1 paper · 1 filter
Chentao Li, Han Guo
During the training of large Transformer models, attention masks regulate the scope and direction of information flow across a sequence. Numerous mask variants exist, and operators…