Pytorch 2.1.0 causes a seg fault when running whisper in a container.

[neilshevlin]

To start: Whisper has no issue running in a container when using the stated torch version 1.10.1

However the latest version of torch(2.1.0) causes a 139 SIGSEGV fault. The issue is not related to the download or write of .pt models for whisper. Whisper by default will download and write the model to ~/.cache/whisper. Improper permissions can cause the segfault; So to avoid this, you can configure with proper write access in your container.

The issue, however, is that Pytorch 2.1.0 causes this fault. The issue pops up in the encoder forward pass when doing a convolution with the spectrogram tensor. Specifically, pytorch seems to be failing when doing the conv1d operation when it hands off to torch.F.conv1d().

This becomes an issue with Pytorch 2.1.0; Works fine with 2.0.1.

[wredan]

This issue also occurs with the PyTorch 2.6.0 release. I had to roll back to version 2.5.1 to resolve it. Hope this helps!