Eval bug: AMD driver 25.11.1 crashes with llamacpp built with Vulkan SDK 1.4.328.1 #17432
Description
Name and Version
.\llama-cli.exe --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon(TM) 8060S Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
version: 7072 (e33e9c0)
built with MSVC 19.44.35213.0 for x64
Operating systems
Windows
GGML backends
Vulkan
Hardware
Ryzen AI Max+ 395 + 128GB Ram
Models
any models, for example: Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf
Problem description & steps to reproduce
Latest master llamacpp (21.11.25) built with Vulkan SDK DK 1.4.328.1.
Model loads fine, but with ctx-size > 512 the AMD driver 25.11.1 crashed with tg step:
First Bad Commit
Relevant log output
llama-bench.exe -m "C:\Users\{{name}}\.cache\huggingface\hub\models--unsloth--Qwen3-Coder-30B-A3B-Instruct-GGUF\snapshots\7ce945e58ed3f09f9cf9c33a2122d86ac979b457\Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf" -ngl 99 --flash-attn on -d 512,4096 -n 128
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon(TM) 8060S Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen3moe 30B.A3B Q4_K - Medium | 17.28 GiB | 30.53 B | Vulkan | 99 | pp512 @ d512 | 753.24 ± 91.97 |
| qwen3moe 30B.A3B Q4_K - Medium | 17.28 GiB | 30.53 B | Vulkan | 99 | tg128 @ d512 | 79.51 ± 0.69 |
| qwen3moe 30B.A3B Q4_K - Medium | 17.28 GiB | 30.53 B | Vulkan | 99 | pp512 @ d4096 | 717.61 ± 2.73 |
C:\Users\{{name}}\AppData\Local\lemonade_server\bin\llama\lastest_vulkan>./llama-cli.exe
^^ no tg128@d4096 in the output and AMD driver pop up appears with crash message !