llama.cpp b9803 ships OpenCL fix, supports Apple Silicon to Windows

ggml-org's llama.cpp, the leading open-source C/C++ implementation for running large language models locally, has released version b9803. The update addresses a specific issue in the OpenCL backend: flushing profiling batches at shutdown for incomplete batches (issue #25016). This ensures that performance profiling data is correctly completed when GPU inference tasks are interrupted or terminated. The release supports a vast array of hardware, including macOS Apple Silicon (arm64) both with and without KleidiAI acceleration, macOS Intel (x64), iOS via XCFramework, Ubuntu (x64 and arm64) with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL FP32/FP16 backends, Windows (x64 CPU, arm64 CPU, OpenCL Adreno, CUDA 12/13, Vulkan, SYCL, HIP), Android arm64, and openEuler with ACL Graph support.

With over 118,000 stars on GitHub, llama.cpp remains the go-to solution for developers and researchers deploying models like LLaMA, Mistral, and Gemma on consumer hardware. This point release, though minor, improves the reliability of performance profiling—a crucial feature for optimization and debugging during local model inference. The extensive platform coverage means that from Apple Silicon Macs to Windows gaming GPUs and Linux servers, users can now trust that their GPU profiling data is complete even after unexpected shutdowns. This update reinforces llama.cpp's position as the most versatile and community-driven LLM inference engine available.