compute attention
reweight context
then predict token