qwen.cpp
qwen.cpp copied to clipboard
[BUG] 多轮对话的 prompt 应该如何构建?
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
在你们给出的cpp示例中,auto QwenTokenizer::build_prompt(const std::vector<std::string> &history) const -> std::string { 这个函数指出,如果是多轮对话时,会重新构建user和assistant的信息。但是构建历史assistant信息的时候使用的是<|im_start|>" << history[i + 1] << "<|im_end|>",而不是使用<|im_start|>assistant\n" << history[i + 1] << "<|im_end|>"。这样会导致无法在多轮对话行中使用kv cache,每次必须要重新构建prompt,然后生成新的kv cache。我的理解是合理的吗?
期望行为 | Expected Behavior
No response
复现方法 | Steps To Reproduce
No response
运行环境 | Environment
- OS:
- Python:
- Transformers:
- PyTorch:
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):
备注 | Anything else?
No response