SADMoE: Exploiting Activation Sparsity with Dynamic-k Gating

Transformer models, despite their impressive performance, often face practical limitations due to their high computational requirements. At the same time, such models exhibit significant activation sparsity, which can be leveraged to reduce the inference cost by transforming parts of the network into Mixture-of-Experts (MoE) layers. However, despite the crucial role of activation sparsity, its impact on this process remains unexplored. In this paper, we enhance the efficiency of MoE conversion through activation sparsity enforcement. Moreover, motivated by the high variance in the number of activated neurons, we propose a more effective dynamic-k expert selection rule that adjusts the number of executed experts on a per-token basis. Finally, we extend this approach to multi-head attention projections, which results in even further savings. The proposed method, Sparsified Activation Dynamic-k Mixture-of-Experts (SADMoE), outperforms existing approaches on common NLP and vision tasks, allowing us to save up to 60% of inference cost without significantly affecting model performance.