A Unified Sparse Attention via Multi-Granularity Compression

2025-12-15

AI Summary

To address the inference bottleneck caused by the quadratic computational complexity $O(n^2)$ of self-attention in long-context scenarios, this paper proposes UniSparse, a unified sparse attention mechanism. Methodologically, UniSparse introduces three key innovations: (1) a novel composite token abstraction enabling training-agnostic, plug-and-play cross-modal sparsification; (2) multi-granularity context compression coupled with dynamic block-level sparse selection to yield hardware-friendly sparsity patterns; and (3) GPU-optimized sparse kernels for efficient execution. Empirically, UniSparse achieves ≥99% accuracy relative to dense attention while accelerating inference by 2.61× over FlashAttention. It significantly outperforms state-of-the-art methods—including MInference and XAttention—across diverse long-context tasks such as multi-turn dialogue and program analysis, demonstrating strong generalization and practical utility.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Computer Vision: Multi-modal Vision

Application Category

• Search and Retrieval-Augmented AI: Personalized, context-aware and across-device search
• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Efficient long-context understanding and reasoning are increasingly vital for large language model (LLM) applications such as multi-turn dialogue and program analysis. However, the core self-attention mechanism scales quadratically with sequence length, creating a fundamental computational bottleneck. Existing sparse attention methods alleviate this issue but face trade-offs: training-based methods are costly and cannot be directly applied as acceleration plugins for other models, while inference-time methods often compromise efficiency or cross-modal generality. To address these limitations, we present UniSparse, a unified mechanism that introduces the notion of composite tokens--compact representations that aggregate multi-granularity contextual information. Building on this abstraction, UniSparse dynamically constructs sparse attention through multi-granularity compression and block-level selection, enabling efficient and hardware-friendly execution on GPU. Across multiple modalities and tasks ranging from synthetic benchmarks to real-world applications, UniSparse consistently surpasses state-of-the-art sparse attention methods (e.g., MInference, XAttention, FlexPrefill) in both accuracy and efficiency, achieving $ge$ 99% of full-attention accuracy and up to 2.61$ imes$ faster attention computation than FlashAttention.

Problem

• Addresses quadratic scaling of self-attention in LLMs
• Unifies training and inference sparse attention methods
• Enables efficient cross-modal long-context understanding

Innovation

• Unified sparse attention via composite tokens
• Multi-granularity compression for dynamic construction
• Hardware-friendly GPU execution with block-level selection