[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-mbridge-perf-moe-dispatcher-selection":3,"mdc--7uhp8n-key":31,"related-org-nvidia-nemo-mbridge-perf-moe-dispatcher-selection":1025,"related-repo-nvidia-nemo-mbridge-perf-moe-dispatcher-selection":1186},{"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-dispatcher-selection","select MoE token dispatchers for NVIDIA hardware","Choose the right MoE token dispatcher (`alltoall`, DeepEP, or HybridEP) for the hardware, EP degree, and optimization stage. Summarizes patterns from DSV3, Qwen3, Qwen3-Next, and VLM bring-up work.",{"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],{"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},2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:28:46.857437","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-dispatcher-selection","---\nname: nemo-mbridge-perf-moe-dispatcher-selection\ndescription: Choose the right MoE token dispatcher (`alltoall`, DeepEP, or HybridEP) for the hardware, EP degree, and optimization stage. Summarizes patterns from DSV3, Qwen3, Qwen3-Next, and VLM bring-up work.\nlicense: Apache-2.0\nwhen_to_use: Choosing a MoE token dispatcher, or tracing a MoE regression or crash to a dispatcher config change; 'which dispatcher', 'alltoall vs DeepEP', 'HybridEP', 'MoE dispatcher', 'flex backend', 'EP dispatcher selection'.\n---\n\n# MoE Dispatcher Selection Guide\n\nStable docs: @docs\u002Ftraining\u002Fmoe-optimization.md\nCard: @skills\u002Fnemo-mbridge-perf-moe-dispatcher-selection\u002Fcard.yaml\n\n## Quick Decision\n\n### By hardware\n\n| Hardware | First choice | Why |\n|---|---|---|\n| H100 | DeepEP, if the runtime package is installed | Strong default for cross-node EP on Hopper |\n| B200 | DeepEP, if the runtime package is installed | Good first choice unless a platform-specific HybridEP path is available |\n| GB200 \u002F GB300 NVL72 | HybridEP, if the runtime package is installed | Best fit for NVLink-domain-aware dispatch and lower memory pressure |\n| Unknown or first bring-up | `alltoall` | Easiest path for correctness and debugging |\n\n### By EP degree\n\n| EP size | Guidance |\n|---|---|\n| Small EP | Dispatcher choice is usually second-order; start with `alltoall` or DeepEP |\n| Medium EP | DeepEP often becomes worthwhile |\n| Large EP | HybridEP is usually the best target on NVL72 systems |\n\n## Model-Family Patterns\n\n| Workload | Common best path | Notes |\n|---|---|---|\n| DSV3 at large scale | HybridEP on GB200 or GB300, DeepEP on H100 | Dispatcher choice matters more as EP and PP both grow |\n| Qwen3 235B | DeepEP on H100, HybridEP on GB200 | HybridEP usually wins on GB200 and often uses less memory |\n| Qwen3 30B | DeepEP | Smaller models still benefit, but the absolute gap is smaller |\n| Qwen3-Next | Close race in BF16, HybridEP stronger in FP8 or memory-tight runs | Good reminder to test, not assume |\n| MoE VLMs | Start simple, then test HybridEP on GB200-class systems | Vision workloads are sensitive to both memory and host overhead |\n\n## Rounded Evidence Summary\n\n### Backend availability gate\n\nDo not interpret a dispatcher timing until the container has proven that the\nselected backend package is available. `--moe_flex_dispatcher_backend None`\nselects the standard `alltoall` dispatcher, while `deepep` and `hybridep`\nselect `moe_token_dispatcher_type=\"flex\"` and then require their corresponding\nruntime packages at model construction time. If DeepEP or HybridEP is missing,\nrecord the import failure as an environment limitation and treat `alltoall` as\nthe only measured correctness fallback for that run.\n\n### Qwen3 30B A3B on H100\n\nA short 2026-05-17 H100 smoke run used Qwen3 30B A3B BF16, 16 GPUs, EP=16,\nthe recipe's Transformer Engine CUDA graph scopes (`moe_router`,\n`moe_preprocess`), and `model.moe_permute_fusion=false` due to a Triton JIT\ncompatibility issue in the run container. The `alltoall` fallback completed five\nsteps with 45.65 s mean step time after warmup, 132.9 mean TFLOP\u002Fs\u002FGPU after\nwarmup, final loss 11.44050, and 61.351 GB peak max allocated memory. DeepEP\nand HybridEP selected the requested flex backend in the dumped configs but\nfailed before the first iteration because the packages were not installed. This\nconfirms the availability gate; it is not a throughput ranking for flex\ndispatchers on H100.\n\n### DSV3 on GB200 or GB300\n\nThe broad trend is more important than any single row in the tracker:\n\n- plain `alltoall` is usually the conservative baseline\n- DeepEP improves that baseline once EP communication becomes visible\n- HybridEP adds another step up on NVL72 systems, especially after CUDA graphs,\n  routing improvements, and CPU-side cleanup are already in place\n\nIn practice, the stack often moves from roughly \"low-teens MFU\" territory with\nan untuned baseline into \"high-teens to low-20s MFU\" territory after the full\ndispatcher and kernel stack is tuned.