[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7CPXFD425VWSD544PP904MJ":3,"skill-same:01M7CPXFD425VWSD544PP904MJ":47},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":31,"updatedAt":32,"install":33,"resources":46},"01M7CPXFD425VWSD544PP904MJ","nemo-mbridge-perf-moe-hardware-configs",null,"Representative MoE training playbooks by hardware platform and model family. Summarizes rounded throughput bands, parallelism patterns, and common tuning stacks.","om-scogo\u002Fskillsh-scraper","om-scogo","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F160466840?v=4","01M79AWB6DFVCQ7RK70CS34B96","Om Achrekar","01M79AWB6ZM9SKR8KDH1WSSM22",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fnvidia-skills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002Fskills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: nemo-mbridge-perf-moe-hardware-configs\ndescription: Representative MoE training playbooks by hardware platform and model family. Summarizes rounded throughput bands, parallelism patterns, and common tuning stacks.\nlicense: Apache-2.0\nwhen_to_use: Hardware-specific MoE playbooks or throughput estimates; 'MoE on H100', 'GB200 config', 'expected throughput', 'MoE hardware playbook', 'parallelism for B200'.\n---\n\n# MoE Hardware Configuration Reference\n\nStable docs: @docs\u002Ftraining\u002Fmoe-optimization.md\nCard: @skills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002Fcard.yaml\n\n## Quick Platform Playbook\n\n| Platform | Typical MoE strategy | What usually matters most |\n|---|---|---|\n| H100 | DeepEP + stronger PP + moderate TP | communication overlap and PP efficiency |\n| B200 | DeepEP + MXFP8 + careful PP layout | container quality and tuned comm settings |\n| GB200 | HybridEP + partial CUDA graphs + CPU cleanup | host overhead, topology-aware dispatch, memory headroom |\n| GB300 | HybridEP + newer FP8 and kernel stack | same GB200 playbook, usually with a higher ceiling |\n\n## First Answer Checklist\n\nFor hardware playbook questions, answer from these canonical rows before adding\nthroughput caveats:\n\n| Workload | Hardware | Dispatcher | Layout |\n|---|---|---|---|\n| DSV3 | H100 | DeepEP | TP=2, EP=64, PP=8, VPP=4 |\n| DSV3 | GB200\u002FGB300 | HybridEP | TP=1, EP=64, PP=4, VPP=4 |\n| Qwen3 235B | H100 | DeepEP | TP=2, EP=32, PP=8, VPP=4 |\n| Qwen3 235B | GB200 | HybridEP | TP=1 or 2, EP=32-64, PP=4, VPP=unspecified |\n\nFor Qwen3 235B on GB200, explicitly say `VPP=unspecified`; do not invent or\nextrapolate `VPP=12` unless a measured row provides it. Include TE-scoped CUDA\ngraph scopes (`attn`, `moe_router`, `moe_preprocess`),\n`CUDA_DEVICE_MAX_CONNECTIONS` selection,\n`PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True`, `NCCL_GRAPH_REGISTER=0`,\nGB200\u002FGB300 CPU-side tuning, and the warning not to cargo-cult tracker rows.\n\n## Rounded Performance Bands\n\nThese are intentionally rounded so the document stays durable as the tracker\nmoves. Treat them as planning ranges, not exact promises.\n\n| Workload family | Hardware | Typical band | Representative shape |\n|---|---|---|---|\n| DSV3, large-scale | H100 | low-to-mid hundreds TFLOPS\u002FGPU, high-teens MFU | TP2, EP64, PP8, DeepEP |\n| DSV3, large-scale | B200 | high-hundreds TFLOPS\u002FGPU, mid-teens MFU | TP1, EP32, PP8, DeepEP |\n| DSV3, large-scale | GB200 | around 1K TFLOPS\u002FGPU, low-20s MFU | TP1, EP64, PP4, HybridEP |\n| DSV3, large-scale | GB300 | above the GB200 band, often mid-20s MFU | TP1, EP64, PP4, HybridEP |\n| Qwen3 235B | H100 | low-300s TFLOPS\u002FGPU, around 30% MFU | TP2, EP32, PP8, DeepEP |\n| Qwen3 235B | GB200 | high-hundreds TFLOPS\u002FGPU in tuned runs | TP1 or TP2, EP32-64, PP4, HybridEP |\n| Qwen3 30B | H100 | low-200s TFLOPS\u002FGPU | TP1, EP8, PP1, DeepEP |\n| Qwen3-Next 80B | GB200 | low-300s TFLOPS\u002FGPU in BF16-class runs | TP1, EP32, PP2, HybridEP |\n\n## Representative Config Families\n\n### DSV3 on H100\n\n```text\nDispatcher: DeepEP\nTP=2  EP=64  PP=8  VPP=4\nRouting: force balance\nRecompute: light-to-moderate selective recompute\nPriority: overlap communication and keep PP efficient\n```\n\n### DSV3 on B200\n\n```text\nDispatcher: DeepEP\nTP=1  EP=32  PP=8  VPP=2 or similar\nPrecision: MXFP8-class\nRecompute: selective recompute around MLA up-projection and MLP-side modules\nPriority: container quality, PP layout, and DeepEP