[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-mbridge-perf-tp-dp-comm-overlap":3,"mdc--dtd6m2-key":34,"related-org-nvidia-nemo-mbridge-perf-tp-dp-comm-overlap":757,"related-repo-nvidia-nemo-mbridge-perf-tp-dp-comm-overlap":917},{"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-tp-dp-comm-overlap","optimize Megatron-Bridge communication overlap","Operational guide for enabling TP, DP, and PP communication overlap in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,17,20],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":9,"slug":8,"type":15},{"name":18,"slug":19,"type":15},"Engineering","engineering",{"name":21,"slug":22,"type":15},"AI Infrastructure","ai-infrastructure",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:27:44.07719","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-tp-dp-comm-overlap","---\nname: nemo-mbridge-perf-tp-dp-comm-overlap\ndescription: Operational guide for enabling TP, DP, and PP communication overlap in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.\nlicense: Apache-2.0\nwhen_to_use: Enabling TP\u002FDP\u002FPP comm overlap, or tracing a throughput regression to a comm overlap config change; 'overlap_param_gather', 'overlap_grad_reduce', 'sequence-parallel overlap', 'TP overlap', 'DP overlap', 'comm overlap'.\n---\n\n# TP \u002F DP \u002F PP Communication Overlap Skill\n\nFor stable background and recommendation level, see:\n\n- @docs\u002Ftraining\u002Fcommunication-overlap.md\n\n## Enablement\n\nMinimal Bridge override:\n\n```python\nfrom megatron.bridge.training.comm_overlap import CommOverlapConfig\n\ncfg.model.tensor_model_parallel_size = 4\ncfg.model.sequence_parallel = True\ncfg.model.pipeline_model_parallel_size = 4\ncfg.model.virtual_pipeline_model_parallel_size = 2\n\ncfg.comm_overlap = CommOverlapConfig(\n    tp_comm_overlap=True,\n)\n\ncfg.ddp.use_distributed_optimizer = True\ncfg.ddp.overlap_grad_reduce = True\ncfg.ddp.overlap_param_gather = True\n```\n\nOptional TP preset:\n\n```python\nfrom megatron.bridge.training.comm_overlap import userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048\n\ncfg.comm_overlap.tp_comm_overlap_cfg = userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048\n```\n\nPrecision knobs belong to mixed precision:\n\n```python\ncfg.mixed_precision.grad_reduce_in_fp32 = False\ncfg.mixed_precision.fp8_param_gather = False\n```\n\n## Code Anchors\n\nBridge overlap gating:\n\n```439:449:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py\nif self.user_comm_overlap_cfg.tp_comm_overlap is True:\n    if model_cfg.tensor_model_parallel_size \u003C 2:\n        ...\n    elif not model_cfg.sequence_parallel:\n        ...\n    elif not HAVE_TE:\n        ...\n```\n\nPP overlap selection:\n\n```451:458:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py\nif model_cfg.pipeline_model_parallel_size > 1:\n    if vp_size > 1:\n        comm_overlap_cfg.overlap_p2p_comm = True\n        comm_overlap_cfg.batch_p2p_comm = False\n    else:\n        comm_overlap_cfg.overlap_p2p_comm = False\n        comm_overlap_cfg.batch_p2p_comm = True\n```\n\nDP overlap defaults:\n\n```572:579:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py\nif self.data_parallel_size > 1:\n    comm_overlap_cfg.bucket_size = 128 * 1024 * 1024\n    comm_overlap_cfg.overlap_grad_reduce = True\n    comm_overlap_cfg.overlap_param_gather = True\n```\n\nLaunch-time env tuning:\n\n```570:609:src\u002Fmegatron\u002Fbridge\u002Frecipes\u002Frun_plugins.py\nexecutor.env_vars[\"CUDA_DEVICE_MAX_CONNECTIONS\"] = str(cuda_device_max_connections)\n...\nexecutor.env_vars[\"NVTE_FWD_LAYERNORM_SM_MARGIN\"] = str(self.layernorm_sm_margin)\nexecutor.env_vars[\"NVTE_BWD_LAYERNORM_SM_MARGIN\"] = str(self.layernorm_sm_margin)\n```\n\n## Pitfalls\n\n1. TP overlap silently disables itself if `sequence_parallel=False` or Transformer Engine is unavailable.