[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-mbridge-perf-moe-long-context":3,"mdc--ubtt9f-key":34,"related-repo-nvidia-nemo-mbridge-perf-moe-long-context":462,"related-org-nvidia-nemo-mbridge-perf-moe-long-context":566},{"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-long-context","optimize long-context MoE training for Megatron Bridge","Long-context MoE training guidance for Megatron Bridge. Covers CP sizing, selective recompute, dispatcher choices, and practical patterns from DSV3, Qwen3, and Qwen3-Next long-context experiments.",{"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},"LLM","llm",{"name":20,"slug":21,"type":15},"Machine Learning","machine-learning",{"name":9,"slug":8,"type":15},2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:30:31.673996","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-long-context","---\nname: nemo-mbridge-perf-moe-long-context\ndescription: Long-context MoE training guidance for Megatron Bridge. Covers CP sizing, selective recompute, dispatcher choices, and practical patterns from DSV3, Qwen3, and Qwen3-Next long-context experiments.\nlicense: Apache-2.0\nwhen_to_use: Training MoE at long sequence lengths, or investigating a commit that caused long-context MoE OOM or degraded throughput; 'long context MoE', '128k tokens', 'CP sizing for long sequences', 'selective recompute long context', 'MoE long-context OOM'.\n---\n\n# MoE Long-Context Training\n\nStable docs: @docs\u002Ftraining\u002Fmoe-optimization.md\nCard: @skills\u002Fnemo-mbridge-perf-moe-long-context\u002Fcard.yaml\n\n## What Changes At Long Context\n\nOnce sequence length moves well past the 4K-class regime, attention memory and\nactivation residency become the dominant constraints. For MoE models, that\nusually means you need some combination of:\n\n- context parallelism\n- selective recompute\n- lower precision\n- CPU offload for optimizer state\n- a dispatcher and PP layout that do not waste the smaller remaining DP budget\n\n## Rounded Scaling Patterns\n\n### DSV3 on H100\n\nThe DSV3 long-context runs show a stable pattern:\n\n- selective recompute works better than full recompute once you move past the\n  shortest contexts\n- throughput stays in a fairly narrow band from mid-length through very long\n  contexts if CP is increased appropriately\n- the trade shifts from \"memory fit\" to \"GPU-count feasibility\" as CP grows\n\nIn other words, long context does not immediately collapse utilization if the\nlayout is chosen well, but it does consume the DP budget very quickly.\n\n### Qwen3-Next on GB200\n\nQwen3-Next behaves more like a memory-sensitive medium-scale model:\n\n- 8K and 32K remain practical with moderate CP\n- 64K is possible, but the throughput drop is noticeable and memory becomes\n  much tighter\n- pipeline layout and grouped-GEMM improvements matter almost as much as CP\n\n### Qwen3 235B on GB200\n\nQwen3 235B shows that long context can still be efficient on NVL72 systems when\nTP, CP, and HybridEP are coordinated. The best 128K-class configurations are\nnot just \"fit-only\" recipes; they can remain highly efficient if routing,\nparallelism, and recompute are balanced.\n\n## CP Sizing Rules Of Thumb\n\n1. **Start from a 4K shard target**: a good first guess is\n   `CP ~= seq_len \u002F 4096`, then round to a practical power-of-two layout.\n\n2. **Keep DP alive if possible**: long-context scaling becomes brittle once CP,\n   EP, TP, and PP together squeeze DP down to the floor.\n\n3. **Prefer selective recompute**: recompute modules such as `up_proj`, `norm`,\n   `moe`, `moe_act`, or `mlp` before reaching for full recompute.\n\n4. **Avoid SDPA-heavy recompute at very long context**: recomputing attention\n   internals can add a lot of work for less memory benefit than recomputing\n   smaller MoE and MLP-side modules.\n\n5. **Use TP as another lever on NVL72 systems**: GB200 and GB300 runs can\n   sometimes trade some CP for TP while still staying efficient.