Validate FastGelu fusion scale node

[Akshay Sonawane (apsonawane)]

This pull request improves the robustness of the FastGelu fusion optimization by ensuring malformed nodes are properly skipped and adds a test to verify this behavior. The main changes include stricter input validation in the fusion logic and a new unit test.

Fusion logic improvements:

• Added explicit checks for the number of inputs (InputDefs().size()) in Mul and Pow nodes within the FastGeluFusion optimizer to ensure only well-formed nodes are considered for fusion. [1][2][3]

Testing enhancements:

• Introduced a new test, FastGeluFusionSkipsMalformedScaleMul, that modifies a model to create a malformed Mul node and verifies that the fusion optimizer correctly skips it (i.e., does not produce a FastGelu node).

[Copilot]

Pull request overview

This PR hardens the FastGeluFusion graph rewrite by adding input-count validation so malformed Mul/Pow nodes are skipped instead of being processed, and adds a regression test to ensure the fusion does not produce a com.microsoft.FastGelu node when the scale Mul is malformed.

Changes:

• Add InputDefs().size() == 2 / != 2 checks for key Mul and Pow nodes used by the FastGelu pattern matcher.
• Fix the first-formula scale-mul validation to actually validate the upstream Mul node being inspected.
• Add a unit test that mutates the FastGelu test model to create a malformed scale Mul and verifies fusion is skipped.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
onnxruntime/core/optimizer/fast_gelu_fusion.cc	Adds stricter input-count validation in FastGelu pattern matching to avoid processing malformed nodes.
onnxruntime/test/optimizer/graph_transform_test.cc	Adds a regression test that forces a malformed scale Mul and checks the fusion optimizer does not emit com.microsoft.FastGelu.

Suppressed comments (1)

onnxruntime/core/optimizer/fast_gelu_fusion.cc:122

• Similar to the pow1_node.InputDefs().size() != 2 guard, the rest of CheckSecondFormula assumes intermediate nodes have 2 inputs and that IndexOfNodeInput returns a valid index before using [(input_index + 1) % 2] (e.g., the checks around lines 135-138 and 174-177). For malformed graphs this can still access out-of-range inputs or match the wrong operand when IndexOfNodeInput returns -1. Consider adding InputDefs().size() == 2 and input_index >= 0 guards before all modulo-based input indexing so the fusion always fails safely.

 if (!graph_utils::IsSupportedOptypeVersionAndDomain(pow1_node, "Pow", {7, 12, 13, 15}) ||
pow1_node.InputDefs().size() != 2 ||
!graph_utils::IsSupportedProvider(pow1_node, GetCompatibleExecutionProviders()) ||
pow1_node.GetOutputEdgesCount() != 1 ||
!IsSupportedDataType(pow1_node)) {

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.