The llama.cpp project released version b9992, which includes a refactor of the MMQ kernel configuration. This update specifically addresses and fixes the configuration for NVIDIA's Blackwell architecture while removing legacy code.

  • Refactored MMQ kernel configuration to improve CUDA support.
  • Fixed Blackwell GPU configuration issues.
  • Removed legacy code from the codebase.

This release provides updated binaries for macOS, Linux, Windows, Android, and openEuler across various hardware backends including CPU, Vulkan, ROCm, OpenVINO, SYCL, and HIP.