Fix GemmTransposeFusion handling of identity Transpose

[Sylvester Kaczmarek (sylvesterkaczmarek)]

Description

Fixes #32228.

GemmTransposeFusion currently treats any Transpose adjacent to Gemm as a matrix transpose and toggles transA or transB without checking the perm attribute. For an identity transpose (perm=[0,1]), removing the node and toggling the Gemm flag changes the computation and can silently produce incorrect results.

This change only folds Transpose nodes that actually swap the two Gemm matrix axes ([1,0], including the rank-2 default permutation). The check is applied to input A, input B, and output Transpose paths.

Tests

Adds regression coverage showing that perm=[0,1] is left intact with transA unchanged, while perm=[1,0] continues to fuse.

[azure-pipelines]

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

[Copilot]

🔵 Needs a closer look

The newly guarded input-B and output-Transpose paths need regression coverage.

Pull request overview

Fixes incorrect GemmTransposeFusion handling of identity transposes.

Changes:

• Restricts fusion to matrix transposes using [1, 0] or the rank-2 default permutation.
• Adds input-A regression coverage.

File summaries

File	Review
onnxruntime/test/optimizer/gemm_transpose_fusion_test.cc	Tests input-A permutations; add identity-permutation coverage for input B and output Transpose paths.
onnxruntime/core/optimizer/gemm_transpose_fusion.cc	Validates permutations across all Gemm paths; add identity and [1, 0] tests for input B and output Transpose behavior.

Review details

Suppressed comments (3)

onnxruntime/core/optimizer/gemm_transpose_fusion.cc:63

• The regression tests only place the Transpose on input A, so the newly guarded input-B path is not exercised. A future omission or regression of this IsMatrixTranspose check would still pass both new tests; add corresponding identity and [1, 0] cases with the Transpose feeding Gemm input 1 and assert transB.

  if (B_node_ptr != nullptr && B_node_ptr->OpType() == "Transpose" && IsMatrixTranspose(*B_node_ptr)) {

onnxruntime/core/optimizer/gemm_transpose_fusion.cc:84

• The output-Transpose permutation check is also new behavior, but the added tests cover only an input Transpose. Add output-side identity and [1, 0] cases so this path verifies that identity remains and a real matrix transpose still fuses with the expected swapped inputs/flags.

      output_node_ptr->OpType() == "Transpose" &&
      IsMatrixTranspose(*output_node_ptr)) {

onnxruntime/test/optimizer/gemm_transpose_fusion_test.cc:72

• These regressions exercise the new permutation check only for input A. The same predicate now gates input B and the output-Transpose rewrite, but neither path verifies that perm={0, 1} is preserved; a future omission or branch-specific regression there would still pass. Please add identity-permutation cases for B and the output path (the existing graph-transform tests cover their successful matrix-transpose fusion).

TEST(GemmTransposeFusionTest, IdentityInputTransposeIsNotFolded) {
  RunInputTransposeTest({0, 1}, {2, 3}, false, 0);
}

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Xavier Dupré (xadupre)]

duplicate of #32429

[Sylvester Kaczmarek (sylvesterkaczmarek)]

Closing as a duplicate of #32429.