[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-mbridge-perf-moe-comm-overlap":3,"mdc-xw2vv9-key":31,"related-repo-nvidia-nemo-mbridge-perf-moe-comm-overlap":914,"related-org-nvidia-nemo-mbridge-perf-moe-comm-overlap":1018},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":20,"repoUrl":21,"updatedAt":22,"license":23,"forks":24,"topics":25,"repo":26,"sourceUrl":29,"mdContent":30},"nemo-mbridge-perf-moe-comm-overlap","optimize MoE communication overlap in Megatron-Bridge","MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch\u002Fcombine overlap, flex dispatcher backends, and expert wgrad scheduling.",{"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],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":9,"slug":8,"type":15},{"name":18,"slug":19,"type":15},"Engineering","engineering",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:26:57.640281","Apache-2.0",281,[],{"repoUrl":21,"stars":20,"forks":24,"topics":27,"description":28},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fnemo-mbridge-perf-moe-comm-overlap","---\nname: nemo-mbridge-perf-moe-comm-overlap\ndescription: MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch\u002Fcombine overlap, flex dispatcher backends, and expert wgrad scheduling.\nlicense: Apache-2.0\nwhen_to_use: Tuning MoE communication overlap, or tracing a MoE throughput regression to a comm-overlap config change; 'overlap_moe_expert_parallel_comm', 'MoE dispatch overlap', 'flex dispatcher', 'DeepEP overlap', 'expert wgrad scheduling'.\n---\n\n# MoE Communication Overlap\n\nFor the higher-level overview, see:\n\n- @docs\u002Ftraining\u002Fcommunication-overlap.md\n- @skills\u002Fnemo-mbridge-perf-moe-comm-overlap\u002Fcard.yaml\n\n## Quick Decision\n\nUse MoE communication overlap when:\n\n- `EP > 1`\n- token dispatch or combine time is visible in the profile\n- the run is already correct and you are now tuning throughput\n\nAvoid turning it on as an early bring-up step. It is easier to validate after\nthe dispatcher, routing mode, and recompute plan are already stable.\n\n## Enablement\n\n```python\ncfg.comm_overlap.overlap_moe_expert_parallel_comm = True\n\n# Optional: delayed wgrad for additional overlap\ncfg.comm_overlap.delay_wgrad_compute = True\n\n# IMPORTANT: disable shared expert overlap when using dispatch overlap\ncfg.model.moe_shared_expert_overlap = False\n```\n\n### Prerequisites\n\n- `expert_model_parallel_size > 1`\n- `num_moe_experts > 1`\n- `moe_token_dispatcher_type` must be `\"alltoall\"` or `\"flex\"`\n- Precision: BF16 or FP16\n- If PP is used, VPP (`virtual_pipeline_model_parallel_size`) must be set (non-`None`)\n\n### Flex dispatcher activation\n\nSetting `moe_flex_dispatcher_backend` alone does **not** activate flex dispatch.\nYou must also set `moe_token_dispatcher_type = \"flex\"`.\n\n## Recompute And CUDA Graph Interaction\n\n- Full recompute is not a good companion for the overlap path.\n- `delay_wgrad_compute` adds further constraints if CUDA-graph scopes include\n  attention or MoE-router work.\n- In practice, selective recompute is the safer pairing when overlap is enabled.\n\n## Measured Short-Run Caveat\n\nA 2026-05-18 current-main H100 x16 smoke on Qwen3 30B-A3B mock pretraining\nused `EP=16`, `alltoall`, global batch size 1024, CUDA graphs disabled, and\n`moe_permute_fusion=false` because the PyTorch 25.11 \u002F TE \u002F Triton stack failed\nin Transformer Engine fused permutation in prior bring-up.\n\nResults were directional rather than release-grade:\n\n- no EP overlap: 41.25s steady-state mean over iterations 3-8\n- EP overlap: 31.31s steady-state mean over iterations 3-8\n- EP overlap plus `delay_wgrad_compute`: 31.20s steady-state mean over\n  iterations 3-8\n\nTreat this as evidence that EP overlap can help an inter-node `alltoall` MoE\nshape when communication is exposed. It is not proof that delayed wgrad is a\nseparate win, and it does not validate the fused permutation path. An earlier\n2026-05-16 short smoke on the same shape showed the same pattern.\n\n## Code Anchors\n\n- Overlap validation: `src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py`\n- Flex dispatcher backend: `src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fflex_dispatcher_backend.py`\n- Config: `src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fconfig.py`\n- Unit tests: `tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py`\n- DeepEP tests: `tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py`\n\n## Pitfalls\n\n1. **Shared expert overlap conflict**: `moe_shared_expert_overlap` and\n   `overlap_moe_expert_parallel_comm` can conflict. Disable shared expert\n   overlap when using the dispatch overlap path.