[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7B5WRMPZE23YXCFMZBM8KJA":3,"skill-same:01M7B5WRMPZE23YXCFMZBM8KJA":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":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":6,"updatedAt":32,"install":33,"resources":46},"01M7B5WRMPZE23YXCFMZBM8KJA","modelopt-qad",null,"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.","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-qad\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: modelopt-qad\ndescription: >-\n  Run explicitly requested ModelOpt Quantization-Aware Distillation (QAD) through\n  Megatron Bridge (recommended) or Hugging Face Transformers to\n  recover a measured BF16-to-PTQ accuracy gap.\n  Use only when the user explicitly asks for QAD, including its topology, data\n  preparation, launch, resume, checkpoint export, or recovery decisions.\n---\n\n# ModelOpt Quantization-Aware Distillation\n\nQAD is expensive. Run it only when the user explicitly authorizes QAD for the\ntarget model or run, with a loop or GPU-hour budget; ask for the budget when it\nis missing. A Day-0, PTQ, evaluation, comparison, or recipe-search request\nalone is not authorization to start QAD.\nThe authorization covers new recipes from a recipe search only when the user\nsays so.\n\n## Confirm the gap\n\nFor both frameworks, reuse only validated, comparable BF16\u002FPTQ results and\nthe exact benchmark configuration from preceding evaluation or recipe\nsearch; run missing, invalid, or non-comparable baselines. Confirm the target\nbenchmarks and their context-length needs. Stop if the PTQ gap to BF16 is\nalready below 1%.\n\n## Follow the supported workflow\n\nRead `docs\u002Fsource\u002Fguides\u002Fquantization_aware_training_and_distillation.rst`\nfor QAD background and framework selection.\nRecommend Megatron Bridge for scalable training and its model-specific tutorials.\nChoose the applicable tutorial from `examples\u002Fmegatron_bridge\u002Ftutorials\u002FREADME.md`.\nHonor an explicit HF choice.\n\nUse the common skill's `environment-setup.md` and `workspace-management.md`;\nfor Slurm, also read `slurm-setup.md` and, when remote, `remote-execution.md`.\n\nTreat the example README and `--help` output as authoritative for mutable flags,\ncommands, containers, and checkpoint formats. Both frameworks can run on local\nGPUs or Slurm; use the user's requested environment.\n\n## Megatron Bridge execution and training policy\n\nFor Megatron Bridge, before constructing commands, read:\n\n- `examples\u002Fmegatron_bridge\u002FREADME.md`, especially PTQ, data preparation, QAD,\n  export, and Slurm usage\n- `examples\u002Fmegatron_bridge\u002F{quantize.py,distill.py}` via `--help`\n\n### Execute in this order\n\n1. **Reproduce PTQ and verify compatibility.** In the target runtime, require\n   `AutoBridge.can_handle()` for the target model and PTQ through `quantize.py`\n   to succeed while preserving the exact preceding PTQ config or recipe:\n   format, layer selection, calibration data\u002Fcount, sequence length, and seed.\n   A changed quantization setting is a new PTQ candidate and must be evaluated\n   before QAD. In the master-rank `.quant_summary.txt`, require finite positive\n   `amax` for enabled static quantizers; accept `dynamic`\u002Fformat-defined `None`\n   only when the recipe intends it. Treat the summary as rank-local under model\n   parallelism.\n2. **Choose topology explicitly.** Derive the smallest fitting node count and\n   TP\u002FPP\u002FCP\u002FEP from student and teacher architecture, the chosen sequence length,\n   and available GPU memory. Prefer CP before TP for small long-context models;\n   keep EP=1 for dense models and ETP=1 because the current `distill.py`\n   workflow does not support expert tensor parallelism. For MoE require:\n\n   - `DP = world_size \u002F (TP * PP * CP)`\n   - `EDP = world_size \u002F (EP * PP)`\n   - integral DP\u002FEDP, `num_experts % EP == 0`, and\n     `GBS % (MBS * DP) == 0`\n\n   Static-block NVFP4 recipes require TP=1; use PP for memory. With PP, balance\n   stages with `--num_layers_in_{first,last}_pipeline_stage` (the first and last\n   stages also hold the embedding and output layer) and recompute only where\n   memory needs it (`--recompute_num_layers_per_stage`). In the first job's\n   allocation, time a few iterations of each candidate layout with short\n   launches before the run starts, and keep the fastest. Give each timing\n   launch a scratch `--output_dir`: `distill.py` resumes from\n   `\u003Coutput_dir>\u002Fcheckpoints`, so a timing checkpoint there would become the\n   run's starting point.