1 paper
Minchul Kang, Changyong Shin, Younghun Go +4
Selecting a parallelism strategy - the configuration of data, tensor, and pipeline parallelism degrees together with micro- and global-batch sizes - largely determines the training…