Detect upstream Quark ONNX changes that affect the ONNX skill family and classify required updates. Use when Quark ONNX docs, custom-op registry, quantization config presets, calibration methods, AutoSearchPro presets, ONNX Runtime install matrix, or source behavior under `quark/onnx/` may have drifted from the skill contracts. Trigger for "check if ONNX skills are up to date", "sync ONNX skills with Quark", "has Quark ONNX changed", "update ONNX skills after Quark upgrade", or when ONNX debugging reveals a mismatch between skill instructions and actual `quark.onnx` behavior.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Detect upstream Quark ONNX changes that affect the ONNX skill family and classify required updates. Use when Quark ONNX docs, custom-op registry, quantization config presets, calibration methods, AutoSearchPro presets, ONNX Runtime install matrix, or source behavior under `quark/onnx/` may have drifted from the skill contracts. Trigger for "check if ONNX skills are up to date", "sync ONNX skills with Quark", "has Quark ONNX changed", "update ONNX skills after Quark upgrade", or when ONNX debugging reveals a mismatch between skill instructions and actual `quark.onnx` behavior.
Audit upstream Quark ONNX source against the assumptions baked into the quark-onnx-* skill family
and report which skills or contracts need updates. ONNX-side drift is dangerous because the
quantization presets, calibration methods, custom-op registry, deployment-target gates, and
AutoSearchPro presets are referenced by name in skill decision tables — a renamed QConfig field,
a removed preset, a new calibration method, or a new custom op silently produces wrong guidance.
# ONNX Skill Sync Report## What to Check### Source Knowledge Dependencies
Each ONNX skill declares `source_knowledge` files. Check that these files still exist and that the
skill's content matches the current source:
| Skill | Source Files to Check |
|-------|----------------------|
| `quark-onnx-install` | `tools/ci/install_onnxruntime.sh`, `docs/source/install.rst`, `requirements.txt`, `quark/onnx/operators/custom_ops/build_custom_ops.py` |
| `quark-onnx-router` | `docs/source/install.rst`, `docs/source/onnx/basic_usage_onnx.rst`, `docs/source/onnx/onnx_examples.rst`, `examples/onnx/model_support.md` |
| `quark-onnx-model-intake` | `quark/onnx/__init__.py`, `quark/onnx/quantization/api.py`, , , , |
| | , , , , , |
| | , , , , , , |
| | , , , |
| | , , , , , |
| | , , , , |
— check for the supported
combinations cited by .
— ,
(, , , ) versions cited by / .
— check for the
preset list (, , , , , , , weights-only INT4 …)
referenced by and .
— fields cited by skill decision tables: , ,
, , , ,
, . Verify each still exists in / .
— , , , , ,
cited by are still present in
(and ).
— , , , ,
, , , cited by the plan and
AutoSearchPro recipe still exist in .
— , , the
family, and the opset domain are still registered under
. The build pipeline in
still matches / expectations.
— , , and the
/ / execution-provider gates referenced by
still match what accepts.
— , , ,
cited by are still defined in
and exposed by .
— the rule that triggers
: a preset name, calibration-method name, custom-op name, ORT version,
or QConfig field rename. Straightforward — update the affected table or list in the skill.
: a workflow pattern changed (e.g. AutoSearchPro now requires a two-stage
search by default; CLE moved from to a pre-pass); a deployment-target gate
was redefined; a custom-op signature changed. Requires careful skill rewriting.
: a referenced preset, calibration method, custom op, AutoSearchPro preset,
ORT package, or field was removed entirely. The skill will produce wrong guidance
until fixed.
: Read from each ONNX skill's frontmatter and
union with the table above.
: Compare current source against what the skills assume. Look for:
New presets / algorithms / calibration methods not mentioned in skills (e.g. a new
— report it and require human confirmation before applying
changes.
— preset additions, calibration-method
name normalizations, and ORT version bumps are safe to apply after review.
, run calibration mode: record the current state of all ONNX
source files as the baseline for future comparisons.
— , , ,
, and are read-only from this skill.
Drift is reported, not patched at the source.
— do not touch skills. Torch-side drift belongs to
; if a finding crosses scopes (e.g. a shared schema field), surface it
but defer the cross-cut change to the torch maintainer.
