[AMD Vulkan] Slow image generation, GPU is not used at 100%

[Karhunkaatolaulu]

When generating images with any model, the GPU is not used at full capacity, the process takes much longer than, for example, in ComfyUI program. In Koboldcpp the generation process takes 10 minutes, while in ComfyUI it takes only 25 seconds. The same issue is observed in the Windows version of the program.
Using official kcppt-files from https://huggingface.co/koboldcpp/kcppt . Image models are z_image_turbo_bf16, krea2_turbo_fp8_scaled, flux-2-klein-9b-fp8 (all .safetensors).

Steps to reproduce:

1. Execute Koboldcpp.
2. Load kcppt-template.
3. On the image generation tab select the manually downloaded model files.
4. "Launch".
5. Run any image generation in StableUI.
6. Check GPU usage and Power Consumption in "nvtop" or similiar software.

Manjaro Linux KDE, Kernel 6.18.39, Mesa 21.6.5 and Windows 11 25H2 with the current drivers.
AMD Ryzen 7 5700X
32 GB DDR4 RAM
AMD Radeon RX 9060 XT 16GB
Nvme SSD
Koboldcpp ver. 1.118.1-nocuda

terminal.log

[LostRuins]

Hmm, why are you loading the bf16 version of the model? I'm not sure if accelerated BF16 ops are available. Can you try with https://huggingface.co/leejet/Z-Image-Turbo-GGUF/resolve/main/z_image_turbo-Q4_0.gguf instead?

Otherwise, maybe @wbruna might have a better idea.

[Karhunkaatolaulu]

I am not satisfied with the quality of the generated images using quantized models.
The results with z_image_turbo-Q4_0.gguf are in a new log file and screenshots, generation took 2 minutes and 23 seconds.

terminal2.log

[wbruna]

It's most likely spilling to GTT memory. If the card can't quite hold the whole model + working memory, the Vulkan driver automatically swaps VRAM out, which is much slower.

One thing that could be tried is sdoffloadcpu, which will keep the VAE weights out of the GPU during generation; that's what I use on my own 16G 7600XT for running z_image_turbo_bf16 . In my case, it's enough because the card isn't running anything else; but if the slowdown is caused by a spill, even a little bit of extra memory may give you a noticeable speed difference.

Also, when comparing with ComfyUI, check if the generation parameters are the same. CFG 2.5 + 20 steps is overkill for Z-Image Turbo; if ComfyUI is running with the recommended CFG 1 and 8-9 steps, it'll cause a 4x speed difference right away. Vulkan can also be slower than ROCm for DiT models like Z-Image Turbo (~60% on my card).

[Karhunkaatolaulu]

Yeah my mistake, I messed up the generation parameters . I ran two new generations with the recommended parameters (same as in ComfyUI); the logs and screenshots are below. Even with the z_image_turbo-Q4_0.gguf, not all of the graphics card’s potential is being used (~135/182 W). Is this normal?

terminal_Q4_0.txt

terminal_BF16.txt