[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7B5WXMDYXCK43YA5ZEF7ECY":3,"skill-same:01M7B5WXMDYXCK43YA5ZEF7ECY":57},{"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":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":6,"updatedAt":32,"install":33,"resources":46},"01M7B5WXMDYXCK43YA5ZEF7ECY","modelopt-quant-recipe-search",null,"Use when the user asks to find, search for, or optimize the best quantization recipe for a model, including direct requests like \"find the best quantization recipe and generate a PTQ checkpoint.\" Guides the multi-candidate loop: choose compute-vs-memory success metrics, select ModelOpt recipe baselines, design AutoQuant\u002Fmanual recipe deltas, interpret sensitivity, and decide next candidates. Do NOT use for a single known PTQ recipe run (use modelopt-ptq), serving (use modelopt-deployment), creating\u002Frunning evals (use modelopt-evaluation or modelopt-launching-evals), monitoring jobs (use modelopt-monitor), MLflow browsing (use modelopt-accessing-mlflow), or comparing completed baseline-vs-candidate scores only (use modelopt-compare-results).","NVIDIA\u002FModel-Optimizer","NVIDIA","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F1728152?v=4","01M6BXXZYTRNBXSTRXNNBB356H","NVIDIA Corporation","01M6BXY02X3PRVHNNATDHA2D0W",5604,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-quant-recipe-search\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: modelopt-quant-recipe-search\ndescription: >-\n  Use when the user asks to find, search for, or optimize the best quantization\n  recipe for a model, including direct requests like \"find the best quantization\n  recipe and generate a PTQ checkpoint.\" Guides the multi-candidate loop:\n  choose compute-vs-memory success metrics, select ModelOpt recipe baselines,\n  design AutoQuant\u002Fmanual recipe deltas, interpret sensitivity, and decide next\n  candidates. Do NOT use for a single known PTQ recipe run (use modelopt-ptq), serving\n  (use modelopt-deployment), creating\u002Frunning evals (use modelopt-evaluation or modelopt-launching-evals),\n  monitoring jobs (use modelopt-monitor), MLflow browsing (use modelopt-accessing-mlflow), or\n  comparing completed baseline-vs-candidate scores only (use modelopt-compare-results).\n---\n\n# Quant Recipe Search\n\nUse this skill when quantization is an iterative recipe search, not a one-off\nPTQ run. The skill owns strategy: define success, choose the search space,\nsequence candidates, and decide the next iteration. It delegates checkpoint\ngeneration, serving, evaluation, monitoring, and metric comparison to the\nexisting execution skills.\n\nTreat a direct request such as \"find the best quantization recipe and generate a\nPTQ checkpoint for this model\" as enough to start. Recover local state first,\nthen ask only for missing decisions that change the search.\n\n## Skill Boundaries\n\n- Use `modelopt-ptq` to produce and validate checkpoints.\n- Use `modelopt-deployment` to serve checkpoints and debug serving-specific flags.\n- Use `modelopt-evaluation` to create NEL configs and submit evals.\n- Use `modelopt-launching-evals` to run, resume, debug, and analyze NEL runs.\n- Use `modelopt-monitor` for active job tracking.\n- Use `modelopt-accessing-mlflow` for MLflow artifact lookup.\n- Use `modelopt-compare-results` for validated baseline-vs-candidate deltas and score-field comparability.\n\nDo not duplicate those workflows here. This skill should leave the user with a clear recipe portfolio, success metric, experiment sequence, and next decision.\n\n## Problem\n\nThe task is to find the best recipe for a user-defined target, not merely to\nproduce a quantized checkpoint. A generated PTQ checkpoint is only a candidate.\nIt becomes a recommended recipe only after evaluation and comparison against the\nmatching baseline.\n\nRequired inputs before planning candidates:\n\n- **Optimization goal:** compute\u002Fthroughput, memory\u002Flatency, or a custom metric.\n- **Primary quantization family:** for example NVFP4, W4A16 NVFP4, FP8\u002FW8A8,\n  INT4\u002FAWQ, or a custom mixed set.\n- **Benchmark set or baseline results:** the user-defined acceptance surface.\n\nIf any of these are missing, ask for them. Do not silently default to FP8\u002FW8A8\nor call a checkpoint \"best\" before evaluation.\n\nDefault success rule: maximize the chosen performance objective while keeping\neach benchmark within 1 percentage point of the matching BF16\u002FFP16 baseline.\nNear-threshold or noisy regressions require reruns before making a decision.\n\n## Search Space\n\nKeep the search space explicit. A candidate recipe is a tuple across these axes:\n\n- **Numeric format:** FP8\u002FW8A8, NVFP4\u002FW4A4, W4A16 NVFP4, INT4\u002FAWQ, or mixed\n  formats such as NVFP4+FP8.\n- **Calibration\u002Fsearch algorithm:** max calibration, MSE calibration, GPTQ,\n  AWQ, AutoQuant scoring, and calibration dataset or sample-count variants.\n- **Selection method:** manual\u002Fheuristic rules, sensitivity-guided manual\n  recipes, AutoQuant selection, or a hybrid of AutoQuant plus manual overrides.\n- **Module family:** attention, MLP, MoE experts, routers\u002Fgates, embeddings,\n  `lm_head`, adapters, vision encoders, and model-specific modules.\n- **Layer position:** first\u002Flast transformer-layer counts or explicit ordinal\n  ranges to keep in BF16. First 3-4 and last 1-2 layers are common starting\n  candidate ranges, not defaults.\n- **Runtime fusion constraints:** modules fused by the inference library must\n  use compatible quantization. Examples: vLLM Qwen `linear_attn.in_proj_qkvz`\n  and fused MoE expert projections such as gate\u002Fup (`w1`\u002F`w3`).\n- **Calibration budget:** dataset mix, sample count, sequence length, and batch\n  settings.\n\nDo not collapse the search to one dimension such as numeric format only. Read\n`references\u002Frecipe_iteration.md` when choosing concrete axes or candidates.\n\n## Design Workflow\n\n1. **Recover state**\n   - Read result tables, recipe logs, AutoQuant states, sensitivity reports, and\n     experiment notes before proposing new work.\n   - Ask `modelopt-monitor`, `modelopt-launching-evals`, or `modelopt-compare-results` to recover active\n     job state and completed metrics when needed.\n\n2. **Define the target**\n   - Confirm the optimization goal, primary quantization family, benchmark set,\n     accuracy-loss threshold, calibration budget, and cost metric.\n   - Include quantization metadata such as scale storage in active-cost or size\n     estimates.\n\n3. **Pick baselines and first candidates**\n   - Always include BF16\u002FFP16 and a near-lossless FP8\u002FW8A8 baseline unless FP8\n     itself is the target.\n   - For ModelOpt work, start from `modelopt_recipes`: model-specific recipes\n     first, then general PTQ presets or recipe fragments.\n   - Add an AutoQuant candidate in the requested primary family when AutoQuant\n     is available. Expect AutoQuant to find a better trade-off than a first\n     manual recipe, but validate that assumption with the same evals.\n   - Add at least one manual or sensitivity-guided candidate so AutoQuant can be\n     compared against controlled ablations and there is a fallback if AutoQuant\n     misses the best frontier or hits runtime constraints.\n   - When sensitivity or model behavior implicates boundary layers, add a\n     controlled first-layer, last-layer, or combined BF16 exclusion candidate.\n     Do not preserve boundary layers without testing the trade-off.\n\n4. **Generate candidates**\n   - Delegate checkpoint generation and PTQ validation to `modelopt-ptq`.\n   - Change one major axis at a time: format, calibration algorithm, module\n     family, layer position, granularity, or calibration data.\n   - Use AutoQuant for broad candidate generation and sensitivity reports; use\n     manual recipes for controlled module-family ablations and overrides.\n   - Resolve positional ordinals against the model's transformer block sequence.\n     For manual recipes, add those blocks to the recipe exclusions. For\n     AutoQuant or hybrid candidates, pass positional exclusions into the\n     selected implementation when supported; otherwise apply a manual override\n     to its result and record the limitation.\n\n5. **Gate before scaling**\n   - Validate checkpoint coverage and metadata.\n   - Reject or rewrite recipes that mix quantization algorithms inside a fused\n     runtime group.\n   - Ensure positional exclusions preserve complete fused runtime groups, then\n     evaluate the candidate against the same BF16 baseline and acceptance\n     criteria as every other recipe.\n   - If the checkpoint is valid but serving fails due to runtime support, do not\n     reject the recipe immediately. Delegate to `modelopt-deployment` \u002F `modelopt-debug` for small\n     patches or flags, then rerun a pipe-clean check.\n\n## Iteration Loop\n\n1. Run cheap screen evals for every candidate that passes the gates.\n2. Compare accuracy, verbosity\u002Ftoken usage, and active cost against baselines.\n3. Rerun noisy or near-threshold results before labeling a regression.\n4. Decide the next candidate:\n   - Accuracy drop: protect or ablate sensitive module families, try MSE\u002FGPTQ,\n     use AutoQuant sensitivity to choose overrides, or test first\u002Flast-layer\n     BF16 exclusions when evidence points to boundary sensitivity.\n   - Poor performance\u002Fcost: quantize the next high-cost active family, adjust\n     active-cost objective, or try a more aggressive format.\n   - AutoQuant underperforms manual recipes: inspect sensitivity reports,\n     achieved bits, excluded modules, and runtime-fusion constraints; keep the\n     manual recipe in the portfolio instead of forcing the AutoQuant result.\n   - Runtime incompatibility: rewrite around fused groups or isolate deployment\n     support from checkpoint quality.\n   - Repeated AutoQuant recipes: inspect achieved bits and recipe hashes, then\n     adjust constraints before launching a larger sweep.\n5. Promote only when `modelopt-compare-results` shows no failed external sanity check,\n   the candidate is comparable to the validated measured baseline, and the\n   user-defined goal is met. An externally unverified baseline is non-blocking.\n\nWhen the user has authorized QAD, a candidate is a recipe plus the QAD settings\nfrom the `modelopt-qad` skill's recipe loop, which ends each recipe with a pass or a\nclear miss. A PTQ recipe that already meets the threshold counts as a pass.\nStep the layout from that outcome:\n\n- After a pass, add one module group to the passing recipe, in order of\n  savings on the success metric; the MTP head is its own group. A group can be\n  a slice: its middle layers with the first and last few kept, or some of its\n  projections (such as GDN `in_proj` only, or attention `linear_qkv` only).\n  Find and evaluate the new layout's best PTQ recipe, starting from the\n  passing recipe's calibration, before its QAD. Step one layout at a time;\n  the `modelopt-qad` recipe loop runs its knob variants in parallel.\n- After a clear miss, shrink the added group to a slice; if that misses too,\n  drop the group and try the next one on the last passing recipe. If no recipe\n  has passed yet, report no QAD pass and return to the accuracy-drop steps\n  above.\n- Stop when every group has been added or dropped, or the QAD budget is spent.\n  Return the passing recipe that best meets the success metric, with its QAD\n  checkpoint.\n\nMaintain a recipe portfolio table with recipe name, objective, active-cost\nestimate, calibration notes, checkpoint path, eval\u002Flog references, accuracy,\nverbosity, positional exclusions, and decision.\n\n## References\n\n- For recipe design, search-space details, sensitivity, and active-cost\n  accounting, read `references\u002Frecipe_iteration.md`.\n- For a concrete prior case study, read `references\u002Fqwen36_case_study.md` only\n  when Qwen3.5\u002FQwen3.6 details are relevant.\n","# Quant Recipe Search\n\nUse this skill when quantization is an iterative recipe search, not a one-off\nPTQ run. The skill owns strategy: define success, choose the search space,\nsequence candidates, and decide the next iteration. It delegates checkpoint\ngeneration, serving, evaluation, monitoring, and metric comparison to the\nexisting execution skills.\n\nTreat a direct request such as \"find the best quantization recipe and generate a\nPTQ checkpoint for this model\" as enough to start. Recover local state first,\nthen ask only for missing decisions that change the search.\n\n## Skill Boundaries\n\n- Use `modelopt-ptq` to produce and validate checkpoints.\n- Use `modelopt-deployment` to serve checkpoints and debug serving-specific flags.\n- Use `modelopt-evaluation` to create NEL configs and submit evals.\n- Use `modelopt-launching-evals` to run, resume, debug, and analyze NEL runs.\n- Use `modelopt-monitor` for active job tracking.\n- Use `modelopt-accessing-mlflow` for MLflow artifact lookup.\n- Use `modelopt-compare-results` for validated baseline-vs-candidate deltas and score-field comparability.\n\nDo not duplicate those workflows here. This skill should leave the user with a clear recipe portfolio, success metric, experiment sequence, and next decision.\n\n## Problem\n\nThe task is to find the best recipe for a user-defined target, not merely to\nproduce a quantized checkpoint. A generated PTQ checkpoint is only a candidate.\nIt becomes a recommended recipe only after evaluation and comparison against the\nmatching baseline.\n\nRequired inputs before planning candidates:\n\n- **Optimization goal:** compute\u002Fthroughput, memory\u002Flatency, or a custom metric.\n- **Primary quantization family:** for example NVFP4, W4A16 NVFP4, FP8\u002FW8A8,\n  INT4\u002FAWQ, or a custom mixed set.\n- **Benchmark set or baseline results:** the user-defined acceptance surface.\n\nIf any of these are missing, ask for them. Do not silently default to FP8\u002FW8A8\nor call a checkpoint \"best\" before evaluation.\n\nDefault success rule: maximize the chosen performance objective while keeping\neach benchmark within 1 percentage point of the matching BF16\u002FFP16 baseline.\nNear-threshold or noisy regressions require reruns before making a decision.\n\n## Search Space\n\nKeep the search space explicit. A candidate recipe is a tuple across these axes:\n\n- **Numeric format:** FP8\u002FW8A8, NVFP4\u002FW4A4, W4A16 NVFP4, INT4\u002FAWQ, or mixed\n  formats such as NVFP4+FP8.\n- **Calibration\u002Fsearch algorithm:** max calibration, MSE calibration, GPTQ,\n  AWQ, AutoQuant scoring, and calibration dataset or sample-count variants.\n- **Selection method:** manual\u002Fheuristic rules, sensitivity-guided manual\n  recipes, AutoQuant selection, or a hybrid of AutoQuant plus manual overrides.\n- **Module family:** attention, MLP, MoE experts, routers\u002Fgates, embeddings,\n  `lm_head`, adapters, vision encoders, and model-specific modules.\n- **Layer position:** first\u002Flast transformer-layer counts or explicit ordinal\n  ranges to keep in BF16. First 3-4 and last 1-2 layers are common starting\n  candidate ranges, not defaults.\n- **Runtime fusion constraints:** modules fused by the inference library must\n  use compatible quantization. Examples: vLLM Qwen `linear_attn.in_proj_qkvz`\n  and fused MoE expert projections such as gate\u002Fup (`w1`\u002F`w3`).\n- **Calibration budget:** dataset mix, sample count, sequence length, and batch\n  settings.\n\nDo not collapse the search to one dimension such as numeric format only. Read\n`references\u002Frecipe_iteration.md` when choosing concrete axes or candidates.\n\n## Design Workflow\n\n1. **Recover state**\n   - Read result tables, recipe logs, AutoQuant states, sensitivity reports, and\n     experiment notes before proposing new work.\n   - Ask `modelopt-monitor`, `modelopt-launching-evals`, or `modelopt-compare-results` to recover active\n     job state and completed metrics when needed.\n\n2. **Define the target**\n   - Confirm the optimization goal, primary quantization family, benchmark set,\n     accuracy-loss threshold, calibration budget, and cost metric.\n   - Include quantization metadata such as scale storage in active-cost or size\n     estimates.\n\n3. **Pick baselines and first candidates**\n   - Always include BF16\u002FFP16 and a near-lossless FP8\u002FW8A8 baseline unless FP8\n     itself is the target.\n   - For ModelOpt work, start from `modelopt_recipes`: model-specific recipes\n     first, then general PTQ presets or recipe fragments.\n   - Add an AutoQuant candidate in the requested primary family when AutoQuant\n     is available. Expect AutoQuant to find a better trade-off than a first\n     manual recipe, but validate that assumption with the same evals.\n   - Add at least one manual or sensitivity-guided candidate so AutoQuant can be\n     compared against controlled ablations and there is a fallback if AutoQuant\n     misses the best frontier or hits runtime constraints.\n   - When sensitivity or model behavior implicates boundary layers, add a\n     controlled first-layer, last-layer, or combined BF16 exclusion candidate.\n     Do not preserve boundary layers without testing the trade-off.\n\n4. **Generate candidates**\n   - Delegate checkpoint generation and PTQ validation to `modelopt-ptq`.\n   - Change one major axis at a time: format, calibration algorithm, module\n     family, layer position, granularity, or calibration data.\n   - Use AutoQuant for broad candidate generation and sensitivity reports; use\n     manual recipes for controlled module-family ablations and overrides.