\n\n### Qwen3 235B on GB200\n\nFor Qwen3 235B, the practical ordering is usually:\n\n1. `alltoall` for initial bring-up\n2. DeepEP if you want a familiar tuned path\n3. HybridEP for the strongest steady-state result on GB200\n\nHybridEP is usually modestly faster than `alltoall` on this workload and often\nhas noticeably better memory headroom.\n\n### Qwen3-Next on GB200\n\nThis family is a good reminder that dispatcher wins are workload-dependent:\n\n- in BF16, `alltoall` and HybridEP can be close\n- in FP8 or memory-constrained settings, HybridEP tends to look better\n- pipeline layout and grouped-GEMM changes can matter almost as much as the\n  dispatcher itself\n\n## Tuning Parameters\n\n### DeepEP\n\nDeepEP is selected by setting\n`moe_token_dispatcher_type=\"flex\"` and `moe_flex_dispatcher_backend=\"deepep\"`.\n\n```bash\n--moe-deepep-num-sms 20\n```\n\nTune the SM count allocated to DeepEP communication kernels (default 20).\nThe optimal value depends on the workload and EP degree.\nFirst confirm the DeepEP package imports in the target container; a missing\npackage fails during model construction, before any dispatcher timing is\navailable.\n\n### HybridEP\n\nHybridEP is selected by setting\n`moe_token_dispatcher_type=\"flex\"` and `moe_flex_dispatcher_backend=\"hybridep\"`.\n\n```bash\n--moe-hybridep-num-sms 16\n```\n\nTune the SM count allocated to HybridEP communication (default 16). The\nperformance harness uses 32 for HybridEP workloads. Sweep between 16 and 32\nfor the target hardware. Set\n`NUM_OF_HYBRID_EP_RANKS_PER_NVLINK_DOMAIN` to match the NVLink domain size of\nthe deployment. If it does not match the actual topology, performance and\nsometimes correctness will suffer.\nFirst confirm the HybridEP package imports in the target container; a missing\npackage fails during model construction, before any dispatcher timing is\navailable.\n\n### Routing mode\n\n```bash\n--moe-router-force-load-balancing\n```\n\nFor performance benchmarking, force-balance routing is the safer default. It\nusually outperforms dropless routing in large-scale benchmarks and makes results\nmore comparable across dispatcher backends.\n\n## Key Interactions\n\n| Feature | Interaction |\n|---|---|\n| CUDA graphs | Best paired with `attn moe_router moe_preprocess` on dropless MoE |\n| EP overlap | Helps when dispatcher time is still visible after backend tuning |\n| FP8 | Often increases the relative importance of communication and host overhead |\n| CPU affinity | Can matter as much as dispatcher choice on GB200 or GB300 |\n| Pipeline layout | Poor PP or VPP layout can erase dispatcher gains |\n\n## When To Use Each\n\n### `alltoall`\n\n- first correctness bring-up\n- small EP configurations\n- debugging communication regressions\n\n### DeepEP\n\n- Hopper or B200 deployments\n- cross-node EP is clearly visible in profiles\n- you want a mature intermediate step before testing HybridEP\n\n### HybridEP\n\n- GB200 or GB300 NVL72 systems\n- large EP degrees\n- memory headroom matters in addition to throughput\n\n## Pitfalls\n\n1. **Do not compare dispatchers on different stacks**: container, routing mode,\n   PP layout, and CUDA-graph scope can move the result as much as the dispatcher.\n\n2. **HybridEP is topology-sensitive**: it is not a universal win outside the\n   hardware it was designed for.\n\n3. **Both dispatchers need SM tuning**: default `moe_deepep_num_sms` (20) and\n   `moe_hybridep_num_sms` (16) are reasonable starting points but rarely optimal.\n\n4. **Force-balance and dropless are not interchangeable baselines**: keep the\n   routing mode fixed when comparing dispatcher backends.\n\n5. **Memory and throughput can trade off differently by model**: Qwen3-style\n   runs may show a smaller speed delta than DSV3, but still justify HybridEP for\n   memory headroom.\n\n6. **Backend import failures are not performance data**: if DeepEP or HybridEP\n   is missing from the container, do not compare its failed job against a\n   completed `alltoall` job. Fix the environment first, then rerun the same\n   stack.\n",{"data":32,"body":34},{"name":4,"description":6,"license":23,"when_to_use":33},"Choosing a MoE token dispatcher, or tracing a MoE regression or crash to a dispatcher config change; 'which dispatcher', 'alltoall vs DeepEP', 'HybridEP', 'MoE dispatcher', 'flex backend', 'EP dispatcher selection'.",{"type":35,"children":36},"root",[37,46,52,59,66,175,181,249,255,372,378,384,435,441,477,483,488,515,520,526,531,555,567,573,578,603,609,613,632,661,666,671,689,709,722,728,742,747,753,848,854,862,880,885,903,908,926,932,1019],{"type":38,"tag":39,"props":40,"children":42},"element","h1",{"id":41},"moe-dispatcher-selection-guide",[43],{"type":44,"value":45},"text","MoE Dispatcher Selection Guide",{"type":38,"tag":47,"props":48,"children":49},"p",{},[50],{"type":44,"value":51},"Stable docs: @docs\u002Ftraining\u002Fmoe-optimization.md\nCard: @skills\u002Fnemo-mbridge-perf-moe-dispatcher-selection\u002Fcard.yaml",{"type":38,"tag":53,"props":54,"children":56},"h2",{"id":55},"quick-decision",[57],{"type":44,"value":58},"Quick Decision",{"type":38,"tag":60,"props":61,"children":63},"h3",{"id":62},"by-hardware",[64],{"type":44,"value":65},"By