Training-Free Long-Context Scaling of Large Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: An, Chenxin, Huang, Fei, Zhang, Jun, Gong, Shansan, Qiu, Xipeng, Zhou, Chang, Kong, Lingpeng
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929363220430848
author An, Chenxin
Huang, Fei
Zhang, Jun
Gong, Shansan
Qiu, Xipeng
Zhou, Chang
Kong, Lingpeng
author_facet An, Chenxin
Huang, Fei
Zhang, Jun
Gong, Shansan
Qiu, Xipeng
Zhou, Chang
Kong, Lingpeng
contents The ability of Large Language Models (LLMs) to process and generate coherent text is markedly weakened when the number of input tokens exceeds their pretraining length. Given the expensive overhead of finetuning large-scale models with longer sequences, we propose Dual Chunk Attention (DCA), which enables Llama2 70B to support context windows of more than 100k tokens without continual training. By decomposing the attention computation for long sequences into chunk-based modules, DCA manages to effectively capture the relative positional information of tokens within the same chunk (Intra-Chunk) and across distinct chunks (Inter-Chunk), as well as integrates seamlessly with Flash Attention. In addition to its impressive extrapolation capability, DCA achieves performance on practical long-context tasks that is comparable to or even better than that of finetuned models. When compared with proprietary models, our training-free 70B model attains 94% of the performance of gpt-3.5-16k, indicating it is a viable open-source alternative. All code and data used in this work are released at \url{https://github.com/HKUNLP/ChunkLlama}.
format Preprint
id arxiv_https___arxiv_org_abs_2402_17463
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Training-Free Long-Context Scaling of Large Language Models
An, Chenxin
Huang, Fei
Zhang, Jun
Gong, Shansan
Qiu, Xipeng
Zhou, Chang
Kong, Lingpeng
Computation and Language
The ability of Large Language Models (LLMs) to process and generate coherent text is markedly weakened when the number of input tokens exceeds their pretraining length. Given the expensive overhead of finetuning large-scale models with longer sequences, we propose Dual Chunk Attention (DCA), which enables Llama2 70B to support context windows of more than 100k tokens without continual training. By decomposing the attention computation for long sequences into chunk-based modules, DCA manages to effectively capture the relative positional information of tokens within the same chunk (Intra-Chunk) and across distinct chunks (Inter-Chunk), as well as integrates seamlessly with Flash Attention. In addition to its impressive extrapolation capability, DCA achieves performance on practical long-context tasks that is comparable to or even better than that of finetuned models. When compared with proprietary models, our training-free 70B model attains 94% of the performance of gpt-3.5-16k, indicating it is a viable open-source alternative. All code and data used in this work are released at \url{https://github.com/HKUNLP/ChunkLlama}.
title Training-Free Long-Context Scaling of Large Language Models
topic Computation and Language
url https://arxiv.org/abs/2402.17463