\n\n2. **PP without VPP**: MoE overlap requires VPP when pipeline parallelism is\n   active. Without it, the overlap scheduling cannot interleave correctly.\n\n3. **Flex != backend flag**: `moe_flex_dispatcher_backend=\"deepep\"` alone\n   does nothing if `moe_token_dispatcher_type` is still `\"alltoall\"`.\n\n4. **Conservative recipe defaults**: Most public recipes leave MoE overlap\n   disabled. You need to explicitly enable it via overrides.\n\n5. **Performance gains are workload-dependent**: overlap helps most when dispatch\n   communication is already a visible slice of step time. It is not guaranteed\n   to help every small or lightly loaded EP run.\n\n## Verification\n\nLook for overlap-related log messages during initialization. The comm overlap\nvalidation in `comm_overlap.py` will raise if prerequisites are not met, so a\nclean startup confirms the feature is active.\n\nFor a short performance-harness smoke, keep the command shape explicit and vary\nonly one overlap knob at a time:\n\n```bash\nuv run python scripts\u002Fperformance\u002Frun_script.py \\\n  -m qwen \\\n  -mr qwen3_30b_a3b \\\n  --task pretrain \\\n  -g h100 \\\n  -c bf16 \\\n  -ng 16 \\\n  -gn 8 \\\n  --max_steps 8 \\\n  --cuda_graph_impl none \\\n  --moe_flex_dispatcher_backend None \\\n  --moe_a2a_overlap false \\\n  --tokenizer_type NullTokenizer \\\n  comm_overlap.overlap_moe_expert_parallel_comm=true \\\n  comm_overlap.delay_wgrad_compute=false \\\n  model.moe_shared_expert_overlap=false\n```\n\nIf fused MoE permutation fails during bring-up, add\n`model.moe_permute_fusion=false` to separate overlap timing from runtime-stack\nvalidation, then retest with the matched production container.\n",{"data":32,"body":34},{"name":4,"description":6,"license":23,"when_to_use":33},"Tuning MoE communication overlap, or tracing a MoE throughput regression to a comm-overlap config change; 'overlap_moe_expert_parallel_comm', 'MoE dispatch overlap', 'flex dispatcher', 'DeepEP overlap', 'expert wgrad scheduling'.",{"type":35,"children":36},"root",[37,46,52,67,74,79,102,107,113,189,196,268,274,303,309,333,339,368,373,398,410,416,474,480,570,576,589,594,895,908],{"type":38,"tag":39,"props":40,"children":42},"element","h1",{"id":41},"moe-communication-overlap",[43],{"type":44,"value":45},"text","MoE Communication Overlap",{"type":38,"tag":47,"props":48,"children":49},"p",{},[50],{"type":44,"value":51},"For the higher-level overview, see:",{"type":38,"tag":53,"props":54,"children":55},"ul",{},[56,62],{"type":38,"tag":57,"props":58,"children":59},"li",{},[60],{"type":44,"value":61},"@docs\u002Ftraining\u002Fcommunication-overlap.md",{"type":38,"tag":57,"props":63,"children":64},{},[65],{"type":44,"value":66},"@skills\u002Fnemo-mbridge-perf-moe-comm-overlap\u002Fcard.yaml",{"type":38,"tag":68,"props":69,"children":71},"h2",{"id":70},"quick-decision",[72],{"type":44,"value":73},"Quick Decision",{"type":38,"tag":47,"props":75,"children":76},{},[77],{"type":44,"value":78},"Use MoE communication overlap when:",{"type":38,"tag":53,"props":80,"children":81},{},[82,92,97],{"type":38,"tag":57,"props":83,"children":84},{},[85],{"type":38,"tag":86,"props":87,"children":89},"code",{"className":88},[],[90],{"type":44,"value":91},"EP > 1",{"type":38,"tag":57,"props":93,"children":94},{},[95],{"type":44,"value":96},"token dispatch or combine time is visible in the profile",{"type":38,"tag":57,"props":98,"children":99},{},[100],{"type":44,"value":101},"the run is already correct and you are now tuning throughput",{"type":38,"tag":47,"props":103,"children":104},{},[105],{"type":44,"value":106},"Avoid turning it on as an early bring-up step. It is easier to validate after\nthe dispatcher, routing mode, and recompute plan are already stable.",{"type":38,"tag":68,"props":108,"children":110},{"id":109},"enablement",[111],{"type":44,"value":112},"Enablement",{"type":38,"tag":114,"props":115,"children":120},"pre",{"className":116,"code":117,"language":118,"meta":119,"style":119},"language-python shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","cfg.comm_overlap.overlap_moe_expert_parallel_comm = True\n\n# Optional: delayed wgrad for additional