Inconsistent constant_pad_nd behavior with negative padding #161014
Description
🕵️♂️ Detected with FACTO
🐛 Describe the bug
The aten op constant_pad_nd allows for negative padding, as long as the absolute value of the padding is smaller than the size, and as long as the resulting output doesn't have negative size
For example:
>>> torch.ops.aten.constant_pad_nd.default(torch.ones([5, 3]), [-1, -2]).shape
torch.Size([5, 0])Notice that the -1, -2 padding applies to the last dimension
We can also pad the first dimension, by adding two more pads:
>>> torch.ops.aten.constant_pad_nd.default(torch.ones([5, 3]), [-1, -2, 0, 0]).shape
torch.Size([5, 0])
>>> torch.ops.aten.constant_pad_nd.default(torch.ones([5, 3]), [-1, -2, -1, -1]).shape
torch.Size([3, 0])However, in the following case, it throws an error:
>>> torch.ops.aten.constant_pad_nd.default(torch.ones([5, 3]), [-1, -2, 1, 1]).shape
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/Users/mcandales/miniconda3/envs/facto/lib/python3.11/site-packages/torch/_ops.py", line 840, in __call__
return self._op(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: The input size 3, plus negative padding -1 and -2 resulted in a negative output size, which is invalid. Check dimension 1 of your input.Notice that, if we make sure the output won't have size 0, we always avoid the bug:
>>> torch.ops.aten.constant_pad_nd.default(torch.ones([5, 3]), [-1, -1, 1, 1]).shape
torch.Size([7, 1])Versions
nightly
cc @malfet