Awesome Distributed Machine Learning System

A curated list of awesome projects and papers for distributed training or inference especially for large model.

Contents

• Awesome Distributed Machine Learning System
  • Contents
  • Open Source Projects
  • Papers
    • Survey
    • Pipeline Parallelism
    • Sequence Parallelism
    • Mixture-of-Experts System
    • Graph Neural Networks System
    • Hybrid Parallelism & Framework
    • Memory Efficient Training
    • Tensor Movement
    • Auto Parallelization
    • Communication Optimization
    • Fault-tolerant Training
    • Inference and Serving
    • Applications
  • Contribute

Open Source Projects

• Megatron-LM: Ongoing Research Training Transformer Models at Scale
• DeepSpeed: A Deep Learning Optimization Library that Makes Distributed Training and Inference Easy, Efficient, and Effective.
• ColossalAI: A Unified Deep Learning System for Large-Scale Parallel Training
• OneFlow: A Performance-Centered and Open-Source Deep Learning Framework
• Mesh TensorFlow: Model Parallelism Made Easier
• FlexFlow: A Distributed Deep Learning Framework that Supports Flexible Parallelization Strategies.
• Alpa: Auto Parallelization for Large-Scale Neural Networks
• Easy Parallel Library: A General and Efficient Deep Learning Framework for Distributed Model Training
• FairScale: PyTorch Extensions for High Performance and Large Scale Training
• TePDist: an HLO-level automatic distributed system for DL models
• EasyDist: Automated Parallelization System and Infrastructure

Papers

Survey

• Demystifying Parallel and Distributed Deep Learning: An In-Depth Concurrency Analysis by Tal Ben-Nun et al., ACM Computing Surveys 2020
• A Survey on Auto-Parallelism of Neural Networks Training by Peng Liang., techrxiv 2022

Pipeline Parallelism

• GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism by Yanping Huang et al., NeurIPS 2019
• PipeDream: generalized pipeline parallelism for DNN training by Deepak Narayanan et al., SOSP 2019
• Memory-Efficient Pipeline-Parallel DNN Training by Deepak Narayanan et al., ICML 2021
• DAPPLE: a pipelined data parallel approach for training large models by Shiqing Fan et al. PPoPP 2021
• Chimera: efficiently training large-scale neural networks with bidirectional pipelines by Shigang Li et al., SC 2021
• Elastic Averaging for Efficient Pipelined DNN Training by Zihao Chen et al. PPoPP 2023
• Mobius: Fine Tuning Large-Scale Models on Commodity GPU Servers by Yangyang Feng et al. ASPLOS 2023
• Hanayo: Harnessing Wave-like Pipeline Parallelism for Enhanced Large Model Training Efficiency by Ziming Liu et al. SC 2023

Sequence Parallelism

• Long Sequence Training from System Perspective by Shenggui Li et al., ACL 2023
• DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer Models by Sam Ade Jacobs et al., arxiv 2023
• Ring Attention with Blockwise Transformers for Near-Infinite Context by Hao Liu et al., NeurIPS 2023 Workshop

Mixture-of-Experts System

• GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding by Dmitry Lepikhin et al., ICLR 2021
• FasterMoE: modeling and optimizing training of large-scale dynamic pre-trained models by Jiaao He et al., PPoPP 2022
• BaGuaLu: targeting brain scale pretrained models with over 37 million cores by Zixuan Ma et al., PPoPP 2022
• DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale by Samyam Rajbhandari et al., ICML 2022
• Tutel: Adaptive Mixture-of-Experts at Scale by Changho Hwang et al., MLSys 2023
• Accelerating Distributed MoE Training and Inference with Lina by Jiamin Li et al., ATC 2023
• SmartMoE: Efficiently Training Sparsely-Activated Models through Combining Static and Dynamic Parallelization by Mingshu Zhai et al., ATC 2023
• MegaBlocks: Efficient Sparse Training with Mixture-of-Experts by Trevor Gale et al., MLSys 2023

Graph Neural Networks System

• PiPAD: Pipelined and Parallel Dynamic GNN Training on GPUs Chunyang Wang et al., PPoPP 2023
• DSP: Efficient GNN Training with Multiple GPUs Zhenkun Cai et al., PPoPP 2023
• Accelerating Graph Neural Networks with Fine-grained intra-kernel Communication-Computation Pipelining on Multi-GPU Platforms Yuke Wang et al., OSDI 2023

Hybrid Parallelism & Framework

• Efficient large-scale language model training on GPU clusters using megatron-LM by Deepak Narayanan et al., SC 2021
• GEMS: GPU-Enabled Memory-Aware Model-Parallelism System for Distributed DNN Training by Arpan Jain et al., SC 2020
• Amazon SageMaker Model Parallelism: A General and Flexible Framework for Large Model Training by Can Karakus et al., arxiv 2021
• OneFlow: Redesign the Distributed Deep Learning Framework from Scratch by Jinhui Yuan et al., arxiv 2021
• Colossal-AI: A Unified Deep Learning System For Large-Scale Parallel Training by Zhengda Bian., arxiv 2021

