why can’t pytorch 2.7.1 use system memory?
Your question is not specific enough to answer properly, so please add more details.
PyTorch can use host RAM and will do so when the default cpu device is used e.g. in torch.empty(1).
I mean when I convert a file with a voice that I have cloned, it will only use my GPU memory. In pytorch 2.0 it could allocate more ram for larger mp3 files. This means that I can’t convert a file of 16 min in the newer pytorch.
PyTorch won’t use the GPU by default unless your script specified the default device to be cuda or moved data explicitly to the GPU. I also assume you are using torchaudio or another 3rd party package to load these files? If so, could you share a minimal and executable code snippet reproducing this behavior?
try downloading the 2 versions of RVC and try the file that I put up in dropbox. I don’t know how to mess with scripts, as I use a screen reader.
rtx 40:
https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC20240604Nvidia.7z?download=true
rtx 50:
mp3 file: