Add error message for complex alpha and non-complex inputs (#54964)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54964
Previously, the following would error out with a strange error message:
```
import torch
x=torch.randn(2)
torch.rsub(x, 1, alpha=2j)
Traceback (most recent call last)
<ipython-input-2-caf2a1c03d0b> in <module>
1 import torch
2 x=torch.randn(2)
----> 3 torch.rsub(x, 1, alpha=2j)
RuntimeError: value cannot be converted to type float without overflow: (-0,-2)
```
The reason why this is happening is because the alpha check doesn't check for if `x` is not complex and `alpha` is complex.
The error gets thrown further along in the implementation of torch.sub,
when it coerces `alpha` to be the same dtype as the input tensor:
https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/cpu/BinaryOpsKernel.cpp#L53
This PR fixes the bad error message by adding a new check to the alpha check.
Test Plan:
- pytest test/test_binary_ufuncs.py
- NB: add, sub, and rsub all share the same alpha check. The test only tests it for torch.add, but that should be sufficient.
Reviewed By: gchanan
Differential Revision: D27504017
Pulled By: zou3519
fbshipit-source-id: 70b9aa75a7a4faaaa93f6ba235cae85998a91697
