[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-mbridge-perf-moe-hardware-configs":3,"mdc--af5mhx-key":34,"related-repo-nvidia-nemo-mbridge-perf-moe-hardware-configs":1027,"related-org-nvidia-nemo-mbridge-perf-moe-hardware-configs":1131},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":29,"sourceUrl":32,"mdContent":33},"nemo-mbridge-perf-moe-hardware-configs","optimize MoE training hardware configurations","Representative MoE training playbooks by hardware platform and model family. Summarizes rounded throughput bands, parallelism patterns, and common tuning stacks.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,22],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":17,"slug":18,"type":15},"Benchmarking","benchmarking",{"name":20,"slug":21,"type":15},"Deep Learning","deep-learning",{"name":9,"slug":8,"type":15},2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:30:46.571588","Apache-2.0",281,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fnemo-mbridge-perf-moe-hardware-configs","---\nname: nemo-mbridge-perf-moe-hardware-configs\ndescription: Representative MoE training playbooks by hardware platform and model family. Summarizes rounded throughput bands, parallelism patterns, and common tuning stacks.\nlicense: Apache-2.0\nwhen_to_use: Hardware-specific MoE playbooks or throughput estimates; 'MoE on H100', 'GB200 config', 'expected throughput', 'MoE hardware playbook', 'parallelism for B200'.\n---\n\n# MoE Hardware Configuration Reference\n\nStable docs: @docs\u002Ftraining\u002Fmoe-optimization.md\nCard: @skills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002Fcard.yaml\n\n## Quick Platform Playbook\n\n| Platform | Typical MoE strategy | What usually matters most |\n|---|---|---|\n| H100 | DeepEP + stronger PP + moderate TP | communication overlap and PP efficiency |\n| B200 | DeepEP + MXFP8 + careful PP layout | container quality and tuned comm settings |\n| GB200 | HybridEP + partial CUDA graphs + CPU cleanup | host overhead, topology-aware dispatch, memory headroom |\n| GB300 | HybridEP + newer FP8 and kernel stack | same GB200 playbook, usually with a higher ceiling |\n\n## First Answer Checklist\n\nFor hardware playbook questions, answer from these canonical rows before adding\nthroughput caveats:\n\n| Workload | Hardware | Dispatcher | Layout |\n|---|---|---|---|\n| DSV3 | H100 | DeepEP | TP=2, EP=64, PP=8, VPP=4 |\n| DSV3 | GB200\u002FGB300 | HybridEP | TP=1, EP=64, PP=4, VPP=4 |\n| Qwen3 235B | H100 | DeepEP | TP=2, EP=32, PP=8, VPP=4 |\n| Qwen3 235B | GB200 | HybridEP | TP=1 or 2, EP=32-64, PP=4, VPP=unspecified |\n\nFor Qwen3 235B on GB200, explicitly say `VPP=unspecified`; do not invent or\nextrapolate `VPP=12` unless a measured row provides it. Include TE-scoped CUDA\ngraph scopes (`attn`, `moe_router`, `moe_preprocess`),\n`CUDA_DEVICE_MAX_CONNECTIONS` selection,\n`PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True`, `NCCL_GRAPH_REGISTER=0`,\nGB200\u002FGB300 CPU-side tuning, and the warning not to cargo-cult tracker rows.\n\n## Rounded Performance Bands\n\nThese are intentionally rounded so the document stays durable as the tracker\nmoves. Treat them as planning ranges, not exact promises.\n\n| Workload family | Hardware | Typical band | Representative shape |\n|---|---|---|---|\n| DSV3, large-scale | H100 | low-to-mid hundreds TFLOPS\u002FGPU, high-teens MFU | TP2, EP64, PP8, DeepEP |\n| DSV3, large-scale | B200 | high-hundreds TFLOPS\u002FGPU, mid-teens MFU | TP1, EP32, PP8, DeepEP |\n| DSV3, large-scale | GB200 | around 1K TFLOPS\u002FGPU, low-20s MFU | TP1, EP64, PP4, HybridEP |\n| DSV3, large-scale | GB300 | above the GB200 band, often mid-20s MFU | TP1, EP64, PP4, HybridEP |\n| Qwen3 235B | H100 | low-300s TFLOPS\u002FGPU, around 30% MFU | TP2, EP32, PP8, DeepEP |\n| Qwen3 235B | GB200 | high-hundreds TFLOPS\u002FGPU in tuned runs | TP1 or TP2, EP32-64, PP4, HybridEP |\n| Qwen3 30B | H100 | low-200s TFLOPS\u002FGPU | TP1, EP8, PP1, DeepEP |\n| Qwen3-Next 80B | GB200 | low-300s TFLOPS\u002FGPU in BF16-class runs | TP1, EP32, PP2, HybridEP |\n\n## Representative Config Families\n\n### DSV3 on H100\n\n```text\nDispatcher: DeepEP\nTP=2  EP=64  PP=8  VPP=4\nRouting: force balance\nRecompute: light-to-moderate selective recompute\nPriority: overlap communication and keep PP efficient\n```\n\n### DSV3 on B200\n\n```text\nDispatcher: DeepEP\nTP=1  EP=32  PP=8  VPP=2 or similar\nPrecision: MXFP8-class\nRecompute: selective recompute around MLA up-projection and MLP-side modules\nPriority: container quality, PP layout, and DeepEP SMS tuning\n```\n\n### DSV3 on GB200 or GB300\n\n```text\nDispatcher: HybridEP\nTP=1  EP=64  PP=4  VPP=4\nPrecision: MXFP8-class\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: HybridEP, CPU