|  | 
|  | 1 | +add_custom_target(deep_gemm) | 
|  | 2 | + | 
|  | 3 | +if(WIN32) | 
|  | 4 | +  return() | 
|  | 5 | +endif() | 
|  | 6 | + | 
|  | 7 | +# Prepare files | 
|  | 8 | +# ============= | 
|  | 9 | + | 
|  | 10 | +# Use DeepGEMM submodule | 
|  | 11 | +set(DEEP_GEMM_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../../../3rdparty/DeepGEMM) | 
|  | 12 | +get_filename_component(DEEP_GEMM_SOURCE_DIR ${DEEP_GEMM_SOURCE_DIR} ABSOLUTE) | 
|  | 13 | + | 
|  | 14 | +if(NOT EXISTS ${DEEP_GEMM_SOURCE_DIR}) | 
|  | 15 | +  message( | 
|  | 16 | +    FATAL_ERROR | 
|  | 17 | +      "DeepGEMM submodule not found at ${DEEP_GEMM_SOURCE_DIR}. Please run: git submodule update --init --recursive" | 
|  | 18 | +  ) | 
|  | 19 | +endif() | 
|  | 20 | + | 
|  | 21 | +# Check if submodules are initialized | 
|  | 22 | +if(NOT EXISTS ${DEEP_GEMM_SOURCE_DIR}/third-party/cutlass/include) | 
|  | 23 | +  message( | 
|  | 24 | +    FATAL_ERROR | 
|  | 25 | +      "DeepGEMM submodules not initialized. Please run: git submodule update --init --recursive" | 
|  | 26 | +  ) | 
|  | 27 | +endif() | 
|  | 28 | + | 
|  | 29 | +# Copy and update python files | 
|  | 30 | +set(DEEP_GEMM_PYTHON_DEST ${CMAKE_CURRENT_BINARY_DIR}/python/deep_gemm) | 
|  | 31 | +file(REMOVE_RECURSE ${DEEP_GEMM_PYTHON_DEST}) | 
|  | 32 | +file(MAKE_DIRECTORY ${DEEP_GEMM_PYTHON_DEST}) | 
|  | 33 | + | 
|  | 34 | +# Copy all files from deep_gemm directory | 
|  | 35 | +file(GLOB_RECURSE DEEP_GEMM_ALL_FILES ${DEEP_GEMM_SOURCE_DIR}/deep_gemm/*) | 
|  | 36 | +configure_file(${DEEP_GEMM_SOURCE_DIR}/LICENSE ${DEEP_GEMM_PYTHON_DEST}/LICENSE | 
|  | 37 | +               COPYONLY) | 
|  | 38 | +foreach(SOURCE_FILE ${DEEP_GEMM_ALL_FILES}) | 
|  | 39 | +  file(RELATIVE_PATH REL_PATH ${DEEP_GEMM_SOURCE_DIR}/deep_gemm ${SOURCE_FILE}) | 
|  | 40 | +  get_filename_component(REL_DIR ${REL_PATH} DIRECTORY) | 
|  | 41 | +  file(MAKE_DIRECTORY ${DEEP_GEMM_PYTHON_DEST}/${REL_DIR}) | 
|  | 42 | + | 
|  | 43 | +  # Check if it's a Python file that needs import renaming | 
|  | 44 | +  get_filename_component(FILE_EXT ${SOURCE_FILE} EXT) | 
|  | 45 | +  if(FILE_EXT STREQUAL ".py") | 
|  | 46 | +    # Read file content and replace module imports for Python files | 
|  | 47 | +    file(READ ${SOURCE_FILE} _content) | 
|  | 48 | +    string(REPLACE "deep_gemm_cpp" "tensorrt_llm.deep_gemm_cpp_tllm" _content | 
|  | 49 | +                   "${_content}") | 
|  | 50 | + | 
|  | 51 | +    # Add adaptation header | 
|  | 52 | +    string( | 
|  | 53 | +      PREPEND | 
|  | 54 | +      _content | 
|  | 55 | +      "# Adapted from https://github.com/deepseek-ai/DeepGEMM/blob/main/deep_gemm/${REL_PATH}\n" | 
|  | 56 | +    ) | 
|  | 57 | + | 
|  | 58 | +    # Write modified content | 
|  | 59 | +    set(_dst "${DEEP_GEMM_PYTHON_DEST}/${REL_PATH}") | 
|  | 60 | +    file(WRITE ${_dst} "${_content}") | 
|  | 61 | +  else() | 
|  | 62 | +    # Copy non-Python files as-is | 
|  | 63 | +    set(_dst "${DEEP_GEMM_PYTHON_DEST}/${REL_PATH}") | 
|  | 64 | +    file(COPY ${SOURCE_FILE} DESTINATION ${DEEP_GEMM_PYTHON_DEST}/${REL_DIR}) | 
|  | 65 | +  endif() | 
|  | 66 | + | 
