Torch script and eager-mode hybrid interface

Hi all,
While I am very familiar with PyTorch, I am a noob when it comes to torch.jit.
I have several questions:

  1. Is it possible to start writing the scripting module now and still train your model following the standard PyTorch training regime?
  2. does torch. Script support forward hooks?
  3. Is it possible to write a hybrid model with some sub-modules that are script/trace while others are just standard nn.Module and still enjoy some benefit?

I’m doing this because my model is still a work-in-progress, and many parts are subjected to change, while some are entirely static. So I want to test what speed-up I can get starting the transition now.