Fix Ming-Image GGUF packed QKV loading

[RealRebelAI]

Summary

Adds Ming-Image GGUF support for quantized Diffusers-format checkpoints that store attention Q/K/V as separate tensors.

ComfyUI's normal Diffusers conversion merges those tensors using ordinary Torch tensor assignment. With quantized GGUF tensors, the underlying physical storage is packed bytes while the GGMLTensor keeps a separate logical tensor shape. This can cause the packed row width to be used during the merge instead of the logical width.

For Ming-Image-0.1-Design this produced:

The expanded size of the tensor (3840) must match the existing size (4080)
Target sizes: [3840, 3840]
Tensor sizes: [3840, 4080]

Fix

When UnetLoaderGGUF detects a quantized Ming-Image state dict, it performs the Ming Diffusers key conversion before ComfyUI's generic model conversion.

The Q/K/V merge:

• keeps the tensors quantized
• concatenates their packed GGUF storage directly
• preserves the GGML quantization type
• preserves the correct logical tensor shape
• preserves is_largest_weight when present

The GGUF model itself does not need to be rebuilt, repaired, dequantized, or requantized.

Tested

Tested successfully with:

• Ming-Image-0.1-Design
• mixed GGUF with BF16 / Q8_0 / Q6_K / Q4_K tensors
• ComfyUI 0.37.0
• ComfyUI-GGUF
• Windows portable ComfyUI
• RTX 3070 8 GB
• --lowvram

Before the change the same GGUF failed during convert_diffusers_mmdit() with the 3840 vs 4080 shape mismatch.

After the change the original GGUF loads successfully and runs without modifying the model file.

Scope

This is deliberately detected using Ming-Image-specific tensor keys rather than changing GGMLTensor behavior globally.

Ming currently shares ComfyUI's z_image_to_diffusers() key mapping internally, but this change is specifically scoped and named for Ming-Image because that is the model and path tested here.

[RealRebelAI]

id also like to add that i have quants uploading for the model for anyone interested. theyll be in my HF repo.

[abdulraheem81]

ComfyUI-GGUF-db80272b3d04f4747a4df00cb0a7ca95f63ec490.zip
I make it work completely in Macos test it Error AttributeError: type object 'GGMLOps' has no attribute' MoEExperts' fix