Result for torch.pow mismatch between PyTorch eager and PyTorch/XLA.

[ysiraichi]

🐛 Bug

The following cases to torch.pow result in mismatched results, when comparing with PyTorch eager:

>>> x = torch.tensor([-3, -2, -1,  0,  1])

>>> torch.pow(x, -1)
# > Expected: RuntimeError: Integers to negative integer powers are not allowed.
# >   Actual: tensor([ 0,  0, -1, -1,  1], device='xla:0')

>>> torch.pow(x, torch.tensor(-1))
# > Expected: tensor([ 0,  0, -1,  0,  1])
# >   Actual: tensor([ 0,  0, -1, -1,  1], device='xla:0')

>>> torch.pow(0, torch.tensor(-1))
# > Expected: tensor(0)
# >   Actual: tensor(-1, device='xla:0')

Interestingly, the result of these other cases match:

>>> torch.pow(x, -1 * torch.ones(x.shape))
>>> torch.pow(0, -1 * torch.ones(x.shape))

Here are some observation:

• 0 ** (-1) should be inf. Not 0 nor 1
• The results seem to mismatch when XLA has to broadcast the result (?)

Environment

• Reproducible on XLA backend [CPU/TPU/CUDA]: CUDA
• torch_xla version: 99c41ba

cc @miladm @JackCaoG