requirements.txt

[eftSharptooth]

I think it should be considered to have a requirements file, as a couple libraries need to be pinned to specific versions to function (mostly for training so far), whereas the current method installs (sometimes incompatible) newest versions. The one that specifically brings this to mind is librosa, which must be pinned around 0.9.2 instead of 0.10 due to the changing of the number of required arguments in 0.10. This causes the training to fail to start.

pip install librosa==0.9.2 fixes the error, just putting it on here for anyone else running into the issue.

[eftSharptooth]

So far I have as requirements the following: nvcc 12.1 or better for windows.

requirements.txt: torch>=2.0.0 torchvision einops fire>=0.6.0 huggingface-hub safetensors sentencepiece transformers tokenizers protobuf requests invisible-watermark diffusers einops diffusers transformers scipy numpy regex tqdm accelerate soundfile unidecode phonemizer torchlibrosa ftfy pandas timm matplotlib thop flash-attn==2.6.3 sentencepiece Pillow #Only required for creating your own datasets progressbar librosa==0.9.2 #required for audiollm, which apparently does not use torchlibrosa

[EnzoVLLC]

respeto maximo :P