DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters
Abstract
Explore new techniques in Microsoft's open source library called DeepSpeed, which advances large model training by improving scale, speed, cost, and usability, unlocking the ability to train 100-billion-parameter models. DeepSpeed is compatible with PyTorch. One piece of our library, called ZeRO, is a new parallelized optimizer that greatly reduces the resources needed for model and data parallelism while massively increasing the number of parameters that can be trained. Researchers have used these breakthroughs to create Turing Natural Language Generation (Turing-NLG), which at the time of its release was the largest publicly known language model at 17 billion parameters. In addition we will also go over our latest transformer kernel advancements that led the DeepSpeed team to achieve the world fastest BERT pretraining record.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
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?
Federated learning allows multiple entities to collaboratively train a model without sharing their data.
“one feasible solution is federated learning (FL) [2,28,36,41,50,66], which allows multiple entities to learn model collaboratively without sharing data.”
From FederatedScope-LLM: A Comprehensive Package for Fine-tuning Large Language Models in Federated Learning (2024) · cites this paper as [50]
Cites
Micikevicius, P., Narang, S., Alben, J., Diamos, G., Elsen, E., Garcia, D., Ginsburg, B., Houston, M., Kuchaiev, O., Venkatesh, G., et al. (2017) · arXiv · doi:10.48550/arxiv.1710.03740 · Cited by 1,179 · see citing works
Flagged first, then by citations
Cited by
Sorted by citations
On the Opportunities and Risks of Foundation Models
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.07258 · Cited by 2,676 · see citing works
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 · see citing works
Pre-trained models: Past, present and future
Han, X., Zhang, Z., Ding, N., Gu, Y., Liu, X., Huo, Y., Qiu, J., Yao, Y., Zhang, A., Zhang, L., et al. (2021) · AI Open · doi:10.1016/j.aiopen.2021.08.002 · Cited by 994 · see citing works
ProtGPT2 is a deep unsupervised language model for protein design
Ferruz, N., Schmidt, S., & Höcker, B. (2022) · Nature Communications · doi:10.1038/s41467-022-32007-7 · Cited by 795 · see citing works
Ahdritz, G., Bouatta, N., Floristean, C., Kadyan, S., Xia, Q., Gerecke, W., O’Donnell, T. J., Berenberg, D., Fisk, I., Zanichelli, N., et al. (2024) · Nature Methods · doi:10.1038/s41592-024-02272-z · Cited by 439 · see citing works
Self-citationToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection
Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., & Kamar, E. (2022) · Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2022.acl-long.234 · Cited by 279 · see citing works
Kuang, W., Qian, B., Li, Z., Chen, D., Gao, D., Pan, X., Xie, Y., Li, Y., Ding, B., & Zhou, J. (2024) · Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining · doi:10.1145/3637528.3671573 · Cited by 142 · see citing works
GenSLMs: Genome-scale language models reveal SARS-CoV-2 evolutionary dynamics
Zvyagin, M., Brace, A., Hippe, K., Deng, Y., Zhang, B., Bohorquez, C. O., Clyde, A., Kale, B., Perez-Rivera, D., Ma, H., et al. (2023) · The International Journal of High Performance Computing Applications · doi:10.1177/10943420231201154 · Cited by 120 · see citing works
GenSLMs: Genome-scale language models reveal SARS-CoV-2 evolutionary dynamics
Zvyagin, M., Brace, A., Hippe, K., Deng, Y., Zhang, B., Bohorquez, C. O., Clyde, A., Kale, B., Perez-Rivera, D., Ma, H., et al. (2022) · doi:10.1101/2022.10.10.511571 · Cited by 120 · see citing works
LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models
Chen, Y., Qian, S., Tang, H., Lai, X., Liu, Z., Han, S., & Jia, J. (2023) · arXiv · doi:10.48550/arxiv.2309.12307 · Cited by 50 · see citing works
Scaling Data-Constrained Language Models
Muennighoff, N., Rush, A., Barak, B., Scao, T. L., Tazi, N., Piktus, A., Pyysalo, S., Wolf, T., & Raffel, C. (2023) · Advances in Neural Information Processing Systems 36 · doi:10.52202/075280-2191 · Cited by 47 · see citing works
Large language models for automated scholarly paper review: A survey
Zhuang, Z., Chen, J., Xu, H., Jiang, Y., & Lin, J. (2025) · Information Fusion · doi:10.1016/j.inffus.2025.103332 · Cited by 34 · see citing works
Language models for materials discovery and sustainability: Progress, challenges, and opportunities
Pei, Z., Yin, J., & Zhang, J. (2025) · Progress in Materials Science · doi:10.1016/j.pmatsci.2025.101495 · Cited by 30 · see citing works
GAIA-1: A Generative World Model for Autonomous Driving
Hu, A., Russell, L., Yeo, H., Murez, Z., Fedoseev, G., Kendall, A., Shotton, J., & Corrado, G. (2023) · arXiv · doi:10.48550/arxiv.2309.17080 · Cited by 28 · see citing works
“To optimize distributed training, we used the Deepspeed ZeRO-2 training strategy [42] with activation checkpointing.”
