torch.min, torch.max, torch.median and torch.mode return indices along with the values when reducing over one dimension. This is incompatible with the Python Array API Standard as described in the following issue #58745 and furthermore makes it difficult to support reducing over multiple dimensions.

Alternatives

• a dedicated operator for computing the extra output such as argmax that only returns that extra output
• a variant of the function, e.g. argmax_with_indices, that returns all output tensors
• an optional parameter return_indices=True to explicitly ask for the extra output, although it is discouraged to have operators with multiple possible return types
• an optional out parameter on which to place the extra output, this avoids the return type changing and creating a combinatorial number of variants of the operator. We already support the out= keyword so it might be possible to extend the code to accept extra optional out tensors.

cc @mruberry @rgommers @heitorschueroff @pmeier @asmeurer @leofang @AnirudhDagar @asi1024 @emcastillo @kmaehashi