optimization, and graph-friendly static shapes\n```\n\n### Qwen3 235B on H100\n\n```text\nDispatcher: DeepEP\nTP=2  EP=32  PP=8  VPP=4\nRecompute: norm and activation-side selective recompute\nPriority: communication overlap and router-path cleanup\n```\n\n### Qwen3 235B on GB200\n\n```text\nDispatcher: HybridEP\nTP=1 or 2  EP=32 to 64  PP=4  VPP=unspecified unless measured\nCUDA Graph: attn + moe_router + moe_preprocess\nRecompute: moe_act, mlp, or norm depending on memory pressure\nPriority: balance throughput against memory headroom\n```\n\n### Qwen3-Next 80B on GB200\n\n```text\nDispatcher: HybridEP\nTP=1  EP=32  PP=2  VPP around 4\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: pipeline layout and grouped GEMM quality\n```\n\n## Cross-Cutting Patterns\n\n### PP layout\n\n- `E` = embedding\n- `t` = transformer\n- `m` = MTP\n- `L` = loss\n- `|` = stage boundary\n\nThe biggest platform difference is usually not just the dispatcher. It is the\ncombination of dispatcher, PP shape, and whether VPP keeps each stage balanced.\n\n### Recompute strategy\n\n| Memory pressure | Starting point |\n|---|---|\n| low | none or a very narrow selective set |\n| moderate | `moe_act`, `mlp`, `norm`, or similar selective modules |\n| high | model-specific up-projection plus selective MoE and MLP modules |\n| extreme or long-context | full recompute only if the selective path still does not fit |\n\n### Environment variables\n\n```bash\nCUDA_DEVICE_MAX_CONNECTIONS=1\nCUDA_DEVICE_MAX_CONNECTIONS=32   # common when EP overlap and CUDA graphs are combined\nPYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\nNCCL_GRAPH_REGISTER=0\n```\n\n### CPU-side tuning\n\nOn GB200 and GB300, CPU affinity and general host-overhead cleanup can move the\nneedle almost as much as a dispatcher swap. Treat them as first-class tuning\nwork, not as afterthoughts.\n\n## Pitfalls\n\n1. **Do not cargo-cult a tracker row**: the winning config usually depends on\n   routing mode, container, and PP layout as much as on hardware name.\n\n2. **Container quality matters**: large regressions can come from the software\n   stack rather than the model recipe.\n\n3. **VPP must be intentional**: a bad VPP split can erase the gain from a better\n   dispatcher.\n\n4. **Compare absolute throughput, not only MFU**: MFU can mislead when switching\n   between BF16, FP8, and other precision modes.\n\n5. **Force-balance routing is the safer benchmark default**: keep routing mode\n   fixed when comparing hardware or dispatcher stacks.\n",{"data":35,"body":37},{"name":4,"description":6,"license":26,"when_to_use":36},"Hardware-specific MoE playbooks or throughput estimates; 'MoE on H100', 'GB200 config', 'expected throughput', 'MoE hardware playbook', 'parallelism for B200'.",{"type":38,"children":39},"root",[40,49,55,62,167,173,178,295,363,369,374,575,581,588,600,606,615,621,630,636,645,651,660,666,675,681,687,747,752,758,852,858,949,955,960,966,1021],{"type":41,"tag":42,"props":43,"children":45},"element","h1",{"id":44},"moe-hardware-configuration-reference",[46],{"type":47,"value":48},"text","MoE Hardware Configuration Reference",{"type":41,"tag":50,"props":51,"children":52},"p",{},[53],{"type":47,"value":54},"Stable docs: @docs\u002Ftraining\u002Fmoe-optimization.md\nCard: @skills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002Fcard.yaml",{"type":41,"tag":56,"props":57,"children":59},"h2",{"id":58},"quick-platform-playbook",[60],{"type":47,"value":61},"Quick Platform Playbook",{"type":41,"tag":63,"props":64,"children":65},"table",{},[66,90],{"type":41,"tag":67,"props":68,"children":69},"thead",{},[70],{"type":41,"tag":71,"props":72,"children":73},"tr",{},[74,80,85],{"type":41,"tag":75,"props":76,"children":77},"th",{},[78],{"type":47,"value":79},"Platform",{"type":41,"tag":75,"props":81,"children":82},{},[83],{"type":47,"value":84},"Typical MoE strategy",{"type":41,"tag":75,"props":86,"children":87},{},[88],{"type":47,"value":89},"What usually matters most",{"type":41,"tag":91,"props":92,"children":93},"tbody",{},[94,113,131,149],{"type":41,"tag":71,"props":95,"children":96},{},[97,103,108],{"type":41,"tag":98,"props":99,"children":100},"td",{},[101],{"type":47,"value":102},"H100",{"type":41,"tag":98,"props":104,"children":105},{},[106],{"type":47,"value":107},"DeepEP + stronger PP + moderate TP",{"type":41,"tag":98,"props":109,"children":110},{},[111],{"type":47,"value":112},"communication overlap and PP efficiency",{"type":41,"tag":71,"props":114,"children":115},{},[116,121,126],{"type":41,"tag":98,"props":117,"children":118},{},[119],{"type":47,"value":120},"B200",{"type":41,"tag":98,"props":122,"children":123},{},[124],{"type":47,"value":125},"DeepEP + MXFP8 + careful PP layout",{"type":41,"tag":98,"props":127,"children":128},{},[129],{"type":47,"value":130},"container