llama.cpp Release b11169 Fixes Token ID Parsing Truncation in llama-grammar
Open-source LLM inference framework llama.cpp released patch version b11169. The update resolves a numeric truncation issue during token_id parsing inside the llama-grammar module (PR #29382). Resolving numeric truncation prevents parsing bugs when enforcing grammar rules on modern LLMs with large vocabulary sizes. This ensures reliable structured text generation, such as JSON outputs, during constrained decoding. The patch specifically targets token ID parsing logic within the `llama-grammar` component. Updated pre-built binaries have been published across supported platforms, including macOS, Linux, Windows, and Android.
## BACKGROUND
`llama.cpp` is a widely used C/C++ library designed for high-performance LLM inference across diverse hardware platforms. Its `llama-grammar` feature utilizes GBNF (GGML Backus-Naur Form) rules to constrain model sampling, forcing outputs to adhere strictly to defined formats such as valid JSON or custom grammars.