Can inf be accepted as a GeLU solution to FLT_MAX input?

[eralmual]

Hi!

We wanted to update the oneDNN EP to the latest version but we noticed errors on the GeLU tests when the input is FLT_MAX, ORT expects the output to be FLT_MAX, while oneDNN outputs inf.

Is there any chance for the tests to be updated and inf accepted as a valid response for the oneDNN EP?

Below is an example of the error log

D:\onnxruntime\oneDNN-v3.2\onnxruntime\test\providers\checkers.cc(272): error: The difference between cur_expected[i] and cur_actual[i] is inf, which exceeds *(params.relative_error) * std::abs(cur_expected[i]), where
cur_expected[i] evaluates to 3.4028234663852886e+38,
cur_actual[i] evaluates to inf, and
*(params.relative_error) * std::abs(cur_expected[i]) evaluates to 3.4028232737618185e+32.
i:10