quality and tuned comm settings",{"type":41,"tag":71,"props":132,"children":133},{},[134,139,144],{"type":41,"tag":98,"props":135,"children":136},{},[137],{"type":47,"value":138},"GB200",{"type":41,"tag":98,"props":140,"children":141},{},[142],{"type":47,"value":143},"HybridEP + partial CUDA graphs + CPU cleanup",{"type":41,"tag":98,"props":145,"children":146},{},[147],{"type":47,"value":148},"host overhead, topology-aware dispatch, memory headroom",{"type":41,"tag":71,"props":150,"children":151},{},[152,157,162],{"type":41,"tag":98,"props":153,"children":154},{},[155],{"type":47,"value":156},"GB300",{"type":41,"tag":98,"props":158,"children":159},{},[160],{"type":47,"value":161},"HybridEP + newer FP8 and kernel stack",{"type":41,"tag":98,"props":163,"children":164},{},[165],{"type":47,"value":166},"same GB200 playbook, usually with a higher ceiling",{"type":41,"tag":56,"props":168,"children":170},{"id":169},"first-answer-checklist",[171],{"type":47,"value":172},"First Answer Checklist",{"type":41,"tag":50,"props":174,"children":175},{},[176],{"type":47,"value":177},"For hardware playbook questions, answer from these canonical rows before adding\nthroughput caveats:",{"type":41,"tag":63,"props":179,"children":180},{},[181,207],{"type":41,"tag":67,"props":182,"children":183},{},[184],{"type":41,"tag":71,"props":185,"children":186},{},[187,192,197,202],{"type":41,"tag":75,"props":188,"children":189},{},[190],{"type":47,"value":191},"Workload",{"type":41,"tag":75,"props":193,"children":194},{},[195],{"type":47,"value":196},"Hardware",{"type":41,"tag":75,"props":198,"children":199},{},[200],{"type":47,"value":201},"Dispatcher",{"type":41,"tag":75,"props":203,"children":204},{},[205],{"type":47,"value":206},"Layout",{"type":41,"tag":91,"props":208,"children":209},{},[210,232,254,275],{"type":41,"tag":71,"props":211,"children":212},{},[213,218,222,227],{"type":41,"tag":98,"props":214,"children":215},{},[216],{"type":47,"value":217},"DSV3",{"type":41,"tag":98,"props":219,"children":220},{},[221],{"type":47,"value":102},{"type":41,"tag":98,"props":223,"children":224},{},[225],{"type":47,"value":226},"DeepEP",{"type":41,"tag":98,"props":228,"children":229},{},[230],{"type":47,"value":231},"TP=2, EP=64, PP=8, VPP=4",{"type":41,"tag":71,"props":233,"children":234},{},[235,239,244,249],{"type":41,"tag":98,"props":236,"children":237},{},[238],{"type":47,"value":217},{"type":41,"tag":98,"props":240,"children":241},{},[242],{"type":47,"value":243},"GB200\u002FGB300",{"type":41,"tag":98,"props":245,"children":246},{},[247],{"type":47,"value":248},"HybridEP",{"type":41,"tag":98,"props":250,"children":251},{},[252],{"type":47,"value":253},"TP=1, EP=64, PP=4, VPP=4",{"type":41,"tag":71,"props":255,"children":256},{},[257,262,266,270],{"type":41,"tag":98,"props":258,"children":259},{},[260],{"type":47,"value":261},"Qwen3 235B",{"type":41,"tag":98,"props":263,"children":264},{},[265],{"type":47,"value":102},{"type":41,"tag":98,"props":267,"children":268},{},[269],{"type":47,"value":226},{"type":41,"tag":98,"props":271,"children":272},{},[273],{"type":47,"value":274},"TP=2, EP=32, PP=8, VPP=4",{"type":41,"tag":71,"props":276,"children":277},{},[278,282,286,290],{"type":41,"tag":98,"props":279,"children":280},{},[281],{"type":47,"value":261},{"type":41,"tag":98,"props":283,"children":284},{},[285],{"type":47,"value":138},{"type":41,"tag":98,"props":287,"children":288},{},[289],{"type":47,"value":248},{"type":41,"tag":98,"props":291,"children":292},{},[293],{"type":47,"value":294},"TP=1 or 2, EP=32-64, PP=4, VPP=unspecified",{"type":41,"tag":50,"props":296,"children":297},{},[298,300,307,309,315,317,323,325,331,332,338,340,346,348,354,355,361],{"type":47,"value":299},"For Qwen3 235B on GB200, explicitly say ",{"type":41,"tag":301,"props":302,"children":304},"code",{"className":303},[],[305],{"type":47,"value":306},"VPP=unspecified",{"type":47,"value":308},"; do not invent or\nextrapolate ",{"type":41,"tag":301,"props":310,"children":312},{"className":311},[],[313],{"type":47,"value":314},"VPP=12",{"type":47,"value":316}," unless a measured row provides it. Include TE-scoped CUDA\ngraph scopes (",{"type":41,"tag":301,"props":318,"children":320},{"className":319},[],[321],{"type":47,"value":322},"attn",{"type":47,"value":324},", ",{"type":41,"tag":301,"props":326,"children":328},{"className":327},[],[329],{"type":47,"value":330},"moe_router",{"type":47,"value":324},{"type":41,"tag":301,"props":333,"children":335},{"className":334},[],[336],{"type":47,"value":337},"moe_preprocess",{"type":47,"value":339},"),\n",{"type":41,"tag":301,"props":341,"children":343},{"className":342},[],[344],{"type":47,"value":345},"CUDA_DEVICE_MAX_CONNECTIONS",{"type":47,"value":347}," selection,\n",{"type":41,"tag":301,"props":349,"children":351},{"className":350},[],[352],{"type":47,"value":353},"PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True",{"type":47,"value":324},{"type":41,"tag":301,"props":356,"children":358},{"className":357},[],[359],{"type":47,"value":360},"NCCL_GRAPH_REGISTER=0",{"type":47,"value":362},",\nGB200\u002FGB300 