aphrodite-engine/CMakeLists.txt

cmake_minimum_required(VERSION 3.21)

project(aphrodite_extensions LANGUAGES CXX)

option(APHRODITE_TARGET_DEVICE "Target device backend for Aphrodite" "cuda")

message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
message(STATUS "Target device: ${APHRODITE_TARGET_DEVICE}")

include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)

#
# Supported python versions.  These versions will be searched in order, the
# first match will be selected.  These should be kept in sync with setup.py.
#
set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11")

# Supported NVIDIA architectures.
set(CUDA_SUPPORTED_ARCHS "6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0")

# Supported AMD GPU architectures.
set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100")

#
# Supported/expected torch versions for CUDA/ROCm.
#
# Currently, having an incorrect pytorch version results in a warning
# rather than an error.
#
# Note: the CUDA torch version is derived from pyproject.toml and various
# requirements.txt files and should be kept consistent.  The ROCm torch
# versions are derived from Dockerfile.rocm
#
set(TORCH_SUPPORTED_VERSION_CUDA "2.3.0")
set(TORCH_SUPPORTED_VERSION_ROCM_5X "2.0.1")
set(TORCH_SUPPORTED_VERSION_ROCM_6X "2.1.1")

#
# Try to find python package with an executable that exactly matches
# `APHRODITE_PYTHON_EXECUTABLE` and is one of the supported versions.
#
if (APHRODITE_PYTHON_EXECUTABLE)
  find_python_from_executable(${APHRODITE_PYTHON_EXECUTABLE} "${PYTHON_SUPPORTED_VERSIONS}")
else()
  message(FATAL_ERROR
    "Please set APHRODITE_PYTHON_EXECUTABLE to the path of the desired python version"
    " before running cmake configure.")
endif()

#
# Update cmake's `CMAKE_PREFIX_PATH` with torch location.
#
append_cmake_prefix_path("torch" "torch.utils.cmake_prefix_path")

# Ensure the 'nvcc' command is in the PATH
find_program(NVCC_EXECUTABLE nvcc)
if (CUDA_FOUND AND NOT NVCC_EXECUTABLE)
    message(FATAL_ERROR "nvcc not found")
endif()

#
# Import torch cmake configuration.
# Torch also imports CUDA (and partially HIP) languages with some customizations,
# so there is no need to do this explicitly with check_language/enable_language,
# etc.
#
find_package(Torch REQUIRED)

#
# Normally `torch.utils.cpp_extension.CUDAExtension` would add
# `libtorch_python.so` for linking against an extension. Torch's cmake
# configuration does not include this library (presumably since the cmake
# config is used for standalone C++ binaries that link against torch).
# The `libtorch_python.so` library defines some of the glue code between
# torch/python via pybind and is required by APHRODITE extensions for this
# reason. So, add it by manually using `append_torchlib_if_found` from
# torch's cmake setup.
#
find_library(torch_python_LIBRARY torch_python PATHS
  "${TORCH_INSTALL_PREFIX}/lib")

#
# Forward the non-CUDA device extensions to external CMake scripts.
#
if (NOT APHRODITE_TARGET_DEVICE STREQUAL "cuda" AND
    NOT APHRODITE_TARGET_DEVICE STREQUAL "rocm")
    if (APHRODITE_TARGET_DEVICE STREQUAL "cpu")
        include(${CMAKE_CURRENT_LIST_DIR}/cmake/cpu_extension.cmake)
    else()
        message(FATAL_ERROR "Unsupported Aphrodite target device: ${APHRODITE_TARGET_DEVICE}")
    endif()
    return()
endif()

#
# Set up GPU language and check the torch version and warn if it isn't
# what is expected.
#
if (NOT HIP_FOUND AND CUDA_FOUND)
  set(APHRODITE_GPU_LANG "CUDA")

  if (NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_CUDA} "
      "expected for CUDA build, saw ${Torch_VERSION} instead.")
  endif()
elseif(HIP_FOUND)
  set(APHRODITE_GPU_LANG "HIP")

  # Importing torch recognizes and sets up some HIP/ROCm configuration but does
  # not let cmake recognize .hip files. In order to get cmake to understand the
  # .hip extension automatically, HIP must be enabled explicitly.
  enable_language(HIP)

  # ROCm 5.x
  if (ROCM_VERSION_DEV_MAJOR EQUAL 5 AND
      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM_5X})
    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM_5X} "
      "expected for ROCMm 5.x build, saw ${Torch_VERSION} instead.")
  endif()

