[6.3 Branch] Training Error on CPU

Hello! I've attempted to train a test reflow model on the 6.3 version branch here using the standard Command Prompt in Windows 10; using Python 3.10.10; with pip 24.0 installed and setuptools and wheel upgraded to their latest versions as of May 19, 2025; and torch 2.4.1 and torchaudio 2.4.1 installed, both on CPU version. In this attempt, I used ContentVec as my Encoder, NSF_HiFiGAN as my Vocoder, and RMVPE as my Pitch Extractor. I implemented the torch.load change in encoder/rmvpe/inference.py from a response I received in Issue #132. I also modified configs/reflow.yaml to: - set the device to cpu in line 31, - set cache_all_data to false in line 45, - set training and validation data paths to different folders to organize data in lines 15 and 16; ...As well as modify reflow/solver.py to import GradScaler from torch.amp instead of torch.cuda.amp in line 9. Fortunately, preprocessing data actually worked! However, I have seemingly one more obstacle to overcome before I get to train the model. Here's my traceback: > Traceback (most recent call last): > File "F:\AI Projects\DDSP-SVC-6.3\train_reflow.py", line 75, in <module> > train(args, initial_global_step, model, optimizer, scheduler, vocoder, loader_train, loader_valid) > File "F:\AI Projects\DDSP-SVC-6.3\reflow\solver.py", line 232, in train > ddsp_loss, reflow_loss=model(data['units'], data['f0'], data['volume'], data['spk_id'], > File "F:\AI Projects\DDSP-SVC-6.3\_ddsp63-win10\lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_impl > return self._call_impl(*args, **kwargs) > File "F:\AI Projects\DDSP-SVC-6.3\_ddsp63-win10\lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_impl > return forward_call(*args, **kwargs) > File "F:\AI Projects\DDSP-SVC-6.3\reflow\vocoder.py", line 195, in forward > ddsp_wav, hidden = self.ddsp_model(units, f0, volume, spk_id=spk_id, spk_mix_dict=spk_mix_dict, aug_shift=aug_shift, infer=infer) > File "F:\AI Projects\DDSP-SVC-6.3\_ddsp63-win10\lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_impl > return self._call_impl(*args, **kwargs) > File "F:\AI Projects\DDSP-SVC-6.3\_ddsp63-win10\lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_impl > return forward_call(*args, **kwargs) > File "F:\AI Projects\DDSP-SVC-6.3\ddsp\vocoder.py", line 586, in forward > src_filter = torch.exp(ctrls['harmonic_magnitude'] + 1.j * np.pi * ctrls['harmonic_phase']) > RuntimeError: "exp_vml_cpu" not implemented for 'ComplexHalf' I believe that's everything I remembered doing in this attempt. Thank you for reading, and have a good (rest of your) day!