fszontagh

Have you tried the CUDA_* environment variables? eg: CUDA_VISIBLE_DEVICES

I didn't tested with sd.cpp, but with pytorch. [This post maybe can help](https://discuss.pytorch.org/t/what-does-export-cuda-visible-devices-1-really-do/90340/3) In theorically you can start the program with different env variables, eg in linux: `CUDA_VISIBLE_DEVICES=0,1,2,3 sd -m...

Interesting. I tried the safetensors model from the link, really not works. (As i remember it worked some months ago). But the pickletensors model file works well: https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_openpose.pth

> Looks like the tensor naming convention is completely diferent. I'm guessing `diffusers` requires different tensor names for control Maybe this one would work? https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/blob/main/control_v11p_sd15_openpose_fp16.safetensors I can't test it right...

> Do we support control nets for other models as well besides 1.5? What about this? > > https://huggingface.co/xinsir/controlnet-union-sdxl-1.0/tree/main Yeah, I tried the CN model you linked, as well as...

> Test with Flux Schnell q3_k at 1792x1792 (biggest I could acheive using shared video memory without crashing), 6 steps How much video memory do you have?

If I'm right, this is a precompiled version of sd.exe. The error you encountered: **"_CUDA error: no kernel image is available for execution on the device_"** This means your GPU's...

> According to this: master-0e86d90 > > stable-diffusion.cpp should work on Nvidia 3xxx series RTX but it still doesn't: > What is the NVIDIA driver and CUDA version?

Yeah, something weird happening here. Your 3050 is sm_86 which is built with the workflow. You driver is too new and the binary is built with CUDA12, but in theory...

> I ended up building an exe from the source and now I test it. Should I post the link to the build here or something if it ends up...