llama.cpp Release b11297 Adds Coerced Array Attribute Support in Jinja Templates
llama.cpp release b11297 introduces support for coerced array attributes within its internal Jinja template parser via pull request #29574. The automated build release includes updated unit tests along with pre-compiled binaries for Linux, Windows, macOS, Android, and multiple hardware backends. Improving Jinja template handling enhances parity with Hugging Face's formatting standard for modern Large Language Models (LLMs). This helps prevent prompt parsing failures when formatting complex multi-turn conversations and tool-use calls during local LLM inference. The patch specifically implements handling for array attributes subjected to type coercion inside the C++ Jinja parsing logic, accompanied by dedicated test cases. The binary distribution provides builds across CUDA 12/13, Vulkan, ROCm, OpenVINO, SYCL, and Snapdragon processors.
## BACKGROUND
llama.cpp is a widely used open-source C/C++ library designed for efficient local inference of Large Language Models on consumer hardware. It relies on Jinja2 templates to convert structured user and system messages into formatted text prompts expected by specific AI model architectures.