Inference

A language model does not output text. At each step it outputs one score per vocabulary token, and a separate piece of code, the sampler, turns those scores int…