subclass_zoo

subclass_zoo copied to clipboard

Aka, https://github.com/pytorch/pytorch/issues/69991 ## Motivation We should make sure all PyTorch operations are "Composite Compliant". This condition is necessary for operators that look out-of-place to preserve the Tensor Subclass when passed...

zou3519

https://github.com/pytorch/pytorch/issues/62884 this is a footgun for python keys who don't expect to see conjugate values

ezyang

Schema access needs to get finished and then dogfooded to see if there are gaps

1

In progress PR at https://github.com/pytorch/pytorch/pull/72729

ezyang

torch function is public API while torch dispatch is not, it may be friendlier to people if they can program against torch function api.

ezyang

https://github.com/pytorch/pytorch/pull/73727 is an attempt at doing this it ran into problems. Fortunately, we have an updated plan at https://github.com/pytorch/pytorch/pull/73727#issuecomment-1059216810 that just needs to be implemented

ezyang

There is a PR at https://github.com/pytorch/pytorch/pull/74469 but it's wrong and we need a better strategy

ezyang

Simple functorch suggests that these are the wrong APIs https://colab.research.google.com/github/albanD/subclass_zoo/blob/main/simple_functorch.ipynb but due to how the dispatcher is implemented it is not easy to prevent people from accidentally going back to...

ezyang

https://github.com/pytorch/pytorch/pull/73942 is blocked because torch function gets disabled in the default implementation and then if you fall through to Python key it is still disabled, which breaks FX tracing. So...

ezyang

This should be as simple as just swapping from CompositeImplicitAutograd to explicit; they are de facto explicit already because there is a codegen special case.

ezyang

In complex numbers example, mm sample impl seems just regular complex number multiplication impl and not mm

1

https://github.com/albanD/subclass_zoo/blob/main/complex_tensor.py#L27-L31 right?

vadimkantorov