[BUG] Qwen72B-Chat基于llama.cpp推理，答案末尾总是追加[PAD151645]

[zhengxingmao]

是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

• [X] 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

• [X] 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

输出结果：{"table": {"columns": ["col3", "count"], "data": [["出", 7], ["进", 3]]}}[PAD151645][PAD151643]

期望行为 | Expected Behavior

{"table": {"columns": ["col3", "count"], "data": [["出", 7], ["进", 3]]}}

复现方法 | Steps To Reproduce

基于llama.cpp进行推理问答，答案后总是追加额外字符

运行环境 | Environment

- OS:Ubuntu 22.04
- Python:3.10
- Transformers:4.36
- PyTorch:2.1.2
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):12.3

备注 | Anything else?

No response

[jklj077]

llama.cpp对qwen tokenizer的支持有限，[PAD151645][PAD151643]分别对应<|im_end|>和<|endoftext|>，做点后处理吧，匹配删除就可以。或者去llama.cpp提issue。