llama.cpp b9813 adds Intel Xe-LPG Plus GPU support with coopmat acceleration

llama.cpp, the popular open-source C++ library for LLM inference, has released version b9813 with significant GPU support enhancements. The update introduces the INTEL_XE1 architecture enum alongside the existing INTEL_PRE_XE2 enum, specifically enabling cooperative matrix operations (coopmat) on Intel Xe-LPG Plus GPUs (Xe1-ARLH). This feature, critical for accelerating matrix multiplications in neural networks, is now available on both integrated and discrete Intel GPUs using the Vulkan backend. The patch also refines bfloat16 handling, removes a driver workaround, and adds a Windows driver version check for stability. Contributions from Intel engineers Jie Xia and Russell Liu ensure compatibility with Intel's hardware and driver stack.

The release has been tested across multiple platforms including macOS (Apple Silicon and Intel), Linux (x64, arm64, Vulkan, ROCm, OpenVINO, SYCL), Windows (x64, arm64, CUDA, Vulkan, OpenVINO, HIP), and Android arm64. Notably, macOS Apple Silicon with KleidiAI enabled is currently disabled, and some other configurations remain untested. For developers and AI practitioners, this means faster, more efficient LLM execution on Intel Xe-LPG Plus GPUs without sacrificing portability. As llama.cpp continues to expand hardware support, the open-source community gains better performance for local LLM deployments across a wider range of consumer and enterprise hardware.