedAt":1329},"nvshmem-docs","fetch official NVSHMEM documentation","Find version-aware official NVSHMEM and NVSHMEM4Py documentation for releases, installation, APIs, runtime settings, transports, containers, and troubleshooting.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1324,1325,1326],{"name":1141,"slug":1142,"type":15},{"name":9,"slug":8,"type":15},{"name":1327,"slug":1328,"type":15},"Reference","reference","2026-08-31T09:19:00.214841",{"slug":1331,"name":1331,"fn":1332,"description":1333,"org":1334,"tags":1335,"stars":20,"repoUrl":21,"updatedAt":1341},"nvshmem-enable-tma","optimize NVSHMEM CUDA kernels for TMA","Prepare or review NVSHMEM CUDA kernels for TMA SMEM registration and direct-SMEM transfers. Do not use for unrelated CUDA tuning.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1336,1339,1340],{"name":1337,"slug":1338,"type":15},"Engineering","engineering",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-08-31T09:18:50.449676",{"slug":1343,"name":1343,"fn":1344,"description":1345,"org":1346,"tags":1347,"stars":20,"repoUrl":21,"updatedAt":1353},"nvshmem-get-started","onboard developers to NVSHMEM","Guide NVSHMEM beginners through fit assessment, mental models, first C\u002FC++ or Python NVSHMEM programs, compilation, launching, and next steps. Use for onboarding.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1348,1349,1350],{"name":1141,"slug":1142,"type":15},{"name":9,"slug":8,"type":15},{"name":1351,"slug":1352,"type":15},"Onboarding","onboarding","2026-08-31T09:18:55.48573",{"slug":1355,"name":1355,"fn":1356,"description":1357,"org":1358,"tags":1359,"stars":20,"repoUrl":21,"updatedAt":1363},"nvshmem-install","install and validate NVSHMEM","Plan and validate NVSHMEM and NVSHMEM4Py installations. Use for package, container, or source deployments.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1360,1361,1362],{"name":1162,"slug":1163,"type":15},{"name":1337,"slug":1338,"type":15},{"name":9,"slug":8,"type":15},"2026-08-31T09:18:50.105521",{"slug":4,"name":4,"fn":5,"description":6,"org":1365,"tags":1366,"stars":20,"repoUrl":21,"updatedAt":22},{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1367,1368,1369],{"name":17,"slug":18,"type":15},{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},9]