[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7B5VDN1KB12HHFMAY9X2YMX":3,"skill-same:01M7B5VDN1KB12HHFMAY9X2YMX":83},{"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},"01M7B5VDN1KB12HHFMAY9X2YMX","modelopt-day0-release",null,"Deterministic end-to-end driver for day-0 quantized-checkpoint releases — chains PTQ → evaluation → comparison with enforced gates between stages (the evaluation stage deploys the checkpoint itself), and returns a publish decision (ACCEPT \u002F REGRESSION \u002F INDETERMINATE \u002F ANOMALOUS \u002F INFEASIBLE). Use when the user asks to \"release a model at day-0\", \"quantize and validate model X is within N% of baseline and tell me if it's publishable\", or \"run the full day-0 workflow\". Do NOT use for single-stage requests — quantizing only (use modelopt-ptq), serving only (use modelopt-deployment), evaluating only (use modelopt-evaluation), or comparing two existing runs (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-day0-release\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: modelopt-day0-release\ndescription: Deterministic end-to-end driver for day-0 quantized-checkpoint releases — chains PTQ → evaluation → comparison with enforced gates between stages (the evaluation stage deploys the checkpoint itself), and returns a publish decision (ACCEPT \u002F REGRESSION \u002F INDETERMINATE \u002F ANOMALOUS \u002F INFEASIBLE). Use when the user asks to \"release a model at day-0\", \"quantize and validate model X is within N% of baseline and tell me if it's publishable\", or \"run the full day-0 workflow\". Do NOT use for single-stage requests — quantizing only (use modelopt-ptq), serving only (use modelopt-deployment), evaluating only (use modelopt-evaluation), or comparing two existing runs (use modelopt-compare-results).\nlicense: Apache-2.0\n---\n\n# Day-0 Release\n\nDrive a model from a pretrained checkpoint to a publish decision for a quantized\ncheckpoint, in a fixed sequence with a gate after every stage. This skill is a\n**conductor**: it sequences the existing domain skills and enforces the gates —\nit does not re-implement quantization, serving, evaluation, or comparison.\n\n**Goal (the default day-0 criterion):** a quantized checkpoint smaller than the\nsource, with accuracy drop within the threshold (default \u003C1%) on the standard\nbenchmark set versus the matching baseline, plus a publish recommendation.\n\n## When to use\n\nUse only for the full goal-driven release. For a single stage, route to the\ndomain skill directly: quantize → **ptq**, serve → **deployment**, evaluate →\n**evaluation**, compare two existing runs → **compare-results**.\n\n## Inputs\n\nResolve these before starting (ask the user for anything missing):\n\n- **Model** — HF handle or checkpoint path.\n- **Recipe \u002F qformat** — e.g. `nvfp4`, `fp8`, or a recipe path. One candidate for v1.\n- **Cluster \u002F launcher** — from `clusters.yaml` (see the common skill's\n  `environment-setup.md`).\n- **Eval set** — defaults to the evaluation skill's AA suite\n  (`recipes\u002Ftasks\u002Faa\u002F`).\n- **Threshold** — max accuracy drop; default `0.01` (1%).\n\n## The chain\n\n```text\nsetup ─▶ PTQ ─▶ canary ─▶ baseline-eval ─▶ quantized-eval ─▶ compare ─▶ verbosity ─▶ closeout\n          │        │           │                │              │           │\n       gate_ptq  \u002Fhealth    gate_run         gate_run    gate_compare  gate_verbosity\n                 + 1 gen\n```\n\nThe **evaluation** skill deploys the model it evaluates (it stands up its own\nendpoint per run), so there is no separate deploy *stage* — a serving failure\nduring evaluation surfaces through the eval gate (`DEPLOYMENT_HEALTH_FAILED`) and\ntriages to the **deployment** skill (see Step 4). The Step 2b **canary** is not\nthat: it runs *before* any evaluation precisely so an unservable checkpoint is\ncaught in ~15 min rather than after a multi-hour eval.\n\nAccuracy (Step 5) and verbosity (Step 5b) are **independent gates**; closeout\nrequires both.\n\nRun each stage by invoking the domain skill, then run its gate before\nproceeding. **Do not advance past a failed gate.** Copy this checklist and track\nprogress:\n\n```text\n- [ ] Step 0: Resolve inputs; confirm threshold and eval set\n- [ ] Step 1: Setup gate — creds present, cluster reachable\n- [ ] Step 2: PTQ (ptq skill) → gate_ptq.py\n- [ ] Step 2b: Serving canary — \u002Fhealth + one generation (deployment skill)\n- [ ] Step 3: Baseline eval (evaluation skill, deploys source) → gate_run.py\n- [ ] Step 4: Quantized eval (evaluation skill, deploys candidate) → gate_run.py\n- [ ] Step 5: Compare (compare-results skill) → external sanity → gate_compare.py → decision\n- [ ] Step 5b: Verbosity gate → gate_verbosity.py\n- [ ] Step 6: Closeout — report + publish recommendation\n```\n\n### Step 1 — Setup gate\n\nUse the common skill's `credentials.md` and `remote-execution.md` to confirm\ncredentials and cluster reachability. If either fails, stop with\n`SYSTEMIC` — do not start PTQ.