llama.cpp b10442 boosts Intel Xe Vulkan performance with coopmat fixes

llama.cpp, the open-source C++ library that powers local LLM inference everywhere from Raspberry Pi to high-end servers, just dropped release b10442. Maintained by ggml-org, this latest tag focuses heavily on the Vulkan backend, specifically targeting Intel Xe GPUs. The headline changes are new cooperative matrix (coopmat) optimizations: SHMEM_STRIDE_PAD and APPLY_SLM_A_RESHAPE are added for mul_mm operations on Intel Xe, alongside fixes to shared memory estimates and cacheline alignment. These tweaks are designed to squeeze more throughput out of Intel's integrated and discrete graphics when running quantized models like those in the mxfp4 format.

Notably, the release also addresses a real bug: an out-of-bounds read in kvalues_mxfp4 initialization after cacheline padding, which could cause crashes on certain hardware. The team restricted the SLM-A reshape optimization to the Intel Windows driver to avoid compatibility issues, and reverted the mxfp4 cacheline padding entirely to ensure stability. Like every llama.cpp release, b10442 ships prebuilt binaries for a massive matrix of targets—macOS Apple Silicon (with optional KleidiAI), Linux x64/arm64/s390x, Android, Windows x64/arm64, and GPU backends including Vulkan, CUDA 12/13, ROCm 7.14, OpenVINO, and SYCL. For developers running local AI on Intel Xe graphics, this is a meaningful step forward in reliability and speed.