b8626

The open-source project llama.cpp, maintained by ggml.ai, has released a new update identified by commit hash b8626. This patch specifically targets a memory leak in the OpenCL compute backend when processing models using the `q8_0` quantization format on devices with Qualcomm Adreno GPUs. The fix, detailed in pull request #21212, is crucial for mobile and embedded developers using Snapdragon platforms, as memory leaks can lead to application instability and crashes during prolonged AI inference sessions.

Alongside this critical fix, the release provides a full suite of updated pre-built binaries for major platforms. For Apple users, it includes builds for macOS on both Apple Silicon (arm64) and Intel (x64) architectures, as well as an iOS XCFramework. Linux developers can access builds for Ubuntu across x64 and arm64 CPUs, with optional backends for Vulkan, ROCm 7.2, and OpenVINO. Windows support is equally comprehensive, covering x64 and arm64 CPU builds, plus specialized versions for CUDA 12.4, CUDA 13.1, Vulkan, SYCL, and HIP. This ensures developers can easily deploy efficient, quantized large language models like Llama 3 across a wide range of hardware without compilation overhead.