RecurFormer: Not All Transformer Heads Need Self-Attention

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yan, Ruiqing, Zheng, Linghan, Du, Xingbo, Zou, Han, Guo, Yufeng, Yang, Jianfei
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909352649031680
author Yan, Ruiqing
Zheng, Linghan
Du, Xingbo
Zou, Han
Guo, Yufeng
Yang, Jianfei
author_facet Yan, Ruiqing
Zheng, Linghan
Du, Xingbo
Zou, Han
Guo, Yufeng
Yang, Jianfei
contents Transformer-based large language models (LLMs) excel in modeling complex language patterns but face significant computational costs during inference, especially with long inputs due to the attention mechanism's memory overhead. We observe that certain attention heads exhibit a distribution where the attention weights concentrate on tokens near the query token, termed as recency aware, which focuses on local and short-range dependencies. Leveraging this insight, we propose RecurFormer, a novel architecture that replaces these attention heads with linear recurrent neural networks (RNNs), specifically the Mamba architecture. This replacement reduces the cache size without evicting tokens, thus maintaining generation quality. RecurFormer retains the ability to model long-range dependencies through the remaining attention heads and allows for reusing pre-trained Transformer-based LLMs weights with continual training. Experiments demonstrate that RecurFormer matches the original model's performance while significantly enhancing inference efficiency. Our approach provides a practical solution to the computational challenges of Transformer-based LLMs inference, making it highly attractive for tasks involving long inputs.
format Preprint
id arxiv_https___arxiv_org_abs_2410_12850
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle RecurFormer: Not All Transformer Heads Need Self-Attention
Yan, Ruiqing
Zheng, Linghan
Du, Xingbo
Zou, Han
Guo, Yufeng
Yang, Jianfei
Computation and Language
Artificial Intelligence
Machine Learning
Transformer-based large language models (LLMs) excel in modeling complex language patterns but face significant computational costs during inference, especially with long inputs due to the attention mechanism's memory overhead. We observe that certain attention heads exhibit a distribution where the attention weights concentrate on tokens near the query token, termed as recency aware, which focuses on local and short-range dependencies. Leveraging this insight, we propose RecurFormer, a novel architecture that replaces these attention heads with linear recurrent neural networks (RNNs), specifically the Mamba architecture. This replacement reduces the cache size without evicting tokens, thus maintaining generation quality. RecurFormer retains the ability to model long-range dependencies through the remaining attention heads and allows for reusing pre-trained Transformer-based LLMs weights with continual training. Experiments demonstrate that RecurFormer matches the original model's performance while significantly enhancing inference efficiency. Our approach provides a practical solution to the computational challenges of Transformer-based LLMs inference, making it highly attractive for tasks involving long inputs.
title RecurFormer: Not All Transformer Heads Need Self-Attention
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2410.12850