#sliding-window

Full attention is O(n²). Sliding window attention is O(n). The trade: lose long-range dependencies, gain linear scaling. Often worth it.