hardware",{"type":38,"tag":67,"props":68,"children":69},"table",{},[70,94],{"type":38,"tag":71,"props":72,"children":73},"thead",{},[74],{"type":38,"tag":75,"props":76,"children":77},"tr",{},[78,84,89],{"type":38,"tag":79,"props":80,"children":81},"th",{},[82],{"type":44,"value":83},"Hardware",{"type":38,"tag":79,"props":85,"children":86},{},[87],{"type":44,"value":88},"First choice",{"type":38,"tag":79,"props":90,"children":91},{},[92],{"type":44,"value":93},"Why",{"type":38,"tag":95,"props":96,"children":97},"tbody",{},[98,117,134,152],{"type":38,"tag":75,"props":99,"children":100},{},[101,107,112],{"type":38,"tag":102,"props":103,"children":104},"td",{},[105],{"type":44,"value":106},"H100",{"type":38,"tag":102,"props":108,"children":109},{},[110],{"type":44,"value":111},"DeepEP, if the runtime package is installed",{"type":38,"tag":102,"props":113,"children":114},{},[115],{"type":44,"value":116},"Strong default for cross-node EP on Hopper",{"type":38,"tag":75,"props":118,"children":119},{},[120,125,129],{"type":38,"tag":102,"props":121,"children":122},{},[123],{"type":44,"value":124},"B200",{"type":38,"tag":102,"props":126,"children":127},{},[128],{"type":44,"value":111},{"type":38,"tag":102,"props":130,"children":131},{},[132],{"type":44,"value":133},"Good first choice unless a platform-specific HybridEP path is available",{"type":38,"tag":75,"props":135,"children":136},{},[137,142,147],{"type":38,"tag":102,"props":138,"children":139},{},[140],{"type":44,"value":141},"GB200 \u002F GB300 NVL72",{"type":38,"tag":102,"props":143,"children":144},{},[145],{"type":44,"value":146},"HybridEP, if the runtime package is installed",{"type":38,"tag":102,"props":148,"children":149},{},[150],{"type":44,"value":151},"Best fit for NVLink-domain-aware dispatch and lower memory pressure",{"type":38,"tag":75,"props":153,"children":154},{},[155,160,170],{"type":38,"tag":102,"props":156,"children":157},{},[158],{"type":44,"value":159},"Unknown or first bring-up",{"type":38,"tag":102,"props":161,"children":162},{},[163],{"type":38,"tag":164,"props":165,"children":167},"code",{"className":166},[],[168],{"type":44,"value":169},"alltoall",{"type":38,"tag":102,"props":171,"children":172},{},[173],{"type":44,"value":174},"Easiest path for correctness and debugging",{"type":38,"tag":60,"props":176,"children":178},{"id":177},"by-ep-degree",[179],{"type":44,"value":180},"By EP degree",{"type":38,"tag":67,"props":182,"children":183},{},[184,200],{"type":38,"tag":71,"props":185,"children":186},{},[187],{"type":38,"tag":75,"props":188,"children":189},{},[190,195],{"type":38,"tag":79,"props":191,"children":192},{},[193],{"type":44,"value":194},"EP size",{"type":38,"tag":79,"props":196,"children":197},{},[198],{"type":44,"value":199},"Guidance",{"type":38,"tag":95,"props":201,"children":202},{},[203,223,236],{"type":38,"tag":75,"props":204,"children":205},{},[206,211],{"type":38,"tag":102,"props":207,"children":208},{},[209],{"type":44,"value":210},"Small EP",{"type":38,"tag":102,"props":212,"children":213},{},[214,216,221],{"type":44,"value":215},"Dispatcher choice is usually second-order; start with ",{"type":38,"tag":164,"props":217,"children":219},{"className":218},[],[220],{"type":44,"value":169},{"type":44,"value":222}," or DeepEP",{"type":38,"tag":75,"props":224,"children":225},{},[226,231],{"type":38,"tag":102,"props":227,"children":228},{},[229],{"type":44,"value":230},"Medium EP",{"type":38,"tag":102,"props":232,"children":233},{},[234],{"type":44,"value":235},"DeepEP often becomes worthwhile",{"type":38,"tag":75,"props":237,"children":238},{},[239,244],{"type":38,"tag":102,"props":240,"children":241},{},[242],{"type":44,"value":243},"Large EP",{"type":38,"tag":102,"props":245,"children":246},{},[247],{"type":44,"value":248},"HybridEP is usually the best target on NVL72 systems",{"type":38,"tag":53,"props":250,"children":252},{"id":251},"model-family-patterns",[253],{"type":44,"value":254},"Model-Family Patterns",{"type":38,"tag":67,"props":256,"children":257},{},[258,279],{"type":38,"tag":71,"props":259,"children":260},{},[261],{"type":38,"tag":75,"props":262,"children":263},{},[264,269,274],{"type":38,"tag":79,"props":265,"children":266},{},[267],{"type":44,"value":268},"Workload",{"type":38,"tag":79,"props":270,"children":271},{},[272],{"type":44,"value":273},"Common best path",{"type":38,"tag":79,"props":275,"children":276},{},[277],{"type":44,"value":278},"Notes",{"type":38,"tag":95,"props":280,"children":281},{},[282,300,318,336,354],{"type":38,"tag":75,"props":283,"children":284},{},[285,290,295],{"type":38,"tag":102,"props":286,"children":287},{},[288],{"type":44,"value":289},"DSV3 at large scale",{"type":38,"tag":102,"props":291,"children":292},{},[293],{"type":44,"value":294},"HybridEP on GB200 or GB300, DeepEP on H100",{"type":38,"tag":102,"props":296,"children":297},{},[298],{"type":44,"value":299},"Dispatcher choice matters more as EP and PP both grow",{"type":38,"tag":75,"props":301,"children":302},{},[303,308,313],{"type":38,"tag":102,"props":304,"children":305},{},[306],{"type":44,"value":307},"Qwen3 235B",{"type":38,"tag":102,"props":309,"children":310},{},[311],{"type":44,"value":312},"DeepEP on