  # ROCm 6.x
  if (ROCM_VERSION_DEV_MAJOR EQUAL 6 AND
      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM_6X})
    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM_6X} "
      "expected for ROCMm 6.x build, saw ${Torch_VERSION} instead.")
  endif()
else()
  message(FATAL_ERROR "Can't find CUDA or HIP installation.")
endif()

#
# Override the GPU architectures detected by cmake/torch and filter them by
# the supported versions for the current language.
# The final set of arches is stored in `APHRODITE_GPU_ARCHES`.
#
override_gpu_arches(APHRODITE_GPU_ARCHES
  ${APHRODITE_GPU_LANG}
  "${${APHRODITE_GPU_LANG}_SUPPORTED_ARCHS}")

#
# Query torch for additional GPU compilation flags for the given
# `APHRODITE_GPU_LANG`.
# The final set of arches is stored in `APHRODITE_GPU_FLAGS`.
#
get_torch_gpu_compiler_flags(APHRODITE_GPU_FLAGS ${APHRODITE_GPU_LANG})

#
# Set nvcc parallelism.
#
if(NVCC_THREADS AND APHRODITE_GPU_LANG STREQUAL "CUDA")
  list(APPEND APHRODITE_GPU_FLAGS "--threads=${NVCC_THREADS}")
endif()

#
# Define extension targets
#

#
# _C extension
#

set(APHRODITE_EXT_SRC
  "kernels/cache_kernels.cu"
  "kernels/attention/attention_kernels.cu"
  "kernels/pos_encoding_kernels.cu"
  "kernels/activation_kernels.cu"
  "kernels/layernorm_kernels.cu"
  "kernels/cuda_utils_kernels.cu"
  "kernels/moe/align_block_size_kernel.cu"
  "kernels/pybind.cpp")

if(APHRODITE_GPU_LANG STREQUAL "CUDA")
  list(APPEND APHRODITE_EXT_SRC
    "kernels/all_reduce/custom_all_reduce.cu")
endif()

define_gpu_extension_target(
  _C
  DESTINATION aphrodite
  LANGUAGE ${APHRODITE_GPU_LANG}
  SOURCES ${APHRODITE_EXT_SRC}
  COMPILE_FLAGS ${APHRODITE_GPU_FLAGS}
  ARCHITECTURES ${APHRODITE_GPU_ARCHES}
  WITH_SOABI)


#
# _quant_C extension
#

set (APHRODITE_QUANT_EXT_SRC
  "kernels/quantization/gptq/q_gemm.cu"
  "kernels/quantization/squeezellm/quant_cuda_kernel.cu"
  "kernels/quantization/exl2/q_matrix.cu"
  "kernels/quantization/exl2/q_gemm_exl2.cu"
  "kernels/quantization/fp8/common.cu"
  "kernels/quantization/compressed_tensors/int8_quant_kernels.cu"
  "kernels/quantization/quant_ops.cpp")

if(APHRODITE_GPU_LANG STREQUAL "CUDA")
include(FetchContent)
  SET(CUTLASS_ENABLE_HEADERS_ONLY=ON)
  FetchContent_Declare(
        cutlass 
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
        # CUTLASS 3.5.0
        GIT_TAG 7d49e6c7e2f8896c47f586706e67e1fb215529dc
  )
  FetchContent_MakeAvailable(cutlass)
  list(APPEND APHRODITE_QUANT_EXT_SRC
    "kernels/quantization/aqlm/gemm_kernels.cu"
    "kernels/quantization/awq/gemm_kernels.cu"
    "kernels/quantization/bitsandbytes/int4_fp16_gemm_kernels.cu"
    "kernels/quantization/bitsandbytes/format.cu"
    "kernels/quantization/bitsandbytes/gemm_s4_f16.cu"
    "kernels/quantization/gguf/gguf_kernel.cu"
    "kernels/quantization/marlin/dense/marlin_cuda_kernel.cu"
    "kernels/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
    "kernels/quantization/gptq_marlin/gptq_marlin.cu"
    "kernels/quantization/gptq_marlin/gptq_marlin_repack.cu"
    "kernels/quantization/quip/origin_order.cu"
    "kernels/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu"
    "kernels/quantization/cutlass_w8a8/scaled_mm_dq_c2x.cu"
    "kernels/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu")

    #
    # The CUTLASS kernels for Hopper require sm90a to be enabled.
    # This is done via the below gencode option, BUT that creates kernels for both sm90 and sm90a.
    # That adds an extra 17MB to compiled binary, so instead we selectively enable it.
    set_source_files_properties(
        "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu"
        PROPERTIES
        COMPILE_FLAGS
        "-gencode arch=compute_90a,code=sm_90a")
endif()

define_gpu_extension_target(
  _quant_C
  DESTINATION aphrodite
  LANGUAGE ${APHRODITE_GPU_LANG}
  SOURCES ${APHRODITE_QUANT_EXT_SRC}
  COMPILE_FLAGS ${APHRODITE_GPU_FLAGS}
  ARCHITECTURES ${APHRODITE_GPU_ARCHES}
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
  WITH_SOABI)