CPU-side tuning, and the warning not to cargo-cult tracker rows.",{"type":41,"tag":56,"props":364,"children":366},{"id":365},"rounded-performance-bands",[367],{"type":47,"value":368},"Rounded Performance Bands",{"type":41,"tag":50,"props":370,"children":371},{},[372],{"type":47,"value":373},"These are intentionally rounded so the document stays durable as the tracker\nmoves. Treat them as planning ranges, not exact promises.",{"type":41,"tag":63,"props":375,"children":376},{},[377,402],{"type":41,"tag":67,"props":378,"children":379},{},[380],{"type":41,"tag":71,"props":381,"children":382},{},[383,388,392,397],{"type":41,"tag":75,"props":384,"children":385},{},[386],{"type":47,"value":387},"Workload family",{"type":41,"tag":75,"props":389,"children":390},{},[391],{"type":47,"value":196},{"type":41,"tag":75,"props":393,"children":394},{},[395],{"type":47,"value":396},"Typical band",{"type":41,"tag":75,"props":398,"children":399},{},[400],{"type":47,"value":401},"Representative shape",{"type":41,"tag":91,"props":403,"children":404},{},[405,427,448,469,489,510,531,553],{"type":41,"tag":71,"props":406,"children":407},{},[408,413,417,422],{"type":41,"tag":98,"props":409,"children":410},{},[411],{"type":47,"value":412},"DSV3, large-scale",{"type":41,"tag":98,"props":414,"children":415},{},[416],{"type":47,"value":102},{"type":41,"tag":98,"props":418,"children":419},{},[420],{"type":47,"value":421},"low-to-mid hundreds TFLOPS\u002FGPU, high-teens MFU",{"type":41,"tag":98,"props":423,"children":424},{},[425],{"type":47,"value":426},"TP2, EP64, PP8, DeepEP",{"type":41,"tag":71,"props":428,"children":429},{},[430,434,438,443],{"type":41,"tag":98,"props":431,"children":432},{},[433],{"type":47,"value":412},{"type":41,"tag":98,"props":435,"children":436},{},[437],{"type":47,"value":120},{"type":41,"tag":98,"props":439,"children":440},{},[441],{"type":47,"value":442},"high-hundreds TFLOPS\u002FGPU, mid-teens MFU",{"type":41,"tag":98,"props":444,"children":445},{},[446],{"type":47,"value":447},"TP1, EP32, PP8, DeepEP",{"type":41,"tag":71,"props":449,"children":450},{},[451,455,459,464],{"type":41,"tag":98,"props":452,"children":453},{},[454],{"type":47,"value":412},{"type":41,"tag":98,"props":456,"children":457},{},[458],{"type":47,"value":138},{"type":41,"tag":98,"props":460,"children":461},{},[462],{"type":47,"value":463},"around 1K TFLOPS\u002FGPU, low-20s MFU",{"type":41,"tag":98,"props":465,"children":466},{},[467],{"type":47,"value":468},"TP1, EP64, PP4, HybridEP",{"type":41,"tag":71,"props":470,"children":471},{},[472,476,480,485],{"type":41,"tag":98,"props":473,"children":474},{},[475],{"type":47,"value":412},{"type":41,"tag":98,"props":477,"children":478},{},[479],{"type":47,"value":156},{"type":41,"tag":98,"props":481,"children":482},{},[483],{"type":47,"value":484},"above the GB200 band, often mid-20s MFU",{"type":41,"tag":98,"props":486,"children":487},{},[488],{"type":47,"value":468},{"type":41,"tag":71,"props":490,"children":491},{},[492,496,500,505],{"type":41,"tag":98,"props":493,"children":494},{},[495],{"type":47,"value":261},{"type":41,"tag":98,"props":497,"children":498},{},[499],{"type":47,"value":102},{"type":41,"tag":98,"props":501,"children":502},{},[503],{"type":47,"value":504},"low-300s TFLOPS\u002FGPU, around 30% MFU",{"type":41,"tag":98,"props":506,"children":507},{},[508],{"type":47,"value":509},"TP2, EP32, PP8, DeepEP",{"type":41,"tag":71,"props":511,"children":512},{},[513,517,521,526],{"type":41,"tag":98,"props":514,"children":515},{},[516],{"type":47,"value":261},{"type":41,"tag":98,"props":518,"children":519},{},[520],{"type":47,"value":138},{"type":41,"tag":98,"props":522,"children":523},{},[524],{"type":47,"value":525},"high-hundreds TFLOPS\u002FGPU in tuned runs",{"type":41,"tag":98,"props":527,"children":528},{},[529],{"type":47,"value":530},"TP1 or TP2, EP32-64, PP4, HybridEP",{"type":41,"tag":71,"props":532,"children":533},{},[534,539,543,548],{"type":41,"tag":98,"props":535,"children":536},{},[537],{"type":47,"value":538},"Qwen3 30B",{"type":41,"tag":98,"props":540,"children":541},{},[542],{"type":47,"value":102},{"type":41,"tag":98,"props":544,"children":545},{},[546],{"type":47,"value":547},"low-200s TFLOPS\u002FGPU",{"type":41,"tag":98,"props":549,"children":550},{},[551],{"type":47,"value":552},"TP1, EP8, PP1, DeepEP",{"type":41,"tag":71,"props":554,"children":555},{},[556,561,565,570],{"type":41,"tag":98,"props":557,"children":558},{},[559],{"type":47,"value":560},"Qwen3-Next 80B",{"type":41,"tag":98,"props":562,"children":563},{},[564],{"type":47,"value":138},{"type":41,"tag":98,"props":566,"children":567},{},[568],{"type":47,"value":569},"low-300s TFLOPS\u002FGPU in BF16-class runs",{"type":41,"tag":98,"props":571,"children":572},{},[573],{"type":47,"value":574},"TP1, EP32, PP2, HybridEP",{"type":41,"tag":56,"props":576,"children":578},{"id":577},"representative-config-families",[579],{"type":47,"value":580},"Representative Config Families",{"type":41,"tag":582,"props":583,"children":585},"h3",{"id":584},"dsv3-on-h100",[586],{"type":47,"value":587},"DSV3 on H100",{"type":41,"tag":589,"props":590,"children":595},"pre",{"className":591,"code":593,"language":47,"meta":594},[592],"language-text","Dispatcher: DeepEP\nTP=2  EP=64  PP=8  VPP=4\nRouting: force balance\nRecompute: light-to-moderate selective recompute\nPriority: overlap communication and keep PP efficient\n","",[596],{"type":41,"tag":301,"props":597,"children":598},{"__ignoreMap":594},[599],{"type":47,"value":593},{"type":41,"tag":582,"props":601,"children":603},{"id":602},"dsv3-on-b200",[604],{"type":47,"value":605},"DSV3 on B200",{"type":41,"tag":589,"props":607,"children":610},{"className":608,"code":609,"language":47,"meta":594},[592],"Dispatcher: DeepEP\nTP=1  EP=32  PP=8  VPP=2 or similar\nPrecision: MXFP8-class\nRecompute: selective recompute around MLA up-projection and MLP-side modules\nPriority: container quality, PP layout, and DeepEP SMS tuning\n",[611],{"type":41,"tag":301,"props":612,"children":613},{"__ignoreMap":594},[614],{"type":47,"value":609},{"type":41,"tag":582,"props":616,"children":618},{"id":617},"dsv3-on-gb200-or-gb300",[619],{"type":47,"value":620},"DSV3 on GB200 or GB300",{"type":41,"tag":589,"props":622,"children":625},{"className":623,"code":624,"language":47,"meta":594},[592],"Dispatcher: HybridEP\nTP=1  EP=64  PP=4  VPP=4\nPrecision: MXFP8-class\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: HybridEP, CPU optimization, and graph-friendly static shapes\n",[626],{"type":41,"tag":301,"props":627,"children":628},{"__ignoreMap":594},[629],{"type":47,"value":624},{"type":41,"tag":582,"props":631,"children":633},{"id":632},"qwen3-235b-on-h100",[634],{"type":47,"value":635},"Qwen3 235B on H100",{"type":41,"tag":589,"props":637,"children":640},{"className":638,"code":639,"language":47,"meta":594},[592],"Dispatcher: DeepEP\nTP=2  EP=32  PP=8  VPP=4\nRecompute: norm and activation-side selective recompute\nPriority: communication overlap and router-path cleanup\n",[641],{"type":41,"tag":301,"props":642,"children":643},{"__ignoreMap":594},[644],{"type":47,"value":639},{"type":41,"tag":582,"props":646,"children":648},{"id":647},"qwen3-235b-on-gb200",[649],{"type":47,"value":650},"Qwen3 235B on GB200",{"type":41,"tag":589,"props":652,"children":655},{"className":653,"code":654,"language":47,"meta":594},[592],"Dispatcher: HybridEP\nTP=1 or 2  EP=32 to 64  PP=4  VPP=unspecified unless measured\nCUDA Graph: attn + moe_router + moe_preprocess\nRecompute: moe_act, mlp, or norm depending on memory pressure\nPriority: balance throughput against memory headroom\n",[656],{"type":41,"tag":301,"props":657,"children":658},{"__ignoreMap":594},[659],{"type":47,"value":654},{"type":41,"tag":582,"props":661,"children":663},{"id":662},"qwen3-next-80b-on-gb200",[664],{"type":47,"value":665},"Qwen3-Next 80B on GB200",{"type":41,"tag":589,"props":667,"children":670},{"className":668,"code":669,"language":47,"meta":594},[592],"Dispatcher: HybridEP\nTP=1  EP=32  PP=2  VPP around 4\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: pipeline layout and grouped GEMM quality\n",[671],{"type":41,"tag":301,"props":672,"children":673},{"__ignoreMap":594},[674],{"type":47,"value":669},{"type":41,"tag":56,"props":676,"children":678},{"id":677},"cross-cutting-patterns",[679],{"type":47,"value":680},"Cross-Cutting Patterns",{"type":41,"tag":582,"props":682,"children":684},{"id":683},"pp-layout",[685],{"type":47,"value":686},"PP layout",{"type":41,"tag":688,"props":689,"children":690},"ul",{},[691,703,714,725,736],{"type":41,"tag":692,"props":693,"children":694},"li",{},[695,701],{"type":41,"tag":301,"props":696,"children":698},{"className":697},[],[699],{"type":47,"value":700},"E",{"type":47,"value":702}," = embedding",{"type":41,"tag":692,"props":704,"children":705},{},[706,712],{"type":41,"tag":301,"props":707,"children":709},{"className":708},[],[710],{"type":47,"value":711},"t",{"type":47,"value":713}," = transformer",{"type":41,"tag":692,"props":715,"children":716},{},[717,723],{"type":41,"tag":301,"props":718,"children":720},{"className":719},[],[721],{"type":47,"value":722},"m",{"type":47,"value":724}," = MTP",{"type":41,"tag":692,"props":726,"children":727},{},[728,734],{"type":41,"tag":301,"props":729,"children":731},{"className":730},[],[732],{"type":47,"value":733},"L",{"type":47,"value":735}," = loss",{"type":41,"tag":692,"props":737,"children":738},{},[739,745],{"type":41,"tag":301,"props":740,"children":742},{"className":741},[],[743],{"type":47,"value":744},"|",{"type":47,"value":746}," = stage