Does it support the gguf format model of Qwen2-VL-2B-Instruct

[helloHKTK]

Is your feature request related to a problem? Please describe.
I would like to support whether the current llama-cpp-python supports the gguf format model of Qwen2-VL-2B-Instruct？

Describe the solution you'd like
Does it support the gguf format model of Qwen2-VL-2B-Instruct.

Describe alternatives you've considered
Does it support the gguf format model of Qwen2-VL-2B-Instruct.

Additional context
Add any other context or screenshots about the feature request here.

[rob-field1]

Yes the latest commit integrates the llama.cpp changes for Qwen2-VL support.
e.g.
llm = Llama( model_path="models\Qwen2-VL-7B-Instruct-Q6_K_L.gguf", n_gpu_layers=30, n_threads=6, seed=42, n_ctx=7500, flash_attn=True, )
Model metadata: {'general.name': 'Qwen2 VL 7B Instruct', 'general.architecture': 'qwen2vl', ...

edit: although I'm having some trouble processing/running inference on images.

[fanfansoft]

Yes the latest commit integrates the llama.cpp changes for Qwen2-VL support. e.g. llm = Llama( model_path="models\Qwen2-VL-7B-Instruct-Q6_K_L.gguf", n_gpu_layers=30, n_threads=6, seed=42, n_ctx=7500, flash_attn=True, ) Model metadata: {'general.name': 'Qwen2 VL 7B Instruct', 'general.architecture': 'qwen2vl', ...

edit: although I'm having some trouble processing/running inference on images.

hello，Is the 'latest commit' referring to the llama-cpp-python version 0.3.6? How should the chat_handler be referenced in llm = Llama()? I looked at the source code and didn't find any support for qwen2_vl_2B.

[MoRocety]

Yes the latest commit integrates the llama.cpp changes for Qwen2-VL support. e.g. llm = Llama( model_path="models\Qwen2-VL-7B-Instruct-Q6_K_L.gguf", n_gpu_layers=30, n_threads=6, seed=42, n_ctx=7500, flash_attn=True, ) Model metadata: {'general.name': 'Qwen2 VL 7B Instruct', 'general.architecture': 'qwen2vl', ...

edit: although I'm having some trouble processing/running inference on images.

To my understanding, the actual llama.cpp lib doesn't actually and never has supported multimodal inference server with qwen2vl so curious as to how you made it work at all?

[rob-field1]

@MoRocety There's been support for a few months (ggerganov/llama.cpp#10361) but the llama-cpp-python prompt format needs some tweaking

[MoRocety]

@MoRocety There's been support for a few months (ggerganov/llama.cpp#10361) but the llama-cpp-python prompt format needs some tweaking

I think they mention in there that llama-server doesn't support multimodal yet, so again curious as to how llama-cpp-python makes it work?