|  | 67 | +  # Add dependency tracking | 
|  | 68 | +  set_property( | 
|  | 69 | +    DIRECTORY | 
|  | 70 | +    APPEND | 
|  | 71 | +    PROPERTY CMAKE_CONFIGURE_DEPENDS ${SOURCE_FILE}) | 
|  | 72 | +endforeach() | 
|  | 73 | + | 
|  | 74 | +# Copy third-party includes (cutlass and fmt) to the include directory | 
|  | 75 | +set(DEEP_GEMM_INCLUDE_DEST ${DEEP_GEMM_PYTHON_DEST}/include) | 
|  | 76 | +file(MAKE_DIRECTORY ${DEEP_GEMM_INCLUDE_DEST}) | 
|  | 77 | +file(COPY ${DEEP_GEMM_SOURCE_DIR}/third-party/cutlass/include/cute | 
|  | 78 | +     DESTINATION ${DEEP_GEMM_INCLUDE_DEST}) | 
|  | 79 | +file(COPY ${DEEP_GEMM_SOURCE_DIR}/third-party/cutlass/include/cutlass | 
|  | 80 | +     DESTINATION ${DEEP_GEMM_INCLUDE_DEST}) | 
|  | 81 | + | 
|  | 82 | +# Find torch_python | 
|  | 83 | +find_library(TORCH_PYTHON_LIB torch_python REQUIRED | 
|  | 84 | +             HINTS ${TORCH_INSTALL_PREFIX}/lib) | 
|  | 85 | + | 
|  | 86 | +# Build deep_gemm_cpp_tllm extension (matching deep_gemm's setup.py) | 
|  | 87 | +set(DEEP_GEMM_SOURCES ${DEEP_GEMM_SOURCE_DIR}/csrc/python_api.cpp) | 
|  | 88 | + | 
|  | 89 | +pybind11_add_module(deep_gemm_cpp_tllm ${DEEP_GEMM_SOURCES}) | 
|  | 90 | +set_target_properties( | 
|  | 91 | +  deep_gemm_cpp_tllm | 
|  | 92 | +  PROPERTIES CXX_STANDARD_REQUIRED ON | 
|  | 93 | +             CXX_STANDARD 17 | 
|  | 94 | +             CXX_SCAN_FOR_MODULES OFF | 
|  | 95 | +             CUDA_SEPARABLE_COMPILATION ON | 
|  | 96 | +             LINK_DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/deep_gemm_cpp_tllm.version | 
|  | 97 | +             INSTALL_RPATH "${TORCH_INSTALL_PREFIX}/lib" | 
|  | 98 | +             BUILD_WITH_INSTALL_RPATH TRUE) | 
|  | 99 | + | 
|  | 100 | +target_compile_options(deep_gemm_cpp_tllm PRIVATE ${TORCH_CXX_FLAGS} -std=c++17 | 
|  | 101 | +                                                  -O3 -fPIC -Wno-psabi) | 
|  | 102 | + | 
|  | 103 | +# Extension name definition | 
|  | 104 | +target_compile_definitions(deep_gemm_cpp_tllm | 
|  | 105 | +                           PRIVATE TORCH_EXTENSION_NAME=deep_gemm_cpp_tllm) | 
|  | 106 | + | 
|  | 107 | +# Include directories matching deep_gemm setup.py | 
|  | 108 | +target_include_directories( | 
|  | 109 | +  deep_gemm_cpp_tllm | 
|  | 110 | +  PRIVATE ${CUDA_INCLUDE_DIRS} ${DEEP_GEMM_SOURCE_DIR}/deep_gemm/include | 
|  | 111 | +          ${DEEP_GEMM_SOURCE_DIR}/third-party/cutlass/include | 
|  | 112 | +          ${DEEP_GEMM_SOURCE_DIR}/third-party/fmt/include) | 
|  | 113 | + | 
|  | 114 | +# Link libraries (matching deep_gemm setup.py: cuda, cudart + torch) | 
|  | 115 | +target_link_libraries( | 
|  | 116 | +  deep_gemm_cpp_tllm PRIVATE ${TORCH_LIBRARIES} ${TORCH_PYTHON_LIB} | 
|  | 117 | +                             CUDA::cuda_driver CUDA::cudart) | 
|  | 118 | + | 
|  | 119 | +# Link directories | 
|  | 120 | +target_link_directories( | 
|  | 121 | +  deep_gemm_cpp_tllm PRIVATE ${CUDA_TOOLKIT_ROOT_DIR}/lib64 | 
|  | 122 | +  ${CUDA_TOOLKIT_ROOT_DIR}/lib64/stubs) | 
|  | 123 | + | 
|  | 124 | +# Set targets | 
|  | 125 | +# =========== | 
|  | 126 | +add_dependencies(deep_gemm deep_gemm_cpp_tllm) | 
0 commit comments