H100, HybridEP on GB200",{"type":38,"tag":102,"props":314,"children":315},{},[316],{"type":44,"value":317},"HybridEP usually wins on GB200 and often uses less memory",{"type":38,"tag":75,"props":319,"children":320},{},[321,326,331],{"type":38,"tag":102,"props":322,"children":323},{},[324],{"type":44,"value":325},"Qwen3 30B",{"type":38,"tag":102,"props":327,"children":328},{},[329],{"type":44,"value":330},"DeepEP",{"type":38,"tag":102,"props":332,"children":333},{},[334],{"type":44,"value":335},"Smaller models still benefit, but the absolute gap is smaller",{"type":38,"tag":75,"props":337,"children":338},{},[339,344,349],{"type":38,"tag":102,"props":340,"children":341},{},[342],{"type":44,"value":343},"Qwen3-Next",{"type":38,"tag":102,"props":345,"children":346},{},[347],{"type":44,"value":348},"Close race in BF16, HybridEP stronger in FP8 or memory-tight runs",{"type":38,"tag":102,"props":350,"children":351},{},[352],{"type":44,"value":353},"Good reminder to test, not assume",{"type":38,"tag":75,"props":355,"children":356},{},[357,362,367],{"type":38,"tag":102,"props":358,"children":359},{},[360],{"type":44,"value":361},"MoE VLMs",{"type":38,"tag":102,"props":363,"children":364},{},[365],{"type":44,"value":366},"Start simple, then test HybridEP on GB200-class systems",{"type":38,"tag":102,"props":368,"children":369},{},[370],{"type":44,"value":371},"Vision workloads are sensitive to both memory and host overhead",{"type":38,"tag":53,"props":373,"children":375},{"id":374},"rounded-evidence-summary",[376],{"type":44,"value":377},"Rounded Evidence Summary",{"type":38,"tag":60,"props":379,"children":381},{"id":380},"backend-availability-gate",[382],{"type":44,"value":383},"Backend availability gate",{"type":38,"tag":47,"props":385,"children":386},{},[387,389,395,397,402,404,410,412,418,420,426,428,433],{"type":44,"value":388},"Do not interpret a dispatcher timing until the container has proven that the\nselected backend package is available. ",{"type":38,"tag":164,"props":390,"children":392},{"className":391},[],[393],{"type":44,"value":394},"--moe_flex_dispatcher_backend None",{"type":44,"value":396},"\nselects the standard ",{"type":38,"tag":164,"props":398,"children":400},{"className":399},[],[401],{"type":44,"value":169},{"type":44,"value":403}," dispatcher, while ",{"type":38,"tag":164,"props":405,"children":407},{"className":406},[],[408],{"type":44,"value":409},"deepep",{"type":44,"value":411}," and ",{"type":38,"tag":164,"props":413,"children":415},{"className":414},[],[416],{"type":44,"value":417},"hybridep",{"type":44,"value":419},"\nselect ",{"type":38,"tag":164,"props":421,"children":423},{"className":422},[],[424],{"type":44,"value":425},"moe_token_dispatcher_type=\"flex\"",{"type":44,"value":427}," and then require their corresponding\nruntime packages at model construction time. If DeepEP or HybridEP is missing,\nrecord the import failure as an environment limitation and treat ",{"type":38,"tag":164,"props":429,"children":431},{"className":430},[],[432],{"type":44,"value":169},{"type":44,"value":434}," as\nthe only measured correctness fallback for that run.",{"type":38,"tag":60,"props":436,"children":438},{"id":437},"qwen3-30b-a3b-on-h100",[439],{"type":44,"value":440},"Qwen3 30B A3B on H100",{"type":38,"tag":47,"props":442,"children":443},{},[444,446,452,454,460,462,468,470,475],{"type":44,"value":445},"A short 2026-05-17 H100 smoke run used Qwen3 30B A3B BF16, 16 GPUs, EP=16,\nthe recipe's Transformer Engine CUDA graph scopes (",{"type":38,"tag":164,"props":447,"children":449},{"className":448},[],[450],{"type":44,"value":451},"moe_router",{"type":44,"value":453},",\n",{"type":38,"tag":164,"props":455,"children":457},{"className":456},[],[458],{"type":44,"value":459},"moe_preprocess",{"type":44,"value":461},"), and ",{"type":38,"tag":164,"props":463,"children":465},{"className":464},[],[466],{"type":44,"value":467},"model.moe_permute_fusion=false",{"type":44,"value":469}," due to a Triton JIT\ncompatibility issue in the run container. The ",{"type":38,"tag":164,"props":471,"children":473},{"className":472},[],[474],{"type":44,"value":169},{"type":44,"value":476}," fallback completed five\nsteps with 45.65 s mean step time after warmup, 132.9 mean TFLOP\u002Fs\u002FGPU after\nwarmup, final loss 11.44050, and 61.351 GB peak max allocated memory. DeepEP\nand HybridEP selected the requested flex backend in the dumped configs but\nfailed before the first iteration because the packages were not installed. This\nconfirms the availability gate; it is not a throughput ranking for flex\ndispatchers on H100.",{"type":38,"tag":60,"props":478,"children":480},{"id":479},"dsv3-on-gb200-or-gb300",[481],{"type":44,"value":482},"DSV3 on GB200 or GB300",{"type":38,"tag":47,"props":484,"children":485},{},[486],{"type":44,"value":487},"The broad trend is more important than any single row in the tracker:",{"type":38,"tag":489,"props":490,"children":491},"ul",{},[492,505,510],{"type":38,"tag":493,"props":494,"children":495},"li",{},[496,498,503],{"type":44,"value":497},"plain ",{"type":38,"tag":164,"props":499,"children":501},{"className":500},[],[502],{"type":44,"value":169},{"type":44,"value":504}," is