Create nvfuser_common package in build system by rdspring1 · Pull Request #4323 · NVIDIA/Fuser

rdspring1 · 2025-04-26T03:53:36Z

This PR moves common utilities from nvfuser package into nvfuser_common package, so nvfuser_next has access.

CI updates for `jit_examples_sinh_libtorch_20`

Change from

import nvfuser.utils;
import torch.utils;
print(nvfuser.utils.cmake_prefix_path, torch.utils.cmake_prefix_path, sep=";")

to

import nvfuser_common.utils;
import torch.utils;
print(nvfuser_common.utils.cmake_prefix_path, torch.utils.cmake_prefix_path, sep=";")

github-actions · 2025-04-26T03:54:31Z

Review updated until commit a59964c

Description

Created nvfuser_common package for shared utilities.
Updated import paths in various files.
Modified CMakeLists.txt to reflect new directory structure.

Changes walkthrough 📝

Relevant files

Enhancement

setup.py `Update import path in setup.py` examples/sinh_extension/setup.py Updated import path for `nvfuser` to `nvfuser_common`.	+1/-1
utils.py `Update cmake_prefix_path in utils.py` python/nvfuser_common/utils.py Updated `cmake_prefix_path` to reflect new package name.	+1/-1
utils.py `Update paths and package data in utils.py` python/utils.py Updated `libnvfuser_path` and `install_prefix` to use `nvfuser_common`. Renamed `nvfuser_package_data` to `nvfuser_common_package_data`. Updated `package_data` dictionary key to `nvfuser_common`.	+8/-4
CMakeLists.txt `Update CMakeLists.txt for new directory structure` CMakeLists.txt Added `NVFUSER_PYTHON_BINDINGS` variable. Updated paths for Python frontend source files.	+12/-11
README.md `Update import path in README.md` examples/sinh_libtorch/README.md Updated import path for `nvfuser` to `nvfuser_common`.	+1/-1

Documentation

__init__.py `Add license header to __init__.py` python/nvfuser_common/init.py Added SPDX license header.	+3/-0

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review

Import Path

Ensure that the new import path nvfuser_common is correctly set up and accessible in all environments where this setup script is used.

nvfuser_spec = importlib.util.find_spec("nvfuser_common")

Package Data

Verify that the package data paths are correctly updated to reflect the new nvfuser_common directory structure.

if config.build_setup:
    cmake(config, relative_path)
if not config.cmake_only:
    # NOTE: package include files for cmake
    # TODO(crcrpar): Better avoid hardcoding `libnvfuser_codegen.so`
    # might can be treated by using `exclude_package_data`.
    nvfuser_common_package_data = [

File Paths

Confirm that the updated file paths in CMakeLists.txt correctly point to the new nvfuser_common directory and that no files are missing.

set(NVFUSER_PYTHON_BINDINGS "${NVFUSER_ROOT}/python/python_frontend")
set(NVFUSER_THIRD_PARTY_DIR "${NVFUSER_ROOT}/third_party")

option(NVFUSER_STANDALONE_BUILD_WITH_UCC "" OFF)
option(NVFUSER_EXPLICIT_ERROR_CHECK "" OFF)

if(NVFUSER_EXPLICIT_ERROR_CHECK)
  add_compile_definitions(NVFUSER_EXPLICIT_ERROR_CHECK)
endif()

option(NVFUSER_BUILD_WITH_ASAN "Build nvFuser with asan" OFF)

include(CMakeDependentOption)
cmake_dependent_option(NVFUSER_DISTRIBUTED "" ON "USE_DISTRIBUTED" OFF)

if(NVFUSER_DISTRIBUTED)
  add_compile_definitions(NVFUSER_DISTRIBUTED)
endif()

message(STATUS "Setting NVFUSER_DISTRIBUTED=${NVFUSER_DISTRIBUTED}")

