Can't get cuda to work.

[manus693]

Whatever I do I get Torch not compiled with CUDA, I've followed the instructions and installed as written.

miniconda3\envs\whisperx\lib\site-packages\torch\cuda_init_.py", line 310, in _lazy_init
raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

When running pip install whisperx it installs torch without cuda enabled. I'm running this inside the conda environment.
I'm not really sure how the get this to work, been trying for ages now.
It also install torch 2.5.0, but the conda install is 2.0.0 before the "pip install whisperx" in the description.

Is Setup in description outdated?

Is there any difference except maybe speed? CPU or Cuda?
I ran whisperx on a movie with int8 enabled, and it was almost correct, some timestamps where completely wrong, just random places.
I could not get float 16 to work, trying CPU and float 32 now.

I have 4060ti.

[mukhituly]

+1, experiencing this issue as well

[NefariousC]

Ran into the same issue, idk but there seems to be a pip bug where installing/updating whisperx automatically pulls the CPU version of torch instead of the CUDA version, regardless of what you had before.

You can try this (bear in mind that i'm on CUDA 12)
This will force reinstall torch and torchaudio with CUDA support along with their dependencies:

pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-cache-dir

• After that, you might run into another error when using whisperx :

Warning

Could not locate cudnn_ops_infer64_8.dll. Please make sure it is in your library path!

1. Download the cuDNN library from nvidia website (for windows)

Important

Since CTranslate 4.4.0 only supports up to cuDNN 8.x, make sure to grab the 8.x version compatible with CUDA 12

2. Locate the files inside the downloaded zip bin folder then extract them to your whisperx environment bin folder ...\envs\whisperx\bin

I think there's a simpler solution somewhere but after searching for hours this is what I got.

[ruze00]

So CUDA 12.1 works? I've been holding off on even trying to install whisperx since it says it supports 11.8 and other versions are YMMV. I'm not going to downgrade my whole env for 1 component.

[Matagon]

This works, FYI, the kernel will die if ran from a Jupyter Notebook.

[NefariousC]

@ruze00 So CUDA 12.1 works? I've been holding off on even trying to install whisperx since it says it supports 11.8 and other versions are YMMV. I'm not going to downgrade my whole env for 1 component.

yes, it works. i've been using it with no problem.

[ruze00]

@ruze00 So CUDA 12.1 works? I've been holding off on even trying to install whisperx since it says it supports 11.8 and other versions are YMMV. I'm not going to downgrade my whole env for 1 component.

yes, it works. i've been using it with no problem.

Thanks, appreciate it.

[zos474]

Thanks @NefariousC your solution worked. I have CUDA 12.5 installed on my Windows 11 machine and had a perfectly working WhisperX 3.1 Conda environment until I decided stupidly to do an upgrade to 3.3.1. I ran into one problem after another, eventually deciding to trash the environment and create a new one, but of course the standard install doesnt work, because even though you have installed torch 2.0.0, the install of whisperx from PyPI overwrites it with 2.6.0 and it is the cpu version... But your steps worked exactly. I made two slight mods: Installed 12.4 pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu124 --force-reinstall --no-cache-dir and copied the cudnn dll files directly into .\miniconda3\envs\whisperx not into a bin subdirectory. Thanks so much. I had wasted most of Sunday trying to get it to all work again. I guess the lesson is dont upgrade what aint broke.

[NefariousC]

@zos474 No worries, I was in the same boat as you. I encountered some issues like OP after blindly upgrading WhisperX. After running into errors even after reinstalling, I had to dig into Google and see what this error was about:

miniconda3\envs\whisperx\lib\site-packages\torch\cuda_init_.py", line 310, in _lazy_init
raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

I was so confused because I thought I DID install the CUDA version, per what the setup docs said:

2. Install PyTorch, e.g. for Linux and Windows CUDA11.8:
   conda install pytorch==2.0.0 torchaudio==2.0.0 pytorch-cuda=11.8 -c pytorch -c nvidia

After surfing through forums, I found out about the pip bug.

Then I remembered WhisperX uses faster-whisper backend, which needs CTranslate2. Went back to the WhisperX docs and found this important bit:

Important

GPU execution requires the NVIDIA libraries cuBLAS 11.x and cuDNN 8.x to be installed on the system. Please refer to the CTranslate2 documentation

Looks like I missed that tidbit when I was just yolo-ing the update. Lesson learned for me too. Have to thoroughly read the documentation before updating I guess 😂.

[jfsbra]

Hi All.
I have been trying to get WhisperX to run on my Win11 PC for days now... Simply no joy!
Whatever i do seems to break things because of dependency incompatibilites.
I'm wondering if someone who has whisperx running correctly could provide a list of their installation (e.g. pip freeze) so I could try to replicate the install - preferably in a Docker container or conda environment.
Thanks a million!

[UdinaDev]

@jfsbra

Had similar dependencies issue yesterday.
Now it's working but my solution is for whisperx==3.1.1 as it has all the features I need

Create venv with your python 3.10 :

python3.10 -m venv whisperx311

Then install whisperx==3.1.1 :

E:\Codes\virtual_envs\whisperx311\Scripts\python -m pip install whisperx==3.1.1

Create following requirements.txt :

--extra-index-url https://download.pytorch.org/whl/cu118
torch==2.0.0+cu118
torchaudio==2.0.1+cu118
torchvision==0.15.1+cu118
faster-whisper==0.10.1
transformers
ctranslate2<4.5.0
pandas
setuptools==65.6.3
nltk
pytorch_lightning==2.0.1
lightning==2.0.1
numpy==1.26.4

Then

E:\Codes\virtual_envs\whisperx311\Scripts\python -m pip install -r requirements.txt 

At this point you should have no dependency resolver errors and a working whisperx venv (whisperx===3.1.1)