| import torchaudio | |
| import torch | |
| from vui.inference import render | |
| from vui.model import Vui | |
| model = Vui.from_pretrained().cuda() | |
| model.decoder = torch.compile(model.decoder, fullgraph=True, mode="max-autotune") | |
| for i in range(10): | |
| waveform = render( | |
| model, | |
| """Hey, here is some random stuff, usually something quite long as the shorter the text the less likely the model can cope! | |
| So cool yeah makes sense, would you be able to help me with something? | |
| Sure what is it?""", | |
| ) | |
| print(waveform.shape) | |
| torchaudio.save("out.opus", waveform[0].cpu(), 22050) | |