Zubnet AILearnWiki › Sparse Attention
Models

Sparse Attention

Local Attention, Sliding Window Attention
Attention mechanisms that process only a subset of token pairs instead of the full N×N attention matrix. Sliding window attention attends to only nearby tokens (within a fixed window). Sparse patterns (like Longformer's combination of local + global attention) let specific tokens attend to everything while most tokens attend locally. These approaches reduce attention's quadratic cost for long sequences.

Why it matters

Sparse attention is how Mistral, Mixtral, and other efficient models handle long sequences without the full cost of dense attention. It's the practical compromise between "attend to everything" (expensive but thorough) and "attend to nothing distant" (cheap but limited). Understanding sparse attention helps you evaluate claims about context length and predict where quality degradation might occur.

Deep Dive

Sliding window attention: each token attends only to tokens within a fixed window (e.g., 4096 tokens). Information from earlier tokens propagates through the layers — layer 1 sees 4096 tokens, layer 2 effectively sees 8192 (two windows worth), and by the final layer, information from the full sequence has had a chance to propagate. Mistral-7B uses a 4096-token sliding window across its 32 layers.

Hybrid Patterns

Longformer combines sliding window (local) attention with global attention on selected tokens (like [CLS] or user-defined positions). BigBird adds random attention connections on top of local and global patterns. These hybrid approaches let models handle 4K–16K tokens with subquadratic cost while maintaining the ability to connect distant tokens through global positions.

The Quality Trade-off

Sparse attention works well for many tasks but can degrade on tasks requiring precise long-range dependencies — referencing a specific fact from the beginning of a long document, maintaining consistency across a long conversation, or following complex instructions that span many tokens. Dense attention (full quadratic) with Flash Attention remains more robust for these cases, which is why most frontier models still use dense attention and rely on Flash Attention for efficiency rather than sparsity.

Related Concepts

← All Terms
← Softmax Sparse Autoencoder →