[ssss141414]

Summary

• add a reusable NPZ-backed ORT calibration reader on top of the existing validated InputDataDataset
• honor calibration_load_path in static QDQ quantization before task-specific or random fallback
• document the NPZ contract and cover sample limiting, invalid limits, rewind, validation, and end-to-end static quantization

Why

Static quantization accepted calibration_load_path in configuration but did not consume it. For the Wav2Vec2 optimization demo this silently selected random calibration, which was not representative of speech and failed the declared tensor-quality gate.

Reproduction

$env:UV_DEFAULT_INDEX = "https://packagefeedproxy.microsoft.io/pypi/simple"
$env:PYTHONPATH = (Resolve-Path .\src).Path
..\.venv\Scripts\python.exe -m pytest tests\unit\datasets\test_input_data.py tests\unit\test_quant_passes.py -q
..\.venv\Scripts\python.exe -m pytest tests\unit\packaging\test_wheel_install.py -q

The immutable end-to-end replay is in demo/deliveries/v2/bundle: run repro.ps1 to verify hashes, WinML 0.4.0, and exact PR HEAD 828fce6dea347360338307a6d863d13edb78a5ef before rebuilding. repro-run.ps1 is the underlying clean public-CLI replay.

Validation

• 57 focused unit tests passed.
• 2 packaging/wheel tests passed with the required corporate package mirror.
• git diff --check origin/main...HEAD passed.
• Full unit suite collected 10,647 tests and reached 83%; it was terminated by a native windowsml access violation while tests/unit/session/conftest.py initialized the EP catalog. This is outside the changed calibration path. The two packaging failures observed before termination were independently rerun with the corporate mirror and passed.
• Two clean public WinML rebuilds produced byte-identical W8A16 artifacts.
• Delivery receipt status: validated; champion SHA-256 316802bd3f3912fa1df24c12181555a98834c1a1efaf998c700e59171f789a92.

Measured result

On Snapdragon X QNN NPU, burst mode and graph finalization mode 3:

• speech-calibrated baseline p50: 8495.1555 ms
• optimized leader p50: 32.5255 ms
• mean paired gain: 99.6151%
• 95% CI: [99.5989%, 99.6313%]
• protocol: four sessions per side, 5 warmups + 30 timed iterations, paired A/B/B/A/A/B/B/A order
• correctness/tensor gates: pass
• task quality: provisional because the frozen speech windows do not have emotion-regression labels

Complexity review

Fallback seven-rung review (Ponytail unavailable): Lean already. Ship. Net removable lines: 0. The implementation reuses existing loading, validation, dtype conversion, batching, and sample slicing; the new class is only the ORT CalibrationDataReader adapter required for get_next and rewind. No dependency, option, or model-specific path was added.

Risks

• NPZ keys must match model input names; existing dataset validation reports mismatches.
• Calibration sample limiting preserves current dataset batching semantics.
• Detail profiling fell back to basic attribution for the multi-partition graph, so hardware/memory/DRAM counters are unavailable; no such counters are inferred.

Rollback

Revert commits 828fce6dea347360338307a6d863d13edb78a5ef and c8c8ad2a5183159a45e48dc5ae1ab200b9fcb453. Static quantization will return to task-specific/random calibration fallback, and callers relying on calibration_load_path will again have that configuration ignored.

Review state

• Graph Scout: NO_MATERIAL_OMISSION
• Complexity review: complete via documented fallback
• First Check-in Reviewer findings resolved: NPZ contract documented, non-positive sample limits rejected, and immutable delivery v2 binds and verifies the exact PR implementation
• Final Check-in Reviewer: READY_FOR_CHECK_IN with no blocking findings
• All nine hosted PR checks passed at HEAD 828fce6dea347360338307a6d863d13edb78a5ef
• PR intentionally remains Draft; do not merge or mark ready without explicit approval

This change was discovered while using the auto-optimize skill on audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim.