AI Inference & Serving
Sliding-Window Attention
An attention pattern that limits each position's direct attention to a bounded neighborhood, often a recent causal window. The exact inclusion rules depend on the architecture. Information may propagate across layers, so a direct window is neither a promise of full-history recall nor necessarily the model's entire effective receptive field.
Reviewed
Sources
Member lesson
The definition and sources are public. The complete practical lesson is for members.