How do you handle long-term dependencies in sequence generation with transformer-based models

Question

Are there strategies or ways to handle long-term dependencies in sequence generation with transformer-based models?