Cmake based build system (#2830)

CMakeLists.txt

+cmake_minimum_required(VERSION 3.21)
+
+project(vllm_extensions LANGUAGES CXX)
+
+message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
+
+include(${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 "7.0;7.5;8.0;8.6;8.9;9.0")
+
+# Supported AMD GPU architectures.
+set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;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 "2.1.2")
+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
+# `VLLM_PYTHON_EXECUTABLE` and is one of the supported versions.
+#
+if (VLLM_PYTHON_EXECUTABLE)
+  find_python_from_executable(${VLLM_PYTHON_EXECUTABLE} "${PYTHON_SUPPORTED_VERSIONS}")
+else()
+  message(FATAL_ERROR
+    "Please set VLLM_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")
+
+#
+# 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 VLLM extensions for this
+# reason. So, add it by manually using `append_torchlib_if_found` from
+# torch's cmake setup.
+#
+append_torchlib_if_found(torch_python)
+
+#
+# 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(VLLM_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(VLLM_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 `VLLM_GPU_ARCHES`.
+#
+override_gpu_arches(VLLM_GPU_ARCHES
+  ${VLLM_GPU_LANG}
+  "${${VLLM_GPU_LANG}_SUPPORTED_ARCHS}")
+
+#
+# Query torch for additional GPU compilation flags for the given
+# `VLLM_GPU_LANG`.
+# The final set of arches is stored in `VLLM_GPU_FLAGS`.
+#
+get_torch_gpu_compiler_flags(VLLM_GPU_FLAGS ${VLLM_GPU_LANG})
+
+#
+# Set nvcc parallelism.
+#
+if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
+endif()
+
+#
+# Define extension targets
+#
+
+#
+# _C extension
+#
+
+set(VLLM_EXT_SRC
+  "csrc/cache_kernels.cu"
+  "csrc/attention/attention_kernels.cu"
+  "csrc/pos_encoding_kernels.cu"
+  "csrc/activation_kernels.cu"
+  "csrc/layernorm_kernels.cu"
+  "csrc/quantization/squeezellm/quant_cuda_kernel.cu"
+  "csrc/quantization/gptq/q_gemm.cu"
+  "csrc/cuda_utils_kernels.cu"
+  "csrc/moe_align_block_size_kernels.cu"
+  "csrc/pybind.cpp")
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_EXT_SRC
+    "csrc/quantization/awq/gemm_kernels.cu"
+    "csrc/quantization/marlin/marlin_cuda_kernel.cu"
+    "csrc/custom_all_reduce.cu")
+endif()
+
+define_gpu_extension_target(
+  _C
+  DESTINATION vllm
+  LANGUAGE ${VLLM_GPU_LANG}
+  SOURCES ${VLLM_EXT_SRC}
+  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+  ARCHITECTURES ${VLLM_GPU_ARCHES}
+  WITH_SOABI)
+
+#
+# _moe_C extension
+#
+
+set(VLLM_MOE_EXT_SRC
+  "csrc/moe/moe_ops.cpp"
+  "csrc/moe/topk_softmax_kernels.cu")
+
+define_gpu_extension_target(
+  _moe_C
+  DESTINATION vllm
+  LANGUAGE ${VLLM_GPU_LANG}
+  SOURCES ${VLLM_MOE_EXT_SRC}
+  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+  ARCHITECTURES ${VLLM_GPU_ARCHES}
+  WITH_SOABI)
+
+#
+# _punica_C extension
+#
+
+set(VLLM_PUNICA_EXT_SRC
+  "csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_bf16_bf16_fp16.cu"
+  "csrc/punica/bgmv/bgmv_bf16_fp16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_bf16_fp16_fp16.cu"
+  "csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu"
+  "csrc/punica/bgmv/bgmv_bf16_fp32_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_bf16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_bf16_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_fp16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_fp32_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_bf16_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_fp16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_fp32_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_fp32_fp16.cu"
+  "csrc/punica/punica_ops.cc")
+
+#
+# Copy GPU compilation flags+update for punica
+#
+set(VLLM_PUNICA_GPU_FLAGS ${VLLM_GPU_FLAGS})
+list(REMOVE_ITEM VLLM_PUNICA_GPU_FLAGS
+  "-D__CUDA_NO_HALF_OPERATORS__"
+  "-D__CUDA_NO_HALF_CONVERSIONS__"
+  "-D__CUDA_NO_BFLOAT16_CONVERSIONS__"
+  "-D__CUDA_NO_HALF2_OPERATORS__")
+
+#
+# Filter out CUDA architectures < 8.0 for punica.
+#
+if (${VLLM_GPU_LANG} STREQUAL "CUDA")
+  set(VLLM_PUNICA_GPU_ARCHES)
+  foreach(ARCH ${VLLM_GPU_ARCHES})
+    string_to_ver(CODE_VER ${ARCH})
+    if (CODE_VER GREATER_EQUAL 8.0)
+      list(APPEND VLLM_PUNICA_GPU_ARCHES ${ARCH})
+    endif()
+  endforeach()
+  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
+endif()
+
+if (VLLM_PUNICA_GPU_ARCHES)
+  define_gpu_extension_target(
+    _punica_C
+    DESTINATION vllm
+    LANGUAGE ${VLLM_GPU_LANG}
+    SOURCES ${VLLM_PUNICA_EXT_SRC}
+    COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
+    ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
+    WITH_SOABI)
+else()
+  message(WARNING "Unable to create _punica_C target because none of the "
+    "requested architectures (${VLLM_GPU_ARCHES}) are supported, i.e. >= 8.0")
+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 -DVLLM_PYTHON_EXECUTABLE=`which python3` -DCMAKE_LIBRARY_OUTPUT_DIRECTORY=../vllm ..
+# cmake --build . --target default
+#
+add_custom_target(default)
+
+if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
+  message(STATUS "Enabling C extension.")
+  add_dependencies(default _C)
+endif()
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  message(STATUS "Enabling moe extension.")
+  add_dependencies(default _moe_C)
+
+  # Enable punica if -DVLLM_INSTALL_PUNICA_KERNELS=ON or
+  # VLLM_INSTALL_PUNICA_KERNELS is set in the environment and
+  # there are supported target arches.
+  if (VLLM_PUNICA_GPU_ARCHES AND
+      (ENV{VLLM_INSTALL_PUNICA_KERNELS} OR VLLM_INSTALL_PUNICA_KERNELS))
+    message(STATUS "Enabling punica extension.")
+    add_dependencies(default _punica_C)
+  endif()
+endif()

