ggml 8de5c82360 broke image generation on Vulkan

[wbruna]

This revision is not yet on sd.cpp, but it's included as part of ggml 0.9.1 (8c18eabe19).

./sd -m ./dreamshaper_8.safetensors -p 'fast car' -o (...).png

Testing master-301, including ROCm as a baseline:

ROCm	ROCm + ggml 8c18eabe19	Vulkan	Vulkan + ggml 8c18eabe19

Bisect points to ggml-org/ggml@8de5c82 (llama.cpp ggml-org/llama.cpp@803dac2 , PR ggml-org/llama.cpp#16056 ) :

Vulkan + ggml d98d650b62	Vulkan + ggml 8de5c82360

Edit: the device is:

ggml_vulkan: 1 = 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

My iGPU is affected too:

ggml_vulkan: 0 = AMD Radeon Vega 11 Graphics (RADV RAVEN) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none

Also happens with amdvlk:

ggml_vulkan: 0 = AMD Radeon RX 7600 XT (AMD open-source driver) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat

This is on Debian 13, mainline amdgpu (both on mainline kernel 6.12.43 and 6.16.3 from backports), and Mesa 25.0.7 .

[0cc4m]

On what hardware?

[wbruna]

On what hardware?

@0cc4m, I've edited the post to include them.

[wbruna]

A few more tests. I first noticed this issue on SDXL:

./sd -m cyberrealisticXL_v60.safetensors -p 'photo of a red car in a parking lot' -o (...).png

d98d650b62	8de5c82360

But Chroma output changes just a little bit:

./sd --diffusion-model Chroma1-HD-Flash-Q4_0.gguf --t5xxl t5xxl_q5_1.gguf --vae ae-f16.gguf -p 'photo of a red car in a parking lot' --steps 13 --cfg-scale 1 --sampling-method euler --scheduler smoothstep -o (...).png

d98d650b62	8de5c82360

[0cc4m]

@wbruna Please give ggml-org/llama.cpp#16151 a try.

[wbruna]

@0cc4m , applying d15ad04a02 (from your PR) on top of ggml 8de5c82360 seems to have fixed it; thanks!

SD1.5	SDXL	Chroma

Curiously, SD1.5 and SDXL generations are identical to those before 8de5c82360, while Chroma kept the slightly different output.

@leejet , maybe it's best to keep this issue open as a reminder until the fix arrives on ggml?

[leejet]

I just updated ggml… Once this fix is synced to GGML, I will update again.

[Green-Sky]

Should be synced.

[leejet]

I’ve updated to the latest ggml, and this issue should be fixed now.

[wbruna]

Did a bunch of tests with 6ad46bb, it seems to be working now!