I’m general, most Deep Learning processes these days are non-deterministic because 1) they only care about statistical correctness 2) there are some speed advantages in ignoring the existence of race condition bugs if you don’t care about being deterministic