b8770
Latest update patches crash when sending images smaller than 2x2 pixels across all major platforms.
The open-source project Llama.cpp, maintained by ggml-org, has rolled out a new stable release tagged b8770. This version is primarily a maintenance update that resolves a specific but critical bug (#21711) where the system would crash when attempting to process an image smaller than 2x2 pixels. This fix is crucial for applications using multimodal features or any image input pipeline, preventing unexpected failures from malformed or tiny image data.
Alongside the bug fix, the release provides a comprehensive suite of pre-built binaries for developers, eliminating the need for manual compilation. The supported platforms now include macOS for both Apple Silicon (with optional KleidiAI acceleration) and Intel, iOS via XCFramework, various Linux distributions (with CPU, Vulkan, ROCm 7.2, and OpenVINO backends), Windows (with CPU, CUDA 12.4/13.1, Vulkan, SYCL, and HIP support), and openEuler for specific Huawei Ascend hardware. This broad compatibility underscores the project's commitment to being a universal, high-performance inference engine for running models like Meta's Llama 3 locally.
- Fixes crash bug (#21711) in multimodal (mtmd) processing for images under 2x2 pixels.
- Provides pre-built binaries for 27 distinct platform/backend combinations including macOS, Windows, Linux, and iOS.
- Expands support for specialized hardware like Huawei Ascend (openEuler) and adds CUDA 13.1 DLLs for Windows.
Why It Matters
Ensures robust, crash-free local AI inference for developers building applications with image inputs on any major platform.