SMS tuning\n```\n\n### DSV3 on GB200 or GB300\n\n```text\nDispatcher: HybridEP\nTP=1  EP=64  PP=4  VPP=4\nPrecision: MXFP8-class\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: HybridEP, CPU optimization, and graph-friendly static shapes\n```\n\n### Qwen3 235B on H100\n\n```text\nDispatcher: DeepEP\nTP=2  EP=32  PP=8  VPP=4\nRecompute: norm and activation-side selective recompute\nPriority: communication overlap and router-path cleanup\n```\n\n### Qwen3 235B on GB200\n\n```text\nDispatcher: HybridEP\nTP=1 or 2  EP=32 to 64  PP=4  VPP=unspecified unless measured\nCUDA Graph: attn + moe_router + moe_preprocess\nRecompute: moe_act, mlp, or norm depending on memory pressure\nPriority: balance throughput against memory headroom\n```\n\n### Qwen3-Next 80B on GB200\n\n```text\nDispatcher: HybridEP\nTP=1  EP=32  PP=2  VPP around 4\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: pipeline layout and grouped GEMM quality\n```\n\n## Cross-Cutting Patterns\n\n### PP layout\n\n- `E` = embedding\n- `t` = transformer\n- `m` = MTP\n- `L` = loss\n- `|` = stage boundary\n\nThe biggest platform difference is usually not just the dispatcher. It is the\ncombination of dispatcher, PP shape, and whether VPP keeps each stage balanced.\n\n### Recompute strategy\n\n| Memory pressure | Starting point |\n|---|---|\n| low | none or a very narrow selective set |\n| moderate | `moe_act`, `mlp`, `norm`, or similar selective modules |\n| high | model-specific up-projection plus selective MoE and MLP modules |\n| extreme or long-context | full recompute only if the selective path still does not fit |\n\n### Environment variables\n\n```bash\nCUDA_DEVICE_MAX_CONNECTIONS=1\nCUDA_DEVICE_MAX_CONNECTIONS=32   # common when EP overlap and CUDA graphs are combined\nPYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\nNCCL_GRAPH_REGISTER=0\n```\n\n### CPU-side tuning\n\nOn GB200 and GB300, CPU affinity and general host-overhead cleanup can move the\nneedle almost as much as a dispatcher swap. Treat them as first-class tuning\nwork, not as afterthoughts.\n\n## Pitfalls\n\n1. **Do not cargo-cult a tracker row**: the winning config usually depends on\n   routing mode, container, and PP layout as much as on hardware name.\n\n2. **Container quality matters**: large regressions can come from the software\n   stack rather than the model recipe.\n\n3. **VPP must be intentional**: a bad VPP split can erase the gain from a better\n   dispatcher.\n\n4. **Compare absolute throughput, not only MFU**: MFU can mislead when switching\n   between BF16, FP8, and other precision modes.\n\n5. **Force-balance routing is the safer benchmark default**: keep routing mode\n   fixed when comparing hardware or dispatcher stacks.\n","# MoE Hardware Configuration Reference\n\nStable docs: @docs\u002Ftraining\u002Fmoe-optimization.md\nCard: @skills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002Fcard.yaml\n\n## Quick Platform Playbook\n\n| Platform | Typical MoE strategy | What usually matters most |\n|---|---|---|\n| H100 | DeepEP + stronger PP + moderate TP | communication overlap and PP efficiency |\n| B200 | DeepEP + MXFP8 + careful PP layout | container quality and tuned comm settings |\n| GB200 | HybridEP + partial CUDA graphs + CPU cleanup | host overhead, topology-aware dispatch, memory headroom |\n| GB300 | HybridEP + newer FP8 and kernel stack | same GB200 playbook, usually with a higher ceiling |\n\n## First Answer Checklist\n\nFor hardware playbook questions, answer from these canonical rows before adding\nthroughput caveats:\n\n| Workload | Hardware | Dispatcher | Layout |\n|---|---|---|---|\n| DSV3 | H100 | DeepEP | TP=2, EP=64, PP=8, VPP=4 |\n| DSV3 | GB200\u002FGB300 | HybridEP | TP=1, EP=64, PP=4, VPP=4 |\n| Qwen3 235B | H100 | DeepEP | TP=2, EP=32, PP=8, VPP=4 |\n| Qwen3 235B | GB200 | HybridEP | TP=1 or 2, EP=32-64, PP=4, VPP=unspecified |\n\nFor Qwen3 235B on GB200, explicitly say `VPP=unspecified`; do not invent or\nextrapolate `VPP=12` unless a measured row provides it. Include TE-scoped CUDA\ngraph scopes (`attn`, `moe_router`, `moe_preprocess`),\n`CUDA_DEVICE_MAX_CONNECTIONS` selection,\n`PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True`, `NCCL_GRAPH_REGISTER=0`,\nGB200\u002FGB300 CPU-side tuning, and the warning not to cargo-cult tracker rows.