ReduceFormer: Attention with Tensor Reduction by Summation
This work addresses the problem of deploying transformer models efficiently in low-latency or high-throughput applications, such as on edge devices or in cloud computing, though it is incremental in its approach.
The authors tackled the computational inefficiency of attention mechanisms in Transformers by introducing ReduceFormer, which uses only reduction and element-wise operations, achieving up to 37% lower latency and 44% higher throughput while maintaining competitive accuracy.
Transformers have excelled in many tasks including vision. However, efficient deployment of transformer models in low-latency or high-throughput applications is hindered by the computation in the attention mechanism which involves expensive operations such as matrix multiplication and Softmax. To address this, we introduce ReduceFormer, a family of models optimized for efficiency with the spirit of attention. ReduceFormer leverages only simple operations such as reduction and element-wise multiplication, leading to greatly simplified architecture and improved inference performance, with up to 37% reduction in latency and 44% improvement in throughput, while maintaining competitive accuracy comparable to other recent methods. The proposed model family is suitable for edge devices where compute resource and memory bandwidth are limited, as well as for cloud computing where high throughput is sought after.