Saved in:
Bibliographic Details
Main Authors: Wang, Zhibin, Ning, Rui, Fang, Chao, Zhang, Zhonghui, Lin, Xi, Ma, Shaobo, Zhou, Mo, Li, Xue, Wang, Zhongfeng, Huan, Chengying, Gu, Rong, Yang, Kun, Chen, Guihai, Zhong, Sheng, Tian, Chen
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2505.17694
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915895796826112
author Wang, Zhibin
Ning, Rui
Fang, Chao
Zhang, Zhonghui
Lin, Xi
Ma, Shaobo
Zhou, Mo
Li, Xue
Wang, Zhongfeng
Huan, Chengying
Gu, Rong
Yang, Kun
Chen, Guihai
Zhong, Sheng
Tian, Chen
author_facet Wang, Zhibin
Ning, Rui
Fang, Chao
Zhang, Zhonghui
Lin, Xi
Ma, Shaobo
Zhou, Mo
Li, Xue
Wang, Zhongfeng
Huan, Chengying
Gu, Rong
Yang, Kun
Chen, Guihai
Zhong, Sheng
Tian, Chen
contents Prefix-sharing among multiple prompts presents opportunities to combine the operations of the shared prefix, while attention computation in the decode stage, which becomes a critical bottleneck with increasing context lengths, is a memory-intensive process requiring heavy memory access on the key-value (KV) cache of the prefixes. Therefore, in this paper, we explore the potential of prefix-sharing in the attention computation of the decode stage. However, the tree structure of the prefix-sharing mechanism presents significant challenges for attention computation in efficiently processing shared KV cache access patterns while managing complex dependencies and balancing irregular workloads. To address the above challenges, we propose a dedicated attention kernel to combine the memory access of shared prefixes in the decoding stage, namely CoDec. CoDec delivers two key innovations: a novel shared-prefix attention kernel that optimizes memory hierarchy and exploits both intra-block and inter-block parallelism, and a comprehensive workload balancing mechanism that efficiently estimates cost, divides tasks, and schedules execution. Experimental results show that CoDec achieves an average $1.9\times$ speedup and $120.9\times$ memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and $3.8\times$ end-to-end time per output token compared to the vLLM.
format Preprint
id arxiv_https___arxiv_org_abs_2505_17694
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CoDec: Prefix-Shared Decoding Kernel for LLMs
Wang, Zhibin
Ning, Rui
Fang, Chao
Zhang, Zhonghui
Lin, Xi
Ma, Shaobo
Zhou, Mo
Li, Xue
Wang, Zhongfeng
Huan, Chengying
Gu, Rong
Yang, Kun
Chen, Guihai
Zhong, Sheng
Tian, Chen
Machine Learning
Prefix-sharing among multiple prompts presents opportunities to combine the operations of the shared prefix, while attention computation in the decode stage, which becomes a critical bottleneck with increasing context lengths, is a memory-intensive process requiring heavy memory access on the key-value (KV) cache of the prefixes. Therefore, in this paper, we explore the potential of prefix-sharing in the attention computation of the decode stage. However, the tree structure of the prefix-sharing mechanism presents significant challenges for attention computation in efficiently processing shared KV cache access patterns while managing complex dependencies and balancing irregular workloads. To address the above challenges, we propose a dedicated attention kernel to combine the memory access of shared prefixes in the decoding stage, namely CoDec. CoDec delivers two key innovations: a novel shared-prefix attention kernel that optimizes memory hierarchy and exploits both intra-block and inter-block parallelism, and a comprehensive workload balancing mechanism that efficiently estimates cost, divides tasks, and schedules execution. Experimental results show that CoDec achieves an average $1.9\times$ speedup and $120.9\times$ memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and $3.8\times$ end-to-end time per output token compared to the vLLM.
title CoDec: Prefix-Shared Decoding Kernel for LLMs
topic Machine Learning
url https://arxiv.org/abs/2505.17694