\n2. PP overlap is not enabled for all PP cases. Bridge only auto-selects `overlap_p2p_comm=True` when `PP > 1` and `VPP > 1`.\n3. `bucket_size` is a parameter-count knob, not a byte-size knob.\n4. `grad_reduce_in_fp32` and `fp8_param_gather` should be set through mixed precision, not as standalone DDP tuning first.\n5. `CUDA_DEVICE_MAX_CONNECTIONS` and LayerNorm SM margin are launch-time plugin settings, not `CommOverlapConfig` fields.\n\n## Verification\n\nUse the checked-in overlap unit coverage first:\n\n```bash\nuv run python -m pytest tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py -q\n```\n\nOptional second check if `nemo_run` is available:\n\n```bash\nuv run python -m pytest tests\u002Funit_tests\u002Frecipes\u002Ftest_run_plugins.py -q\n```\n\nSuccess criteria:\n\n- first command reports `26 passed`\n- second command validates plugin-owned env wiring when not skipped\n",{"data":35,"body":37},{"name":4,"description":6,"license":26,"when_to_use":36},"Enabling TP\u002FDP\u002FPP comm overlap, or tracing a throughput regression to a comm overlap config change; 'overlap_param_gather', 'overlap_grad_reduce', 'sequence-parallel overlap', 'TP overlap', 'DP overlap', 'comm overlap'.",{"type":38,"children":39},"root",[40,49,55,65,72,77,216,221,251,256,279,285,290,353,358,423,428,469,474,515,521,615,621,626,675,688,727,732,751],{"type":41,"tag":42,"props":43,"children":45},"element","h1",{"id":44},"tp-dp-pp-communication-overlap-skill",[46],{"type":47,"value":48},"text","TP \u002F DP \u002F PP Communication Overlap Skill",{"type":41,"tag":50,"props":51,"children":52},"p",{},[53],{"type":47,"value":54},"For stable background and recommendation level, see:",{"type":41,"tag":56,"props":57,"children":58},"ul",{},[59],{"type":41,"tag":60,"props":61,"children":62},"li",{},[63],{"type":47,"value":64},"@docs\u002Ftraining\u002Fcommunication-overlap.md",{"type":41,"tag":66,"props":67,"children":69},"h2",{"id":68},"enablement",[70],{"type":47,"value":71},"Enablement",{"type":41,"tag":50,"props":73,"children":74},{},[75],{"type":47,"value":76},"Minimal Bridge override:",{"type":41,"tag":78,"props":79,"children":84},"pre",{"className":80,"code":81,"language":82,"meta":83,"style":83},"language-python shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","from megatron.bridge.training.comm_overlap import CommOverlapConfig\n\ncfg.model.tensor_model_parallel_size = 4\ncfg.model.sequence_parallel = True\ncfg.model.pipeline_model_parallel_size = 4\ncfg.model.virtual_pipeline_model_parallel_size = 2\n\ncfg.comm_overlap = CommOverlapConfig(\n    tp_comm_overlap=True,\n)\n\ncfg.ddp.use_distributed_optimizer = True\ncfg.ddp.overlap_grad_reduce = True\ncfg.ddp.overlap_param_gather = True\n","python","",[85],{"type":41,"tag":86,"props":87,"children":88},"code",{"__ignoreMap":83},[89,100,110,119,128,137,146,154,163,172,181,189,198,207],{"type":41,"tag":90,"props":91,"children":94},"span",{"class":92,"line":93},"line",1,[95],{"type":41,"tag":90,"props":96,"children":97},{},[98],{"type":47,"value":99},"from megatron.bridge.training.comm_overlap import CommOverlapConfig\n",{"type":41,"tag":90,"props":101,"children":103},{"class":92,"line":102},2,[104],{"type":41,"tag":90,"props":105,"children":107},{"emptyLinePlaceholder":106},true,[108],{"type":47,"value":109},"\n",{"type":41,"tag":90,"props":111,"children":113},{"class":92,"line":112},3,[114],{"type":41,"tag":90,"props":115,"children":116},{},[117],{"type":47,"value":118},"cfg.model.tensor_model_parallel_size = 4\n",{"type":41,"tag":90,"props":120,"children":122},{"class":92,"line":121},4,[123],{"type":41,"tag":90,"props":124,"children":125},{},[126],{"type":47,"value":127},"cfg.model.sequence_parallel = True\n",{"type":41,"tag":90,"props":129,"children":131},{"class":92,"line":130},5,[132],{"type":41,"tag":90,"props":133,"children":134},{},[135],{"type":47,"value":136},"cfg.model.pipeline_model_parallel_size = 