llama.cpp b10215 Released with Vulkan Driver Checks for Intel GPUs
The llama.cpp project has released version b10215, which introduces a driver version check for Windows Intel GPUs to mitigate Vulkan-related crashes. The update also removes a previous crash guard for Intel GPUs, as the issue has been resolved starting from driver version 32.0.101.8860. This release improves the stability of local LLM inference on Windows systems utilizing Intel integrated or discrete graphics via the Vulkan API. It ensures users with updated drivers can run models without unnecessary performance guards, while protecting those on older drivers from crashes. The driver version check converts the version directly from `driverVersion` rather than parsing it as a string, and the update allows GPUs other than the Xe2+ architecture. Additionally, this release disables the KleidiAI-enabled build for macOS Apple Silicon due to an open pull request.
## BACKGROUND
llama.cpp is a popular open-source C/C++ library designed for efficient local inference of large language models (LLMs) like Llama. Vulkan is a cross-platform graphics and compute API that llama.cpp uses to accelerate model execution on various GPUs, including those from Intel, AMD, and Nvidia.