\n   - Resolve positional ordinals against the model's transformer block sequence.\n     For manual recipes, add those blocks to the recipe exclusions. For\n     AutoQuant or hybrid candidates, pass positional exclusions into the\n     selected implementation when supported; otherwise apply a manual override\n     to its result and record the limitation.\n\n5. **Gate before scaling**\n   - Validate checkpoint coverage and metadata.\n   - Reject or rewrite recipes that mix quantization algorithms inside a fused\n     runtime group.\n   - Ensure positional exclusions preserve complete fused runtime groups, then\n     evaluate the candidate against the same BF16 baseline and acceptance\n     criteria as every other recipe.\n   - If the checkpoint is valid but serving fails due to runtime support, do not\n     reject the recipe immediately. Delegate to `modelopt-deployment` \u002F `modelopt-debug` for small\n     patches or flags, then rerun a pipe-clean check.\n\n## Iteration Loop\n\n1. Run cheap screen evals for every candidate that passes the gates.\n2. Compare accuracy, verbosity\u002Ftoken usage, and active cost against baselines.\n3. Rerun noisy or near-threshold results before labeling a regression.\n4. Decide the next candidate:\n   - Accuracy drop: protect or ablate sensitive module families, try MSE\u002FGPTQ,\n     use AutoQuant sensitivity to choose overrides, or test first\u002Flast-layer\n     BF16 exclusions when evidence points to boundary sensitivity.\n   - Poor performance\u002Fcost: quantize the next high-cost active family, adjust\n     active-cost objective, or try a more aggressive format.\n   - AutoQuant underperforms manual recipes: inspect sensitivity reports,\n     achieved bits, excluded modules, and runtime-fusion constraints; keep the\n     manual recipe in the portfolio instead of forcing the AutoQuant result.\n   - Runtime incompatibility: rewrite around fused groups or isolate deployment\n     support from checkpoint quality.\n   - Repeated AutoQuant recipes: inspect achieved bits and recipe hashes, then\n     adjust constraints before launching a larger sweep.\n5. Promote only when `modelopt-compare-results` shows no failed external sanity check,\n   the candidate is comparable to the validated measured baseline, and the\n   user-defined goal is met. An externally unverified baseline is non-blocking.\n\nWhen the user has authorized QAD, a candidate is a recipe plus the QAD settings\nfrom the `modelopt-qad` skill's recipe loop, which ends each recipe with a pass or a\nclear miss. A PTQ recipe that already meets the threshold counts as a pass.\nStep the layout from that outcome:\n\n- After a pass, add one module group to the passing recipe, in order of\n  savings on the success metric; the MTP head is its own group. A group can be\n  a slice: its middle layers with the first and last few kept, or some of its\n  projections (such as GDN `in_proj` only, or attention `linear_qkv` only).\n  Find and evaluate the new layout's best PTQ recipe, starting from the\n  passing recipe's calibration, before its QAD. Step one layout at a time;\n  the `modelopt-qad` recipe loop runs its knob variants in parallel.\n- After a clear miss, shrink the added group to a slice; if that misses too,\n  drop the group and try the next one on the last passing recipe. If no recipe\n  has passed yet, report no QAD pass and return to the accuracy-drop steps\n  above.\n- Stop when every group has been added or dropped, or the QAD budget is spent.\n  Return the passing recipe that best meets the success metric, with its QAD\n  checkpoint.\n\nMaintain a recipe portfolio table with recipe name, objective, active-cost\nestimate, calibration notes, checkpoint path, eval\u002Flog references, accuracy,\nverbosity, positional exclusions, and decision.\n\n## References\n\n- For recipe design, search-space details, sensitivity, and active-cost\n  accounting, read `references\u002Frecipe_iteration.md`.\n- For a concrete prior case study, read `references\u002Fqwen36_case_study.md` only\n  when Qwen3.5\u002FQwen3.6 details are relevant.