The llama.cpp project released version b9939, introducing support for CONV_2D_DW (depthwise 2D convolution) on the Metal backend. This update includes a new channel-tiled kernel for non-contiguous layouts and fixes F16 kernel support for CONV_2D_DW on CPU.
- Added CONV_2D_DW support to the Metal backend with 3D dispatch and channel-tiled kernels.
- Fixed F16 kernel support for CONV_2D_DW on CPU.
- Added performance test cases and F16 tests for CONV_2D_DW.
- Provided binaries for macOS, iOS, Linux, Android, Windows, and openEuler across various hardware backends including CUDA, ROCm, Vulkan, and SYCL.