usually the conservative baseline",{"type":38,"tag":493,"props":506,"children":507},{},[508],{"type":44,"value":509},"DeepEP improves that baseline once EP communication becomes visible",{"type":38,"tag":493,"props":511,"children":512},{},[513],{"type":44,"value":514},"HybridEP adds another step up on NVL72 systems, especially after CUDA graphs,\nrouting improvements, and CPU-side cleanup are already in place",{"type":38,"tag":47,"props":516,"children":517},{},[518],{"type":44,"value":519},"In practice, the stack often moves from roughly \"low-teens MFU\" territory with\nan untuned baseline into \"high-teens to low-20s MFU\" territory after the full\ndispatcher and kernel stack is tuned.",{"type":38,"tag":60,"props":521,"children":523},{"id":522},"qwen3-235b-on-gb200",[524],{"type":44,"value":525},"Qwen3 235B on GB200",{"type":38,"tag":47,"props":527,"children":528},{},[529],{"type":44,"value":530},"For Qwen3 235B, the practical ordering is usually:",{"type":38,"tag":532,"props":533,"children":534},"ol",{},[535,545,550],{"type":38,"tag":493,"props":536,"children":537},{},[538,543],{"type":38,"tag":164,"props":539,"children":541},{"className":540},[],[542],{"type":44,"value":169},{"type":44,"value":544}," for initial bring-up",{"type":38,"tag":493,"props":546,"children":547},{},[548],{"type":44,"value":549},"DeepEP if you want a familiar tuned path",{"type":38,"tag":493,"props":551,"children":552},{},[553],{"type":44,"value":554},"HybridEP for the strongest steady-state result on GB200",{"type":38,"tag":47,"props":556,"children":557},{},[558,560,565],{"type":44,"value":559},"HybridEP is usually modestly faster than ",{"type":38,"tag":164,"props":561,"children":563},{"className":562},[],[564],{"type":44,"value":169},{"type":44,"value":566}," on this workload and often\nhas noticeably better memory headroom.",{"type":38,"tag":60,"props":568,"children":570},{"id":569},"qwen3-next-on-gb200",[571],{"type":44,"value":572},"Qwen3-Next on GB200",{"type":38,"tag":47,"props":574,"children":575},{},[576],{"type":44,"value":577},"This family is a good reminder that dispatcher wins are workload-dependent:",{"type":38,"tag":489,"props":579,"children":580},{},[581,593,598],{"type":38,"tag":493,"props":582,"children":583},{},[584,586,591],{"type":44,"value":585},"in BF16, ",{"type":38,"tag":164,"props":587,"children":589},{"className":588},[],[590],{"type":44,"value":169},{"type":44,"value":592}," and HybridEP can be close",{"type":38,"tag":493,"props":594,"children":595},{},[596],{"type":44,"value":597},"in FP8 or memory-constrained settings, HybridEP tends to look better",{"type":38,"tag":493,"props":599,"children":600},{},[601],{"type":44,"value":602},"pipeline layout and grouped-GEMM changes can matter almost as much as the\ndispatcher itself",{"type":38,"tag":53,"props":604,"children":606},{"id":605},"tuning-parameters",[607],{"type":44,"value":608},"Tuning Parameters",{"type":38,"tag":60,"props":610,"children":611},{"id":409},[612],{"type":44,"value":330},{"type":38,"tag":47,"props":614,"children":615},{},[616,618,623,624,630],{"type":44,"value":617},"DeepEP is selected by setting\n",{"type":38,"tag":164,"props":619,"children":621},{"className":620},[],[622],{"type":44,"value":425},{"type":44,"value":411},{"type":38,"tag":164,"props":625,"children":627},{"className":626},[],[628],{"type":44,"value":629},"moe_flex_dispatcher_backend=\"deepep\"",{"type":44,"value":631},".",{"type":38,"tag":633,"props":634,"children":639},"pre",{"className":635,"code":636,"language":637,"meta":638,"style":638},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","--moe-deepep-num-sms 20\n","bash","",[640],{"type":38,"tag":164,"props":641,"children":642},{"__ignoreMap":638},[643],{"type":38,"tag":644,"props":645,"children":648},"span",{"class":646,"line":647},"line",1,[649,655],{"type":38,"tag":644,"props":650,"children":652},{"style":651},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[653],{"type":44,"value":654},"--moe-deepep-num-sms",{"type":38,"tag":644,"props":656,"children":658},{"style":657},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[659],{"type":44,"value":660}," 20\n",{"type":38,"tag":47,"props":662,"children":663},{},[664],{"type":44,"value":665},"Tune the SM count allocated to DeepEP communication kernels (default 20).\nThe optimal value depends on the workload and EP degree.