Saved in:
Bibliographic Details
Main Authors: Fang, Jiarui, Zhao, Shangchun
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2405.07719
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911939682107392
author Fang, Jiarui
Zhao, Shangchun
author_facet Fang, Jiarui
Zhao, Shangchun
contents Sequence parallelism (SP), which divides the sequence dimension of input tensors across multiple computational devices, is becoming key to unlocking the long-context capabilities of generative AI models. This paper investigates the state-of-the-art SP approaches, i.e. DeepSpeed-Ulysses and Ring-Attention, and proposes a unified SP approach, which is more robust to transformer model architectures and network hardware topology. This paper compares the communication and memory cost of SP and existing parallelism, including data/tensor/zero/pipeline parallelism, and discusses the best practices for designing hybrid 4D parallelism involving SP. We achieved 47% MFU on two 8xA800 nodes using SP for the LLAMA3-8B model training using sequence length 208K. Our code is publicly available at https://github.com/feifeibear/long-context-attention.
format Preprint
id arxiv_https___arxiv_org_abs_2405_07719
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle USP: A Unified Sequence Parallelism Approach for Long Context Generative AI
Fang, Jiarui
Zhao, Shangchun
Machine Learning
Artificial Intelligence
Sequence parallelism (SP), which divides the sequence dimension of input tensors across multiple computational devices, is becoming key to unlocking the long-context capabilities of generative AI models. This paper investigates the state-of-the-art SP approaches, i.e. DeepSpeed-Ulysses and Ring-Attention, and proposes a unified SP approach, which is more robust to transformer model architectures and network hardware topology. This paper compares the communication and memory cost of SP and existing parallelism, including data/tensor/zero/pipeline parallelism, and discusses the best practices for designing hybrid 4D parallelism involving SP. We achieved 47% MFU on two 8xA800 nodes using SP for the LLAMA3-8B model training using sequence length 208K. Our code is publicly available at https://github.com/feifeibear/long-context-attention.
title USP: A Unified Sequence Parallelism Approach for Long Context Generative AI
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2405.07719