[Theodore Cooper (the0cp)]

Description

Fix QDQ MatMul fusion that can produce an unsupported QLinearMatMul or incorrect results when per-axis quantization parameters are used.

The selector now validates that:

• A and Y use scalar or single-element quantization parameters.

• B uses per-tensor or supported 2D per-column quantization, with a supported axis and shapes.

Unsupported cases fall back to the original DQ → MatMul → Q path by skipping fusion, preserving correctness without changing the kernels.

fix: #33333

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

The kernel-specific checks also restrict shared execution-provider grouping, breaking previously supported QNN quantized MatMul offload.

1 open issue

• Restrict QLinearMatMul checks to optimizer MatMulSelector · New

What changed in this PR

This PR adds quantization-parameter validation to ONNX Runtime’s QDQ MatMul fusion to avoid unsupported QLinearMatMul nodes or incorrect results.

Changes:

• Require scalar or single-element quantization parameters for A and Y.
• Allow B’s per-tensor or supported two-dimensional per-column quantization, skipping fusion otherwise.
• Add five regression tests covering supported and rejected parameter shapes and axes.

File	Description
onnxruntime/​test/​optimizer/​qdq_transformer_test.cc	Adds fusion and fallback tests for MatMul quantization parameters.
onnxruntime/​core/​optimizer/​qdq_transformer/​selectors_actions/​qdq_selectors.cc	Validates parameter shapes and B’s quantization axis before fusion.

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

The focused validation matches kernel restrictions, preserves generic selection, and includes regression coverage for the affected behavior.

No open issues 🎉

1 resolved since last review

• Restrict QLinearMatMul checks to optimizer MatMulSelector

_{Review effort: Balanced}