RuntimeError: CUDA failed with error CUDA driver version is insufficient for CUDA runtime version

[Floydius]

1. Thank you for this tool!

2. I've been struggling with this for several days and can't solve. If this is not the right place, please let me know and I'll direct elsewhere.

faster-whisper seems to be working fine on cpu, but I cannot get it to run on gpu for anything.
based on everything I've found, my best guess is that i've got some combination of the packages below mismatched.
these are my currently-installed versions. (GPU is RTX 2070 super, 8gb)

faster-whisper 1.2.1
torch 2.9.1
ctranslate2 4.6.3
Nvidia Driver 580.95.05
CUDA Version 13.0
nvcc 12.0.140
cuBLAS 12.0.2.224~12.0.1-4build4 amd64
cuDNN 8.9.2.26~cuda12+3

Based on the requirements in the readme, Ctranslate2 only works with CUDA 12; does that mean >= 12, or 13 is too high?

I also found this issue.
it looks like that user was dealing with a different architecture, but i also saw a suggestion to set an environment variable: NVIDIA_VISIBLE_DEVICES

I did set mine to all, but no difference in result.

printenv NVIDIA_VISIBLE_DEVICES returns all

thanks in advance for any help!

[Purfview]

Try to install cuDNN 9.x

[Floydius]

Thanks for the suggestion.
Successfully updated to 9.17.1, but still getting that same error:

RuntimeError: CUDA failed with error CUDA driver version is insufficient for CUDA runtime version

going to try and downgrade my nvidia driver to get to CUDA 12, but if anyone else has ideas, any help is appreciated.

[Purfview]

Successfully updated to 9.17.1, but still getting that same error

Try to update cuBLAS too.

[Floydius]

if anyone has this working properly on Linux, would you be willing to share your versions for the packages I mentioned in the original post?

for convenience, how to get them:

pip show faster-whisper
pip show torch
pip show ctranslate2
nvidia-smi (this should give you Nvidia Driver and CUDA version)
nvcc --version
apt list --installed | grep cublas
apt list --installed | grep cudnn

[Floydius]

simplifying the question to see if anyone can help;

Does anyone have faster-whisper working with nvidia/cuda enabled on linux?

[monsieurlinux]

Similar problem here on Linux:

CUDA Version: 13.0
Driver Version: 580.126.09

nvidia-cublas-cu12: 12.8.4.1
nvidia-cudnn-cu12: 9.10.2.21
torch: 2.10.0+cu128 (but I don't think faster-whisper needs PyTorch?)

Maybe CUDA 13 is not supported yet?