Application fails to start due to incompatible cuDNN version on Linux

[yusufipk]

Hello,

I encountered an error while trying to install and run the script on Linux. Script ran fine but when I tried to start the app I got an error message.

Traceback (most recent call last):
File "[...]/app/app.py", line 10, in <module>
from modules.whisper.whisper_factory import WhisperFactory
File "[...]/app/modules/whisper/whisper_factory.py", line 6, in <module>
from modules.whisper.faster_whisper_inference import FasterWhisperInference
File "[...]/app/modules/whisper/faster_whisper_inference.py", line 4, in <module>
import torch
File "[...]/app/env/lib/python3.10/site-packages/torch/init.py", line 239, in <module>
from torch._C import * # noqa: F403
ImportError: libcudnn.so.8: cannot open shared object file: No such file or directory

To resolve this issue, I had to manually install the following packages in the Conda environment:

1. conda install nccl
2. conda install cudnn=8.2

After installing these packages, I was able to run the application successfully.

It appears that the application requires a specific version of cuDNN (8.x) which conflicts with the version installed by default (9.x).

• OS: Linux

[jhj0517]

Hi. There's a version incompatibility between faster-whisper and torch >= 2.4 now.

I fixed it in jhj0517/Whisper-WebUI#318 by downgrading torch to 2.3.1, so it should work fine if you re install the app.

I just tested on pinokio and it works fine in my environment.

[jhj0517]

+) faster-whisper is now updated and requires CUDA 12.4 and torch>=2.5.0+cu121.
The latest whisper-webui works fine if you have CUDA 12.4 installed.

For more info:
SYSTRAN/faster-whisper#1086