llama.cpp b9843 reverts synchronisation fix for split compute stability
New release rolls back a controversial optimization to address potential crashes in multi-GPU setups.
The latest llama.cpp release, b9843, marks a significant stability-focused update for the popular open-source LLM inference engine. The core change is a reversion of commit #20793, which had introduced a less synchronized approach during split compute operations. That earlier optimization aimed to reduce overhead when distributing model computation across multiple GPUs, but apparently introduced regressions that could cause hangs or incorrect results. By rolling back to the previous synchronization logic, b9843 prioritizes correctness and reliability over raw throughput, especially in multi-device configurations.
This release is signed with GitHub's verified signature (GPG key B5690EEEBB952194) and supports an extensive range of platforms. Apple users get builds for macOS Apple Silicon (with and without KleidiAI), Intel x64, and iOS XCFramework. Linux users have options for x64 CPU, arm64 CPU, s390x CPU, and GPU backends: Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16. Windows offers x64 CPU, arm64 CPU, and GPU variants including CUDA 12, CUDA 13, Vulkan, OpenVINO, SYCL, and HIP. Android arm64 CPU and Windows arm64 with OpenCL Adreno are also available. Developers running LLMs like Llama, Mistral, or Gemma on local hardware will appreciate the focus on deterministic behavior, even if it means slightly more synchronization overhead.
- Reverts commit #20793 that reduced synchronizations during split compute, restoring original sync logic for stability.
- Release b9843 was created on June 30 and signed with GitHub's verified GPG key (B5690EEEBB952194).
- Available across 20+ build targets including macOS, Linux, Windows, and Android, with multiple GPU backends (CUDA 12/13, Vulkan, ROCm, OpenVINO, SYCL, HIP).
Why It Matters
Prioritises correctness over speed in multi-GPU LLM inference, preventing silent errors for production workloads.