The DOI matches its Crossref record; 37 of the 39 references it cites were found in a registry or an index.
Efficient large-scale language model training on GPU clusters using megatron-LM
Open accessAbstract
Large language models have led to state-of-the-art accuracies across several tasks. However, training these models efficiently is challenging because: a) GPU memory capacity is limited, making it impossible to fit large models on even a multi-GPU server, and b) the number of compute operations required can result in unrealistically long training times. Consequently, new methods of model parallelism such as tensor and pipeline parallelism have been proposed. Unfortunately, naive usage of these methods leads to scaling issues at thousands of GPUs. In this paper, we show how tensor, pipeline, and data parallelism can be composed to scale to thousands of GPUs. We propose a novel interleaved pipelining schedule that can improve throughput by 10+% with memory footprint comparable to existing approaches. Our approach allows us to perform training iterations on a model with 1 trillion parameters at 502 petaFLOP/s on 3072 GPUs (per-GPU throughput of 52% of theoretical peak).
Overview
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Current implementations have constraints on supported batch size per GPU, requiring it to be divisible by the number of pipeline stages.”
From The Llama 3 Herd of Models (2024) · cites this paper as Narayanan et al., 2021
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
PyTorch: An Imperative Style, High-Performance Deep Learning Library
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019) · arXiv · doi:10.48550/arxiv.1912.01703 · Cited by 31,473
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
RoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
XLNet: Generalized Autoregressive Pretraining for Language Understanding
Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R., & Le, Q. V. (2019) · arXiv · doi:10.48550/arxiv.1906.08237 · Cited by 6,687
Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
Goyal, P., Dollár, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y., & He, K. (2017) · arXiv · doi:10.48550/arxiv.1706.02677 · Cited by 2,294
PipeDream: generalized pipeline parallelism for DNN training
Narayanan, D., Harlap, A., Phanishayee, A., Seshadri, V., Devanur, N. R., Ganger, G., Gibbons, P. B., & Zaharia, M. (2019) · Symposium on Operating Systems Principles · doi:10.1145/3341301.3359646 · Cited by 1,362
Self-citationGPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism
Huang, Y., Cheng, Y., Bapna, A., Firat, O., Chen, M. X., Chen, D., Lee, H., Ngiam, J., Le, Q. V., Wu, Y., et al. (2018) · arXiv · doi:10.48550/arxiv.1811.06965 · Cited by 1,180
Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
Fedus, W., Zoph, B., & Shazeer, N. (2021) · arXiv · doi:10.48550/arxiv.2101.03961 · Cited by 1,005
Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
Shoeybi, M., Patwary, M., Puri, R., LeGresley, P., Casper, J., & Catanzaro, B. (2019) · arXiv · doi:10.48550/arxiv.1909.08053 · Cited by 893
Self-citationZeRO: Memory Optimization Towards Training A Trillion Parameter ModelsPreprint
Rajbhandari, S., Rasley, J., Ruwase, O., & He, Y. (2019) · arXiv.org · doi:10.48550/arxiv.1910.02054 · Cited by 612
Training Deep Nets with Sublinear Memory Cost
Chen, T., Xu, B., Zhang, C., & Guestrin, C. (2016) · arXiv · doi:10.48550/arxiv.1604.06174 · Cited by 561
Griewank, A., & Walther, A. (2000) · ACM Transactions on Mathematical Software · doi:10.1145/347837.347846 · Cited by 529
You, Y., Zhang, Z., Hsieh, C., Demmel, J. W., & Keutzer, K. (2018) · Proceedings of the International Conference on Parallel Processing · doi:10.1145/3225058.3225069 · Cited by 463
Mesh-TensorFlow: Deep Learning for SupercomputersPreprint
Shazeer, N., Cheng, Y.-L., Parmar, N., Tran, D., Vaswani, A., Koanantakool, P., Hawkins, P. G., Lee, H., Hong, M., Young, C., et al. (2018) · Neural Information Processing Systems · doi:10.48550/arxiv.1811.02084 · Cited by 452
Petuum: A New Platform for Distributed Machine Learning on Big Data
Xing, E. P., Ho, Q., Dai, W., Kim, J. K., Wei, J., Lee, S., Zheng, X., Xie, P., Kumar, A., & Yu, Y. (2015) · IEEE Transactions on Big Data · doi:10.1109/tbdata.2015.2472014 · Cited by 400
Beyond Data and Model Parallelism for Deep Neural NetworksPreprint
Jia, Z., Zaharia, M., & Aiken, A. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1807.05358 · Cited by 354
Self-citationFan, S., Rong, Y., Chen, M., Cao, Z., Wang, S., Zheng, Z., Wu, C., Long, G., Yang, J., Xia, L., et al. (2021) · ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming · doi:10.1145/3437801.3441593 · Cited by 276
Mattson, P., Cheng, C., Coleman, C., Diamos, G., Micikevicius, P., Patterson, D., Tang, H., Wei, G.-Y., Bailis, P., Bittorf, V., et al. (2019) · arXiv · doi:10.48550/arxiv.1910.01500 · Cited by 206 · cited 1×
“6 [28], sharded data parallelism [24, 44], where optimizer state is sharded over data-parallel workers, was introduced.”