\n\n### Step 2 — PTQ\n\nInvoke the **ptq** skill to produce the quantized checkpoint. Then gate:\n\n```bash\n# The ptq skill's post-PTQ validation produces a validation-summary JSON (size\n# ratio + layer-precision counts + metadata diffs; see the ptq skill's\n# references\u002Fcheckpoint-validation.md). v1 gates on that summary:\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_ptq.py\" --summary \u003Cvalidation-summary.json>\n#   add `--recipe \u003Cqformat>` to override the recipe recorded in the summary\n```\n\n`gate_ptq.py` returns JSON `{pass, failure_class, detail}`. On `pass: false`,\nbranch on `failure_class` (see **Triage** below). Do not evaluate an\nunvalidated checkpoint.\n\n#### Step 2b — serving canary (MANDATORY before Step 3)\n\nThe canary itself is **already specified** by the ptq skill: see\n`modelopt-ptq\u002Freferences\u002Fcheckpoint-validation.md` (required gate, canary query and the\n`Serving canary` row of its report table) and `modelopt-ptq\u002FSKILL.md`. Run it there rather\nthan re-deriving it here — `gate_ptq.py` checks size, coverage and metadata, not\nwhether the checkpoint *loads*, and skipping the canary has cost a full baseline\neval against a checkpoint the serving stack could never load.\n\nOn failure use `failure_class: CHECKPOINT_NOT_SERVABLE` and drop to the\n**deployment** skill; do not proceed to Step 3.\n\n**What that spec does not cover: writing a canary that cannot lie.** Both of these\nproduced a wrong verdict on a large MoE, in opposite directions:\n\n- **Poll ceiling > load time**, with headroom. A 50 min poll against a 51 min load\n  reported failure for a checkpoint that serves fine. Large MoE loads are\n  CPU-bound fp8 dequant (~50 min); 0% GPU during *load* is normal.\n- **Print an explicit `RESULT:` on every path** and exit non-zero on failure. A\n  poll loop that falls through to the generation test exits 0 and reads as PASS.\n- Server log on shared storage, not node-local `\u002Ftmp`.\n- Canary the **as-exported** artifact, not a copy you modified to make it work.\n\n### Step 3 — Baseline eval\n\nThe baseline is the **source** (pre-quantization) model on the same task set and\nsampling params. Always run a fresh baseline via the **evaluation** skill,\nwhich deploys the source model itself. Gate every run with `gate_run.py`.\n\nBefore submitting either side, record per-task run counts, uncertainty method,\nand stopping rule. For retry-level Welch, plan at least two fresh independent\nfull-benchmark runs per side (SciCode: at least eight), increasing these fixed\ncounts up front if more precision is needed. One AA-suite invocation does not\nsupply multiple full runs for GPQA or IFBench: in-run repeats are not independent\nretry scores. A task-specific method using repeat\u002Ftrial data instead must justify\nits observation unit and dependence assumptions.\n\nComplete the declared plan before the final assessment. Result-dependent reruns,\nrepeated looks, or selecting candidates on these results are exploratory: require\nindependent confirmation under a fresh declared plan, or a justified\nsequential\u002Fselection-aware method, before confidence-based ACCEPT.\n\n### Step 4 — Quantized eval\n\nInvoke the **evaluation** skill on the quantized checkpoint, matching the\nbaseline's task set, sampling params, and declared run plan. The evaluation skill\nstands up the serving endpoint itself (it builds the `deployment.command`, e.g. a\n`vllm serve …`), so a serving failure surfaces here as a failed `gate_run.py`\nwith `DEPLOYMENT_HEALTH_FAILED`. When that happens, **drop to the deployment\nskill** to reproduce and debug serving in isolation (serve the checkpoint\nstandalone, confirm `\u002Fhealth` + one generation, iterate on flags \u002F TP \u002F image \u002F\nenv vars) rather than burning full eval cycles on a broken endpoint — then carry\nthe working command back into NEL's `deployment.command` and resume the eval. If\nthe checkpoint genuinely can't serve, `POINT_INFEASIBLE`.\n\n**Before submitting: assert baseline\u002Fcandidate config parity.** The candidate\nconfig must differ from the baseline's in nothing but checkpoint path and served\nmodel name. Diff mechanically — an eyeball pass misses this:\n\n```bash\ndiff \u003C(grep -vE 'checkpoint_path|served_model_name' baseline.yaml) \\\n     \u003C(grep -vE 'checkpoint_path|served_model_name' candidate.yaml)\n```\n\nAny other difference biases the comparison and invalidates the gate: a mismatched\n`parallelism` between the two sides was worth ~2 pp, enough to invert the sign of\nthe delta. If the model card splits sampling params per scenario, apply the same\nsplit to both sides. Never set an unbounded `request_timeout` (`1e9`) — it turns a\ntransient stall into a job that holds its GPUs until the wall clock kills it.\n\nGate:\n\n```bash\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_run.py\" --run \u003Crun-summary.json>\n```\n\nA `pass: false` here means the run is incomplete or invalid (judge\u002Fparse error,\ndropped samples) — do **not** compare scores from it.\n\n### Step 5 — Compare\n\nInvoke the **compare-results** skill. It must perform the shared external\nbaseline sanity check before the candidate-delta gate. A failed check is\n`ANOMALOUS` with failure class `EXTERNAL_BASELINE_MISMATCH`: investigate and\nrerun the baseline. If no credible comparable external score exists, record the\nbaseline as externally unverified and continue using the validated measured\nbaseline.\n\n#### Uncertainty — check BEFORE trusting any per-task verdict\n\nUse the [compare-results assessment](..