\n\n3. **Prepare the full capped dataset once.** Use suitable user-provided data, or\n   copy `examples\u002Fmegatron_bridge\u002Fdata\u002Fnemotron-cascade-2-blend.yaml` as the\n   default. Set the target tokenizer and workspace path, then materialize the\n   randomly sampled subset before training. Pack the chosen sequence length;\n   Megatron's `99,1,0` split creates the 1% validation holdout from the same\n   data. Keep every source mixed across training, never ordered as a\n   curriculum, so early stops see the whole blend.\n4. **Run and monitor QAD.** Run one training job at a time per QAD run;\n   separate runs may train in parallel. Fold startup validation into the first\n   job; do not submit separate GPU preflight jobs or split at recovery\n   iterations. Let training continue while evaluating saved\n   checkpoints, and cancel it when a stop condition below is met.\n\n### Default training policy\n\n| Setting | Default |\n| --- | --- |\n| Sequence length | 32768; adjust for target benchmarks |\n| Peak \u002F minimum LR | `1e-5` \u002F `1e-6` |\n| LR schedule | cosine |\n| LR horizon (`train_iters`) | 1000 iterations |\n| Global batch size | 512 |\n| Dataset | `nvidia\u002FNemotron-Cascade-2-SFT-Data` by default |\n| Materialized token budget | 17.3B at 32K; cover `train_iters` at the chosen length |\n| Training validation | every 25 iterations; deterministic 1% holdout; 2 batches |\n| Checkpoint interval | 50 iterations |\n| Loss logging | every 10 iterations |\n| Recovery benchmark | from the KD-loss knee, then every checkpoint |\n| Slurm duration exit | 220 minutes for a 4-hour allocation |\n\n### Run policy\n\n- `train_iters` sets the cosine LR horizon and is a loop knob: with 1000, the\n  LR stays near its peak through iteration 150. `distill.py` also stops\n  training there, with or without `exit_interval`. Stop runs by the decisions\n  below and leave `exit_interval` unset. A run that reaches `train_iters` while\n  the decision is still to continue ends as a Miss from under-recovery.\n- On Slurm, use initial step timing to submit only enough sequential jobs to reach\n  a few checkpoints past the expected KD-loss knee; never submit through\n  `train_iters` upfront. At each recovery checkpoint, submit to the next only\n  after its targeted evaluation and any triggered full suite, and only while\n  the decision is to continue.\n- On Slurm, give all training jobs the same run-specific job name and\n  `--dependency=singleton`; record job IDs and, on any stop, cancel pending jobs\n  before the active job.\n- Cancel on non-finite loss, repeated skipped iterations, or a sustained spike.\n  At iteration 50, require the loss aggregate to be lower than at iteration 10.\n- At each recovery checkpoint, first evaluate the one to three benchmarks with\n  the largest PTQ drops or that failed in earlier loops. Run the remaining\n  original PTQ suite at that checkpoint once the targeted benchmarks are within\n  run noise of 1%, or after recovery beyond run noise.\n- Start benchmark evaluation at the knee of the KD loss curve, where its steep\n  drop ends, rounded to the next checkpoint, then evaluate every checkpoint\n  while the scores still change. Before the knee, use loss only to catch\n  drift; after it, judge and pick checkpoints by benchmarks alone, since KD\n  loss keeps falling after they peak. Between near ties, the benchmark most\n  sensitive to quantization decides.\n- Run noise is 2x the standard error of a benchmark's gap to BF16: paired per\n  question when both sides score the same questions, otherwise from both\n  sides' per-sample standard errors, and pooled across runs once repeats\n  exist. Repeat a benchmark only when its gap is within run noise of 1%, up to\n  8 runs per side; the compare-results run floors still apply before a Pass.