#
# _moe_C extension
#

set(APHRODITE_MOE_EXT_SRC
  "kernels/moe/moe_ops.cpp"
  "kernels/moe/softmax.cu")

define_gpu_extension_target(
  _moe_C
  DESTINATION aphrodite
  LANGUAGE ${APHRODITE_GPU_LANG}
  SOURCES ${APHRODITE_MOE_EXT_SRC}
  COMPILE_FLAGS ${APHRODITE_GPU_FLAGS}
  ARCHITECTURES ${APHRODITE_GPU_ARCHES}
  WITH_SOABI)

#
# _punica_C extension
#

set(APHRODITE_PUNICA_EXT_SRC
  "kernels/punica/bgmv/bgmv_bf16_bf16_bf16.cu"
  "kernels/punica/bgmv/bgmv_bf16_bf16_fp16.cu"
  "kernels/punica/bgmv/bgmv_bf16_fp16_bf16.cu"
  "kernels/punica/bgmv/bgmv_bf16_fp16_fp16.cu"
  "kernels/punica/bgmv/bgmv_bf16_fp32_bf16.cu"
  "kernels/punica/bgmv/bgmv_bf16_fp32_fp16.cu"
  "kernels/punica/bgmv/bgmv_fp16_bf16_bf16.cu"
  "kernels/punica/bgmv/bgmv_fp16_bf16_fp16.cu"
  "kernels/punica/bgmv/bgmv_fp16_fp16_bf16.cu"
  "kernels/punica/bgmv/bgmv_fp16_fp16_fp16.cu"
  "kernels/punica/bgmv/bgmv_fp16_fp32_bf16.cu"
  "kernels/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
  "kernels/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
  "kernels/punica/bgmv/bgmv_fp32_bf16_fp16.cu"
  "kernels/punica/bgmv/bgmv_fp32_fp16_bf16.cu"
  "kernels/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
  "kernels/punica/bgmv/bgmv_fp32_fp32_bf16.cu"
  "kernels/punica/bgmv/bgmv_fp32_fp32_fp16.cu"
  "kernels/punica/punica_ops.cu"
  "kernels/punica/punica_pybind.cpp")

#
# Copy GPU compilation flags+update for punica
#
set(APHRODITE_PUNICA_GPU_FLAGS ${APHRODITE_GPU_FLAGS})
list(REMOVE_ITEM APHRODITE_PUNICA_GPU_FLAGS
  "-D__CUDA_NO_HALF_OPERATORS__"
  "-D__CUDA_NO_HALF_CONVERSIONS__"
  "-D__CUDA_NO_BFLOAT16_CONVERSIONS__"
  "-D__CUDA_NO_HALF2_OPERATORS__")

#
# Filter out CUDA architectures < 8.0 for punica.
#
if (${APHRODITE_GPU_LANG} STREQUAL "CUDA")
  set(APHRODITE_PUNICA_GPU_ARCHES)
  foreach(ARCH ${APHRODITE_GPU_ARCHES})
    string_to_ver(CODE_VER ${ARCH})
    if (CODE_VER GREATER_EQUAL 8.0)
      list(APPEND APHRODITE_PUNICA_GPU_ARCHES ${ARCH})
    endif()
  endforeach()
  message(STATUS "Punica target arches: ${APHRODITE_PUNICA_GPU_ARCHES}")
elseif(${APHRODITE_GPU_LANG} STREQUAL "HIP")
  set(APHRODITE_PUNICA_GPU_ARCHES ${APHRODITE_GPU_ARCHES})
  message(STATUS "Punica target arches: ${APHRODITE_PUNICA_GPU_ARCHES}")
endif()

if (APHRODITE_PUNICA_GPU_ARCHES)
  define_gpu_extension_target(
    _punica_C
    DESTINATION aphrodite
    LANGUAGE ${APHRODITE_GPU_LANG}
    SOURCES ${APHRODITE_PUNICA_EXT_SRC}
    COMPILE_FLAGS ${APHRODITE_PUNICA_GPU_FLAGS}
    ARCHITECTURES ${APHRODITE_PUNICA_GPU_ARCHES}
    WITH_SOABI)
else()
  message(WARNING "Unable to create _punica_C target because none of the "
    "requested architectures (${APHRODITE_GPU_ARCHES}) are supported, i.e. >= 8.0")
endif()

