[Bug] LoRAs will not work when using quantized models???

[BingLiHanShuang]

Git commit

Inside https://pypi.org/project/stable-diffusion-cpp-python/, there is:
LoRAs will not work when using quantized models. You must instead use a full precision .safetensors model.
But LoRA in python code doesn't have this limitation?

By the way, could qwen edit model use lora_model_dir() function?

Operating System & Version

Windows11

GGML backends

CUDA

Command-line arguments used

python app.py

Steps to reproduce

from stable_diffusion_cpp import StableDiffusion

stable_diffusion = StableDiffusion(
diffusion_model_path="D:/qwen/QwenEdit/Qwen-Rapid-NSFW-v23_Q4_K.gguf",
llm_path="D:/qwen/QwenEdit/Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf",
vae_path="D:/qwen/QwenEdit/qwen_image_vae.safetensors",
clip_vision_path="D:/qwen/QwenEdit/Qwen2.5-VL-7B-Instruct-mmproj-BF16.gguf",
# lora_model_dir="D:/qwen/QwenEdit/LORA",
qwen_image_zero_cond_t=True,
diffusion_flash_attn=True,
offload_params_to_cpu=True,
)

prompts = [
"lora:F2P:1,
........

What you expected to happen

Not work?

What actually happened

Work?

Logs / error messages / stack trace

No response

Additional context / environment details

No response

[wbruna]

Sounds like you've opened this bug on the wrong project: this is mainline stable-diffusion.cpp. Try william-murray1204/stable-diffusion-cpp-python.

And while we are at it, that information is outdated: LoRAs do work with quantized models.

[JohnLoveJoy]

I can confirm that Loras work with GGUF quantized models.