Eval bug:[5808-] qwen3 30B vulkan run with GGG

[wszgrcy]

Name and Version

load_backend: loaded RPC backend from D:\139-conf\llama\llama\llama-b5844-bin-win-vulkan-x64\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 7900 XT (AMD proprietary driver) | uma: 0 | fp16: 1 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
load_backend: loaded Vulkan backend from D:\139-conf\llama\llama\llama-b5844-bin-win-vulkan-x64\ggml-vulkan.dll
load_backend: loaded CPU backend from D:\139-conf\llama\llama\llama-b5844-bin-win-vulkan-x64\ggml-cpu-haswell.dll
version: 5844 (17a1f0d)
built with clang version 19.1.5 for x86_64-pc-windows-msvc

Operating systems

Windows

GGML backends

Vulkan

Hardware

5900x+7900xt

Models

qwen3 30b

Problem description & steps to reproduce

in 5688 version ,The chat can proceed normally.But in the latest version, conversations are limited to the first few hundred characters (300-500?)
Then it will output GGGGG at a slow speed (Approximately every ten seconds, a G is generated)

First Bad Commit

I think is 5808
I tested multiple versions and found that it worked fine under 5804, but then the issue mentioned above appeared

Relevant log output

llama.log

[bennmann]

Happens to me too for b5845 for multiple model types

example command:
llama-server.exe -m F:\hunyuan-a13b-instruct-hf-WIP-IQ4_NL.gguf -ngl 33 -c 16384 --override-tensor "([1][0-9]|[2-9][0-9]|[0-9][0-9][1-9]).ffn_.exps.=CPU,([0-9]).ffn._exps.=Vulkan0" --flash-attn --cache-type-k q8_0 --cache-type-v q8_0 --batch-size 128

Get about 256 tokens in or so then repetition of a single token in llama server GUI frontend.

EDIT: AMD RX 6900 XT in case it helps