llama.cpp Release b11376 Fixes Flaky Float16 CI Test
The open-source LLM inference engine llama.cpp released version b11376. This update resolves a flaky float16 (f16) test in continuous integration by adjusting the tolerance threshold for fused addition operations. Fixing non-deterministic or flaky automated tests improves continuous integration reliability and prevents false test pipeline failures. This ensures smoother code contribution and maintenance for developers without affecting runtime inference functionality. The release specifically addresses floating-point precision tolerances for the `ADD_ADD` float16 operation under PR #29904. Pre-built binaries are provided across multiple platforms including Linux, macOS, Windows, Android, and Snapdragon target architectures.
## BACKGROUND
llama.cpp is a popular C/C++ library designed for local LLM inference across diverse hardware platforms. In continuous integration (CI) systems, float16 mathematical operations can sometimes cause false test failures due to minor floating-point hardware differences, requiring calibrated tolerance bounds.