boundary",{"type":41,"tag":50,"props":748,"children":749},{},[750],{"type":47,"value":751},"The biggest platform difference is usually not just the dispatcher. It is the\ncombination of dispatcher, PP shape, and whether VPP keeps each stage balanced.",{"type":41,"tag":582,"props":753,"children":755},{"id":754},"recompute-strategy",[756],{"type":47,"value":757},"Recompute strategy",{"type":41,"tag":63,"props":759,"children":760},{},[761,777],{"type":41,"tag":67,"props":762,"children":763},{},[764],{"type":41,"tag":71,"props":765,"children":766},{},[767,772],{"type":41,"tag":75,"props":768,"children":769},{},[770],{"type":47,"value":771},"Memory pressure",{"type":41,"tag":75,"props":773,"children":774},{},[775],{"type":47,"value":776},"Starting point",{"type":41,"tag":91,"props":778,"children":779},{},[780,793,826,839],{"type":41,"tag":71,"props":781,"children":782},{},[783,788],{"type":41,"tag":98,"props":784,"children":785},{},[786],{"type":47,"value":787},"low",{"type":41,"tag":98,"props":789,"children":790},{},[791],{"type":47,"value":792},"none or a very narrow selective set",{"type":41,"tag":71,"props":794,"children":795},{},[796,801],{"type":41,"tag":98,"props":797,"children":798},{},[799],{"type":47,"value":800},"moderate",{"type":41,"tag":98,"props":802,"children":803},{},[804,810,811,817,818,824],{"type":41,"tag":301,"props":805,"children":807},{"className":806},[],[808],{"type":47,"value":809},"moe_act",{"type":47,"value":324},{"type":41,"tag":301,"props":812,"children":814},{"className":813},[],[815],{"type":47,"value":816},"mlp",{"type":47,"value":324},{"type":41,"tag":301,"props":819,"children":821},{"className":820},[],[822],{"type":47,"value":823},"norm",{"type":47,"value":825},", or similar selective modules",{"type":41,"tag":71,"props":827,"children":828},{},[829,834],{"type":41,"tag":98,"props":830,"children":831},{},[832],{"type":47,"value":833},"high",{"type":41,"tag":98,"props":835,"children":836},{},[837],{"type":47,"value":838},"model-specific up-projection plus selective MoE and MLP modules",{"type":41,"tag":71,"props":840,"children":841},{},[842,847],{"type":41,"tag":98,"props":843,"children":844},{},[845],{"type":47,"value":846},"extreme or long-context",{"type":41,"tag":98,"props":848,"children":849},{},[850],{"type":47,"value":851},"full recompute only if the selective path still does not fit",{"type":41,"tag":582,"props":853,"children":855},{"id":854},"environment-variables",[856],{"type":47,"value":857},"Environment variables",{"type":41,"tag":589,"props":859,"children":863},{"className":860,"code":861,"language":862,"meta":594,"style":594},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","CUDA_DEVICE_MAX_CONNECTIONS=1\nCUDA_DEVICE_MAX_CONNECTIONS=32   # common when EP overlap and CUDA graphs are combined\nPYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\nNCCL_GRAPH_REGISTER=0\n","bash",[864],{"type":41,"tag":301,"props":865,"children":866},{"__ignoreMap":594},[867,890,913,931],{"type":41,"tag":868,"props":869,"children":872},"span",{"class":870,"line":871},"line",1,[873,878,884],{"type":41,"tag":868,"props":874,"children":876},{"style":875},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[877],{"type":47,"value":345},{"type":41,"tag":868,"props":879,"children":881},{"style":880},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[882],{"type":47,"value":883},"=",{"type":41,"tag":868,"props":885,"children":887},{"style":886},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[888],{"type":47,"value":889},"1\n",{"type":41,"tag":868,"props":891,"children":893},{"class":870,"line":892},2,[894,898,902,907],{"type":41,"tag":868,"props":895,"children":896},{"style":875},[897],{"type":47,"value":345},{"type":41,"tag":868,"props":899,"children":900},{"style":880},[901],{"type":47,"value":883},{"type":41,"tag":868,"props":903,"children":904},{"style":886},[905],{"type":47,"value":906},"32",{"type":41,"tag":868,"props":908,"children":910},{"style":909},"--shiki-light:#90A4AE;--shiki-light-font-style:italic;--shiki-default:#546E7A;--shiki-default-font-style:italic;--shiki-dark:#676E95;--shiki-dark-font-style:italic",[911],{"type":47,"value":912},"   # common when EP overlap and CUDA graphs are