The llama.cpp project has released version b10000, which includes a key update to the ggml backend. This change uniformizes the im2col destination type and casting logic across all convolution operations, allowing functions like im2col_f16 to accept any kernel type.

  • Uniformized im2col dst_type for all conv ops in ggml
  • Updated ggml/src/ggml.c to standardize casting logic
  • Fixed im2col_f16 to accept any kernel type
  • macOS/iOS binaries available for Apple Silicon, Intel, and iOS XCFramework
  • Linux builds provided for Ubuntu x64, arm64, s390x, Vulkan, ROCm 7.2, OpenVINO, and SYCL
  • Android arm64 (CPU) binaries included
  • Windows binaries support CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP
  • openEuler builds for x86 and aarch64 with ACL Graph support are available
  • UI binary package is also released