llama.cpp with ZLUDA fails on Qwen3.5-35B-A3B-UD-Q4_K_XL model

zluda_trace logs (tarball/zip file)

zluda.zip

Description

Qwen3.5-35B-A3B quantized version from Unsloth. Fails on model prewarm.

Additional info:

• With "--flash-attn off" prewarm was successful, but again fails after inference starting.

I also tried Qwen3.5-27B-UD-Q4_K_XL.

• Without "--flash-attn off" it fails the same error.
• With "--flash-attn off" works fine.

All models work well with official llama.cpp HIP b8329 build.

Steps to reproduce

1. Install all dependecies. HIP SDK 7.1.1
2. Build llama.cpp from b8370 with CUDA SDK 13.2. Command cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="86" -DGGML_CUDA_FORCE_CUBLAS=true in Visual Studio 2022 terminal
3. cmake --build ./build
4. Copy all libraries from zluda path into llama.cpp path.
5. Run llama.cpp ./zluda/zluda.exe -- ./llama_cuda_zluda/llama-server.exe --model ./Qwen3.5-35B-A3B-GGUF/Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf --mmproj ./Qwen3.5-35B-A3B-GGUF/mmproj-F16.gguf --temp 1.0 --top-p 0.95 --min-p 0.00 --top-k 20 --host 0.0.0.0 --port 23321

ZLUDA version

v6-preview.62

Operating System

Windows 11 version 10.0.26200

GPU

AMD Radeon RX 7900 XTX