overlap\ncfg.comm_overlap.delay_wgrad_compute = True\n\n# IMPORTANT: disable shared expert overlap when using dispatch overlap\ncfg.model.moe_shared_expert_overlap = False\n","python","",[121],{"type":38,"tag":86,"props":122,"children":123},{"__ignoreMap":119},[124,135,145,154,163,171,180],{"type":38,"tag":125,"props":126,"children":129},"span",{"class":127,"line":128},"line",1,[130],{"type":38,"tag":125,"props":131,"children":132},{},[133],{"type":44,"value":134},"cfg.comm_overlap.overlap_moe_expert_parallel_comm = True\n",{"type":38,"tag":125,"props":136,"children":138},{"class":127,"line":137},2,[139],{"type":38,"tag":125,"props":140,"children":142},{"emptyLinePlaceholder":141},true,[143],{"type":44,"value":144},"\n",{"type":38,"tag":125,"props":146,"children":148},{"class":127,"line":147},3,[149],{"type":38,"tag":125,"props":150,"children":151},{},[152],{"type":44,"value":153},"# Optional: delayed wgrad for additional overlap\n",{"type":38,"tag":125,"props":155,"children":157},{"class":127,"line":156},4,[158],{"type":38,"tag":125,"props":159,"children":160},{},[161],{"type":44,"value":162},"cfg.comm_overlap.delay_wgrad_compute = True\n",{"type":38,"tag":125,"props":164,"children":166},{"class":127,"line":165},5,[167],{"type":38,"tag":125,"props":168,"children":169},{"emptyLinePlaceholder":141},[170],{"type":44,"value":144},{"type":38,"tag":125,"props":172,"children":174},{"class":127,"line":173},6,[175],{"type":38,"tag":125,"props":176,"children":177},{},[178],{"type":44,"value":179},"# IMPORTANT: disable shared expert overlap when using dispatch overlap\n",{"type":38,"tag":125,"props":181,"children":183},{"class":127,"line":182},7,[184],{"type":38,"tag":125,"props":185,"children":186},{},[187],{"type":44,"value":188},"cfg.model.moe_shared_expert_overlap = False\n",{"type":38,"tag":190,"props":191,"children":193},"h3",{"id":192},"prerequisites",[194],{"type":44,"value":195},"Prerequisites",{"type":38,"tag":53,"props":197,"children":198},{},[199,208,217,242,247],{"type":38,"tag":57,"props":200,"children":201},{},[202],{"type":38,"tag":86,"props":203,"children":205},{"className":204},[],[206],{"type":44,"value":207},"expert_model_parallel_size > 1",{"type":38,"tag":57,"props":209,"children":210},{},[211],{"type":38,"tag":86,"props":212,"children":214},{"className":213},[],[215],{"type":44,"value":216},"num_moe_experts > 1",{"type":38,"tag":57,"props":218,"children":219},{},[220,226,228,234,236],{"type":38,"tag":86,"props":221,"children":223},{"className":222},[],[224],{"type":44,"value":225},"moe_token_dispatcher_type",{"type":44,"value":227}," must be ",{"type":38,"tag":86,"props":229,"children":231},{"className":230},[],[232],{"type":44,"value":233},"\"alltoall\"",{"type":44,"value":235}," or ",{"type":38,"tag":86,"props":237,"children":239},{"className":238},[],[240],{"type":44,"value":241},"\"flex\"",{"type":38,"tag":57,"props":243,"children":244},{},[245],{"type":44,"value":246},"Precision: BF16 or FP16",{"type":38,"tag":57,"props":248,"children":249},{},[250,252,258,260,266],{"type":44,"value":251},"If PP is used, VPP (",{"type":38,"tag":86,"props":253,"children":255},{"className":254},[],[256],{"type":44,"value":257},"virtual_pipeline_model_parallel_size",{"type":44,"value":259},") must be set (non-",{"type":38,"tag":86,"props":261,"children":263},{"className":262},[],[264],{"type":44,"value":265},"None",{"type":44,"value":267},")",{"type":38,"tag":190,"props":269,"children":271},{"id":270},"flex-dispatcher-activation",[272],{"type":44,"value":273},"Flex dispatcher activation",{"type":38,"tag":47,"props":275,"children":276},{},[277,279,285,287,293,295,301],{"type":44,"value":278},"Setting ",{"type":38,"tag":86,"props":280,"children":282},{"className":281},[],[283],{"type":44,"value":284},"moe_flex_dispatcher_backend",{"type":44,"value":286}," alone does ",{"type":38,"tag":288,"props":289,"children":290},"strong",{},[291],{"type":44,"value":292},"not",{"type":44,"value":294}," activate flex dispatch.