4\n",{"type":41,"tag":90,"props":138,"children":140},{"class":92,"line":139},6,[141],{"type":41,"tag":90,"props":142,"children":143},{},[144],{"type":47,"value":145},"cfg.model.virtual_pipeline_model_parallel_size = 2\n",{"type":41,"tag":90,"props":147,"children":149},{"class":92,"line":148},7,[150],{"type":41,"tag":90,"props":151,"children":152},{"emptyLinePlaceholder":106},[153],{"type":47,"value":109},{"type":41,"tag":90,"props":155,"children":157},{"class":92,"line":156},8,[158],{"type":41,"tag":90,"props":159,"children":160},{},[161],{"type":47,"value":162},"cfg.comm_overlap = CommOverlapConfig(\n",{"type":41,"tag":90,"props":164,"children":166},{"class":92,"line":165},9,[167],{"type":41,"tag":90,"props":168,"children":169},{},[170],{"type":47,"value":171},"    tp_comm_overlap=True,\n",{"type":41,"tag":90,"props":173,"children":175},{"class":92,"line":174},10,[176],{"type":41,"tag":90,"props":177,"children":178},{},[179],{"type":47,"value":180},")\n",{"type":41,"tag":90,"props":182,"children":184},{"class":92,"line":183},11,[185],{"type":41,"tag":90,"props":186,"children":187},{"emptyLinePlaceholder":106},[188],{"type":47,"value":109},{"type":41,"tag":90,"props":190,"children":192},{"class":92,"line":191},12,[193],{"type":41,"tag":90,"props":194,"children":195},{},[196],{"type":47,"value":197},"cfg.ddp.use_distributed_optimizer = True\n",{"type":41,"tag":90,"props":199,"children":201},{"class":92,"line":200},13,[202],{"type":41,"tag":90,"props":203,"children":204},{},[205],{"type":47,"value":206},"cfg.ddp.overlap_grad_reduce = True\n",{"type":41,"tag":90,"props":208,"children":210},{"class":92,"line":209},14,[211],{"type":41,"tag":90,"props":212,"children":213},{},[214],{"type":47,"value":215},"cfg.ddp.overlap_param_gather = True\n",{"type":41,"tag":50,"props":217,"children":218},{},[219],{"type":47,"value":220},"Optional TP preset:",{"type":41,"tag":78,"props":222,"children":224},{"className":80,"code":223,"language":82,"meta":83,"style":83},"from megatron.bridge.training.comm_overlap import userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048\n\ncfg.comm_overlap.tp_comm_overlap_cfg = userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048\n",[225],{"type":41,"tag":86,"props":226,"children":227},{"__ignoreMap":83},[228,236,243],{"type":41,"tag":90,"props":229,"children":230},{"class":92,"line":93},[231],{"type":41,"tag":90,"props":232,"children":233},{},[234],{"type":47,"value":235},"from megatron.bridge.training.comm_overlap import userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048\n",{"type":41,"tag":90,"props":237,"children":238},{"class":92,"line":102},[239],{"type":41,"tag":90,"props":240,"children":241},{"emptyLinePlaceholder":106},[242],{"type":47,"value":109},{"type":41,"tag":90,"props":244,"children":245},{"class":92,"line":112},[246],{"type":41,"tag":90,"props":247,"children":248},{},[249],{"type":47,"value":250},"cfg.comm_overlap.tp_comm_overlap_cfg = userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048\n",{"type":41,"tag":50,"props":252,"children":253},{},[254],{"type":47,"value":255},"Precision knobs belong to mixed precision:",{"type":41,"tag":78,"props":257,"children":259},{"className":80,"code":258,"language":82,"meta":83,"style":83},"cfg.mixed_precision.grad_reduce_in_fp32 = False\ncfg.mixed_precision.fp8_param_gather = False\n",[260],{"type":41,"tag":86,"props":261,"children":262},{"__ignoreMap":83},[263,271],{"type":41,"tag":90,"props":264,"children":265},{"class":92,"line":93},[266],{"type":41,"tag":90,"props":267,"children":268},{},[269],{"type":47,"value":270},"cfg.mixed_precision.grad_reduce_in_fp32 = False\n",{"type":41,"tag":90,"props":272,"children":273},{"class":92,"line":102},[274],{"type":41,"tag":90,"props":275,"children":276},{},[277],{"type":47,"value":278},"cfg.mixed_precision.fp8_param_gather = False\n",{"type":41,"tag":66,"props":280,"children":282},{"id":281},"code-anchors",[283],{"type":47,"value":284},"Code Anchors",{"type":41,"tag":50,"props":286,"children":287},{},[288],{"type":47,"value":289},"Bridge overlap gating:",{"type":41,"tag":78,"props":291,"children":295},{"className":292,"code":293,"language":294,"meta":83,"style":83},"language-439:449:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","if self.user_comm_overlap_cfg.tp_comm_overlap is True:\n    if model_cfg.tensor_model_parallel_size \u003C 2:\n        ...