PreonStar inference engine update targets Vulkan and dense 27B models on AMD BC-250
PreonStar, a native inference engine for Qwen models, receives updates for Vulkan, dense 27B speeds, and IsValorum compatibility.
The latest commit to PreonStar, a small native C99 inference engine for Qwen models designed for hardware like the AMD BC-250 board, brings updates for its Vulkan backend. The revision updates documentation regarding IsValorum model compatibility and benchmarks for dense 27B models.
Changes introduce short-batch prefill, fixed-row MTP verification, and long-context decode attention paths for Vulkan. PreonStar runs models like Qwen3.6-35B-A3B and Qwen3.8-27B without Python dependencies or heavy frameworks, relying on a direct C codebase built with make.
To try the update on Linux with a Vulkan-capable setup:
- Install dependencies such as
build-essential,glslang-tools, andlibvulkan-dev. - Clone the repository and run
make -j"$(nproc)"to compile the binaries. - Download a model using
./download_model.sh q2-imatrixand execute./p36to run local prompts.