mirror of
https://github.com/ollama/ollama.git
synced 2026-07-23 09:10:53 -05:00
runner: Remove CGO engines, use llama-server exclusively for GGML models (#16031)
* broad lint fixes to sidestep CI scope glitch * runner: Remove CGO engines, use llama-server exclusively for GGML models Remove the vendored GGML and llama.cpp backend, CGO runner, Go model implementations, and sample. llama-server (built from upstream llama.cpp via FetchContent) is now the sole inference engine for GGUF-based models. (Safetensor based models continue to run on the new MLX engine.) This allows us to more rapidly pick up new capabilities and fixes from llama.cpp as they come out. On windows this now requires recent AMD driver versions to support ROCm v7 as llama.cpp currently does not support building against v6. * llama/compat: load Ollama-format GGUFs in llama-server Squashed from upstream/jmorganca/llama-compat on 2026-04-29. Source tip:0c33775d37. Original source commits: -25223160dllama/compat: add in-memory shim so llama-server can load Ollama-format GGUFs -7449b539allm,server: route Ollama-format gemma3 blobs through llama/compat -436f2e2b1llama/compat: make patch-apply idempotent -8c2c9d4c8llama/compat: extend gemma3 handler to cover 1B and 270M blobs -021389f7bllama/compat: shrink clip.cpp injection from 18 lines to 1 -61b367ec2llama/compat: shrink patch to pure call-site hooks (34 -> 20 lines) -36049361cllama/compat: simplify shim (gemma3-tested) -8fa664865llama/compat: add qwen35moe text handler -db0c74530llama/compat: add qwen35moe vision (clip) support -2a388da77llama/compat: split shared infra into a util TU -9a69a17dcllama/compat: document non-public API dependencies -d0f38a915llama/compat: add gpt-oss and lfm2 handlers -086071822llama/compat: add mistral3 text handler (vision TODO) -63bde9ff7llama/compat: add mistral3 vision (clip) support -3a57b89d5llama/compat: apply LLaMA RoPE permute to mistral3 vision Q/K -99cb87439llama/compat: add qwen35, gemma4, deepseek-ocr handlers -2c7850dballama/compat: add nemotron_h_moe handler (latent FFN + MTP skip) -9e3b54225llama/compat: add llama4 text + clip handlers -034fee349llama/compat: add gemma4 clip handler (gemma4v projector) -9945c5a93server: remove dhiltgen/* compat redirect table -5d4539101llama/compat: rewrite gemma4 tokenizer model to BPE -7e0765327llama/compat: add glm-ocr text handler + text-loader load-op hook -f1bd1a25allama/compat: add glm-ocr clip handler (glm4v projector) -4b5cf3420llama/compat: collapse text-loader hook back to one new patch line -eb4ecf4fcllama/compat: extend gemma4 clip handler to gemma4a (audio) -a23a5e76fllama/compat: fix gemma4a per-block norm tensor mapping -cd2dcaff4llama/compat: add embeddinggemma handler -1ce8a6b26llama/compat: add qwen3-vl + qwen2.5-vl handlers -fd98ffa1ellama/compat: add gemma3n + glm4moelite handlers -cc7bdf0bcllama/compat: handle null buft in maybe_load_tensor -0c33775d3llama/compat: disable mmap when load_op transforms text-side tensors * refine implementation * ci: fix windows MLX build * ci: fix windows llama-server build * ci: fix windows rocm build * ci: windows mlx tuning Shorten long-tail on build, and get OllamaSetup.exe back under 2g limit * ci: fix windows dependencies * win: fix dependency gathering * disable openmp * win: arm64 cross-compile build also DRY out CI steps * scheduler improvements * ci: improvements from #15982 * win: favor ninja for faster developer builds * win: fix build * win: fix arm64 cross-compile * win: avoid spaces in compiler path * misc discovery fixes, and bos handling * lint fixes * win: fix arm cross-compile build/CI bugs * llama.cpp update * win: handle multiple CRT dirs * vulkan: add windows iGPU detection * fix creation bugs for patched models, other