\n    elif not model_cfg.sequence_parallel:\n        ...\n    elif not HAVE_TE:\n        ...\n","439:449:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py",[296],{"type":41,"tag":86,"props":297,"children":298},{"__ignoreMap":83},[299,307,315,323,331,338,346],{"type":41,"tag":90,"props":300,"children":301},{"class":92,"line":93},[302],{"type":41,"tag":90,"props":303,"children":304},{},[305],{"type":47,"value":306},"if self.user_comm_overlap_cfg.tp_comm_overlap is True:\n",{"type":41,"tag":90,"props":308,"children":309},{"class":92,"line":102},[310],{"type":41,"tag":90,"props":311,"children":312},{},[313],{"type":47,"value":314},"    if model_cfg.tensor_model_parallel_size \u003C 2:\n",{"type":41,"tag":90,"props":316,"children":317},{"class":92,"line":112},[318],{"type":41,"tag":90,"props":319,"children":320},{},[321],{"type":47,"value":322},"        ...\n",{"type":41,"tag":90,"props":324,"children":325},{"class":92,"line":121},[326],{"type":41,"tag":90,"props":327,"children":328},{},[329],{"type":47,"value":330},"    elif not model_cfg.sequence_parallel:\n",{"type":41,"tag":90,"props":332,"children":333},{"class":92,"line":130},[334],{"type":41,"tag":90,"props":335,"children":336},{},[337],{"type":47,"value":322},{"type":41,"tag":90,"props":339,"children":340},{"class":92,"line":139},[341],{"type":41,"tag":90,"props":342,"children":343},{},[344],{"type":47,"value":345},"    elif not HAVE_TE:\n",{"type":41,"tag":90,"props":347,"children":348},{"class":92,"line":148},[349],{"type":41,"tag":90,"props":350,"children":351},{},[352],{"type":47,"value":322},{"type":41,"tag":50,"props":354,"children":355},{},[356],{"type":47,"value":357},"PP overlap selection:",{"type":41,"tag":78,"props":359,"children":363},{"className":360,"code":361,"language":362,"meta":83,"style":83},"language-451:458:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","if model_cfg.pipeline_model_parallel_size > 1:\n    if vp_size > 1:\n        comm_overlap_cfg.overlap_p2p_comm = True\n        comm_overlap_cfg.batch_p2p_comm = False\n    else:\n        comm_overlap_cfg.overlap_p2p_comm = False\n        comm_overlap_cfg.batch_p2p_comm = True\n","451:458:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py",[364],{"type":41,"tag":86,"props":365,"children":366},{"__ignoreMap":83},[367,375,383,391,399,407,415],{"type":41,"tag":90,"props":368,"children":369},{"class":92,"line":93},[370],{"type":41,"tag":90,"props":371,"children":372},{},[373],{"type":47,"value":374},"if model_cfg.pipeline_model_parallel_size > 1:\n",{"type":41,"tag":90,"props":376,"children":377},{"class":92,"line":102},[378],{"type":41,"tag":90,"props":379,"children":380},{},[381],{"type":47,"value":382},"    if vp_size > 1:\n",{"type":41,"tag":90,"props":384,"children":385},{"class":92,"line":112},[386],{"type":41,"tag":90,"props":387,"children":388},{},[389],{"type":47,"value":390},"        comm_overlap_cfg.overlap_p2p_comm = True\n",{"type":41,"tag":90,"props":392,"children":393},{"class":92,"line":121},[394],{"type":41,"tag":90,"props":395,"children":396},{},[397],{"type":47,"value":398},"        comm_overlap_cfg.batch_p2p_comm = False\n",{"type":41,"tag":90,"props":400,"children":401},{"class":92,"line":130},[402],{"type":41,"tag":90,"props":403,"children":404},{},[405],{"type":47,"value":406},"    else:\n",{"type":41,"tag":90,"props":408,"children":409},{"class":92,"line":139},[410],{"type":41,"tag":90,"props":411,"children":412},{},[413],{"type":47,"value":414},"        