`aminmax` will trigger INTERNAL ASSERT if input is empty on cuda

[payphon]

🐛 Describe the bug

aminmax will trigger INTERNAL ASSERT if input is empty on cuda

import torch

def fn(input):
    return input.aminmax(dim=0)

device = "cuda"

input = torch.empty([36, 0], dtype=torch.float32)
input.uniform_(-16, 3)

fn(input.to(device))
# RuntimeError: iter.numel() > 0 && iter.ntensors() - iter.noutputs() == 1 && iter.noutputs() >= 1 INTERNAL ASSERT FAILED at "/opt/conda/conda-bld/pytorch_1663571524876/work/aten/src/ATen/native/cuda/Reduce.cuh":1134, please report a bug to PyTorch.

On cpu, it will just return empty tensor

Versions

torch: '1.13.0.dev20220919'

cc @pbelevich

[lezcano]

In fact, I think both the CPU and GPU implementations are wrong (how about that?). I think that, in this case, we should error out, as we do in amin and amax. See for example:

>>> torch.amin(torch.rand(3,0), dim=1)
IndexError: amin(): Expected reduction dim 1 to have non-zero size.

This is because the amin and amax functions do not have an identity element**.

** In fact they do have one. For amin the identity element would be +inf for floats and int_max for ints, but we resolved at some point that this would be an unexpected behaviour, so we figured it'd be best to throw an error.

[lezcano]

We would accept a PR fixing this issue :)

[payphon]

Yes, I think throwing an error is the correct way!

[nmacchioni]

In fact, I think both the CPU and GPU implementations are wrong (how about that?). I think that, in this case, we should error out, as we do in amin and amax. See for example:

>>> torch.amin(torch.rand(3,0), dim=1)
IndexError: amin(): Expected reduction dim 1 to have non-zero size.

This is because the amin and amax functions do not have an identity element**.

** In fact they do have one. For amin the identity element would be +inf for floats and int_max for ints, but we resolved at some point that this would be an unexpected behaviour, so we figured it'd be best to throw an error.

From the PyTorch docs it seems that RuntimeError is the appropriate response when the reduced dim has size 0 (https://pytorch.org/docs/stable/generated/torch.aminmax.html), in which case the CPU path needs to be fixed, but I do agree that torch.amin/torch.amax/torch.aminmax should be consistent in the returned error.

[lezcano]

Exactly, both the CPU and GPU paths should throw the same error, similar to the one in amin and amax.

[ShisuiUzumaki]

I would like to try my hand on this. Could you guide me on this?

[lezcano]

A good start would be to git grep for the amin() error to see how amin checks these things, and replicate these in the aminmax function. To see what are the possible entry points for the aminmax function for the different backends, please have a look at the native_functions.yaml. That's the file that binds the Python frontend and the C++ backend.

[ShisuiUzumaki]

@lezcano Okay, I will start with amin(). I am assuming that aminmax() behaviour should be same as amin() in case of empty tensor

[nmacchioni]

Yes, aminmax() should return an error similar to that returned from amin() when the reduced dimension is 0-dim.