Wrapper subclass serialization should preserve all _make_wrapper_subclass() args about pytorch

[ezyang]

This probably needs a design doc

[jbschlosser]

From internal discussion: there is a preference towards avoiding actual serialization of dispatch-related args. We should store them as class properties instead -> query them from the cls arg within _make_wrapper_subclass(). This has a few benefits:

• No need to change the serialization format
• Decreased BC surface - serializing dispatch keys in particular is not something we want to do

We also discussed whether a default impl for subclasses that just works is valuable, as another option is to require a manual __reduce_ex__() impl for each new subclass. @albanD has the opinion that it is valuable and should be present; there are already quite a few things to think about for subclasses.

[ezyang]

What are you using the serialization for? Is it just multiprocessing?

[jbschlosser]

The serialization is used for both multiprocessing and torch.load() / torch.save():

pytorch/torch/_tensor.py

Lines 394 to 404 in 13462ec

	arg_wrapper_subclass = (
	type(self),
	self.dtype,
	tuple(self.size()),
	self.stride(),
	self.storage_offset(),
	self.layout,
	self.device,
	self.requires_grad,
	)
	return (torch._utils._rebuild_wrapper_subclass, arg_wrapper_subclass)