if varret is analogous to vararg, it should be appended after the previous args. So it'd be nice to do so for generality

sure, let's make it nice :-)

so the change here to align returning tuple, right? is it why all of the expect files change? are you sure it's a safe change?

I'm still getting confused by how we treat tuple return vs multiple args tbh. cc @suo @zdevito

after some digging, I think this diff is fine with existing function schema parser. Here is the reasoning:

1. when parsing the results, we automatically remove the outer pair of parentheses. related code This code also tells us multiple returns should be something like (int, float), one example is aten::divmod(int x, int y) -> (int, int) defined here

2. since we use the outer pair parentheses to represent multi-returns, to return a tuple, we should use two pairs of parentheses. My updates on the expect files are actually correct, because the returns are all single tuples, one example

The test in #23208 already make sure that all the serialized schemas can be correctly parsed (i.e., parsed schemas are exactly equal to the original ones.)

also executed the following code:

        s1 = parse_schema('any(Tensor self, int dim, bool keepdim=False) -> (Tensor, Tensor)')
        s2 = parse_schema('any(Tensor self, int dim, bool keepdim=False) -> ((Tensor, Tensor))')
        print(len(s1.returns))
        print(len(s2.returns))

output is

2
1

this also shows that, for tuple, we should use '(())'

Yeah, existing parsing is a bit non-pythonic (as python would treat it as a tuple). However, I think we can land it avoid changing existing parser and schemas