MixPipe: Efficient Bidirectional Pipeline Parallelism for Training Large-Scale Models
Author:
Affiliation:
1. Chinese Academy of Sciences,Institute of Information Engineering
Publisher
IEEE
Link
http://xplorestaging.ieee.org/ielx7/10247654/10247655/10247730.pdf?arnumber=10247730
Reference17 articles.
1. Chimera
2. DAPPLE
3. GEMS: GPU-Enabled Memory-Aware Model-Parallelism System for Distributed DNN Training
4. vDNN: Virtualized deep neural networks for scalable, memory-efficient neural network design
5. Memory-efficient pipeline-parallel dnn training;narayanan;PMLR,2021
Cited by 4 articles. 订阅此论文施引文献 订阅此论文施引文献，注册后可以免费订阅5篇论文的施引文献，订阅后可以查看论文全部施引文献
1. HCPipe: Hidden Computation Pipeline Model Parallelism for Large-scale DCNN Model Training;2024 5th International Conference on Computer Engineering and Application (ICCEA);2024-04-12
2. CAPSlog: Scalable Memory-Centric Partitioning for Pipeline Parallelism;2024 32nd Euromicro International Conference on Parallel, Distributed and Network-Based Processing (PDP);2024-03-20
3. Training and Serving System of Foundation Models: A Comprehensive Survey;IEEE Open Journal of the Computer Society;2024
4. Quartet: A Holistic Hybrid Parallel Framework for Training Large Language Models;Lecture Notes in Computer Science;2024
1.学者识别学者识别
2.学术分析学术分析
3.人才评估人才评估
"同舟云学术"是以全球学者为主线，采集、加工和组织学术论文而形成的新型学术文献查询和分析系统，可以对全球学者进行文献检索和人才价值评估。用户可以通过关注某些学科领域的顶尖人物而持续追踪该领域的学科进展和研究前沿。经过近期的数据扩容，当前同舟云学术共收录了国内外主流学术期刊6万余种，收集的期刊论文及会议论文总量共计约1.5亿篇，并以每天添加12000余篇中外论文的速度递增。我们也可以为用户提供个性化、定制化的学者数据。欢迎来电咨询！咨询电话：010-8811{复制后删除}0370
www.globalauthorid.com
TOP
Copyright © 2019-2024 北京同舟云网络信息技术有限公司 京公网安备11010802033243号 京ICP备18003416号-3