Compare ONNX skill contracts and guidance against current Quark ONNX documentation and source entry points. Use when maintainers need to verify that ONNX install docs, custom-op registry, QConfig fields, preset and calibration lists, AutoSearchPro presets, deployment-target gates, or example-script invocations still match upstream Quark ONNX reality. Trigger for "check ONNX doc drift", "are ONNX skills still accurate", "verify against Quark ONNX docs", "fact-check quark-onnx-* skills", or after a Quark release when upstream ONNX documentation may have changed.
Instrucciones de origen · Vista previa de solo lectura
name
quark-onnx-doc-drift-check
description
Compare ONNX skill contracts and guidance against current Quark ONNX documentation and source entry points. Use when maintainers need to verify that ONNX install docs, custom-op registry, QConfig fields, preset and calibration lists, AutoSearchPro presets, deployment-target gates, or example-script invocations still match upstream Quark ONNX reality. Trigger for "check ONNX doc drift", "are ONNX skills still accurate", "verify against Quark ONNX docs", "fact-check quark-onnx-* skills", or after a Quark release when upstream ONNX documentation may have changed.
Provide a lightweight ONNX-side governance check focused on documentation accuracy. While
quark-onnx-skill-sync audits source-code changes broadly and applies the actual updates, this
skill focuses specifically on whether the ONNX skills' user-facing guidance — preset names,
calibration methods, custom-op names, QConfig field names, AutoSearchPro presets, ORT install
matrices, example-script invocations — still matches what Quark's own ONNX docs and source say.
Think of this as a fact-checker for the quark-onnx-* skill family.
Inputs
Quark upstream ONNX docs and source code (read-only)
Current ONNX skill SKILL.md files under .claude/skills-impl/{l1-atomic,l2-workflows,l3-recipes}/onnx/
ONNX-side shared schemas under .claude/skills-impl/shared/contracts/
Outputs: validation_report.md
Lists where current ONNX skill guidance has drifted from upstream Quark ONNX docs and source.
# ONNX Documentation Drift Report## What to Check### 1. Quantization Presets Still Exist
Skills reference specific preset names in their decision tables. Verify each is still defined in
`quark/onnx/quantization/config/custom_config.py`:
**Critical presets to verify** (referenced by `quark-onnx-quant-plan`, `quark-onnx-ptq-workflow`,
`quark-onnx-autosearch-pro`):
-`XINT8` (and the `EnableNPUCnn=True` companion flag)
-`A8W8`, `A16W8`-`BF16`, `BFP16`-`MX*` family (e.g. `MXINT8`, `MXFP8_E4M3`, `MXFP8_E5M2`, `MXFP4`, `MXFP6_E3M2`, )
Weights-only INT4 path (e.g. )
Check that (and ) still contain every
calibration method named in :
Verify that still exposes every algorithm config
class referenced by the skills:
,
, ,
, ,
The surface is copied into skill decision tables, generated scripts, and the workflow
example. Verify every field is still present in / :
, ,
,
,
Skills reference Quark ONNX custom ops by name in error messages, debug guidance, and the
opset domain check used by . Verify each is still
registered in :
family
The domain string
Check that the recipe's preset names still exist in
:
, , ,
Compare 's ORT package/version matrix with:
(the authoritative matrix)
installation instructions
(GPU/EP guidance)
core ONNX deps (, , )
Verify the supported tuples cited by the skill still match.
Common EP names that must be consistent: , ,
, .
lists a deployment-target compatibility table (CPU / CUDA / ROCm /
AMD NPU CNN / AMD NPU Transformer). Verify the gating logic still matches what
accepts (e.g. BFP16 forbidden on NPU CNN, XINT8 + ).
Verify that the JSON schemas in match what the ONNX
workflow actually produces:
— matches the ONNX with ?
— includes ONNX-specific fields used by (preset, calibrationconfig, EnableNPUCnn, usedatanodes)?
— matches 's manifest (generated script path, exact command, resolved )?
Check that the example invocation in and the
walkthrough in still agree on:
the field names used in the worked example,
the argument order,
the imports from ,
the calibration data reader pattern from .
Verify the rule cited by and
the workflow still matches what / enforce.
— report findings but do not modify skills. Modifications go through
.
— preset names, calibration-method names, custom-op names, QConfig
field names, ORT versions, and example-script signatures matter most because users will copy-paste
them into scripts the workflow will run.
— a skill that recommends a deleted preset, calibration method,
custom op, or QConfig field will cause immediate user failures at script generation or runtime.
— a new preset, calibration method, or AutoSearchPro preset
that is not yet documented in skills is a coverage gap, not an error.
— never touch skills. Cross-cut findings (e.g. a shared
field) are surfaced but deferred to the torch maintainer.
Quantization presets: 11 checked, 10 match, 1 new (not in skills)
Calibration methods: 6 checked, 6 match
Algorithm configs: 8 checked, 7 match, 1 drift
QConfig fields: 8 checked, 8 match
Custom-op registry: 4 checked, 4 match
AutoSearchPro presets: 4 checked, 4 match
ORT install matrix: 5 rows checked, 4 match, 1 drift
Deployment-target gates: 5 checked, 5 match
in skills → now in (case change)
: Generated scripts in and will
fail at (ImportError)
: Update the import + plan tables in and
SKILL.md, then re-validate the YOLOv8 worked example
Skills recommend
now pins
: Copy-paste install commands in will downgrade users on a
freshly installed environment
: Update the ROCm row in 's ORT install matrix
added to
: Users asking about MXFP6 won't see it in 's preset table
: Add row to the preset table; consider whether AutoSearchPro should expose
a matching
now accepts a mapping
: Skill mentions but not the new parameter
: Optional — extend calibration row when a real user request lands
Select scope: Full check or focused on specific ONNX skills / facts (e.g. only AutoSearchPro presets)?
Run checks: Compare ONNX skill content against upstream source and docs using the grep-driven extraction patterns above.
Classify findings: Critical (broken guidance), warning (outdated), info (gap in coverage).
Report: Produce the drift report as validation_report.md.
Hand off: If fixes are needed, route to quark-onnx-skill-sync for the actual updates — never patch skills from this check.
Recovery
If upstream ONNX docs are inaccessible, report which checks could not be performed and which ONNX skills are therefore unverified.
If quark/onnx/operators/custom_ops/build_custom_ops.py cannot be parsed (e.g. moved or renamed), mark quark-onnx-install and quark-onnx-debug as potentially affected even when no string drift is observed — custom-op load failures are runtime-only.
If drift is detected, hand off to quark-onnx-skill-sync with the specific findings so it can apply targeted fixes; do not edit skills from this check.
If a finding spans both backends (e.g. a shared/contracts/ schema change), report it but defer cross-cut fixes to the torch maintainer rather than editing torch skills from here.