# We try to update which C++ standard we use together in lockstep across all
# built libraries, and these variables control which that is. Generally we are
# on C++20, but we still support a version of CUDA (11) that does not recognize
# C++20 and so we drop back to 17 there. Also, we allow all of these to be
# overridden by the user.
# Note we do not use a global set_property on e.g. CXX_STANDARD. CMake globals
# are footguns that should generally be avoided, because they are difficult to
# target where and *only* where they are needed. See e.g.:
# https://cliutils.gitlab.io/modern-cmake/chapters/intro/dodonot.html
set(NVFUSER_C_STANDARD 20 CACHE STRING "C standard to use for C code")
set(NVFUSER_CPP_STANDARD 20 CACHE STRING "C++ standard to use for C++ code")
set(NVFUSER_CUDA_STANDARD 17 CACHE STRING "C++ standard to use for CUDA code")

if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
  # TODO: gcc 11.4 has been end of life according to https://gcc.gnu.org/
  # I believe we should bump up the version below to 12.x.
  # However, because gcc 11.4 is well tested and stable, let's defer this
  # rejection until the day that we find a bug in gcc 11.4.
  if(CMAKE_CXX_COMPILER_VERSION VERSION_LESS 11.4)
    message(FATAL_ERROR "GCC < 11.4 has compiler bugs and can not compile nvFuser.")
  endif()
endif()

string(APPEND CMAKE_CXX_FLAGS " -Wno-psabi")

find_package(Torch REQUIRED)
find_package(Python REQUIRED Development.Module Interpreter)
find_package(pybind11 REQUIRED)

# need this since the pytorch execution uses a different name
set(PYTHON_EXECUTABLE ${Python_EXECUTABLE})

# CXX flags is necessary since https://github.com/pytorch/pytorch/issues/98093
string(APPEND CMAKE_CXX_FLAGS " ${TORCH_CXX_FLAGS}")
include(cmake/FlatBuffers.cmake)
include(cmake/Dependencies.cmake)

# set CUDA_ARCH for cu tests.
if(TORCH_CUDA_ARCH_LIST)
  set(ARCH_FLAGS)
  cuda_select_nvcc_arch_flags(ARCH_FLAGS ${TORCH_CUDA_ARCH_LIST})
  list(APPEND CUDA_NVCC_FLAGS ${ARCH_FLAGS})
endif()

add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/lib/dynamic_type)

# TODO: fix MSVC
if(NOT MSVC)
  find_library(LIBCUPTI libcupti.so PATHS ${CUDA_TOOLKIT_ROOT_DIR}/extras/CUPTI/lib64/ ${CUDA_TOOLKIT_ROOT_DIR}/lib64/)
endif()

