Unable to load any of {libcudnn_cnn.so.9.1.0, libcudnn_cnn.so.9.1, libcudnn_cnn.so.9, libcudnn_cnn.so}

[Leandrocnf]

Please, help me. I don't know much about programming, but I use whisperx via google colab because it helps me a lot with my work subtitling, but yesterday it started giving me this message:

/usr/local/lib/python3.10/dist-packages/pyannote/audio/utils/reproducibility.py:74: ReproducibilityWarning: TensorFloat-32 (TF32) has been disabled as it might lead to reproducibility issues and lower accuracy.
It can be re-enabled by calling

import torch
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
See pyannote/pyannote-audio#1370 for more details.

warnings.warn(
Unable to load any of {libcudnn_cnn.so.9.1.0, libcudnn_cnn.so.9.1, libcudnn_cnn.so.9, libcudnn_cnn.so}
Invalid handle. Cannot load symbol cudnnCreateConvolutionDescriptor

I don't know what happened. I didn't change anything. I always use a simple command, it's just:

! whisperx (path to the videofile) --model large-v3 --output_dir (path where the file will be saved) --language en --output_format srt

I tried using another model, but it's still not working and I can't really understand the error message to try to fix whatever the problem is. If anyone could help me, I really need this to make my work easier.

[chriskathumbi]

Same issue here. Started today

[DSobrino]

I encountered the same issue. A temporary fix can be found here: SYSTRAN/faster-whisper#1083

In my case, I added the following to my requirements.txt, and it is working now:

ctranslate2>=4.0,<=4.4.0

[Leandrocnf]

I solved by downgrading ctranslate 2. I used this command:

pip install ctranslate2==4.4.0

[chriskathumbi]

This works. Thanks

[CREESTL]

Thanks, @Leandrocnf, this helped me:

pip install ctranslate2==4.4.0

[Jenniuss]

Thanks, @Leandrocnf, it works!!!

[ak4off]

didn't work

[jon-fox]

Didn't work for me either :/

ctranslate2==4.4.0
whisperx==3.1.5

[bhautik-pithadiya]

@jon-fox This error occurred because your cuDNN file is not set in the environment.

1. Ensure cuDNN is Installed Correctly
   Verify that the cuDNN library files are present on your system. They are usually located in the directory where CUDA is installed. For example:
   -> /usr/lib/x86_64-linux-gnu/ (Debian-based systems)
   -> /usr/local/cuda/lib64/

   Run the following command to locate cuDNN files:

    find /usr -name 'libcudnn*'

If the files are missing, you need to reinstall cuDNN.

2. Check Library Paths
   Ensure the paths to the cuDNN libraries are correctly set in your environment variables:

export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH

Copy these lines and add them to ~/.bashrc or ~/.zshrc. Then run:

source ~/.bashrc

These steps worked for me.

[jon-fox]

Yeah I've tried these with no luck. My app was working for a few months with whisperx until the lib updates.

I ended up just switching over to using faster whisper and am not having any issues.

[GUUser91]

I was able to fix this problem while using ctranslate2 version 4.5.0. First I create a virtual python environment.

python3.11 -m venv venv

Then I activate the environment.

source /home/user/WhisperX/venv/bin/activate

I install whisperx

pip install git+https://github.com/m-bain/whisperx.git

Install ctranslate2 4.5.0

pip install ctranslate2==4.5.0

Then I export the cuDNN libraries that are located in the virtual python enviroment

export LD_LIBRARY_PATH=/home/user/WhisperX/venv/lib/python3.11/site-packages/nvidia/cudnn/lib/

Now I can launch it.

whisperx --model large-v3 --language en The.Prime.of.Miss.Jean.Brodie.1969.flac

I open the bashrc file to create a alias command so I make a shorcut out of it.

wx='source /home/user/WhisperX/venv/bin/activate && export LD_LIBRARY_PATH=/home/user/WhisperX/venv/lib/python3.11/site-packages/nvidia/cudnn/lib/ && whisperx --model large-v3 --language en'

I'm using torch and torchaudio version 2.6.0 and cuda 12.8.0-1 on my artix linux distro.