[Ti-Tai Wang (titaiwangms)]

Describe your changes

Fix numerical errors in shared weight quantization and KQuant candidate selection:

• Symmetric RTN range: all-nonnegative groups previously used half the range of their sign-flipped equivalents. Use the absolute bound for both signs, and divide directly by half the integer range to avoid overflowing an intermediate doubled bound.
• Quantization rounding: perform FP16/BF16 weight-to-scale division in at least FP32, preserving higher promoted precision. Round the quotient before adding the integer zero point, avoiding low-precision false ties and zero-point-dependent tie-to-even results.
• KQuant storage-aware scoring: evaluate candidates after integer-zero-point conversion, scale conversion to the original weight dtype, code regeneration, and dequantization. The previous free-offset objective could rank candidates differently from their stored representation. Include same-settings RTN as a per-group baseline, preferring RTN on ties.
• KQuant intermediate range: normalize each group consistently for RMS, refinement, and scoring, then restore scale units before storage. This prevents magnitude-driven overflow in otherwise representable groups.

Preserves upstream INT3 support, the minimum-positive-scale floor, complete-group chunking, optional compute-device selection, output device/dtype, and checkpoint packing formats. Existing quantized checkpoints are not rewritten on load. Newly quantized codes and scales can change.

Validation

Fresh offline CPU validation of this port onto c063dd3bd071eb1bcb3eeb7e6d102454c4a33947:

• 978 passed, 6 CUDA skips across test/common/quant, local KQuant/RTN pass tests, test_quant_utils.py, and test_rtn_mixed_qmoe.py.
• Excluded only the two Hub-dependent test functions test_kquant.py::test_kquant and test_rtn.py::test_gptq (8 parameterizations each); all other collected tests in the selected files ran or skipped normally.
• 156 new regression cases, included above: sign-flip invariance, overflow boundaries, rounding/division precision, stored-candidate ranking reversal, packed reconstruction, RTN weighted-error bounds, nonfinite inputs, subnormal scale-floor preservation, and exact chunked/unchunked agreement.
• Pylint: 10.00/10; Ruff formatting, git diff --check, package-import provenance, and CLI help checks passed.
• Ran lintrunner. The local Ruff executable is 0.16.2, versus the repository pin 0.15.11, and reports CPY001 for the existing yearless Microsoft copyright header. The identical failure reproduces on untouched olive/common/quant/hf_utils.py. Supplemental Ruff checks pass with only that baseline diagnostic excluded; no repository lint configuration or existing headers were changed.

Earlier model evidence (not a rerun of this upstream port)

The same numerical fixes in the earlier implementation were evaluated on Gemma4-26B-A4B-it with a fixed 500-question MMLU-Pro subset and the same evaluation recipe:

Quantizer	Before	After
KQuant	395/500 (79.0%)	398/500 (79.6%)
RTN	392/500 (78.4%)	397/500 (79.4%)

These are bounded observations, not a claim that the remaining BF16-to-INT4 accuracy gap is fixed. The RTN fallback bounds the computed per-group weighted reconstruction objective (subject to floating-point scoring), not downstream task accuracy. Full-model inference, CUDA qualification, and quantization-time/peak-memory measurements were not repeated for this latest-main port.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass. Selected offline tests pass; Hub/CUDA coverage is excluded or skipped as described above.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a. Formatting/Pylint passed; full local lintrunner has the baseline Ruff-version issue described above.
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

Release note: Correct RTN symmetric ranges and low-precision quantization rounding; rank KQuant candidates using their stored representation with a per-group RTN baseline and normalized numerical scoring. Checkpoint formats remain unchanged.

(Optional) Issue link

None.

[Copilot]

🔵 Needs a closer look

Precision-sensitive shared quantization changes warrant human review because CUDA behavior, quantization time, and peak memory remain unqualified for this port.

No open issues 🎉

What changed in this PR

This PR corrects numerical behavior in Olive’s shared weight quantization and makes KQuant candidate selection reflect the stored checkpoint representation, without changing checkpoint formats.

Changes:

• Correct symmetric RTN ranges and promote low-precision division before rounding and zero-point addition.
• Score KQuant candidates using stored parameters, with per-group normalization and an RTN fallback.
• Add numerical regression tests and document the behavior and compatibility guarantees.

File	Description
test/​common/​quant/​test_quantizer_numerics.py	Adds coverage for rounding, overflow, stored-error selection, RTN bounds, and chunking.
olive/​passes/​pytorch/​kquant.py	Adds normalized, storage-aware candidate scoring and a per-group RTN fallback.
olive/​common/​quant/​utils.py	Fixes symmetric scale calculation and quantization rounding precision.
docs/​source/​features/​quantization.md	Documents numerical changes, scoring guarantees, and checkpoint compatibility.

_{Review effort: Balanced}