\u002Fmodelopt-compare-results\u002Freferences\u002Fassess-results.md)\nagainst the allowed loss, in matching units. A point estimate within threshold\nor an SE below it does not establish acceptance; require the entire signed-loss\nconfidence interval above the minimum acceptable margin.\n\nApply Step 3's declared run\u002Fstopping plan.\n\n**Re-running does not guarantee fresh samples.** With a warm NEL response cache a \"re-run\" can\nreplay cached responses — two runs came back bit-identical to 16 digits. But an equal score is a\nsignal to check provenance (invocation id, output dir, response artifacts), **not** proof of replay:\ndiscrete-scored tasks tie legitimately. Confirm replay before discarding a run as a duplicate.\n\nAfter recording the external status, produce per-task deltas and run:\n\n```bash\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_compare.py\" \\\n    --baseline \u003Cbaseline_scores.json> --candidate \u003Ccandidate_scores.json> \\\n    --threshold 0.01\n```\n\nThe threshold is a fraction of each task's score scale. Most AA tasks report\n0-100, but some (e.g. `tau2_bench_telecom` `Result`) report 0-1; the gate infers\neach task's scale (0-1 if both scores are within [0, 1], else 0-100) and\nnormalizes the drop accordingly, so `--threshold 0.01` means \"≤1 pt on a 0-100\ntask \u002F ≤0.01 on a 0-1 task\" uniformly. Pass `--scales '{\"task\": max}'` to\noverride inference if a task's scores happen to fall in an ambiguous range.\n\n`gate_compare.py` is a preliminary point-estimate check, not the final accuracy\nverdict. It cannot override external sanity or the interval assessment. Combined\ndecision (external sanity failures and implausible scores take precedence):\n\n- **ACCEPT (accuracy only)** — no external check failed and every task's\n  assessment is **Within margin** → continue to **Step 5b**. Advance to Step 6\n  only once the verbosity gate also passes. A missing comparable external score\n  is not a failure; report it as externally unverified.\n- **REGRESSION** — at least one task's assessment is **Substantial regression**.\n  **v1 stops here and reports** which tasks regressed by how much.\n- **INDETERMINATE** — otherwise, any task's target assessment is unresolved or\n  unavailable (including insufficient valid runs). Stop with `do-not-publish`;\n  report the missing evidence rather than accepting the point-estimate verdict.\n- **ANOMALOUS** — external baseline sanity failed, or scores are otherwise\n  implausible (e.g. baseline lower than candidate by a large margin, or a task\n  score is outside its valid range) → correct the baseline or surface it.\n\n#### Step 5b — Verbosity gate (MANDATORY; independent of accuracy)\n\n`gate_compare.py` does not measure verbosity, so stopping after Step 5 leaves a\nhard gate unmeasured. Step 5's ACCEPT is accuracy-only — Step 6 requires both.\n\n```bash\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_verbosity.py\" \\\n    --baseline \u003Cbaseline_eval_root> --candidate \u003Ccandidate_eval_root> \\\n    --glob 'eval_*' --exclude _high --threshold 0.05\n```\n\nExit codes follow the other gates: `0` pass, `1` the gate failed, `2` it could not read its input.\nA `2` means fix the invocation, not the checkpoint — check `--baseline`\u002F`--candidate`, `--glob`\n(`no_files_matched` names the pattern that missed), and `--exclude`. If the detail names\n`collapsed_keys`, the artifact tree gave several tasks one name: point the gate at a tree whose\ntask dirs are `\u003Charness>.\u003Ctask>` rather than `\u003Cinvocation_id>.\u003Cjob_index>`, or re-sync so each\n`artifacts\u002F` carries its own `config.yml`.\n\nRead `response_stats.avg_completion_tokens` from each task's\n`artifacts\u002Feval_factory_metrics.json`. Do **not** use the `reasoning.*` fields:\n`reasoning.*_tokens` are always `0` (only the reasoning\u002Fcontent split is missing)\nand the `*_words` siblings are a proxy that can disagree with the gate — one task\nread +6.10% FAIL in words and +1.32% PASS in tokens.\n\nThe gate is two-sided; a large drop in output length is also a change.\n\nTwo filters are mandatory: same reasoning effort (pass `--exclude` explicitly — it is empty by\ndefault, since which tier is canonical is run-specific) and complete runs only —\nruns within 1% of the largest `successful_count` that **both** sides can match (exact equality\nwould make one dropped sample unmeasurable). A task with no such count on both sides is\n`not_comparable`. When the matched count is below a run\none side has, the task carries `truncated_comparison`.\n\nTasks sharing no sample count are `not_comparable`, not a delta. Means under\n~1000 tokens carry a `short_output_warning` — read absolute counts there.\n\n### Step 6 — Closeout\n\nReport the decision with: source vs output size + ratio, per-task baseline \u002F\ncandidate \u002F signed-loss CI \u002F margin \u002F assessment, **the verbosity verdict from Step 5b**,\nexternal source and sanity status, MLflow run IDs, and a publish recommendation\n(publish \u002F do-not-publish). Archive artifacts to the workspace.\n\n**Publish exactly what was evaluated.** Verify mechanically, not by path\nconvention: inode-compare a shard in the evaluated directory against the one being\npublished. A day-0 run leaves near-identical sibling exports\nthat differ by one calibration suffix — prefix rejected ones `REJECTED-` so the\nartifact cannot be picked by autocomplete, and re-run the Step 2b canary against\nthe final path after any move.