The current normalizer transform when operating on GPU tensors uses cupy, which assumes device=0 unless otherwise stated. This results in a bug where trying to run ffcv pipelines with Normalize on any other gpu causes an error to be raised claiming that the tensors are on different devices.
The current normalizer transform when operating on GPU tensors uses cupy, which assumes device=0 unless otherwise stated. This results in a bug where trying to run ffcv pipelines with Normalize on any other gpu causes an error to be raised claiming that the tensors are on different devices.