[Bug] Hidream-dev produces broken images on Vulkan with quantized models

[wbruna]

Git commit

0665a7f
(master-602-0665a7f)

Operating System & Version

Debian 13

GGML backends

Vulkan

Command-line arguments used

sd-cli -m HiDream-O1-Image-Dev-Q8_0.gguf --cfg-scale 1 -p 'flower'

Steps to reproduce

Vulkan seems to be broken for any parameter combination, right from the first preview:

[DEBUG] ggml_extend.hpp:60 - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:60 - ggml_vulkan: 0 = AMD Radeon RX 7600 XT (RADV NAVI33) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat

What you expected to happen

A pretty flower

What actually happened

A pretty noise pattern

Logs / error messages / stack trace

No response

Additional context / environment details

Same hardware and quantization as #1495 .

[wbruna]

Vulkan seems to be sensitive to the mlp.down_proj layer quantization: keeping it as-is avoids this issue.

From HiDream-O1-Image-Dev-BF16.safetensors:

• --tensor-type-rules "^model\.language_model\.layers\.[0-9]+\.mlp\.(gate|up)_proj\.weight=q8_0" works (~12.4G)
• adding ,^model\.language_model\.layers\.[0-9]+\.self_attn\.[kqvo]_proj\.weight=q8_0 works too (~11.1G)

[wbruna]

For existing quantized models, generating with:
--tensor-type-rules "^model\.language_model\.layers\.[0-9]+\.mlp\.down_proj\.weight=bf16"
seems to work too.

[LostRuins]

I have the same issue, except I don't even get areas of coherence