Dockerfile

@@ -38,6 +38,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 # copy input files
 COPY csrc csrc
 COPY setup.py setup.py
+COPY cmake cmake
+COPY CMakeLists.txt CMakeLists.txt
 COPY requirements.txt requirements.txt
 COPY pyproject.toml pyproject.toml
 COPY vllm/__init__.py vllm/__init__.py

MANIFEST.in

 include LICENSE
 include requirements.txt
+include CMakeLists.txt
 
+recursive-include cmake *
 recursive-include csrc *

cmake/hipify.py

+#!/usr/bin/env python3
+
+#
+# A command line tool for running pytorch's hipify preprocessor on CUDA
+# source files.
+#
+# See https://github.com/ROCm/hipify_torch
+# and <torch install dir>/utils/hipify/hipify_python.py
+#
+
+import argparse
+import shutil
+import os
+
+from torch.utils.hipify.hipify_python import hipify
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+
+    # Project directory where all the source + include files live.
+    parser.add_argument(
+        "-p",
+        "--project_dir",
+        help="The project directory.",
+    )
+
+    # Directory where hipified files are written.
+    parser.add_argument(
+        "-o",
+        "--output_dir",
+        help="The output directory.",
+    )
+
+    # Source files to convert.
+    parser.add_argument("sources",
+                        help="Source files to hipify.",
+                        nargs="*",
+                        default=[])
+
+    args = parser.parse_args()
+
+    # Limit include scope to project_dir only
+    includes = [os.path.join(args.project_dir, '*')]
+
+    # Get absolute path for all source files.
+    extra_files = [os.path.abspath(s) for s in args.sources]
+
+    # Copy sources from project directory to output directory.
+    # The directory might already exist to hold object files so we ignore that.
+    shutil.copytree(args.project_dir, args.output_dir, dirs_exist_ok=True)
+
+    hipify_result = hipify(project_directory=args.project_dir,
+                           output_directory=args.output_dir,
+                           header_include_dirs=[],
+                           includes=includes,
+                           extra_files=extra_files,
+                           show_detailed=True,
+                           is_pytorch_extension=True,
+                           hipify_extra_files_only=True)
+
+    hipified_sources = []
+    for source in args.sources:
+        s_abs = os.path.abspath(source)
+        hipified_s_abs = (hipify_result[s_abs].hipified_path if
+                          (s_abs in hipify_result
+                           and hipify_result[s_abs].hipified_path is not None)
+                          else s_abs)
+        hipified_sources.append(hipified_s_abs)
+
+    assert (len(hipified_sources) == len(args.sources))
+
+    # Print hipified source files.
+    print("\n".join(hipified_sources))

