[Feature Request] MOSS-Image GGUF: Support separate Q/K/V projection weights #2158

Feature Summary

[Feature Request] MOSS-Image GGUF: Support Separate Q/K/V Weights

Detailed Description

Feature Request

Please add support for MOSS-Image GGUF models that store Q, K, and V projection weights separately.

Model

https://huggingface.co/realrebelai/MOSS-Image_GGUFs

Example: MOSS-Image-SFT-Q4_K_M.gguf

Issue

The GGUF model uses separate attention projection tensors:

• layers.0.proj.q.weight
• layers.0.proj.k.weight
• layers.0.proj.v.weight

However, the current MOSS-Image implementation in stable-diffusion.cpp appears to expect a fused proj.qkv projection.

The original OpenMOSS model also uses separate Q/K/V weights.

This appears to be a weight-layout compatibility issue rather than a Q4_K_M quantization limitation.

Requested Improvement

• Automatically detect fused and separate Q/K/V weight layouts.
• Support both formats without requiring manual model conversion.
• Support loading existing quantized MOSS-Image GGUF files directly.
• Include the improvement in official prebuilt Windows ROCm releases.

Environment

• OS: Windows
• GPU: AMD Radeon RX 9070 16GB
• Backend: ROCm
• stable-diffusion.cpp: f0aac52
• Model: MOSS-Image-SFT Q4_K_M GGUF

Would it be possible to add support for separate Q/K/V projections so existing MOSS-Image GGUF models can run directly with stable-diffusion.cpp?

Thank you!

Alternatives you considered

No response

Additional context

No response