Trainer cannot find available GPUS

[omrishac]

Hey,

I am trying to use Huyvnphan's implementation of CIFAR-10 classifier with Pytorch on Colab.
In his implementation he uses pytorch-lightning and for some reason when I try to train the model PL raises an error:
"pytorch_lightning.utilities.exceptions.MisconfigurationException: GPUs requested but none are available."

note that when I run "torch.cuda.is_available()" I get True and "torch.cuda.device_count()" is 1.

The Trainer call looks like that:

    trainer = Trainer(
        fast_dev_run=bool(args.dev),
        logger=logger if not bool(args.dev + args.test_phase) else None,
        gpus=-1,
        deterministic=True,
        weights_summary=None,
        log_every_n_steps=1,
        max_epochs=args.max_epochs,
        checkpoint_callback=checkpoint,
        precision=args.precision,
    )

please look at his implementation for more info:
https://github.com/huyvnphan/PyTorch_CIFAR10

Thanks in advance,
OS

[lilylizhengzhou]

have exactly same issue when I try to run this repo

[akihironitta]

@lilylizhengzhou Lightning uses this simple condition to check the cuda device availability (in master):
https://github.com/Lightning-AI/lightning/blob/4508c7004c0f8075e556b6c6678c60b0e958719e/src/lightning_lite/accelerators/cuda.py#L100-L122

I'd recommend recreating your environment and installing PyTorch and Lightning again.