[Question]: How to install CUDA version of torchcodec on Windows using uv (not conda)?

[RakeshReddyKondeti]

Hi,
I am working on a project that requires the CUDA-enabled version of torchcodec on Windows. I am using a uv virtual environment (not conda). Is there a way to install the CUDA version of torchcodec using uv? If so, what is the correct command or configuration? If not, are there any workarounds for getting CUDA support for torchcodec in a uv/pip-based environment on Windows?

Thanks!

[NicolasHug]

Hi @RakeshReddyKondeti , we only ship CPU Windows wheels at the moment, not CUDA Windows wheels. Which means you can only install the CPU Windows version of TorchCodec with pip or uv. For the time being, only conda (with conda-forge) supports installing the CUDA version of TorchCodec Windows.

[traversaro]

I would also add that if you want a uv-like experience with a tool that can be install both conda packages and pip packages, you can look into pixi, see https://pixi.sh/dev/python/pytorch/ for some pytorch-related documentation that also applies to torchcodec.

Alternatively, you can try to compile torchcodec with CUDA support on Windows yourself, for reference the scripts that we use for that in conda-forge is available at https://github.com/conda-forge/torchcodec-feedstock/blob/main/recipe/build.bat , but that is a complex operation with a lot of things that can go wrong.

[earetaurus]

plus one was forced to switch to conda to use this and abbandon my rocm uv workflow kindly make it detect existing ffmpeg installs on windows