“For models with dense attention, memory fetch time scales linearly with context length.”