\n\n## Triage (gate failure → decision)\n\nFor completed evaluations, apply\n[evaluation run validation](..\u002Fmodelopt-evaluation\u002Freferences\u002Frun-validation.md) before\ntriage; its policy overrides retry actions below. Warn but do not retry solely\nfor failures below tolerance. For runs that fail tolerance or an independent\nvalidation check, return findings and a recommendation to the parent\u002Fuser for\nreview; do not automatically resubmit.\n\nMap a gate's `failure_class` to the next action:\n\n| `failure_class` | Action |\n| --- | --- |\n| `INFRA_TRANSIENT` | Retry the stage once; if it recurs, `SYSTEMIC`. |\n| `MODEL_UNSUPPORTED` | PATCH: fix the recipe pattern \u002F add model support (ptq skill owns the patch loop), then retry. If unpatchable, `POINT_INFEASIBLE`. |\n| `QUANT_COVERAGE_FAILURE` | PATCH: fix the recipe wildcard so intended layers are covered; re-run PTQ. |\n| `SIZE_NOT_REDUCED` | The output is not smaller than the source. If the recipe cannot shrink the source (e.g. mxfp4 under nvfp4, or fp8 under fp8), record `source_precision` in the validation summary and re-run the gate — that states *why* the growth is expected. Otherwise treat it as a real compression failure: check that the recipe matched the intended parameter mass and that the exporter did not retain the original tensors. `accept_size_growth: true` waives it unconditionally (no growth bound) and is a last resort, not the fix — it records no reason, so prefer `source_precision` where the claim is checkable. |\n| `CHECKPOINT_NOT_SERVABLE` | The Step 2b canary could not load\u002Fgenerate. Usually a tensor-naming or config-schema mismatch between the exporter and the serving stack, or missing\u002Fdangling auxiliary files (tokenizer). Fix the export; do not evaluate. |\n| `VERBOSITY_EXCEEDED` | Re-check run hygiene first (mixed reasoning effort, partial runs, unequal sample counts) — that has explained every false positive so far. If the delta survives matched, complete runs, it is a real behavioural change; do not publish on accuracy alone. |\n| `DEPLOYMENT_HEALTH_FAILED` | Drop to the **deployment** skill: reproduce serving standalone (`\u002Fhealth` + one generation), debug flags \u002F image \u002F TP \u002F env, then carry the working command into NEL's `deployment.command` and retry the eval. If it can't serve, `POINT_INFEASIBLE`. |\n| `EVAL_JUDGE_FAILED` | For completed runs, apply evaluation run validation above; return tolerance failures to the parent\u002Fuser without automatic retry. Otherwise investigate auth\u002Frate limits before retrying. |\n| `SAMPLE_ACCOUNTING_FAILED` | Investigate dropped\u002Ffailed samples before trusting scores. |\n| `EXTERNAL_BASELINE_MISMATCH` | Investigate baseline configuration, correct it, rerun the baseline, and repeat external sanity before comparison. |\n| `USER_CONFIG_ERROR` | Correct it from the request, workspace, or model\u002Fconfig metadata and retry; if irrecoverable, return `ANOMALOUS` with evidence. |\n| `UNKNOWN` | Investigate with the owning domain skill; if unresolved, return `ANOMALOUS` with the evidence and next automated retry or patch action. |\n\n`gate_ptq.py` also emits non-blocking `notes` (present on every result). Size growth is\n**blocking by default**, waived to a note only when the validation summary's `source_precision`\nis already at or below the recipe's target bits and the growth is within what that explains. Recording\n`source_precision` is part of the ptq skill's validation table, so the waiver is reachable from the\nnormal pipeline. A BF16 source that failed to compress still fails, which is the case this check\nexists for.\n\n`SYSTEMIC` (cluster down, dataset unavailable) aborts the whole run.\n`POINT_INFEASIBLE` means this (model, recipe) can't work as configured.\n\n## Output\n\nReturn a decision, not a raw artifact:\n\n- `ACCEPT` + report + publish recommendation\n- `REGRESSION` + which tasks failed the threshold and by how much\n- `INDETERMINATE` + unresolved tasks, missing evidence, and `do-not-publish`\n- `ANOMALOUS` \u002F `INFEASIBLE` + reason and next automated action\n- Always: workspace path + MLflow run IDs for traceability\n\n## Scope (v1)\n\nIn v1: the linear chain + gates + report. On `REGRESSION`, v1 reports and stops.\nDeferred to a follow-up: the evaluator-optimizer recipe loop (compare → pick the\nnext recipe → re-run PTQ), which needs the bigpareto integration and a shared\nconfig\u002Fresult schema.\n","# Day-0 Release\n\nDrive a model from a pretrained checkpoint to a publish decision for a quantized\ncheckpoint, in a fixed sequence with a gate after every stage. This skill is a\n**conductor**: it sequences the existing domain skills and enforces the gates —\nit does not re-implement quantization, serving, evaluation, or comparison.\n\n**Goal (the default day-0 criterion):** a quantized checkpoint smaller than the\nsource, with accuracy drop within the threshold (default \u003C1%) on the standard\nbenchmark set versus the matching baseline, plus a publish recommendation.