I have a GPU and CUDA installed in Windows 10 but Pytorch's torch.cuda.is_available() returns false; how can I correct this?

Question

I have PyTorch installed on a Windows 10 machine with a Nvidia GTX 1050 GPU. I have installed the CUDA Toolkit and tested it using Nvidia instructions and that has gone smoothly, including execution of the suggested tests.

However, torch.cuda.is_available() returns False. How can I fix this?

Answer 1

I just spent about an hour fighting this problem, breaking down and building up at least four different conda environments. I finally got something to work using the same matrix selector at their web site but selected conda, because conda seems to be working hard on getting a conda installation to work. And it worked! Here is the command that worked in my conda command line:

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

Obviously these details may change, but the command generated worked! Such a great feeling to see that True after all this time.

Answer 2

I had the same issue, and it turned out that I installed a CPU-only version by running the command provided by https://pytorch.org/get-started/locally/.

If you have CUDA 10.2 installed like me, the website would likely give you pip install torch===1.7.1 torchvision===0.8.2 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html, which doesn't explicitly specify CPU or GPU.

In my case, it routed me to cpu/torch-1.7.1%2Bcpu-cp38-cp38-win_amd64.whl, which is a CPU version, instead of cu102/torch-1.7.1-cp38-cp38-win_amd64.whl for my CUDA 10.2.

My workaround is that, directly go to https://download.pytorch.org/whl/torch_stable.html and download the binary file matching your environment, and simply install from downloads folder:

pip install --no-cache-dir --force-reinstall torch===1.7.1 torchvision===0.8.2 -f .\Downloads\

Answer 3

Restarting your runtime is advisable, if it appears as if the other solutions did not work.

Answer 4

The reason for torch.cuda.is_available() resulting False is the incompatibility between the versions of pytorch and cudatoolkit.

As on Jun-2022, the current version of pytorch is compatible with cudatoolkit=11.3 whereas the current cuda toolkit version = 11.7. Source

Solution:

1. Run conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch to install pytorch.
2. Install CUDA 11.3 version from https://developer.nvidia.com/cuda-11.3.0-download-archive.

You are good to go.

The original answer is posted here: https://stackoverflow.com/a/72650265/10468354. This is just for quick reference:

Answer 5

I also had the same issue. And running this => a=torch.cuda.FloatTensor(), gave the assertion error AssertionError: Torch not compiled with CUDA enabled . ...which kind of cleared that i was running pytorch without cuda.

Steps:

1. Make sure you have un-installed Pytorch by invoking the following command:

   pip uninstall torch

2. Go to https://pytorch.org/get-started/locally/ and select your system configurations(as shown in the figure).

3. Copy the exact command from the Run this command dialog and run it on your terminal.