llama.cpp Release b10749 Adds Context Autoscaling for YaRN Scaling
Open-source LLM inference engine llama.cpp released build b10749, introducing automatic context scaling for the training context length parameter (n_ctx_train) when YaRN scaling parameters are specified. This patch simplifies context window management when running models with extended context techniques, preventing user configuration errors when extending token limits. The release integrates Pull Request #28030, which automatically computes the effective training context length from the specified YaRN context scaling configuration during model context setup.
## BACKGROUND
llama.cpp is a C/C++ inference framework designed to run Large Language Models locally across diverse hardware platforms. YaRN (Yet another RoPE extensioN) is a method for extending the context window of Rotary Position Embedding (RoPE) based models with minimal additional fine-tuning.