Why is GPU not found after installing PyTorch via Anaconda?

Locales: Data Centers & Campus RC University of Alabama at Birmingham
, ,
1

I’ve installed PyTorch in an Anaconda environment but can’t find the GPU. What is going wrong, why, and how can I fix it?

2

There are two common reasons a GPU can’t be found when using PyTorch.

  1. Be sure you’ve requested a GPU.
    • On Cheaha you will need to use the flags --partition=pascalnodes or --partition=pascalnodes-medium and --gres=gpu:1 for a single GPU.
    • With Open OnDemand Jupyter Notebook, you will need to select the pascalnodes or pascalnodes medium partition. At this time exactly one GPU is automatically requested. It is not possible to request more.
  2. Be sure you’ve installed the GPU flavor of PyTorch. Please see https://docs.rc.uab.edu/cheaha/slurm/gpu/#pytorch-compatibility for more information about how to install PyTorch with GPU compatibility.
3

Make sure that you have installed PyTorch following the guide on their website (Start Locally | PyTorch). For CUDA-aware PyTorch they recommend using:

conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia

which should pull from the pytorch and nvidia channels to get the right packages.

1 Like
4

First, I would check that you are on a node with a GPU assuming it is a Nvidia GPU.

type in the command: nvidia-smi (check the CUDA version and make sure it is the same you used during pytorch installation)

Second, I would check with the conda list command that the pytorch version is not a cpu version.

5

In Addition to that. Always check for the device being used by the application. Whether Tensorflow or Pythorch:

In Pytorch, you can list number of available GPUs using torch.cuda.device_count().

  • If you specify cpu as a device such as torch.device("cpu"), this means all available CPUs/cores and memory will be used in the computation of tensors.
  • If you specify cuda as a device such as torch.device("cuda"), it is the same as torch.device("cuda:0") which is the first GPU and memory.
  • Otherwise, specify which GPU you want using torch.device(f"cuda:{i}") for the ith device.