Is double dispatch automatically invoked for tensors to the appropriate kernels or does it have to be registered?
Could you give an example of your use case?
Are you writing a custom extension or are you referring to the PyTorch backend in general?
I am writing my own extension and I was wondering just for tensor creation torch.randn(2, 2, device)
is dispatch stub always needed?
I noticed that not all devices in the device type / backend enums are handled in dispatch stub.
Ah OK.
Sorry, I’m not familiar enough with the DispatchStub
implementation especially for new types.