Tools: supported models architectures - any support for Qwen-Image (and both Image-Edit Qwens) ?

[Dino-Zavrus]

Hello City96
thank you for your great modules.

Recently i reinstalled tools and Llama.cpp to quantize CosmosPredict2 and eventually it worked, though building Llama.cpp is always a pain for me :(
I have examined your llama patch to figure out the following architectures are supported:

• FLUX
• SD1.5
• SDXL
• SD3 (3.5)
• AuraFlow
• LTXV
• Hunyuan Video
• WAN
• HiDream
• Cosmos

Do you plan to add Qwen Image models support (t2i and both i2i) for quantization tools?
Could be handy to have the ability to quantize Qwens locally.
So that is kinda "feature request"

[phil2sat]

i searched alot but take a look at this, tool_auto.py is your friend
https://github.com/city96/ComfyUI-GGUF/tree/auto_convert/tools

[Dino-Zavrus]

oh, thanks a lot, my friend !!!
saved the code locally to understand and implement the change

just for my curiosity: do i get it right this is the dead branch which will never make it to "production" main branch?

[phil2sat]

i dont get it also...
but for edit 2509 after applying the lcpp.patch maybe replace the part around 18176 with this:

    // first/last block high precision test
    if (arch == LLM_ARCH_QWEN_IMAGE){
        if (
            (name.find("transformer_blocks.0.") != std::string::npos) ||
            (name.find("transformer_blocks.59.") != std::string::npos) // this should be dynamic
        ) {
            new_type = GGML_TYPE_Q8_0;  // Always Q8_0 for first/last blocks
        }
    }

in llama.cpp before building, it fixes the garbled image on lower quants, here i set Q8_0 everything over, including Q5_K works but i saw a little bit better quality, maybe next ill try to skip the blocks entirely and leave them original.

see here: https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/discussions/6 and https://huggingface.co/city96/Qwen-Image-gguf/discussions/10#68a5ca5c25a18b29ad50794d