inference code is effectively trivial to port at this time

everyone understands cuda well enough anyway