b8721

The open-source project llama.cpp, maintained by the ggml-org, has released a new update identified by commit hash b8721. This is a targeted maintenance release focused on fixing a specific bug (#21633) in the model loading logic. The issue was that when selecting files for a GGUF format model—a common format for quantized Llama-family models—the system could incorrectly prioritize non-primary split files. Since GGUF models are often split into multiple parts for easier handling, this bug could lead to a failure to load the model correctly. The fix, contributed by Adrien Gallouët, ensures the loader correctly identifies and uses the primary file, making the loading process more robust, especially for users working with large, multi-file models.

Alongside this core fix, the release includes a comprehensive set of 27 pre-built binaries, showcasing the project's extensive cross-platform support. These binaries cater to a wide range of hardware and operating systems, including macOS on both Apple Silicon and Intel, various Linux distributions (with CPU, Vulkan, ROCm, and OpenVINO backends), Windows (with CPU, CUDA 12/13, Vulkan, SYCL, and HIP support), and even specialized builds for openEuler with Huawei Ascend AI processor support. This single commit, while small in code change, significantly improves the stability of a fundamental operation for thousands of developers and researchers who rely on llama.cpp for efficient, local LLM inference.