[Bug] HiDream-O1-image in Q8 quant creating fully white images on metal

[banddans]

Git commit

version: stable-diffusion.cpp version master-810-db99efd, commit db99efd

Operating System & Version

macOS Tahoe 26.6

GGML backends

Metal

Command-line arguments used

./bin/sd-cli -m models/hidream_o1_image_q8_0.gguf -p "a lovely cat holding a sign says 'hidream o1 cpp'" -v -H 1024 -W 1024 --cfg-scale 1.0 --preview proj

Steps to reproduce

1. I downloaded the non-quantized version of the model from their official huggingface collection
2. Quantized it to gguf: ./bin/sd-cli -M convert \ -m models/hidream_o1_image_bf16.safetensors \ -o models/hidream_o1_image_q8_0.gguf \ -v --type q8_0
3. I ran the exact command from the README:
   ./bin/sd-cli -m models/hidream_o1_image_q8_0.gguf -p "a lovely cat holding a sign says 'hidream o1 cpp'" -v -H 1024 -W 1024 --cfg-scale 1.

What you expected to happen

I expected it to generate an image like in the docs...

What actually happened

I got a fully white image.

Logs / error messages / stack trace

Full command history with verbose logs:
logs.txt

Additional context / environment details

I'm on a mac mini with 16GB unified memory. This may have to do with the quantization but I can't test it as the original model is too large for me. Here I used Q8_0.

[wbruna]

Possibly related to #1496 .

[banddans]

The --tensor-type-rules "^model\.language_model\.layers\.[0-9]+\.mlp\.(gate|up)_proj\.weight=q8_0" trick seems to work for me so probably yes.