cmake/utils.cmake

+#
+# Attempt to find the python package that uses the same python executable as
+# `EXECUTABLE` and is one of the `SUPPORTED_VERSIONS`.
+#
+macro (find_python_from_executable EXECUTABLE SUPPORTED_VERSIONS)
+  file(REAL_PATH ${EXECUTABLE} EXECUTABLE)
+  set(Python_EXECUTABLE ${EXECUTABLE})
+  find_package(Python COMPONENTS Interpreter Development.Module)
+  if (NOT Python_FOUND)
+    message(FATAL_ERROR "Unable to find python matching: ${EXECUTABLE}.")
+  endif()
+  set(_VER "${Python_VERSION_MAJOR}.${Python_VERSION_MINOR}")
+  set(_SUPPORTED_VERSIONS_LIST ${SUPPORTED_VERSIONS} ${ARGN})
+  if (NOT _VER IN_LIST _SUPPORTED_VERSIONS_LIST)
+    message(FATAL_ERROR
+      "Python version (${_VER}) is not one of the supported versions: "
+      "${_SUPPORTED_VERSIONS_LIST}.")
+  endif()
+  message(STATUS "Found python matching: ${EXECUTABLE}.")
+endmacro()
+
+#
+# Run `EXPR` in python.  The standard output of python is stored in `OUT` and
+# has trailing whitespace stripped.  If an error is encountered when running
+# python, a fatal message `ERR_MSG` is issued.
+#
+function (run_python OUT EXPR ERR_MSG)
+  execute_process(
+    COMMAND
+    "${Python_EXECUTABLE}" "-c" "${EXPR}"
+    OUTPUT_VARIABLE PYTHON_OUT
+    RESULT_VARIABLE PYTHON_ERROR_CODE
+    ERROR_VARIABLE PYTHON_STDERR
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+  if(NOT PYTHON_ERROR_CODE EQUAL 0)
+    message(FATAL_ERROR "${ERR_MSG}: ${PYTHON_STDERR}")
+  endif()
+  set(${OUT} ${PYTHON_OUT} PARENT_SCOPE)
+endfunction()
+
+# Run `EXPR` in python after importing `PKG`. Use the result of this to extend
+# `CMAKE_PREFIX_PATH` so the torch cmake configuration can be imported.
+macro (append_cmake_prefix_path PKG EXPR)
+  run_python(_PREFIX_PATH
+    "import ${PKG}; print(${EXPR})" "Failed to locate ${PKG} path")
+  list(APPEND CMAKE_PREFIX_PATH ${_PREFIX_PATH})
+endmacro()
+
+#
+# Add a target named `hipify${NAME}` that runs the hipify preprocessor on a set
+# of CUDA source files. The names of the corresponding "hipified" sources are
+# stored in `OUT_SRCS`.
+#
+function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
+  #
+  # Split into C++ and non-C++ (i.e. CUDA) sources.
+  #
+  set(SRCS ${ORIG_SRCS})
+  set(CXX_SRCS ${ORIG_SRCS})
+  list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)$")
+  list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)$")
+
+  #
+  # Generate ROCm/HIP source file names from CUDA file names.
+  # Since HIP files are generated code, they will appear in the build area
+  # `CMAKE_CURRENT_BINARY_DIR` directory rather than the original csrc dir.
+  #
+  set(HIP_SRCS)
+  foreach (SRC ${SRCS})
+    string(REGEX REPLACE "\.cu$" "\.hip" SRC ${SRC})
+    string(REGEX REPLACE "cuda" "hip" SRC ${SRC})
+    list(APPEND HIP_SRCS "${CMAKE_CURRENT_BINARY_DIR}/${SRC}")
+  endforeach()
+
+  set(CSRC_BUILD_DIR ${CMAKE_CURRENT_BINARY_DIR}/csrc)
+  add_custom_target(
+    hipify${NAME}
+    COMMAND ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
+    DEPENDS ${CMAKE_SOURCE_DIR}/cmake/hipify.py ${SRCS}
+    BYPRODUCTS ${HIP_SRCS}
+    COMMENT "Running hipify on ${NAME} extension source files.")
+
