Fix int32 bias adjust for Gemm with 2D bias

[mcollinswisc]

Description

The weight scale adjust for int32 bias currently assumes that the bias is 1D. Gemm nodes may have a 2D bias: the only requirement is that it be broadcastable to the output shape:

https://onnx.ai/onnx/operators/onnx__Gemm.html

This change adds logic that gets the largest bias element corresponding to each scale in the per-channel quantization, and computes the adjustment based on that. It also fixes the axis chosen for per-channel quantization of the bias. This allows it to support Gemm's 2D bias.

Along with the logic, this change adds tests that cover the weight scale adjust for int32 bias for Gemm nodes with 2D bias, previously only Conv was covered. It also factors out a common helper for building the weights in these unit tests.

Motivation and Context

This weight scale adjust was added in: 4f6993d5 and 5e4d8dc3.

However, this new logic breaks and raises an exception when given a Gemm that has bias of shape (1, N) instead of (N,). The 2D shape can arise from the preprocessing that fuses MatMul and Add nodes (since the Add has explicit broadcasting).

#24815

[azure-pipelines]

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

[Copilot]

Pull request overview

This PR fixes int32-bias weight-scale adjustment logic in the Python quantization tooling so it correctly handles Gemm nodes whose bias input is 2D (e.g. (1, N)), which can arise after MatMul+Add fusion and previously triggered exceptions.

Changes:

• Add bias_abs_max_per_channel() helper to compute per-output-channel bias magnitudes for broadcastable (incl. 2D) Gemm biases, and use it in both QDQ and QOperator weight-scale-adjust paths.
• Fix the per-channel bias dequantization axis selection to use the bias tensor’s last axis (supporting 2D Gemm biases).
• Add unit tests covering Gemm with 2D bias for both QDQ and QOperator formats, and factor out a shared tiny-weight initializer builder for tests.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
onnxruntime/python/tools/quantization/quant_utils.py	Adds bias_abs_max_per_channel() helper for per-channel bias magnitude reduction across broadcastable Gemm bias shapes.
onnxruntime/python/tools/quantization/qdq_quantizer.py	Uses the new bias reduction helper during weight-scale adjustment; sets bias dequant axis to last dim for per-channel bias.
onnxruntime/python/tools/quantization/onnx_quantizer.py	Mirrors the QDQ changes for the non-QDQ quantizer path (weight-scale adjustment + last-axis bias dequant).
onnxruntime/test/python/quantization/op_test_utils.py	Adds build_tiny_weights() helper to build tiny-magnitude weights used by multiple tests.
onnxruntime/test/python/quantization/test_qdq.py	Adds Gemm (1, N) bias test coverage for QDQ weight-scale adjustment and uses build_tiny_weights().
onnxruntime/test/python/quantization/test_qoperator_adjust_int32_bias.py	Adds Gemm (1, N) bias test coverage for QOperator weight-scale adjustment and uses build_tiny_weights().