\n- Export each checkpoint to evaluate and pass the ptq skill's\n  `references\u002Fcheckpoint-validation.md` gate on it before evaluation. Run the\n  serving canary on the run's first export only; later exports change only\n  weights, so on those the gate passes without the canary: record its\n  serving-canary row as `skipped; first export \u003Cpath> passed`.\n- At each evaluated checkpoint, decide:\n  - **Pass:** every benchmark of the full original PTQ suite is within 1% of\n    BF16, each at its compare-results run floor (such as SciCode's 8 runs per\n    side). Keep the checkpoint and stop the run.\n  - **Miss:** two evaluated checkpoints past the knee bring no gain beyond run\n    noise, or two consecutive checkpoints decline beyond it. Stop and change\n    one QAD knob in the [recipe loop](#recipe-loop).\n  - **Otherwise continue** to the next checkpoint.\n- After a duration exit, resume the latest QAD checkpoint in the same output\n  directory with unchanged prepared data paths\u002Fcache, seed, topology, optimizer,\n  scheduler, iteration, and consumed-sample state; do not restart from PTQ.\n- Report, per loop, the PTQ recipe and calibration, data sample, QAD settings,\n  training topology, loss\u002Fstate, checkpoints, and comparable BF16\u002FPTQ\u002FQAD\n  results.\n\n### Recipe loop\n\nEach run trains one PTQ recipe and differs from that recipe's control in one\nQAD knob. Start from the best PTQ recipe of the preceding PTQ evaluation or\nrecipe search; its control uses the default policy. When GPUs allow, start\none-knob variants as separate runs alongside the control instead of waiting\nfor its miss. This loop chooses the QAD settings; the quant-recipe-search\nskill chooses the next recipe from its outcome. Stop when the QAD budget is\nspent and report the best result so far.\n\n1. **On a miss, change one QAD knob.** Read each missed benchmark against PTQ:\n   worse than PTQ is drift (lower the peak LR or add matching data); better\n   than PTQ but short is under-recovery (keep the LR high longer with a larger\n   `train_iters`). Knobs, in order of priority when GPUs are limited: peak LR,\n   `train_iters`, data blend, then teacher-generated data. A changed `train_iters` is a fresh run from PTQ, with\n   data materialized to cover it: resuming would change the decay length the\n   checkpoint's LR scheduler expects.\n2. **Compare a knob with its control at the same checkpoints** past the knee,\n   on benchmark scores pooled over fresh runs. A rejection holds only for that\n   PTQ recipe: rerun the control when the recipe changes.\n3. **End the loop on a recipe** with a pass, or with a clear miss once every\n   knob has missed or as soon as QAD stays flat against PTQ or far from\n   passing. Return the outcome, the best checkpoint, and its scores.\n\n## Hugging Face execution and training policy\n\nBefore constructing commands, read `examples\u002Fllm_qat\u002FREADME.md`, its\n`ARGUMENTS.md`, selected training\u002Fdataset\u002FAccelerate configs, and\n`{quantize.py,train.py,export.py} --help`.\nFollow its quantize → distill → export flow: the PTQ checkpoint is the student;\nthe original unquantized model is the teacher. Adapt the QAD training and\nAccelerate configs to the model and budget.\nResolve tutorial-relative commands and config paths from `examples\u002Fllm_qat\u002F`.\nUse the HF tutorial's evaluation and checkpoint resume guidance.\n\nReuse the Bridge procedure's PTQ-recipe preservation and one-training-job-at-a-time\nguidance. Apply the [Run policy](#run-policy)'s loss-stability and\nbenchmark-recovery stop conditions and report the same fields.\nUse HF evaluation\u002Fcheckpoint intervals; Bridge-specific topology, data defaults,\niteration thresholds, and Slurm job chaining are not HF requirements.\n","# ModelOpt Quantization-Aware Distillation\n\nQAD is expensive. Run it only when the user explicitly authorizes QAD for the\ntarget model or run, with a loop or GPU-hour budget; ask for the budget when it\nis missing. A Day-0, PTQ, evaluation, comparison, or recipe-search request\nalone is not authorization to start QAD.