For writing your own kernels, sure, but a PyTorch built-in is almost always going to be faster than an implementation consisting of a series of simpler PyTorch functions.

→ View original post on X — @id_aa_carmack