comm_overlap_cfg.overlap_p2p_comm = False\n",{"type":41,"tag":90,"props":416,"children":417},{"class":92,"line":148},[418],{"type":41,"tag":90,"props":419,"children":420},{},[421],{"type":47,"value":422},"        comm_overlap_cfg.batch_p2p_comm = True\n",{"type":41,"tag":50,"props":424,"children":425},{},[426],{"type":47,"value":427},"DP overlap defaults:",{"type":41,"tag":78,"props":429,"children":433},{"className":430,"code":431,"language":432,"meta":83,"style":83},"language-572:579:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","if self.data_parallel_size > 1:\n    comm_overlap_cfg.bucket_size = 128 * 1024 * 1024\n    comm_overlap_cfg.overlap_grad_reduce = True\n    comm_overlap_cfg.overlap_param_gather = True\n","572:579:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py",[434],{"type":41,"tag":86,"props":435,"children":436},{"__ignoreMap":83},[437,445,453,461],{"type":41,"tag":90,"props":438,"children":439},{"class":92,"line":93},[440],{"type":41,"tag":90,"props":441,"children":442},{},[443],{"type":47,"value":444},"if self.data_parallel_size > 1:\n",{"type":41,"tag":90,"props":446,"children":447},{"class":92,"line":102},[448],{"type":41,"tag":90,"props":449,"children":450},{},[451],{"type":47,"value":452},"    comm_overlap_cfg.bucket_size = 128 * 1024 * 1024\n",{"type":41,"tag":90,"props":454,"children":455},{"class":92,"line":112},[456],{"type":41,"tag":90,"props":457,"children":458},{},[459],{"type":47,"value":460},"    comm_overlap_cfg.overlap_grad_reduce = True\n",{"type":41,"tag":90,"props":462,"children":463},{"class":92,"line":121},[464],{"type":41,"tag":90,"props":465,"children":466},{},[467],{"type":47,"value":468},"    comm_overlap_cfg.overlap_param_gather = True\n",{"type":41,"tag":50,"props":470,"children":471},{},[472],{"type":47,"value":473},"Launch-time env tuning:",{"type":41,"tag":78,"props":475,"children":479},{"className":476,"code":477,"language":478,"meta":83,"style":83},"language-570:609:src\u002Fmegatron\u002Fbridge\u002Frecipes\u002Frun_plugins.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","executor.env_vars[\"CUDA_DEVICE_MAX_CONNECTIONS\"] = str(cuda_device_max_connections)\n...\nexecutor.env_vars[\"NVTE_FWD_LAYERNORM_SM_MARGIN\"] = str(self.layernorm_sm_margin)\nexecutor.env_vars[\"NVTE_BWD_LAYERNORM_SM_MARGIN\"] = str(self.layernorm_sm_margin)\n","570:609:src\u002Fmegatron\u002Fbridge\u002Frecipes\u002Frun_plugins.py",[480],{"type":41,"tag":86,"props":481,"children":482},{"__ignoreMap":83},[483,491,499,507],{"type":41,"tag":90,"props":484,"children":485},{"class":92,"line":93},[486],{"type":41,"tag":90,"props":487,"children":488},{},[489],{"type":47,"value":490},"executor.env_vars[\"CUDA_DEVICE_MAX_CONNECTIONS\"] = str(cuda_device_max_connections)\n",{"type":41,"tag":90,"props":492,"children":493},{"class":92,"line":102},[494],{"type":41,"tag":90,"props":495,"children":496},{},[497],{"type":47,"value":498},"...\n",{"type":41,"tag":90,"props":500,"children":501},{"class":92,"line":112},[502],{"type":41,"tag":90,"props":503,"children":504},{},[505],{"type":47,"value":506},"executor.env_vars[\"NVTE_FWD_LAYERNORM_SM_MARGIN\"] = str(self.layernorm_sm_margin)\n",{"type":41,"tag":90,"props":508,"children":509},{"class":92,"line":121},[510],{"type":41,"tag":90,"props":511,"children":512},{},[513],{"type":47,"value":514},"executor.env_vars[\"NVTE_BWD_LAYERNORM_SM_MARGIN\"] = str(self.layernorm_sm_margin)\n",{"type":41,"tag":66,"props":516,"children":518},{"id":517},"pitfalls",[519],{"type":47,"value":520},"Pitfalls",{"type":41,"tag":522,"props":523,"children":524},"ol",{},[525,538,567,578,596],{"type":41,"tag":60,"props":526,"children":527},{},[528,530,536],{"type":47,"value":529},"TP overlap silently disables itself if ",{"type":41,"tag":86,"props":531,"children":533},{"className":532},[],[534],{"type":47,"value":535},"sequence_parallel=False",{"type":47,"value":537}," or