_prepare_teacher_forcing_mask silently ignored local_attn_size while the block-wise causal mask honors it, so a configured attention window was dropped on the teacher-forcing path. Raise NotImplementedError instead of training with a mask that contradicts the config.