Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning
This addresses the challenge of dynamic efficiency-accuracy trade-offs in deploying long-context LLMs, offering a practical improvement over existing methods.
The paper tackles the problem of fixed-budget attention sparsity in long-context LLMs by proposing Twilight, a framework that adaptively prunes tokens using hierarchical top-p pruning, achieving up to 98% token pruning, 15.4x acceleration in self-attention, and 3.9x acceleration in end-to-end latency.
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 $15.4\times$ acceleration in self-attention operations and $3.9\times$ acceleration in end-to-end per token latency in long context LLM decoding.