pytorch lightning self.device has incorrect type

[ismael-mendoza]

For now a fixed this with #typing: ignore but curious to know if there is a better solution

I do think however that using self.device is a bit strange (since pytorch lightning automatically moves tensors to GPU) and perhaps should be avoided in cases where it's possible.

[dereklhansen]

If it's just a typing problem, it's okay to ignore since the typing ecosystem isn't perfect yet in Python. Is there a specific location where this is happening and it's an issue?

As far as I am aware, PyTorch lightning only makes sure that tensors from the DataLoader and the LightningModule are on the same device during training. Thus, we still need to get the device for any new tensors being created during calculations. That said, relying on the device attribute of the LightningModule or LightningDataset is often not necessary, because we can often figure out the device from other tensors in scope.