Multi gpu, named tuple converts to tuple

[Vozf]

🐛 Bug

When setting multiple gpu training, named tuples are converted to tuples and this breaks training resulting in error

    preds = self.model(batch.image)
AttributeError: 'tuple' object has no attribute 'image'

Where batch is a named tuple and works fine as namedtuple when training with single gpu

[Vozf]

This may be connected to #1588

[edenlightning]

Thanks for the report!
This should be resolved by PyTorch 1.8. This isn't an issue in Lightning.

A workaround in the meantime: use dictionaries instead, or extract the values from the tuple, and the first index will be the image.