Can Glow compile training of dynamically scheduled networks?

If I understand things correctly, Glow can be used to compile both the inference and the training of statically scheduled networks.
Can it also be used to compile the training of dynamically scheduled networks?

Can you define what exactly you mean by static and dynamic scheduling here? Mostly likely the answer is no – currently Glow does not support dynamic shapes or control flow, so in order to support e.g. different batch sizes you need to compile different models (though you could dynamically dispatch to different compiled models for different batch sizes at runtime).