\n\n6. **Assume GBS will need to shrink**: as CP rises and DP falls, you may need\n   to reduce global batch size or accept higher GA.\n\n## Representative Config Families\n\n### DSV3 at 128K on H100\n\n```text\nTP=1  CP=32  EP=32  PP=8  VPP=4\nPrecision: FP8-class\nDispatcher: DeepEP\nRecompute: up_proj, norm, moe, mlp\nExtra memory help: optimizer CPU offload\n```\n\n### DSV3 at 256K on H100\n\n```text\nTP=1  CP=64  EP=32  PP=8  EDP=2  VPP=4\nPrecision: FP8-class\nDispatcher: DeepEP\nRecompute: up_proj, norm, moe, mlp\nExtra memory help: optimizer CPU offload\n```\n\n### Qwen3 235B at 128K on GB200\n\n```text\nTP=4  CP=4  EP=32  PP=4  VPP=12\nPrecision: BF16 or MXFP8\nDispatcher: HybridEP\nRecompute: moe_act, norm\nCUDA Graph: attn + moe_router + moe_preprocess\n```\n\n## Recompute And CUDA Graph Guidance\n\nFor long-context MoE training:\n\n- start with selective recompute\n- add CUDA graphs only after the shapes and routing path are stable\n- keep sequence length and MBS fixed when using CUDA graphs\n- if the run depends on highly dynamic batches, prefer eager execution\n\nUseful references:\n\n- @docs\u002Ftraining\u002Factivation-recomputation.md\n- @skills\u002Fnemo-mbridge-perf-cuda-graphs\u002FSKILL.md\n\n## Pitfalls\n\n1. **CP does not replace EP or PP**: it adds another dimension; it does not make\n   the others disappear.\n\n2. **A good 4K baseline can still be a bad long-context baseline**: routing mode,\n   recompute choice, and offload strategy often need to change.\n\n3. **GPU-count feasibility becomes the real constraint**: very long context can\n   look fine in a single recipe, then become impossible once EP and PP are added\n   honestly across the full model.\n\n4. **CUDA graphs need static shapes**: variable-length batches and opportunistic\n   padding strategies can silently break the path.\n\n5. **Container and kernel support matters more at 128K+**: long-context paths\n   tend to rely on newer kernels and bug fixes than short-context bring-up does.\n",{"data":35,"body":37},{"name":4,"description":6,"license":26,"when_to_use":36},"Training MoE at long sequence lengths, or investigating a commit that caused long-context MoE OOM or degraded throughput; 'long context MoE', '128k tokens', 'CP sizing for long sequences', 'selective recompute long context', 'MoE long-context OOM'.",{"type":38,"children":39},"root",[40,49,55,62,67,97,103,110,115,133,138,144,149,167,173,178,184,297,303,309,321,327,336,342,351,357,362,385,390,403,409],{"type":41,"tag":42,"props":43,"children":45},"element","h1",{"id":44},"moe-long-context-training",[46],{"type":47,"value":48},"text","MoE Long-Context Training",{"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-long-context\u002Fcard.yaml",{"type":41,"tag":56,"props":57,"children":59},"h2",{"id":58},"what-changes-at-long-context",[60],{"type":47,"value":61},"What Changes At Long Context",{"type":41,"tag":50,"props":63,"children":64},{},[65],{"type":47,"value":66},"Once sequence length moves well past the 4K-class regime, attention memory and\nactivation residency become the dominant constraints. For MoE models, that\nusually means you need some combination of:",{"type":41,"tag":68,"props":69,"children":70},"ul",{},[71,77,82,87,92],{"type":41,"tag":72,"props":73,"children":74},"li",{},[75],{"type":47,"value":76},"context parallelism",{"type":41,"tag":72,"props":78,"children":79},{},[80],{"type":47,"value":81},"selective recompute",{"type":41,"tag":72,"props":83,"children":84},{},[85],{"type":47,"value":86},"lower precision",{"type":41,"tag":72,"props":88,"children":89},{},[90],{"type":47,"value":91},"CPU offload for optimizer state",{"type":41,"tag":72,"props":93,"children":94},{},[95],{"type":47,"value":96},"a dispatcher and PP layout that do not waste the smaller remaining DP budget",{"type":41,"tag":56,"props":98,"children":100},{"id":99},"rounded-scaling-patterns",[101],{"type":47,"value":102},"Rounded Scaling Patterns",{"type":41,"tag":104,"props":105,"children":107},"h3",{"id":106},"dsv3-on-h100",[108],{"type":47,"value":109},"DSV3 on H100",{"type":41,"tag":50,"props":111,"children":112},{},[113],{"type":47,"value":114},"The DSV3 long-context runs show a stable pattern:",{"type":41,"tag":68,"props":116,"children":117},{},[118,123,128],{"type":41,"tag":72,"props":119,"children":120},{},[121],{"type":47,"value":122},"selective recompute works better than full recompute once you move past the\nshortest contexts",{"type":41,"tag":72,"props":124,"children":125},{},[126],{"type":47,"value":127},"throughput stays in a fairly narrow band from mid-length through very long\ncontexts if CP is increased appropriately",{"type":41,"tag":72,"props":129,"children":130},{},[131],{"type":47,"value":132},"the trade shifts from \"memory fit\" to \"GPU-count feasibility\" as CP grows",{"type":41,"tag":50,"props":134,"children":135},{},[136],{"type":47,"value":137},"In other words, long context does not immediately collapse utilization if the\nlayout is chosen well, but it does consume the DP budget very quickly.",{"type":41,"tag":104,"props":139,"children":141},{"id":140},"qwen3-next-on-gb200",[142],{"type":47,"value":143},"Qwen3-Next on GB200",{"type":41,"tag":50,"props":145,"children":146},{},[147],{"type":47,"value":148},"Qwen3-Next behaves more like a memory-sensitive medium-scale model:",{"type":41,"tag":68,"props":150,"children":151},{},[152,157,162],{"type":41,"tag":72,"props":153,"children":154},{},[155],{"type":47,"value":156},"8K and 32K remain practical with moderate CP",{"type":41,"tag":72,"props":158,"children":159},{},[160],{"type":47,"value":161},"64K is possible, but the throughput drop is noticeable and memory becomes\nmuch tighter",{"type":41,"tag":72,"props":163,"children":164},{},[165],{"type":47,"value":166},"pipeline layout and grouped-GEMM improvements matter almost as much as CP",{"type":41,"tag":104,"props":168,"children":170},{"id":169},"qwen3-235b-on-gb200",[171],{"type":47,"value":172},"Qwen3 235B on GB200",{"type":41,"tag":50,"props":174,"children":175},{},[176],{"type":47,"value":177},"Qwen3 235B shows that long context can still be efficient on NVL72 systems when\nTP, CP, and HybridEP are coordinated. The best 128K-class configurations are\nnot just \"fit-only\" recipes; they can remain highly efficient if routing,\nparallelism, and recompute are balanced.",{"type":41,"tag":56,"props":179,"children":181},{"id":180},"cp-sizing-rules-of-thumb",[182],{"type":47,"value":183},"CP Sizing Rules Of Thumb",{"type":41,"tag":185,"props":186,"children":187},"ol",{},[188,208,218,267,277,287],{"type":41,"tag":72,"props":189,"children":190},{},[191,197,199,206],{"type":41,"tag":192,"props":193,"children":194},"strong",{},[195],{"type":47,"value":196},"Start from a 4K shard target",{"type":47,"value":198},": a good first guess is\n",{"type":41,"tag":200,"props":201,"children":203},"code",{"className":202},[],[204],{"type":47,"value":205},"CP ~= seq_len \u002F 4096",{"type":47,"value":207},", then round to a practical power-of-two layout.",{"type":41,"tag":72,"props":209,"children":210},{},[211,216],{"type":41,"tag":192,"props":212,"children":213},{},[214],{"type":47,"value":215},"Keep DP alive if possible",{"type":47,"value":217},": long-context scaling becomes brittle once CP,\nEP, TP, and PP together squeeze DP down to the floor.",{"type":41,"tag":72,"props":219,"children":220},{},[221,226,228,234,236,242,244,250,251,257,259,265],{"type":41,"tag":192,"props":222,"children":223},{},[224],{"type":47,"value":225},"Prefer selective recompute",{"type":47,"value":227},": recompute modules such as ",{"type":41,"tag":200,"props":229,"children":231},{"className":230},[],[232],{"type":47,"value":233},"up_proj",{"type":47,"value":235},", ",{"type":41,"tag":200,"props":237,"children":239},{"className":238},[],[240],{"type":47,"value":241},"norm",{"type":47,"value":243},",\n",{"type":41,"tag":200,"props":245,"children":247},{"className":246},[],[248],{"type":47,"value":249},"moe",{"type":47,"value":235},{"type":41,"tag":200,"props":252,"children":254},{"className":253},[],[255],{"type":47,"value":256},"moe_act",{"type":47,"value":258},", or ",{"type":41,"tag":200,"props":260,"children":262},{"className":261},[],[263],{"type":47,"value":264},"mlp",{"type":47,"value":266}," before reaching for full recompute.",{"type":41,"tag":72,"props":268,"children":269},{},[270,275],{"type":41,"tag":192,"props":271,"children":272},{},[273],{"type":47,"value":274},"Avoid SDPA-heavy recompute at very long context",{"type":47,"value":276},": recomputing attention\ninternals can add a lot of work for less memory benefit than recomputing\nsmaller MoE and MLP-side modules.",{"type":41,"tag":72,"props":278,"children":279},{},[280,285],{"type":41,"tag":192,"props":281,"children":282},{},[283],{"type":47,"value":284},"Use TP as another lever on NVL72 systems",{"type":47,"value":286},": GB200 and GB300 runs can\nsometimes trade some CP for TP while still staying efficient.",{"type":41,"tag":72,"props":288,"children":289},{},[290,295],{"type":41,"tag":192,"props":291,"children":292},{},[293],{"type":47,"value":294},"Assume GBS will need to shrink",{"type":47,"value":296},": as CP rises and DP falls, you may need\nto reduce global batch size or accept higher GA.",{"type":41,"tag":56,"props":298,"children":300},{"id":299},"representative-config-families",[301],{"type":47,"value":302},"Representative Config Families",{"type":41,"tag":104,"props":304,"children":306},{"id":305},"dsv3-at-128k-on-h100",[307],{"type":47,"value":308},"DSV3 at 128K on H100",{"type":41,"tag":310,"props":311,"children":316},"pre",{"className":312,"code":314,"language":47,"meta":315},[313],"language-text","TP=1  CP=32  EP=32  PP=8  VPP=4\nPrecision: FP8-class\nDispatcher: DeepEP\nRecompute: up_proj, norm, moe, mlp\nExtra memory help: optimizer CPU offload\n","",[317],{"type":41,"tag":200,"props":318,"children":319},{"__ignoreMap":315},[320],{"type":47,"value":314},{"type":41,"tag":104,"props":322,"children":324},{"id":323},"dsv3-at-256k-on-h100",[325],{"type":47,"value":326},"DSV3 at 256K on H100",{"type":41,"tag":310,"props":328,"children":331},{"className":329,"code":330,"language":47,"meta":315},[313],"TP=1  CP=64  EP=32  PP=8  EDP=2  VPP=4\nPrecision: FP8-class\nDispatcher: DeepEP\nRecompute: up_proj, norm, moe, mlp\nExtra memory help: optimizer CPU offload\n",[332],{"type":41,"tag":200,"props":333,"children":334},{"__ignoreMap":315},[335],{"type":47,"value":330},{"type":41,"tag":104,"props":337,"children":339},{"id":338},"qwen3-235b-at-128k-on-gb200",[340],{"type":47,"value":341},"Qwen3 235B at 128K on GB200",{"type":41,"tag":310,"props":343,"children":346},{"className":344,"code":345,"language":47,"meta":315},[313],"TP=4  CP=4  EP=32  PP=4  VPP=12\nPrecision: BF16 or MXFP8\nDispatcher: HybridEP\nRecompute: moe_act, norm\nCUDA Graph: attn + moe_router + moe_preprocess\n",[347],{"type":41,"tag":200,"props":348,"children":349},{"__ignoreMap":315},[350],{"type":47,"value":345},{"type":41,"tag":56,"props":352,"children":354},{"id":353},"recompute-and-cuda-graph-guidance",[355],{"type":47,"value":356},"Recompute And CUDA Graph Guidance",{"type":41,"tag":50,"props":358,"children":359},{},[360],{"type":47,"value":361},"For long-context MoE training:",{"type":41,"tag":68,"props":363,"children":364},{},[365,370,375,380],{"type":41,"tag":72,"props":366,"children":367},{},[368],{"type":47,"value":369},"start with selective recompute",{"type":41,"tag":72,"props":371,"children":372},{},[373],{"type":47,"value":374},"add CUDA graphs only after the shapes and routing path are stable",{"type":41,"tag":72,"props":376,"children":377},{},[378],{"type":47,"value":379},"keep sequence length and MBS fixed when using CUDA graphs",{"type":41,"tag":72,"props":381,"children":382},{},[383],{"type":47,"value":384},"if the run depends on highly dynamic batches, prefer eager execution",{"type":41,"tag":50,"props":386,"children":387},{},[388],{"type":47,"value":389},"Useful