I guess we should add tests if this is the official C++ API (I'm not sure if it is documented in this level of detail).

Looking at #20866, the JIT does extra work to expand single integer arguments into lists, see e.g.:

@weak_script
def _max_pool2d(input, kernel_size, stride=None, padding=0, dilation=1,
                ceil_mode=False, return_indices=False):
    # type: (Tensor, BroadcastingList2[int], Optional[BroadcastingList2[int]], BroadcastingList2[int], BroadcastingList2[int], bool, bool) -> Tensor  # noqa
    if stride is None:
        stride = torch.jit.annotate(List[int], [])
    return torch.max_pool2d(
        input, kernel_size, stride, padding, dilation, ceil_mode)

So the JIT takes care not to pass single integers to torch::max_pool2d. The question I have is whether it does so for internal convenience or because the C++ API was originally intended that way.

[I think it is reasonable to permit single integers in the C++ API, but it would be nice to sync behavior with the JIT.]

I guess we should add tests if this is the official C++ API.

Good point! I am inclined to think C++ should have the same behavior as Python, i.e., auto-expanding kernel_size, but I am not sure either. @gchanan has this been discussed before?

Agreed, I also think that C++ should have the same behavior as Python, mostly because a lot of other Python idioms work on the C++ level.

I think ideally it would work the same way in C++.

I'm landing the PR. if you want to send tests, please do so in a follow-up PR

@mrshenli merged this pull request in 2372e7e.

Yes, tests will be part of #22277 (more functions need to be changed in ATen).