ChunkAttention: Efficient Attention on KV Cache with Chunking Sharing and Batching

Lu Ye, Ze Tao, Yong Huang, Yang Li

Primary Area: infrastructure, software libraries, hardware, etc.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: large language model, model inference, self attention

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Batching long shared prompt prefixes to speed up self-attention based on chunked KV cache

Abstract: Self-attention is an essential component of GPT-style models and a significant cause of LLM inference latency for long sequences. In multi-tenant LLM inference servers, the compute and memory operation cost of self-attention can be amortized by making use of the probability that sequences from users may share long prompt prefixes. This paper introduces ChunkAttention, a unique self-attention kernel built on chunking, sharing the KV cache, and batching the attention computation. ChunkAttention recognizes matching prompt prefixes across several sequences and shares their KV cache in memory by chunking the KV cache and structuring it into the auxiliary prefix tree. To significantly improve the memory reuse of KV cache and consequently the speed of self-attention for long shared prompts, we design an efficient computation kernel on this new storage structure, where two-phased partitioning is implemented to reduce memory operations on shared KV cache during self-attention. Experiments show that ChunkAttention can speed up self-attention of long shared prompts 1.6-3 times, with lengths ranging from 1024 to 8192.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7819