Twilight: Adaptive Attention Sparsity with Hierarchical Top-p Pruning

2025/02/04 by Lin, Chaofan, Tang, Jiaming, Shuo Yang +13 · 7 citations
Computer Science · Decision Sciences · #Intelligent Tutoring Systems and Adaptive Learning #Online Learning and Analytics #Advanced Bandit Algorithms Research

paper · pdf · doi:10.48550/arxiv.2502.02770

Abstract

Leveraging attention sparsity to accelerate long-context large language models (LLMs) has been a hot research topic. However, current algorithms such as sparse attention or key-value (KV) cache compression tend to use a fixed budget, which presents a significant challenge during deployment because it fails to account for the dynamic nature of real-world scenarios, where the optimal balance between accuracy and efficiency can vary greatly. In this paper, we find that borrowing top-p sampling (nucleus sampling) to sparse attention can surprisingly achieve adaptive budgeting. Based on this, we propose Twilight, a framework to bring adaptive sparsity to any existing sparse attention algorithm without sacrificing their accuracy. Empirical results show that Twilight can adaptively prune at most 98% of redundant tokens, leading to 15.4× acceleration in self-attention operations and 3.9× acceleration in end-to-end per token latency in long context LLM decoding.

Cited by

• Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification
• LOCKS: Page-Local Compact Key Summaries for Efficient Long-Context Decoding
• Focus: A Streaming Concentration Architecture for Efficient Vision-Language Models
• A Preliminary Study on the Promises and Challenges of Native Top-k Sparse Attention
• Accelerating Large-Scale Reasoning Model Inference with Sparse Self-Speculative Decoding
• OmniSparse: Training-Aware Fine-Grained Sparse Attention for Long-Video MLLMs
• SpecAttn: Speculating Sparse Attention

Related

• Tactic: Adaptive Sparse Attention with Clustering and Distribution Fitting for Long-Context LLMs
• Progressive Sparse Attention: Algorithm and System Co-design for Efficient Attention in LLM Serving
• SampleAttention: Near-Lossless Acceleration of Long Context LLM Inference with Adaptive Structured Sparse Attention
• Sparser Block-Sparse Attention via Token Permutation
• Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention