AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

H3T: Efficient Integration of Memory Optimization and Parallelism for Large-scale Transformer Training

Yuzhong Wang, Xu Han, Weilin Zhao, Guoyang Zeng, Zhiyuan Liu, Maosong Sun

NeurIPS 2023 · Architectures and efficiency › RNNs and normalization · approach: Distributed LLM inference systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

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
Replace manual all-switches-on strategy with H3T DP solverBMTrain · BERT-1.8B 8xRTX2080Ti bs128time per step (s): 7.1 -> 6.0Improved
Replace manual strategy with H3T DP solverBMTrain · BERT-6B 8xRTX2080Ti bs8time per step (s): 22.1 -> 19.4Improved
Replace manual strategy with H3T DP solver when memory is plentifulBMTrain · BERT-13B 64xA100 bs1024time per step (s): 7.8 -> 7.4Improved
Replace manual strategy with H3T DP solverBMTrain · BERT-100B 64xA100 bs2048time per step (s): 61.8 -> 61.6No clear effect
Use greedy solver instead of DP solver (ablation)H3T · BERT-6B 8xRTX2080Ti bs32time per step (s): 19.8 -> 20.5Hurt
Use greedy solver instead of DP solver (ablation)H3T · BERT-1.8B 8xRTX2080Ti bs128time per step (s): 6.0 -> 5.8Improved
Use H3T instead of Megatron-DeepSpeed ZeRO-2BERT training · BERT-13B 8xA100 bs128time per step (s): 17.7 -> 10.0Improved
Use H3T instead of Megatron-DeepSpeed ZeRO-3BERT training · BERT-13B 8xA100 bs128time per step (s): 43.2 -> 10.0Improved
Use H3T instead of Megatron-DeepSpeed ZeRO-2BERT training · BERT-6B 8xA100 bs128time per step (s): 6.8 -> 5.5Improved
Larger model size widens H3T advantage over Megatron-DeepSpeedH3T · BERT-6B vs BERT-13B, 8xA100 bs128speedup over ZeRO-2: 6.8 -> 5.5 (6B); 17.7 -> 10.0 (13B)Improved
Use H3T instead of Megatron-DeepSpeed for minimal memoryBERT training · Minimal memory, all settingsGPU memory saving: 34.6% ~ 70.7%Improved
Use H3T with random solver (ablation)H3T · Actual-run settingsfeasibilityHurt

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