It is mentioned in https://pytorch.org/2018/05/02/road-to-1.0.html that “exporting to C+±only runtimes for use in larger projects” will be supported in PyTorch 1.0. It is categorized as one of the "production-support"s.
But “production support” means only “forward pass” in common sense.
However, we may also use backward for online-training (fast adaptation of neural networks, such as MAML). Can it also be supported?