BackgroundSelf-citationZeRO-Offload: Democratizing Billion-Scale Model Training
Ren, J., Rajbhandari, S., Aminabadi, R. Y., Ruwase, O., Yang, S., Zhang, M., Li, D., & He, Y. (2021) · arXiv · doi:10.48550/arxiv.2101.06840 · Cited by 151
PipeMare: Asynchronous Pipeline Parallel DNN TrainingPreprint
Yang, B.-W., Zhang, J., Li, J., Ré, C., Aberger, C. R., & Sa, C. D. (2019) · Conference on Machine Learning and Systems · doi:10.48550/arxiv.1910.05124 · Cited by 141
Xing, E. P., Ho, Q., Dai, W., Kim, J.-K., Wei, J., Lee, S., Zheng, X., Xie, P., Kumar, A., & Yu, Y. (2015) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/2783258.2783323 · Cited by 126
Memory-Efficient Pipeline-Parallel DNN TrainingPreprint
Narayanan, D., Phanishayee, A., Shi, K., Xie, C., & Zaharia, M. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.09503 · Cited by 118
Self-citationPyTorch Distributed: Experiences on Accelerating Data Parallel TrainingPreprint
Shen, L., Zhao, Y., Varma, R., Salpekar, O., Noordhuis, P., Li, T., Paszke, A., Smith, J. S., Vaughan, B., Damania, P., et al. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2006.15704 · Cited by 113
Efficient Algorithms for Device Placement of DNN Graph OperatorsPreprint
Tarnawski, J., Phanishayee, A., Devanur, N. R., Mahajan, D., & Paravecino, F. N. (2020) · Neural Information Processing Systems · doi:10.48550/arxiv.2006.16423 · Cited by 87
Self-citationPark, J., Yun, G., Yi, C. M., Nguyen, N. T., Lee, S., Choi, J., Noh, S. H., & Choi, Y.-R. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2005.14038 · Cited by 85
Integrated Model, Batch, and Domain Parallelism in Training Neural Networks
Gholami, A., Azad, A., Jin, P., Keutzer, K., & Buluç, A. (2018) · ACM Symposium on Parallelism in Algorithms and Architectures (SPAA) · doi:10.1145/3210377.3210394 · Cited by 80
ZeRO: Memory Optimization Towards Training A Trillion Parameter Models.