refactoring work * tune batch size for better performance * ci and lint fixes * fix repeat_last_n bug * build: revamp build for better developer UX * amd, sampler, qwen3next fixes * version bump * fix mlx build * revamp GPU discovery Scanning the output of llama-server is turning out to be too error prone across llama.cpp updates, so this switches to a thin dynamic library load against the bundled GGML libraries so more details can be gathered from the API. * version bump * missing file * ci: fix cache miss on rocm build * refine vulkan dep handling * fix ps reporting bug on full GPU load * improve cmake wiring for customized local builds * version bump * docker build arg cleanup * improve windows exit error logs * fix community gemma4 support and ci flakes * fix mlx unit test * tighten up ps logic to avoid double counting fit log lines * version bump * fix ps view for full gpu layer offload * add MTP wiring for llama-server and create with GGUFs * pick best template by capabilities * version bump * ci: harden apt repos * remove unused cpu core discovery * adjust batch default logic to reduce OOMs * support larger tool calls * fix audio support, template show * qwen35 mtp patch support * flesh out dtypes * rocm deps * version bump * lint fix * block broken gfx1150 on windows * fix qwen3.5 moe mtp tensors in patch * mmproj oom fallback and vulkan on by default * qwen MTP compat fix * version bump * ci: fix WoA cross-compile * ci: workaround ui tool in cross-compile * version bump * win: enable OpenMP for CPU builds * build: improve developer UX * ci: windows path workaround for CPU build * win: fix WoA dependencies * win: fix large offset reads for mmproj patched loads * version bump * fix vulkan dup detection * add OLLAMA_IGPU_ENABLE and largely disable iGPUs by default * opt-in MTP, win large offset, integraton fixes * fix unit test scheduler interaction hang * fix multi-gpu filtering * version bump * review comments * fix thinking level * fix linux rocm ordering and granite 3.3 template * version bump * ci fix - non-shallow MLX checkout * bypass linux sysfs unit test on windows --------- Co-authored-by: jmorganca <jmorganca@gmail.com>
This commit is contained in:
Daniel Hiltgen
353
CMakeLists.txt
353
CMakeLists.txt
@@ -1,4 +1,4 @@
|
||||
cmake_minimum_required(VERSION 3.21)
|
||||
cmake_minimum_required(VERSION 3.24)
|
||||
|
||||
project(Ollama C CXX)
|
||||
|
||||
@@ -23,30 +23,23 @@ include(GNUInstallDirs)
|
||||
|
||||
find_package(Threads REQUIRED)
|
||||
|
||||
set(CMAKE_BUILD_TYPE Release)
|
||||
set(BUILD_SHARED_LIBS ON)
|
||||
if(NOT CMAKE_CONFIGURATION_TYPES AND NOT CMAKE_BUILD_TYPE)
|
||||
set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
|
||||
endif()
|
||||
|
||||
# These defaults can be overridden by presets (e.g., for static macOS llama-server builds)
|
||||
if(NOT DEFINED BUILD_SHARED_LIBS)
|
||||
set(BUILD_SHARED_LIBS ON)
|
||||
endif()
|
||||
|
||||
set(CMAKE_CXX_STANDARD 17)
|
||||
set(CMAKE_CXX_STANDARD_REQUIRED ON)
|
||||
set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
|
||||
set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX require gnu++17 extensions to compile properly
|
||||
|
||||
set(GGML_BUILD ON)
|
||||
set(GGML_SHARED ON)
|
||||
set(GGML_CCACHE ON)
|
||||
set(GGML_BACKEND_DL ON)
|
||||
set(GGML_BACKEND_SHARED ON)
|
||||
set(GGML_SCHED_MAX_COPIES 4)
|
||||
|
||||
set(GGML_LLAMAFILE ON)
|
||||
set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
|
||||
set(GGML_CUDA_GRAPHS ON)
|
||||
set(GGML_CUDA_FA ON)
|
||||
set(GGML_CUDA_COMPRESSION_MODE default)
|
||||
|
||||
if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
|
||||
OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
|
||||
set(GGML_CPU_ALL_VARIANTS ON)
|
||||
endif()
