Saved in:
Bibliographic Details
Main Authors: Lv, Bo, Zhou, Quan, Ding, Xuanang, Wang, Yan, Ma, Zeming
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2409.11057
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914950787629056
author Lv, Bo
Zhou, Quan
Ding, Xuanang
Wang, Yan
Ma, Zeming
author_facet Lv, Bo
Zhou, Quan
Ding, Xuanang
Wang, Yan
Ma, Zeming
contents The bottleneck associated with the key-value(KV) cache presents a significant challenge during the inference processes of large language models. While depth pruning accelerates inference, it requires extensive recovery training, which can take up to two weeks. On the other hand, width pruning retains much of the performance but offers slight speed gains. To tackle these challenges, we propose KVPruner to improve model efficiency while maintaining performance. Our method uses global perplexity-based analysis to determine the importance ratio for each block and provides multiple strategies to prune non-essential KV channels within blocks. Compared to the original model, KVPruner reduces runtime memory usage by 50% and boosts throughput by over 35%. Additionally, our method requires only two hours of LoRA fine-tuning on small datasets to recover most of the performance.
format Preprint
id arxiv_https___arxiv_org_abs_2409_11057
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle KVPruner: Structural Pruning for Faster and Memory-Efficient Large Language Models
Lv, Bo
Zhou, Quan
Ding, Xuanang
Wang, Yan
Ma, Zeming
Computation and Language
The bottleneck associated with the key-value(KV) cache presents a significant challenge during the inference processes of large language models. While depth pruning accelerates inference, it requires extensive recovery training, which can take up to two weeks. On the other hand, width pruning retains much of the performance but offers slight speed gains. To tackle these challenges, we propose KVPruner to improve model efficiency while maintaining performance. Our method uses global perplexity-based analysis to determine the importance ratio for each block and provides multiple strategies to prune non-essential KV channels within blocks. Compared to the original model, KVPruner reduces runtime memory usage by 50% and boosts throughput by over 35%. Additionally, our method requires only two hours of LoRA fine-tuning on small datasets to recover most of the performance.
title KVPruner: Structural Pruning for Faster and Memory-Efficient Large Language Models
topic Computation and Language
url https://arxiv.org/abs/2409.11057