There is no option pytorch for CUDA 12.2 and latest 12.3


I want to install pytorch with GPU in WSL Linux Ubuntu in my Windows computer. When I go to the start locally, it only has option for CUDA 11.8 and CUDA 12.1. I use CUDA 12.2, which option should I choose.

Minh Nguyen

The PyTorch binaries ship with their own CUDA runtime dependencies and you can select any version you want. For Hopper GPUs we recommend sticking to CUDA 12. If you need CUDA 12.3 for some reasons you would need to build from source.

Thank you very much.

“If you need CUDA 12.3 for some reasons you would need to build from source.”

How can I build from source? I don’t know what build from source means.

I will explain my problem from the beginning. I recently downloaded pytorch for AI work. This morning when I looked at pytorch, I saw that it was using the CPU (also I didn’t have CUDA on my computer). Then, I deleted all pytorch versions and all pytorch related packages from my computer, downloaded the latest CUDA (with CUDA toolkit) for my video card (RTX 3050 8GB) and got version “12.3.103” (aka “12.3”). But when I looked at pytorch’s own website, I saw that the latest version available for download is 11.8 CUDA. Is there a torch for version 12.3 CUDA? Where can I download it?

You don’t need a locally installed CUDA toolkit as PyTorch ships with all needed CUDA runtime dependencies as already mentioned. Only a properly installed NVIDIA driver is needed.