Eval bug: gemma-3n-E4B-it-Q8_0.gguf is speaking nonsense

[rowheel]

Name and Version

$ llama-server --version
version: 5995 (9b8f3c6c)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

Operating systems

Linux

GGML backends

Vulkan

Hardware

CPU: AMD Ryzen 7 8845H (16) @ 5.14 GHz
GPU: AMD Radeon 780M Graphics [Integrated]

Models

ggml-org/gemma-3n-E4B-it-GGUF

Problem description & steps to reproduce

$ llama-cli -ngl 255 -m gemma-3n-E4B-it-Q8_0.gguf -c 0
> Hello, kindly make a self-introduction
Hello, kindly, and *for* you invisible, tobitoushtem pocketmemsimn/b! Helloz! 👋V‌-‌[SymbolX-lockH0\000:F1]‌ 👋�-██_�7�7��8������𘬏���𬑬�𐏬�𐔠�𐧬�𐔊𠬔𘛈𠺠�𐬬�𐧔�𐔊��͋�𐭠��𐬓�𐬊�𐠘��𐬔��𐬑Helloz! I'm Gemm
a> o-tecaiaa de Blob-T
Helloz! 👋�-�-�-��ら�𿠠�𿑠���𿠺�𿠠���𿬠�𿣠���
> 

First Bad Commit

No response

Relevant log output

$ llama-server --version
version: 5995 (9b8f3c6c)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
$ llama-cli -ngl 255 -m gemma-3n-E4B-it-Q8_0.gguf -c 0
> Hello, kindly make a self-introduction
Hello, kindly, and *for* you invisible, tobitoushtem pocketmemsimn/b! Helloz! 👋V‌-‌[SymbolX-lockH0\000:F1]‌ 👋�-██_�7�7��8������𘬏���𬑬�𐏬�𐔠�𐧬�𐔊𠬔𘛈𠺠�𐬬�𐧔�𐔊��͋�𐭠��𐬓�𐬊�𐠘��𐬔��𐬑Helloz! I'm Gemm
a> o-tecaiaa de Blob-T
Helloz! 👋�-�-�-��ら�𿠠�𿑠���𿠺�𿠠���𿬠�𿣠���
> 

[0cc4m]

I think this is the same as #14525

[rowheel]

I think this is the same as #14525

Yup, looks like this is a Vulkan over APU specific issue - running E4B solely on CPU is fine, speaking nonsense when it's accelerated with Vulkan.

Should I close this issue as duplicated? Or leave it open until the other thread revolves.

[0cc4m]

Please participate in the other thread to help with finding a solution.