Mixture-of-Depths: Dynamically allocating compute in transformer-based language models

MoE
Re-assign community