Pytorch and CUDA 11.4

Hello everyone, i’m aware that there are already a few post for this topic, but none of the answer i considered seems to help.

I’m trying to use pytorch on a machine with CUDA and CUDNN already installed.
I have not the permission to change the CUDA and CUDNN versions.
nividia-smi produces:
“NVIDIA-SMI 470.57.02 Driver Version: 470.57.02 CUDA Version: 11.4”

I can’t find the right pytorch version for using gpus.
“torch.cuda_is_available() returns False”.

I really need that system to be able to work.
A help is really needed, thank you.

I assume you are installing the PyTorch binaries and are selecting a proper CUDA runtime version.
If so, note that your local CUDA toolkit will only be used if you are building custom CUDA extensions or build PyTorch from source. The binaries will use their own CUDA runtime.

Based on the output of cuda.is_available() it seems you’ve either installed the CPU-only binaries (the install log would show it) or PyTorch is unable to communicate with the NVIDIA driver.