PyTorch Version Inconsistency: VAE Music Reconstruction Fails in 2.6.0 but Works in 2.8.0

[Zjy0401]

Thank you for open-sourcing such an amazing model! While testing the reconstruction of the VAE model, I found that in the torch 2.6.0 environment, it couldn't reconstruct music properly. However, in the 2.8.0 environment, the reconstruction quality was perfect. I just wanted to ask, what might be the reasons for this phenomenon? Thank you!

[tencent-ailab]

Could you share some bad cases? We haven't encountered this situation before, or is it a dependency conflict issue?

[Zjy0401]

Sure. I only change my torch 2.8.0 conda environment(which is properly) by use pip install torch==2.6.0 torchaudio==2.6.0 torchvision==0.21.0 -i https://pypi.tuna.tsinghua.edu.cn/simple The reconstruct bad cases is like：

004162_seg9_torch2.6.mp3

The input case is like:

004162_seg9.mp3

When I change my torch version to 2.8, the output case is like:

004162_seg9_torch2.8.mp3

I think these samples have the same problem in torch 2.6.0.

[tencent-ailab]

The example is just noise, it looks like the model parameters aren't being loaded correctly. However, we didn't reproduce this problem in torch 2.6.0 yet.