combined\n",{"type":41,"tag":868,"props":914,"children":916},{"class":870,"line":915},3,[917,922,926],{"type":41,"tag":868,"props":918,"children":919},{"style":875},[920],{"type":47,"value":921},"PYTORCH_CUDA_ALLOC_CONF",{"type":41,"tag":868,"props":923,"children":924},{"style":880},[925],{"type":47,"value":883},{"type":41,"tag":868,"props":927,"children":928},{"style":886},[929],{"type":47,"value":930},"expandable_segments:True\n",{"type":41,"tag":868,"props":932,"children":934},{"class":870,"line":933},4,[935,940,944],{"type":41,"tag":868,"props":936,"children":937},{"style":875},[938],{"type":47,"value":939},"NCCL_GRAPH_REGISTER",{"type":41,"tag":868,"props":941,"children":942},{"style":880},[943],{"type":47,"value":883},{"type":41,"tag":868,"props":945,"children":946},{"style":886},[947],{"type":47,"value":948},"0\n",{"type":41,"tag":582,"props":950,"children":952},{"id":951},"cpu-side-tuning",[953],{"type":47,"value":954},"CPU-side tuning",{"type":41,"tag":50,"props":956,"children":957},{},[958],{"type":47,"value":959},"On GB200 and GB300, CPU affinity and general host-overhead cleanup can move the\nneedle almost as much as a dispatcher swap. Treat them as first-class tuning\nwork, not as afterthoughts.",{"type":41,"tag":56,"props":961,"children":963},{"id":962},"pitfalls",[964],{"type":47,"value":965},"Pitfalls",{"type":41,"tag":967,"props":968,"children":969},"ol",{},[970,981,991,1001,1011],{"type":41,"tag":692,"props":971,"children":972},{},[973,979],{"type":41,"tag":974,"props":975,"children":976},"strong",{},[977],{"type":47,"value":978},"Do not cargo-cult a tracker row",{"type":47,"value":980},": the winning config usually depends on\nrouting mode, container, and PP layout as much as on hardware name.",{"type":41,"tag":692,"props":982,"children":983},{},[984,989],{"type":41,"tag":974,"props":985,"children":986},{},[987],{"type":47,"value":988},"Container quality matters",{"type":47,"value":990},": large regressions can come from the software\nstack rather than the model recipe.",{"type":41,"tag":692,"props":992,"children":993},{},[994,999],{"type":41,"tag":974,"props":995,"children":996},{},[997],{"type":47,"value":998},"VPP must be intentional",{"type":47,"value":1000},": a bad VPP split can erase the gain from a better\ndispatcher.",{"type":41,"tag":692,"props":1002,"children":1003},{},[1004,1009],{"type":41,"tag":974,"props":1005,"children":1006},{},[1007],{"type":47,"value":1008},"Compare absolute throughput, not only MFU",{"type":47,"value":1010},": MFU can mislead when switching\nbetween BF16, FP8, and other precision modes.",{"type":41,"tag":692,"props":1012,"children":1013},{},[1014,1019],{"type":41,"tag":974,"props":1015,"children":1016},{},[1017],{"type":47,"value":1018},"Force-balance routing is the safer benchmark default",{"type":47,"value":1020},": keep routing mode\nfixed when comparing hardware or dispatcher stacks.",{"type":41,"tag":1022,"props":1023,"children":1024},"style",{},[1025],{"type":47,"value":1026},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":1028,"total":1130},[1029,1044,1058,1072,1084,1101,1116],{"slug":1030,"name":1030,"fn":1031,"description":1032,"org":1033,"tags":1034,"stars":23,"repoUrl":24,"updatedAt":1043},"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},[1035,1038,1041,1042],{"name":1036,"slug":1037,"type":15},"Data Analysis","data-analysis",{"name":1039,"slug":1040,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":13,"slug":14,"type":15},"2026-07-14T05:28:43.176466",{"slug":1045,"name":1045,"fn":1046,"description":1047,"org":1048,"tags":1049,"stars":23,"repoUrl":24,"updatedAt":1057},"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},[1050,1053,1056],{"name":1051,"slug":1052,"type":15},"Deployment","deployment",{"name":1054,"slug":1055,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1059,"name":1059,"fn":1060,"description":1061,"org":1062,"tags":1063,"stars":23,"repoUrl":24,"updatedAt":1071},"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},[1064,1067,1068],{"name":1065,"slug":1066,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1069,"slug":1070,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1073,"name":1073,"fn":1074,"description":1075,"org":1076,"tags":1077,"stars":23,"repoUrl":24,"updatedAt":1083},"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},[1078,1079,1080],{"name":1036,"slug":1037,"type":15},{"name":9,"slug":8,"type":15},{"name":1081,"slug":1082,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":1085,"name":1085,"fn":1086,"description":1087,"org":1088,"tags":1089,"stars":23,"repoUrl":24,"updatedAt":1100},"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},[1090,1093,1096,1097],{"name":1091,"slug":1092,"type":15},"Automation","automation",{"name":1094,"slug":1095,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":1098,"slug":1099,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":1102,"name":1102,"fn":1103,"description":1104,"org":1105,"tags":1106,"stars":23,"repoUrl":24,"updatedAt":1115},"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},[1107,1108,1111,1112],{"name":1051,"slug":1052,"type":15},{"name":1109,"slug":1110,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1113,"slug":1114,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1117,"name":1117,"fn":1118,"description":1119,"org":1120,"tags":1121,"stars":23,"repoUrl":24,"updatedAt":1129},"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},[1122,1123,1126],{"name":9,"slug":8,"type":15},{"name":1124,"slug":1125,"type":15},"Quantum