The llama.cpp project released build b9894, which includes a fix for the Vulkan backend to prevent failures in GGML_OP_SET_ROWS caused by unimplemented f16 support.

  • The Vulkan implementation now checks the src0 type in GGML_OP_SET_ROWS to avoid crashes due to missing f16 support.
  • macOS builds are available for Apple Silicon (arm64) and Intel (x64), with KleidiAI disabled on Apple Silicon.
  • Linux binaries cover Ubuntu x64, arm64, and s390x across CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL backends.
  • Android support includes arm64 CPU builds.
  • Windows binaries provide CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP options.
  • openEuler x86 and aarch64 builds with ACL Graph are available, while openEuler x86 (310p) is disabled.

This release ensures stability for users relying on Vulkan inference by addressing the specific type-checking issue.