references:",{"type":41,"tag":68,"props":391,"children":392},{},[393,398],{"type":41,"tag":72,"props":394,"children":395},{},[396],{"type":47,"value":397},"@docs\u002Ftraining\u002Factivation-recomputation.md",{"type":41,"tag":72,"props":399,"children":400},{},[401],{"type":47,"value":402},"@skills\u002Fnemo-mbridge-perf-cuda-graphs\u002FSKILL.md",{"type":41,"tag":56,"props":404,"children":406},{"id":405},"pitfalls",[407],{"type":47,"value":408},"Pitfalls",{"type":41,"tag":185,"props":410,"children":411},{},[412,422,432,442,452],{"type":41,"tag":72,"props":413,"children":414},{},[415,420],{"type":41,"tag":192,"props":416,"children":417},{},[418],{"type":47,"value":419},"CP does not replace EP or PP",{"type":47,"value":421},": it adds another dimension; it does not make\nthe others disappear.",{"type":41,"tag":72,"props":423,"children":424},{},[425,430],{"type":41,"tag":192,"props":426,"children":427},{},[428],{"type":47,"value":429},"A good 4K baseline can still be a bad long-context baseline",{"type":47,"value":431},": routing mode,\nrecompute choice, and offload strategy often need to change.",{"type":41,"tag":72,"props":433,"children":434},{},[435,440],{"type":41,"tag":192,"props":436,"children":437},{},[438],{"type":47,"value":439},"GPU-count feasibility becomes the real constraint",{"type":47,"value":441},": very long context can\nlook fine in a single recipe, then become impossible once EP and PP are added\nhonestly across the full model.",{"type":41,"tag":72,"props":443,"children":444},{},[445,450],{"type":41,"tag":192,"props":446,"children":447},{},[448],{"type":47,"value":449},"CUDA graphs need static shapes",{"type":47,"value":451},": variable-length batches and opportunistic\npadding strategies can silently break the path.",{"type":41,"tag":72,"props":453,"children":454},{},[455,460],{"type":41,"tag":192,"props":456,"children":457},{},[458],{"type":47,"value":459},"Container and kernel support matters more at 128K+",{"type":47,"value":461},": long-context paths\ntend to rely on newer kernels and bug fixes than short-context bring-up does.",{"items":463,"total":565},[464,479,493,507,519,536,551],{"slug":465,"name":465,"fn":466,"description":467,"org":468,"tags":469,"stars":23,"repoUrl":24,"updatedAt":478},"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},[470,473,476,477],{"name":471,"slug":472,"type":15},"Data Analysis","data-analysis",{"name":474,"slug":475,"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":480,"name":480,"fn":481,"description":482,"org":483,"tags":484,"stars":23,"repoUrl":24,"updatedAt":492},"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},[485,488,491],{"name":486,"slug":487,"type":15},"Deployment","deployment",{"name":489,"slug":490,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":494,"name":494,"fn":495,"description":496,"org":497,"tags":498,"stars":23,"repoUrl":24,"updatedAt":506},"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},[499,502,503],{"name":500,"slug":501,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":504,"slug":505,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":508,"name":508,"fn":509,"description":510,"org":511,"tags":512,"stars":23,"repoUrl":24,"updatedAt":518},"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},[513,514,515],{"name":471,"slug":472,"type":15},{"name":9,"slug":8,"type":15},{"name":516,"slug":517,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":520,"name":520,"fn":521,"description":522,"org":523,"tags":524,"stars":23,"repoUrl":24,"updatedAt":535},"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},[525,528,531,532],{"name":526,"slug":527,"type":15},"Automation","automation",{"name":529,"slug":530,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":533,"slug":534,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":537,"name":537,"fn":538,"description":539,"org":540,"tags":541,"stars":23,"repoUrl":24,"updatedAt":550},"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},[542,543,546,547],{"name":486,"slug":487,"type":15},{"name":544,"slug":545,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":548,"slug":549,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":552,"name":552,"fn":553,"description":554,"org":555,"tags":556,"stars":23,"repoUrl":24,"updatedAt":564},"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},[557,558,561],{"name":9,"slug":8,"type":15},{"name":559,"slug":560,"type":15},"Quantum