\nThe authorization covers new recipes from a recipe search only when the user\nsays so.\n\n## Confirm the gap\n\nFor both frameworks, reuse only validated, comparable BF16\u002FPTQ results and\nthe exact benchmark configuration from preceding evaluation or recipe\nsearch; run missing, invalid, or non-comparable baselines. Confirm the target\nbenchmarks and their context-length needs. Stop if the PTQ gap to BF16 is\nalready below 1%.\n\n## Follow the supported workflow\n\nRead `docs\u002Fsource\u002Fguides\u002Fquantization_aware_training_and_distillation.rst`\nfor QAD background and framework selection.\nRecommend Megatron Bridge for scalable training and its model-specific tutorials.\nChoose the applicable tutorial from `examples\u002Fmegatron_bridge\u002Ftutorials\u002FREADME.md`.\nHonor an explicit HF choice.\n\nUse the common skill's `environment-setup.md` and `workspace-management.md`;\nfor Slurm, also read `slurm-setup.md` and, when remote, `remote-execution.md`.\n\nTreat the example README and `--help` output as authoritative for mutable flags,\ncommands, containers, and checkpoint formats. Both frameworks can run on local\nGPUs or Slurm; use the user's requested environment.\n\n## Megatron Bridge execution and training policy\n\nFor Megatron Bridge, before constructing commands, read:\n\n- `examples\u002Fmegatron_bridge\u002FREADME.md`, especially PTQ, data preparation, QAD,\n  export, and Slurm usage\n- `examples\u002Fmegatron_bridge\u002F{quantize.py,distill.py}` via `--help`\n\n### Execute in this order\n\n1. **Reproduce PTQ and verify compatibility.** In the target runtime, require\n   `AutoBridge.can_handle()` for the target model and PTQ through `quantize.py`\n   to succeed while preserving the exact preceding PTQ config or recipe:\n   format, layer selection, calibration data\u002Fcount, sequence length, and seed.\n   A changed quantization setting is a new PTQ candidate and must be evaluated\n   before QAD. In the master-rank `.quant_summary.txt`, require finite positive\n   `amax` for enabled static quantizers; accept `dynamic`\u002Fformat-defined `None`\n   only when the recipe intends it. Treat the summary as rank-local under model\n   parallelism.\n2. **Choose topology explicitly.** Derive the smallest fitting node count and\n   TP\u002FPP\u002FCP\u002FEP from student and teacher architecture, the chosen sequence length,\n   and available GPU memory. Prefer CP before TP for small long-context models;\n   keep EP=1 for dense models and ETP=1 because the current `distill.py`\n   workflow does not support expert tensor parallelism. For MoE require:\n\n   - `DP = world_size \u002F (TP * PP * CP)`\n   - `EDP = world_size \u002F (EP * PP)`\n   - integral DP\u002FEDP, `num_experts % EP == 0`, and\n     `GBS % (MBS * DP) == 0`\n\n   Static-block NVFP4 recipes require TP=1; use PP for memory. With PP, balance\n   stages with `--num_layers_in_{first,last}_pipeline_stage` (the first and last\n   stages also hold the embedding and output layer) and recompute only where\n   memory needs it (`--recompute_num_layers_per_stage`). In the first job's\n   allocation, time a few iterations of each candidate layout with short\n   launches before the run starts, and keep the fastest. Give each timing\n   launch a scratch `--output_dir`: `distill.py` resumes from\n   `\u003Coutput_dir>\u002Fcheckpoints`, so a timing checkpoint there would become the\n   run's starting point.\n\n3. **Prepare the full capped dataset once.** Use suitable user-provided data, or\n   copy `examples\u002Fmegatron_bridge\u002Fdata\u002Fnemotron-cascade-2-blend.yaml` as the\n   default. Set the target tokenizer and workspace path, then materialize the\n   randomly sampled subset before training. Pack the chosen sequence length;\n   Megatron's `99,1,0` split creates the 1% validation holdout from the same\n   data. Keep every source mixed across training, never ordered as a\n   curriculum, so early stops see the whole blend.\n4. **Run and monitor QAD.