upstream-archive/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-23 09:10:53 -05:00
Code Issues Packages Projects Releases Wiki Activity

MLX: add header vendoring and remove go build tag (#14642)

Browse Source 
* prefer rocm v6 on windows

Avoid building with v7 - more changes are needed

* MLX: add header vendoring and remove go build tag

This switches to using a vendoring approach for the mlx-c headers so that Go
can build without requiring a cmake first.  This enables building the new MLX
based code by default.  Every time cmake runs, the headers are refreshed, so we
can easily keep them in sync when we bump mlx versions.  Basic Windows
and Linux support are verified.

* ci: harden for flaky choco repo servers

CI sometimes fails due to choco not actually installing cache.  Since it just speeds up the build, we can proceed without.

* review comments
This commit is contained in:
Daniel Hiltgen
2026-03-09 17:24:45 -07:00
committed by GitHub
parent 3e06bde643
commit 10e51c5177
142 changed files with 5350 additions and 1064 deletions
Download Patch File Download Diff File Expand all files Collapse all files

106
CMakeLists.txt
View File

@@ -64,10 +64,15 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${OLLAMA_BUILD_DIR})
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_DEBUG ${OLLAMA_BUILD_DIR})
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_RELEASE ${OLLAMA_BUILD_DIR})
include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include)
include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu)
include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx)
# Store ggml include paths for use with target_include_directories later.
# We avoid global include_directories() to prevent polluting the include path
# for other projects like MLX (whose openblas dependency has its own common.h).
set(GGML_INCLUDE_DIRS
${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src
${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/include
${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu
${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx
)
add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
@@ -87,6 +92,14 @@ if(NOT CPU_VARIANTS)
set(CPU_VARIANTS "ggml-cpu")
endif()
# Apply ggml include directories to ggml targets only (not globally)
target_include_directories(ggml-base PRIVATE ${GGML_INCLUDE_DIRS})
foreach(variant ${CPU_VARIANTS})
if(TARGET ${variant})
target_include_directories(${variant} PRIVATE ${GGML_INCLUDE_DIRS})
endif()
endforeach()
install(TARGETS ggml-base ${CPU_VARIANTS}
RUNTIME_DEPENDENCIES
PRE_EXCLUDE_REGEXES ".*"
@@ -103,6 +116,7 @@ if(CMAKE_CUDA_COMPILER)
find_package(CUDAToolkit)
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cuda)
target_include_directories(ggml-cuda PRIVATE ${GGML_INCLUDE_DIRS})
install(TARGETS ggml-cuda
RUNTIME_DEPENDENCIES
DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
@@ -134,6 +148,7 @@ if(CMAKE_HIP_COMPILER)
if(AMDGPU_TARGETS)
find_package(hip REQUIRED)
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
target_include_directories(ggml-hip PRIVATE ${GGML_INCLUDE_DIRS})
if (WIN32)
target_compile_definitions(ggml-hip PRIVATE GGML_CUDA_NO_PEER_COPY)
@@ -168,6 +183,7 @@ if(NOT APPLE)
find_package(Vulkan)
if(Vulkan_FOUND)
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
target_include_directories(ggml-vulkan PRIVATE ${GGML_INCLUDE_DIRS})
install(TARGETS ggml-vulkan
RUNTIME_DEPENDENCIES
PRE_INCLUDE_REGEXES vulkan
@@ -179,7 +195,6 @@ if(NOT APPLE)
endif()
option(MLX_ENGINE "Enable MLX backend" OFF)
if(MLX_ENGINE)
message(STATUS "Setting up MLX (this takes a while...)")
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/imagegen/mlx)
@@ -187,10 +202,36 @@ if(MLX_ENGINE)
# Find CUDA toolkit if MLX is built with CUDA support
find_package(CUDAToolkit)
# Build list of directories for runtime dependency resolution
set(MLX_RUNTIME_DIRS ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR})
