TidalDecode: Fast and Accurate LLM Decoding with Position Persistent Sparse Attention

TL;DR

TidalDecode introduces a position persistent sparse attention mechanism that enhances LLM decoding speed and accuracy by leveraging spatial coherence and selective full attention, significantly reducing latency without quality loss.

Contribution

It proposes a novel position persistent sparse attention algorithm that improves decoding efficiency and accuracy in large language models by combining sparse and full attention layers.

Findings

Reduces LLM decoding latency by up to 2.1x

Maintains high-quality generation comparable to full attention

Effectively leverages spatial coherence in token selection

Abstract

Large language models (LLMs) have driven significant advancements across diverse NLP tasks, with long-context models gaining prominence for handling extended inputs. However, the expanding key-value (KV) cache size required by Transformer architectures intensifies the memory constraints, particularly during the decoding phase, creating a significant bottleneck. Existing sparse attention mechanisms designed to address this bottleneck have two limitations: (1) they often fail to reliably identify the most relevant tokens for attention, and (2) they overlook the spatial coherence of token selection across consecutive Transformer layers, which can lead to performance degradation and substantial overhead in token selection. This paper introduces TidalDecode, a simple yet effective algorithm and system for fast and accurate LLM decoding through position persistent sparse attention. TidalDecode leverages the spatial coherence of tokens selected by existing sparse attention methods and introduces a few token selection layers that perform full attention to identify the tokens with the highest attention scores, while all other layers perform sparse attention with the pre-selected tokens. This design enables TidalDecode to substantially reduce the overhead of token selection for sparse attention without sacrificing the quality of the generated results. Evaluation on a diverse set of LLMs and tasks shows that TidalDecode closely matches the generative performance of full attention methods while reducing the LLM decoding latency by up to 2.1x.