\n",{"name":5,"description":7},"main","e169b4c2bc62a32c45aaef793cde7599dee2c1b2","skills\u002Fmodelopt-quant-recipe-search",806,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer","https:\u002F\u002Fnvidia.github.io\u002FModel-Optimizer\u002F","https:\u002F\u002Fgithub.com\u002FNVIDIA","https:\u002F\u002Fnvidia.com","2026-11-15 02:45:59",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add NVIDIA\u002FModel-Optimizer --skill modelopt-quant-recipe-search",{"tool":39,"command":40},"bunx","bunx skills add NVIDIA\u002FModel-Optimizer --skill modelopt-quant-recipe-search",{"tool":42,"command":43},"pnpm","pnpm dlx skills add NVIDIA\u002FModel-Optimizer --skill modelopt-quant-recipe-search","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-quant-recipe-search\u002FSKILL.md",true,[47,53],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"references\u002Fqwen36_case_study.md","REFERENCE","参考文档",1497,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-quant-recipe-search\u002Freferences\u002Fqwen36_case_study.md",{"path":54,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":55,"sourceHtmlUrl":56},"references\u002Frecipe_iteration.md",12566,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-quant-recipe-search\u002Freferences\u002Frecipe_iteration.md",[58,65,72,79,86,93],{"skillNo":59,"name":60,"nameZh":6,"description":61,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":62,"categories":63,"keywords":64,"publishedAt":19},"01M7B5XD9M965H2TAV4ZJMPF6N","modelopt-speculative-decoding","Train, debug, and validate a speculative decoding draft model (EAGLE3, DFlash, DSpark, Domino) through the ModelOpt launcher pipeline. Use when the user wants to add a new model to a draft-training pipeline, asks why a pipeline run failed, wants experiment logs reviewed, or wants to check whether a run's acceptance rate meets threshold. Triggers on \"EAGLE3\", \"DFlash\", \"DSpark\", \"Domino\", \"draft model\", \"acceptance rate\", \"speculative decoding pipeline\". Do NOT use for quantizing a model (use modelopt-ptq) or serving a checkpoint (use modelopt-deployment).","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-speculative-decoding\u002FSKILL.md",[],[],{"skillNo":66,"name":67,"nameZh":6,"description":68,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M7B5X9SG5FHW174VFK979DQ9","modelopt-run-agentx","Run the AgentX agentic serving benchmark with the SemiAnalysis harness. Use for AgentX setup, dataset downloads, concurrency sweeps, or serving latency and throughput comparisons.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-run-agentx\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M7B5X4C0XGVA8N0BXF9N4HCX","modelopt-release-cherry-pick","Audit merged bug-fix PRs and release NVBugs for missing cherry-pick labels, then cherry-pick labeled PRs into a release branch and open a PR. Use when asked to \"cherry-pick PRs for release\u002FX.Y.Z\", \"pick PRs to release branch\", \"verify cherry-pick labels\", or \"cherry-pick labeled PRs\".","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-release-cherry-pick\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M7B5WRMPZE23YXCFMZBM8KJA","modelopt-qad","Run explicitly requested ModelOpt Quantization-Aware Distillation (QAD) through Megatron Bridge (recommended) or Hugging Face Transformers to recover a measured BF16-to-PTQ accuracy gap. Use only when the user explicitly asks for QAD, including its topology, data preparation, launch, resume, checkpoint export, or recovery decisions.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-qad\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M7B5WJHB4D787EZCY4RNN44T","modelopt-ptq","Use when the user asks to \"quantize a model\", \"run PTQ\", \"post-training quantization\", \"NVFP4 quantization\", \"FP8 quantization\", \"INT8 quantization\", \"INT4 AWQ\", \"quantize LLM\", \"quantize MoE\", \"quantize VLM\", or needs to produce a quantized HuggingFace checkpoint from a pretrained model using ModelOpt. Do NOT use for multi-candidate recipe exploration or optimization (use modelopt-quant-recipe-search).","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-ptq\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":19},"01M7B5WDB7KYNXD84PQRA69RPT","modelopt-monitor","Monitor submitted jobs (PTQ, evaluation, deployment) on SLURM clusters. Use when the user asks \"check job status\", \"is my job done\", \"monitor my evaluation\", \"what's the status of the PTQ\", \"check on job\" followed by a SLURM job ID, or after any skill submits a long-running job. Also triggers on \"nel status\", \"squeue\", or any request to check progress of a previously submitted job.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-monitor\u002FSKILL.md",[],[]]