The model does the work, not the code. The inference code should be generic autoregressive decoding that would work with any transformer checkpoint. If your generation loop contains addition-specific logic — manually pairing digits, threading carry state, indexing into specific positions — then the Python code is solving the problem, not the model.
with: #anyMessage: -> [:pattern | pattern beBinary ];
，详情可参考WPS官方版本下载
Creator: $24/month，这一点在91视频中也有详细论述
merge(arr, left, mid, right); // 合并两部分