\nYou must also set ",{"type":38,"tag":86,"props":296,"children":298},{"className":297},[],[299],{"type":44,"value":300},"moe_token_dispatcher_type = \"flex\"",{"type":44,"value":302},".",{"type":38,"tag":68,"props":304,"children":306},{"id":305},"recompute-and-cuda-graph-interaction",[307],{"type":44,"value":308},"Recompute And CUDA Graph Interaction",{"type":38,"tag":53,"props":310,"children":311},{},[312,317,328],{"type":38,"tag":57,"props":313,"children":314},{},[315],{"type":44,"value":316},"Full recompute is not a good companion for the overlap path.",{"type":38,"tag":57,"props":318,"children":319},{},[320,326],{"type":38,"tag":86,"props":321,"children":323},{"className":322},[],[324],{"type":44,"value":325},"delay_wgrad_compute",{"type":44,"value":327}," adds further constraints if CUDA-graph scopes include\nattention or MoE-router work.",{"type":38,"tag":57,"props":329,"children":330},{},[331],{"type":44,"value":332},"In practice, selective recompute is the safer pairing when overlap is enabled.",{"type":38,"tag":68,"props":334,"children":336},{"id":335},"measured-short-run-caveat",[337],{"type":44,"value":338},"Measured Short-Run Caveat",{"type":38,"tag":47,"props":340,"children":341},{},[342,344,350,352,358,360,366],{"type":44,"value":343},"A 2026-05-18 current-main H100 x16 smoke on Qwen3 30B-A3B mock pretraining\nused ",{"type":38,"tag":86,"props":345,"children":347},{"className":346},[],[348],{"type":44,"value":349},"EP=16",{"type":44,"value":351},", ",{"type":38,"tag":86,"props":353,"children":355},{"className":354},[],[356],{"type":44,"value":357},"alltoall",{"type":44,"value":359},", global batch size 1024, CUDA graphs disabled, and\n",{"type":38,"tag":86,"props":361,"children":363},{"className":362},[],[364],{"type":44,"value":365},"moe_permute_fusion=false",{"type":44,"value":367}," because the PyTorch 25.11 \u002F TE \u002F Triton stack failed\nin Transformer Engine fused permutation in prior bring-up.",{"type":38,"tag":47,"props":369,"children":370},{},[371],{"type":44,"value":372},"Results were directional rather than release-grade:",{"type":38,"tag":53,"props":374,"children":375},{},[376,381,386],{"type":38,"tag":57,"props":377,"children":378},{},[379],{"type":44,"value":380},"no EP overlap: 41.25s steady-state mean over iterations 3-8",{"type":38,"tag":57,"props":382,"children":383},{},[384],{"type":44,"value":385},"EP overlap: 31.31s steady-state mean over iterations 3-8",{"type":38,"tag":57,"props":387,"children":388},{},[389,391,396],{"type":44,"value":390},"EP overlap plus ",{"type":38,"tag":86,"props":392,"children":394},{"className":393},[],[395],{"type":44,"value":325},{"type":44,"value":397},": 31.20s steady-state mean over\niterations 3-8",{"type":38,"tag":47,"props":399,"children":400},{},[401,403,408],{"type":44,"value":402},"Treat this as evidence that EP overlap can help an inter-node ",{"type":38,"tag":86,"props":404,"children":406},{"className":405},[],[407],{"type":44,"value":357},{"type":44,"value":409}," MoE\nshape when communication is exposed. It is not proof that delayed wgrad is a\nseparate win, and it does not validate the fused permutation path. An earlier\n2026-05-16 short smoke on the same shape showed the same pattern.",{"type":38,"tag":68,"props":411,"children":413},{"id":412},"code-anchors",[414],{"type":44,"value":415},"Code Anchors",{"type":38,"tag":53,"props":417,"children":418},{},[419,430,441,452,463],{"type":38,"tag":57,"props":420,"children":421},{},[422,424],{"type":44,"value":423},"Overlap validation: ",{"type":38,"tag":86,"props":425,"children":427},{"className":426},[],[428],{"type":44,"value":429},"src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py",{"type":38,"tag":57,"props":431,"children":432},{},[433,435],{"type":44,"value":434},"Flex dispatcher backend: ",{"type":38,"tag":86,"props":436,"children":438},{"className":437},[],[439],{"type":44,"value":440},"src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fflex_dispatcher_backend.py",{"type":38,"tag":57,"props":442,"children":443},{},[444,446],{"type":44,"value":445},"Config: ",{"type":38,"tag":86,"props":447,"children":449},{"className":448},[],[450],{"type":44,"value":451},"src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fconfig.py",{"type":38,"tag":57,"props":453,"children":454},{},[455,457],{"type":44,"value":456},"Unit tests: ",{"type":38,"tag":86,"props":458,"children":460},{"className":459},[],[461],{"type":44,"value":462},"tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py",{"type":38,"tag":57,"props":464,"children":465},{},[466,468],{"type":44,"value":467},"DeepEP tests: ",{"type":38,"tag":86,"props":469,"children":471},{"className":470},[],[472],{"type":44,"value":473},"tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py",{"type":38,"tag":68,"props":475,"children":477},{"id":476},"pitfalls",[478],{"type":44,"value":479},"Pitfalls",{"type":38,"tag":481,"props":482,"children":483},"ol",{},[484,510,520,550,560],{"type":38,"tag":57,"props":485,"children":486},{},[487,492,494,500,502,508],{"type":38,"tag":288,"props":488,"children":489},{},[490],{"type":44,"value":491},"Shared