Can't disable CMAKE ARG on Apple: GGML_METAL=OFF

Issue はメインの CMakeLists.txt ファイルの 95 行目を指しています。その行と周辺のロジックを調べ、Apple プラットフォームで GGML_METAL がどのように強制的に有効化されているかを理解してください。目的は、デフォルトの動作を維持しながら、ユーザーが指定した GGML_METAL=OFF フラグが尊重されるように CMake のロジックを変更することです。指定された CMAKE_ARGS を使ってビルドし、Vulkan が動作し、Metal が無効になることを確認して変更をテストしてください。

Expected Behavior

I can disable GGML_METAL and enable GGML_VULKAN.

Current Behavior

Due to this line in the CMAKE config I cannot set GGML_METAL=OFF as it is forced.

Environment and Context

I'm trying to build GGML_VULKAN=ON with Metal off as I want to use MoltenVK on my Intel MBP with an AMD GPU. I have already successfully don't this directly with llama.cpp.

I am having to resort to cloning the repo and patching the lines by hand, which is not ideal.

• Physical (or virtual) hardware you are using, e.g. for Linux:

2019 Intel MBP

• Operating System, e.g. for Linux:

MacOS 15.5

• SDK version, e.g. for Linux:

Python 3.11.9

GNU Make 3.81

Apple clang version 16.0.0 (clang-1600.0.26.4)
Target: x86_64-apple-darwin24.5.0
Thread model: posix

Steps to Reproduce

CMAKE_ARGS="-DGGML_METAL=OFF -DGGML_VULKAN=ON \
-DVulkan_INCLUDE_DIR=/Users/brenden/VulkanSDK/1.4.309.0/macOS/include \
-DVulkan_LIBRARY=/Users/brenden/VulkanSDK/1.4.309.0/macOS/lib/libMoltenVK.dylib \
-DOpenMP_ROOT=$(brew --prefix)/opt/libomp \
-DVulkan_GLSLC_EXECUTABLE=$(brew --prefix)/opt/shaderc/bin/glslc \
-DVulkan_GLSLANG_VALIDATOR_EXECUTABLE=$(brew --prefix)/opt/glslang/bin/glslangValidator \
-DOpenMP_C_FLAGS=-fopenmp=lomp \
-DOpenMP_CXX_FLAGS=-fopenmp=lomp \
-DOpenMP_C_LIB_NAMES=libomp \
-DOpenMP_CXX_LIB_NAMES=libomp \
-DOpenMP_libomp_LIBRARY=$(brew --prefix)/opt/libomp/lib/libomp.dylib \
-DOpenMP_CXX_FLAGS=\"-Xpreprocessor -fopenmp $(brew --prefix)/opt/libomp/lib/libomp.dylib -I$(brew --prefix)/opt/libomp/include\" \
-DOpenMP_CXX_LIB_NAMES=libomp \
-DOpenMP_C_FLAGS=\"-Xpreprocessor -fopenmp $(brew --prefix)/opt/libomp/lib/libomp.dylib -I$(brew --prefix)/opt/libomp/include\"" FORCE_CMAKE=1 pip install --no-binary=llama-cpp-python --no-cache-dir -v llama-cpp-python

(NOTE: These args are derived from what I used to compile a MoltenVK/Vulkan build of llama.cpp using a version of llama.cpp that is from June 2025.

GGML Metal backend runs when you run the server.