{"title":"MPMoE:利用自适应管道并行性预训练模型的内存效率 MoE","authors":"Zheng Zhang;Yaqi Xia;Hulin Wang;Donglin Yang;Chuang Hu;Xiaobo Zhou;Dazhao Cheng","doi":"10.1109/TPDS.2024.3385639","DOIUrl":null,"url":null,"abstract":"In recent years, the Mixture-of-Experts (MoE) technique has gained widespread popularity as a means to scale pre-trained models to exceptionally large sizes. Dynamic activation of experts allows for conditional computation, increasing the number of parameters of neural networks, which is critical for absorbing the vast amounts of knowledge available in many deep learning areas. However, despite the existing system and algorithm optimizations, there are significant challenges to be tackled when it comes to the inefficiencies of communication and memory consumption. In this paper, we present the design and implementation of MPMoE, a high-performance library that accelerates MoE training with adaptive and memory-efficient pipeline parallelism. Inspired by that the MoE training procedure can be divided into multiple independent sub-stages. We design a pipeline parallelism method for reducing communication latency by overlapping with computation operations. Further, we analyze the memory footprint breakdown of MoE training and identify that activations and temporary buffers are the primary contributors to the overall memory footprint. Toward memory efficiency, we propose memory reuse strategies to reduce memory requirements by eliminating memory redundancies. Finally, to optimize pipeline granularity and memory reuse strategies jointly, we propose a profile-based algorithm and a performance model to determine the configurations of MPMoE at runtime. We implement MPMoE upon PyTorch and evaluate it with common MoE models in two physical clusters, including 64 NVIDIA A100 GPU cards and 16 NVIDIA V100 GPU cards. Compared with the state-of-art approach, MPMoE achieves up to 2.3× speedup while reducing more than 30% memory footprint for training large models.","PeriodicalId":13257,"journal":{"name":"IEEE Transactions on Parallel and Distributed Systems","volume":null,"pages":null},"PeriodicalIF":5.6000,"publicationDate":"2024-04-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"MPMoE: Memory Efficient MoE for Pre-Trained Models With Adaptive Pipeline Parallelism\",\"authors\":\"Zheng Zhang;Yaqi Xia;Hulin Wang;Donglin Yang;Chuang Hu;Xiaobo Zhou;Dazhao Cheng\",\"doi\":\"10.1109/TPDS.2024.3385639\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"In recent years, the Mixture-of-Experts (MoE) technique has gained widespread popularity as a means to scale pre-trained models to exceptionally large sizes. Dynamic activation of experts allows for conditional computation, increasing the number of parameters of neural networks, which is critical for absorbing the vast amounts of knowledge available in many deep learning areas. However, despite the existing system and algorithm optimizations, there are significant challenges to be tackled when it comes to the inefficiencies of communication and memory consumption. In this paper, we present the design and implementation of MPMoE, a high-performance library that accelerates MoE training with adaptive and memory-efficient pipeline parallelism. Inspired by that the MoE training procedure can be divided into multiple independent sub-stages. We design a pipeline parallelism method for reducing communication latency by overlapping with computation operations. Further, we analyze the memory footprint breakdown of MoE training and identify that activations and temporary buffers are the primary contributors to the overall memory footprint. Toward memory efficiency, we propose memory reuse strategies to reduce memory requirements by eliminating memory redundancies. Finally, to optimize pipeline granularity and memory reuse strategies jointly, we propose a profile-based algorithm and a performance model to determine the configurations of MPMoE at runtime. We implement MPMoE upon PyTorch and evaluate it with common MoE models in two physical clusters, including 64 NVIDIA A100 GPU cards and 16 NVIDIA V100 GPU cards. Compared with the state-of-art approach, MPMoE achieves up to 2.3× speedup while reducing more than 30% memory footprint for training large models.\",\"PeriodicalId\":13257,\"journal\":{\"name\":\"IEEE Transactions on Parallel and Distributed Systems\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":5.6000,\"publicationDate\":\"2024-04-08\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"IEEE Transactions on Parallel and Distributed Systems\",\"FirstCategoryId\":\"94\",\"ListUrlMain\":\"https://ieeexplore.ieee.org/document/10494556/\",\"RegionNum\":2,\"RegionCategory\":\"计算机科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"COMPUTER SCIENCE, THEORY & METHODS\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Transactions on Parallel and Distributed Systems","FirstCategoryId":"94","ListUrlMain":"https://ieeexplore.ieee.org/document/10494556/","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, THEORY & METHODS","Score":null,"Total":0}
MPMoE: Memory Efficient MoE for Pre-Trained Models With Adaptive Pipeline Parallelism
In recent years, the Mixture-of-Experts (MoE) technique has gained widespread popularity as a means to scale pre-trained models to exceptionally large sizes. Dynamic activation of experts allows for conditional computation, increasing the number of parameters of neural networks, which is critical for absorbing the vast amounts of knowledge available in many deep learning areas. However, despite the existing system and algorithm optimizations, there are significant challenges to be tackled when it comes to the inefficiencies of communication and memory consumption. In this paper, we present the design and implementation of MPMoE, a high-performance library that accelerates MoE training with adaptive and memory-efficient pipeline parallelism. Inspired by that the MoE training procedure can be divided into multiple independent sub-stages. We design a pipeline parallelism method for reducing communication latency by overlapping with computation operations. Further, we analyze the memory footprint breakdown of MoE training and identify that activations and temporary buffers are the primary contributors to the overall memory footprint. Toward memory efficiency, we propose memory reuse strategies to reduce memory requirements by eliminating memory redundancies. Finally, to optimize pipeline granularity and memory reuse strategies jointly, we propose a profile-based algorithm and a performance model to determine the configurations of MPMoE at runtime. We implement MPMoE upon PyTorch and evaluate it with common MoE models in two physical clusters, including 64 NVIDIA A100 GPU cards and 16 NVIDIA V100 GPU cards. Compared with the state-of-art approach, MPMoE achieves up to 2.3× speedup while reducing more than 30% memory footprint for training large models.
期刊介绍:
IEEE Transactions on Parallel and Distributed Systems (TPDS) is published monthly. It publishes a range of papers, comments on previously published papers, and survey articles that deal with the parallel and distributed systems research areas of current importance to our readers. Particular areas of interest include, but are not limited to:
a) Parallel and distributed algorithms, focusing on topics such as: models of computation; numerical, combinatorial, and data-intensive parallel algorithms, scalability of algorithms and data structures for parallel and distributed systems, communication and synchronization protocols, network algorithms, scheduling, and load balancing.
b) Applications of parallel and distributed computing, including computational and data-enabled science and engineering, big data applications, parallel crowd sourcing, large-scale social network analysis, management of big data, cloud and grid computing, scientific and biomedical applications, mobile computing, and cyber-physical systems.
c) Parallel and distributed architectures, including architectures for instruction-level and thread-level parallelism; design, analysis, implementation, fault resilience and performance measurements of multiple-processor systems; multicore processors, heterogeneous many-core systems; petascale and exascale systems designs; novel big data architectures; special purpose architectures, including graphics processors, signal processors, network processors, media accelerators, and other special purpose processors and accelerators; impact of technology on architecture; network and interconnect architectures; parallel I/O and storage systems; architecture of the memory hierarchy; power-efficient and green computing architectures; dependable architectures; and performance modeling and evaluation.
d) Parallel and distributed software, including parallel and multicore programming languages and compilers, runtime systems, operating systems, Internet computing and web services, resource management including green computing, middleware for grids, clouds, and data centers, libraries, performance modeling and evaluation, parallel programming paradigms, and programming environments and tools.