tokenize input

manage context length

understand attention cost

choose decoding strategy

optimize inference

control output quality