llama.cpp Release b11010 Adds Vulkan Workaround for Nvidia Driver Bug
Open-source LLM inference framework llama.cpp released patch version b11010, introducing a targeted Vulkan backend workaround for an Nvidia driver bug in argsort_large.comp. This patch prevents unexpected crashes or errors when running LLM inference via the Vulkan backend on systems equipped with Nvidia graphics cards. The patch addresses pull request #28975 by modifying the compute shader logic in `argsort_large.comp`. Updated pre-compiled binaries are provided for various platforms, including Windows, Linux, macOS, iOS, and Android across CPU, CUDA, Vulkan, ROCm, OpenVINO, and SYCL backends.
## BACKGROUND
llama.cpp is a widely used C/C++ software library designed for high-performance Large Language Model (LLM) inference on consumer-grade hardware. Vulkan is a cross-platform, low-overhead API for graphics and compute that enables hardware-accelerated GPU inference across diverse GPU vendors.