Brian Hirsh

Hey! The API run at https://github.com/pytorch/xla/blob/5a113aff98ce42420891c724843ccb30691dc24a/torch_xla/csrc/aten_xla_type.cpp#L3639-L3645 is just re-invoking the dispatcher, and attempting to run whatever decomposition for `pixel_shuffle` is registered there. My guess for the python vs C++ difference...

@wonjoolee95 the way I would categorize the decomps we have is that: (1) there are a ton of decomps, for many (most?) ops in ATen. A compiler backend to torch.compile...

I updated the PR with @soulitzer's idea to not add a new API: it seems like `torch._C._functions.DelayedError` should do the job. Adding that ErrorNode also caused some test failures, which...

Hmm... the errors I'm getting are because `y = torch._C._functions.DelayedError(1)(x)` don't "propagate" the fact that x is a fake tensor (it returns a plain tensor)

oh it's probably because the SparseTensorImpl C++ subclasses don't override shallow_copy_and_detach properly... We probably are not able to handle "tensor subclass holding a fake sparse tensor" today for other reasons,...

cc @albanD who would e a good person to look at custom backend + distributed changes?

@pytorchbot merge

More context here: https://github.com/pytorch/pytorch/pull/122434#issuecomment-2087794247

Turned into an error by https://github.com/pytorch/pytorch/pull/122434

@pytorchbot merge