[P] easy-torch-tpu: Making it easy to train PyTorch-based models on Google TPUs

About this article

I've been working with Google TPU clusters for a few months now, and using PyTorch/XLA to train PyTorch-based models on them has frankly been a pain in the neck. To make it easier for everyone else, I'm releasing the training framework that I developed to support my own research: aklein4/easy-torch-tpu This framework is designed to be an alternative to the sprawling and rigid Hypercomputer/torchprime repo. The design of easy-torch-tpu prioritizes: Simplicity Flexibility Customizability Ease o...