The llama.cpp project has released version b10004, introducing full support for f16 as a source in SET_ROWS operations for both Vulkan and CPU backends. This update also includes additional backend tests and specific fixes for Intel hardware.

  • Adds full support for f16 SET_ROWS (equivalent to f32) to Vulkan and CPU backends.
  • Sets DenormPreserve 16 when supported to address failures on Intel processors.
  • Tunes error thresholds and updates Metal supports_op.
  • Provides binaries for macOS, Linux, Android, Windows, and openEuler across CPU, Vulkan, CUDA, ROCm, OpenVINO, SYCL, and HIP backends.

This release improves precision handling in specific backend operations and expands platform coverage with new binary distributions.