llama.cpp Release b10903 Fixes Vulkan Backend Memory Issues
llama.cpp release b10903 resolves a data race condition in the inner loop of the argsort function and fixes out-of-bounds (OOB) memory access in argsort_large within the Vulkan backend (PR #28705). Fixing memory corruption bugs improves stability for users running LLMs on cross-platform GPUs via Vulkan, while helping ensure reliable build passes in continuous integration (CI) pipelines. The data race in argsort was identified using Vulkan Validation Layers (VVL), while the OOB access bug in argsort_large was patched to eliminate suspected automated CI test failures despite being difficult to reproduce locally.
## BACKGROUND
llama.cpp is a popular open-source C/C++ framework for running LLM inference locally across hardware backends including CPU, CUDA, and Vulkan. Vulkan is a low-level graphics and compute API that gives developers explicit control over GPU operations, where tools like Vulkan Validation Layers (VVL) help detect illegal memory accesses and thread synchronization bugs.