b8913
Llamacpp's latest release fixes a critical shader bug and extends hardware support.
The llama.cpp project, a popular open-source library for running large language models locally, has released version b8913. This release primarily fixes a shader bug related to buffer aliasing in the RMS fuse operation, which could cause incorrect results or crashes on certain GPU configurations. The fix, implemented in commit e5f070a, addresses a subtle memory management issue in the Vulkan and other GPU backends.
Beyond the bug fix, this release significantly expands the project's cross-platform support. Pre-built binaries are now available for macOS Apple Silicon (both standard and KleidiAI-optimized), macOS Intel, iOS XCFramework, Windows x64 and ARM64, and multiple Linux variants including Ubuntu (x64, ARM64, s390x) with support for Vulkan, ROCm 7.2, OpenVINO, and SYCL. This broad support ensures developers and end-users can run LLMs efficiently on nearly any modern hardware, from consumer laptops to enterprise servers.
- Fixes a buffer aliasing bug in the RMS fuse shader for GPU inference.
- Adds pre-built binaries for macOS Apple Silicon (with KleidiAI), Windows ARM64, and Linux s390x.
- Supports multiple GPU backends: CUDA 12/13, Vulkan, ROCm 7.2, OpenVINO, and SYCL.
Why It Matters
llama.cpp continues to democratize local LLM inference by fixing bugs and expanding hardware support for developers and users.