+  # Swap out original extension sources with hipified sources.
+  list(APPEND HIP_SRCS ${CXX_SRCS})
+  set(${OUT_SRCS} ${HIP_SRCS} PARENT_SCOPE)
+endfunction()
+
+#
+# Get additional GPU compiler flags from torch.
+#
+function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
+  if (${GPU_LANG} STREQUAL "CUDA")
+    #
+    # Get common NVCC flags from torch.
+    #
+    run_python(GPU_FLAGS
+      "from torch.utils.cpp_extension import COMMON_NVCC_FLAGS; print(';'.join(COMMON_NVCC_FLAGS))"
+      "Failed to determine torch nvcc compiler flags")
+
+    if (CUDA_VERSION VERSION_GREATER_EQUAL 11.8)
+      list(APPEND GPU_FLAGS "-DENABLE_FP8_E5M2")
+    endif()
+
+  elseif(${GPU_LANG} STREQUAL "HIP")
+    #
+    # Get common HIP/HIPCC flags from torch.
+    #
+    run_python(GPU_FLAGS
+      "import torch.utils.cpp_extension as t; print(';'.join(t.COMMON_HIP_FLAGS + t.COMMON_HIPCC_FLAGS))"
+      "Failed to determine torch nvcc compiler flags")
+
+    list(APPEND GPU_FLAGS
+      "-DUSE_ROCM"
+      "-U__HIP_NO_HALF_CONVERSIONS__"
+      "-U__HIP_NO_HALF_OPERATORS__"
+      "-fno-gpu-rdc")
+
+  endif()
+  set(${OUT_GPU_FLAGS} ${GPU_FLAGS} PARENT_SCOPE)
+endfunction()
+
+# Macro for converting a `gencode` version number to a cmake version number.
+macro(string_to_ver OUT_VER IN_STR)
+  string(REGEX REPLACE "\([0-9]+\)\([0-9]\)" "\\1.\\2" ${OUT_VER} ${IN_STR})
+endmacro()
+
+#
+# Override the GPU architectures detected by cmake/torch and filter them by
+# `GPU_SUPPORTED_ARCHES`. Sets the final set of architectures in
+# `GPU_ARCHES`.
+#
+# Note: this is defined as a macro since it updates `CMAKE_CUDA_FLAGS`.
+#
+macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
+  set(_GPU_SUPPORTED_ARCHES_LIST ${GPU_SUPPORTED_ARCHES} ${ARGN})
+  message(STATUS "${GPU_LANG} supported arches: ${_GPU_SUPPORTED_ARCHES_LIST}")
+
+  if (${GPU_LANG} STREQUAL "HIP")
+    #
+    # `GPU_ARCHES` controls the `--offload-arch` flags.
+    # `CMAKE_HIP_ARCHITECTURES` is set up by torch and can be controlled
+    # via the `PYTORCH_ROCM_ARCH` env variable.
+    #
+
+    #
+    # Find the intersection of the supported + detected architectures to
+    # set the module architecture flags.
+    #
+    set(${GPU_ARCHES})
+    foreach (_ARCH ${CMAKE_HIP_ARCHITECTURES})
+      if (_ARCH IN_LIST _GPU_SUPPORTED_ARCHES_LIST)
+        list(APPEND ${GPU_ARCHES} ${_ARCH})
+      endif()
+    endforeach()
+
+    if(NOT ${GPU_ARCHES})
+      message(FATAL_ERROR
+        "None of the detected ROCm architectures: ${CMAKE_HIP_ARCHITECTURES} is"
+        " supported. Supported ROCm architectures are: ${_GPU_SUPPORTED_ARCHES_LIST}.")
+    endif()
+
+  elseif(${GPU_LANG} STREQUAL "CUDA")
+    #
+    # Setup/process CUDA arch flags.
+    #
+    # The torch cmake setup hardcodes the detected architecture flags in
+    # `CMAKE_CUDA_FLAGS`.  Since `CMAKE_CUDA_FLAGS` is a "global" variable, it
+    # can't modified on a per-target basis, e.g. for the `punica` extension.
+    # So, all the `-gencode` flags need to be extracted and removed from
+    # `CMAKE_CUDA_FLAGS` for processing so they can be passed by another method.
+    # Since it's not possible to use `target_compiler_options` for adding target
