Smart apply produces unusable code in the editor

[manos-msdn]

Describe the bug

Smart apply pastes code that is not usable, see image.

Information about your version

0.24.0 - latest stable as of today Feb 17th 2025 windows binary cuda 122 tabby_x86_64-windows-msvc-cuda122

Information about your GPU

NVIDIA-SMI 572.16 Driver Version: 572.16 CUDA Version: 12.8

Additional context

First time trying this today. Windows 10, used the binary with the standard small models. Installation went smoothly, and the integration with vs code works.

Tried the example with the string conversion. Standard "apply in editor" works fine, but smart apply shows the below output which is not usable

I could be doing something wrong as I just started but other than that, looks good, it suggests code snippets, and on localhost 8080 I can see the stats and accepted completions and all

[icycodes]

Hi @manos-msdn
Thank you for reporting this issue.

Could you please share which chat model you are using?
I recommend trying a model like Qwen2.5-Coder-7B-Instruct with the smart apply feature.

[manos-msdn]

Hi @manos-msdn Thank you for reporting this issue.

Could you please share which chat model you are using? I recommend trying a model like Qwen2.5-Coder-7B-Instruct with the smart apply feature.

hi, I was using the standard small models that come with the default installation (according to the quick start guide). I can try the qwen when I find some time, thanks

[georgedorn]

I can confirm this continues to happen using Qwen2.5-Coder-14B as the coding model. It makes a real hash of the file, with left-in diff markers and leftover bits of XML from the response. The same thing happens with the 7B version of that model. I don't think this has anything to do with the model itself, but the extension itself.

[gitwittidbit]

I have a similar issue. For me, Tabby inserts its whole thinking process into the code which also makes the code unusable.