3 ms·

Completely irrelevant since the model is run on a custom chip. So no part of the pytorch code actually ever makes it into production.