CUDNN with RMVPE pitch

[Hayri-maker]

When I have to do an inference in RVC, I have to turn off CUDNN with the following line:
torch.backends.cudnn.enabled=False
It seems that RMVPE does not support CUDNN with excessively long files. I am using the rtx 5090 version.