fix: auto-upgrade model opset to 21 for int16/uint16 QDQ quantization

[Rishi-Dave]

Summary

• Extends existing update_opset_version helper to auto-bump opset from < 21 to 21 when QUInt16/QInt16 weight quantization is requested
• Mirrored after the existing float8 quantization opset upgrade pattern
• Adds test coverage with parametric subtests for 16-bit int quantization

Motivation

Fixes #25223.

Users exporting models from torch.export with uint16/int16 quantization hit a gap where models below opset 21 were not being upgraded. Mirroring the existing float8 branch gives users a consistent, predictable upgrade path for 16-bit QDQ.

Changes

• onnxruntime/python/tools/quantization/quant_utils.py: new elif branch in update_opset_version that bumps opset to 21 when weight_quant_type is INT16 / UINT16 and current opset is < 21. Emits a warning matching the existing float8 branch style.
• onnxruntime/test/python/quantization/test_quant_util.py: new test_update_opset_version_16bit with parametric subtests covering QUInt16 / QInt16 bumping from opset 20 → 21 and a no-op regression check for models already at opset 21.

Test Plan

python -m pytest onnxruntime/test/python/quantization/test_quant_util.py -v

All tests pass. lintrunner -a produces no changes.

[Copilot]

Pull request overview

Extends ONNX Runtime’s Python quantization utilities to automatically upgrade an input model’s ONNX opset to 21 when 16-bit integer (INT16/UINT16) QDQ quantization is requested, aligning behavior with the existing float8 opset auto-upgrade logic and adding regression tests.

Changes:

• Add an update_opset_version branch to auto-bump opset < 21 to 21 for INT16/UINT16 quantization types (with a warning).
• Add unit tests validating opset 20 → 21 upgrade and no-op behavior at opset 21 for QUInt16/QInt16.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
onnxruntime/python/tools/quantization/quant_utils.py	Adds opset auto-upgrade logic for INT16/UINT16 quantization to ensure ONNX-native QDQ compatibility.
onnxruntime/test/python/quantization/test_quant_util.py	Adds test coverage for the new opset upgrade behavior for 16-bit quantization types.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tianleiwu]

Thanks for the follow-up on the activation-type gap. I found one remaining workflow issue: the config-based static quantization path can still carry an auto-derived UseQDQContribOps flag from the original opset, so it may keep emitting contrib-domain Q/DQ after the model is bumped to opset 21. Please align that path with the direct quantize_static behavior before merge.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tianleiwu]

Re-review after commit ef62c23

The main concern from the prior round — get_qdq_config() pre-setting UseQDQContribOps=True for 16-bit types before the opset bump — is now fixed. The narrowed condition correctly limits the auto-set to 4-bit types only, and the new end-to-end test (test_quantize_via_config_int16_opset_lt21_uses_native_qdq) validates that the config-based path produces opset-21 models with native ONNX Q/DQ.

One minor suggestion on the overrides check below.

LGTM — nice work addressing the feedback.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Rishi-Dave]

Thanks for the careful review. Pushed 1d4160a addressing all four points:

• quant_utils.py opset-bump helper now also scans TensorQuantOverrides (including per-tensor convert.quant_type) for QInt16/QUInt16, so models with default 8-bit base types but 16-bit overrides take the native opset-21 path instead of falling back to contrib Q/DQ.
• Generalized the opset-bump warning wording so it reads correctly for quantize_dynamic() callers, not just QDQ static.
• quantize.py now recomputes UseQDQContribOps after the opset bump, so 16-bit/4-bit override-driven bumps no longer latch the model to com.microsoft Q/DQ.
• Added two regression tests in test_get_qdq_config.py: one for opset<21 + QUInt16 TensorQuantOverrides (asserts opset==21 and Q/DQ in default domain), and one for the mixed 16-bit + 4-bit override case.

lintrunner -a clean, all 10 test_get_qdq_config.py tests pass locally.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Rishi-Dave]

