1 paper · 1 filter
Mincong Huang, Chao Wang, Chi Ma +3
Pipeline parallelism is an essential technique in the training of large-scale Transformer models. However, it suffers from imbalanced memory consumption, leading to insufficient me…