Fastai on TPU?

[ilovescience (Tanishq)]

This is a thread for fastai v1:

How to create a callback using torch.multiprocessing (TPU)

I am creating a callback for distributed TPU training based on the pytorch XLA library. It will require the training loop to be passed into a multiprocessing spawning function to spawn training loops on each of the TPU cores. I have worked with the callbacks before but I am unsure if it is possible to change the training loop with a callback. Is this possible? @sgugger any input?

Development for fastai v2 is slow but ongoing.