HSMLA: Hierarchical Softmax Multi-scale Linear Attention for Efficient Vision Transformers figure
AlphaXiv 中文概览(可滚动查看)