Hey @ysiraichi, thanks for the ping and sorry to make you wait for a review. Now let's talk about as_strided and empty_strided and use this opportunity to correct and modernize their documentation.

First, I don't think the documentation for this function is correct, and that's part of what has made this PR hard to review. Let's look at an example program:

t = torch.tensor((1, 2, 3, 4, 5, 6, 7, 8, 9, 10))[::2]
# t = tensor([1, 3, 5, 7, 9])
v = torch.as_strided(t, size=[2], stride=[1])
# v = tensor([1, 2])

In what sense is v a view of t? It contains a value not found in t!

And when we look at the implementation of as_strided this makes sense:

because it seems like as_strided creates a view of :attr:`input`'s storage, not a view of :attr:`input`. So first we need to explain that.

Now let's talk about warning. It's trying to tell us two things:

• creating an "overlapped" tensor where multiple indices refer to the same element is a bad idea
• you probably shouldn't use this function

This seems like a strange warning to me, so let's dig into its history: #22842 (comment). Seems like there were concerns about exposing as_strided at all, which is reasonable, but I think we can do a better job with this warning.

What would you think of changing the warning to this, instead:

"Prefer using other view functions, like :meth:`torch.Tensor.expand`, to setting a view's strides manually with `as_strided`, as this function's behavior depends on the implementation of a tensor's storage. The constructed view of the storage must only refer to elements within the storage or a runtime error will be thrown, and if the view is "overlapped" (with multiple indices referring to the same element in memory) its behavior is undefined."

This more clearly communicates why this function shouldn't be used -- because storage is an implementation detail and relying on it is inherently risky and strange for users. This also simplifies the other warnings.

That sounds way better! Thanks, @mruberry. I will change it right away!

In light of the above discussion we should carefully update the documentation of empty_strided, too.

Let's start with the first paragraph:

"Creates a tensor with the specified :attr:`size` and :attr:`stride` and filled with undefined data."

This modernizes the language, preserves this PR's fix, and doesn't include implementation details like how much memory is allocated.

I don't know if a note about about how much memory an implementation of this function could use is interesting to users. So maybe we could drop it?

Now let's talk about the warning for this function. This warning is intended to say "don't create an overlapped tensor", and we can do that more easily (following the above):

"If the constructed tensor is "overlapped" (with multiple indices referring to the same element in memory) its behavior is undefined."

We don't declare "overlapped" behaviors as having undefined behavior in a note (yet), but we probably should.

I agree. The amount of allocated memory doesn't sound that useful here.