Transformer Engine is unavailable.",{"type":41,"tag":60,"props":539,"children":540},{},[541,543,549,551,557,559,565],{"type":47,"value":542},"PP overlap is not enabled for all PP cases. Bridge only auto-selects ",{"type":41,"tag":86,"props":544,"children":546},{"className":545},[],[547],{"type":47,"value":548},"overlap_p2p_comm=True",{"type":47,"value":550}," when ",{"type":41,"tag":86,"props":552,"children":554},{"className":553},[],[555],{"type":47,"value":556},"PP > 1",{"type":47,"value":558}," and ",{"type":41,"tag":86,"props":560,"children":562},{"className":561},[],[563],{"type":47,"value":564},"VPP > 1",{"type":47,"value":566},".",{"type":41,"tag":60,"props":568,"children":569},{},[570,576],{"type":41,"tag":86,"props":571,"children":573},{"className":572},[],[574],{"type":47,"value":575},"bucket_size",{"type":47,"value":577}," is a parameter-count knob, not a byte-size knob.",{"type":41,"tag":60,"props":579,"children":580},{},[581,587,588,594],{"type":41,"tag":86,"props":582,"children":584},{"className":583},[],[585],{"type":47,"value":586},"grad_reduce_in_fp32",{"type":47,"value":558},{"type":41,"tag":86,"props":589,"children":591},{"className":590},[],[592],{"type":47,"value":593},"fp8_param_gather",{"type":47,"value":595}," should be set through mixed precision, not as standalone DDP tuning first.",{"type":41,"tag":60,"props":597,"children":598},{},[599,605,607,613],{"type":41,"tag":86,"props":600,"children":602},{"className":601},[],[603],{"type":47,"value":604},"CUDA_DEVICE_MAX_CONNECTIONS",{"type":47,"value":606}," and LayerNorm SM margin are launch-time plugin settings, not ",{"type":41,"tag":86,"props":608,"children":610},{"className":609},[],[611],{"type":47,"value":612},"CommOverlapConfig",{"type":47,"value":614}," fields.",{"type":41,"tag":66,"props":616,"children":618},{"id":617},"verification",[619],{"type":47,"value":620},"Verification",{"type":41,"tag":50,"props":622,"children":623},{},[624],{"type":47,"value":625},"Use the checked-in overlap unit coverage first:",{"type":41,"tag":78,"props":627,"children":631},{"className":628,"code":629,"language":630,"meta":83,"style":83},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","uv run python -m pytest tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py -q\n","bash",[632],{"type":41,"tag":86,"props":633,"children":634},{"__ignoreMap":83},[635],{"type":41,"tag":90,"props":636,"children":637},{"class":92,"line":93},[638,644,650,655,660,665,670],{"type":41,"tag":90,"props":639,"children":641},{"style":640},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[642],{"type":47,"value":643},"uv",{"type":41,"tag":90,"props":645,"children":647},{"style":646},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[648],{"type":47,"value":649}," run",{"type":41,"tag":90,"props":651,"children":652},{"style":646},[653],{"type":47,"value":654}," python",{"type":41,"tag":90,"props":656,"children":657},{"style":646},[658],{"type":47,"value":659}," -m",{"type":41,"tag":90,"props":661,"children":662},{"style":646},[663],{"type":47,"value":664}," pytest",{"type":41,"tag":90,"props":666,"children":667},{"style":646},[668],{"type":47,"value":669}," tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py",{"type":41,"tag":90,"props":671,"children":672},{"style":646},[673],{"type":47,"value":674}," -q\n",{"type":41,"tag":50,"props":676,"children":677},{},[678,680,686],{"type":47,"value":679},"Optional second check if ",{"type":41,"tag":86,"props":681,"children":683},{"className":682},[],[684],{"type":47,"value":685},"nemo_run",{"type":47,"value":687}," is available:",{"type":41,"tag":78,"props":689,"children":691},{"className":628,"code":690,"language":630,"meta":83,"style":83},"uv run python -m pytest tests\u002Funit_tests\u002Frecipes\u002Ftest_run_plugins.py -q\n",[692],{"type":41,"tag":86,"props":693,"children":694},{"__ignoreMap":83},[695],{"type":41,"tag":90,"props":696,"children":697},{"class":92,"line":93},[698,702,706,710,714,718,723],{"type":41,"tag":90,"props":699,"children":700},{"style":640},[701],{"type":47,"value":643},{"type":41,"tag":90,"props":703,"children":704},{"style":646},[705],{"type":47,"value":649},{"type":41,"tag":90,"props":707,"children":708},{"style":646},[709],{"type":47,"value":654},{"type":41,"tag":90,"props":711,"children":712},{"style":646},[713],{"type":47,"value":659},{"type":41,"tag":90,"props":715,"children":716},{"style":646},[717],{"type":47,"value":664},{"type":41,"tag":90,"props":719,"children":720},{"style":646},[721],{"type":47,"value":722}," tests\u002Funit_tests\u002Frecipes\u002Ftest_run_plugins.py",{"type":41,"tag":90,"props":724,"children":725},{"style":646},[726],{"type":47,"value":674},{"type":41,"tag":50,"props":728,"children":729},{},[730],{"type":47,"value":731},"Success criteria:",{"type":41,"tag":56,"props":733,"children":734},{},[735,746],{"type":41,"tag":60,"props":736,"children":737},{},[738,740],{"type":47,"value":739},"first command reports ",{"type":41,"tag":86,"props":741,"children":743},{"className":742},[],[744],{"type":47,"value":745},"26 passed",{"type":41,"tag":60,"props":747,"children":748},{},[749],{"type":47,"value":750},"second command validates plugin-owned env wiring when not skipped",{"type":41,"tag":752,"props":753,"children":754},"style",{},[755],{"type":47,"value":756},"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":758,"total":916},[759,777,794,805,817,831,844,858,871,882,896,905],{"slug":760,"name":760,"fn":761,"description":762,"org":763,"tags":764,"stars":774,"repoUrl":775,"updatedAt":776},"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},[765,768,771],{"name":766,"slug":767,"type":15},"Documentation","documentation",{"name":769,"slug":770,"type":15},"MCP","mcp",{"name":772,"slug":773,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":778,"name":778,"fn":779,"description":780,"org":781,"tags":782,"stars":791,"repoUrl":792,"updatedAt":793},"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},[783,786,789],{"name":784,"slug":785,"type":15},"Containers","containers",{"name":787,"slug":788,"type":15},"Deployment","deployment",{"name":790,"slug":82,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":795,"name":795,"fn":796,"description":797,"org":798,"tags":799,"stars":791,"repoUrl":792,"updatedAt":804},"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},[800,803],{"name":801,"slug":802,"type":15},"CI\u002FCD","ci-cd",{"name":787,"slug":788,"type":15},"2026-07-14T05:25:59.97109",{"slug":806,"name":806,"fn":807,"description":808,"org":809,"tags":810,"stars":791,"repoUrl":792,"updatedAt":816},"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},[811,812,813],{"name":801,"slug":802,"type":15},{"name":787,"slug":788,"type":15},{"name":814,"slug":815,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":818,"name":818,"fn":819,"description":820,"org":821,"tags":822,"stars":791,"repoUrl":792,"updatedAt":830},"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},[823,826,827],{"name":824,"slug":825,"type":15},"Debugging","debugging",{"name":814,"slug":815,"type":15},{"name":828,"slug":829,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":832,"name":832,"fn":833,"description":834,"org":835,"tags":836,"stars":791,"repoUrl":792,"updatedAt":843},"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},[837,840],{"name":838,"slug":839,"type":15},"Best Practices","best-practices",{"name":841,"slug":842,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":845,"name":845,"fn":846,"description":847,"org":848,"tags":849,"stars":791,"repoUrl":792,"updatedAt":857},"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},[850,853,856],{"name":851,"slug":852,"type":15},"Machine