[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-mbridge-perf-expert-parallel-overlap":3,"mdc--wh13o3-key":34,"related-repo-nvidia-nemo-mbridge-perf-expert-parallel-overlap":2350,"related-org-nvidia-nemo-mbridge-perf-expert-parallel-overlap":2454},{"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-expert-parallel-overlap","optimize expert-parallel communication in Megatron-Bridge","Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,20],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":17,"slug":18,"type":15},"Deep Learning","deep-learning",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"Engineering","engineering",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:29:50.642546","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-expert-parallel-overlap","---\nname: nemo-mbridge-perf-expert-parallel-overlap\ndescription: Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP.\nlicense: Apache-2.0\nwhen_to_use: Enabling EP overlap to hide dispatch\u002Fcombine latency, or tracing a throughput regression to an EP overlap config change; 'overlap_moe_expert_parallel_comm', 'delay_wgrad_compute', 'flex dispatcher', 'DeepEP overlap', 'HybridEP overlap'.\n---\n\n# MoE Expert-Parallel Overlap Skill\n\n## References\n\n- Stable docs: @docs\u002Ftraining\u002Fcommunication-overlap.md\n- Structured metadata: @skills\u002Fnemo-mbridge-perf-expert-parallel-overlap\u002Fcard.yaml\n\n## What It Is\n\nExpert-parallel (EP) overlap hides the cost of token dispatch\u002Fcombine all-to-all\ncommunication by running it concurrently with expert FFN compute. Optionally,\ndelayed expert weight-gradient computation (`delay_wgrad_compute`) provides\nadditional overlap by deferring wgrad to overlap with the next layer's forward.\n\nBridge supports two dispatcher paths:\n\n| Dispatcher | Backend | When to use |\n|---|---|---|\n| `alltoall` | Standard MoE all-to-all | Default, broadest compatibility |\n| `flex` | DeepEP or HybridEP | Higher overlap on Ampere\u002FHopper\u002FBlackwell |\n\n## Quick Decision\n\nUse EP overlap when:\n\n- the model is MoE with `EP > 1`\n- expert dispatch\u002Fcombine communication is a meaningful part of step time\n- you have memory headroom and are tuning for throughput\n\nPrefer:\n\n- `alltoall` dispatcher for the first rollout (broader compatibility)\n- `flex` + DeepEP\u002FHybridEP when running on supported GPUs and seeking\n  additional gains\n\nAvoid EP overlap when:\n\n- full activation recompute is enabled\n- `moe_shared_expert_overlap` is enabled\n- the run is still being brought up for correctness\n- PyTorch \u003C 2.6.0\n\nExpected outcome:\n\n- if all-to-all dispatch is a clear profile bottleneck, overlap can produce a\n  modest to meaningful speedup\n- if the run is tiny, communication-light, or dominated by another wall, the\n  gain may be negligible\n\n## Correctness-First alltoall Benchmark\n\nFor the plain EP-overlap isolation benchmark, keep flex dispatch and delayed\nwgrad disabled. The measured shape was Qwen3 MoE 30B-A3B SFT on 16 H100 GPUs:\n`EP=16`, `alltoall`, BF16, global batch size 1024, CUDA graphs disabled,\n`moe_permute_fusion=false`, measured over iterations 3-8.\n\nUse these overrides for the plain-overlap case:\n\n```bash\n--cuda_graph_impl none \\\n--moe_flex_dispatcher_backend None \\\n--moe_a2a_overlap false \\\ncomm_overlap.overlap_moe_expert_parallel_comm=true \\\ncomm_overlap.delay_wgrad_compute=false \\\nmodel.moe_shared_expert_overlap=false\n```\n\nDo not use `--moe_a2a_overlap true` for this isolation test: the performance\nharness helper enables both `overlap_moe_expert_parallel_comm` and\n`delay_wgrad_compute`, so it does not isolate plain EP overlap.\n\nSteady-window timing from that benchmark:\n\n| Case | Steady mean | Relative |\n|---|---:|---:|\n| no EP overlap | 41.25s | 1.000x |\n| EP overlap | 31.31s | 1.317x |\n| EP overlap plus `delay_wgrad_compute` | 31.20s | 1.322x |\n\nThis is evidence for enabling plain EP overlap on this inter-node all-to-all\nshape. It does not show a meaningful independent win from delayed wgrad, and it\ndoes not validate fused MoE permutation because that path was disabled for the\nruntime stack.\n\n## Enablement\n\n### alltoall dispatcher\n\n```python\ncfg.comm_overlap.overlap_moe_expert_parallel_comm = True\ncfg.comm_overlap.delay_wgrad_compute = False\ncfg.model.moe_shared_expert_overlap = False\n\ncfg.model.expert_model_parallel_size = 8\ncfg.model.num_moe_experts = 64\ncfg.model.moe_token_dispatcher_type = \"alltoall\"\ncfg.model.bf16 = True\ncfg.model.fp16 = False\n```\n\nEnable `delay_wgrad_compute=True` only after the plain overlap path is known to\nwork and its extra compatibility constraints have been checked.\n\n### flex dispatcher (DeepEP or HybridEP)\n\n```python\nfrom megatron.bridge.training.flex_dispatcher_backend import apply_flex_dispatcher_backend\n\ncfg.comm_overlap.overlap_moe_expert_parallel_comm = True\ncfg.comm_overlap.delay_wgrad_compute = True\ncfg.model.moe_shared_expert_overlap = False\n\napply_flex_dispatcher_backend(cfg.model, moe_flex_dispatcher_backend=\"deepep\")\n# or: apply_flex_dispatcher_backend(cfg.model, moe_flex_dispatcher_backend=\"hybridep\")\n```\n\n## Compatibility And Constraints\n\n- `expert_model_parallel_size > 1`\n- `num_moe_experts > 1`\n- `moe_token_dispatcher_type` must be `\"alltoall\"` or `\"flex\"`\n- `moe_shared_expert_overlap = False`\n- Base precision is BF16 or FP16\n- PyTorch `>= 2.6.0`\n- If `PP > 1`, `virtual_pipeline_model_parallel_size` must be set\n- `recompute_granularity != \"full\"`, `recompute_method = None`,\n  `recompute_num_layers = None`\n- `mtp_num_layers` must be `None` or `1`\n- `delay_wgrad_compute` requires `overlap_moe_expert_parallel_comm` as a\n  prerequisite\n- `delay_wgrad_compute` with `overlap_grad_reduce` requires TE >= 2.7.0\n- `delay_wgrad_compute` with `gradient_accumulation_fusion` requires TE >= 2.7.0\n- CUDA graph `attn` scope + `delay_wgrad_compute` requires TE >= 2.12.0,\n  `gradient_accumulation_fusion = True`, and no attention bias\n- DeepEP: Ampere, Hopper, B200, B300 GPUs only\n- HybridEP: Ampere, Hopper, B200, B300, GB200\u002FGB300 with NVL72\n\n## Minimal Working Config\n\n```python\ncfg.comm_overlap.overlap_moe_expert_parallel_comm = True\ncfg.comm_overlap.delay_wgrad_compute = False\ncfg.model.expert_model_parallel_size = 4\ncfg.model.num_moe_experts = 64\ncfg.model.moe_token_dispatcher_type = \"alltoall\"\ncfg.model.moe_shared_expert_overlap = False\ncfg.model.bf16 = True\n```\n\nUse this as the correctness-first starting point. Add delayed wgrad, flex\ndispatch, and CUDA-graph interactions only after the plain overlap path is\nknown to work.\n\n## Minimal Runnable Command\n\nPerformance harness example inside a Slurm allocation. Keep the model,\nparallelism, dispatcher, and runtime fixed, and vary only the two overlap\noverrides:\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\nDo not use `--moe_a2a_overlap true` when separating plain EP overlap from\ndelayed wgrad: the performance harness helper enables both\n`overlap_moe_expert_parallel_comm` and `delay_wgrad_compute`.\n\nUnit test verification:\n\n```bash\nuv run python -m pytest \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py -k \"moe\" \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py -q\n```\n\n## Verification\n\n### Unit tests\n\n```bash\nuv run python -m pytest \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py -q\n```\n\n### Log checks\n\nAfter a successful run with EP overlap:\n\n1. Confirm no assertion errors during `CommOverlapConfig` finalization\n2. Confirm `overlap_moe_expert_parallel_comm` appears as `True` in the logged\n   config\n3. If using flex dispatcher, confirm `moe_token_dispatcher_type = \"flex\"` and\n   the correct backend in logs\n\n### Success criteria\n\n- Config validation passes for the selected dispatcher and overlap settings\n- Training runs complete without hangs or assertion failures\n- Throughput improves or at least does not regress for the target workload\n- Loss trajectory matches baseline (overlap should not affect convergence)\n\n## Code Anchors\n\n### Bridge overlap validation\n\n```470:505:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py\nif self.user_comm_overlap_cfg.overlap_moe_expert_parallel_comm is True:\n    assert model_cfg.expert_model_parallel_size > 1, ...\n    assert model_cfg.num_moe_experts > 1, ...\n    assert model_cfg.moe_token_dispatcher_type in [\"alltoall\", \"flex\"], ...\n    assert model_cfg.bf16 or model_cfg.fp16, ...\n    assert is_torch_min_version(\"2.6.0\"), ...\n    # ... PP + VPP check, recompute checks, shared_expert_overlap check ...\n```\n\n### Delayed wgrad validation\n\n```507:557:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py\nif self.user_comm_overlap_cfg.delay_wgrad_compute is True:\n    # TE version checks for overlap_grad_reduce and gradient_accumulation_fusion\n    # CUDA graph scope validations for delayed wgrad\n    assert overlap_moe_expert_parallel_comm, ...