llama.cpp Releases Version b10286 with Grammar Repetition Adjustment
The llama.cpp project has released version b10286, which introduces a minor update to its grammar parsing logic. Specifically, it degrades the maximum repetition limit to unbounded when the value is set to 2000 or greater. This release is a minor, routine update that optimizes how llama.cpp handles large repetition constraints in formal grammars. It helps prevent potential parsing issues or overhead when developers define highly repetitive patterns for constraining LLM outputs. The change is implemented in pull request #26613, which automatically treats any grammar repetition limit of 2000 or more as unbounded. Additionally, this release disables the KleidiAI-enabled macOS Apple Silicon build due to an ongoing pull request.
## BACKGROUND
llama.cpp uses GGML Backus-Naur Form (GBNF) grammars to constrain the output of large language models, ensuring they adhere to specific formats like JSON. These grammars define rules for how tokens can be generated, including repetition limits for specific patterns. KleidiAI is an open-source library by Arm designed to optimize AI workload performance on Arm CPUs.