Learning","machine-learning",{"name":854,"slug":855,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":859,"name":859,"fn":860,"description":861,"org":862,"tags":863,"stars":791,"repoUrl":792,"updatedAt":870},"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},[864,867],{"name":865,"slug":866,"type":15},"QA","qa",{"name":868,"slug":869,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":872,"name":872,"fn":873,"description":874,"org":875,"tags":876,"stars":791,"repoUrl":792,"updatedAt":881},"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},[877,878],{"name":787,"slug":788,"type":15},{"name":879,"slug":880,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":883,"name":883,"fn":884,"description":885,"org":886,"tags":887,"stars":791,"repoUrl":792,"updatedAt":895},"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},[888,891,892],{"name":889,"slug":890,"type":15},"Code Review","code-review",{"name":814,"slug":815,"type":15},{"name":893,"slug":894,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":897,"name":897,"fn":898,"description":899,"org":900,"tags":901,"stars":791,"repoUrl":792,"updatedAt":904},"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},[902,903],{"name":865,"slug":866,"type":15},{"name":868,"slug":869,"type":15},"2026-07-14T05:25:54.928983",{"slug":906,"name":906,"fn":907,"description":908,"org":909,"tags":910,"stars":791,"repoUrl":792,"updatedAt":915},"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},[911,914],{"name":912,"slug":913,"type":15},"Automation","automation",{"name":801,"slug":802,"type":15},"2026-07-30T05:29:03.275638",496,{"items":918,"total":1012},[919,934,944,958,968,983,998],{"slug":920,"name":920,"fn":921,"description":922,"org":923,"tags":924,"stars":23,"repoUrl":24,"updatedAt":933},"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},[925,928,931,932],{"name":926,"slug":927,"type":15},"Data Analysis","data-analysis",{"name":929,"slug":930,"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":935,"name":935,"fn":936,"description":937,"org":938,"tags":939,"stars":23,"repoUrl":24,"updatedAt":943},"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},[940,941,942],{"name":787,"slug":788,"type":15},{"name":879,"slug":880,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":945,"name":945,"fn":946,"description":947,"org":948,"tags":949,"stars":23,"repoUrl":24,"updatedAt":957},"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},[950,953,954],{"name":951,"slug":952,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":955,"slug":956,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":959,"name":959,"fn":960,"description":961,"org":962,"tags":963,"stars":23,"repoUrl":24,"updatedAt":967},"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},[964,965,966],{"name":926,"slug":927,"type":15},{"name":9,"slug":8,"type":15},{"name":868,"slug":869,"type":15},"2026-07-17T05:29:03.913266",{"slug":969,"name":969,"fn":970,"description":971,"org":972,"tags":973,"stars":23,"repoUrl":24,"updatedAt":982},"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},[974,975,978,979],{"name":912,"slug":913,"type":15},{"name":976,"slug":977,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":980,"slug":981,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":984,"name":984,"fn":985,"description":986,"org":987,"tags":988,"stars":23,"repoUrl":24,"updatedAt":997},"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},[989,990,993,994],{"name":787,"slug":788,"type":15},{"name":991,"slug":992,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":995,"slug":996,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":999,"name":999,"fn":1000,"description":1001,"org":1002,"tags":1003,"stars":23,"repoUrl":24,"updatedAt":1011},"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},[1004,1005,1008],{"name":9,"slug":8,"type":15},{"name":1006,"slug":1007,"type":15},"Quantum Computing","quantum-computing",{"name":1009,"slug":1010,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]