+    # specific `-gencode` arguments, the target's `CUDA_ARCHITECTURES` property
+    # must be used instead.  This requires repackaging the architecture flags
+    # into a format that cmake expects for `CUDA_ARCHITECTURES`.
+    #
+    # This is a bit fragile in that it depends on torch using `-gencode` as opposed
+    # to one of the other nvcc options to specify architectures.
+    #
+    # Note: torch uses the `TORCH_CUDA_ARCH_LIST` environment variable to override
+    # detected architectures.
+    #
+    message(DEBUG "initial CMAKE_CUDA_FLAGS: ${CMAKE_CUDA_FLAGS}")
+
+    # Extract all `-gencode` flags from `CMAKE_CUDA_FLAGS`
+    string(REGEX MATCHALL "-gencode arch=[^ ]+" _CUDA_ARCH_FLAGS
+      ${CMAKE_CUDA_FLAGS})
+
+    # Remove all `-gencode` flags from `CMAKE_CUDA_FLAGS` since they will be modified
+    # and passed back via the `CUDA_ARCHITECTURES` property.
+    string(REGEX REPLACE "-gencode arch=[^ ]+ *" "" CMAKE_CUDA_FLAGS
+      ${CMAKE_CUDA_FLAGS})
+
+    # If this error is triggered, it might mean that torch has changed how it sets
+    # up nvcc architecture code generation flags.
+    if (NOT _CUDA_ARCH_FLAGS)
+      message(FATAL_ERROR
+        "Could not find any architecture related code generation flags in "
+        "CMAKE_CUDA_FLAGS. (${CMAKE_CUDA_FLAGS})")
+    endif()
+
+    message(DEBUG "final CMAKE_CUDA_FLAGS: ${CMAKE_CUDA_FLAGS}")
+    message(DEBUG "arch flags: ${_CUDA_ARCH_FLAGS}")
+
+    # Initialize the architecture lists to empty.
+    set(${GPU_ARCHES})
+
+    # Process each `gencode` flag.
+    foreach(_ARCH ${_CUDA_ARCH_FLAGS})
+      # For each flag, extract the version number and whether it refers to PTX
+      # or native code.
+      # Note: if a regex matches then `CMAKE_MATCH_1` holds the binding
+      # for that match.
+
+      string(REGEX MATCH "arch=compute_\([0-9]+a?\)" _COMPUTE ${_ARCH})
+      if (_COMPUTE)
+        set(_COMPUTE ${CMAKE_MATCH_1})
+      endif()
+
+      string(REGEX MATCH "code=sm_\([0-9]+a?\)" _SM ${_ARCH})
+      if (_SM)
+        set(_SM ${CMAKE_MATCH_1})
+      endif()
+
+      string(REGEX MATCH "code=compute_\([0-9]+a?\)" _CODE ${_ARCH})
+      if (_CODE)
+        set(_CODE ${CMAKE_MATCH_1})
+      endif()
+
+      # Make sure the virtual architecture can be matched.
+      if (NOT _COMPUTE)
+        message(FATAL_ERROR
+          "Could not determine virtual architecture from: ${_ARCH}.")
+      endif()
+
+      # One of sm_ or compute_ must exist.
+      if ((NOT _SM) AND (NOT _CODE))
+        message(FATAL_ERROR
+          "Could not determine a codegen architecture from: ${_ARCH}.")
+      endif()
+
+      if (_SM)
+        set(_VIRT "")
+        set(_CODE_ARCH ${_SM})
+      else()
+        set(_VIRT "-virtual")
+        set(_CODE_ARCH ${_CODE})
+      endif()
+
+      # Check if the current version is in the supported arch list.
+      string_to_ver(_CODE_VER ${_CODE_ARCH})
+      if (NOT _CODE_VER IN_LIST _GPU_SUPPORTED_ARCHES_LIST)
+        message(STATUS "discarding unsupported CUDA arch ${_VER}.")
+        continue()
+      endif()
+
+      # Add it to the arch list.
+      list(APPEND ${GPU_ARCHES} "${_CODE_ARCH}${_VIRT}")
+    endforeach()
+  endif()
