Transformers and LLMs

Sliding-Window Attention, explained

Sliding-window attention restricts each token to attend only to a nearby window of tokens.
Back to Glossary

Part of the Semantic Notion technical glossary.

Short Definition

Sliding-window attention restricts each token to attend only to a nearby window of tokens.

Intuition

Most local language structure is nearby, so a model can save compute by not attending globally in every layer.

Technical Definition

Sliding-window attention masks attention to a fixed local neighborhood, reducing attention cost relative to full sequence attention.

Example

A long-context model may use local attention in most layers and global attention in selected layers.

Common Misunderstandings

Sliding-window attention can miss distant dependencies unless combined with other mechanisms.
It is an efficiency tradeoff, not a universal replacement for global attention.

Same Category

More from Transformers and LLMs

AttentionAttention is a neural network mechanism that lets a model decide which parts of an input are most relevant when producing a representation or output.Context WindowA context window is the maximum amount of input and generated tokens a model can consider at one time.Cross-AttentionCross-attention lets one sequence attend to information from a different sequence or modality.Hybrid AttentionHybrid attention combines different attention patterns, such as local and global attention, inside one model.KV CacheA KV cache stores attention keys and values from previous tokens so autoregressive generation does not recompute them every step.Multimodal ModelA multimodal model can process or generate more than one kind of data, such as text, images, audio, or video.

Related Reading

Essays that use this concept

Transformers and LLMsInkling: What Thinking Machines Actually Shipped

Start here

Need the broader concept map?

Return to the glossary index to move from this definition into adjacent AI, ML, mathematics, and computer science terms.

Browse Glossary