Rajbhandari, S., Rasley, J., Ruwase, O., & He, Y. (2019) · Cited by 69
ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep LearningPreprint
Rajbhandari, S., Ruwase, O., Rasley, J., Smith, S., & He, Y. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2104.07857 · Cited by 60
TeraPipe: Token-Level Pipeline Parallelism for Training Large-Scale Language ModelsPreprint
Li, Z., Zhuang, S., Guo, S., Zhuo, D., Zhang, H., Song, D. X., & Stoica, I. G. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2102.07988 · Cited by 59
Automatic Cross-Replica Sharding of Weight Update in Data-Parallel TrainingPreprint
Xu, Y.-Z., Lee, H., Chen, D., Choi, H., Hechtman, B. A., & Wang, S.-B. (2020) · arXiv.org · doi:10.48550/arxiv.2004.13336 · Cited by 53
PipeTransformer: Automated Elastic Pipelining for Distributed Training of TransformersPreprint
He, C., Li, S., Soltanolkotabi, M., & Avestimehr, S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2102.03161 · Cited by 34
Scale MLPerf-0.6 models on Google TPU-v3 PodsPreprint
Kumar, S., Bittorf, V., Chen, D., Chou, C., Hechtman, B. A., Lee, H., Kumar, N., Mattson, P. R., Wang, S., Wang, T., et al. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1909.09756 · Cited by 31
Pipelined Backpropagation at Scale: Training Large Models without BatchesPreprint
Kosson, A., Chiley, V., Venigalla, A., Hestness, J., & Köster, U. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2003.11666 · Cited by 21
Research works printed but not linked: 3
- Paras Jain , Ajay Jain , Aniruddha Nrusimha , Amir Gholami , Pieter Abbeel , Joseph Gonzalez , Kurt Keutzer , and Ion Stoica. Breaking the Memory Wall with Optimal Tensor Rematerialization. In Proceedings of Machine Learning and Systems 2020 , pages 497 -- 511 . 2020 . Paras Jain, Ajay Jain, Aniruddha Nrusimha, Amir Gholami, Pieter Abbeel, Joseph Gonzalez, Kurt Keutzer, and Ion Stoica. Breaking the Memory Wall with Optimal Tensor Rematerialization. In Proceedings of Machine Learning and Systems 2020, pages 497--511. 2020.
- Alec Radford , Karthik Narasimhan , Tim Salimans , and Ilya Sutskever . Improving Language Understanding by Generative Pre-Training , 2018 . Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving Language Understanding by Generative Pre-Training, 2018.
- Alec Radford , Jeffrey Wu , Rewon Child , David Luan , Dario Amodei , and Ilya Sutskever . Language Models are Unsupervised Multitask Learners . OpenAI Blog , 1 ( 8 ): 9 , 2019 . Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Blog, 1(8):9, 2019.
Cited by
Sorted by citations
PaLM: Scaling Language Modeling with Pathways
Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., Gehrmann, S., et al. (2022) · arXiv · doi:10.48550/arxiv.2204.02311 · Cited by 2,590
“Many other works aim to increase of the scale of models, while limiting communication overheads (Rajbhandari et al., 2020; Lepikhin et al., 2020; Li et al., 2020; Rasley et al., 2020; Rajbhandari et al., 2021; Ren et al., 2021; Narayanan et al., 2021a).”
BackgroundMethodInfluentialKimi K2.5: Visual Agentic Intelligence
Kimi Team, Bai, T., Bai, Y., Bao, Y., Cai, S. H., Cao, Y., Chai, Z., Charles, Y., Che, H. S., Chen, C., et al. (2026) · arXiv · doi:10.48550/arxiv.2602.02276 · Cited by 770
Holistic Evaluation of Language Models
Bommasani, R., Liang, P., & Lee, T. (2023) · Annals of the New York Academy of Sciences · doi:10.1111/nyas.15007 · Cited by 480
Pre-Trained Language Models and Their Applications
Wang, H., Li, J., Wu, H. R., Hovy, E. H., & Sun, Y. (2022) · Engineering · doi:10.1016/j.eng.2022.04.024 · Cited by 435
Siriwardhana, S., Weerasekera, R., Wen, E., Kaluarachchi, T., Rana, R., & Nanayakkara, S. (2023) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00530 · Cited by 337
Metadata
- Authors
- Narayanan, D., Shoeybi, M., Casper, J., LeGresley, P., Patwary, M., Korthikanti, V., Vainbrand, D., Kashinkunti, P., Bernauer, J., Catanzaro, B., Phanishayee, A., & Zaharia, M.
- Affiliations
- NVIDIA and Microsoft Research
- Nvidia (United Kingdom) (GB)
- Microsoft (United States) (US)
- Stanford University (US)
- ORCID
- Venue
- Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis
- Publisher
- ACM
- Published
- 2021-11-13
- Pages
- 1-15
- ISSN
- 2167-4337 (ISSN-L)
- OpenAlex
- W3187255235
- Cited by
- 639
- References
- 39
- Citations by year
- 2026: 151 · 2025: 211 · 2024: 152 · 2023: 90 · 2022: 33 · 2021: 2
- FWCI
- 50.34
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- International Conference for High Performance Computing, Networking, Storage and Analysis (SC) (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Science Foundation (DGE-165651 8, 1651570, 1656518, CAREER)
- Topics
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- large language model training, GPU clusters, tensor parallelism, pipeline parallelism, data parallelism, memory efficiency, training throughput
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
