fix: skip DQ->MatMulNBits fusion when weight/scale initializer is shared

[Rishi-Dave]

Summary

• Reject the DQ→MatMulNBits fusion when the weight or scale initializer is shared by multiple consumers (e.g. tied-embedding pattern).
• Prevents a crash in TransposeDQWeightsForMatMulNBits ("Missing required scale") when loading models like Whisper's decoder_model_merged_uint8.onnx.
• Adds a regression test covering Int4x2 and UInt4x2 with and without zero-points.

Motivation

Fixes #28306.

ORT 1.25 regressed on quantized Whisper decoder models. PR #27769 broadened the DQ→MatMulNBits fusion, but its CheckOutputEdges guard only checks DQ output edges; it does not catch the case where two DQ nodes share the same weight + scale initializers at their inputs. The first fusion consumes the shared initializer; the second fusion then asserts in qdq_actions.cc:136 because the initializer has been removed from the graph.

Changes

• onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selectors.cc: in DQMatMulNodeGroupSelector::Check, after the existing CheckOutputEdges guard, return false if either the weight initializer or the scale initializer has more than one consumer node. This conservatively preserves both DQ nodes as DequantizeLinear + MatMul rather than attempting a fusion whose first half would invalidate the second.
• onnxruntime/test/optimizer/qdq_matmulnbits_transformer_test.cc: adds DQMatMulNotConvertedToMatMulNBits_SharedWeight (4 variants: Int4x2/UInt4x2 × with/without zero-points). Builds two DQ nodes pointing at a shared weight + shared scale initializer, runs the QDQ transformer at TransformerLevel::Level2, and asserts no MatMulNBits is emitted (MatMul=2, DequantizeLinear=2, MatMulNBits=0). Without the fix the second fusion crashes before the assertion runs, so this is a real regression guard.

Test Plan

• C++ unit tests: DQMatMulNotConvertedToMatMulNBits_SharedWeight is included in onnxruntime_test_all and will run in CI.
• Local lintrunner: lintrunner -a clean on the diff.
• Manual: loading onnx-community/whisper-tiny decoder_model_merged_uint8.onnx via InferenceSession no longer asserts in TransposeDQWeightsForMatMulNBits.

Fixes #28306

[Copilot]

Pull request overview

This PR fixes a regression in the QDQ transformer by preventing DequantizeLinear -> MatMul/Gemm from being fused into MatMulNBits when the DQ’s weight or scale initializer is shared across multiple consumers (e.g., tied embeddings), which previously could lead to a crash during TransposeDQWeightsForMatMulNBits.

Changes:

• Add a selector-side guard to skip DQ→MatMulNBits fusion when the weight or scale initializer has multiple consumers.
• Add a regression test that constructs two DQ→MatMul paths sharing the same weight+scale initializers and asserts that no MatMulNBits nodes are produced.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selectors.cc	Adds a consumer-count guard in DQMatMulNodeGroupSelector::Check to reject fusion when weight/scale initializers are shared.
onnxruntime/test/optimizer/qdq_matmulnbits_transformer_test.cc	Adds a regression test ensuring shared weight/scale initializers prevent fusion and avoid the prior crash.

---

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