Salvato in:
Dettagli Bibliografici
Autori principali: Li, Qingyuan, Zhang, Bo, Ye, Liang, Zhang, Yifan, Wu, Wei, Sun, Yerui, Ma, Lin, Xie, Yuchen
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:https://arxiv.org/abs/2412.04964
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915058041225216
author Li, Qingyuan
Zhang, Bo
Ye, Liang
Zhang, Yifan
Wu, Wei
Sun, Yerui
Ma, Lin
Xie, Yuchen
author_facet Li, Qingyuan
Zhang, Bo
Ye, Liang
Zhang, Yifan
Wu, Wei
Sun, Yerui
Ma, Lin
Xie, Yuchen
contents The ever-increasing sizes of large language models necessitate distributed solutions for fast inference that exploit multi-dimensional parallelism, where computational loads are split across various accelerators such as GPU clusters. However, this approach often introduces significant communication overhead, especially on devices with limited bandwidth. In this paper, we introduce Flash Communication, a novel low-bit compression technique designed to alleviate the tensor-parallelism communication bottleneck during inference. Our method substantially boosts intra-node communication speed by more than 3x and reduces the time-to-first-token by 2x, with nearly no sacrifice in model accuracy. Extensive experiments on various up-to-date LLMs demonstrate the effectiveness of our approach.
format Preprint
id arxiv_https___arxiv_org_abs_2412_04964
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Flash Communication: Reducing Tensor Parallelization Bottleneck for Fast Large Language Model Inference
Li, Qingyuan
Zhang, Bo
Ye, Liang
Zhang, Yifan
Wu, Wei
Sun, Yerui
Ma, Lin
Xie, Yuchen
Artificial Intelligence
The ever-increasing sizes of large language models necessitate distributed solutions for fast inference that exploit multi-dimensional parallelism, where computational loads are split across various accelerators such as GPU clusters. However, this approach often introduces significant communication overhead, especially on devices with limited bandwidth. In this paper, we introduce Flash Communication, a novel low-bit compression technique designed to alleviate the tensor-parallelism communication bottleneck during inference. Our method substantially boosts intra-node communication speed by more than 3x and reduces the time-to-first-token by 2x, with nearly no sacrifice in model accuracy. Extensive experiments on various up-to-date LLMs demonstrate the effectiveness of our approach.
title Flash Communication: Reducing Tensor Parallelization Bottleneck for Fast Large Language Model Inference
topic Artificial Intelligence
url https://arxiv.org/abs/2412.04964