1 paper · 1 filter
Jiashu Zhang, Zihan Pan, Molly +3
The occurrence of bubbles in pipeline parallelism is an inherent limitation that can account for more than 40% of the large language model (LLM) training time and is one of the mai…