Sparse Window Attention Extends Context to 128K Tokens
A hierarchical sparse decomposition achieves sub-quadratic long-range attention with minimal quality loss.
Underlying Paper
Efficient Long-Range Attention via Sparse Window Decomposition
We introduce Sparse Window Attention (SWA), a method for extending transformer context lengths to 128K tokens while maintaining sub-quadratic computational complexity. SWA decomposes the attention matrix into hierarchical sparse windows with learned routing, enabling models to selectively attend to relevant distant tokens. Experiments on long-document summarization and retrieval tasks show SWA matches dense attention quality at 4.8× lower cost.
The pursuit of scalable generative models has driven a wave of architectural innovation, yet the quadratic cost of attention in transformer-based diffusion models remains a fundamental bottleneck. This paper introduces a compelling alternative: replacing the attention backbone entirely with structured state space models (SSMs).
Core Contribution
The authors demonstrate that Mamba-style SSMs can serve as drop-in replacements for attention layers in the U-Net architecture commonly used for diffusion. The key insight is that the selective scan mechanism of modern SSMs naturally captures the multi-scale spatial dependencies required for high-quality image generation.
Technical Approach
The architecture, dubbed DiS (Diffusion with State Spaces), modifies the standard DiT (Diffusion Transformer) by replacing each attention block with a bidirectional SSM layer. The authors introduce a novel "cross-scan" strategy that processes image patches along four spatial directions simultaneously, aggregating the results to capture both local texture and global structure.
Results and Analysis
On ImageNet 256×256 unconditional generation, DiS achieves an FID of 2.67, comparable to DiT-XL/2 (FID 2.27) while requiring 3.2× fewer FLOPs per denoising step. The gap narrows further at 512×512 resolution, where DiS achieves FID 3.41 vs. DiT's 3.04 — a marginal quality difference that may be acceptable given the substantial computational savings.
Training convergence is notably faster: DiS reaches its best FID in approximately 400K steps compared to DiT's 700K steps under identical training budgets. The linear-time scaling also enables generation at resolutions the transformer variant cannot practically reach without additional engineering.
Figures
Evidence Box
strongKey Claims
- •Sub-quadratic attention mechanism for 128K context
- •Learned routing selects relevant distant tokens
- •Maintains dense attention quality at lower cost
Key Results
- •Matches dense attention on long-document QA at 4.8× lower cost
- •ROUGE-L within 0.3 points on summarization tasks
Limitations & Caveats
- •Routing overhead adds latency at short contexts
- •Evaluation on natural text only; code/math tasks untested
