@lanluo-nvidia @narendasan

From the CI logs, it seems the issue might be caused by the change I made to generate integer random inputs for dynamic shapes. Previously, the inputs were all zeros, so the problem didn’t surface. I’ll look into fixing this part.

If strong type support for atan2 is urgent, I can remove the dynamic shape input change from this PR and submit it separately. Let me know what you think.

self = Input(shape=(2, 3, 4), dtype=dtype.b, format=memory_format.linear, domain=[0.0, 2.0))
optimization_profile_field = None

def example_tensor(
    self, optimization_profile_field: Optional[str] = None
) -> torch.Tensor:
    """
    Get an example tensor of the shape specified by the Input object

    Args:
        optimization_profile_field (Optional(str)): Name of the field to use for shape in the case the Input is dynamically shaped

    Returns:
        A PyTorch Tensor
    """
    if self.shape_mode == Input._ShapeMode.STATIC:
        if optimization_profile_field is not None:
            raise ValueError(
                "Specified a optimization profile field but the input is static"
            )
        else:
            if isinstance(self.shape, tuple):
                shape = self.shape
                dtype = self.dtype.to(torch.dtype, use_default=True)
                if dtype.is_floating_point:
                    return torch.rand(shape).to(dtype=dtype)
                else:
                    # For integer types, use randint to get a better range of values for testing

                  return torch.randint(-10, 10, shape, dtype=dtype)

E torch._dynamo.exc.BackendCompilerFailed: backend='torch_tensorrt_backend' raised:
E RuntimeError: from is out of bounds for bool
E
E Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"
E
E
E To execute this test, run the following from the base repo dir:
E python test_decompositions.py TestLowering.test_masked_scatter_1_float16_3d
E
E This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

/opt/python/cp39-cp39/lib/python3.9/site-packages/torch_tensorrt/Input.py:389: BackendCompilerFailed
------------------------------ Captured log call -------------------------------
CRITICAL torch_tensorrt.dynamo.backend.backends:backends.py:180 Halting compilation on build failure since pass_through_build_failures was specified as True. To return the default Torch implementation and avoid halting compilation on engine build failures, specify pass_through_build_failures=False.
=============================== warnings summary ===============================
tests/py/dynamo/lowering/test_decompositions.py: 68 warnings
tests/py/dynamo/lowering/test_aten_lowering_passes.py: 3 warnings
/opt/python/cp39-cp39/lib/python3.9/site-packages/torch_tensorrt/dynamo/utils.py:275: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.detach().clone() or sourceTensor.detach().clone().requires_grad(True), rather than torch.tensor(sourceTensor).
torch.tensor(inputs),

tests/py/dynamo/lowering/test_decompositions.py: 68 warnings
tests/py/dynamo/lowering/test_aten_lowering_passes.py: 3 warnings
/opt/python/cp39-cp39/lib/python3.9/site-packages/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py:976: DeprecationWarning: Use Deprecated in TensorRT 10.12. Superseded by strong typing. instead.
output.dtype = output_dtype.to(trt.DataType, use_default=True)

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html

• generated xml file: /tmp/test_results/tests_py_dynamo_core_lowering_test_results.xml -
  =========================== short test summary info ============================
  FAILED lowering/test_decompositions.py::TestLowering::test_masked_scatter_0_float32_2d - torch._dynamo.exc.BackendCompilerFailed: backend='torch_tensorrt_backend' raised:
  RuntimeError: from is out of bounds for bool

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

To execute this test, run the following from the base repo dir:
python test_decompositions.py TestLowering.test_masked_scatter_0_float32_2d

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
FAILED lowering/test_decompositions.py::TestLowering::test_masked_scatter_1_float16_3d - torch._dynamo.exc.BackendCompilerFailed: backend='torch_tensorrt_backend' raised:
RuntimeError: from is out of bounds for bool

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

To execute this test, run the following from the base repo dir:
python test_decompositions.py TestLowering.test_masked_scatter_1_float16_3d

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
================= 2 failed, 77 passed, 142 warnings in 59.93s ==================

I will create a separate issue for py/torch_tensorrt/_Input.py as revising the masked_scatter lowering and converter will take more time.