pytorch training guide
04/27/2022The PyTorch Training Performance Guide is an introduction to and reference on PyTorch training peformance optimization.
Many of the large-scale deep learning models used in production today have training times measured in days. As a result, techniques for training these models quickly and effectively are an important area of optimization in applied settings.
Techniques discussed in this guide include: mixed-precision training, distributed training, model pruning, gradient checkpoints, and just-in-time compilation.
This is a practitioner-oriented guide which grew out of a sequence of blog posts I wrote at Spell. It includes many ample code samples and benchmarks demonstrating the concepts discussed.