CodeQwen returns extra white space for code completion

[ycclnn]

Other models like deepseek works without a problem.

[wsxiaoys]

Thanks for reporting the issue - I have also observed the issue and am looking into the debug process. It seems that the tokenizer treats _ as a space, regardless of the context for CodeQwen.

[wsxiaoys]

Can confirm this presents in upstream llama.cpp as well:

cross posted at: ggerganov/llama.cpp#7050

[ycclnn]

Can confirm this presents in upstream llama.cpp as well:

cross posted at: ggerganov/llama.cpp#7050

yeah, white space exists for using vllm as well so its a model thing rather than serving framework thing. I was thinking shiftleft the cursor a few chars and then check the completion result with overlapping substrings to overcome this and thats the only workaround in my mind.

[ycclnn]

Observed that codeqwen not always returns extra white space, and sometimes the white space is meaningful. Therefore, trim the leading white space may not be an approach.