Deprecate torch.stft returning real-valued tensors and torch.istft accepting real-valued inputs #55948
Description
For the PyTorch 1.9 release we can:
- update torch.stft to always return a complex tensor, and make setting return_complex=False a runtime error. The return_complex kwarg can remain on torch.stft and moved to kwarg-only, but it should no longer be documented. This will let current torch.stft users continue to use the op without disruption.
- update torch.istft to stop accepting real-valued inputs
And this should complete the deprecation process.