Modify scale & offset of WhereDummyDq

[qti-hungjuiw]

Description

• Update WhereDummyDq QDQ transformer to be more selective before inserting a dummy DequantizeLinear around Where.
  • SatisfyCondition now requires the Where output to have exactly one consumer and that consumer must be QuantizeLinear (Q). Otherwise, the transform is skipped.
  • InsertDummyDQ additionally checks element type consistency between the upstream DQ input tensor type and the downstream Q output tensor type; if they differ, the transform returns without modifying the graph.
• Update the implementation of WhereDummyDq to avoid negative or zero scale value. The change maps the float value to the boundary of integer domain to ensure the scale value is positive.

  • If WhereOp get a float scalar xf and a DequantizeLinear as its two inputs, WhereDummyDq insert DQ to ensure xf = DQ(xq, scale, zp)

  • The xq, scale and zp are determined with the following table.

    	uint8	uint16	int8	int16
    xf > 0
    xq	255	65535	127	32767
    zp	127	32767	0	0
    xf < 0
    xq	0	0	-128	-32768
    zp	127	32767	0	0
    xf = 0
    xq	127	32767	0	0
    zp	127	32767	0	0

  • scale = xf / (xq - zp) if xq != zp else 1

Motivation and Context

• Negative or zero scale value is not friendly for various EP and backend such as QNN-EP.
• Inserting an additional DQ is only useful when it forms a valid QDQ “node unit” pattern. If the Where output is not followed by a single QuantizeLinear (e.g., multiple consumers or a non-Q consumer), adding a dummy DQ cannot create the intended pattern and may lead to non-fusible/undesired graph structures.

[qti-hungjuiw]

Hi @edgchen1, this fix is for a critical model. Could you help trigger the CI and review the PR?

[Copilot]

Pull request overview

Updates the WhereDummyDq QDQ transformer to be more selective about when it inserts a dummy DequantizeLinear on Where scalar inputs, and adjusts the dummy quantization parameters to avoid non-positive scales (improving compatibility with EPs/backends).

Changes:

• Tighten WhereDummyDq::SatisfyCondition to only apply when Where is immediately followed by a single QuantizeLinear.
• Update dummy (xq, zp, scale) construction so computed scale is positive by mapping the float scalar to integer-domain boundaries.
• Fix the unit test to use the correct zero-point type for the downstream QuantizeLinear.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
onnxruntime/test/optimizer/qdq_transformer_test.cc	Adjusts WhereDummyDq test model to use ZpType for QuantizeLinear zero-point so types match the test’s quantized dtype.
onnxruntime/core/optimizer/qdq_transformer/where_dummy_dq.h	Expands class documentation to describe the dummy quantization parameter selection and scale computation.
onnxruntime/core/optimizer/qdq_transformer/where_dummy_dq.cc	Implements stricter pattern checks and new dummy scale/zp/xq selection to keep dummy scale positive.

---

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

[adrianlizarraga]

Hi @qti-hungjuiw,

I think there are some merge conflicts with main. May have to sync your branch with latest main.

Also, there are still a couple copilot review comments that are unresolved. Could you please take a look at those? It looks like some of those may be valid.

Thank you for fixing this!

[qti-hungjuiw]

Hi @qti-hungjuiw,

I think there are some merge conflicts with main. May have to sync your branch with latest main.

Also, there are still a couple copilot review comments that are unresolved. Could you please take a look at those? It looks like some of those may be valid.

Thank you for fixing this!

Hi @adrianlizarraga, sure, I've resolved the conflict and addressed copilot's comments. Please let me know if there are any issues.

[edgchen1]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).