** Run one training job at a time per QAD run;\n   separate runs may train in parallel. Fold startup validation into the first\n   job; do not submit separate GPU preflight jobs or split at recovery\n   iterations. Let training continue while evaluating saved\n   checkpoints, and cancel it when a stop condition below is met.\n\n### Default training policy\n\n| Setting | Default |\n| --- | --- |\n| Sequence length | 32768; adjust for target benchmarks |\n| Peak \u002F minimum LR | `1e-5` \u002F `1e-6` |\n| LR schedule | cosine |\n| LR horizon (`train_iters`) | 1000 iterations |\n| Global batch size | 512 |\n| Dataset | `nvidia\u002FNemotron-Cascade-2-SFT-Data` by default |\n| Materialized token budget | 17.3B at 32K; cover `train_iters` at the chosen length |\n| Training validation | every 25 iterations; deterministic 1% holdout; 2 batches |\n| Checkpoint interval | 50 iterations |\n| Loss logging | every 10 iterations |\n| Recovery benchmark | from the KD-loss knee, then every checkpoint |\n| Slurm duration exit | 220 minutes for a 4-hour allocation |\n\n### Run policy\n\n- `train_iters` sets the cosine LR horizon and is a loop knob: with 1000, the\n  LR stays near its peak through iteration 150. `distill.py` also stops\n  training there, with or without `exit_interval`. Stop runs by the decisions\n  below and leave `exit_interval` unset. A run that reaches `train_iters` while\n  the decision is still to continue ends as a Miss from under-recovery.\n- On Slurm, use initial step timing to submit only enough sequential jobs to reach\n  a few checkpoints past the expected KD-loss knee; never submit through\n  `train_iters` upfront. At each recovery checkpoint, submit to the next only\n  after its targeted evaluation and any triggered full suite, and only while\n  the decision is to continue.\n- On Slurm, give all training jobs the same run-specific job name and\n  `--dependency=singleton`; record job IDs and, on any stop, cancel pending jobs\n  before the active job.\n- Cancel on non-finite loss, repeated skipped iterations, or a sustained spike.\n  At iteration 50, require the loss aggregate to be lower than at iteration 10.\n- At each recovery checkpoint, first evaluate the one to three benchmarks with\n  the largest PTQ drops or that failed in earlier loops. Run the remaining\n  original PTQ suite at that checkpoint once the targeted benchmarks are within\n  run noise of 1%, or after recovery beyond run noise.\n- Start benchmark evaluation at the knee of the KD loss curve, where its steep\n  drop ends, rounded to the next checkpoint, then evaluate every checkpoint\n  while the scores still change. Before the knee, use loss only to catch\n  drift; after it, judge and pick checkpoints by benchmarks alone, since KD\n  loss keeps falling after they peak. Between near ties, the benchmark most\n  sensitive to quantization decides.\n- Run noise is 2x the standard error of a benchmark's gap to BF16: paired per\n  question when both sides score the same questions, otherwise from both\n  sides' per-sample standard errors, and pooled across runs once repeats\n  exist. Repeat a benchmark only when its gap is within run noise of 1%, up to\n  8 runs per side; the compare-results run floors still apply before a Pass.\n- Export each checkpoint to evaluate and pass the ptq skill's\n  `references\u002Fcheckpoint-validation.md` gate on it before evaluation. Run the\n  serving canary on the run's first export only; later exports change only\n  weights, so on those the gate passes without the canary: record its\n  serving-canary row as `skipped; first export \u003Cpath> passed`.\n- At each evaluated checkpoint, decide:\n  - **Pass:** every benchmark of the full original PTQ suite is within 1% of\n    BF16, each at its compare-results run floor (such as SciCode's 8 runs per\n    side). Keep the checkpoint and stop the run.\n  - **Miss:** two evaluated checkpoints past the knee bring no gain beyond run\n    noise, or two consecutive checkpoints decline beyond it. Stop and change\n    one QAD knob in the [recipe loop](#recipe-loop).\n  - **Otherwise continue** to the next checkpoint.\n- After a duration exit, resume the latest QAD checkpoint in the same output\n  directory with unchanged prepared data paths\u002Fcache, seed, topology, optimizer,\n  scheduler, iteration, and consumed-sample state; do not restart from PTQ.