llama.cpp Release b10938 Adds Workaround for NVIDIA Vulkan Driver Bug
Release b10938 of llama.cpp adds a temporary workaround for an NVIDIA driver bug affecting Vulkan execution. The patch wraps Vulkan queue submission operations in a mutex to prevent internal synchronization failures. This update prevents crashes and synchronization issues for users running LLM inference using the Vulkan backend on NVIDIA hardware. It ensures reliable cross-platform execution without waiting for NVIDIA to issue an official driver fix. The driver bug causes internal synchronization failures when two command queues on the same `VkDevice` submit work simultaneously. The release introduces a mutex wrapper around `queuesubmit` calls as a temporary fix.
## BACKGROUND
llama.cpp is a popular open-source C/C++ framework for running Large Language Model (LLM) inference locally across various hardware backends. Vulkan is a low-level, cross-platform graphics and compute API that manages GPU operations through command queues submitted to execution devices.