\nFirst confirm the DeepEP package imports in the target container; a missing\npackage fails during model construction, before any dispatcher timing is\navailable.",{"type":38,"tag":60,"props":667,"children":668},{"id":417},[669],{"type":44,"value":670},"HybridEP",{"type":38,"tag":47,"props":672,"children":673},{},[674,676,681,682,688],{"type":44,"value":675},"HybridEP is selected by setting\n",{"type":38,"tag":164,"props":677,"children":679},{"className":678},[],[680],{"type":44,"value":425},{"type":44,"value":411},{"type":38,"tag":164,"props":683,"children":685},{"className":684},[],[686],{"type":44,"value":687},"moe_flex_dispatcher_backend=\"hybridep\"",{"type":44,"value":631},{"type":38,"tag":633,"props":690,"children":692},{"className":635,"code":691,"language":637,"meta":638,"style":638},"--moe-hybridep-num-sms 16\n",[693],{"type":38,"tag":164,"props":694,"children":695},{"__ignoreMap":638},[696],{"type":38,"tag":644,"props":697,"children":698},{"class":646,"line":647},[699,704],{"type":38,"tag":644,"props":700,"children":701},{"style":651},[702],{"type":44,"value":703},"--moe-hybridep-num-sms",{"type":38,"tag":644,"props":705,"children":706},{"style":657},[707],{"type":44,"value":708}," 16\n",{"type":38,"tag":47,"props":710,"children":711},{},[712,714,720],{"type":44,"value":713},"Tune the SM count allocated to HybridEP communication (default 16). The\nperformance harness uses 32 for HybridEP workloads. Sweep between 16 and 32\nfor the target hardware. Set\n",{"type":38,"tag":164,"props":715,"children":717},{"className":716},[],[718],{"type":44,"value":719},"NUM_OF_HYBRID_EP_RANKS_PER_NVLINK_DOMAIN",{"type":44,"value":721}," to match the NVLink domain size of\nthe deployment. If it does not match the actual topology, performance and\nsometimes correctness will suffer.\nFirst confirm the HybridEP package imports in the target container; a missing\npackage fails during model construction, before any dispatcher timing is\navailable.",{"type":38,"tag":60,"props":723,"children":725},{"id":724},"routing-mode",[726],{"type":44,"value":727},"Routing mode",{"type":38,"tag":633,"props":729,"children":731},{"className":635,"code":730,"language":637,"meta":638,"style":638},"--moe-router-force-load-balancing\n",[732],{"type":38,"tag":164,"props":733,"children":734},{"__ignoreMap":638},[735],{"type":38,"tag":644,"props":736,"children":737},{"class":646,"line":647},[738],{"type":38,"tag":644,"props":739,"children":740},{"style":651},[741],{"type":44,"value":730},{"type":38,"tag":47,"props":743,"children":744},{},[745],{"type":44,"value":746},"For performance benchmarking, force-balance routing is the safer default. It\nusually outperforms dropless routing in large-scale benchmarks and makes results\nmore comparable across dispatcher backends.",{"type":38,"tag":53,"props":748,"children":750},{"id":749},"key-interactions",[751],{"type":44,"value":752},"Key Interactions",{"type":38,"tag":67,"props":754,"children":755},{},[756,772],{"type":38,"tag":71,"props":757,"children":758},{},[759],{"type":38,"tag":75,"props":760,"children":761},{},[762,767],{"type":38,"tag":79,"props":763,"children":764},{},[765],{"type":44,"value":766},"Feature",{"type":38,"tag":79,"props":768,"children":769},{},[770],{"type":44,"value":771},"Interaction",{"type":38,"tag":95,"props":773,"children":774},{},[775,796,809,822,835],{"type":38,"tag":75,"props":776,"children":777},{},[778,783],{"type":38,"tag":102,"props":779,"children":780},{},[781],{"type":44,"value":782},"CUDA graphs",{"type":38,"tag":102,"props":784,"children":785},{},[786,788,794],{"type":44,"value":787},"Best paired with ",{"type":38,"tag":164,"props":789,"children":791},{"className":790},[],[792],{"type":44,"value":793},"attn moe_router moe_preprocess",{"type":44,"value":795}," on dropless MoE",{"type":38,"tag":75,"props":797,"children":798},{},[799,804],{"type":38,"tag":102,"props":800,"children":801},{},[802],{"type":44,"value":803},"EP overlap",{"type":38,"tag":102,"props":805,"children":806},{},[807],{"type":44,"value":808},"Helps when dispatcher time is still visible after backend tuning",{"type":38,"tag":75,"props":810,"children":811},{},[812,817],{"type":38,"tag":102,"props":813,"children":814},{},[815],{"type":44,"value":816},"FP8",{"type":38,"tag":102,"props":818,"children":819},{},[820],{"type":44,"value":821},"Often increases the relative importance of communication and host overhead",{"type":38,"tag":75,"props":823,"children":824},{},[825,830],{"type":38,"tag":102,"props":826,"children":827},{},[828],{"type":44,"value":829},"CPU affinity",{"type":38,"tag":102,"props":831,"children":832},{},[833],{"type":44,"value":834},"Can matter as much as dispatcher choice on GB200 or GB300",{"type":38,"tag":75,"props":836,"children":837},{},[838,843],{"type":38,"tag":102,"props":839,"children":840},{},[841],{"type":44,"value":842},"Pipeline layout",{"type":38,"tag":102,"props":844,"children":845},{},[846],{"type":44,"value":847},"Poor PP or VPP layout can erase dispatcher gains",{"type":38,"tag":53,"props":849,"children":851},{"id":850},"when-to-use-each",[852],{"type":44,"value":853},"When To Use Each",{"type":38,"tag":60,"props":855,"children":856},{"id":169},[857],{"type":38,"tag":164,"props":858,"children":860},{"className":859},[],[861],{"type":44,"value":169},{"type":38,"tag":489,"props":863,"children":864},{},[865,870,875],{"type":38,"tag":493,"props":866,"children":867},{},[868],{"type":44,"value":869},"first correctness bring-up",{"type":38,"tag":493,"props":871,"children":872},{},[873],{"type":44,"value":874},"small EP configurations",{"type":38,"tag":493,"props":876,"children":877},{},[878],{"type":44,"value":879},"debugging communication regressions",{"type":38,"tag":60,"props":881,"children":883},{"id":882},"deepep-1",[884],{"type":44,"value":330},{"type":38,"tag":489,"props":886,"children":887},{},[888,893,898],{"type":38,"tag":493,"props":889,"children":890},{},[891],{"type":44,"value":892},"Hopper