Addressed the three review comments in d1ae1ab:

• quant_utils.py: wrapped the TensorQuantOverrides scan loop in try/except (AttributeError, TypeError): pass so malformed input (non-list values, non-dict convert entries) falls through to TensorQuantOverridesHelper.is_valid() and raises the original deterministic ValueError. KeyError cannot occur (all reads use .get()), and ValueError is intentionally not caught.
• test_tensor_quant_overrides_option.py: renamed test_16bit_overrides_set_ms_domain to test_16bit_overrides_bump_opset_to_21 and flipped its assertions to expect ai.onnx opset >= 21 with native-domain Q/DQ.
• Added test_16bit_convert_quant_type_bumps_opset_to_21 covering the convert.quant_type branch specifically: top-level quant_type=QUInt8, convert={'quant_type': QInt16}, opset=20 input → bumped to >=21 with native Q/DQ.

Full file passes 21/21. lintrunner clean.

[tianleiwu]

Re-review after commit d1ae1ab

The previous concern from my earlier round (splitting the 16-bit vs 4-bit override checks in get_qdq_config) has been fully addressed by the refactored logic that separates will_bump_to_opset21 from needs_contrib_ops. Resolved that thread.

The overall approach is sound — auto-bumping opset to 21 for 16-bit types and removing the unnecessary UseQDQContribOps fallback is the right design. Test coverage is thorough with good edge-case scenarios (mixed 16-bit + 4-bit, convert.quant_type, end-to-end model verification).

One remaining item below on the silent except clause (also flagged by CodeQL).

[Rishi-Dave]

Thanks for the catch. Pushed 140d631 which replaces the empty except (AttributeError, TypeError): pass block in quant_utils.py with your suggested snippet:

except (AttributeError, TypeError):
    # Malformed overrides; structural validation is deferred to
    # TensorQuantOverridesHelper.is_valid(). Skip bump heuristic.
    logging.debug("Skipping 16-bit opset bump heuristic for TensorQuantOverrides: structure not as expected.")

import logging was already present in the module so no new imports were needed. lintrunner -a is clean. This should resolve the CodeQL "Empty except" finding as well.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

[Rishi-Dave]

Thanks for the review. Pushed da1271c8 addressing all four points:

1. quantize_static() — guarded extra_options.get("TensorQuantOverrides") with (extra_options or {}).get(...) so the path is safe when extra_options=None.
2. test_tensor_quant_overrides_option.py — replaced the bare next(...) opset lookup with the production get_opset_version() helper, so a missing ai.onnx import raises a clear ValueError instead of StopIteration. Applied to all three 16-bit opset-bump tests.
3. Same file — tightened the Q/DQ domain assertion from assertNotEqual(node.domain, ms_domain) to assertEqual(node.domain, "") with an informative failure message.
4. test_get_qdq_config.py — extended test_overrides_mixed_16bit_4bit_opset_lt21_no_contrib_ops to actually run quantize(...), load the output model, and assert opset ≥ 21 with all Q/DQ nodes in the default ONNX domain.

[tianleiwu]

Re-review after commits 140d631 and da1271c

All concerns from prior review rounds have been addressed:

1. logging.debug for malformed overrides — The bare except (AttributeError, TypeError): pass is now replaced with a logging.debug() message. Resolves both my earlier comment and the CodeQL finding.

2. extra_options None guard — (extra_options or {}).get("TensorQuantOverrides") correctly handles the case where extra_options might be None.

3. Strengthened tests — End-to-end tests now run quantize() and verify the output model opset is 21 with native ONNX domain Q/DQ nodes. Coverage for convert.quant_type and mixed 16-bit+4-bit overrides is solid.

The logic is consistent across update_opset_version() (scans quant_type and convert.quant_type in overrides) and get_qdq_config() (uses overrides_helper.get_quant_types() which checks the same fields). The will_bump_to_opset21 / needs_contrib_ops separation in get_qdq_config() is clean and correct.

LGTM.