Prism: Spectral-Aware Block-Sparse Attention
This work addresses efficiency issues in long-context LLM pre-filling for AI researchers and practitioners, offering a novel solution to a known bottleneck.
The paper tackled the bottleneck of efficiently identifying relevant blocks in block-sparse attention for long-context LLM pre-filling by addressing inaccuracies from mean pooling with RoPE, resulting in a method that maintains accuracy parity with full attention while achieving up to 5.1× speedup.
Block-sparse attention is promising for accelerating long-context LLM pre-filling, yet identifying relevant blocks efficiently remains a bottleneck. Existing methods typically employ coarse-grained attention as a proxy for block importance estimation, but often resort to expensive token-level searching or scoring, resulting in significant selection overhead. In this work, we trace the inaccuracy of standard coarse-grained attention via mean pooling to a theoretical root cause: the interaction between mean pooling and Rotary Positional Embeddings (RoPE). We prove that mean pooling acts as a low-pass filter that induces destructive interference in high-frequency dimensions, effectively creating a "blind spot" for local positional information (e.g., slash patterns). To address this, we introduce Prism, a training-free spectral-aware approach that decomposes block selection into high-frequency and low-frequency branches. By applying energy-based temperature calibration, Prism restores the attenuated positional signals directly from pooled representations, enabling block importance estimation using purely block-level operations, thereby improving efficiency. Extensive evaluations confirm that Prism maintains accuracy parity with full attention while delivering up to $\mathbf{5.1\times}$ speedup.