xla
xla copied to clipboard
Automatically move CUDA non XLA Tensors to XLA Device and back to CUDA device
Currently only works for inference. The assumptions don't hold for training with Autograd yet.
Is this an experimental pr or you want to merge this?
Is this an experimental pr or you want to merge this?
Ideally we would merge this, or is there a reason not to?
Sorry for being late. It's looking good!