# Add cuDNN bin paths for DLLs (Windows MLX CUDA builds)
# CUDNN_ROOT_DIR is the standard CMake variable for cuDNN location
if(DEFINED ENV{CUDNN_ROOT_DIR})
# cuDNN 9.x has versioned subdirectories under bin/ (e.g., bin/13.0/)
file(GLOB CUDNN_BIN_SUBDIRS "$ENV{CUDNN_ROOT_DIR}/bin/*")
list(APPEND MLX_RUNTIME_DIRS ${CUDNN_BIN_SUBDIRS})
endif()
# Add build output directory and MLX dependency build directories
list(APPEND MLX_RUNTIME_DIRS ${OLLAMA_BUILD_DIR})
# OpenBLAS DLL location (pre-built zip extracts into openblas-src/bin/)
list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/openblas-src/bin)
# NCCL: on Linux, if real NCCL is found, cmake bundles libnccl.so via the
# regex below. If NCCL is not found, MLX links a static stub (OBJECT lib)
# so there is no runtime dependency. This path covers the stub build dir
# for windows so we include the DLL in our dependencies.
list(APPEND MLX_RUNTIME_DIRS ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/distributed/nccl/nccl_stub-prefix/src/nccl_stub-build/Release)
# Base regexes for runtime dependencies (cross-platform)
set(MLX_INCLUDE_REGEXES cublas cublasLt cudart cufft nvrtc nvrtc-builtins cudnn nccl openblas gfortran)
# On Windows, also include dl.dll (dlfcn-win32 POSIX emulation layer)
if(WIN32)
list(APPEND MLX_INCLUDE_REGEXES "^dl\\.dll$")
endif()
install(TARGETS mlx mlxc
RUNTIME_DEPENDENCIES
DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc nvrtc-builtins cudnn nccl openblas gfortran
DIRECTORIES ${MLX_RUNTIME_DIRS}
PRE_INCLUDE_REGEXES ${MLX_INCLUDE_REGEXES}
PRE_EXCLUDE_REGEXES ".*"
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
@@ -205,13 +246,54 @@ if(MLX_ENGINE)
COMPONENT MLX)
endif()
# Manually install cudart and cublas since they might not be picked up as direct dependencies
# Install CCCL headers for NVRTC JIT compilation at runtime.
# MLX's own install rules use the default component so they get skipped by
# --component MLX. Headers are installed alongside libmlx in OLLAMA_INSTALL_DIR.
# On Linux, MLX's jit_module.cpp resolves CCCL via
# current_binary_dir().parent_path() / "include" / "cccl", so we create a
# symlink from lib/ollama/include -> ${OLLAMA_RUNNER_DIR}/include
# This will need refinement if we add multiple CUDA versions for MLX in the future.
if(EXISTS ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda)
install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/cuda
DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
COMPONENT MLX)
install(DIRECTORY ${CMAKE_BINARY_DIR}/_deps/cccl-src/include/nv
DESTINATION ${OLLAMA_INSTALL_DIR}/include/cccl
COMPONENT MLX)
if(NOT WIN32 AND NOT APPLE)
install(CODE "
set(_link \"${CMAKE_INSTALL_PREFIX}/lib/ollama/include\")
set(_target \"${OLLAMA_RUNNER_DIR}/include\")
if(NOT EXISTS \${_link})
execute_process(COMMAND \${CMAKE_COMMAND} -E create_symlink \${_target} \${_link})
endif()
" COMPONENT MLX)
endif()
endif()
# On Windows, explicitly install dl.dll (dlfcn-win32 POSIX dlopen emulation)
# RUNTIME_DEPENDENCIES auto-excludes it via POST_EXCLUDE_FILES_STRICT because
# dlfcn-win32 is a known CMake target with its own install rules (which install
# to the wrong destination). We must install it explicitly here.
if(WIN32)
install(FILES ${OLLAMA_BUILD_DIR}/dl.dll
DESTINATION ${OLLAMA_INSTALL_DIR}
COMPONENT MLX)
endif()
# Manually install CUDA runtime libraries that MLX loads via dlopen
# (not detected by RUNTIME_DEPENDENCIES since they aren't link-time deps)
if(CUDAToolkit_FOUND)
file(GLOB CUDART_LIBS
file(GLOB MLX_CUDA_LIBS
"${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
"${CUDAToolkit_LIBRARY_DIR}/libcublas.so*")
if(CUDART_LIBS)
install(FILES ${CUDART_LIBS}
"${CUDAToolkit_LIBRARY_DIR}/libcublas.so*"
"${CUDAToolkit_LIBRARY_DIR}/libcublasLt.so*"
"${CUDAToolkit_LIBRARY_DIR}/libnvrtc.so*"
"${CUDAToolkit_LIBRARY_DIR}/libnvrtc-builtins.so*"
"${CUDAToolkit_LIBRARY_DIR}/libcufft.so*"
"${CUDAToolkit_LIBRARY_DIR}/libcudnn.so*")
if(MLX_CUDA_LIBS)
install(FILES ${MLX_CUDA_LIBS}
DESTINATION ${OLLAMA_INSTALL_DIR}
COMPONENT MLX)
endif()