KoboldCPP.exe v1.105.2 - SDUI Image Generation Error with Z Image Turbo - Vulkan and CLBlast

[alan-l]

Error as described in title. Like in my previous issue #1896 , I have to go back to v1.103.

Running on Intel Iris Xe.

The only model that ever worked with SDUI running Vulkan was DreamShaperXL Turbo. All other models always generated errors unless I ran CLBlast. Now CLBlast is not working too.

I tested this with 1.105, 1.105.1, and 1.105.2. They all generate the error described below.

---

Error when Vulkan selected:

GGML_ASSERT_CONTINUE: ggml/src/ggml-vulkan/ggml-vulkan.cpp:6023: wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] && wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] && wg2 <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]
...
...
...
GGML_ASSERT_CONTINUE: ggml/src/ggml-vulkan/ggml-vulkan.cpp:6023: wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] && wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] && wg2 <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]
GGML_ASSERT_CONTINUE: ggml/src/ggml-vulkan/ggml-vulkan.cpp:6023: wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] && wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] && wg2 <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]
exception: access violation reading 0x0000007683550000
Generate Image: The response could not be sent, maybe connection was terminated?

---

Error when CLBlast selected:

Generating Image (9 steps)

OpenCL: Unsupported Tensor Type Detected: 23
Note that CLBlast only works on classic quants and K-quants!
otherarch/ggml_v3b-opencl.cpp:1821: GGML_ASSERT(to_fp32_cl != nullptr) failed

---

[alan-l]

To append to this, running on CUDA playing around with Wan 2.2 for vidgen, testing out your new TAE SD feature generated the following error and exits after running through all the requested steps:

D:\a\koboldcpp\koboldcpp\ggml\src\ggml.c:1716: GGML_ASSERT(obj_new) failed

[LostRuins]

In 1.103 was vulkan working for you?

[alan-l]

In 1.103 was vulkan working for you?

@LostRuins Just finished testing 1.103 with both Vulkan and CLBlast on Intel Iris Xe with 8GB shared (16GB total), both completed successfully at 512x512.

[alan-l]

Also tested 1024x1024 using 1.103 with both Vulkan and CLBlast on Intel Iris Xe. Works as well.

[LostRuins]

That sounds confusing, there's no way it would have worked in CLBlast if its the same model, CLBlast doesn't support IQ4_XS which seems to be what you're using. Something must have changed in your run parameters.

[alan-l]

That sounds confusing, there's no way it would have worked in CLBlast if its the same model, CLBlast doesn't support IQ4_XS which seems to be what you're using. Something must have changed in your run parameters.

I'm running z_image_turbo-Q4_K.gguf from https://huggingface.co/leejet/Z-Image-Turbo-GGUF , flux1vae.safetensors for VAE, and Qwen3-4B-Instruct-2507-UD-Q4_K_XL.gguf for Clip-1.

I also run KoboldCPP from kcpp files, so the config from version to version remains constant.

[LostRuins]

@alan-l can you try this test build on Vulkan
https://github.com/LostRuins/koboldcpp/actions/runs/20708872505

let me know if it works or not, and share the output it prints to the terminal.

[alan-l]

@alan-l can you try this test build on Vulkan https://github.com/LostRuins/koboldcpp/actions/runs/20708872505

let me know if it works or not, and share the output it prints to the terminal.

Works for Z-Image Turbo at 512x512 for both Vulkan and CLBlast. Doesn't work for Qwen Image Edit Rapid AIO.

Logs attached as requested. Hope they help!

KCPP-Debug_Z-Image_Turbo_512_Vulkan.txt
KCPP-Output_Qwen_512_CLBlast.log
KCPP-Output_Qwen_512_CLBlast_GGUF.log
KCPP-Output_Qwen_512_Vulkan.log
KCPP-Output_Z-Image_Turbo_512_CLBlast.log
KCPP-Output_Z-Image_Turbo_512_CPU.log

[alan-l]

@alan-l can you try this test build on Vulkan https://github.com/LostRuins/koboldcpp/actions/runs/20708872505

let me know if it works or not, and share the output it prints to the terminal.

One more log from testing, using Q4_K version of safetensors.

KCPP-Output_Z-Image_Turbo_512_Vulkan.log

[LostRuins]

Alright can you try this build https://github.com/LostRuins/koboldcpp/actions/runs/20748615692 on vulkan z-image.

[alan-l]

Alright can you try this build https://github.com/LostRuins/koboldcpp/actions/runs/20748615692 on vulkan z-image.

Here you go! Let me know what else you need!

KCPP-Output_Qwen_512_CLBlast.log
KCPP-Output_Qwen_512_CLBlast_safetensors.log
KCPP-Output_Qwen_512_Vulkan.log
KCPP-Output_Qwen_512_Vulkan_safetensors.log
KCPP-Output_Z-Image_Turbo_512_Vulkan.log
KCPP-Output_Z-Image_Turbo_1024_Vulkan.log

[LostRuins]

So it worked for you then? I see no errors.

[alan-l]

So it worked for you then? I see no errors.

Thank you for the Z-Vulkan fix, @LostRuins ! Qwen image edit doesn't work under Vulkan, works under CLBlast. Should I open another ticket for QIEdit issue under Vulkan?