Re-evaluating the Memory-balanced Pipeline Parallelism: BPipe
This work addresses memory inefficiency in pipeline parallelism for AI model training, but it is incremental as it critiques and extends an existing method.
The paper re-evaluates the BPipe technique for memory-balanced pipeline parallelism in large-scale Transformer training, finding that it does not benefit LLaMA training and offers negligible gains for GPT-3 with flash attention, while analyzing the causes and introducing a performance estimation method.
Pipeline parallelism is an essential technique in the training of large-scale Transformer models. However, it suffers from imbalanced memory consumption, leading to insufficient memory utilization. The BPipe technique was proposed to address this issue and has proven effective in the GPT-3 model. Nevertheless, our experiments have not yielded similar benefits for LLaMA training. Additionally, BPipe only yields negligible benefits for GPT-3 training when applying flash attention. We analyze the underlying causes of the divergent performance of BPipe on GPT-3 and LLaMA. Furthermore, we introduce a novel method to estimate the performance of BPipe.