#
# _hadamard_C extension
#

set(APHRODITE_HADAMARD_EXT_SRC
  "kernels/hadamard/fast_hadamard_transform.cpp"
  "kernels/hadamard/fast_hadamard_transform_cuda.cu")

#
# Copy GPU compilation flags+update for hadamard
#
set(APHRODITE_HADAMARD_GPU_FLAGS ${APHRODITE_GPU_FLAGS})
list(APPEND APHRODITE_HADAMARD_GPU_FLAGS
  "-U__CUDA_NO_HALF_OPERATORS__"
  "-U__CUDA_NO_HALF_CONVERSIONS__"
  "-U__CUDA_NO_BFLOAT16_OPERATORS__"
  "-U__CUDA_NO_BFLOAT16_CONVERSIONS__"
  "-U__CUDA_NO_BFLOAT162_OPERATORS__"
  "-U__CUDA_NO_BFLOAT162_CONVERSIONS__"
  "--expt-relaxed-constexpr"
  "--expt-extended-lambda"
  "--use_fast_math"
  "-lineinfo")

#
# Filter out CUDA architectures < 7.0 for hadamard.
#
if (${APHRODITE_GPU_LANG} STREQUAL "CUDA")
  set(APHRODITE_HADAMARD_GPU_ARCHES)
  foreach(ARCH ${APHRODITE_GPU_ARCHES})
    string_to_ver(CODE_VER ${ARCH})
    if (CODE_VER GREATER_EQUAL 6.0)
      list(APPEND APHRODITE_HADAMARD_GPU_ARCHES ${ARCH})
    endif()
  endforeach()
  message(STATUS "Hadamard target arches: ${APHRODITE_HADAMARD_GPU_ARCHES}")
endif()

if (APHRODITE_HADAMARD_GPU_ARCHES)
  define_gpu_extension_target(
    _hadamard_C
    DESTINATION aphrodite
    LANGUAGE ${APHRODITE_GPU_LANG}
    SOURCES ${APHRODITE_HADAMARD_EXT_SRC}
    COMPILE_FLAGS ${APHRODITE_HADAMARD_GPU_FLAGS}
    ARCHITECTURES ${APHRODITE_HADAMARD_GPU_ARCHES}
    WITH_SOABI)
else()
  message(WARNING "Unable to create _hadamard_C target because none of the "
    "requested architectures (${APHRODITE_GPU_ARCHES}) are supported, i.e. >= 6.0")
endif()

#
# Add the `default` target which detects which extensions should be
# built based on platform/architecture.  This is the same logic that
# setup.py uses to select which extensions should be built and should
# be kept in sync.
#
# The `default` target makes direct use of cmake easier since knowledge
# of which extensions are supported has been factored in, e.g.
#
# mkdir build && cd build
# cmake -G Ninja -DAPHRODITE_PYTHON_EXECUTABLE=`which python3` -DCMAKE_LIBRARY_OUTPUT_DIRECTORY=../aphrodite ..
# cmake --build . --target default
#
add_custom_target(default)

if(APHRODITE_GPU_LANG STREQUAL "CUDA" OR APHRODITE_GPU_LANG STREQUAL "HIP")
  message(STATUS "Enabling C extension.")
  add_dependencies(default _C)
  # Enable punica if -DAPHRODITE_INSTALL_PUNICA_KERNELS=ON or
  # APHRODITE_INSTALL_PUNICA_KERNELS is set in the environment and
  # there are supported target arches.
  if (APHRODITE_QUANT_EXT_SRC AND
      (ENV{APHRODITE_INSTALL_QUANT_KERNELS} OR APHRODITE_INSTALL_QUANT_KERNELS))
    message(STATUS "Enabling quant extension.")
    add_dependencies(default _quant_C)
  endif()
  if (APHRODITE_PUNICA_GPU_ARCHES AND
      (ENV{APHRODITE_INSTALL_PUNICA_KERNELS} OR APHRODITE_INSTALL_PUNICA_KERNELS))
    message(STATUS "Enabling punica extension.")
    add_dependencies(default _punica_C)
  endif()
  if (APHRODITE_HADAMARD_GPU_ARCHES AND
      (ENV{APHRODITE_INSTALL_HADAMARD_KERNELS} OR APHRODITE_INSTALL_HADAMARD_KERNELS))
    message(STATUS "Enabling hadamard extension.")
    add_dependencies(default _hadamard_C)
  endif()
endif()

if(APHRODITE_GPU_LANG STREQUAL "CUDA")
  message(STATUS "Enabling moe extension.")
  add_dependencies(default _moe_C)
endif()