Comment by mruts

PyTorch wraps THNN, not Torch. Moreover, even if this was true, it wouldn't matter at all. Practically 0% of the overhead is related to Python in the first place, all of the time is dominated by the underlying C implementation.