llama.cpp b10436 patches multimodal bugs and reverts GGUF fix

llama.cpp, the wildly popular open-source project from ggml-org, shipped a new tagged release b10436 on August 15 via GitHub Actions. With 124k stars and 21.7k forks, it remains the go-to engine for running LLMs locally on consumer hardware. This update is a classic maintenance release: it applies fixes for mtmd (a multimodal transformer architecture) and common utilities, as referenced in PR #27071. Notably, the release also reverts a recent GGUF fix—likely to address regressions that surfaced after the previous change.

For users, b10436 delivers prebuilt binaries for a wide range of platforms, including macOS (Apple Silicon and Intel), Linux (x86/ARM/s390x with CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, OpenCL, Vulkan, SYCL, ROCm), Android, and iOS. This means developers can pull an updated binary and continue running models locally without compiling from source. While no headline-grabbing feature lands here, the patch-and-revert pattern underscores the project's active refinement of core file formats and model support—important for anyone relying on llama.cpp for production workflows or local experimentation.