Checked: 8 ONNX skills, 17 source files, 4 contract schemas
Mechanical drift: 2 findings
Semantic drift: 1 finding
Breaking drift: 1 finding
: is a supported calibration method
: removed from (replaced by )
: Users selecting will hit at plan-to-script translation
: Drop row from calibration table; update workflow examples
: now runs a coarse + fine pass; old single-stage flag deprecated
: Skill's "single-shot search" framing is now misleading; trial budget interpretation changed
: Rewrite the "Search budget" section and the preset table to reflect two-stage semantics
: now exposes preset
: Users asking about MXFP6 won't see it in the preset list
: Add row to the preset table
: now pins (was 1.19.2)
: Skill recommends 1.19.2; users on 1.20.0 are told to downgrade
: Update ROCm row in the ORT install matrix
`use_external_data_format=True` in `quark-onnx-model-intake` and the workflow still matches
what `api.py` / `input_check.py` enforce.
11. **Already-quantized detection** — the QDQ / `com.amd.quark` domain checks that intake uses
to stop before re-quantizing still match `input_check.py`.
12. **YOLOv8 example pin** — `examples/onnx/yolo_quantization/quantize_yolo.py` still matches
the `quark-onnx-ptq-workflow` worked example (`example-xint8-yolov8n.md`).
### Drift Classification
-
**Mechanical drift**
-
**Semantic drift**
`algo_config`
-
**Breaking drift**
`QConfig`
## Audit Process
1.
**Discover dependencies**
`source_knowledge`
2.
**Check for changes**
-
`MXFP6` preset or a new `FastFinetune` knob).
- Removed presets / methods / custom ops still mentioned in skills (e.g. deprecated
`BFP16Spec` field).
- Renamed QConfig fields (e.g. `EnableNPUCnn` → `enable_npu_cnn`).
- Changed value ranges (e.g. AutoSearchPro trial budget defaults).
- Custom-op binary names / load paths that drift from `build_custom_ops.py`.
- ORT EP names changing (`CUDAExecutionProvider` / `ROCMExecutionProvider` / `VitisAIExecutionProvider`).
3. **Classify each drift**: mechanical, semantic, or breaking.
4. **Report**: Produce a `validation_report.md` with the findings.
## Rules
-
**Do not auto-fix breaking drift**
-
**Mechanical drift can be flagged for batch update**
### quark-onnx-autosearch-pro: search now two-stage by default
-
**Source**
`auto_search_pro.py`
-
**Impact**
-
**Fix**
## Mechanical Drift
### quark-onnx-quant-plan: new preset added
-
**Source**
`custom_config.py`
`MXFP6_E3M2`
-
**Impact**
-
**Fix**
`MXFP6_E3M2`
### quark-onnx-install: ORT version bump
-
**Source**
`tools/ci/install_onnxruntime.sh`
`onnxruntime-rocm==1.20.0`
-
**Impact**
-
**Fix**
Interaction Flow
Collect: Identify all source-knowledge dependencies across ONNX skills (union of frontmatter source_knowledge + the table above).
Audit: Check each source file for changes relevant to ONNX skill content (presets, algo configs, calibration methods, custom ops, ORT matrix, AutoSearchPro presets, deployment-target gates).
Classify: Tag each finding as mechanical, semantic, or breaking.
Report: Present findings in a structured validation_report.md.
Confirm: Get approval before applying any fixes; never auto-apply breaking drift.
Recovery
If a source file no longer exists (was moved or deleted), this is breaking drift. Report the missing file and suggest where the information might have moved to (e.g. custom_config.py → config/presets/).
If tools/ci/install_onnxruntime.sh is missing or replaced, surface the gap to quark-onnx-install so its ORT matrix is not silently stale.
If the ONNX custom-op build pipeline (build_custom_ops.py) has changed, 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 the upstream Quark repository is not accessible, report the blocked audit and which ONNX skills could not be verified.
If drift crosses backends (e.g. a shared/contracts/ schema field used by both torch and onnx skills), report it but hand off the cross-cut change to the torch maintainer rather than editing torch skills from here.
