Saved in:
Bibliographic Details
Main Authors: Chan, Brian J, Huang, MaoXun, Cheng, Jui-Hung, Chen, Chao-Ting, Huang, Hen-Hsen
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2502.00085
Tags: Add Tag
No Tags, Be the first to tag this record!
Table of Contents:
  • This work presents a novel trie (prefix-tree)-based parallel decoding method that addresses the memory inefficiency of batch-based beam search. By sharing a single KV cache across beams with common prefixes, our approach dramatically reduces memory usage and enables efficient decoding. We evaluated our method across three attention architectures, Multi-Head Attention (Phi-3.5-mini-instruct), Grouped Query Attention (Llama-3.1-8B-Instruct), and Sliding Window Attention (Mistral-Small-24B-Instruct-2501), using CNN/DailyMail for abstractive summarization and HumanEval for code generation. Our experiments demonstrate substantial memory savings (4--8$\times$) and up to 2.4$\times$ faster decoding, without compromising generation quality. These results highlight our method's suitability for memory-constrained environments and large-scale deployments.