"KCPP SD Failed to create context!" after trying to load flux.1-schnell

[tororon1231]

Downloaded weights for FLUX.1-schnell. After that this error was shown:

---

Welcome to KoboldCpp - Version 1.78
For command line arguments, please refer to --help

---

Auto Selected CUDA Backend...

Attempting to use CPU library.
Initializing dynamic library: koboldcpp_default.dll

ImageGen Init - Load Model: C:\Users\flux1-schnell-q4_k.gguf

Error: KCPP SD Failed to create context!
Load Image Model OK: False

Error: Could not load image model: C:\Users\flux1-schnell-q4_k.gguf

[LostRuins]

Did you load the t5xxl, clip-l and vae as well? Those are required for flux.
https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

Also make sure you load the image model in the right place (NOT the text model)

[tororon1231]

@LostRuins Nice it worked! I used quantized text encorders, so it doesn't use much space on my machine. Unfortunately the GUI only supports .safetensors, so I needed to change the file extension from .gguf to .safetensors. Could you please add support for selecting .gguf file extensions for text encoders in the launch GUI?

[LostRuins]

Oh, I didn't know anyone actually made quantized text encoders. Sure, I can do that.

Glad flux is working well for you. Do share and help others as I've noticed many people struggling to get it to work.

[Pyrooogenic]

I'm getting this exact same error on KoboldCPP-v1.78.yr0-ROCm, having tried several of the exact same files as this post suggested and even went and tried it on Vulkan through koboldcpp-1.78 just to see if maybe the problem is isolated to that version. But just about everything I've done is not making this work.

I even made a post about it on the ROCm discussion board. Same problem, but I also noted that KCPP SD in general just stopped working for me entirely on this latest version.

[LostRuins]

Run it with --debugmode , see if any additional information is displayed. Are you sure you've selected all the correct text encoders? Is it only flux that's broken or is sd1.5 also broken?

[Pyrooogenic]

Run it with --debugmode , see if any additional information is displayed. Are you sure you've selected all the correct text encoders? Is it only flux that's broken or is sd1.5 also broken?

So I made sure to go back and retest to see if I was not crazy on the text encoder side of things.

Same exact setup on both ends just using Vulkan on normal version, HIP on ROCm ofc.

And they both worked.

Something I did not account for was the possibility of a bad model, the vae, t5xxl, clip_l were all supposed to be from flux.1-schnell which I used, but apparently that didn't help the case of the original model I wanted to use.

I made sure to go back and test both the schnell GGUF I tried before and artsyliteQ4KS, the schnell one now works, but artsy doesn't. Originally schnell itself didn't work with the exact same setup I used, idk what changed.

Same with artsyDream, so I was just led astray by my own incompetence.

![image](https://github.com/user-attachments/assets/1054316a-7369-48c1-8494-792bd2ce768f)

For SD1.5 and SDXL (By extension the PonyXL model), my usual main gotos, I made sure to retest since you asked.

![Untitled](https://github.com/user-attachments/assets/7f6b3678-ca0a-4bc1-8495-930ba4e6e9ce)

Works too. 
![388777675-02880a9d-b4ef-447c-8208-2e4dfa5d89d6](https://github.com/user-attachments/assets/2c6692bb-1098-4a08-b0cf-65c80eb9ce63)
While earlier I was experiencing this. Which I'm unsure how to interpret it. I commented here since it really WAS NOT working at all earlier even with the exact same setup.

So I can't really explain it besides just, it happened.

[LostRuins]

Another possibility is a bad download (you can check SHA256 hashes against the original), or perhaps you were running something else at the time and had insufficient or occupied VRAM. In any case, glad you got it working, and do close this issue if its resolved.