I understand that backward passes are serialized for a particular model with multiple loss tensors for example. But if I have independent modules can I run backward pass in parallel?
Hi,
No you cannot do that at the moment I’m afraid.
Can you share more about your use case and why it would be interesting for you?