According to the docs, input_tensor.to(device) doesn't move the tensor to the device but returns a tensor that's on the device.
input_tensor.to(device)
Fixes GPU predictions