3 ms·

inference code is effectively trivial to port at this time everyone understands cuda well enough anyway