AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale

Samyam Rajbhandari, Conglong Li, Zhewei Yao, Minjia Zhang, Reza Yazdani Aminabadi, Ammar Ahmad Awan, Jeff Rasley, Yuxiong He

ICML 2022 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization · 54 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add sparse MoE experts to the 1.3B dense model1.3B NLG model · LAMBADAaccuracy: 63.65 -> 69.84Improved
Add sparse MoE experts to the 1.3B dense model1.3B NLG model · TriviaQAaccuracy: 10.05 -> 31.29Improved
Replace standard MoE with pyramid-residual MoE1.3B+MoE-128 · BoolQaccuracy: 64.92 -> 67.16Improved
Replace standard MoE with pyramid-residual MoE1.3B+MoE-128 · TriviaQAaccuracy: 31.29 -> 28.86Hurt
Reduce PR-MoE depth from 24 to 21 layers without distillation (ablation)350M PR-MoE · BoolQaccuracy: 59.88 -> 52.35Hurt
Add staged knowledge distillation to the 21-layer student (ablation)350M PR-MoE · BoolQaccuracy: 52.35 -> 58.07Improved
Apply knowledge distillation throughout training rather than stopping it early (ablation)350M PR-MoE student · LAMBADAaccuracy: 63.46 vs. 61.56Hurt
Apply knowledge distillation throughout training rather than stopping it early (ablation)1.3B PR-MoE student · BoolQaccuracy: 65.66 vs. 64.16Hurt
Replace top-2 expert routing with a fixed dense branch plus one selected expert (ablation)350M MoE · Validation loss and training speedtraining speed: more than 10%Improved
Use residual-MoE rather than top-2 routing (ablation)350M MoE · Validation lossvalidation lossNo clear effect
Use staged KD rather than KD for the entire pre-training process (ablation)PR-MoE student · Pre-training validation lossvalidation lossImproved
Combine pyramid-MoE and residual-MoE (ablation)350M MoE · Validation lossvalidation loss: around 0.01Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy