Why does WhisperX seem to install non-CUDA pytorch?

[FriedGenera]

When I'm using pip to install WhisperX it seems to install the cpu version of Pytorch 2.8.0 instead of the CUDA version even though I have CUDA 12.8 installed? Also is it necessary to install CUDA beforehand since the Pytorch installation usually installs it's own CUDA libraries?

[Bibblebrig]

I seem to have this happen as well. With the base install of whisperx (having already installed the CUDA 12.8 compatible version of Pytorch prior) I get the line

 ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchvision 0.24.1+cu128 requires torch==2.9.1, but you have torch 2.8.0 which is incompatible.

But when I attempt to reinstall the CUDA version I'm given

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
whisperx 3.7.4 requires torch~=2.8.0, but you have torch 2.9.1+cu128 which is incompatible.
whisperx 3.7.4 requires torchaudio~=2.8.0, but you have torchaudio 2.9.1+cu128 which is incompatible.

[guslatho]

Had the same issue - downgrading to whisperx 3.7.2 seemed to fix the issue for me

[arcticloud]

Two separate questions in there:

Why CPU torch? The default torch wheel on PyPI is the CPU build, so a plain pip install whisperx (which pulls torch as a dependency) gives you CPU torch unless you install the CUDA build from PyTorch's own index. Do that first, then whisperX:

pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install whisperx

The ERROR: pip's dependency resolver... message the other commenter saw is just pip noting that whisperX's pin disagrees with your pre-installed cu128 torch — usually harmless. If pip actually tries to replace your CUDA torch with the CPU one, install whisperX with --no-deps (and add its remaining deps manually) so it can't override torch.

Do you need CUDA installed beforehand? No. The PyTorch CUDA wheels bundle their own CUDA runtime, so you don't need a system CUDA toolkit — only a recent enough NVIDIA driver. (Heads-up: whisperX's transcription backend is CTranslate2/faster-whisper, which needs cuBLAS + cuDNN; those ship with the ctranslate2 wheel, but if inference fails with a cuDNN error it's usually a driver/version mismatch, not a missing toolkit.)