pytorch - 💡(How to fix) Fix AArch64 Unit Test Failure - inductor/test_aot_inductor_arrayref.py test_deconv_freezing_cpu_with_stack_allocation [2 comments, 3 participants]

Traceback (most recent call last): File "/builds/software-machine-learning-infra-frameworks-workspaces-robhar02/pytorch/test/inductor/test_aot_inductor.py", line 910, in test_deconv_freezing self.check_model(Model(self.device), example_inputs) File "/builds/software-machine-learning-infra-frameworks-workspaces-robhar02/pytorch/test/inductor/test_aot_inductor_utils.py", line 259, in check_model self.assertEqual(actual, expected, atol=atol, rtol=rtol) File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual return super().assertEqual(x, y, *args, **kwargs) File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 4365, in assertEqual raise error_metas.pop()[0].to_error( # type: ignore[index] AssertionError: Tensor-likes are not close!

Mismatched elements: 9 / 32 (28.1%) Greatest absolute difference: 4.3125 at index (0, 1, 1, 1) (up to 1e-05 allowed) Greatest relative difference: 3.390625 at index (0, 1, 3, 0) (up to 0.016 allowed)

To execute this test, run the following from the base repo dir: python test/inductor/test_aot_inductor_arrayref.py AOTInductorTestABICompatibleCpuWithStackAllocation.test_deconv_freezing_cpu_with_stack_allocation

Fix Plan

The fix involves updating the tolerance values in the assertEqual method to account for the differences in floating-point calculations on the Neoverse-V1 architecture.

• Update the atol and rtol values in the check_model function:

self.assertEqual(actual, expected, atol=1e-4, rtol=1e-3)

• Alternatively, consider using a more robust comparison method, such as torch.allclose:

self.assertTrue(torch.allclose(actual, expected, atol=1e-4, rtol=1e-3))

Verification

To verify the fix, re-run the test using the updated tolerance values or comparison method:

python test/inductor/test_aot_inductor_arrayref.py AOTInductorTestABICompatibleCpuWithStackAllocation.test_deconv_freezing_cpu_with_stack_allocation

If the test passes, it indicates that the updated tolerance values or comparison method have resolved the issue.

Extra Tips

• When working with floating-point calculations, it's essential to consider the architecture-specific differences in rounding and precision.
• Using more robust comparison methods, such as torch.allclose, can help mitigate issues related to floating-point precision.