●Change8

b8804

Breaking Changes
📦 llama-cppView on GitHub →
⚠ 1 breaking✨ 2 features

Summary

This release introduces a security enhancement by requiring explicit opt-in for CUDA P2P access and provides extensive pre-built binaries across multiple platforms and hardware accelerators.

⚠️ Breaking Changes

  • CUDA P2P (Peer-to-Peer) access is now disabled by default and requires explicit opt-in to enhance security and control.

Migration Steps

  1. If your application relied on CUDA P2P access, you must now explicitly enable it.

✨ New Features

  • Added support for CUDA P2P access via explicit opt-in.
  • Provided pre-built binaries for macOS (Apple Silicon and Intel), Linux (various architectures and backends like Vulkan, ROCm 7.2, OpenVINO), Windows (CPU, CUDA 12.4, CUDA 13.1, Vulkan, SYCL, HIP), and openEuler.