expert overlap conflict",{"type":44,"value":493},": ",{"type":38,"tag":86,"props":495,"children":497},{"className":496},[],[498],{"type":44,"value":499},"moe_shared_expert_overlap",{"type":44,"value":501}," and\n",{"type":38,"tag":86,"props":503,"children":505},{"className":504},[],[506],{"type":44,"value":507},"overlap_moe_expert_parallel_comm",{"type":44,"value":509}," can conflict. Disable shared expert\noverlap when using the dispatch overlap path.",{"type":38,"tag":57,"props":511,"children":512},{},[513,518],{"type":38,"tag":288,"props":514,"children":515},{},[516],{"type":44,"value":517},"PP without VPP",{"type":44,"value":519},": MoE overlap requires VPP when pipeline parallelism is\nactive. Without it, the overlap scheduling cannot interleave correctly.",{"type":38,"tag":57,"props":521,"children":522},{},[523,528,529,535,537,542,544,549],{"type":38,"tag":288,"props":524,"children":525},{},[526],{"type":44,"value":527},"Flex != backend flag",{"type":44,"value":493},{"type":38,"tag":86,"props":530,"children":532},{"className":531},[],[533],{"type":44,"value":534},"moe_flex_dispatcher_backend=\"deepep\"",{"type":44,"value":536}," alone\ndoes nothing if ",{"type":38,"tag":86,"props":538,"children":540},{"className":539},[],[541],{"type":44,"value":225},{"type":44,"value":543}," is still ",{"type":38,"tag":86,"props":545,"children":547},{"className":546},[],[548],{"type":44,"value":233},{"type":44,"value":302},{"type":38,"tag":57,"props":551,"children":552},{},[553,558],{"type":38,"tag":288,"props":554,"children":555},{},[556],{"type":44,"value":557},"Conservative recipe defaults",{"type":44,"value":559},": Most public recipes leave MoE overlap\ndisabled. You need to explicitly enable it via overrides.",{"type":38,"tag":57,"props":561,"children":562},{},[563,568],{"type":38,"tag":288,"props":564,"children":565},{},[566],{"type":44,"value":567},"Performance gains are workload-dependent",{"type":44,"value":569},": overlap helps most when dispatch\ncommunication is already a visible slice of step time. It is not guaranteed\nto help every small or lightly loaded EP run.",{"type":38,"tag":68,"props":571,"children":573},{"id":572},"verification",[574],{"type":44,"value":575},"Verification",{"type":38,"tag":47,"props":577,"children":578},{},[579,581,587],{"type":44,"value":580},"Look for overlap-related log messages during initialization. The comm overlap\nvalidation in ",{"type":38,"tag":86,"props":582,"children":584},{"className":583},[],[585],{"type":44,"value":586},"comm_overlap.py",{"type":44,"value":588}," will raise if prerequisites are not met, so a\nclean startup confirms the feature is active.",{"type":38,"tag":47,"props":590,"children":591},{},[592],{"type":44,"value":593},"For a short performance-harness smoke, keep the command shape explicit and vary\nonly one overlap knob at a time:",{"type":38,"tag":114,"props":595,"children":599},{"className":596,"code":597,"language":598,"meta":119,"style":119},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","uv run python scripts\u002Fperformance\u002Frun_script.py \\\n  -m qwen \\\n  -mr qwen3_30b_a3b \\\n  --task pretrain \\\n  -g h100 \\\n  -c bf16 \\\n  -ng 16 \\\n  -gn 8 \\\n  --max_steps 8 \\\n  --cuda_graph_impl none \\\n  --moe_flex_dispatcher_backend None \\\n  --moe_a2a_overlap false \\\n  --tokenizer_type NullTokenizer \\\n  comm_overlap.overlap_moe_expert_parallel_comm=true \\\n  comm_overlap.delay_wgrad_compute=false \\\n  model.moe_shared_expert_overlap=false\n","bash",[600],{"type":38,"tag":86,"props":601,"children":602},{"__ignoreMap":119},[603,634,651,668,685,702,719,737,755,772,790,808,827,845,863,881],{"type":38,"tag":125,"props":604,"children":605},{"class":127,"line":128},[606,612,618,623,628],{"type":38,"tag":125,"props":607,"children":609},{"style":608},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[610],{"type":44,"value":611},"uv",{"type":38,"tag":125,"props":613,"children":615},{"style":614},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[616],{"type":44,"value":617}," run",{"type":38,"tag":125,"props":619,"children":620},{"style":614},[621],{"type":44,"value":622}," python",{"type":38,"tag":125,"props":624,"children":625},{"style":614},[626],{"type":44,"value":627}," scripts\u002Fperformance\u002Frun_script.py",{"type":38,"tag":125,"props":629,"children":631},{"style":630},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[632],{"type":44,"value":633}," \\\n",{"type":38,"tag":125,"props":635,"children":636},{"class":127,"line":137},[637,642,647],{"type":38,"tag":125,"props":638,"children":639},{"style":614},[640],{"type":44,"value":641},"  -m",{"type":38,"tag":125,"props":643,"children":644},{"style":614},[645],{"type":44,"value":646}," qwen",{"type":38,"tag":125,"props":648,"children":649},{"style":630},[650],{"type":44,"value":633},{"type":38,"tag":125,"props":652,"children":653},{"class":127,"line":147},[654,659,664],{"type":38,"tag":125,"props":655,"children":656},{"style":614},[657],{"type":44,"value":658},"  -mr",{"type":38,"tag":125,"props":660,"children":661},{"style":614},[662],{"type":44,"value":663}," qwen3_30b_a3b",{"type":38,"tag":125,"props":665,"children":666},{"style":630},[667],{"type":44,"value":633},{"type":38,"tag":125,"props":669,"children":670},{"class":127,"line":156},[671,676,681],{"type":38,"tag":125,"props":672,"children":673},{"style":614},[674],{"type":44,"value":675},"  --task",{"type":38,"tag":125,"props":677,"children":678},{"style":614},[679],{"type":44,"value":680}," pretrain",{"type":38,"tag":125,"props":682,"children":683},{"style":630},[684],{"type":44,"value":633},{"type":38,"tag":125,"props":686,"children":687},{"class":127,"line":165},[688,693,698],{"type":38,"tag":125,"props":689,"children":690},{"style":614},[691],{"type":44,"value":692},"  -g",{"type":38,"tag":125,"props":694,"children":695},{"style":614},[696],{"type":44,"value":697}," h100",{"type":38,"tag":125,"props":699,"children":700},{"style":630},[701],{"type":44,"value":633},{"type":38,"tag":125,"props":703,"children":704},{"class":127,"line":173},[705,710,715],{"type":38,"tag":125,"props":706,"children":707},{"style":614},[708],{"type":44,"value":709},"  -c",{"type":38,"tag":125,"props":711,"children":712},{"style":614},[713],{"type":44,"value":714}," bf16",{"type":38,"tag":125,"props":716,"children":717},{"style":630},[718],{"type":44,"value":633},{"type":38,"tag":125,"props":720,"children":721},{"class":127,"line":182},[722,727,733],{"type":38,"tag":125,"props":723,"children":724},{"style":614},[725],{"type":44,"value":726},"  -ng",{"type":38,"tag":125,"props":728,"children":730},{"style":729},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[731],{"type":44,"value":732}," 16",{"type":38,"tag":125,"props":734,"children":735},{"style":630},[736],{"type":44,"value":633},{"type":38,"tag":125,"props":738,"children":740},{"class":127,"line":739},8,[741,746,751],{"type":38,"tag":125,"props":742,"children":743},{"style":614},[744],{"type":44,"value":745},"  -gn",{"type":38,"tag":125,"props":747,"children":748},{"style":729},[749],{"type":44,"value":750}," 8",{"type":38,"tag":125,"props":752,"children":753},{"style":630},[754],{"type":44,"value":633},{"type":38,"tag":125,"props":756,"children":758},{"class":127,"line":757},9,[759,764,768],{"type":38,"tag":125,"props":760,"children":761},{"style":614},[762],{"type":44,"value":763},"  --max_steps",{"type":38,"tag":125,"props":765,"children":766},{"style":729},[767],{"type":44,"value":750},{"type":38,"tag":125,"props":769,"children":770},{"style":630},[771],{"type":44,"value":633},{"type":38,"tag":125,"props":773,"children":775},{"class":127,"line":774},10,[776,781,786],{"type":38,"tag":125,"props":777,"children":778},{"style":614},[779],{"type":44,"value":780},"  --cuda_graph_impl",{"type":38,"tag":125,"props":782,"children":783},{"style":614},[784],{"type":44,"value":785}," none",{"type":38,"tag":125,"props":787,"children":788},{"style":630},[789],{"type":44,"value":633},{"type":38,"tag":125,"props":791,"children":793},{"class":127,"line":792},11,[794,799,804],{"type":38,"tag":125,"props":795,"children":796},{"style":614},[797],{"type":44,"value":798},"  --moe_flex_dispatcher_backend",{"type":38,"tag":125,"props":800,"children":801},{"style":614},[802],{"type":44,"value":803}," None",{"type":38,"tag":125,"props":805,"children":806},{"style":630},[807],{"type":44,"value":633},{"type":38,"tag":125,"props":809,"children":811},{"class":127,"line":810},12,[812,817,823],{"type":38,"tag":125,"props":813,"children":814},{"style":614},[815],{"type":44,"value":816},"  --moe_a2a_overlap",{"type":38,"tag":125,"props":818,"children":820},{"style":819},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[821],{"type":44,"value":822}," false",{"type":38,"tag":125,"props":824,"children":825},{"style":630},[826],{"type":44,"value":633},{"type":38,"tag":125,"props":828,"children":830},{"class":127,"line":829},13,[831,836,841],{"type":38,"tag":125,"props":832,"children":833},{"style":614},[834],{"type":44,"value":835},"  --tokenizer_type",{"type":38,"tag":125,"props":837,"children":838},{"style":614},[839],{"type":44,"value":840}," NullTokenizer",{"type":38,"tag":125,"props":842,"children":843},{"style":630},[844],{"type":44,"value":633},{"type":38,"tag":125,"props":846,"children":848},{"class":127,"line":847},14,[849,854,859],{"type":38,"tag":125,"props":850,"children":851},{"style":614},[852],{"type":44,"value":853},"  comm_overlap.overlap_moe_expert_parallel_comm=",{"type":38,"tag":125,"props":855,"children":856},{"style":819},[857],{"type":44,"value":858},"true",{"type":38,"tag":125,"props":860,"children":861},{"style":630},[862],{"type":44,"value":633},{"type":38,"tag":125,"props":864,"children":866},{"class":127,"line":865},15,[867,872,877],{"type":38,"tag":125,"props":868,"children":869},{"style":614},[870],{"type":44,"value":871},"  comm_overlap.delay_wgrad_compute=",{"type":38,"tag":125,"props":873,"children":874},{"style":819},[875],{"type":44,"value":876},"false",{"type":38,"tag":125,"props":878,"children":879},{"style":630},[880],{"type":44,"value":633},{"type":38,"tag":125,"props":882,"children":884},{"class":127,"line":883},16,[885,890],{"type":38,"tag":125,"props":886,"children":887},{"style":614},[888],{"type":44,"value":889},"  model.moe_shared_expert_overlap=",{"type":38,"tag":125,"props":891,"children":892},{"style":819},[893],{"type":44,"value":894},"false\n",{"type":38,"tag":47,"props":896,"children":897},{},[898,900,906],{"type":44,"value":899},"If fused MoE permutation fails during bring-up, add\n",{"type":38,"tag":86,"props":901,"children":903},{"className":902},[],[904],{"type":44,"value":905},"model.moe_permute_fusion=false",{"type":44,"value":907}," to separate overlap timing from runtime-stack\nvalidation, then retest with the matched production container.",{"type":38,"tag":909,"props":910,"children":911},"style",{},[912],{"type":44,"value":913},"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":915,"total":1017},[916,931,945,959,971,988,1003],{"slug":917,"name":917,"fn":918,"description":919,"org":920,"tags":921,"stars":20,"repoUrl":21,"updatedAt":930},"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},[922,925,928,929],{"name":923,"slug":924,"type":15},"Data Analysis","data-analysis",{"name":926,"slug":927,"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":932,"name":932,"fn":933,"description":934,"org":935,"tags":936,"stars":20,"repoUrl":21,"updatedAt":944},"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},[937,940,943],{"name":938,"slug":939,"type":15},"Deployment","deployment",{"name":941,"slug":942,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":946,"name":946,"fn":947,"description":948,"org":949,"tags":950,"stars":20,"repoUrl":21,"updatedAt":958},"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},[951,954,955],{"name":952,"slug":953,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":956,"slug":957,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":960,"name":960,"fn":961,"description":962,"org":963,"tags":964,"stars":20,"repoUrl":21,"updatedAt":970},"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},[965,966,967],{"name":923,"slug":924,"type":15},{"name":9,"slug":8,"type":15},{"name":968,"slug":969,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":972,"name":972,"fn":973,"description":974,"org":975,"tags":976,"stars":20,"repoUrl":21,"updatedAt":987},"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},[977,980,983,984],{"name":978,"slug":979,"type":15},"Automation","automation",{"name":981,"slug":982,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":985,"slug":986,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":989,"name":989,"fn":990,"description":991,"org":992,"tags":993,"stars":20,"repoUrl":21,"updatedAt":1002},"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},[994,995,998,999],{"name":938,"slug":939,"type":15},{"name":996,"slug":997,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1000,"slug":1001,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1004,"name":1004,"fn":1005,"description":1006,"org":1007,"tags":1008,"stars":20,"repoUrl":21,"updatedAt":1016},"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},[1009,1010,1013],{"name":9,"slug":8,"type":15},{"name":1011,"slug":1012,"type":15},"Quantum