|
||||
# GGML backend for inference is provided by llama-server (built separately via
|
||||
# llama/server/CMakeLists.txt using FetchContent from the pinned llama.cpp source).
|
||||
# The root CMake project is the orchestration entrypoint; backend-specific
|
||||
# build rules live in subprojects under cmake/.
|
||||
|
||||
if(APPLE)
|
||||
set(CMAKE_BUILD_RPATH "@loader_path")
|
||||
@@ -55,7 +48,8 @@ if(APPLE)
|
||||
endif()
|
||||
|
||||
set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
|
||||
set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama/${OLLAMA_RUNNER_DIR})
|
||||
set(OLLAMA_LIB_DIR "lib/ollama" CACHE STRING "Install destination for Ollama runtime payloads")
|
||||
set(OLLAMA_INSTALL_DIR ${OLLAMA_LIB_DIR}/${OLLAMA_RUNNER_DIR})
|
||||
|
||||
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${OLLAMA_BUILD_DIR})
|
||||
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG ${OLLAMA_BUILD_DIR})
|
||||
@@ -64,314 +58,9 @@ 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})
|
||||
|
||||
# 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
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cpu/amx
|
||||
)
|
||||
|
||||
add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
|
||||
|
||||
# Define GGML version variables for shared library SOVERSION
|
||||
# These are required by ggml/src/CMakeLists.txt for proper library versioning
|
||||
set(GGML_VERSION_MAJOR 0)
|
||||
set(GGML_VERSION_MINOR 0)
|
||||
set(GGML_VERSION_PATCH 0)
|
||||
set(GGML_VERSION "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
|
||||
|
||||
set(GGML_CPU ON)
|
||||
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
|
||||
set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)
|
||||
|
||||
get_target_property(CPU_VARIANTS ggml-cpu MANUALLY_ADDED_DEPENDENCIES)
|
||||
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 ".*"
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
|
||||
FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CPU
|
||||
)
|
||||
|
||||
check_language(CUDA)
|
||||
if(CMAKE_CUDA_COMPILER)
|
||||
if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24" AND NOT CMAKE_CUDA_ARCHITECTURES)
|
||||
set(CMAKE_CUDA_ARCHITECTURES "native")
|
||||
endif()
|
||||
|
||||
find_package(CUDAToolkit)
|
||||
add_subdirectory(${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}
|
||||
PRE_INCLUDE_REGEXES cublas cublasLt cudart
|
||||
PRE_EXCLUDE_REGEXES ".*"
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT CUDA
|
||||
)
|
||||
endif()
|
||||
|
||||
set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(908|90a|1200|1201):xnack[+-]$"
|
||||
CACHE STRING
|
||||
"Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(908|90a|1200|1201):xnack[+-]$\"."
|
||||
)
|
||||
|
||||
check_language(HIP)
|
||||
if(CMAKE_HIP_COMPILER)
|
||||
set(HIP_PLATFORM "amd")
|
||||
|
||||
if(NOT AMDGPU_TARGETS)
|
||||
find_package(hip REQUIRED)
|
||||
list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(94[012]|101[02]|1030|110[012]|120[01])$")
|
||||
endif()
|
||||
|
||||
if(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
|
||||
list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
|
||||
endif()
|
||||
|
||||
if(AMDGPU_TARGETS)
|
||||
find_package(hip REQUIRED)
|
||||
add_subdirectory(${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)
|
||||
endif()
|
||||
|
||||
target_compile_definitions(ggml-hip PRIVATE GGML_HIP_NO_VMM)
|
||||
|
||||
install(TARGETS ggml-hip
|
||||
RUNTIME_DEPENDENCY_SET rocm
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
|
||||
)
|
||||
install(RUNTIME_DEPENDENCY_SET rocm
|
||||
DIRECTORIES ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR}
|
||||
PRE_INCLUDE_REGEXES hipblas rocblas amdhip64 rocsolver amd_comgr hsa-runtime64 rocsparse tinfo rocprofiler-register roctx64 rocroller drm drm_amdgpu numa elf
|
||||
PRE_EXCLUDE_REGEXES ".*"
|
||||
POST_EXCLUDE_REGEXES "system32"
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP
|
||||
)
|
||||
|
||||
foreach(HIP_LIB_BIN_INSTALL_DIR IN ITEMS ${HIP_BIN_INSTALL_DIR} ${HIP_LIB_INSTALL_DIR})
|
||||
if(EXISTS ${HIP_LIB_BIN_INSTALL_DIR}/rocblas)
|
||||
install(DIRECTORY ${HIP_LIB_BIN_INSTALL_DIR}/rocblas DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT HIP)
|
||||
break()
|
||||
endif()
|
||||
endforeach()
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if(NOT APPLE)
|
||||
find_package(Vulkan)
|
||||
if(Vulkan_FOUND)
|
||||
add_subdirectory(${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
|
||||
PRE_EXCLUDE_REGEXES ".*"
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
|
||||
)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
option(MLX_ENGINE "Enable MLX backend" OFF)
|
||||
if(MLX_ENGINE)
|
||||
message(STATUS "Setting up MLX (this takes a while...)")
|
||||
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/imagegen/mlx)
|
||||
|
||||
# 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()
|
||||
|
||||
# Split mlx/mlxc libraries from runtime deps to avoid stripping deps
|
||||
install(TARGETS mlx mlxc
|
||||
RUNTIME_DEPENDENCY_SET mlx_runtime_deps
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
|
||||
FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
|
||||
)
|
||||
install(RUNTIME_DEPENDENCY_SET mlx_runtime_deps
|
||||
DIRECTORIES ${MLX_RUNTIME_DIRS}
|
||||
PRE_INCLUDE_REGEXES ${MLX_INCLUDE_REGEXES}
|
||||
PRE_EXCLUDE_REGEXES ".*"
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX_VENDOR
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX_VENDOR
|
||||
)
|
||||
|
||||
if(TARGET jaccl)
|
||||
install(TARGETS jaccl
|
||||
RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
|
||||
LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
|
||||
FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
|
||||
)
|
||||
endif()
|
||||
|
||||
# Install the Metal library for macOS arm64 (must be colocated with the binary)
|
||||
# Metal backend is only built for arm64, not x86_64
|
||||
if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
|
||||
install(FILES ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/backend/metal/kernels/mlx.metallib
|
||||
DESTINATION ${OLLAMA_INSTALL_DIR}
|
||||
COMPONENT MLX)
|
||||
endif()
