Change8

b9571

Breaking Changes
📦 llama-cppView on GitHub →
5 breaking2 features🔧 1 symbols

Summary

This release focuses on updating pre-built binaries across multiple platforms and removing the specific case handling for GGML_TYPE_Q4_K in mvvq.cu. Several platform-specific builds have been disabled.

⚠️ Breaking Changes

  • The case for GGML_TYPE_Q4_K has been removed in mvvq.cu, which may break compatibility if code explicitly relied on this specific quantization type handling.
  • macOS Apple Silicon (arm64) build with KleidiAI enabled is DISABLED.
  • Ubuntu x64 (SYCL FP32) build is DISABLED.
  • Windows x64 (SYCL) build is DISABLED.
  • openEuler builds are DISABLED.

Migration Steps

  1. If relying on GGML_TYPE_Q4_K handling in mvvq.cu, update quantization logic as this specific case was removed.
  2. Users requiring KleidiAI on macOS arm64, SYCL FP32 on Ubuntu, SYCL on Windows, or openEuler builds should note these are currently disabled and may require manual compilation or alternative solutions.

✨ New Features

  • New pre-built binaries provided for various platforms including macOS (Apple Silicon/Intel), Linux (CPU, Vulkan, ROCm 7.2, OpenVINO), Android, and Windows (CPU, CUDA 12.4, CUDA 13.3, Vulkan, HIP).
  • CUDA 12.4 and CUDA 13.3 DLLs provided separately for Windows builds.

Affected Symbols