ggerganov/llama.cpp Released Version b11757 with CUDA Graph Updates
Release b11757 of llama.cpp introduces a CUDA update that clears CUDA graphs after legacy memory pool flushes alongside standard multi-platform binaries. This routine update improves memory management and stability for users running large language models on NVIDIA hardware via CUDA. The release includes pull request #26771 which handles the clearing of CUDA graphs and provides precompiled build artifacts for various platforms including macOS, Linux, Windows, and Android.
## BACKGROUND
llama.cpp is an open-source library that enables efficient inference of large language models in C/C++, serving as the core engine for many local AI tools. CUDA graphs are a feature of NVIDIA's CUDA Toolkit designed to reduce GPU launch overhead by capturing sequences of operations into an executable graph.