\n\n## When to use\n\nUse only for the full goal-driven release. For a single stage, route to the\ndomain skill directly: quantize → **ptq**, serve → **deployment**, evaluate →\n**evaluation**, compare two existing runs → **compare-results**.\n\n## Inputs\n\nResolve these before starting (ask the user for anything missing):\n\n- **Model** — HF handle or checkpoint path.\n- **Recipe \u002F qformat** — e.g. `nvfp4`, `fp8`, or a recipe path. One candidate for v1.\n- **Cluster \u002F launcher** — from `clusters.yaml` (see the common skill's\n  `environment-setup.md`).\n- **Eval set** — defaults to the evaluation skill's AA suite\n  (`recipes\u002Ftasks\u002Faa\u002F`).\n- **Threshold** — max accuracy drop; default `0.01` (1%).\n\n## The chain\n\n```text\nsetup ─▶ PTQ ─▶ canary ─▶ baseline-eval ─▶ quantized-eval ─▶ compare ─▶ verbosity ─▶ closeout\n          │        │           │                │              │           │\n       gate_ptq  \u002Fhealth    gate_run         gate_run    gate_compare  gate_verbosity\n                 + 1 gen\n```\n\nThe **evaluation** skill deploys the model it evaluates (it stands up its own\nendpoint per run), so there is no separate deploy *stage* — a serving failure\nduring evaluation surfaces through the eval gate (`DEPLOYMENT_HEALTH_FAILED`) and\ntriages to the **deployment** skill (see Step 4). The Step 2b **canary** is not\nthat: it runs *before* any evaluation precisely so an unservable checkpoint is\ncaught in ~15 min rather than after a multi-hour eval.\n\nAccuracy (Step 5) and verbosity (Step 5b) are **independent gates**; closeout\nrequires both.\n\nRun each stage by invoking the domain skill, then run its gate before\nproceeding. **Do not advance past a failed gate.** Copy this checklist and track\nprogress:\n\n```text\n- [ ] Step 0: Resolve inputs; confirm threshold and eval set\n- [ ] Step 1: Setup gate — creds present, cluster reachable\n- [ ] Step 2: PTQ (ptq skill) → gate_ptq.py\n- [ ] Step 2b: Serving canary — \u002Fhealth + one generation (deployment skill)\n- [ ] Step 3: Baseline eval (evaluation skill, deploys source) → gate_run.py\n- [ ] Step 4: Quantized eval (evaluation skill, deploys candidate) → gate_run.py\n- [ ] Step 5: Compare (compare-results skill) → external sanity → gate_compare.py → decision\n- [ ] Step 5b: Verbosity gate → gate_verbosity.py\n- [ ] Step 6: Closeout — report + publish recommendation\n```\n\n### Step 1 — Setup gate\n\nUse the common skill's `credentials.md` and `remote-execution.md` to confirm\ncredentials and cluster reachability. If either fails, stop with\n`SYSTEMIC` — do not start PTQ.\n\n### Step 2 — PTQ\n\nInvoke the **ptq** skill to produce the quantized checkpoint. Then gate:\n\n```bash\n# The ptq skill's post-PTQ validation produces a validation-summary JSON (size\n# ratio + layer-precision counts + metadata diffs; see the ptq skill's\n# references\u002Fcheckpoint-validation.md). v1 gates on that summary:\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_ptq.py\" --summary \u003Cvalidation-summary.json>\n#   add `--recipe \u003Cqformat>` to override the recipe recorded in the summary\n```\n\n`gate_ptq.py` returns JSON `{pass, failure_class, detail}`. On `pass: false`,\nbranch on `failure_class` (see **Triage** below). Do not evaluate an\nunvalidated checkpoint.\n\n#### Step 2b — serving canary (MANDATORY before Step 3)\n\nThe canary itself is **already specified** by the ptq skill: see\n`modelopt-ptq\u002Freferences\u002Fcheckpoint-validation.md` (required gate, canary query and the\n`Serving canary` row of its report table) and `modelopt-ptq\u002FSKILL.md`. Run it there rather\nthan re-deriving it here — `gate_ptq.py` checks size, coverage and metadata, not\nwhether the checkpoint *loads*, and skipping the canary has cost a full baseline\neval against a checkpoint the serving stack could never load.\n\nOn failure use `failure_class: CHECKPOINT_NOT_SERVABLE` and drop to the\n**deployment** skill; do not proceed to Step 3.\n\n**What that spec does not cover: writing a canary that cannot lie.** Both of these\nproduced a wrong verdict on a large MoE, in opposite directions:\n\n- **Poll ceiling > load time**, with headroom. A 50 min poll against a 51 min load\n  reported failure for a checkpoint that serves fine. Large MoE loads are\n  CPU-bound fp8 dequant (~50 min); 0% GPU during *load* is normal.\n- **Print an explicit `RESULT:` on every path** and exit non-zero on failure. A\n  poll loop that falls through to the generation test exits 0 and reads as PASS.\n- Server log on shared storage, not node-local `\u002Ftmp`.\n- Canary the **as-exported** artifact, not a copy you modified to make it work.\n\n### Step 3 — Baseline eval\n\nThe baseline is the **source** (pre-quantization) model on the same task set and\nsampling params. Always run a fresh baseline via the **evaluation** skill,\nwhich deploys the source model itself. Gate every run with `gate_run.py`.\n\nBefore submitting either side, record per-task run counts, uncertainty method,\nand stopping rule. For retry-level Welch, plan at least two fresh independent\nfull-benchmark runs per side (SciCode: at least eight), increasing these fixed\ncounts up front if more precision is needed. One AA-suite invocation does not\nsupply multiple full runs for GPQA or IFBench: in-run repeats are not independent\nretry scores. A task-specific method using repeat\u002Ftrial data instead must justify\nits observation unit and dependence assumptions.