llama.cpp Release b10952 Fixes SYCL oneDNN Scratchpad Memory Issue
llama.cpp version b10952 has been released featuring a targeted bug fix for SYCL execution backends. The update resolves an issue where oneDNN scratchpad memory allocation violated the expected pool free order. This bug fix ensures stable and predictable memory management when running llama.cpp on hardware accelerators leveraging SYCL and oneDNN, such as Intel GPUs. It prevents memory corruption or crashes that could occur during long-running inference sessions. The release includes automated pre-compiled binary packages across Windows, Linux, macOS, and Android systems for various CPU and GPU architectures. Notably, builds for macOS KleidiAI and certain openEuler hardware paths remain disabled in this release.
## BACKGROUND
llama.cpp is a high-performance, open-source C/C++ engine designed for running large language models across diverse user hardware. SYCL is an open, C++-based programming model for heterogeneous computing across various hardware accelerators, and oneDNN is an optimized deep learning performance library commonly used alongside SYCL on Intel architecture. Scratchpad memory serves as temporary working space during neural network operations and must be allocated and freed in strict order to avoid memory leaks or runtime errors.