FLASHSWIN: Unlocking Large Windows and Dense Tokens in Swin Vision Transformers with Memory Efficient Attention

Abstract

High-resolution vision backbones have long been forced to trade away local token density to afford larger receptive fields. Hierarchical Swin transformers impose this compromise because standard windowed attention materializes an $M^2\times M^2$ score matrix per window, incurring $O(M^4)$ memory as windows or token grids grow. Furthermore, Swin adds a learned relative-position bias elementwise to attention scores, requiring full materialization of the score matrix and its gradient. This keeps Swin and SwinV2 trapped in a small-window($M=8,16$), coarse-token regime with patch size $4\times4$ ($p=4$), limiting performance for fine-grained tasks. We introduce FLASHSWIN, which replaces standard windowed attention with a FlashAttention implementation that computes exact softmax attention without materializing the score matrix, reducing per-window memory from $O(M^4)$ to $O(M^2)$. This enables higher token density and larger receptive fields without inflating memory overhead. Training memory is flat across window sizes: at a $32\times32$ window, FLASHSWIN-T requires only $12.4$\,GB, unchanged from $8\times8$, compared to $70/90$\,GB for SwinV2/V1-T. However, applying FlashAttention directly to Swin creates a trade-off: bypassing the score matrix precludes Swin's additive relative-position bias, forfeiting spatial information in exchange for memory efficiency. FLASHSWIN restores position information as window-local learnable 2D RoPE, making large windows and dense token grids both affordable and accurate. At matched scale, FLASHSWIN-T outperforms Swin variants. With dense tokens and wide windows ($p=2,M=32$), the same Tiny model reaches $84.1\%$ ImageNet-1K, $44.1$ COCO box AP, and $47.28$ ADE20K mIoU---gains of $+1.3$, $+5.1$, and $+1.82$ over SwinV2-T at $M=16$, respectively. At fixed $M=32$, halving the patch size yields roughly $3\times$ larger gains in boundary quality than in mIoU.