[Rewriter]: Add ∘ MatMul -> Gemm

[AyoubMDL]

A Rewriter rule that transforms MatMul(Add) to Gemm.

[AyoubMDL]

@justinchuby I wrote the rewriter using functions (as in gemm_to_matmul_add). Tell me if you want me to do it using Classes.

[Copilot]

Pull Request Overview

This PR introduces a rewriter rule that transforms a MatMul followed by an Add operation into a Gemm operator for streamlined execution. The changes include implementation of the transformation rule in the rewriter module and accompanying tests to validate both standard and edge-case behaviors.

• Added a new transformation rule in onnxscript/rewriter/matmul_add_to_gemm.py.
• Introduced tests in onnxscript/rewriter/matmul_add_to_gemm_test.py for regular inputs, initializers inclusion, and incompatible shapes.

Reviewed Changes

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

File	Description
onnxscript/rewriter/matmul_add_to_gemm_test.py	Test cases ensuring the fusion of MatMul and Add into Gemm are correct.
onnxscript/rewriter/matmul_add_to_gemm.py	Implementation of the rewriter rule for converting MatMul+Add to Gemm.

Comments suppressed due to low confidence (1)

onnxscript/rewriter/matmul_add_to_gemm_test.py:12

• [nitpick] The test class name 'MatMulAddToMatMulTest' is misleading given that the transformation targets Gemm. Consider renaming it to 'MatMulAddToGemmTest' for clarity.

class MatMulAddToMatMulTest(unittest.TestCase):

[codecov]

Codecov Report

Attention: Patch coverage is 98.80240% with 2 lines in your changes missing coverage. Please review.

Project coverage is 70.36%. Comparing base (321cb41) to head (9b5bec4).
Report is 4 commits behind head on main.

Files with missing lines	Patch %	Lines
onnxscript/rewriter/matmul_add_to_gemm_test.py	98.36%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2356      +/-   ##
==========================================
+ Coverage   70.15%   70.36%   +0.20%     
==========================================
  Files         197      199       +2     
  Lines       24985    25152     +167     
  Branches     2669     2681      +12     
==========================================
+ Hits        17529    17698     +169     
+ Misses       6529     6528       -1     
+ Partials      927      926       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[justinchuby]

@justinchuby I wrote the rewriter using functions (as in gemm_to_matmul_add). Tell me if you want me to do it using Classes.

Would be nice to use the class based approach just for consistency, thanks!

[AyoubMDL]

Added the support for transposed inputs

I think it’s possible to detect the number of users of the weight. Sounds like if Transpose isn’t the only consumer of the weight value then we can skip the rule.

I think I dont need to check the users, the rewriter does check the number of users by default (if i am not mistaken)

You can allow the rule to permute input orders

With the current implementation, I don't think I need it, right ?

[justinchuby]

Nicely written and tested, thanks!

[justinchuby]

@titaiwangms could you approve? Thanks