\n- Report, per loop, the PTQ recipe and calibration, data sample, QAD settings,\n  training topology, loss\u002Fstate, checkpoints, and comparable BF16\u002FPTQ\u002FQAD\n  results.\n\n### Recipe loop\n\nEach run trains one PTQ recipe and differs from that recipe's control in one\nQAD knob. Start from the best PTQ recipe of the preceding PTQ evaluation or\nrecipe search; its control uses the default policy. When GPUs allow, start\none-knob variants as separate runs alongside the control instead of waiting\nfor its miss. This loop chooses the QAD settings; the quant-recipe-search\nskill chooses the next recipe from its outcome. Stop when the QAD budget is\nspent and report the best result so far.\n\n1. **On a miss, change one QAD knob.** Read each missed benchmark against PTQ:\n   worse than PTQ is drift (lower the peak LR or add matching data); better\n   than PTQ but short is under-recovery (keep the LR high longer with a larger\n   `train_iters`). Knobs, in order of priority when GPUs are limited: peak LR,\n   `train_iters`, data blend, then teacher-generated data. A changed `train_iters` is a fresh run from PTQ, with\n   data materialized to cover it: resuming would change the decay length the\n   checkpoint's LR scheduler expects.\n2. **Compare a knob with its control at the same checkpoints** past the knee,\n   on benchmark scores pooled over fresh runs. A rejection holds only for that\n   PTQ recipe: rerun the control when the recipe changes.\n3. **End the loop on a recipe** with a pass, or with a clear miss once every\n   knob has missed or as soon as QAD stays flat against PTQ or far from\n   passing. Return the outcome, the best checkpoint, and its scores.\n\n## Hugging Face execution and training policy\n\nBefore constructing commands, read `examples\u002Fllm_qat\u002FREADME.md`, its\n`ARGUMENTS.md`, selected training\u002Fdataset\u002FAccelerate configs, and\n`{quantize.py,train.py,export.py} --help`.\nFollow its quantize → distill → export flow: the PTQ checkpoint is the student;\nthe original unquantized model is the teacher. Adapt the QAD training and\nAccelerate configs to the model and budget.\nResolve tutorial-relative commands and config paths from `examples\u002Fllm_qat\u002F`.\nUse the HF tutorial's evaluation and checkpoint resume guidance.\n\nReuse the Bridge procedure's PTQ-recipe preservation and one-training-job-at-a-time\nguidance. Apply the [Run policy](#run-policy)'s loss-stability and\nbenchmark-recovery stop conditions and report the same fields.\nUse HF evaluation\u002Fcheckpoint intervals; Bridge-specific topology, data defaults,\niteration thresholds, and Slurm job chaining are not HF requirements.\n",{"name":5,"description":7},"main","e169b4c2bc62a32c45aaef793cde7599dee2c1b2","skills\u002Fmodelopt-qad",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-qad",{"tool":39,"command":40},"bunx","bunx skills add NVIDIA\u002FModel-Optimizer --skill modelopt-qad",{"tool":42,"command":43},"pnpm","pnpm dlx skills add NVIDIA\u002FModel-Optimizer --skill modelopt-qad","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-qad\u002FSKILL.md",true,[],[48,55,62,69,76,83],{"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":15,"licenseFileUrl":6,"sourceHtmlUrl":52,"categories":53,"keywords":54,"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":56,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":59,"categories":60,"keywords":61,"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":63,"name":64,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"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":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":19},"01M7B5WXMDYXCK43YA5ZEF7ECY","modelopt-quant-recipe-search","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).","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-quant-recipe-search\u002FSKILL.md",[],[],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"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":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"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",[],[]]