\n\nComplete the declared plan before the final assessment. Result-dependent reruns,\nrepeated looks, or selecting candidates on these results are exploratory: require\nindependent confirmation under a fresh declared plan, or a justified\nsequential\u002Fselection-aware method, before confidence-based ACCEPT.\n\n### Step 4 — Quantized eval\n\nInvoke the **evaluation** skill on the quantized checkpoint, matching the\nbaseline's task set, sampling params, and declared run plan. The evaluation skill\nstands up the serving endpoint itself (it builds the `deployment.command`, e.g. a\n`vllm serve …`), so a serving failure surfaces here as a failed `gate_run.py`\nwith `DEPLOYMENT_HEALTH_FAILED`. When that happens, **drop to the deployment\nskill** to reproduce and debug serving in isolation (serve the checkpoint\nstandalone, confirm `\u002Fhealth` + one generation, iterate on flags \u002F TP \u002F image \u002F\nenv vars) rather than burning full eval cycles on a broken endpoint — then carry\nthe working command back into NEL's `deployment.command` and resume the eval. If\nthe checkpoint genuinely can't serve, `POINT_INFEASIBLE`.\n\n**Before submitting: assert baseline\u002Fcandidate config parity.** The candidate\nconfig must differ from the baseline's in nothing but checkpoint path and served\nmodel name. Diff mechanically — an eyeball pass misses this:\n\n```bash\ndiff \u003C(grep -vE 'checkpoint_path|served_model_name' baseline.yaml) \\\n     \u003C(grep -vE 'checkpoint_path|served_model_name' candidate.yaml)\n```\n\nAny other difference biases the comparison and invalidates the gate: a mismatched\n`parallelism` between the two sides was worth ~2 pp, enough to invert the sign of\nthe delta. If the model card splits sampling params per scenario, apply the same\nsplit to both sides. Never set an unbounded `request_timeout` (`1e9`) — it turns a\ntransient stall into a job that holds its GPUs until the wall clock kills it.\n\nGate:\n\n```bash\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_run.py\" --run \u003Crun-summary.json>\n```\n\nA `pass: false` here means the run is incomplete or invalid (judge\u002Fparse error,\ndropped samples) — do **not** compare scores from it.\n\n### Step 5 — Compare\n\nInvoke the **compare-results** skill. It must perform the shared external\nbaseline sanity check before the candidate-delta gate. A failed check is\n`ANOMALOUS` with failure class `EXTERNAL_BASELINE_MISMATCH`: investigate and\nrerun the baseline. If no credible comparable external score exists, record the\nbaseline as externally unverified and continue using the validated measured\nbaseline.\n\n#### Uncertainty — check BEFORE trusting any per-task verdict\n\nUse the [compare-results assessment](..\u002Fmodelopt-compare-results\u002Freferences\u002Fassess-results.md)\nagainst the allowed loss, in matching units. A point estimate within threshold\nor an SE below it does not establish acceptance; require the entire signed-loss\nconfidence interval above the minimum acceptable margin.\n\nApply Step 3's declared run\u002Fstopping plan.\n\n**Re-running does not guarantee fresh samples.** With a warm NEL response cache a \"re-run\" can\nreplay cached responses — two runs came back bit-identical to 16 digits. But an equal score is a\nsignal to check provenance (invocation id, output dir, response artifacts), **not** proof of replay:\ndiscrete-scored tasks tie legitimately. Confirm replay before discarding a run as a duplicate.\n\nAfter recording the external status, produce per-task deltas and run:\n\n```bash\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_compare.py\" \\\n    --baseline \u003Cbaseline_scores.json> --candidate \u003Ccandidate_scores.json> \\\n    --threshold 0.01\n```\n\nThe threshold is a fraction of each task's score scale. Most AA tasks report\n0-100, but some (e.g. `tau2_bench_telecom` `Result`) report 0-1; the gate infers\neach task's scale (0-1 if both scores are within [0, 1], else 0-100) and\nnormalizes the drop accordingly, so `--threshold 0.01` means \"≤1 pt on a 0-100\ntask \u002F ≤0.01 on a 0-1 task\" uniformly. Pass `--scales '{\"task\": max}'` to\noverride inference if a task's scores happen to fall in an ambiguous range.\n\n`gate_compare.py` is a preliminary point-estimate check, not the final accuracy\nverdict. It cannot override external sanity or the interval assessment. Combined\ndecision (external sanity failures and implausible scores take precedence):\n\n- **ACCEPT (accuracy only)** — no external check failed and every task's\n  assessment is **Within margin** → continue to **Step 5b**. Advance to Step 6\n  only once the verbosity gate also passes. A missing comparable external score\n  is not a failure; report it as externally unverified.\n- **REGRESSION** — at least one task's assessment is **Substantial regression**.\n  **v1 stops here and reports** which tasks regressed by how much.\n- **INDETERMINATE** — otherwise, any task's target assessment is unresolved or\n  unavailable (including insufficient valid runs). Stop with `do-not-publish`;\n  report the missing evidence rather than accepting the point-estimate verdict.\n- **ANOMALOUS** — external baseline sanity failed, or scores are otherwise\n  implausible (e.g. baseline lower than candidate by a large margin, or a task\n  score is outside its valid range) → correct the baseline or surface it.