llama.cpp b9570 adds WebGPU formatting, CUDA 13 support
116K-starred open-source LLM engine expands build matrix with CUDA 13 and openEuler
llama.cpp, the popular open-source C/C++ library for LLM inference, tagged release b9570 on June 9. The standout addition is a clang-format job for ggml-webgpu, enforcing consistent code style across the WebGPU backend. This release also ships with a verified signature (GPG key B5690EEEBB952194) and continues the project's tradition of broad platform support.
Hardware support sees notable expansion: Windows builds now include CUDA 13.3 DLLs (alongside the existing CUDA 12.4 option), and Linux users gain ROCm 7.2, OpenVINO, and SYCL FP32 (disabled) targets. openEuler joins the list with x86 and aarch64 builds using ACL Graph (310p and 910b). GPU backends cover Vulkan and HIP, while CPU builds span Ubuntu s390x, arm64, and more. Mobile platforms are not forgotten — iOS XCFramework and Android arm64 (CPU) are included. This breadth makes llama.cpp a go-to for developers running LLMs on diverse hardware.
- New ggml-webgpu clang-format job ensures code consistency in the WebGPU backend.
- Windows builds now support CUDA 13.3 DLLs alongside the existing CUDA 12.4 option.
- openEuler Linux with ACL Graph (910b, 310p) and ROCm 7.2 join the supported targets.
Why It Matters
llama.cpp continues to lower barriers for local LLM inference across an ever-wider range of hardware.