\n```\n\n### Flex-dispatcher activation\n\n```27:72:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fflex_dispatcher_backend.py\ndef apply_flex_dispatcher_backend(...):\n    # GPU architecture check for DeepEP \u002F HybridEP\n    model_config.moe_token_dispatcher_type = \"flex\"\n    model_config.moe_flex_dispatcher_backend = moe_flex_dispatcher_backend\n    model_config.moe_shared_expert_overlap = False\n```\n\n### Perf harness override\n\n```149:156:scripts\u002Fperformance\u002Futils\u002Foverrides.py\ndef _set_moe_a2a_overlap_overrides(recipe, moe_a2a_overlap=False):\n    if moe_a2a_overlap:\n        recipe.comm_overlap.overlap_moe_expert_parallel_comm = True\n        recipe.comm_overlap.delay_wgrad_compute = True\n        recipe.model.moe_shared_expert_overlap = False\n```\n\n### Tests\n\n| File | Coverage |\n|---|---|\n| `tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py` | EP overlap validation, delayed wgrad, CUDA graph + wgrad interaction |\n| `tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py` | DeepEP\u002FHybridEP helper activation and GPU gating |\n\n## Failure Diagnosis\n\n| Symptom | Likely Cause | How To Confirm | Fix |\n|---|---|---|---|\n| assert `expert_model_parallel_size > 1` | EP not configured | Check `expert_model_parallel_size` | Set EP > 1 |\n| assert `moe_token_dispatcher_type` | Wrong dispatcher | Check dispatcher type | Use `\"alltoall\"` or `\"flex\"` |\n| assert on BF16\u002FFP16 | Wrong precision | Check `bf16` and `fp16` | Set `bf16 = True` |\n| hang during training | PyTorch \u003C 2.6 | Check PyTorch version | Upgrade to >= 2.6.0 |\n| assert `virtual_pipeline_model_parallel_size` | PP > 1 without VPP | Check PP and VPP config | Set VPP when PP > 1 |\n| assert `recompute_granularity` | Full recompute enabled | Check recompute settings | Disable full recompute |\n| assert `overlap_moe_expert_parallel_comm required` | delayed wgrad without EP overlap | Check `delay_wgrad_compute` without overlap | Enable EP overlap first |\n| assert `gradient_accumulation_fusion` | CUDA graph + delayed wgrad | Check graph scope + wgrad settings | Enable `gradient_accumulation_fusion` |\n| assert on attention bias | CUDA graph attn + delayed wgrad + bias | Check `add_bias_linear` \u002F `add_qkv_bias` | Disable attention bias |\n| no throughput gain from flex dispatcher | `apply_flex_dispatcher_backend` not called | Check `moe_token_dispatcher_type` in logs | Call `apply_flex_dispatcher_backend(...)` |\n| DeepEP\u002FHybridEP silently skipped | Unsupported GPU | Check warning logs | Run on Ampere\u002FHopper\u002FBlackwell |\n\n## Known Limitations\n\n- Setting `moe_flex_dispatcher_backend` alone does not activate flex dispatch —\n  you must call `apply_flex_dispatcher_backend(...)`.\n- Public recipes are often conservative and leave MoE overlap disabled by\n  default.\n- End-to-end throughput gains have not yet been measured in a controlled Bridge\n  experiment for every model family. Code validation is stronger than a single\n  universal performance claim.\n- MoE overlap and shared-expert overlap are mutually exclusive.\n- CUDA graph plus delayed wgrad is a multi-constraint path that requires\n  careful TE version and scope validation.\n",{"data":35,"body":37},{"name":4,"description":6,"license":26,"when_to_use":36},"Enabling EP overlap to hide dispatch\u002Fcombine latency, or tracing a throughput regression to an EP overlap config change; 'overlap_moe_expert_parallel_comm', 'delay_wgrad_compute', 'flex dispatcher', 'DeepEP overlap', 'HybridEP overlap'.",{"type":38,"children":39},"root",[40,49,56,71,77,92,97,174,180,185,209,214,237,242,271,276,289,295,323,328,427,455,460,547,552,558,565,650,663,669,736,742,968,974,1031,1036,1042,1047,1337,1362,1367,1448,1454,1460,1516,1522,1527,1577,1583,1606,1612,1618,1683,1689,1730,1736,1785,1791,1840,1846,1902,1908,2296,2302,2344],{"type":41,"tag":42,"props":43,"children":45},"element","h1",{"id":44},"moe-expert-parallel-overlap-skill",[46],{"type":47,"value":48},"text","MoE Expert-Parallel Overlap Skill",{"type":41,"tag":50,"props":51,"children":53},"h2",{"id":52},"references",[54],{"type":47,"value":55},"References",{"type":41,"tag":57,"props":58,"children":59},"ul",{},[60,66],{"type":41,"tag":61,"props":62,"children":63},"li",{},[64],{"type":47,"value":65},"Stable docs: @docs\u002Ftraining\u002Fcommunication-overlap.md",{"type":41,"tag":61,"props":67,"children":68},{},[69],{"type":47,"value":70},"Structured metadata: @skills\u002Fnemo-mbridge-perf-expert-parallel-overlap\u002Fcard.yaml",{"type":41,"tag":50,"props":72,"children":74},{"id":73},"what-it-is",[75],{"type":47,"value":76},"What It Is",{"type":41,"tag":78,"props":79,"children":80},"p",{},[81,83,90],{"type":47,"value":82},"Expert-parallel (EP) overlap hides the cost of token dispatch\u002Fcombine all-to-all\ncommunication by running it concurrently with expert FFN compute. Optionally,\ndelayed expert weight-gradient computation (",{"type":41,"tag":84,"props":85,"children":87},"code",{"className":86},[],[88],{"type":47,"value":89},"delay_wgrad_compute",{"type":47,"value":91},") provides\nadditional overlap by deferring wgrad to overlap with the next layer's forward.",{"type":41,"tag":78,"props":93,"children":94},{},[95],{"type":47,"value":96},"Bridge supports two dispatcher paths:",{"type":41,"tag":98,"props":99,"children":100},"table",{},[101,125],{"type":41,"tag":102,"props":103,"children":104},"thead",{},[105],{"type":41,"tag":106,"props":107,"children":108},"tr",{},[109,115,120],{"type":41,"tag":110,"props":111,"children":112},"th",{},[113],{"type":47,"value":114},"Dispatcher",{"type":41,"tag":110,"props":116,"children":117},{},[118],{"type":47,"value":119},"Backend",{"type":41,"tag":110,"props":121,"children":122},{},[123],{"type":47,"value":124},"When to use",{"type":41,"tag":126,"props":127,"children":128},"tbody",{},[129,152],{"type":41,"tag":106,"props":130,"children":131},{},[132,142,147],{"type":41,"tag":133,"props":134,"children":135},"td",{},[136],{"type":41,"tag":84,"props":137,"children":139},{"className":138},[],[140],{"type":47,"value":141},"alltoall",{"type":41,"tag":133,"props":143,"children":144},{},[145],{"type":47,"value":146},"Standard MoE all-to-all",{"type":41,"tag":133,"props":148,"children":149},{},[150],{"type":47,"value":151},"Default, broadest compatibility",{"type":41,"tag":106,"props":153,"children":154},{},[155,164,169],{"type":41,"tag":133,"props":156,"children":157},{},[158],{"type":41,"tag":84,"props":159,"children":161},{"className":160},[],[162],{"type":47,"value":163},"flex",{"type":41,"tag":133,"props":165,"children":166},{},[167],{"type":47,"value":168},"DeepEP or HybridEP",{"type":41,"tag":133,"props":170,"children":171},{},[172],{"type":47,"value":173},"Higher overlap on Ampere\u002FHopper\u002FBlackwell",{"type":41,"tag":50,"props":175,"children":177},{"id":176},"quick-decision",[178],{"type":47,"value":179},"Quick Decision",{"type":41,"tag":78,"props":181,"children":182},{},[183],{"type":47,"value":184},"Use EP overlap when:",{"type":41,"tag":57,"props":186,"children":187},{},[188,199,204],{"type":41,"tag":61,"props":189,"children":190},{},[191,193],{"type":47,"value":192},"the model is MoE with ",{"type":41,"tag":84,"props":194,"children":196},{"className":195},[],[197],{"type":47,"value":198},"EP > 1",{"type":41,"tag":61,"props":200,"children":201},{},[202],{"type":47,"value":203},"expert dispatch\u002Fcombine communication is a meaningful part of step time",{"type":41,"tag":61,"props":205,"children":206},{},[207],{"type":47,"value":208},"you have memory headroom and are tuning for throughput",{"type":41,"tag":78,"props":210,"children":211},{},[212],{"type":47,"value":213},"Prefer:",{"type":41,"tag":57,"props":215,"children":216},{},[217,227],{"type":41,"tag":61,"props":218,"children":219},{},[220,225],{"type":41,"tag":84,"props":221,"children":223},{"className":222},[],[224],{"type":47,"value":141},{"type":47,"value":226}," dispatcher for the first rollout (broader compatibility)",{"type":41,"tag":61,"props":228,"children":229},{},[230,235],{"type":41,"tag":84,"props":231,"children":233},{"className":232},[],[234],{"type":47,"value":163},{"type":47,"value":236}," + DeepEP\u002FHybridEP when running on supported GPUs and seeking\nadditional gains",{"type":41,"tag":78,"props":238,"children":239},{},[240],{"type":47,"value":241},"Avoid EP overlap when:",{"type":41,"tag":57,"props":243,"children":244},{},[245,250,261,266],{"type":41,"tag":61,"props":246,"children":247},{},[248],{"type":47,"value":249},"full activation recompute is enabled",{"type":41,"tag":61,"props":251,"children":252},{},[253,259],{"type":41,"tag":84,"props":254,"children":256},{"className":255},[],[257],{"type":47,"value":258},"moe_shared_expert_overlap",{"type":47,"value":260}," is enabled",{"type":41,"tag":61,"props":262,"children":263},{},[264],{"type":47,"value":265},"the run is still being brought up for correctness",{"type":41,"tag":61,"props":267,"children":268},{},[269],{"type":47,"value":270},"PyTorch \u003C 2.6.0",{"type":41,"tag":78,"props":272,"children":273},{},[274],{"type":47,"value":275},"Expected outcome:",{"type":41,"tag":57,"props":277,"children":278},{},[279,284],{"type":41,"tag":61,"props":280,"children":281},{},[282],{"type":47,"value":283},"if all-to-all dispatch is a clear profile bottleneck, overlap can produce a\nmodest to meaningful speedup",{"type":41,"tag":61,"props":285,"children":286},{},[287],{"type":47,"value":288},"if the run is tiny, communication-light, or dominated by another wall, the\ngain may be negligible",{"type":41,"tag":50,"props":290,"children":292},{"id":291},"correctness-first-alltoall-benchmark",[293],{"type":47,"value":294},"Correctness-First alltoall Benchmark",{"type":41,"tag":78,"props":296,"children":297},{},[298,300,306,308,313,315,321],{"type":47,"value":299},"For the plain EP-overlap isolation benchmark, keep flex dispatch and delayed\nwgrad disabled. The measured shape was Qwen3 MoE 30B-A3B SFT on 16 H100 GPUs:\n",{"type":41,"tag":84,"props":301,"children":303},{"className":302},[],[304],{"type":47,"value":305},"EP=16",{"type":47,"value":307},", ",{"type":41,"tag":84,"props":309,"children":311},{"className":310},[],[312],{"type":47,"value":141},{"type":47,"value":314},", BF16, global batch size 1024, CUDA graphs disabled,\n",{"type":41,"tag":84,"props":316,"children":318},{"className":317},[],[319],{"type":47,"value":320},"moe_permute_fusion=false",{"type":47,"value":322},", measured over iterations 3-8.",{"type":41,"tag":78,"props":324,"children":325},{},[326],{"type":47,"value":327},"Use these overrides for the plain-overlap case:",{"type":41,"tag":329,"props":330,"children":335},"pre",{"className":331,"code":332,"language":333,"meta":334,"style":334},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","--cuda_graph_impl