\n\n#### Step 5b — Verbosity gate (MANDATORY; independent of accuracy)\n\n`gate_compare.py` does not measure verbosity, so stopping after Step 5 leaves a\nhard gate unmeasured. Step 5's ACCEPT is accuracy-only — Step 6 requires both.\n\n```bash\npython \"$SKILL_DIR\u002Fscripts\u002Fgate_verbosity.py\" \\\n    --baseline \u003Cbaseline_eval_root> --candidate \u003Ccandidate_eval_root> \\\n    --glob 'eval_*' --exclude _high --threshold 0.05\n```\n\nExit codes follow the other gates: `0` pass, `1` the gate failed, `2` it could not read its input.\nA `2` means fix the invocation, not the checkpoint — check `--baseline`\u002F`--candidate`, `--glob`\n(`no_files_matched` names the pattern that missed), and `--exclude`. If the detail names\n`collapsed_keys`, the artifact tree gave several tasks one name: point the gate at a tree whose\ntask dirs are `\u003Charness>.\u003Ctask>` rather than `\u003Cinvocation_id>.\u003Cjob_index>`, or re-sync so each\n`artifacts\u002F` carries its own `config.yml`.\n\nRead `response_stats.avg_completion_tokens` from each task's\n`artifacts\u002Feval_factory_metrics.json`. Do **not** use the `reasoning.*` fields:\n`reasoning.*_tokens` are always `0` (only the reasoning\u002Fcontent split is missing)\nand the `*_words` siblings are a proxy that can disagree with the gate — one task\nread +6.10% FAIL in words and +1.32% PASS in tokens.\n\nThe gate is two-sided; a large drop in output length is also a change.\n\nTwo filters are mandatory: same reasoning effort (pass `--exclude` explicitly — it is empty by\ndefault, since which tier is canonical is run-specific) and complete runs only —\nruns within 1% of the largest `successful_count` that **both** sides can match (exact equality\nwould make one dropped sample unmeasurable). A task with no such count on both sides is\n`not_comparable`. When the matched count is below a run\none side has, the task carries `truncated_comparison`.\n\nTasks sharing no sample count are `not_comparable`, not a delta. Means under\n~1000 tokens carry a `short_output_warning` — read absolute counts there.\n\n### Step 6 — Closeout\n\nReport the decision with: source vs output size + ratio, per-task baseline \u002F\ncandidate \u002F signed-loss CI \u002F margin \u002F assessment, **the verbosity verdict from Step 5b**,\nexternal source and sanity status, MLflow run IDs, and a publish recommendation\n(publish \u002F do-not-publish). Archive artifacts to the workspace.\n\n**Publish exactly what was evaluated.** Verify mechanically, not by path\nconvention: inode-compare a shard in the evaluated directory against the one being\npublished. A day-0 run leaves near-identical sibling exports\nthat differ by one calibration suffix — prefix rejected ones `REJECTED-` so the\nartifact cannot be picked by autocomplete, and re-run the Step 2b canary against\nthe final path after any move.\n\n## Triage (gate failure → decision)\n\nFor completed evaluations, apply\n[evaluation run validation](..\u002Fmodelopt-evaluation\u002Freferences\u002Frun-validation.md) before\ntriage; its policy overrides retry actions below. Warn but do not retry solely\nfor failures below tolerance. For runs that fail tolerance or an independent\nvalidation check, return findings and a recommendation to the parent\u002Fuser for\nreview; do not automatically resubmit.\n\nMap a gate's `failure_class` to the next action:\n\n| `failure_class` | Action |\n| --- | --- |\n| `INFRA_TRANSIENT` | Retry the stage once; if it recurs, `SYSTEMIC`. |\n| `MODEL_UNSUPPORTED` | PATCH: fix the recipe pattern \u002F add model support (ptq skill owns the patch loop), then retry. If unpatchable, `POINT_INFEASIBLE`. |\n| `QUANT_COVERAGE_FAILURE` | PATCH: fix the recipe wildcard so intended layers are covered; re-run PTQ. |\n| `SIZE_NOT_REDUCED` | The output is not smaller than the source. If the recipe cannot shrink the source (e.g. mxfp4 under nvfp4, or fp8 under fp8), record `source_precision` in the validation summary and re-run the gate — that states *why* the growth is expected. Otherwise treat it as a real compression failure: check that the recipe matched the intended parameter mass and that the exporter did not retain the original tensors. `accept_size_growth: true` waives it unconditionally (no growth bound) and is a last resort, not the fix — it records no reason, so prefer `source_precision` where the claim is checkable. |\n| `CHECKPOINT_NOT_SERVABLE` | The Step 2b canary could not load\u002Fgenerate. Usually a tensor-naming or config-schema mismatch between the exporter and the serving stack, or missing\u002Fdangling auxiliary files (tokenizer). Fix the export; do not evaluate. |\n| `VERBOSITY_EXCEEDED` | Re-check run hygiene first (mixed reasoning effort, partial runs, unequal sample counts) — that has explained every false positive so far. If the delta survives matched, complete runs, it is a real behavioural change; do not publish on accuracy alone. |\n| `DEPLOYMENT_HEALTH_FAILED` | Drop to the **deployment** skill: reproduce serving standalone (`\u002Fhealth` + one generation), debug flags \u002F image \u002F TP \u002F env, then carry the working command into NEL's `deployment.command` and retry the eval. If it can't serve, `POINT_INFEASIBLE`. |\n| `EVAL_JUDGE_FAILED` | For