transformer_layer.TransformerLayer:
  relative_position_type = "t5"

decoder_stack.DecoderStack:
  use_absolute_positions = False