llama.cpp b10451 adds LoRA bounds check to prevent memory bugs

On August 16, ggml-org released llama.cpp b10451, the latest iteration of the popular C/C++ inference engine for running local LLMs. The headline change is a validation fix for LoRA (Low-Rank Adaptation) tensor data — specifically, a new bounds check ensures that tensor data in a LoRA adapter file is within the file's declared boundaries before loading. This addresses PR #27056, which lives in src/llama-adapter.cpp and was co-authored with Sigbjørn Skjæret from Hugging Face. Since LoRA adapters are commonly shared via Hugging Face Hub and other third-party sources, malformed or malicious files could previously cause out-of-bounds memory access, leading to crashes or potential security exploits. This patch closes that hole and is a recommended upgrade for anyone loading fine-tuned adapters.

The release also shows llama.cpp's massive cross-platform footprint. Prebuilt binaries are available for macOS Apple Silicon (including a KleidiAI-enabled variant), iOS XCFramework, Linux on x64/arm64/s390x, Windows x64/arm64, and Android arm64. Backend support spans CPU, Vulkan, ROCm, OpenVINO, SYCL, OpenCL Adreno, and CUDA 12/13, with CUDA 13 builds marked as preview. Several targets (openEuler, ROCm 7.14) are disabled in this release, likely due to build breakages. The commit is signed with a verified GitHub GPG key, and builds are distributed through llama.app. For developers running local models or production inference, b10451 is a low-risk, high-value patch that keeps LoRA workflows safe and stable across every major platform.