|
||||
|
||||
# Install 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.
|
||||
#
|
||||
# Layout:
|
||||
# ${OLLAMA_INSTALL_DIR}/include/cccl/{cuda,nv}/ — CCCL headers
|
||||
# ${OLLAMA_INSTALL_DIR}/include/*.h — CUDA toolkit headers
|
||||
#
|
||||
# MLX's jit_module.cpp resolves CCCL via
|
||||
# current_binary_dir()[.parent_path()] / "include" / "cccl"
|
||||
# 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.
|
||||
# CUDA runtime headers are found via CUDA_PATH env var (set by mlxrunner).
|
||||
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()
|
||||
|
||||
# Install minimal CUDA toolkit headers needed by MLX JIT kernels.
|
||||
# These are the transitive closure of includes from mlx/backend/cuda/device/*.cuh.
|
||||
# The Go mlxrunner sets CUDA_PATH to OLLAMA_INSTALL_DIR so MLX finds them at
|
||||
# $CUDA_PATH/include/*.h via NVRTC --include-path.
|
||||
if(CUDAToolkit_FOUND)
|
||||
# CUDAToolkit_INCLUDE_DIRS may be a semicolon-separated list
|
||||
# (e.g. ".../include;.../include/cccl"). Find the entry that
|
||||
# contains the CUDA runtime headers we need.
|
||||
set(_cuda_inc "")
|
||||
foreach(_dir ${CUDAToolkit_INCLUDE_DIRS})
|
||||
if(EXISTS "${_dir}/cuda_runtime_api.h")
|
||||
set(_cuda_inc "${_dir}")
|
||||
break()
|
||||
endif()
|
||||
endforeach()
|
||||
if(NOT _cuda_inc)
|
||||
message(WARNING "Could not find cuda_runtime_api.h in CUDAToolkit_INCLUDE_DIRS: ${CUDAToolkit_INCLUDE_DIRS}")
|
||||
else()
|
||||
set(_dst "${OLLAMA_INSTALL_DIR}/include")
|
||||
set(_MLX_JIT_CUDA_HEADERS
|
||||
builtin_types.h
|
||||
cooperative_groups.h
|
||||
cuda_bf16.h
|
||||
cuda_bf16.hpp
|
||||
cuda_device_runtime_api.h
|
||||
cuda_fp16.h
|
||||
cuda_fp16.hpp
|
||||
cuda_fp8.h
|
||||
cuda_fp8.hpp
|
||||
cuda_runtime_api.h
|
||||
device_types.h
|
||||
driver_types.h
|
||||
math_constants.h
|
||||
surface_types.h
|
||||
texture_types.h
|
||||
vector_functions.h
|
||||
vector_functions.hpp
|
||||
vector_types.h
|
||||
)
|
||||
foreach(_hdr ${_MLX_JIT_CUDA_HEADERS})
|
||||
install(FILES "${_cuda_inc}/${_hdr}"
|
||||
DESTINATION ${_dst}
|
||||
COMPONENT MLX)
|
||||
endforeach()
|
||||
# Subdirectory headers
|
||||
install(DIRECTORY "${_cuda_inc}/cooperative_groups"
|
||||
DESTINATION ${_dst}
|
||||
COMPONENT MLX
|
||||
FILES_MATCHING PATTERN "*.h")
|
||||
install(FILES "${_cuda_inc}/crt/host_defines.h"
|
||||
DESTINATION "${_dst}/crt"
|
||||
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 ${OLLAMA_INSTALL_DIR}
|
||||
COMPONENT MLX)
|
||||
endif()
|
||||
|
||||
# Manually install CUDA runtime libraries that MLX loads via dlopen
|
||||
# (not detected by RUNTIME_DEPENDENCIES since they aren't link-time deps)
|
||||
if(CUDAToolkit_FOUND)
|
||||
file(GLOB MLX_CUDA_LIBS
|
||||
"${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
|
||||
"${CUDAToolkit_LIBRARY_DIR}/libcublas.so*"
|
||||
"${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 ${OLLAMA_INSTALL_DIR}
|
||||
COMPONENT MLX_VENDOR)
|
||||
endif()
|
||||
endif()
|
||||
if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/llama/server/CMakeLists.txt")
|
||||
set(OLLAMA_HAVE_LLAMA_SERVER TRUE)
|
||||
else()
|
||||
set(OLLAMA_HAVE_LLAMA_SERVER FALSE)
|
||||
endif()
|
||||
include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/local.cmake)
|
||||
|
||||
Reference in New Issue
Block a user