We'd like levanter to be roughly on par with the best torch implementations on GPU.

This may or may not be doable quickly, but we'll see