To avoid the error TypeError: TransformerWithHead.forward() missing 1 required positional argument: 'input_ids', similar to https://github.com/pytorch/pytorch/issues/15161 when using dataparallel and the last minibatch gets compressed, we add right-padding to input_ids and then take off the extra logits/labels that this creates.
To avoid the error
TypeError: TransformerWithHead.forward() missing 1 required positional argument: 'input_ids', similar to https://github.com/pytorch/pytorch/issues/15161 when using dataparallel and the last minibatch gets compressed, we add right-padding to input_ids and then take off the extra logits/labels that this creates.