llama.cpp b10045 adds text-only slot save/restore for multi-modal models

The latest release of llama.cpp, version b10045, brings a new server capability: text-only slot save/restore with mtmd (multi-modal token detection). This feature allows the inference server to preserve and restore slots that contain only text, even when operating in a multi-modal context that also processes images or other modalities. Under the hood, mtmd likely stands for 'multi-modal token detection,' enabling the server to differentiate between text and non-text tokens when saving and loading conversation slots. This is a significant quality-of-life improvement for developers running multi-modal LLMs (e.g., LLaVA-style models) because it reduces memory waste and avoids unnecessary re-processing of text-only interactions when the model expects multi-modal input.

The release also continues llama.cpp's tradition of extensive platform support. Pre-built binaries are available for macOS (Apple Silicon with and without KleidiAI, Intel x64), Linux (x64 and arm64 on CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (x64 and arm64 on CPU, CUDA 12/13, Vulkan, OpenCL Adreno, HIP), Android (arm64 CPU), and iOS (XCFramework). The commit is signed with GitHub's verified signature (GPG key B5690EEEBB952194), ensuring authenticity. For developers building from source, the tag b10045 is ready for checkout. This update underscores llama.cpp's position as the go-to local LLM inference engine for professionals needing maximum performance across diverse hardware.