APB: Accelerating Distributed Long-Context Inference by Passing Compressed Context Blocks across GPUs

TL;DR

APB is a novel framework that significantly accelerates long-context inference in large language models by passing compressed context blocks across GPUs, reducing compute and increasing parallelism without sacrificing performance.

Contribution

We propose APB, a new long-context inference method that uses multi-host approximate attention and communication of key-value pairs to improve speed and scalability.

Findings

Achieves up to 9.2x speedup over FlashAttn

Maintains task performance while accelerating inference

Supports diverse models and parallelism configurations

Abstract

While long-context inference is crucial for advancing large language model (LLM) applications, its prefill speed remains a significant bottleneck. Current approaches, including sequence parallelism strategies and compute reduction through approximate attention mechanisms, still fall short of delivering optimal inference efficiency. This hinders scaling the inputs to longer sequences and processing long-context queries in a timely manner. To address this, we introduce APB, an efficient long-context inference framework that leverages multi-host approximate attention to enhance prefill speed by reducing compute and enhancing parallelism simultaneously. APB introduces a communication mechanism for essential key-value pairs within a sequence parallelism framework, enabling a faster inference speed while maintaining task performance. We implement APB by incorporating a tailored FlashAttn kernel alongside optimized distribution strategies, supporting diverse models and parallelism configurations. APB achieves speedups of up to 9.2x, 4.2x, and 1.6x compared with FlashAttn, RingAttn, and StarAttn, respectively, without any observable task performance degradation. We provide the implementation and experiment code of APB in https://github.com/thunlp/APB.