EGGROLL

YouTube

• No backprop

• Non differentiable architectures

• No need for sync (rng state)

• Can train on inference optimized wuantized versions of the model

• Use on prod inference workoads