none \\\n--moe_flex_dispatcher_backend None \\\n--moe_a2a_overlap false \\\ncomm_overlap.overlap_moe_expert_parallel_comm=true \\\ncomm_overlap.delay_wgrad_compute=false \\\nmodel.moe_shared_expert_overlap=false\n","bash","",[336],{"type":41,"tag":84,"props":337,"children":338},{"__ignoreMap":334},[339,363,381,400,409,418],{"type":41,"tag":340,"props":341,"children":344},"span",{"class":342,"line":343},"line",1,[345,351,357],{"type":41,"tag":340,"props":346,"children":348},{"style":347},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[349],{"type":47,"value":350},"--cuda_graph_impl",{"type":41,"tag":340,"props":352,"children":354},{"style":353},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[355],{"type":47,"value":356}," none",{"type":41,"tag":340,"props":358,"children":360},{"style":359},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[361],{"type":47,"value":362}," \\\n",{"type":41,"tag":340,"props":364,"children":366},{"class":342,"line":365},2,[367,372,377],{"type":41,"tag":340,"props":368,"children":369},{"style":359},[370],{"type":47,"value":371},"--moe_flex_dispatcher_backend ",{"type":41,"tag":340,"props":373,"children":374},{"style":353},[375],{"type":47,"value":376},"None",{"type":41,"tag":340,"props":378,"children":379},{"style":359},[380],{"type":47,"value":362},{"type":41,"tag":340,"props":382,"children":384},{"class":342,"line":383},3,[385,390,396],{"type":41,"tag":340,"props":386,"children":387},{"style":359},[388],{"type":47,"value":389},"--moe_a2a_overlap ",{"type":41,"tag":340,"props":391,"children":393},{"style":392},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[394],{"type":47,"value":395},"false",{"type":41,"tag":340,"props":397,"children":398},{"style":359},[399],{"type":47,"value":362},{"type":41,"tag":340,"props":401,"children":403},{"class":342,"line":402},4,[404],{"type":41,"tag":340,"props":405,"children":406},{"style":359},[407],{"type":47,"value":408},"comm_overlap.overlap_moe_expert_parallel_comm=true \\\n",{"type":41,"tag":340,"props":410,"children":412},{"class":342,"line":411},5,[413],{"type":41,"tag":340,"props":414,"children":415},{"style":359},[416],{"type":47,"value":417},"comm_overlap.delay_wgrad_compute=false \\\n",{"type":41,"tag":340,"props":419,"children":421},{"class":342,"line":420},6,[422],{"type":41,"tag":340,"props":423,"children":424},{"style":359},[425],{"type":47,"value":426},"model.moe_shared_expert_overlap=false\n",{"type":41,"tag":78,"props":428,"children":429},{},[430,432,438,440,446,448,453],{"type":47,"value":431},"Do not use ",{"type":41,"tag":84,"props":433,"children":435},{"className":434},[],[436],{"type":47,"value":437},"--moe_a2a_overlap true",{"type":47,"value":439}," for this isolation test: the performance\nharness helper enables both ",{"type":41,"tag":84,"props":441,"children":443},{"className":442},[],[444],{"type":47,"value":445},"overlap_moe_expert_parallel_comm",{"type":47,"value":447}," and\n",{"type":41,"tag":84,"props":449,"children":451},{"className":450},[],[452],{"type":47,"value":89},{"type":47,"value":454},", so it does not isolate plain EP overlap.",{"type":41,"tag":78,"props":456,"children":457},{},[458],{"type":47,"value":459},"Steady-window timing from that benchmark:",{"type":41,"tag":98,"props":461,"children":462},{},[463,485],{"type":41,"tag":102,"props":464,"children":465},{},[466],{"type":41,"tag":106,"props":467,"children":468},{},[469,474,480],{"type":41,"tag":110,"props":470,"children":471},{},[472],{"type":47,"value":473},"Case",{"type":41,"tag":110,"props":475,"children":477},{"align":476},"right",[478],{"type":47,"value":479},"Steady mean",{"type":41,"tag":110,"props":481,"children":482},{"align":476},[483],{"type":47,"value":484},"Relative",{"type":41,"tag":126,"props":486,"children":487},{},[488,506,524],{"type":41,"tag":106,"props":489,"children":490},{},[491,496,501],{"type":41,"tag":133,"props":492,"children":493},{},[494],{"type":47,"value":495},"no EP overlap",{"type":41,"tag":133,"props":497,"children":498},{"align":476},[499],{"type":47,"value":500},"41.25s",{"type":41,"tag":133,"props":502,"children":503},{"align":476},[504],{"type":47,"value":505},"1.000x",{"type":41,"tag":106,"props":507,"children":508},{},[509,514,519],{"type":41,"tag":133,"props":510,"children":511},{},[512],{"type":47,"value":513},"EP overlap",{"type":41,"tag":133,"props":515,"children":516},{"align":476},[517],{"type":47,"value":518},"31.31s",{"type":41,"tag":133,"props":520,"children":521},{"align":476},[522],{"type":47,"value":523},"1.317x",{"type":41,"tag":106,"props":525,"children":526},{},[527,537,542],{"type":41,"tag":133,"props":528,"children":529},{},[530,532],{"type":47,"value":531},"EP overlap plus ",{"type":41,"tag":84,"props":533,"children":535},{"className":534},[],[536],{"type":47,"value":89},{"type":41,"tag":133,"props":538,"children":539},{"align":476},[540],{"type":47,"value":541},"31.20s",{"type":41,"tag":133,"props":543,"children":544},{"align":476},[545],{"type":47,"value":546},"1.322x",{"type":41,"tag":78,"props":548,"children":549},{},[550],{"type":47,"value":551},"This is evidence for enabling plain EP overlap on this inter-node all-to-all\nshape. It does not show a meaningful independent win from delayed wgrad, and it\ndoes not validate fused MoE permutation because that path was disabled for the\nruntime stack.",{"type":41,"tag":50,"props":553,"children":555},{"id":554},"enablement",[556],{"type":47,"value":557},"Enablement",{"type":41,"tag":559,"props":560,"children":562},"h3",{"id":561},"alltoall-dispatcher",[563],{"type":47,"value":564},"alltoall dispatcher",{"type":41,"tag":329,"props":566,"children":570},{"className":567,"code":568,"language":569,"meta":334,"style":334},"language-python shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","cfg.comm_overlap.overlap_moe_expert_parallel_comm = True\ncfg.comm_overlap.delay_wgrad_compute = False\ncfg.model.moe_shared_expert_overlap = False\n\ncfg.model.expert_model_parallel_size = 8\ncfg.model.num_moe_experts = 64\ncfg.model.moe_token_dispatcher_type = \"alltoall\"\ncfg.model.bf16 = True\ncfg.model.fp16 = False\n","python",[571],{"type":41,"tag":84,"props":572,"children":573},{"__ignoreMap":334},[574,582,590,598,607,615,623,632,641],{"type":41,"tag":340,"props":575,"children":576},{"class":342,"line":343},[577],{"type":41,"tag":340,"props":578,"children":579},{},[580],{"type":47,"value":581},"cfg.comm_overlap.overlap_moe_expert_parallel_comm = True\n",{"type":41,"tag":340,"props":583,"children":584},{"class":342,"line":365},[585],{"type":41,"tag":340,"props":586,"children":587},{},[588],{"type":47,"value":589},"cfg.comm_overlap.delay_wgrad_compute = False\n",{"type":41,"tag":340,"props":591,"children":592},{"class":342,"line":383},[593],{"type":41,"tag":340,"props":594,"children":595},{},[596],{"type":47,"value":597},"cfg.model.moe_shared_expert_overlap = False\n",{"type":41,"tag":340,"props":599,"children":600},{"class":342,"line":402},[601],{"type":41,"tag":340,"props":602,"children":604},{"emptyLinePlaceholder":603},true,[605],{"type":47,"value":606},"\n",{"type":41,"tag":340,"props":608,"children":609},{"class":342,"line":411},[610],{"type":41,"tag":340,"props":611,"children":612},{},[613],{"type":47,"value":614},"cfg.model.expert_model_parallel_size = 8\n",{"type":41,"tag":340,"props":616,"children":617},{"class":342,"line":420},[618],{"type":41,"tag":340,"props":619,"children":620},{},[621],{"type":47,"value":622},"cfg.model.num_moe_experts = 64\n",{"type":41,"tag":340,"props":624,"children":626},{"class":342,"line":625},7,[627],{"type":41,"tag":340,"props":628,"children":629},{},[630],{"type":47,"value":631},"cfg.model.moe_token_dispatcher_type = \"alltoall\"\n",{"type":41,"tag":340,"props":633,"children":635},{"class":342,"line":634},8,[636],{"type":41,"tag":340,"props":637,"children":638},{},[639],{"type":47,"value":640},"cfg.model.bf16 = True\n",{"type":41,"tag":340,"props":642,"children":644},{"class":342,"line":643},9,[645],{"type":41,"tag":340,"props":646,"children":647},{},[648],{"type":47,"value":649},"cfg.model.fp16 = False\n",{"type":41,"tag":78,"props":651,"children":652},{},[653,655,661],{"type":47,"value":654},"Enable ",{"type":41,"tag":84,"props":656,"children":658},{"className":657},[],[659],{"type":47,"value":660},"delay_wgrad_compute=True",{"type":47,"value":662}," only after the plain overlap path is known to\nwork and its extra compatibility constraints have been checked.",{"type":41,"tag":559,"props":664,"children":666},{"id":665},"flex-dispatcher-deepep-or-hybridep",[667],{"type":47,"value":668},"flex dispatcher (DeepEP or HybridEP)",{"type":41,"tag":329,"props":670,"children":672},{"className":567,"code":671,"language":569,"meta":334,"style":334},"from megatron.bridge.training.flex_dispatcher_backend import apply_flex_dispatcher_backend\n\ncfg.comm_overlap.overlap_moe_expert_parallel_comm = True\ncfg.comm_overlap.delay_wgrad_compute = True\ncfg.model.moe_shared_expert_overlap = False\n\napply_flex_dispatcher_backend(cfg.model, moe_flex_dispatcher_backend=\"deepep\")\n# or: apply_flex_dispatcher_backend(cfg.model, moe_flex_dispatcher_backend=\"hybridep\")\n",[673],{"type":41,"tag":84,"props":674,"children":675},{"__ignoreMap":334},[676,684,691,698,706,713,720,728],{"type":41,"tag":340,"props":677,"children":678},{"class":342,"line":343},[679],{"type":41,"tag":340,"props":680,"children":681},{},[682],{"type":47,"value":683},"from megatron.bridge.training.flex_dispatcher_backend import apply_flex_dispatcher_backend\n",{"type":41,"tag":340,"props":685,"children":686},{"class":342,"line":365},[687],{"type":41,"tag":340,"props":688,"children":689},{"emptyLinePlaceholder":603},[690],{"type":47,"value":606},{"type":41,"tag":340,"props":692,"children":693},{"class":342,"line":383},[694],{"type":41,"tag":340,"props":695,"children":696},{},[697],{"type":47,"value":581},{"type":41,"tag":340,"props":699,"children":700},{"class":342,"line":402},[701],{"type":41,"tag":340,"props":702,"children":703},{},[704],{"type":47,"value":705},"cfg.comm_overlap.delay_wgrad_compute = True\n",{"type":41,"tag":340,"props":707,"children":708},{"class":342,"line":411},[709],{"type":41,"tag":340,"props":710,"children":711},{},[712],{"type":47,"value":597},{"type":41,"tag":340,"props":714,"children":715},{"class":342,"line":420},[716],{"type":41,"tag":340,"props":717,"children":718},{"emptyLinePlaceholder":603},[719],{"type":47,"value":606},{"type":41,"tag":340,"props":721,"children":722},{"class":342,"line":625},[723],{"type":41,"tag":340,"props":724,"children":725},{},[726],{"type":47,"value":727},"apply_flex_dispatcher_backend(cfg.model, moe_flex_dispatcher_backend=\"deepep\")\n",{"type":41,"tag":340,"props":729,"children":730},{"class":342,"line":634},[731],{"type":41,"tag":340,"props":732,"children":733},{},[734],{"type":47,"value":735},"# or: apply_flex_dispatcher_backend(cfg.model, moe_flex_dispatcher_backend=\"hybridep\")\n",{"type":41,"tag":50,"props":737,"children":739},{"id":738},"compatibility-and-constraints",[740],{"type":47,"value":741},"Compatibility And Constraints",{"type":41,"tag":57,"props":743,"children":744},{},[745,754,763,788,797,802,813,833,857,879,896,914,930,958,963],{"type":41,"tag":61,"props":746,"children":747},{},[748],{"type":41,"tag":84,"props":749,"children":751},{"className":750},[],[752],{"type":47,"value":753},"expert_model_parallel_size > 1",{"type":41,"tag":61,"props":755,"children":756},{},[757],{"type":41,"tag":84,"props":758,"children":760},{"className":759},[],[761],{"type":47,"value":762},"num_moe_experts > 1",{"type":41,"tag":61,"props":764,"children":765},{},[766,772,774,780,782],{"type":41,"tag":84,"props":767,"children":769},{"className":768},[],[770],{"type":47,"value":771},"moe_token_dispatcher_type",{"type":47,"value":773}," must be ",{"type":41,"tag":84,"props":775,"children":777},{"className":776},[],[778],{"type":47,"value":779},"\"alltoall\"",{"type":47,"value":781}," or ",{"type":41,"tag":84,"props":783,"children":785},{"className":784},[],[786],{"type":47,"value":787},"\"flex\"",{"type":41,"tag":61,"props":789,"children":790},{},[791],{"type":41,"tag":84,"props":792,"children":794},{"className":793},[],[795],{"type":47,"value":796},"moe_shared_expert_overlap = False",{"type":41,"tag":61,"props":798,"children":799},{},[800],{"type":47,"value":801},"Base precision is BF16 or FP16",{"type":41,"tag":61,"props":803,"children":804},{},[805,807],{"type":47,"value":806},"PyTorch ",{"type":41,"tag":84,"props":808,"children":810},{"className":809},[],[811],{"type":47,"value":812},">= 2.6.0",{"type":41,"tag":61,"props":814,"children":815},{},[816,818,824,825,831],{"type":47,"value":817},"If ",{"type":41,"tag":84,"props":819,"children":821},{"className":820},[],[822],{"type":47,"value":823},"PP > 1",{"type":47,"value":307},{"type":41,"tag":84,"props":826,"children":828},{"className":827},[],[829],{"type":47,"value":830},"virtual_pipeline_model_parallel_size",{"type":47,"value":832}," must be set",{"type":41,"tag":61,"props":834,"children":835},{},[836,842,843,849,851],{"type":41,"tag":84,"props":837,"children":839},{"className":838},[],[840],{"type":47,"value":841},"recompute_granularity != \"full\"",{"type":47,"value":307},{"type":41,"tag":84,"props":844,"children":846},{"className":845},[],[847],{"type":47,"value":848},"recompute_method = None",{"type":47,"value":850},",\n",{"type":41,"tag":84,"props":852,"children":854},{"className":853},[],[855],{"type":47,"value":856},"recompute_num_layers = None",{"type":41,"tag":61,"props":858,"children":859},{},[860,866,867,872,873],{"type":41,"tag":84,"props":861,"children":863},{"className":862},[],[864],{"type":47,"value":865},"mtp_num_layers",{"type":47,"value":773},{"type":41,"tag":84,"props":868,"children":870},{"className":869},[],[871],{"type":47,"value":376},{"type":47,"value":781},{"type":41,"tag":84,"props":874,"children":876},{"className":875},[],[877],{"type":47,"value":878},"1",{"type":41,"tag":61,"props":880,"children":881},{},[882,887,889,894],{"type":41,"tag":84,"props":883,"children":885},{"className":884},[],[886],{"type":47,"value":89},{"type":47,"value":888}," requires ",{"type":41,"tag":84,"props":890,"children":892},{"className":891},[],[893],{"type":47,"value":445},{"type":47,"value":895}," as a\nprerequisite",{"type":41,"tag":61,"props":897,"children":898},{},[899,904,906,912],{"type":41,"tag":84,"props":900,"children":902},{"className":901},[],[903],{"type":47,"value":89},{"type":47,"value":905}," with ",{"type":41,"tag":84,"props":907,"children":909},{"className":908},[],[910],{"type":47,"value":911},"overlap_grad_reduce",{"type":47,"value":913}," requires TE >= 2.7.0",{"type":41,"tag":61,"props":915,"children":916},{},[917,922,923,929],{"type":41,"tag":84,"props":918,"children":920},{"className":919},[],[921],{"type":47,"value":89},{"type":47,"value":905},{"type":41,"tag":84,"props":924,"children":926},{"className":925},[],[927],{"type":47,"value":928},"gradient_accumulation_fusion",{"type":47,"value":913},{"type":41,"tag":61,"props":931,"children":932},{},[933,935,941,943,948,950,956],{"type":47,"value":934},"CUDA graph ",{"type":41,"tag":84,"props":936,"children":938},{"className":937},[],[939],{"type":47,"value":940},"attn",{"type":47,"value":942}," scope + ",{"type":41,"tag":84,"props":944,"children":946},{"className":945},[],[947],{"type":47,"value":89},{"type":47,"value":949}," requires TE >= 2.12.0,\n",{"type":41,"tag":84,"props":951,"children":953},{"className":952},[],[954],{"type":47,"value":955},"gradient_accumulation_fusion = True",{"type":47,"value":957},", and no attention bias",{"type":41,"tag":61,"props":959,"children":960},{},[961],{"type":47,"value":962},"DeepEP: Ampere, Hopper, B200, B300 GPUs only",{"type":41,"tag":61,"props":964,"children":965},{},[966],{"type":47,"value":967},"HybridEP: Ampere, Hopper, B200, B300, GB200\u002FGB300 with NVL72",{"type":41,"tag":50,"props":969,"children":971},{"id":970},"minimal-working-config",[972],{"type":47,"value":973},"Minimal Working Config",{"type":41,"tag":329,"props":975,"children":977},{"className":567,"code":976,"language":569,"meta":334,"style":334},"cfg.comm_overlap.overlap_moe_expert_parallel_comm = True\ncfg.comm_overlap.delay_wgrad_compute = False\ncfg.model.expert_model_parallel_size = 4\ncfg.model.num_moe_experts = 64\ncfg.model.moe_token_dispatcher_type = \"alltoall\"\ncfg.model.moe_shared_expert_overlap = False\ncfg.model.bf16 = True\n",[978],{"type":41,"tag":84,"props":979,"children":980},{"__ignoreMap":334},[981,988,995,1003,1010,1017,1024],{"type":41,"tag":340,"props":982,"children":983},{"class":342,"line":343},[984],{"type":41,"tag":340,"props":985,"children":986},{},[987],{"type":47,"value":581},{"type":41,"tag":340,"props":989,"children":990},{"class":342,"line":365},[991],{"type":41,"tag":340,"props":992,"children":993},{},[994],{"type":47,"value":589},{"type":41,"tag":340,"props":996,"children":997},{"class":342,"line":383},[998],{"type":41,"tag":340,"props":999,"children":1000},{},[1001],{"type":47,"value":1002},"cfg.model.expert_model_parallel_size = 4\n",{"type":41,"tag":340,"props":1004,"children":1005},{"class":342,"line":402},[1006],{"type":41,"tag":340,"props":1007,"children":1008},{},[1009],{"type":47,"value":622},{"type":41,"tag":340,"props":1011,"children":1012},{"class":342,"line":411},[1013],{"type":41,"tag":340,"props":1014,"children":1015},{},[1016],{"type":47,"value":631},{"type":41,"tag":340,"props":1018,"children":1019},{"class":342,"line":420},[1020],{"type":41,"tag":340,"props":1021,"children":1022},{},[1023],{"type":47,"value":597},{"type":41,"tag":340,"props":1025,"children":1026},{"class":342,"line":625},[1027],{"type":41,"tag":340,"props":1028,"children":1029},{},[1030],{"type":47,"value":640},{"type":41,"tag":78,"props":1032,"children":1033},{},[1034],{"type":47,"value":1035},"Use this as the correctness-first starting point. Add delayed wgrad, flex\ndispatch, and CUDA-graph interactions only after the plain overlap path is\nknown to work.",{"type":41,"tag":50,"props":1037,"children":1039},{"id":1038},"minimal-runnable-command",[1040],{"type":47,"value":1041},"Minimal Runnable Command",{"type":41,"tag":78,"props":1043,"children":1044},{},[1045],{"type":47,"value":1046},"Performance harness example inside a Slurm allocation. Keep the model,\nparallelism, dispatcher, and runtime fixed, and vary only the two overlap\noverrides:",{"type":41,"tag":329,"props":1048,"children":1050},{"className":331,"code":1049,"language":333,"meta":334,"style":334},"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",[1051],{"type":41,"tag":84,"props":1052,"children":1053},{"__ignoreMap":334},[1054,1081,1098,1115,1132,1149,1166,1184,1201,1217,1234,1252,1270,1288,1306,1323],{"type":41,"tag":340,"props":1055,"children":1056},{"class":342,"line":343},[1057,1062,1067,1072,1077],{"type":41,"tag":340,"props":1058,"children":1059},{"style":347},[1060],{"type":47,"value":1061},"uv",{"type":41,"tag":340,"props":1063,"children":1064},{"style":353},[1065],{"type":47,"value":1066}," run",{"type":41,"tag":340,"props":1068,"children":1069},{"style":353},[1070],{"type":47,"value":1071}," python",{"type":41,"tag":340,"props":1073,"children":1074},{"style":353},[1075],{"type":47,"value":1076}," scripts\u002Fperformance\u002Frun_script.py",{"type":41,"tag":340,"props":1078,"children":1079},{"style":359},[1080],{"type":47,"value":362},{"type":41,"tag":340,"props":1082,"children":1083},{"class":342,"line":365},[1084,1089,1094],{"type":41,"tag":340,"props":1085,"children":1086},{"style":353},[1087],{"type":47,"value":1088},"  -m",{"type":41,"tag":340,"props":1090,"children":1091},{"style":353},[1092],{"type":47,"value":1093}," qwen",{"type":41,"tag":340,"props":1095,"children":1096},{"style":359},[1097],{"type":47,"value":362},{"type":41,"tag":340,"props":1099,"children":1100},{"class":342,"line":383},[1101,1106,1111],{"type":41,"tag":340,"props":1102,"children":1103},{"style":353},[1104],{"type":47,"value":1105},"  -mr",{"type":41,"tag":340,"props":1107,"children":1108},{"style":353},[1109],{"type":47,"value":1110}," qwen3_30b_a3b",{"type":41,"tag":340,"props":1112,"children":1113},{"style":359},[1114],{"type":47,"value":362},{"type":41,"tag":340,"props":1116,"children":1117},{"class":342,"line":402},[1118,1123,1128],{"type":41,"tag":340,"props":1119,"children":1120},{"style":353},[1121],{"type":47,"value":1122},"  --task",{"type":41,"tag":340,"props":1124,"children":1125},{"style":353},[1126],{"type":47,"value":1127}," pretrain",{"type":41,"tag":340,"props":1129,"children":1130},{"style":359},[1131],{"type":47,"value":362},{"type":41,"tag":340,"props":1133,"children":1134},{"class":342,"line":411},[1135,1140,1145],{"type":41,"tag":340,"props":1136,"children":1137},{"style":353},[1138],{"type":47,"value":1139},"  -g",{"type":41,"tag":340,"props":1141,"children":1142},{"style":353},[1143],{"type":47,"value":1144}," h100",{"type":41,"tag":340,"props":1146,"children":1147},{"style":359},[1148],{"type":47,"value":362},{"type":41,"tag":340,"props":1150,"children":1151},{"class":342,"line":420},[1152,1157,1162],{"type":41,"tag":340,"props":1153,"children":1154},{"style":353},[1155],{"type":47,"value":1156},"  -c",{"type":41,"tag":340,"props":1158,"children":1159},{"style":353},[1160],{"type":47,"value":1161}," bf16",{"type":41,"tag":340,"props":1163,"children":1164},{"style":359},[1165],{"type":47,"value":362},{"type":41,"tag":340,"props":1167,"children":1168},{"class":342,"line":625},[1169,1174,1180],{"type":41,"tag":340,"props":1170,"children":1171},{"style":353},[1172],{"type":47,"value":1173},"  -ng",{"type":41,"tag":340,"props":1175,"children":1177},{"style":1176},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[1178],{"type":47,"value":1179}," 16",{"type":41,"tag":340,"props":1181,"children":1182},{"style":359},[1183],{"type":47,"value":362},{"type":41,"tag":340,"props":1185,"children":1186},{"class":342,"line":634},[1187,1192,1197],{"type":41,"tag":340,"props":1188,"children":1189},{"style":353},[1190],{"type":47,"value":1191},"  -gn",{"type":41,"tag":340,"props":1193,"children":1194},{"style":1176},[1195],{"type":47,"value":1196}," 8",{"type":41,"tag":340,"props":1198,"children":1199},{"style":359},[1200],{"type":47,"value":362},{"type":41,"tag":340,"props":1202,"children":1203},{"class":342,"line":643},[1204,1209,1213],{"type":41,"tag":340,"props":1205,"children":1206},{"style":353},[1207],{"type":47,"value":1208},"  --max_steps",{"type":41,"tag":340,"props":1210,"children":1211},{"style":1176},[1212],{"type":47,"value":1196},{"type":41,"tag":340,"props":1214,"children":1215},{"style":359},[1216],{"type":47,"value":362},{"type":41,"tag":340,"props":1218,"children":1220},{"class":342,"line":1219},10,[1221,1226,1230],{"type":41,"tag":340,"props":1222,"children":1223},{"style":353},[1224],{"type":47,"value":1225},"  --cuda_graph_impl",{"type":41,"tag":340,"props":1227,"children":1228},{"style":353},[1229],{"type":47,"value":356},{"type":41,"tag":340,"props":1231,"children":1232},{"style":359},[1233],{"type":47,"value":362},{"type":41,"tag":340,"props":1235,"children":1237},{"class":342,"line":1236},11,[1238,1243,1248],{"type":41,"tag":340,"props":1239,"children":1240},{"style":353},[1241],{"type":47,"value":1242},"  --moe_flex_dispatcher_backend",{"type":41,"tag":340,"props":1244,"children":1245},{"style":353},[1246],{"type":47,"value":1247}," None",{"type":41,"tag":340,"props":1249,"children":1250},{"style":359},[1251],{"type":47,"value":362},{"type":41,"tag":340,"props":1253,"children":1255},{"class":342,"line":1254},12,[1256,1261,1266],{"type":41,"tag":340,"props":1257,"children":1258},{"style":353},[1259],{"type":47,"value":1260},"  --moe_a2a_overlap",{"type":41,"tag":340,"props":1262,"children":1263},{"style":392},[1264],{"type":47,"value":1265}," false",{"type":41,"tag":340,"props":1267,"children":1268},{"style":359},[1269],{"type":47,"value":362},{"type":41,"tag":340,"props":1271,"children":1273},{"class":342,"line":1272},13,[1274,1279,1284],{"type":41,"tag":340,"props":1275,"children":1276},{"style":353},[1277],{"type":47,"value":1278},"  --tokenizer_type",{"type":41,"tag":340,"props":1280,"children":1281},{"style":353},[1282],{"type":47,"value":1283}," NullTokenizer",{"type":41,"tag":340,"props":1285,"children":1286},{"style":359},[1287],{"type":47,"value":362},{"type":41,"tag":340,"props":1289,"children":1291},{"class":342,"line":1290},14,[1292,1297,1302],{"type":41,"tag":340,"props":1293,"children":1294},{"style":353},[1295],{"type":47,"value":1296},"  comm_overlap.overlap_moe_expert_parallel_comm=",{"type":41,"tag":340,"props":1298,"children":1299},{"style":392},[1300],{"type":47,"value":1301},"true",{"type":41,"tag":340,"props":1303,"children":1304},{"style":359},[1305],{"type":47,"value":362},{"type":41,"tag":340,"props":1307,"children":1309},{"class":342,"line":1308},15,[1310,1315,1319],{"type":41,"tag":340,"props":1311,"children":1312},{"style":353},[1313],{"type":47,"value":1314},"  comm_overlap.delay_wgrad_compute=",{"type":41,"tag":340,"props":1316,"children":1317},{"style":392},[1318],{"type":47,"value":395},{"type":41,"tag":340,"props":1320,"children":1321},{"style":359},[1322],{"type":47,"value":362},{"type":41,"tag":340,"props":1324,"children":1326},{"class":342,"line":1325},16,[1327,1332],{"type":41,"tag":340,"props":1328,"children":1329},{"style":353},[1330],{"type":47,"value":1331},"  model.moe_shared_expert_overlap=",{"type":41,"tag":340,"props":1333,"children":1334},{"style":392},[1335],{"type":47,"value":1336},"false\n",{"type":41,"tag":78,"props":1338,"children":1339},{},[1340,1341,1346,1348,1353,1355,1360],{"type":47,"value":431},{"type":41,"tag":84,"props":1342,"children":1344},{"className":1343},[],[1345],{"type":47,"value":437},{"type":47,"value":1347}," when separating plain EP overlap from\ndelayed wgrad: the performance harness helper enables both\n",{"type":41,"tag":84,"props":1349,"children":1351},{"className":1350},[],[1352],{"type":47,"value":445},{"type":47,"value":1354}," and ",{"type":41,"tag":84,"props":1356,"children":1358},{"className":1357},[],[1359],{"type":47,"value":89},{"type":47,"value":1361},".",{"type":41,"tag":78,"props":1363,"children":1364},{},[1365],{"type":47,"value":1366},"Unit test verification:",{"type":41,"tag":329,"props":1368,"children":1370},{"className":331,"code":1369,"language":333,"meta":334,"style":334},"uv run python -m pytest \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py -k \"moe\" \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py -q\n",[1371],{"type":41,"tag":84,"props":1372,"children":1373},{"__ignoreMap":334},[1374,1403,1435],{"type":41,"tag":340,"props":1375,"children":1376},{"class":342,"line":343},[1377,1381,1385,1389,1394,1399],{"type":41,"tag":340,"props":1378,"children":1379},{"style":347},[1380],{"type":47,"value":1061},{"type":41,"tag":340,"props":1382,"children":1383},{"style":353},[1384],{"type":47,"value":1066},{"type":41,"tag":340,"props":1386,"children":1387},{"style":353},[1388],{"type":47,"value":1071},{"type":41,"tag":340,"props":1390,"children":1391},{"style":353},[1392],{"type":47,"value":1393}," -m",{"type":41,"tag":340,"props":1395,"children":1396},{"style":353},[1397],{"type":47,"value":1398}," pytest",{"type":41,"tag":340,"props":1400,"children":1401},{"style":359},[1402],{"type":47,"value":362},{"type":41,"tag":340,"props":1404,"children":1405},{"class":342,"line":365},[1406,1411,1416,1421,1426,1431],{"type":41,"tag":340,"props":1407,"children":1408},{"style":353},[1409],{"type":47,"value":1410},"  tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py",{"type":41,"tag":340,"props":1412,"children":1413},{"style":353},[1414],{"type":47,"value":1415}," -k",{"type":41,"tag":340,"props":1417,"children":1418},{"style":392},[1419],{"type":47,"value":1420}," \"",{"type":41,"tag":340,"props":1422,"children":1423},{"style":353},[1424],{"type":47,"value":1425},"moe",{"type":41,"tag":340,"props":1427,"children":1428},{"style":392},[1429],{"type":47,"value":1430},"\"",{"type":41,"tag":340,"props":1432,"children":1433},{"style":359},[1434],{"type":47,"value":362},{"type":41,"tag":340,"props":1436,"children":1437},{"class":342,"line":383},[1438,1443],{"type":41,"tag":340,"props":1439,"children":1440},{"style":353},[1441],{"type":47,"value":1442},"  tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py",{"type":41,"tag":340,"props":1444,"children":1445},{"style":353},[1446],{"type":47,"value":1447}," -q\n",{"type":41,"tag":50,"props":1449,"children":1451},{"id":1450},"verification",[1452],{"type":47,"value":1453},"Verification",{"type":41,"tag":559,"props":1455,"children":1457},{"id":1456},"unit-tests",[1458],{"type":47,"value":1459},"Unit tests",{"type":41,"tag":329,"props":1461,"children":1463},{"className":331,"code":1462,"language":333,"meta":334,"style":334},"uv run python -m pytest \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py \\\n  tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py -q\n",[1464],{"type":41,"tag":84,"props":1465,"children":1466},{"__ignoreMap":334},[1467,1494,1505],{"type":41,"tag":340,"props":1468,"children":1469},{"class":342,"line":343},[1470,1474,1478,1482,1486,1490],{"type":41,"tag":340,"props":1471,"children":1472},{"style":347},[1473],{"type":47,"value":1061},{"type":41,"tag":340,"props":1475,"children":1476},{"style":353},[1477],{"type":47,"value":1066},{"type":41,"tag":340,"props":1479,"children":1480},{"style":353},[1481],{"type":47,"value":1071},{"type":41,"tag":340,"props":1483,"children":1484},{"style":353},[1485],{"type":47,"value":1393},{"type":41,"tag":340,"props":1487,"children":1488},{"style":353},[1489],{"type":47,"value":1398},{"type":41,"tag":340,"props":1491,"children":1492},{"style":359},[1493],{"type":47,"value":362},{"type":41,"tag":340,"props":1495,"children":1496},{"class":342,"line":365},[1497,1501],{"type":41,"tag":340,"props":1498,"children":1499},{"style":353},[1500],{"type":47,"value":1410},{"type":41,"tag":340,"props":1502,"children":1503},{"style":359},[1504],{"type":47,"value":362},{"type":41,"tag":340,"props":1506,"children":1507},{"class":342,"line":383},[1508,1512],{"type":41,"tag":340,"props":1509,"children":1510},{"style":353},[1511],{"type":47,"value":1442},{"type":41,"tag":340,"props":1513,"children":1514},{"style":353},[1515],{"type":47,"value":1447},{"type":41,"tag":559,"props":1517,"children":1519},{"id":1518},"log-checks",[1520],{"type":47,"value":1521},"Log checks",{"type":41,"tag":78,"props":1523,"children":1524},{},[1525],{"type":47,"value":1526},"After a successful run with EP overlap:",{"type":41,"tag":1528,"props":1529,"children":1530},"ol",{},[1531,1544,1564],{"type":41,"tag":61,"props":1532,"children":1533},{},[1534,1536,1542],{"type":47,"value":1535},"Confirm no assertion errors during ",{"type":41,"tag":84,"props":1537,"children":1539},{"className":1538},[],[1540],{"type":47,"value":1541},"CommOverlapConfig",{"type":47,"value":1543}," finalization",{"type":41,"tag":61,"props":1545,"children":1546},{},[1547,1549,1554,1556,1562],{"type":47,"value":1548},"Confirm ",{"type":41,"tag":84,"props":1550,"children":1552},{"className":1551},[],[1553],{"type":47,"value":445},{"type":47,"value":1555}," appears as ",{"type":41,"tag":84,"props":1557,"children":1559},{"className":1558},[],[1560],{"type":47,"value":1561},"True",{"type":47,"value":1563}," in the logged\nconfig",{"type":41,"tag":61,"props":1565,"children":1566},{},[1567,1569,1575],{"type":47,"value":1568},"If using flex dispatcher, confirm ",{"type":41,"tag":84,"props":1570,"children":1572},{"className":1571},[],[1573],{"type":47,"value":1574},"moe_token_dispatcher_type = \"flex\"",{"type":47,"value":1576}," and\nthe correct backend in logs",{"type":41,"tag":559,"props":1578,"children":1580},{"id":1579},"success-criteria",[1581],{"type":47,"value":1582},"Success criteria",{"type":41,"tag":57,"props":1584,"children":1585},{},[1586,1591,1596,1601],{"type":41,"tag":61,"props":1587,"children":1588},{},[1589],{"type":47,"value":1590},"Config validation passes for the selected dispatcher and overlap settings",{"type":41,"tag":61,"props":1592,"children":1593},{},[1594],{"type":47,"value":1595},"Training runs complete without hangs or assertion failures",{"type":41,"tag":61,"props":1597,"children":1598},{},[1599],{"type":47,"value":1600},"Throughput improves or at least does not regress for the target workload",{"type":41,"tag":61,"props":1602,"children":1603},{},[1604],{"type":47,"value":1605},"Loss trajectory matches baseline (overlap should not affect convergence)",{"type":41,"tag":50,"props":1607,"children":1609},{"id":1608},"code-anchors",[1610],{"type":47,"value":1611},"Code Anchors",{"type":41,"tag":559,"props":1613,"children":1615},{"id":1614},"bridge-overlap-validation",[1616],{"type":47,"value":1617},"Bridge overlap validation",{"type":41,"tag":329,"props":1619,"children":1623},{"className":1620,"code":1621,"language":1622,"meta":334,"style":334},"language-470:505:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","if self.user_comm_overlap_cfg.overlap_moe_expert_parallel_comm is True:\n    assert model_cfg.expert_model_parallel_size > 1, ...\n    assert model_cfg.num_moe_experts > 1, ...\n    assert model_cfg.moe_token_dispatcher_type in [\"alltoall\", \"flex\"], ...\n    assert model_cfg.bf16 or model_cfg.fp16, ...\n    assert is_torch_min_version(\"2.6.0\"), ...\n    # ... PP + VPP check, recompute checks, shared_expert_overlap check ...\n","470:505:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py",[1624],{"type":41,"tag":84,"props":1625,"children":1626},{"__ignoreMap":334},[1627,1635,1643,1651,1659,1667,1675],{"type":41,"tag":340,"props":1628,"children":1629},{"class":342,"line":343},[1630],{"type":41,"tag":340,"props":1631,"children":1632},{},[1633],{"type":47,"value":1634},"if self.user_comm_overlap_cfg.overlap_moe_expert_parallel_comm is True:\n",{"type":41,"tag":340,"props":1636,"children":1637},{"class":342,"line":365},[1638],{"type":41,"tag":340,"props":1639,"children":1640},{},[1641],{"type":47,"value":1642},"    assert model_cfg.expert_model_parallel_size > 1, ...\n",{"type":41,"tag":340,"props":1644,"children":1645},{"class":342,"line":383},[1646],{"type":41,"tag":340,"props":1647,"children":1648},{},[1649],{"type":47,"value":1650},"    assert model_cfg.num_moe_experts > 1, ...\n",{"type":41,"tag":340,"props":1652,"children":1653},{"class":342,"line":402},[1654],{"type":41,"tag":340,"props":1655,"children":1656},{},[1657],{"type":47,"value":1658},"    assert model_cfg.moe_token_dispatcher_type in [\"alltoall\", \"flex\"], ...\n",{"type":41,"tag":340,"props":1660,"children":1661},{"class":342,"line":411},[1662],{"type":41,"tag":340,"props":1663,"children":1664},{},[1665],{"type":47,"value":1666},"    assert model_cfg.bf16 or model_cfg.fp16, ...\n",{"type":41,"tag":340,"props":1668,"children":1669},{"class":342,"line":420},[1670],{"type":41,"tag":340,"props":1671,"children":1672},{},[1673],{"type":47,"value":1674},"    assert is_torch_min_version(\"2.6.0\"), ...\n",{"type":41,"tag":340,"props":1676,"children":1677},{"class":342,"line":625},[1678],{"type":41,"tag":340,"props":1679,"children":1680},{},[1681],{"type":47,"value":1682},"    # ... PP + VPP check, recompute checks, shared_expert_overlap check ...\n",{"type":41,"tag":559,"props":1684,"children":1686},{"id":1685},"delayed-wgrad-validation",[1687],{"type":47,"value":1688},"Delayed wgrad validation",{"type":41,"tag":329,"props":1690,"children":1694},{"className":1691,"code":1692,"language":1693,"meta":334,"style":334},"language-507:557:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","if self.user_comm_overlap_cfg.delay_wgrad_compute is True:\n    # TE version checks for overlap_grad_reduce and gradient_accumulation_fusion\n    # CUDA graph scope validations for delayed wgrad\n    assert overlap_moe_expert_parallel_comm, ...\n","507:557:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fcomm_overlap.py",[1695],{"type":41,"tag":84,"props":1696,"children":1697},{"__ignoreMap":334},[1698,1706,1714,1722],{"type":41,"tag":340,"props":1699,"children":1700},{"class":342,"line":343},[1701],{"type":41,"tag":340,"props":1702,"children":1703},{},[1704],{"type":47,"value":1705},"if self.user_comm_overlap_cfg.delay_wgrad_compute is True:\n",{"type":41,"tag":340,"props":1707,"children":1708},{"class":342,"line":365},[1709],{"type":41,"tag":340,"props":1710,"children":1711},{},[1712],{"type":47,"value":1713},"    # TE version checks for overlap_grad_reduce and gradient_accumulation_fusion\n",{"type":41,"tag":340,"props":1715,"children":1716},{"class":342,"line":383},[1717],{"type":41,"tag":340,"props":1718,"children":1719},{},[1720],{"type":47,"value":1721},"    # CUDA graph scope validations for delayed wgrad\n",{"type":41,"tag":340,"props":1723,"children":1724},{"class":342,"line":402},[1725],{"type":41,"tag":340,"props":1726,"children":1727},{},[1728],{"type":47,"value":1729},"    assert overlap_moe_expert_parallel_comm, ...\n",{"type":41,"tag":559,"props":1731,"children":1733},{"id":1732},"flex-dispatcher-activation",[1734],{"type":47,"value":1735},"Flex-dispatcher activation",{"type":41,"tag":329,"props":1737,"children":1741},{"className":1738,"code":1739,"language":1740,"meta":334,"style":334},"language-27:72:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fflex_dispatcher_backend.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","def apply_flex_dispatcher_backend(...):\n    # GPU architecture check for DeepEP \u002F HybridEP\n    model_config.moe_token_dispatcher_type = \"flex\"\n    model_config.moe_flex_dispatcher_backend = moe_flex_dispatcher_backend\n    model_config.moe_shared_expert_overlap = False\n","27:72:src\u002Fmegatron\u002Fbridge\u002Ftraining\u002Fflex_dispatcher_backend.py",[1742],{"type":41,"tag":84,"props":1743,"children":1744},{"__ignoreMap":334},[1745,1753,1761,1769,1777],{"type":41,"tag":340,"props":1746,"children":1747},{"class":342,"line":343},[1748],{"type":41,"tag":340,"props":1749,"children":1750},{},[1751],{"type":47,"value":1752},"def apply_flex_dispatcher_backend(...):\n",{"type":41,"tag":340,"props":1754,"children":1755},{"class":342,"line":365},[1756],{"type":41,"tag":340,"props":1757,"children":1758},{},[1759],{"type":47,"value":1760},"    # GPU architecture check for DeepEP \u002F HybridEP\n",{"type":41,"tag":340,"props":1762,"children":1763},{"class":342,"line":383},[1764],{"type":41,"tag":340,"props":1765,"children":1766},{},[1767],{"type":47,"value":1768},"    model_config.moe_token_dispatcher_type = \"flex\"\n",{"type":41,"tag":340,"props":1770,"children":1771},{"class":342,"line":402},[1772],{"type":41,"tag":340,"props":1773,"children":1774},{},[1775],{"type":47,"value":1776},"    model_config.moe_flex_dispatcher_backend = moe_flex_dispatcher_backend\n",{"type":41,"tag":340,"props":1778,"children":1779},{"class":342,"line":411},[1780],{"type":41,"tag":340,"props":1781,"children":1782},{},[1783],{"type":47,"value":1784},"    model_config.moe_shared_expert_overlap = False\n",{"type":41,"tag":559,"props":1786,"children":1788},{"id":1787},"perf-harness-override",[1789],{"type":47,"value":1790},"Perf harness override",{"type":41,"tag":329,"props":1792,"children":1796},{"className":1793,"code":1794,"language":1795,"meta":334,"style":334},"language-149:156:scripts\u002Fperformance\u002Futils\u002Foverrides.py shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","def _set_moe_a2a_overlap_overrides(recipe, moe_a2a_overlap=False):\n    if moe_a2a_overlap:\n        recipe.comm_overlap.overlap_moe_expert_parallel_comm = True\n        