[Conformance] Fix input shape generation for Broadcast (f32)

[TempestBirds729804]

Fix Broadcast operation conformance test failure (f32)

Description

In this PR, I fixed a conformance test failure in the Broadcast operation (opset3, type=f32), where the input tensor shape did not meet NumPy broadcasting rules and caused shape inference errors.

In the original test scenario, input[0] was a data tensor with incompatible shape (e.g., [37, 15, 15]), and input[1] : the target shape was a Constant (e.g., [37, 37, 128]). The generated input[0] could not be broadcast correctly, leading to inference failure in the TEMPLATE plugin.

Solution

A custom input shape adjustment logic has been added in generate_inputs.cpp. Specifically:

• If the Broadcast operation uses NumPy mode,
• And the output shape is static,
• The generator adjusts input[0] by replacing incompatible dimensions with 1 (when they do not match the target shape or output shape), ensuring compliance with NumPy broadcasting rules.

Test Result

The Broadcast.3_Type=f32 conformance test now passes successfully.

However, the Broadcast.3_Type=i64 test case still fails. Based on the analysis, the cause is different:

• In this case, input[1] (target shape) is a Parameter instead of a Constant.
• The test framework auto-generates values for this Parameter, which may not conform to broadcasting rules.
• As the i64 failure is related to the framework’s input generation for input[1], and not input[0], this PR focuses solely on fixing the f32 case.

Note: I'm happy to follow up with a separate fix for the i64 case if needed (e.g., by improving test model generation or adjusting framework behavior). Feedback and suggestions are welcome! 🙂

Related Issue

Fixes #23549

[p-wysocki]

Thanks for the PR!

bump @iefode

[TempestBirds729804]

Thanks for the PR!

bump @iefode
U welcome! Have a nice day bro🙂