completed runs, apply evaluation run validation above; return tolerance failures to the parent\u002Fuser without automatic retry. Otherwise investigate auth\u002Frate limits before retrying. |\n| `SAMPLE_ACCOUNTING_FAILED` | Investigate dropped\u002Ffailed samples before trusting scores. |\n| `EXTERNAL_BASELINE_MISMATCH` | Investigate baseline configuration, correct it, rerun the baseline, and repeat external sanity before comparison. |\n| `USER_CONFIG_ERROR` | Correct it from the request, workspace, or model\u002Fconfig metadata and retry; if irrecoverable, return `ANOMALOUS` with evidence. |\n| `UNKNOWN` | Investigate with the owning domain skill; if unresolved, return `ANOMALOUS` with the evidence and next automated retry or patch action. |\n\n`gate_ptq.py` also emits non-blocking `notes` (present on every result). Size growth is\n**blocking by default**, waived to a note only when the validation summary's `source_precision`\nis already at or below the recipe's target bits and the growth is within what that explains. Recording\n`source_precision` is part of the ptq skill's validation table, so the waiver is reachable from the\nnormal pipeline. A BF16 source that failed to compress still fails, which is the case this check\nexists for.\n\n`SYSTEMIC` (cluster down, dataset unavailable) aborts the whole run.\n`POINT_INFEASIBLE` means this (model, recipe) can't work as configured.\n\n## Output\n\nReturn a decision, not a raw artifact:\n\n- `ACCEPT` + report + publish recommendation\n- `REGRESSION` + which tasks failed the threshold and by how much\n- `INDETERMINATE` + unresolved tasks, missing evidence, and `do-not-publish`\n- `ANOMALOUS` \u002F `INFEASIBLE` + reason and next automated action\n- Always: workspace path + MLflow run IDs for traceability\n\n## Scope (v1)\n\nIn v1: the linear chain + gates + report. On `REGRESSION`, v1 reports and stops.\nDeferred to a follow-up: the evaluator-optimizer recipe loop (compare → pick the\nnext recipe → re-run PTQ), which needs the bigpareto integration and a shared\nconfig\u002Fresult schema.\n",{"name":5,"license":15,"description":7},"main","e169b4c2bc62a32c45aaef793cde7599dee2c1b2","skills\u002Fmodelopt-day0-release",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-day0-release",{"tool":39,"command":40},"bunx","bunx skills add NVIDIA\u002FModel-Optimizer --skill modelopt-day0-release",{"tool":42,"command":43},"pnpm","pnpm dlx skills add NVIDIA\u002FModel-Optimizer --skill modelopt-day0-release","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002FSKILL.md",true,[47,53,57,61,65,71,75,79],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"scripts\u002Fgate_compare.py","SCRIPT","脚本",8337,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Fscripts\u002Fgate_compare.py",{"path":54,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":55,"sourceHtmlUrl":56},"scripts\u002Fgate_ptq.py",14325,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Fscripts\u002Fgate_ptq.py",{"path":58,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":59,"sourceHtmlUrl":60},"scripts\u002Fgate_run.py",5834,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Fscripts\u002Fgate_run.py",{"path":62,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":63,"sourceHtmlUrl":64},"scripts\u002Fgate_verbosity.py",22990,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Fscripts\u002Fgate_verbosity.py",{"path":66,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":69,"sourceHtmlUrl":70},"tests\u002Fevals.json","REFERENCE","参考文档",6774,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Ftests\u002Fevals.json",{"path":72,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":73,"sourceHtmlUrl":74},"tests\u002Ftest_agent_definitions.py",3411,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Ftests\u002Ftest_agent_definitions.py",{"path":76,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":77,"sourceHtmlUrl":78},"tests\u002Ftest_gates.py",33430,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Ftests\u002Ftest_gates.py",{"path":80,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":81,"sourceHtmlUrl":82},"tests\u002Ftest_skill_discovery.py",2299,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FModel-Optimizer\u002Fblob\u002Fmain\u002Fskills\u002Fmodelopt-day0-release\u002Ftests\u002Ftest_skill_discovery.py",[84,91,98,105,112,119],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"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":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"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":99,"name":100,"nameZh":6,"description":101,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":102,"categories":103,"keywords":104,"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":106,"name":107,"nameZh":6,"description":108,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":109,"categories":110,"keywords":111,"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":113,"name":114,"nameZh":6,"description":115,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":116,"categories":117,"keywords":118,"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":120,"name":121,"nameZh":6,"description":122,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":125,"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",[],[]]