[Question] PandaGPT with llama.cpp

[ningshanwutuobang]

I have tried to use llama.cpp for PandaGPT in panda_gpt_llama_cpp.
The script get poor performance. Is there any thing wrong for the procedure? Or is it just the limit of the model or q4_1 precision?

The following are my steps.

1. Obtain vicuna v0. Use FastChat@v0.1.10 to merge llama-13b-hf and vicuna-13b-delta-v0.
2. Merge lora weights to vicuna v0.
3. Convert it to ggml format and quantize it to q4_1. The result is ggml-pandagpt-vicuna-merge.
4. The script is located in panda_gpt_llama_cpp.

The model seems to recognize <Img>...</Img> labels.