Computing","quantum-computing",{"name":1014,"slug":1015,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":1019,"total":1169},[1020,1038,1053,1064,1076,1090,1103,1117,1128,1137,1151,1160],{"slug":1021,"name":1021,"fn":1022,"description":1023,"org":1024,"tags":1025,"stars":1035,"repoUrl":1036,"updatedAt":1037},"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},[1026,1029,1032],{"name":1027,"slug":1028,"type":15},"Documentation","documentation",{"name":1030,"slug":1031,"type":15},"MCP","mcp",{"name":1033,"slug":1034,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1039,"name":1039,"fn":1040,"description":1041,"org":1042,"tags":1043,"stars":1050,"repoUrl":1051,"updatedAt":1052},"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},[1044,1047,1048],{"name":1045,"slug":1046,"type":15},"Containers","containers",{"name":938,"slug":939,"type":15},{"name":1049,"slug":118,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1054,"name":1054,"fn":1055,"description":1056,"org":1057,"tags":1058,"stars":1050,"repoUrl":1051,"updatedAt":1063},"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},[1059,1062],{"name":1060,"slug":1061,"type":15},"CI\u002FCD","ci-cd",{"name":938,"slug":939,"type":15},"2026-07-14T05:25:59.97109",{"slug":1065,"name":1065,"fn":1066,"description":1067,"org":1068,"tags":1069,"stars":1050,"repoUrl":1051,"updatedAt":1075},"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},[1070,1071,1072],{"name":1060,"slug":1061,"type":15},{"name":938,"slug":939,"type":15},{"name":1073,"slug":1074,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1077,"name":1077,"fn":1078,"description":1079,"org":1080,"tags":1081,"stars":1050,"repoUrl":1051,"updatedAt":1089},"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},[1082,1085,1086],{"name":1083,"slug":1084,"type":15},"Debugging","debugging",{"name":1073,"slug":1074,"type":15},{"name":1087,"slug":1088,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1091,"name":1091,"fn":1092,"description":1093,"org":1094,"tags":1095,"stars":1050,"repoUrl":1051,"updatedAt":1102},"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},[1096,1099],{"name":1097,"slug":1098,"type":15},"Best Practices","best-practices",{"name":1100,"slug":1101,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1104,"name":1104,"fn":1105,"description":1106,"org":1107,"tags":1108,"stars":1050,"repoUrl":1051,"updatedAt":1116},"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},[1109,1112,1115],{"name":1110,"slug":1111,"type":15},"Machine Learning","machine-learning",{"name":1113,"slug":1114,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1118,"name":1118,"fn":1119,"description":1120,"org":1121,"tags":1122,"stars":1050,"repoUrl":1051,"updatedAt":1127},"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},[1123,1126],{"name":1124,"slug":1125,"type":15},"QA","qa",{"name":968,"slug":969,"type":15},"2026-07-14T05:25:53.673039",{"slug":1129,"name":1129,"fn":1130,"description":1131,"org":1132,"tags":1133,"stars":1050,"repoUrl":1051,"updatedAt":1136},"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},[1134,1135],{"name":938,"slug":939,"type":15},{"name":941,"slug":942,"type":15},"2026-07-14T05:25:49.362534",{"slug":1138,"name":1138,"fn":1139,"description":1140,"org":1141,"tags":1142,"stars":1050,"repoUrl":1051,"updatedAt":1150},"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},[1143,1146,1147],{"name":1144,"slug":1145,"type":15},"Code Review","code-review",{"name":1073,"slug":1074,"type":15},{"name":1148,"slug":1149,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1152,"name":1152,"fn":1153,"description":1154,"org":1155,"tags":1156,"stars":1050,"repoUrl":1051,"updatedAt":1159},"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},[1157,1158],{"name":1124,"slug":1125,"type":15},{"name":968,"slug":969,"type":15},"2026-07-14T05:25:54.928983",{"slug":1161,"name":1161,"fn":1162,"description":1163,"org":1164,"tags":1165,"stars":1050,"repoUrl":1051,"updatedAt":1168},"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},[1166,1167],{"name":978,"slug":979,"type":15},{"name":1060,"slug":1061,"type":15},"2026-07-30T05:29:03.275638",496]