pytorch

1fae890a - fix grad silent correctness issue from view fn followed by an inplace fn (#85374)

fix grad silent correctness issue from view fn followed by an inplace fn (#85374) From https://github.com/pytorch/functorch/issues/1007, which was an issue where we would wrap aliases of unwrapped tensors and miss the inplace error message where we should have gotten it. Instead of keeping aliases unwrapped like I had originally wanted, this simplifies it slightly such that: (1) All tensors that were previously wrapped are still wrapped. This is occasionally important because of the 1-1 relationship between a tensor and autograd meta. By keeping the same number of wrapper tensors before, we'll never have autograd try to write multiple autograd metas to the same tensor when it wouldn't before (2) The tensors that either were unwrapped tensors or aliases of unwrapped tensors now get a flag on them (now called `alias_of_unwrapped`). This way, they are still wrapper tensors (and don't have to potentially break autograd) but we can identify that they should be treated like an unwrapped tensor Pull Request resolved: https://github.com/pytorch/pytorch/pull/85374 Approved by: https://github.com/zou3519