or B200 deployments",{"type":38,"tag":493,"props":894,"children":895},{},[896],{"type":44,"value":897},"cross-node EP is clearly visible in profiles",{"type":38,"tag":493,"props":899,"children":900},{},[901],{"type":44,"value":902},"you want a mature intermediate step before testing HybridEP",{"type":38,"tag":60,"props":904,"children":906},{"id":905},"hybridep-1",[907],{"type":44,"value":670},{"type":38,"tag":489,"props":909,"children":910},{},[911,916,921],{"type":38,"tag":493,"props":912,"children":913},{},[914],{"type":44,"value":915},"GB200 or GB300 NVL72 systems",{"type":38,"tag":493,"props":917,"children":918},{},[919],{"type":44,"value":920},"large EP degrees",{"type":38,"tag":493,"props":922,"children":923},{},[924],{"type":44,"value":925},"memory headroom matters in addition to throughput",{"type":38,"tag":53,"props":927,"children":929},{"id":928},"pitfalls",[930],{"type":44,"value":931},"Pitfalls",{"type":38,"tag":532,"props":933,"children":934},{},[935,946,956,982,992,1002],{"type":38,"tag":493,"props":936,"children":937},{},[938,944],{"type":38,"tag":939,"props":940,"children":941},"strong",{},[942],{"type":44,"value":943},"Do not compare dispatchers on different stacks",{"type":44,"value":945},": container, routing mode,\nPP layout, and CUDA-graph scope can move the result as much as the dispatcher.",{"type":38,"tag":493,"props":947,"children":948},{},[949,954],{"type":38,"tag":939,"props":950,"children":951},{},[952],{"type":44,"value":953},"HybridEP is topology-sensitive",{"type":44,"value":955},": it is not a universal win outside the\nhardware it was designed for.",{"type":38,"tag":493,"props":957,"children":958},{},[959,964,966,972,974,980],{"type":38,"tag":939,"props":960,"children":961},{},[962],{"type":44,"value":963},"Both dispatchers need SM tuning",{"type":44,"value":965},": default ",{"type":38,"tag":164,"props":967,"children":969},{"className":968},[],[970],{"type":44,"value":971},"moe_deepep_num_sms",{"type":44,"value":973}," (20) and\n",{"type":38,"tag":164,"props":975,"children":977},{"className":976},[],[978],{"type":44,"value":979},"moe_hybridep_num_sms",{"type":44,"value":981}," (16) are reasonable starting points but rarely optimal.",{"type":38,"tag":493,"props":983,"children":984},{},[985,990],{"type":38,"tag":939,"props":986,"children":987},{},[988],{"type":44,"value":989},"Force-balance and dropless are not interchangeable baselines",{"type":44,"value":991},": keep the\nrouting mode fixed when comparing dispatcher backends.",{"type":38,"tag":493,"props":993,"children":994},{},[995,1000],{"type":38,"tag":939,"props":996,"children":997},{},[998],{"type":44,"value":999},"Memory and throughput can trade off differently by model",{"type":44,"value":1001},": Qwen3-style\nruns may show a smaller speed delta than DSV3, but still justify HybridEP for\nmemory headroom.",{"type":38,"tag":493,"props":1003,"children":1004},{},[1005,1010,1012,1017],{"type":38,"tag":939,"props":1006,"children":1007},{},[1008],{"type":44,"value":1009},"Backend import failures are not performance data",{"type":44,"value":1011},": if DeepEP or HybridEP\nis missing from the container, do not compare its failed job against a\ncompleted ",{"type":38,"tag":164,"props":1013,"children":1015},{"className":1014},[],[1016],{"type":44,"value":169},{"type":44,"value":1018}," job. Fix the environment first, then rerun the same\nstack.",{"type":38,"tag":1020,"props":1021,"children":1022},"style",{},[1023],{"type":44,"value":1024},"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":1026,"total":1185},[1027,1045,1063,1074,1086,1100,1113,1127,1140,1151,1165,1174],{"slug":1028,"name":1028,"fn":1029,"description":1030,"org":1031,"tags":1032,"stars":1042,"repoUrl":1043,"updatedAt":1044},"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},[1033,1036,1039],{"name":1034,"slug":1035,"type":15},"Documentation","documentation",{"name":1037,"slug":1038,"type":15},"MCP","mcp",{"name":1040,"slug":1041,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1046,"name":1046,"fn":1047,"description":1048,"org":1049,"tags":1050,"stars":1060,"repoUrl":1061,"updatedAt":1062},"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},[1051,1054,1057],{"name":1052,"slug":1053,"type":15},"Containers","containers",{"name":1055,"slug":1056,"type":15},"Deployment","deployment",{"name":1058,"slug":1059,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1064,"name":1064,"fn":1065,"description":1066,"org":1067,"tags":1068,"stars":1060,"repoUrl":1061,"updatedAt":1073},"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},[1069,1072],{"name":1070,"slug":1071,"type":15},"CI\u002FCD","ci-cd",{"name":1055,"slug":1056,"type":15},"2026-07-14T05:25:59.97109",{"slug":1075,"name":1075,"fn":1076,"description":1077,"org":1078,"tags":1079,"stars":1060,"repoUrl":1061,"updatedAt":1085},"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},[1080,1081,1082],{"name":1070,"slug":1071,"type":15},{"name":1055,"slug":1056,"type":15},{"name":1083,"slug":1084,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1087,"name":1087,"fn":1088,"description":1089,"org":1090,"tags":1091,"stars":1060,"repoUrl":1061,"updatedAt":1099},"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},[1092,1095,1096],{"name":1093,"slug":1094,"type":15},"Debugging","debugging",{"name":1083,"slug":1084,"type":15},{"name":1097,"slug":1098,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1101,"name":1101,"fn":1102,"description":1103,"org":1104,"tags":1105,"stars":1060,"repoUrl":1061,"updatedAt":1112},"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},[1106,1109],{"name":1107,"slug":1108,"type":15},"Best