Jethro's Braindump

Gpipe

Gpipe is a scalable pipeline parallelism library published by Google Brain, which allows for efficient training of large, memory-consuming models (Huang et al., 2018). Pipeline parallelism allows for §fast_nn_training.

In Gpipe, neural networks with sequential layers are partitioned across accelerators. The pipeline parallelism divides each input mini-batch into smaller micro-batches, enabling different accelerators to work on different micro-batches simultaneously. This is especially useful in §large_batch_training.

Bibliography

Huang, Y., Cheng, Y., Bapna, A., Firat, O., Chen, M. X., Chen, D., Lee, H., …, Gpipe: efficient training of giant neural networks using pipeline parallelism, CoRR, (), (2018).

Icon by Laymik from The Noun Project. Website built with ♥ with Org-mode, Hugo, and Netlify.