Is torch2.4.1 compatible with CUDA12.5?

[Pied-Piper1]

I locally installed my CUDA Toolkit12.5. And everything went well (I can successfully call torch.cuda.is_available() and returns ture) before I install torchvision with cmd conda install torchvision -c pytorch . After installing torchvision, I even couldn’d import torch with error OSError: [WinError 127] Error loading "D:\anaconda3\envs\pytorch2.4.1\Lib\site-packages\torch\lib\torch_cuda.dll" or one of its dependencies. And torch 2.4.1+cu124,torchvision 0.19.1+cu124 are results of pip list | findstr torch. How can I do to tackle with this issue?

I wish if there exits a way that just requires degrading the cuda Toolkit version from 12.5 to 12.4 that works with torch and torchvision. Beacuse I don't want to remove torch and redownload it and just want to keep the current version of torch and torchvision .

[LuluW8071]

No, cuda version must match with the torch as per instructed in conda/pip docs.

You need to re install 12.4 version of CUDA. No need to reinstall torch again. Conda should have previous version of torch and other dependencies downloaded in cache so no need to download again. It should directly collect from cache wheel and install.

[Pied-Piper1]

Thanks!