Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

mlx: enable CUDA backend in CUDA builds (#17688)

Daniel Hiltgen641df5e5adbd
diff --git a/cmake/mlx/CMakeLists.txt b/cmake/mlx/CMakeLists.txt
index c9cd82eea..30fd941fa 100644
--- a/cmake/mlx/CMakeLists.txt
+++ b/cmake/mlx/CMakeLists.txt
@@ -50,6 +50,11 @@ endif()
option(OLLAMA_MLX_GENERATE_WRAPPERS "Regenerate MLX Go wrappers" OFF)
message(STATUS "Setting up MLX (this takes a while...)")
+foreach(_cudnn_var CUDNN_INCLUDE_PATH CUDNN_LIBRARY_PATH)
+ if((NOT DEFINED ${_cudnn_var} OR "${${_cudnn_var}}" STREQUAL "") AND DEFINED ENV{${_cudnn_var}})
+ set(${_cudnn_var} "$ENV{${_cudnn_var}}" CACHE PATH "")
+ endif()
+endforeach()
add_subdirectory(${OLLAMA_SOURCE_DIR}/x/mlxrunner/mlx ${CMAKE_BINARY_DIR}/x/mlxrunner/mlx)
# Find CUDA toolkit if MLX is built with CUDA support.
@@ -65,6 +70,7 @@ elseif(DEFINED ENV{CUDNN_ROOT_DIR})
set(_cudnn_root "$ENV{CUDNN_ROOT_DIR}")
endif()
if(_cudnn_root)
+ file(TO_CMAKE_PATH "${_cudnn_root}" _cudnn_root)
# cuDNN 9.x has versioned subdirectories under bin/ (e.g., bin/13.0/).
file(GLOB CUDNN_BIN_SUBDIRS "${_cudnn_root}/bin/*")
list(APPEND MLX_RUNTIME_DIRS ${CUDNN_BIN_SUBDIRS})
@@ -310,7 +316,25 @@ if(CUDAToolkit_FOUND)
"${CUDAToolkit_LIBRARY_DIR}/libnvrtc.so*"
"${CUDAToolkit_LIBRARY_DIR}/libnvrtc-builtins.so*"
"${CUDAToolkit_LIBRARY_DIR}/libcufft.so*"
- "${CUDAToolkit_LIBRARY_DIR}/libcudnn.so*")
+ "${CUDAToolkit_LIBRARY_DIR}/libcudnn*.so*")
+ if(WIN32)
+ file(GLOB MLX_CUDA_DLLS
+ "${CUDAToolkit_BIN_DIR}/nvrtc-builtins64_*.dll"
+ "${CUDAToolkit_BIN_DIR}/x64/nvrtc-builtins64_*.dll")
+ list(APPEND MLX_CUDA_LIBS ${MLX_CUDA_DLLS})
+ endif()
+ find_library(MLX_CUDNN_LIBRARY NAMES cudnn HINTS "$ENV{CUDNN_LIBRARY_PATH}")
+ if(MLX_CUDNN_LIBRARY)
+ get_filename_component(MLX_CUDNN_LIBRARY_DIR "${MLX_CUDNN_LIBRARY}" DIRECTORY)
+ file(GLOB MLX_CUDNN_LIBS "${MLX_CUDNN_LIBRARY_DIR}/libcudnn*.so*")
+ list(APPEND MLX_CUDA_LIBS ${MLX_CUDNN_LIBS})
+ endif()
+ if(WIN32 AND _cudnn_root)
+ file(GLOB MLX_CUDNN_DLLS
+ "${_cudnn_root}/bin/${CUDAToolkit_VERSION_MAJOR}.0/cudnn*.dll"
+ "${_cudnn_root}/bin/x64/cudnn*.dll")
+ list(APPEND MLX_CUDA_LIBS ${MLX_CUDNN_DLLS})
+ endif()
if(MLX_CUDA_LIBS)
install(FILES ${MLX_CUDA_LIBS}
DESTINATION ${OLLAMA_INSTALL_DIR}
diff --git a/cmake/mlx/CMakePresets.json b/cmake/mlx/CMakePresets.json
index 981dc89e5..3f23d51fe 100644
--- a/cmake/mlx/CMakePresets.json
+++ b/cmake/mlx/CMakePresets.json
@@ -17,6 +17,8 @@
"inherits": [ "default" ],
"cacheVariables": {
"CMAKE_CUDA_FLAGS": "-t 2",
+ "MLX_BUILD_CUDA": "ON",
+ "MLX_BUILD_METAL": "OFF",
"OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
}
},