> you don't re-compute all attention passes for all tokens for each next token to predict. You don't? I imagine the attention maps could be pretty different be

> you don't re-compute all attention passes for all tokens for each next token to predict. You don't? I imagine the attention maps could be pretty different between n and n+1 tokens. Edit: Or maybe you just meant you don't compute attention Σ(n) times for each new token?