Computing","quantum-computing",{"name":562,"slug":563,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":567,"total":716},[568,586,602,613,625,639,652,664,675,684,698,707],{"slug":569,"name":569,"fn":570,"description":571,"org":572,"tags":573,"stars":583,"repoUrl":584,"updatedAt":585},"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},[574,577,580],{"name":575,"slug":576,"type":15},"Documentation","documentation",{"name":578,"slug":579,"type":15},"MCP","mcp",{"name":581,"slug":582,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":587,"name":587,"fn":588,"description":589,"org":590,"tags":591,"stars":599,"repoUrl":600,"updatedAt":601},"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},[592,595,596],{"name":593,"slug":594,"type":15},"Containers","containers",{"name":486,"slug":487,"type":15},{"name":597,"slug":598,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":603,"name":603,"fn":604,"description":605,"org":606,"tags":607,"stars":599,"repoUrl":600,"updatedAt":612},"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},[608,611],{"name":609,"slug":610,"type":15},"CI\u002FCD","ci-cd",{"name":486,"slug":487,"type":15},"2026-07-14T05:25:59.97109",{"slug":614,"name":614,"fn":615,"description":616,"org":617,"tags":618,"stars":599,"repoUrl":600,"updatedAt":624},"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},[619,620,621],{"name":609,"slug":610,"type":15},{"name":486,"slug":487,"type":15},{"name":622,"slug":623,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":626,"name":626,"fn":627,"description":628,"org":629,"tags":630,"stars":599,"repoUrl":600,"updatedAt":638},"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},[631,634,635],{"name":632,"slug":633,"type":15},"Debugging","debugging",{"name":622,"slug":623,"type":15},{"name":636,"slug":637,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":640,"name":640,"fn":641,"description":642,"org":643,"tags":644,"stars":599,"repoUrl":600,"updatedAt":651},"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},[645,648],{"name":646,"slug":647,"type":15},"Best Practices","best-practices",{"name":649,"slug":650,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":653,"name":653,"fn":654,"description":655,"org":656,"tags":657,"stars":599,"repoUrl":600,"updatedAt":663},"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},[658,659,662],{"name":20,"slug":21,"type":15},{"name":660,"slug":661,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":665,"name":665,"fn":666,"description":667,"org":668,"tags":669,"stars":599,"repoUrl":600,"updatedAt":674},"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},[670,673],{"name":671,"slug":672,"type":15},"QA","qa",{"name":516,"slug":517,"type":15},"2026-07-14T05:25:53.673039",{"slug":676,"name":676,"fn":677,"description":678,"org":679,"tags":680,"stars":599,"repoUrl":600,"updatedAt":683},"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},[681,682],{"name":486,"slug":487,"type":15},{"name":489,"slug":490,"type":15},"2026-07-14T05:25:49.362534",{"slug":685,"name":685,"fn":686,"description":687,"org":688,"tags":689,"stars":599,"repoUrl":600,"updatedAt":697},"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},[690,693,694],{"name":691,"slug":692,"type":15},"Code Review","code-review",{"name":622,"slug":623,"type":15},{"name":695,"slug":696,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":699,"name":699,"fn":700,"description":701,"org":702,"tags":703,"stars":599,"repoUrl":600,"updatedAt":706},"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},[704,705],{"name":671,"slug":672,"type":15},{"name":516,"slug":517,"type":15},"2026-07-14T05:25:54.928983",{"slug":708,"name":708,"fn":709,"description":710,"org":711,"tags":712,"stars":599,"repoUrl":600,"updatedAt":715},"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},[713,714],{"name":526,"slug":527,"type":15},{"name":609,"slug":610,"type":15},"2026-07-30T05:29:03.275638",496]