Fixed SVD ignoring "some/full_matrices" flag for empty inputs (#51109)
Summary:
For empty inputs `torch.svd` (and `torch.linalg.svd`) was returning incorrect results for `some=True` (`full_matrices=False`).
Behaviour on master branch:
```python
In [1]: import torch
In [2]: a = torch.randn(0, 7)
In [3]: a.svd()
Out[3]:
torch.return_types.svd(
U=tensor([], size=(0, 0)),
S=tensor([]),
V=tensor([[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.]]))
In [4]: a.svd(some=False)
Out[4]:
torch.return_types.svd(
U=tensor([], size=(0, 0)),
S=tensor([]),
V=tensor([[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0.]]))
```
`some` flag is ignored and 7x7 `V` matrix is returned in both cases. `V` should have 7x0 shape when `some=True`.
This PR fixes that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51109
Reviewed By: ngimel
Differential Revision: D26170897
Pulled By: mruberry
fbshipit-source-id: 664c09ca27bb375fabef2a046d0a09ca57b01aac