MethodLearning Performance-Improving Code Edits
Shypula, A., Madaan, A., Zeng, Y., Alon, U., Gardner, J., Hashemi, M., Neubig, G., Ranganathan, P., Bastani, O., & Yazdanbakhsh, A. (2023) · arXiv · doi:10.48550/arxiv.2302.07867 · Cited by 23 · see citing works
RWKV: Reinventing RNNs for the Transformer Era
Peng, B., Alcaide, E., Anthony, Q., Albalak, A., Arcadinho, S., Biderman, S., Cao, H., Cheng, X., Chung, M., Grella, M., et al. (2023) · arXiv · doi:10.48550/arxiv.2305.13048 · Cited by 17 · see citing works
“Custom Kernels To address inefficiencies in the W KV computation due to the sequential nature of the task when using standard deep learning frameworks, we implement a custom CUDA kernel so as to launch a single compute kernel in training accelerators.”
MethodCompAct: Compressing Retrieved Documents Actively for Question Answering
Yoon, C., Lee, T., Hwang, H., Jeong, M., & Kang, J. (2024) · arXiv · doi:10.48550/arxiv.2407.09014 · Cited by 14
Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models
Karamcheti, S., Nair, S., Balakrishna, A., Liang, P., Kollar, T., & Sadigh, D. (2024) · arXiv · doi:10.48550/arxiv.2402.07865 · Cited by 7
“5 training implementation: when benchmarked on the same hardware (an AWS p4de.24xlarge node with 8 A100 GPUs), we observe 20% faster step times with our FSDP-backed implementation, a notable gain given LLaVa leverages the well-optimized DeepSpeed ZeRO library (Rasley et al., 2020).”
MethodA Slice and Dice Approach to Accelerate Compound Sparse Attention on GPU
Li, H., Choi, J., & Ahn, J. H. (2022) · 2022 IEEE International Symposium on Workload Characterization (IISWC) · doi:10.1109/iiswc55918.2022.00019 · Cited by 3 · see citing works
Reviewer2: Optimizing Review Generation Through Prompt Generation
Gao, Z., Brantley, K., & Joachims, T. (2024) · arXiv · doi:10.48550/arxiv.2402.10886 · Cited by 3 · see citing works
Mammoth2: Scaling instructions from the web
Yue, X., Zheng, T., Zhang, G., & Chen, W. (2024) · Advances in Neural Information Processing Systems 37 · doi:10.52202/079017-2877 · Cited by 2
World Simulation with Video Foundation Models for Physical AI
NVIDIA, :, Ali, A., Bai, J., Bala, M., Balaji, Y., Blakeman, A., Cai, T., Cao, J., Cao, T., et al. (2025) · arXiv · doi:10.48550/arxiv.2511.00062 · Cited by 1 · see citing works
“We further show that [Cosmos-Transfer2.”
Zweiger, A., Pari, J., Guo, H., Akyürek, E., Kim, Y., & Agrawal, P. (2025) · arXiv · Cited by 3 · see citing works
“We use DeepSpeed ZeRO-3 [83] for SFT in ReST EM training.”
- Authors
- Rasley, J., Rajbhandari, S., Ruwase, O., & He, Y.
- Affiliations
- Microsoft (United States) (US)
- Venue
- Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining
- Publisher
- ACM
- Published
- 2020-08-20
- Pages
- 3505-3506
- OpenAlex
- W3081168214
- Cited by
- 823
- References
- 1
- Citations by year
- 2026: 177 · 2025: 268 · 2024: 182 · 2023: 107 · 2022: 45 · 2021: 38 · 2020: 6
- FWCI
- 36.65
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Stochastic Gradient Optimization Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- DeepSpeed, large language model training, distributed training, data parallelism, model parallelism, PyTorch
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
