PyTorch removes torch.testing.assert_allclose in breaking change

The PyTorch project has announced the removal of the torch.testing.assert_allclose function in favor of torch.testing.assert_close, as detailed in the project's GitHub release notes. This change is backward-incompatible and requires developers to update their code. While the two functions share overlapping keyword arguments and similar functionality, there are notable differences in default behavior. Specifically, torch.testing.assert_allclose had check_dtype=False and eq_nan=True as defaults, whereas torch.testing.assert_close defaults to check_dtype=True and eq_nan=False. Developers must account for these differences when transitioning to the new function. For example, replacing torch.testing.assert_allclose(a, b) with torch.testing.assert_close(a, b, check_dtype=False) would preserve the original behavior, or simply torch.testing.assert_close(a, b) to enable dtype checking.

Operators should review any testing scripts or frameworks that rely on torch.testing.assert_allclose and verify whether dtype checking or NaN equivalence is critical in their workflows. This change could impact test results or introduce subtle issues if the default behaviors are not explicitly handled during migration.