llama.cpp b9486 adds CUDA 13, ROCm 7.2, drops macOS Intel
New release supports CUDA 13 DLLs and ROCm 7.2, macOS Intel builds disabled.
ggml-org released llama.cpp b9486 on June 3, a maintenance update focusing on build system improvements and broader hardware compatibility. The release adds support for CUDA 13 DLLs (CUDA 13.3) on Windows x64 alongside existing CUDA 12.4, and introduces ROCm 7.2 on Ubuntu x64. Other GPU backends like Vulkan, OpenVINO, and SYCL remain supported across Linux and Windows.
Notably, the release disables several build targets that were previously experimental or problematic: macOS Intel (x64), openEuler (x86 and aarch64), and Windows x64 SYCL. This suggests a push toward stable, well-tested configurations. Android arm64 CPU builds and iOS XCFramework are included. The CI fix (disabling ccache for MSVC Windows release jobs) aims to reduce build failures. For developers, this means more reliable cross-platform deployment options for running LLaMA models locally.
- New CUDA 13 DLLs (13.3) and ROCm 7.2 support added
- macOS Intel (x64) builds disabled; Apple Silicon remains
- Windows arm64 CPU support introduced, openEuler and SYCL disabled
Why It Matters
llama.cpp b9486 streamlines GPU support for running local LLMs, prioritizing stable backends like CUDA 13 and ROCm 7.2.