\n\n## Rounded Performance Bands\n\nThese are intentionally rounded so the document stays durable as the tracker\nmoves. Treat them as planning ranges, not exact promises.\n\n| Workload family | Hardware | Typical band | Representative shape |\n|---|---|---|---|\n| DSV3, large-scale | H100 | low-to-mid hundreds TFLOPS\u002FGPU, high-teens MFU | TP2, EP64, PP8, DeepEP |\n| DSV3, large-scale | B200 | high-hundreds TFLOPS\u002FGPU, mid-teens MFU | TP1, EP32, PP8, DeepEP |\n| DSV3, large-scale | GB200 | around 1K TFLOPS\u002FGPU, low-20s MFU | TP1, EP64, PP4, HybridEP |\n| DSV3, large-scale | GB300 | above the GB200 band, often mid-20s MFU | TP1, EP64, PP4, HybridEP |\n| Qwen3 235B | H100 | low-300s TFLOPS\u002FGPU, around 30% MFU | TP2, EP32, PP8, DeepEP |\n| Qwen3 235B | GB200 | high-hundreds TFLOPS\u002FGPU in tuned runs | TP1 or TP2, EP32-64, PP4, HybridEP |\n| Qwen3 30B | H100 | low-200s TFLOPS\u002FGPU | TP1, EP8, PP1, DeepEP |\n| Qwen3-Next 80B | GB200 | low-300s TFLOPS\u002FGPU in BF16-class runs | TP1, EP32, PP2, HybridEP |\n\n## Representative Config Families\n\n### DSV3 on H100\n\n```text\nDispatcher: DeepEP\nTP=2  EP=64  PP=8  VPP=4\nRouting: force balance\nRecompute: light-to-moderate selective recompute\nPriority: overlap communication and keep PP efficient\n```\n\n### DSV3 on B200\n\n```text\nDispatcher: DeepEP\nTP=1  EP=32  PP=8  VPP=2 or similar\nPrecision: MXFP8-class\nRecompute: selective recompute around MLA up-projection and MLP-side modules\nPriority: container quality, PP layout, and DeepEP SMS tuning\n```\n\n### DSV3 on GB200 or GB300\n\n```text\nDispatcher: HybridEP\nTP=1  EP=64  PP=4  VPP=4\nPrecision: MXFP8-class\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: HybridEP, CPU optimization, and graph-friendly static shapes\n```\n\n### Qwen3 235B on H100\n\n```text\nDispatcher: DeepEP\nTP=2  EP=32  PP=8  VPP=4\nRecompute: norm and activation-side selective recompute\nPriority: communication overlap and router-path cleanup\n```\n\n### Qwen3 235B on GB200\n\n```text\nDispatcher: HybridEP\nTP=1 or 2  EP=32 to 64  PP=4  VPP=unspecified unless measured\nCUDA Graph: attn + moe_router + moe_preprocess\nRecompute: moe_act, mlp, or norm depending on memory pressure\nPriority: balance throughput against memory headroom\n```\n\n### Qwen3-Next 80B on GB200\n\n```text\nDispatcher: HybridEP\nTP=1  EP=32  PP=2  VPP around 4\nCUDA Graph: attn + moe_router + moe_preprocess\nPriority: pipeline layout and grouped GEMM quality\n```\n\n## Cross-Cutting Patterns\n\n### PP layout\n\n- `E` = embedding\n- `t` = transformer\n- `m` = MTP\n- `L` = loss\n- `|` = stage boundary\n\nThe biggest platform difference is usually not just the dispatcher. It is the\ncombination of dispatcher, PP shape, and whether VPP keeps each stage balanced.\n\n### Recompute strategy\n\n| Memory pressure | Starting point |\n|---|---|\n| low | none or a very narrow selective set |\n| moderate | `moe_act`, `mlp`, `norm`, or similar selective modules |\n| high | model-specific up-projection plus selective MoE and MLP modules |\n| extreme or long-context | full recompute only if the selective path still does not fit |\n\n### Environment variables\n\n```bash\nCUDA_DEVICE_MAX_CONNECTIONS=1\nCUDA_DEVICE_MAX_CONNECTIONS=32   # common when EP overlap and CUDA graphs are combined\nPYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\nNCCL_GRAPH_REGISTER=0\n```\n\n### CPU-side tuning\n\nOn GB200 and GB300, CPU affinity and general host-overhead cleanup can move the\nneedle almost as much as a dispatcher swap. Treat them as first-class tuning\nwork, not as afterthoughts.\n\n## Pitfalls\n\n1. **Do not cargo-cult a tracker row**: the winning config usually depends on\n   routing mode, container, and PP layout as much as on hardware name.\n\n2. **Container quality matters**: large regressions can come from the software\n   stack rather than the model recipe.\n\n3. **VPP must be intentional**: a bad VPP split can erase the gain from a better\n   dispatcher.\n\n4. **Compare absolute throughput, not only MFU**: MFU can mislead when switching\n   between BF16, FP8, and other precision modes.\n\n5. **Force-balance routing is the safer benchmark default**: keep routing mode\n   fixed when comparing hardware or dispatcher stacks.