b8948

The llama.cpp project, an open-source library for running large language models locally, has released version b8948. This maintenance update addresses a type casting bug in the unaccounted memory calculation, improving memory tracking accuracy during LLM inference. The fix ensures that developers can more precisely monitor and manage memory usage when running models on local hardware, which is critical for performance and stability, especially on resource-constrained systems.

b8948 significantly expands platform support, adding new builds for ROCm 7.2 on AMD GPUs, OpenVINO for Intel hardware, and SYCL for both FP32 and FP16 precision on Windows. It also introduces HIP support for AMD GPUs on Windows and openEuler builds for both x86 and aarch64 architectures. Existing builds for macOS (Apple Silicon, Intel), iOS, Linux (Ubuntu, s390x), Android, and Windows (CPU, CUDA, Vulkan) remain supported. This broad compatibility makes llama.cpp more versatile for AI developers and enthusiasts running models across diverse hardware setups.