decoder_stack.DecoderStack: dstack_window_length = 0 recurrent_layer_indices = (-3,) feedback_recurrence = False transformer_layer.TransformerLayer: window_length = 512 recurrent_num_states = 512 recurrent_gate_type = "lstm" recurrent_single_gate = False recurrent_skip_ffn = True training_loop.Trainer: generate_every_steps = 0 # Not yet supported with recurrence.