llama.cpp b10435 fixes quadratic Jinja slowdown in template parsing

llama.cpp, the widely-used open-source C++ runtime for running large language models on local hardware, has released a new maintenance version: b10435. The update from ggml-org focuses on a single but impactful performance fix: eliminating quadratic cost in gather_string_parts for Jinja templates. Jinja is the templating engine used by many models (like Llama and Mistral) to format chat prompts. Previously, as template parts accumulated, the gathering operation scaled quadratically, meaning longer prompts caused disproportionately longer processing times.

This patch replaces that inefficient logic with a linear-time approach, significantly reducing CPU overhead during prompt tokenization. While the benchmark improvement depends on template complexity and length, users running chat models with custom instruction templates should notice lower latency and better responsiveness. The release ships prebuilt binaries across major platforms—including macOS (Apple Silicon and Intel), Linux (x86/ARM/s390x), Windows (CPU/CUDA/Vulkan), Android, and iOS—with support for hardware accelerators like Vulkan, CUDA 12/13, ROCm, OpenVINO, and SYCL. This update is particularly relevant for developers building local AI assistants or using llama.cpp as a backend for tools like Ollama and LM Studio.