close
Jump to content

DeepSpeed

From Wikipedia, the free encyclopedia
DeepSpeed
Original authorMicrosoft Research
DeveloperMicrosoft
ReleaseMay 18, 2020; 6 years ago (2020-05-18)
Stable release
v0.19.2 / June 16, 2026; 27 days ago (2026-06-16)
Written inPython, CUDA, C++
TypeSoftware library
LicenseApache License 2.0
Websitedeepspeed.ai
Repositorygithub.com/microsoft/DeepSpeed

DeepSpeed' is an open-source optimization library for the distributed training and inference of deep learning models using PyTorch.[1]

Library

[edit]

The library is designed to reduce computing power and memory use and to train large distributed models with better parallelism on existing computer hardware.[2][3] DeepSpeed is optimized for low latency, high throughput training. It includes the Zero Redundancy Optimizer (ZeRO) for training models with 1 trillion or more parameters.[4] Features include mixed precision training, single-GPU, multi-GPU, and multi-node training as well as custom model parallelism. The DeepSpeed source code is licensed under Apache License and available on GitHub.[5]

The team claimed to achieve up to a 6.2x throughput improvement, 2.8x faster convergence, and 4.6x less communication.[6]

See also

[edit]

References

[edit]
  1. "Microsoft Updates Windows, Azure Tools with an Eye on The Future". PCMag UK. May 22, 2020.
  2. Yegulalp, Serdar (February 10, 2020). "Microsoft speeds up PyTorch with DeepSpeed". InfoWorld.
  3. "Microsoft unveils "fifth most powerful" supercomputer in the world". Neowin. 18 June 2023.
  4. "Microsoft trains world's largest Transformer language model". February 10, 2020.
  5. "microsoft/DeepSpeed". July 10, 2020 via GitHub.
  6. "DeepSpeed: Accelerating large-scale model inference and training via system optimizations and compression". Microsoft Research. 2021-05-24. Retrieved 2021-06-19.

Further reading

[edit]
  • Rajbhandari, Samyam; Rasley, Jeff; Ruwase, Olatunji; He, Yuxiong (2019). "ZeRO: Memory Optimization Towards Training A Trillion Parameter Models". arXiv:1910.02054 [cs.LG].
[edit]