Running examples on CPU

[uturuncoglu]

I am trying to get familiar to NVIDIA/modulus but I have no access to any GPU. Is it possible to run the tests (like examples/cfd/darcy_fno) on CPU? I have just download the Docker container and started with the docker run --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 --rm -it nvcr.io/nvidia/modulus/modulus:24.01 bash command since I have no GPU access on my local laptop. It seems I could run very simple torch code which is available in the documentation. But I am getting runtime error like RuntimeError: "cuda" device requested but CUDA is not supported by the hardware or driver from darcy_fno test. I think I need to modify the code to work on CPU (probably DistributedManager in darcy_trainer) but don't know how. Anyway, if you guide me about that that would be great.

[sudoCommit]

Hi @uturuncoglu,
I checked the train_fno_darcy.py and there are few places were we check for device, like in model and this dist is defined in DistributedManager where we have device agnostic code, so it should use CPU when GPU is not available.

I guess I found the culprit, its when dataloader is defined you need to pass device otherwise it'll default to GPU in Darcy2D
So updating the code it should work.

dataloader = Darcy2D(
    resolution=cfg.training.resolution,
    batch_size=cfg.training.batch_size,
    normaliser=normaliser,
    device=torch.device("cpu")      # or device=dist.device
)

Optional if above dosen't work
1 way to try is when you launch the container, try explicitly blocking GPU visibility

docker run \
  --shm-size=1g \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  --rm -it \
  -e CUDA_VISIBLE_DEVICES="" \
  nvcr.io/nvidia/modulus/modulus:24.01 bash