How a transformer actually predicts the next token

A from-scratch walk through attention, KV caches, and sampling — the parts that matter when you're trying to make one fast in production.