[sailingparrot]

Indeed what I meant. The LLM isn’t a blank slate at the beginning of each new token during autoregression as the kv cache is there.