ResDecode: Accelerating Large Language Models Inference via Residual Decoding Heads

Large language Models (LLMs) have immense potential to enhance the capabilities of Cyber-Physical-Social Intelligence (CPSI) systems, enabling them to better engage with complex cyber, physical, and social environments. However, the high inference latency of LLMs, which is inherited from the autoreg...

Full description

Saved in:
Bibliographic Details
Main Authors: Ziqian Zeng, Jiahong Yu, Qianshi Pang, Zihao Wang, Huiping Zhuang, Fan Yu, Hongen Shao, Xiaofeng Zou
Format: Article
Language:English
Published: Tsinghua University Press 2025-06-01
Series:Big Data Mining and Analytics
Subjects:
Online Access:https://www.sciopen.com/article/10.26599/BDMA.2024.9020074
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849714744932433920
author Ziqian Zeng
Jiahong Yu
Qianshi Pang
Zihao Wang
Huiping Zhuang
Fan Yu
Hongen Shao
Xiaofeng Zou
author_facet Ziqian Zeng
Jiahong Yu
Qianshi Pang
Zihao Wang
Huiping Zhuang
Fan Yu
Hongen Shao
Xiaofeng Zou
author_sort Ziqian Zeng
collection DOAJ
description Large language Models (LLMs) have immense potential to enhance the capabilities of Cyber-Physical-Social Intelligence (CPSI) systems, enabling them to better engage with complex cyber, physical, and social environments. However, the high inference latency of LLMs, which is inherited from the autoregressive decoding process, hinders their wide application in CPSI systems. To address this challenge, current approaches have incorporated speculative decoding to enable parallel prediction of multiple subsequent tokens, thereby achieving inference acceleration. Nevertheless, the accuracy of these decoding heads falls short of the autoregressive decoding approach. In light of these limitations, we propose ResDecode, a novel speculative decoding method characterized by its efficient and accurate decoding heads. Within the lightweight draft model, we propose a residual decoding head to compensate for the full context encoder’s limited capability on long-range dependencies, thus improving accuracy. ResDecode demonstrates impressive results, achieving a maximum speedup ratio of 3.2× on the MT-bench compared to vanilla autoregressive decoding.
format Article
id doaj-art-9b7c6bd2e5064f4cbc0f542e5681b461
institution DOAJ
issn 2096-0654
2097-406X
language English
publishDate 2025-06-01
publisher Tsinghua University Press
record_format Article
series Big Data Mining and Analytics
spelling doaj-art-9b7c6bd2e5064f4cbc0f542e5681b4612025-08-20T03:13:36ZengTsinghua University PressBig Data Mining and Analytics2096-06542097-406X2025-06-018477979310.26599/BDMA.2024.9020074ResDecode: Accelerating Large Language Models Inference via Residual Decoding HeadsZiqian Zeng0Jiahong Yu1Qianshi Pang2Zihao Wang3Huiping Zhuang4Fan Yu5Hongen Shao6Xiaofeng Zou7Shien Ming Wu School of Intelligent Engineering, South China University of Technology, Guangzhou 511442, ChinaSchool of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, ChinaSchool of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, ChinaDepartment of Computer Science and Engineering, School of Engineering, The Hong Kong University of Science and Technology, Hong Kong 999077, ChinaShien Ming Wu School of Intelligent Engineering, South China University of Technology, Guangzhou 511442, ChinaHuawei Technologies Co. Ltd., Hangzhou 310000, ChinaSchool of Future Technology, South China University of Technology, Guangzhou 511442, ChinaSchool of Future Technology, South China University of Technology, Guangzhou 511442, ChinaLarge language Models (LLMs) have immense potential to enhance the capabilities of Cyber-Physical-Social Intelligence (CPSI) systems, enabling them to better engage with complex cyber, physical, and social environments. However, the high inference latency of LLMs, which is inherited from the autoregressive decoding process, hinders their wide application in CPSI systems. To address this challenge, current approaches have incorporated speculative decoding to enable parallel prediction of multiple subsequent tokens, thereby achieving inference acceleration. Nevertheless, the accuracy of these decoding heads falls short of the autoregressive decoding approach. In light of these limitations, we propose ResDecode, a novel speculative decoding method characterized by its efficient and accurate decoding heads. Within the lightweight draft model, we propose a residual decoding head to compensate for the full context encoder’s limited capability on long-range dependencies, thus improving accuracy. ResDecode demonstrates impressive results, achieving a maximum speedup ratio of 3.2× on the MT-bench compared to vanilla autoregressive decoding.https://www.sciopen.com/article/10.26599/BDMA.2024.9020074speculative decodingefficient inferencelarge language models (llms)
spellingShingle Ziqian Zeng
Jiahong Yu
Qianshi Pang
Zihao Wang
Huiping Zhuang
Fan Yu
Hongen Shao
Xiaofeng Zou
ResDecode: Accelerating Large Language Models Inference via Residual Decoding Heads
Big Data Mining and Analytics
speculative decoding
efficient inference
large language models (llms)
title ResDecode: Accelerating Large Language Models Inference via Residual Decoding Heads
title_full ResDecode: Accelerating Large Language Models Inference via Residual Decoding Heads
title_fullStr ResDecode: Accelerating Large Language Models Inference via Residual Decoding Heads
title_full_unstemmed ResDecode: Accelerating Large Language Models Inference via Residual Decoding Heads
title_short ResDecode: Accelerating Large Language Models Inference via Residual Decoding Heads
title_sort resdecode accelerating large language models inference via residual decoding heads
topic speculative decoding
efficient inference
large language models (llms)
url https://www.sciopen.com/article/10.26599/BDMA.2024.9020074
work_keys_str_mv AT ziqianzeng resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads
AT jiahongyu resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads
AT qianshipang resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads
AT zihaowang resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads
AT huipingzhuang resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads
AT fanyu resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads
AT hongenshao resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads
AT xiaofengzou resdecodeacceleratinglargelanguagemodelsinferenceviaresidualdecodingheads