Practices","best-practices",{"name":1110,"slug":1111,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1114,"name":1114,"fn":1115,"description":1116,"org":1117,"tags":1118,"stars":1060,"repoUrl":1061,"updatedAt":1126},"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},[1119,1122,1125],{"name":1120,"slug":1121,"type":15},"Machine Learning","machine-learning",{"name":1123,"slug":1124,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1128,"name":1128,"fn":1129,"description":1130,"org":1131,"tags":1132,"stars":1060,"repoUrl":1061,"updatedAt":1139},"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},[1133,1136],{"name":1134,"slug":1135,"type":15},"QA","qa",{"name":1137,"slug":1138,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1141,"name":1141,"fn":1142,"description":1143,"org":1144,"tags":1145,"stars":1060,"repoUrl":1061,"updatedAt":1150},"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},[1146,1147],{"name":1055,"slug":1056,"type":15},{"name":1148,"slug":1149,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1152,"name":1152,"fn":1153,"description":1154,"org":1155,"tags":1156,"stars":1060,"repoUrl":1061,"updatedAt":1164},"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},[1157,1160,1161],{"name":1158,"slug":1159,"type":15},"Code Review","code-review",{"name":1083,"slug":1084,"type":15},{"name":1162,"slug":1163,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1166,"name":1166,"fn":1167,"description":1168,"org":1169,"tags":1170,"stars":1060,"repoUrl":1061,"updatedAt":1173},"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},[1171,1172],{"name":1134,"slug":1135,"type":15},{"name":1137,"slug":1138,"type":15},"2026-07-14T05:25:54.928983",{"slug":1175,"name":1175,"fn":1176,"description":1177,"org":1178,"tags":1179,"stars":1060,"repoUrl":1061,"updatedAt":1184},"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},[1180,1183],{"name":1181,"slug":1182,"type":15},"Automation","automation",{"name":1070,"slug":1071,"type":15},"2026-07-30T05:29:03.275638",496,{"items":1187,"total":1281},[1188,1203,1213,1227,1237,1252,1267],{"slug":1189,"name":1189,"fn":1190,"description":1191,"org":1192,"tags":1193,"stars":20,"repoUrl":21,"updatedAt":1202},"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},[1194,1197,1200,1201],{"name":1195,"slug":1196,"type":15},"Data Analysis","data-analysis",{"name":1198,"slug":1199,"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":1204,"name":1204,"fn":1205,"description":1206,"org":1207,"tags":1208,"stars":20,"repoUrl":21,"updatedAt":1212},"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},[1209,1210,1211],{"name":1055,"slug":1056,"type":15},{"name":1148,"slug":1149,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1214,"name":1214,"fn":1215,"description":1216,"org":1217,"tags":1218,"stars":20,"repoUrl":21,"updatedAt":1226},"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},[1219,1222,1223],{"name":1220,"slug":1221,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1224,"slug":1225,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1228,"name":1228,"fn":1229,"description":1230,"org":1231,"tags":1232,"stars":20,"repoUrl":21,"updatedAt":1236},"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},[1233,1234,1235],{"name":1195,"slug":1196,"type":15},{"name":9,"slug":8,"type":15},{"name":1137,"slug":1138,"type":15},"2026-07-17T05:29:03.913266",{"slug":1238,"name":1238,"fn":1239,"description":1240,"org":1241,"tags":1242,"stars":20,"repoUrl":21,"updatedAt":1251},"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},[1243,1244,1247,1248],{"name":1181,"slug":1182,"type":15},{"name":1245,"slug":1246,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":1249,"slug":1250,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":1253,"name":1253,"fn":1254,"description":1255,"org":1256,"tags":1257,"stars":20,"repoUrl":21,"updatedAt":1266},"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},[1258,1259,1262,1263],{"name":1055,"slug":1056,"type":15},{"name":1260,"slug":1261,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1264,"slug":1265,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1268,"name":1268,"fn":1269,"description":1270,"org":1271,"tags":1272,"stars":20,"repoUrl":21,"updatedAt":1280},"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},[1273,1274,1277],{"name":9,"slug":8,"type":15},{"name":1275,"slug":1276,"type":15},"Quantum Computing","quantum-computing",{"name":1278,"slug":1279,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]