Computing","quantum-computing",{"name":1127,"slug":1128,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":1132,"total":1283},[1133,1151,1167,1178,1190,1204,1217,1231,1242,1251,1265,1274],{"slug":1134,"name":1134,"fn":1135,"description":1136,"org":1137,"tags":1138,"stars":1148,"repoUrl":1149,"updatedAt":1150},"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},[1139,1142,1145],{"name":1140,"slug":1141,"type":15},"Documentation","documentation",{"name":1143,"slug":1144,"type":15},"MCP","mcp",{"name":1146,"slug":1147,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1152,"name":1152,"fn":1153,"description":1154,"org":1155,"tags":1156,"stars":1164,"repoUrl":1165,"updatedAt":1166},"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},[1157,1160,1161],{"name":1158,"slug":1159,"type":15},"Containers","containers",{"name":1051,"slug":1052,"type":15},{"name":1162,"slug":1163,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1168,"name":1168,"fn":1169,"description":1170,"org":1171,"tags":1172,"stars":1164,"repoUrl":1165,"updatedAt":1177},"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},[1173,1176],{"name":1174,"slug":1175,"type":15},"CI\u002FCD","ci-cd",{"name":1051,"slug":1052,"type":15},"2026-07-14T05:25:59.97109",{"slug":1179,"name":1179,"fn":1180,"description":1181,"org":1182,"tags":1183,"stars":1164,"repoUrl":1165,"updatedAt":1189},"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},[1184,1185,1186],{"name":1174,"slug":1175,"type":15},{"name":1051,"slug":1052,"type":15},{"name":1187,"slug":1188,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1191,"name":1191,"fn":1192,"description":1193,"org":1194,"tags":1195,"stars":1164,"repoUrl":1165,"updatedAt":1203},"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},[1196,1199,1200],{"name":1197,"slug":1198,"type":15},"Debugging","debugging",{"name":1187,"slug":1188,"type":15},{"name":1201,"slug":1202,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1205,"name":1205,"fn":1206,"description":1207,"org":1208,"tags":1209,"stars":1164,"repoUrl":1165,"updatedAt":1216},"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},[1210,1213],{"name":1211,"slug":1212,"type":15},"Best Practices","best-practices",{"name":1214,"slug":1215,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1218,"name":1218,"fn":1219,"description":1220,"org":1221,"tags":1222,"stars":1164,"repoUrl":1165,"updatedAt":1230},"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},[1223,1226,1229],{"name":1224,"slug":1225,"type":15},"Machine Learning","machine-learning",{"name":1227,"slug":1228,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1232,"name":1232,"fn":1233,"description":1234,"org":1235,"tags":1236,"stars":1164,"repoUrl":1165,"updatedAt":1241},"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},[1237,1240],{"name":1238,"slug":1239,"type":15},"QA","qa",{"name":1081,"slug":1082,"type":15},"2026-07-14T05:25:53.673039",{"slug":1243,"name":1243,"fn":1244,"description":1245,"org":1246,"tags":1247,"stars":1164,"repoUrl":1165,"updatedAt":1250},"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},[1248,1249],{"name":1051,"slug":1052,"type":15},{"name":1054,"slug":1055,"type":15},"2026-07-14T05:25:49.362534",{"slug":1252,"name":1252,"fn":1253,"description":1254,"org":1255,"tags":1256,"stars":1164,"repoUrl":1165,"updatedAt":1264},"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},[1257,1260,1261],{"name":1258,"slug":1259,"type":15},"Code Review","code-review",{"name":1187,"slug":1188,"type":15},{"name":1262,"slug":1263,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1266,"name":1266,"fn":1267,"description":1268,"org":1269,"tags":1270,"stars":1164,"repoUrl":1165,"updatedAt":1273},"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},[1271,1272],{"name":1238,"slug":1239,"type":15},{"name":1081,"slug":1082,"type":15},"2026-07-14T05:25:54.928983",{"slug":1275,"name":1275,"fn":1276,"description":1277,"org":1278,"tags":1279,"stars":1164,"repoUrl":1165,"updatedAt":1282},"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},[1280,1281],{"name":1091,"slug":1092,"type":15},{"name":1174,"slug":1175,"type":15},"2026-07-30T05:29:03.275638",496]