Incorrect info about overriding torch tensors in version 1.7.0

[jemisjoky]

📚 Documentation

I've been working on extending the torch.Tensor class, where the information provided in the torch._overrides has proven really helpful! After recently upgrading to torch version 1.7.0 however, I noticed that the attribute _overrides is no longer present in torch. Two questions/comments about this:

1. Is the information that used to be in torch._overrides still present somewhere else in the library?
2. The documentation for torch 1.7.0 still refers to torch._overrides, so should be updated with the correct info.

Thank you for your help with this :)

cc @ezyang @gchanan @zou3519 @bdhirsh @hameerabbasi @rgommers @jlin27 @mruberry

[gchanan]

Thanks for pointing this out @jemisjoky, it is indeed an oversight. It looks like torch._overrides is now torch.overrides and we should update the documentation.

[jemisjoky]

Haha if I had just looked a bit further down in dir(torch) it would have been obvious where it had moved to. Thanks a bunch for that info @gchanan, super helpful!

[rgommers]

@hameerabbasi did we consciously decide to move to the non-underscored name? torch.overrides isn't in the docs as a public submodule, so that's a little inconsistent.

[hameerabbasi]

Yes, this was done in #37091, specifically in b422a1b, in the context of there being various utility methods there that were already well-tested and needed to be exposed, such as, for example, is_tensor_method_or_property. I can't seem to find the conversation, maybe it happened in a meeting.

[hameerabbasi]

I've created the docs issue in #48087.