Memory Efficient Training

• Training deep nets with sublinear memory cost by Tianqi Chen et al., arxiv 2016
• ZeRO: memory optimizations toward training trillion parameter models by Samyam Rajbhandari et al., SC 2020
• Checkmate: Breaking the Memory Wall with Optimal Tensor Rematerialization by Paras Jain et al., MLSys 2020
• Dynamic Tensor Rematerialization by Marisa Kirisame et al., ICLR 2021
• ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Training by Jianfei Chen et al., ICML 2021
• GACT: Activation Compressed Training for Generic Network Architectures by Xiaoxuan Liu et al., ICML 2022

Tensor Movement

• Superneurons: dynamic GPU memory management for training deep neural networks by Linnan Wang et al., PPoPP 2018
• Capuchin: Tensor-based GPU Memory Management for Deep Learning by Xuan Peng et al., ASPLOS 2020
• SwapAdvisor: Pushing Deep Learning Beyond the GPU Memory Limit via Smart Swapping by Chien-Chin Huang et al., ASPLOS 2020
• ZeRO-Offload: Democratizing Billion-Scale Model Training by Jie Ren et al., ATC 2021
• ZeRO-infinity: breaking the GPU memory wall for extreme scale deep learning by Samyam Rajbhandari et al., SC 2021
• PatrickStar: Parallel Training of Pre-trained Models via Chunk-based Memory Management by Jiarui Fang et al., TPDS 2023
• MegTaiChi: dynamic tensor-based memory management optimization for DNN training by Zhongzhe Hu et al., ICS 2022
• Tensor Movement Orchestration In Multi-GPU Training Systems Shao-Fu Lin et al., HPCA 2023

Auto Parallelization

• Mesh-tensorflow: Deep learning for supercomputers by Noam Shazeer et al., NeurIPS 2018
• Exploring Hidden Dimensions in Parallelizing Convolutional Neural Networks by Zhihao Jia et al., ICML 2018
• Beyond Data and Model Parallelism for Deep Neural Networks by Zhihao Jia et al., MLSys 2019
• Supporting Very Large Models using Automatic Dataflow Graph Partitioning by Minjie Wang et al., EuroSys 2019
• GSPMD: General and Scalable Parallelization for ML Computation Graphs by Yuanzhong Xu et al., arxiv 2021
• Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning by Lianmin Zheng et al., OSDI 2022
• Unity: Accelerating DNN Training Through Joint Optimization of Algebraic Transformations and Parallelization by Colin Unger, Zhihao Jia, et al., OSDI 2022
• Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism by Xupeng Miao, et al., VLDB 2023
• Auto-Parallelizing Large Models with Rhino: A Systematic Approach on Production AI Platform by Shiwei Zhang, Lansong Diao, et al., arxiv 2023

Communication Optimization

• Blink: Fast and Generic Collectives for Distributed ML by Guanhua Wang et al., MLSys 2020
• Synthesizing optimal collective algorithms by Zixian Cai et al., PPoPP 2021
• Breaking the computation and communication abstraction barrier in distributed machine learning workloads by Abhinav Jangda et al., ASPLOS 2022
• MSCCLang: Microsoft Collective Communication Language by Meghan Cowan et al., ASPLOS 2023
• Overlap Communication with Dependent Computation via Decomposition in Large Deep Learning Models by Shibo Wang et al., ASPLOS 2023
• Logical/Physical Topology-Aware Collective Communication in Deep Learning Training Jo Sanghoon et al., HPCA 2023

Fault-tolerant Training

• Oobleck: Resilient Distributed Training of Large Models Using Pipeline Templates by Insu Jang et al., SOSP 2023
• Bamboo: Making Preemptible Instances Resilient for Affordable Training of Large DNNs by John Thorpe et al., NSDI 2023
• Varuna: scalable, low-cost training of massive deep learning models by Sanjith Athlur et al., EuroSys 2022

Inference and Serving

• DeepSpeed Inference: Enabling Efficient Inference of Transformer Models at Unprecedented Scale by Reza Yazdani Aminabadi et al., SC 2022
• EnergonAI: An Inference System for 10-100 Billion Parameter Transformer Models by Jiangsu Du et al., arxiv 2022
• Efficiently Scaling Transformer Inference by Reiner Pope et al., MLSys 2022
• Beta: Statistical Multiplexing with Model Parallelism for Deep Learning Serving by Zhuohan Li et al., OSDI 2023
• Fast inference from transformers via speculative decoding by Yaniv Leviathan et al., ICML 2023
• FlexGen: High-throughput Generative Inference of Large Language Models with a Single GPU by Ying Sheng et al., ICML 2023
• Liger: Interleaving Intra- and Inter-Operator Parallelism for Distributed Large Model Inference by Jiangsu Du et al., PPoPP 2024

Applications

• NASPipe: High Performance and Reproducible Pipeline Parallel Supernet Training via Causal Synchronous Parallelism by Shixiong Zhao et al., ASPLOS 2022
• AthenaRL: Distributed Reinforcement Learning with Dataflow Fragments by Huanzhou Zhu et al., ATC 2023
• Hydro: Surrogate-Based Hyperparameter Tuning Service in the Datacenter by Qinghao Hu et al., OSDI 2023
• FastFold: Optimizing AlphaFold Training and Inference on GPU Clusters by Shenggan Cheng et al., PPoPP 2024

Contribute

All contributions to this repository are welcome. Open an issue or send a pull request.