Using both Qwen Instruct + Qwen infill via LoRA

[ngxson]

The author of Qwen model confirm that infill capability is only possible with Qwen-coder (non-Instruct): https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/discussions/2#6731a45e0e39be0605a0df20

This will limit the capability of the model to /infill only, so it cannot be used with /chat/completions

However, we know that the instruct version is indeed fine-tuned from non-instruct, see the technical report: https://arxiv.org/pdf/2409.12186

To make the model usable with both chat and infill, one solution is to extract the difference between 2 models to a LoRA adapter. This can be done via something like mergekit-extract-lora, then we can set lora scale at runtime (i.e. set to 0.0 on infill and 1.0 on chat)

[ggerganov]

Good idea! Btw, shouldn't we implement a LoRA extractor in llama.cpp?

[ngxson]

(Another idea would be to use ggml_build_backward_expand to "optimize" lora_A * lora_B to match the delta between 2 models, but I'm not even sure if this works)

[ggerganov]

The power method in pt. 3 of this paper https://watermark.silverchair.com/300268.pdf seems quite simple to implement with ggml graphs. What do you think?

[ngxson]

I can't access the given link, could you give me the name of the paper? (or a screenshot)

[ggerganov]

Should be accessible from here: https://academic.oup.com/comjnl/article/30/3/268/364791 (click on PDF button)

[ngxson]

Hmm yeah that sounds feasible. I also found an implementation using numpy here: https://gist.github.com/Zhenye-Na/cbf4e534b44ef94fdbad663ef56dd333