llama.cpp b10050 fixes Vulkan async race for stable local LLM inference
New release addresses Vulkan race condition on transfer queue async copies, boosting stability.
The ggml-org/llama.cpp project, a leading C++ implementation for running large language models locally, has released version b10050. This update addresses a critical race condition in Vulkan async copies when using a transfer queue, with the fix syncing on event_wait to prevent data races. The release, tagged b10050 and signed with GitHub's verified signature, ensures smoother and more reliable GPU-accelerated inference for users leveraging the Vulkan backend.
Beyond the Vulkan fix, llama.cpp b10050 continues its tradition of broad platform support. Official builds are available for macOS (both Apple Silicon and Intel x64), iOS (XCFramework), various Linux distributions (x64 and arm64 with CPU, Vulkan, ROCm, OpenVINO, and SYCL backends), Android (arm64 CPU), and Windows (x64/arm64 CPU, CUDA 12/13, Vulkan, OpenCL Adreno, OpenVINO, SYCL, and HIP). The openEuler Linux variant also receives builds. This release ensures that developers and enthusiasts running local AI on anything from a Raspberry Pi to a high-end GPU setup benefit from the latest stability improvements.
- Fixes Vulkan race condition in async copies via event_wait synchronization
- Provides builds across macOS, iOS, Linux, Windows, Android, and openEuler
- Supports backends including CPU, Vulkan, CUDA, ROCm, OpenVINO, SYCL, HIP, and OpenCL
Why It Matters
Improves reliability of Vulkan-accelerated local LLM inference, critical for self-hosted AI applications.