Technology Radar
Published : Nov 05, 2025
NOT ON THE CURRENT EDITION
This blip is not on the current edition of the Radar. If it was on one of the last few editions, it is likely that it is still relevant. If the blip is older, it might no longer be relevant and our assessment might be different today. Unfortunately, we simply don't have the bandwidth to continuously review blips from previous editions of the Radar.
Understand more
Nov 2025
Assess
DeepSpeed 是一个 Python 库，用于优化分布式深度学习的训练和推理。对于训练，它集成了 Zero Redundancy Optimizer (ZeRO) 和 3D 并行等技术，以高效地在数千 GPU 上扩展模型。对于推理，它结合了张量并行、流水线并行、专家并行和 ZeRO 并行，并通过自定义内核和通信优化来最小化延迟。DeepSpeed 支持世界上一些最大的语言模型，包括 Megatron-Turing NLG（530B）和 BLOOM（176B）。它兼容稠密模型和稀疏模型，提供高系统吞吐量，并允许在多 GPU 资源受限的环境下进行训练或推理。该库可与流行的 Hugging Face Transformers、PyTorch Lightning 和 Accelerate 无缝集成，是大规模或资源受限深度学习工作负载的高效解决方案。
