tokenize prompt

run decoder

get logits from LM Head

apply temperature

filter with top-k or top-p

sample or choose token

append token

repeat