\n",{"name":5,"license":15,"description":7,"when_to_use":24},"Hardware-specific MoE playbooks or throughput estimates; 'MoE on H100', 'GB200 config', 'expected throughput', 'MoE hardware playbook', 'parallelism for B200'.","main","8859218955d9504cd18b59e7892100718039d3b4","data\u002Fskills\u002Fnvidia-skills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002Fskills\u002Fnemo-mbridge-perf-moe-hardware-configs",1,"https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper","https:\u002F\u002Fgithub.com\u002Fom-scogo","SDE-1 @ Scogo ","2026-07-18 12:22:46",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add om-scogo\u002Fskillsh-scraper --skill nemo-mbridge-perf-moe-hardware-configs",{"tool":39,"command":40},"bunx","bunx skills add om-scogo\u002Fskillsh-scraper --skill nemo-mbridge-perf-moe-hardware-configs",{"tool":42,"command":43},"pnpm","pnpm dlx skills add om-scogo\u002Fskillsh-scraper --skill nemo-mbridge-perf-moe-hardware-configs","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fnvidia-skills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002Fskills\u002Fnemo-mbridge-perf-moe-hardware-configs\u002FSKILL.md",true,[],[48,56,61,68,73,80],{"skillNo":49,"name":50,"nameZh":6,"description":51,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":52,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M7D6MC9FAAEV4CMNYX8W7JTG","arxiv-search","Search arXiv physics, math, and computer science preprints using natural language queries. Powered by Valyu semantic search.","MIT","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fyorkeccak-scientific-skills\u002Farxiv-search\u002Fskills\u002Farxiv-search\u002FSKILL.md",[],[],{"skillNo":57,"name":50,"nameZh":6,"description":51,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":52,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":19},"01M7D6M85CKKN87TA98VCQV4H7","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fyorkeccak-scientific-skills\u002Farxiv-search\u002FSkill.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":52,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7D4ZTCS4DR3CE4JY06X84T2","working-with-legacy-code","Safely change and test untested codebases using Feathers' \"Working Effectively with Legacy Code\". Use when the user mentions \"legacy code\", \"no tests\", \"untested codebase\", \"how do I test this\", \"seams\", \"characterization tests\", \"golden master\", \"sprout method\", \"afraid to change this code\", \"monster method\", \"dependency breaking\", or \"inherited a messy codebase\". Also trigger when changing code without tests safely, getting a class under test when constructors, statics, or singletons block it, adding features to tangled modules, or planning incremental test coverage for an old codebase. Covers the legacy-code change algorithm, seams, characterization tests, sprout\u002Fwrap, and dependency-breaking techniques. For refactoring code that already has tests, see refactoring-patterns. For day-to-day code quality, see clean-code.","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fworking-with-legacy-code\u002Fworking-with-legacy-code\u002FSKILL.md",[],[],{"skillNo":69,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":52,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M7D4ZHZKJD2RSNGAF63V14P4","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fworking-with-legacy-code\u002FSkill.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":52,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M7D4ZB4CJBFW5JE67CPH2QRV","web-typography","Select, pair, and implement typefaces for web projects. Use when the user mentions \"font pairing\", \"which typeface\", \"line height\", \"responsive typography\", \"web font loading\", \"type hierarchy\", \"variable fonts\", \"FOUT\u002FFOIT\", \"typographic scale\", or \"the text is hard to read\". Also trigger when choosing between system fonts and web fonts, optimizing font-loading performance, or designing readable long-form content. Covers readability evaluation, CSS implementation, and performance optimization. For overall UI design systems, see refactoring-ui. For dramatic typographic experiences, see top-design.","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fweb-typography\u002Fweb-typography\u002FSKILL.md",[],[],{"skillNo":81,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":52,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M7D4Z4MVYVY3NMJ18BNV6WNA","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fweb-typography\u002FSkill.md",[],[]]