+  message(STATUS "${GPU_LANG} target arches: ${${GPU_ARCHES}}")
+endmacro()
+
+#
+# Define a target named `GPU_MOD_NAME` for a single extension. The
+# arguments are:
+#
+# DESTINATION <dest>         - Module destination directory.
+# LANGUAGE <lang>            - The GPU language for this module, e.g CUDA, HIP,
+#                              etc.
+# SOURCES <sources>          - List of source files relative to CMakeLists.txt
+#                              directory.
+#
+# Optional arguments:
+#
+# ARCHITECTURES <arches>     - A list of target GPU architectures in cmake
+#                              format.
+#                              Refer `CMAKE_CUDA_ARCHITECTURES` documentation
+#                              and `CMAKE_HIP_ARCHITECTURES` for more info.
+#                              ARCHITECTURES will use cmake's defaults if
+#                              not provided.
+# COMPILE_FLAGS <flags>      - Extra compiler flags passed to NVCC/hip.
+# INCLUDE_DIRECTORIES <dirs> - Extra include directories.
+# LINK_LIBRARIES <libraries> - Extra link libraries.
+# WITH_SOABI                 - Generate library with python SOABI suffix name.
+#
+# Note: optimization level/debug info is set via cmake build type.
+#
+function (define_gpu_extension_target GPU_MOD_NAME)
+  cmake_parse_arguments(PARSE_ARGV 1
+    GPU
+    "WITH_SOABI"
+    "DESTINATION;LANGUAGE"
+    "SOURCES;ARCHITECTURES;COMPILE_FLAGS;INCLUDE_DIRECTORIES;LIBRARIES")
+
+  # Add hipify preprocessing step when building with HIP/ROCm.
+  if (GPU_LANGUAGE STREQUAL "HIP")
+    hipify_sources_target(GPU_SOURCES ${GPU_MOD_NAME} "${GPU_SOURCES}")
+  endif()
+
+  if (GPU_WITH_SOABI)
+    set(GPU_WITH_SOABI WITH_SOABI)
+  else()
+    set(GPU_WITH_SOABI)
+  endif()
+
+  Python_add_library(${GPU_MOD_NAME} MODULE "${GPU_SOURCES}" ${GPU_WITH_SOABI})
+
+  if (GPU_LANGUAGE STREQUAL "HIP")
+    # Make this target dependent on the hipify preprocessor step.
+    add_dependencies(${GPU_MOD_NAME} hipify${GPU_MOD_NAME})
+  endif()
+
+  if (GPU_ARCHITECTURES)
+    set_target_properties(${GPU_MOD_NAME} PROPERTIES
+      ${GPU_LANGUAGE}_ARCHITECTURES "${GPU_ARCHITECTURES}")
+  endif()
+
+  set_property(TARGET ${GPU_MOD_NAME} PROPERTY CXX_STANDARD 17)
+
+  target_compile_options(${GPU_MOD_NAME} PRIVATE
+    $<$<COMPILE_LANGUAGE:${GPU_LANGUAGE}>:${GPU_COMPILE_FLAGS}>)
+
+  target_compile_definitions(${GPU_MOD_NAME} PRIVATE
+    "-DTORCH_EXTENSION_NAME=${GPU_MOD_NAME}")
+
+  target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
+    ${GPU_INCLUDE_DIRECTORIES})
+
+  target_link_libraries(${GPU_MOD_NAME} PRIVATE ${TORCH_LIBRARIES}
+    ${GPU_LIBRARIES})
+
+  install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION})
+endfunction()

pyproject.toml

 [build-system]
 # Should be mirrored in requirements-build.txt
 requires = [
+    "cmake>=3.21",
     "ninja",
     "packaging",
     "setuptools >= 49.4.0",

requirements-build.txt

 # Should be mirrored in pyproject.toml
+cmake>=3.21
 ninja
 packaging
 setuptools>=49.4.0

requirements-rocm.txt

+cmake>=3.21
 ninja  # For faster builds.
 typing-extensions>=4.8.0
 starlette

requirements.txt

+cmake>=3.21
 ninja  # For faster builds.
 psutil
 ray >= 2.9