reshape, broadcast_to, and permute_dims do not accept a single integer

[asmeurer]

reshape, broadcast_to, and permute_dims (I think these are all but I could have missed one), all accept a Tuple[int, ...]. The shape parameter for reshape and broadcast_to for shape, and axis for permute_dims

Every other function that accepts this kind of parameter accepts Union[int, Tuple[int, ...]].

It was pointed out in data-apis/array-api-compat#28 that NumPy currently accepts ints for all of these (and indeed, numpy.array_api accidentally does), but PyTorch does not.

[kgryte]

IIRC, the use of an integer is not the dominant use case and is more of shorthand in NumPy. In #42, we explicitly favored Torch (and somewhat TF) in restricting to a tuple for reshape.

For permute_dims, providing an integer does not make sense as a shorthand, except for the trivial case of a 1d array.

In general, I'd argue that a shape kwarg should be a tuple. If NumPy wants to support an integer shorthand, that's its prerogative, but I don't think this should be codified in the spec.

[rgommers]

Every other function that accepts this kind of parameter accepts Union[int, Tuple[int, ...]].

The only instances of shape parameters are for creation functions (ones & co) and manipulation functions (the ones you mentioned. The former accept ints, because ones(3) is a super common shorthand. The latter all require a tuple. So I'm inclined to agree with @kgryte here and say that the current state is fine, and preferred over allowing reshape(3).

[rgommers]

Since this was reviewed in gh-42, there are good reasons for this, and data-apis/array-api-compat#28 is already closed, let's close this one too.

Thanks @asmeurer and @kgryte