recipe.comm_overlap.delay_wgrad_compute = True\n        recipe.model.moe_shared_expert_overlap = False\n","149:156:scripts\u002Fperformance\u002Futils\u002Foverrides.py",[1797],{"type":41,"tag":84,"props":1798,"children":1799},{"__ignoreMap":334},[1800,1808,1816,1824,1832],{"type":41,"tag":340,"props":1801,"children":1802},{"class":342,"line":343},[1803],{"type":41,"tag":340,"props":1804,"children":1805},{},[1806],{"type":47,"value":1807},"def _set_moe_a2a_overlap_overrides(recipe, moe_a2a_overlap=False):\n",{"type":41,"tag":340,"props":1809,"children":1810},{"class":342,"line":365},[1811],{"type":41,"tag":340,"props":1812,"children":1813},{},[1814],{"type":47,"value":1815},"    if moe_a2a_overlap:\n",{"type":41,"tag":340,"props":1817,"children":1818},{"class":342,"line":383},[1819],{"type":41,"tag":340,"props":1820,"children":1821},{},[1822],{"type":47,"value":1823},"        recipe.comm_overlap.overlap_moe_expert_parallel_comm = True\n",{"type":41,"tag":340,"props":1825,"children":1826},{"class":342,"line":402},[1827],{"type":41,"tag":340,"props":1828,"children":1829},{},[1830],{"type":47,"value":1831},"        recipe.comm_overlap.delay_wgrad_compute = True\n",{"type":41,"tag":340,"props":1833,"children":1834},{"class":342,"line":411},[1835],{"type":41,"tag":340,"props":1836,"children":1837},{},[1838],{"type":47,"value":1839},"        recipe.model.moe_shared_expert_overlap = False\n",{"type":41,"tag":559,"props":1841,"children":1843},{"id":1842},"tests",[1844],{"type":47,"value":1845},"Tests",{"type":41,"tag":98,"props":1847,"children":1848},{},[1849,1865],{"type":41,"tag":102,"props":1850,"children":1851},{},[1852],{"type":41,"tag":106,"props":1853,"children":1854},{},[1855,1860],{"type":41,"tag":110,"props":1856,"children":1857},{},[1858],{"type":47,"value":1859},"File",{"type":41,"tag":110,"props":1861,"children":1862},{},[1863],{"type":47,"value":1864},"Coverage",{"type":41,"tag":126,"props":1866,"children":1867},{},[1868,1885],{"type":41,"tag":106,"props":1869,"children":1870},{},[1871,1880],{"type":41,"tag":133,"props":1872,"children":1873},{},[1874],{"type":41,"tag":84,"props":1875,"children":1877},{"className":1876},[],[1878],{"type":47,"value":1879},"tests\u002Funit_tests\u002Ftraining\u002Ftest_comm_overlap.py",{"type":41,"tag":133,"props":1881,"children":1882},{},[1883],{"type":47,"value":1884},"EP overlap validation, delayed wgrad, CUDA graph + wgrad interaction",{"type":41,"tag":106,"props":1886,"children":1887},{},[1888,1897],{"type":41,"tag":133,"props":1889,"children":1890},{},[1891],{"type":41,"tag":84,"props":1892,"children":1894},{"className":1893},[],[1895],{"type":47,"value":1896},"tests\u002Funit_tests\u002Ftraining\u002Ftest_deepep.py",{"type":41,"tag":133,"props":1898,"children":1899},{},[1900],{"type":47,"value":1901},"DeepEP\u002FHybridEP helper activation and GPU gating",{"type":41,"tag":50,"props":1903,"children":1905},{"id":1904},"failure-diagnosis",[1906],{"type":47,"value":1907},"Failure Diagnosis",{"type":41,"tag":98,"props":1909,"children":1910},{},[1911,1937],{"type":41,"tag":102,"props":1912,"children":1913},{},[1914],{"type":41,"tag":106,"props":1915,"children":1916},{},[1917,1922,1927,1932],{"type":41,"tag":110,"props":1918,"children":1919},{},[1920],{"type":47,"value":1921},"Symptom",{"type":41,"tag":110,"props":1923,"children":1924},{},[1925],{"type":47,"value":1926},"Likely Cause",{"type":41,"tag":110,"props":1928,"children":1929},{},[1930],{"type":47,"value":1931},"How To Confirm",{"type":41,"tag":110,"props":1933,"children":1934},{},[1935],{"type":47,"value":1936},"Fix",{"type":41,"tag":126,"props":1938,"children":1939},{},[1940,1974,2012,2053,2076,2103,2131,2165,2196,2232,2273],{"type":41,"tag":106,"props":1941,"children":1942},{},[1943,1953,1958,1969],{"type":41,"tag":133,"props":1944,"children":1945},{},[1946,1948],{"type":47,"value":1947},"assert ",{"type":41,"tag":84,"props":1949,"children":1951},{"className":1950},[],[1952],{"type":47,"value":753},{"type":41,"tag":133,"props":1954,"children":1955},{},[1956],{"type":47,"value":1957},"EP not configured",{"type":41,"tag":133,"props":1959,"children":1960},{},[1961,1963],{"type":47,"value":1962},"Check ",{"type":41,"tag":84,"props":1964,"children":1966},{"className":1965},[],[1967],{"type":47,"value":1968},"expert_model_parallel_size",{"type":41,"tag":133,"props":1970,"children":1971},{},[1972],{"type":47,"value":1973},"Set EP > 1",{"type":41,"tag":106,"props":1975,"children":1976},{},[1977,1986,1991,1996],{"type":41,"tag":133,"props":1978,"children":1979},{},[1980,1981],{"type":47,"value":1947},{"type":41,"tag":84,"props":1982,"children":1984},{"className":1983},[],[1985],{"type":47,"value":771},{"type":41,"tag":133,"props":1987,"children":1988},{},[1989],{"type":47,"value":1990},"Wrong dispatcher",{"type":41,"tag":133,"props":1992,"children":1993},{},[1994],{"type":47,"value":1995},"Check dispatcher type",{"type":41,"tag":133,"props":1997,"children":1998},{},[1999,2001,2006,2007],{"type":47,"value":2000},"Use ",{"type":41,"tag":84,"props":2002,"children":2004},{"className":2003},[],[2005],{"type":47,"value":779},{"type":47,"value":781},{"type":41,"tag":84,"props":2008,"children":2010},{"className":2009},[],[2011],{"type":47,"value":787},{"type":41,"tag":106,"props":2013,"children":2014},{},[2015,2020,2025,2042],{"type":41,"tag":133,"props":2016,"children":2017},{},[2018],{"type":47,"value":2019},"assert on BF16\u002FFP16",{"type":41,"tag":133,"props":2021,"children":2022},{},[2023],{"type":47,"value":2024},"Wrong precision",{"type":41,"tag":133,"props":2026,"children":2027},{},[2028,2029,2035,2036],{"type":47,"value":1962},{"type":41,"tag":84,"props":2030,"children":2032},{"className":2031},[],[2033],{"type":47,"value":2034},"bf16",{"type":47,"value":1354},{"type":41,"tag":84,"props":2037,"children":2039},{"className":2038},[],[2040],{"type":47,"value":2041},"fp16",{"type":41,"tag":133,"props":2043,"children":2044},{},[2045,2047],{"type":47,"value":2046},"Set ",{"type":41,"tag":84,"props":2048,"children":2050},{"className":2049},[],[2051],{"type":47,"value":2052},"bf16 = True",{"type":41,"tag":106,"props":2054,"children":2055},{},[2056,2061,2066,2071],{"type":41,"tag":133,"props":2057,"children":2058},{},[2059],{"type":47,"value":2060},"hang during training",{"type":41,"tag":133,"props":2062,"children":2063},{},[2064],{"type":47,"value":2065},"PyTorch \u003C 2.6",{"type":41,"tag":133,"props":2067,"children":2068},{},[2069],{"type":47,"value":2070},"Check PyTorch version",{"type":41,"tag":133,"props":2072,"children":2073},{},[2074],{"type":47,"value":2075},"Upgrade to >= 2.6.0",{"type":41,"tag":106,"props":2077,"children":2078},{},[2079,2088,2093,2098],{"type":41,"tag":133,"props":2080,"children":2081},{},[2082,2083],{"type":47,"value":1947},{"type":41,"tag":84,"props":2084,"children":2086},{"className":2085},[],[2087],{"type":47,"value":830},{"type":41,"tag":133,"props":2089,"children":2090},{},[2091],{"type":47,"value":2092},"PP > 1 without VPP",{"type":41,"tag":133,"props":2094,"children":2095},{},[2096],{"type":47,"value":2097},"Check PP and VPP config",{"type":41,"tag":133,"props":2099,"children":2100},{},[2101],{"type":47,"value":2102},"Set VPP when PP > 1",{"type":41,"tag":106,"props":2104,"children":2105},{},[2106,2116,2121,2126],{"type":41,"tag":133,"props":2107,"children":2108},{},[2109,2110],{"type":47,"value":1947},{"type":41,"tag":84,"props":2111,"children":2113},{"className":2112},[],[2114],{"type":47,"value":2115},"recompute_granularity",{"type":41,"tag":133,"props":2117,"children":2118},{},[2119],{"type":47,"value":2120},"Full recompute enabled",{"type":41,"tag":133,"props":2122,"children":2123},{},[2124],{"type":47,"value":2125},"Check recompute settings",{"type":41,"tag":133,"props":2127,"children":2128},{},[2129],{"type":47,"value":2130},"Disable full recompute",{"type":41,"tag":106,"props":2132,"children":2133},{},[2134,2144,2149,2160],{"type":41,"tag":133,"props":2135,"children":2136},{},[2137,2138],{"type":47,"value":1947},{"type":41,"tag":84,"props":2139,"children":2141},{"className":2140},[],[2142],{"type":47,"value":2143},"overlap_moe_expert_parallel_comm required",{"type":41,"tag":133,"props":2145,"children":2146},{},[2147],{"type":47,"value":2148},"delayed wgrad without EP overlap",{"type":41,"tag":133,"props":2150,"children":2151},{},[2152,2153,2158],{"type":47,"value":1962},{"type":41,"tag":84,"props":2154,"children":2156},{"className":2155},[],[2157],{"type":47,"value":89},{"type":47,"value":2159}," without overlap",{"type":41,"tag":133,"props":2161,"children":2162},{},[2163],{"type":47,"value":2164},"Enable EP overlap first",{"type":41,"tag":106,"props":2166,"children":2167},{},[2168,2177,2182,2187],{"type":41,"tag":133,"props":2169,"children":2170},{},[2171,2172],{"type":47,"value":1947},{"type":41,"tag":84,"props":2173,"children":2175},{"className":2174},[],[2176],{"type":47,"value":928},{"type":41,"tag":133,"props":2178,"children":2179},{},[2180],{"type":47,"value":2181},"CUDA graph + delayed wgrad",{"type":41,"tag":133,"props":2183,"children":2184},{},[2185],{"type":47,"value":2186},"Check graph scope + wgrad settings",{"type":41,"tag":133,"props":2188,"children":2189},{},[2190,2191],{"type":47,"value":654},{"type":41,"tag":84,"props":2192,"children":2194},{"className":2193},[],[2195],{"type":47,"value":928},{"type":41,"tag":106,"props":2197,"children":2198},{},[2199,2204,2209,2227],{"type":41,"tag":133,"props":2200,"children":2201},{},[2202],{"type":47,"value":2203},"assert on attention bias",{"type":41,"tag":133,"props":2205,"children":2206},{},[2207],{"type":47,"value":2208},"CUDA graph attn + delayed wgrad + bias",{"type":41,"tag":133,"props":2210,"children":2211},{},[2212,2213,2219,2221],{"type":47,"value":1962},{"type":41,"tag":84,"props":2214,"children":2216},{"className":2215},[],[2217],{"type":47,"value":2218},"add_bias_linear",{"type":47,"value":2220}," \u002F ",{"type":41,"tag":84,"props":2222,"children":2224},{"className":2223},[],[2225],{"type":47,"value":2226},"add_qkv_bias",{"type":41,"tag":133,"props":2228,"children":2229},{},[2230],{"type":47,"value":2231},"Disable attention bias",{"type":41,"tag":106,"props":2233,"children":2234},{},[2235,2240,2251,2262],{"type":41,"tag":133,"props":2236,"children":2237},{},[2238],{"type":47,"value":2239},"no throughput gain from flex dispatcher",{"type":41,"tag":133,"props":2241,"children":2242},{},[2243,2249],{"type":41,"tag":84,"props":2244,"children":2246},{"className":2245},[],[2247],{"type":47,"value":2248},"apply_flex_dispatcher_backend",{"type":47,"value":2250}," not called",{"type":41,"tag":133,"props":2252,"children":2253},{},[2254,2255,2260],{"type":47,"value":1962},{"type":41,"tag":84,"props":2256,"children":2258},{"className":2257},[],[2259],{"type":47,"value":771},{"type":47,"value":2261}," in logs",{"type":41,"tag":133,"props":2263,"children":2264},{},[2265,2267],{"type":47,"value":2266},"Call ",{"type":41,"tag":84,"props":2268,"children":2270},{"className":2269},[],[2271],{"type":47,"value":2272},"apply_flex_dispatcher_backend(...)",{"type":41,"tag":106,"props":2274,"children":2275},{},[2276,2281,2286,2291],{"type":41,"tag":133,"props":2277,"children":2278},{},[2279],{"type":47,"value":2280},"DeepEP\u002FHybridEP silently skipped",{"type":41,"tag":133,"props":2282,"children":2283},{},[2284],{"type":47,"value":2285},"Unsupported GPU",{"type":41,"tag":133,"props":2287,"children":2288},{},[2289],{"type":47,"value":2290},"Check warning logs",{"type":41,"tag":133,"props":2292,"children":2293},{},[2294],{"type":47,"value":2295},"Run on Ampere\u002FHopper\u002FBlackwell",{"type":41,"tag":50,"props":2297,"children":2299},{"id":2298},"known-limitations",[2300],{"type":47,"value":2301},"Known Limitations",{"type":41,"tag":57,"props":2303,"children":2304},{},[2305,2324,2329,2334,2339],{"type":41,"tag":61,"props":2306,"children":2307},{},[2308,2310,2316,2318,2323],{"type":47,"value":2309},"Setting ",{"type":41,"tag":84,"props":2311,"children":2313},{"className":2312},[],[2314],{"type":47,"value":2315},"moe_flex_dispatcher_backend",{"type":47,"value":2317}," alone does not activate flex dispatch —\nyou must call ",{"type":41,"tag":84,"props":2319,"children":2321},{"className":2320},[],[2322],{"type":47,"value":2272},{"type":47,"value":1361},{"type":41,"tag":61,"props":2325,"children":2326},{},[2327],{"type":47,"value":2328},"Public recipes are often conservative and leave MoE overlap disabled by\ndefault.",{"type":41,"tag":61,"props":2330,"children":2331},{},[2332],{"type":47,"value":2333},"End-to-end throughput gains have not yet been measured in a controlled Bridge\nexperiment for every model family. Code validation is stronger than a single\nuniversal performance claim.",{"type":41,"tag":61,"props":2335,"children":2336},{},[2337],{"type":47,"value":2338},"MoE overlap and shared-expert overlap are mutually exclusive.",{"type":41,"tag":61,"props":2340,"children":2341},{},[2342],{"type":47,"value":2343},"CUDA graph plus delayed wgrad is a multi-constraint path that requires\ncareful TE version and scope validation.",{"type":41,"tag":2345,"props":2346,"children":2347},"style",{},[2348],{"type":47,"value":2349},"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":2351,"total":2453},[2352,2367,2381,2395,2407,2424,2439],{"slug":2353,"name":2353,"fn":2354,"description":2355,"org":2356,"tags":2357,"stars":23,"repoUrl":24,"updatedAt":2366},"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},[2358,2361,2364,2365],{"name":2359,"slug":2360,"type":15},"Data Analysis","data-analysis",{"name":2362,"slug":2363,"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":2368,"name":2368,"fn":2369,"description":2370,"org":2371,"tags":2372,"stars":23,"repoUrl":24,"updatedAt":2380},"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},[2373,2376,2379],{"name":2374,"slug":2375,"type":15},"Deployment","deployment",{"name":2377,"slug":2378,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":2382,"name":2382,"fn":2383,"description":2384,"org":2385,"tags":2386,"stars":23,"repoUrl":24,"updatedAt":2394},"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},[2387,2390,2391],{"name":2388,"slug":2389,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":2392,"slug":2393,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":2396,"name":2396,"fn":2397,"description":2398,"org":2399,"tags":2400,"stars":23,"repoUrl":24,"updatedAt":2406},"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},[2401,2402,2403],{"name":2359,"slug":2360,"type":15},{"name":9,"slug":8,"type":15},{"name":2404,"slug":2405,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":2408,"name":2408,"fn":2409,"description":2410,"org":2411,"tags":2412,"stars":23,"repoUrl":24,"updatedAt":2423},"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},[2413,2416,2419,2420],{"name":2414,"slug":2415,"type":15},"Automation","automation",{"name":2417,"slug":2418,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":2421,"slug":2422,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":2425,"name":2425,"fn":2426,"description":2427,"org":2428,"tags":2429,"stars":23,"repoUrl":24,"updatedAt":2438},"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},[2430,2431,2434,2435],{"name":2374,"slug":2375,"type":15},{"name":2432,"slug":2433,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":2436,"slug":2437,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":2440,"name":2440,"fn":2441,"description":2442,"org":2443,"tags":2444,"stars":23,"repoUrl":24,"updatedAt":2452},"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},[2445,2446,2449],{"name":9,"slug":8,"type":15},{"name":2447,"slug":2448,"type":15},"Quantum Computing","quantum-computing",{"name":2450,"slug":2451,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":2455,"total":2605},[2456,2474,2489,2500,2512,2526,2539,2553,2564,2573,2587,2596],{"slug":2457,"name":2457,"fn":2458,"description":2459,"org":2460,"tags":2461,"stars":2471,"repoUrl":2472,"updatedAt":2473},"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},[2462,2465,2468],{"name":2463,"slug":2464,"type":15},"Documentation","documentation",{"name":2466,"slug":2467,"type":15},"MCP","mcp",{"name":2469,"slug":2470,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":2475,"name":2475,"fn":2476,"description":2477,"org":2478,"tags":2479,"stars":2486,"repoUrl":2487,"updatedAt":2488},"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},[2480,2483,2484],{"name":2481,"slug":2482,"type":15},"Containers","containers",{"name":2374,"slug":2375,"type":15},{"name":2485,"slug":569,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":2490,"name":2490,"fn":2491,"description":2492,"org":2493,"tags":2494,"stars":2486,"repoUrl":2487,"updatedAt":2499},"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},[2495,2498],{"name":2496,"slug":2497,"type":15},"CI\u002FCD","ci-cd",{"name":2374,"slug":2375,"type":15},"2026-07-14T05:25:59.97109",{"slug":2501,"name":2501,"fn":2502,"description":2503,"org":2504,"tags":2505,"stars":2486,"repoUrl":2487,"updatedAt":2511},"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},[2506,2507,2508],{"name":2496,"slug":2497,"type":15},{"name":2374,"slug":2375,"type":15},{"name":2509,"slug":2510,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":2513,"name":2513,"fn":2514,"description":2515,"org":2516,"tags":2517,"stars":2486,"repoUrl":2487,"updatedAt":2525},"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},[2518,2521,2522],{"name":2519,"slug":2520,"type":15},"Debugging","debugging",{"name":2509,"slug":2510,"type":15},{"name":2523,"slug":2524,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":2527,"name":2527,"fn":2528,"description":2529,"org":2530,"tags":2531,"stars":2486,"repoUrl":2487,"updatedAt":2538},"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},[2532,2535],{"name":2533,"slug":2534,"type":15},"Best Practices","best-practices",{"name":2536,"slug":2537,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":2540,"name":2540,"fn":2541,"description":2542,"org":2543,"tags":2544,"stars":2486,"repoUrl":2487,"updatedAt":2552},"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},[2545,2548,2551],{"name":2546,"slug":2547,"type":15},"Machine Learning","machine-learning",{"name":2549,"slug":2550,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":2554,"name":2554,"fn":2555,"description":2556,"org":2557,"tags":2558,"stars":2486,"repoUrl":2487,"updatedAt":2563},"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},[2559,2562],{"name":2560,"slug":2561,"type":15},"QA","qa",{"name":2404,"slug":2405,"type":15},"2026-07-14T05:25:53.673039",{"slug":2565,"name":2565,"fn":2566,"description":2567,"org":2568,"tags":2569,"stars":2486,"repoUrl":2487,"updatedAt":2572},"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},[2570,2571],{"name":2374,"slug":2375,"type":15},{"name":2377,"slug":2378,"type":15},"2026-07-14T05:25:49.362534",{"slug":2574,"name":2574,"fn":2575,"description":2576,"org":2577,"tags":2578,"stars":2486,"repoUrl":2487,"updatedAt":2586},"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},[2579,2582,2583],{"name":2580,"slug":2581,"type":15},"Code Review","code-review",{"name":2509,"slug":2510,"type":15},{"name":2584,"slug":2585,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":2588,"name":2588,"fn":2589,"description":2590,"org":2591,"tags":2592,"stars":2486,"repoUrl":2487,"updatedAt":2595},"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},[2593,2594],{"name":2560,"slug":2561,"type":15},{"name":2404,"slug":2405,"type":15},"2026-07-14T05:25:54.928983",{"slug":2597,"name":2597,"fn":2598,"description":2599,"org":2600,"tags":2601,"stars":2486,"repoUrl":2487,"updatedAt":2604},"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},[2602,2603],{"name":2414,"slug":2415,"type":15},{"name":2496,"slug":2497,"type":15},"2026-07-30T05:29:03.275638",496]