Eval bug: Qwen3 Q4_0 not working with SYCL

[invent00]

Name and Version

version: 5215(5f5e39e)
built with MSVC 19343.34808.0

Operating systems

Windows

GGML backends

SYCL

Hardware

Core Ultra 5 125U 32GB mem(ThinkPad X1 Carbon Gen12)
Driver Version: 32.0.101.6739

Models

Qwen3-4B-gguf Q4_0 (https://huggingface.co/unsloth/Qwen3-4B-GGUF/tree/main)

Problem description & steps to reproduce

When attempting inference with the model, the screen briefly goes black and fails to function properly. However, the Q4_K_M model operates normally.

in addition, cuda build (cu11.7,b5215) work properly with Q4_0.

how to reproduce:

1. llama-cli.exe -ngl 99 -m Qwen3-4B-Q4_0.gguf
2. input question
3. Black out occur

in event log, llama-cli.exe shows following application error

First Bad Commit

No response

Relevant log output

1. llama-cli.exe -ngl 99 -m Qwen3-4B-Q4_0.gguf
2. input question
3. Black out occur

[Sketchfellow]

I've also had issues with Q4_0 quants on SYCL resulting in the screen going black and crashing for my Arc A770M. I experienced this on Gemma 3 12B QAT as well as Llama 2 7B when running performance benchmarks. I believe the SYCL Q4_0 reorder optimizations resulted in this as setting GGML_SYCL_DISABLE_OPT=1 allowed things to run normally again.

[qnixsynapse]

I believe the SYCL Q4_0 reorder optimizations resulted in this as setting GGML_SYCL_DISABLE_OPT=1 allowed things to run normally again

cc @Rbiessy @NeoZhangJianyu @Alcpz ^

[invent00]

Hi @Sketchfellow,
Thank you for your advice. after set GGML_SYCL_DISABLE_OPT=1 , it works properly.

[Alcpz]

I've been able to reproduce the issue, but only on Windows. Linux seems unaffected. As reported, GGML_SYCL_DISABLE_OPT=1 works without problem. There seems to be something wrong with the reorder, but I would need to have a deeper look at it.