Please teach stable-diffusion.cpp team how to make widely compatible Vulkan builds!

[alex-ie]

I like Linux Mint 21 (aka LM21, Ubuntu 22.04 based), it is smaller and often faster on my old hardware vs. LM22 (Ubuntu 24.04) based. And kcpp works on LM21 and LM22!

But sd-cli Vuklan version from releases does not run on LM21 (and they do not release CUDA versions at all for Linux), though it works on LM 22. I have looked at building sd.cpp Vulkan myself, but seems installing Vulkan SDK is not very easy task.

Please teach stable-diffusion.cpp team how to make more widely compatible Vulkan builds (for older OSes)!

[LostRuins]

You could use the image gen from KoboldCpp, its using the same engine as sd.cpp. If you self compile it can build for sd CLI as well. what features are you missing currently?

[alex-ie]

what features are you missing currently

I wanted to try Qwen Image 2.1, sd.cpp claimed 1st day support Sep 20 IIRC. But 1.122.1 just (IIRC, I was several days ago): "Load Image Model OK: False." Maybe I got some parts of the model wrong, but debug output was empty except for that abrupt end.

[LostRuins]

It's working in the rolling builds : try this https://github.com/LostRuins/koboldcpp/releases/tag/rolling

Make sure you load all the models files

sdmodel: qwen_image_2.1-Q4_K.gguf
sdclip1: Qwen3VL-8B-Instruct-Q4_K_M.gguf
sdclip2: mmproj-Qwen3VL-8B-Instruct-Q8_0.gguf
vae: qwen_image_2.1_vae_bf16.safetensors

[alex-ie]

Hm, maybe I missed "sdclip2: mmproj-Qwen3VL-8B-Instruct-Q8_0.gguf", thanks! I will wait for release support, I am in not big hurry. As for sd,cpp, I did got used to sd-cli command line way.