# ------------------------------
# build nvfuser_codegen library
# ------------------------------

# nvfuser codegen sources
set(NVFUSER_SRCS)
list(APPEND NVFUSER_SRCS
  ${NVFUSER_SRCS_DIR}/alias_analysis.cpp
  ${NVFUSER_SRCS_DIR}/codegen.cpp
  ${NVFUSER_SRCS_DIR}/compute_at.cpp
  ${NVFUSER_SRCS_DIR}/compute_at_map.cpp
  ${NVFUSER_SRCS_DIR}/contiguity.cpp
  ${NVFUSER_SRCS_DIR}/debug.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/bank_conflict.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/circular_buffer.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/device_version.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/divisible_split.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/fused_reduction.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/index_compute.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/predicate_elimination.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/sync_information.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/tensor_memory.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/tensor_producer_aliases.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/thread_predicate.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/tma.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/analysis/trivial_broadcast.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/lower2device.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/alias_memory.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/allocation.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/circular_buffer.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/expr_sort.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/fusion_simplifier.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/grid_serialization.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/index.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/inline_ptx.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/insert_syncs.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/instrument.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/loop_rotation.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/loops.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/magic_zero.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/predicate.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/replace_size.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/rng.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/scalar_hoist.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/unroll.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/vectorize_welford.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/pass/warp_reduce.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/utils.cpp
  ${NVFUSER_SRCS_DIR}/device_lower/validation.cpp
  ${NVFUSER_SRCS_DIR}/dispatch.cpp
  ${NVFUSER_SRCS_DIR}/driver_api.cpp
  ${NVFUSER_SRCS_DIR}/dynamic_transform.cpp
  ${NVFUSER_SRCS_DIR}/evaluator_common.cpp
  ${NVFUSER_SRCS_DIR}/exceptions.cpp
  ${NVFUSER_SRCS_DIR}/expr_evaluator.cpp
  ${NVFUSER_SRCS_DIR}/expr_simplifier.cpp
  ${NVFUSER_SRCS_DIR}/fusion.cpp
  ${NVFUSER_SRCS_DIR}/fusion_guard.cpp
  ${NVFUSER_SRCS_DIR}/fusion_segmenter.cpp
  ${NVFUSER_SRCS_DIR}/global_allocator.cpp
  ${NVFUSER_SRCS_DIR}/grouped_reduction.cpp
  ${NVFUSER_SRCS_DIR}/host_ir/container.cpp
  ${NVFUSER_SRCS_DIR}/host_ir/executor.cpp
  ${NVFUSER_SRCS_DIR}/host_ir/host_ir.cpp
  ${NVFUSER_SRCS_DIR}/host_ir/lower.cpp
  ${NVFUSER_SRCS_DIR}/id_model/circular_buffer_indexing.cpp
  ${NVFUSER_SRCS_DIR}/id_model/contiguity.cpp
  ${NVFUSER_SRCS_DIR}/id_model/id_model.cpp
  ${NVFUSER_SRCS_DIR}/id_model/id_model_index_compute.cpp
  ${NVFUSER_SRCS_DIR}/id_model/indexing.cpp
  ${NVFUSER_SRCS_DIR}/id_model/indexing_traversal.cpp
  ${NVFUSER_SRCS_DIR}/id_model/loop_promotion.cpp
  ${NVFUSER_SRCS_DIR}/id_model/predicate_indexing.cpp
  ${NVFUSER_SRCS_DIR}/id_model/schedule.cpp
  ${NVFUSER_SRCS_DIR}/id_model/to_string.cpp
  ${NVFUSER_SRCS_DIR}/id_model/transform_replay.cpp
  ${NVFUSER_SRCS_DIR}/id_model/validation_utils.cpp
  ${NVFUSER_SRCS_DIR}/index_compute.cpp
  ${NVFUSER_SRCS_DIR}/instrumentation.cpp
  ${NVFUSER_SRCS_DIR}/interval_analysis.cpp
  ${NVFUSER_SRCS_DIR}/ir/base_nodes.cpp
  ${NVFUSER_SRCS_DIR}/ir/builder.cpp
  ${NVFUSER_SRCS_DIR}/ir/cloner.cpp
  ${NVFUSER_SRCS_DIR}/ir/container.cpp
  ${NVFUSER_SRCS_DIR}/ir/graphviz.cpp
  ${NVFUSER_SRCS_DIR}/ir/iostream.cpp
  ${NVFUSER_SRCS_DIR}/ir/nodes.cpp
  ${NVFUSER_SRCS_DIR}/ir/utils.cpp
  ${NVFUSER_SRCS_DIR}/iter_visitor.cpp
  ${NVFUSER_SRCS_DIR}/kernel.cpp
  ${NVFUSER_SRCS_DIR}/kernel_db/kernel_db.cpp
  ${NVFUSER_SRCS_DIR}/kernel_db/utils.cpp
  ${NVFUSER_SRCS_DIR}/kernel_ir.cpp
  ${NVFUSER_SRCS_DIR}/kernel_ir_dispatch.cpp
  ${NVFUSER_SRCS_DIR}/logical_domain_map.cpp
  ${NVFUSER_SRCS_DIR}/mma_type.cpp
  ${NVFUSER_SRCS_DIR}/multidevice/communication.cpp
  ${NVFUSER_SRCS_DIR}/multidevice/communicator.cpp
  ${NVFUSER_SRCS_DIR}/multidevice/cuda_p2p.cpp
  ${NVFUSER_SRCS_DIR}/multidevice/ipc_handle.cpp
  ${NVFUSER_SRCS_DIR}/multidevice/device_mesh.cpp
  ${NVFUSER_SRCS_DIR}/multidevice/executor.cpp
  ${NVFUSER_SRCS_DIR}/multidevice/utils.cpp
  ${NVFUSER_SRCS_DIR}/mutator.cpp
  ${NVFUSER_SRCS_DIR}/non_divisible_split.cpp
  ${NVFUSER_SRCS_DIR}/ops/alias.cpp
  ${NVFUSER_SRCS_DIR}/ops/arith.cpp
  ${NVFUSER_SRCS_DIR}/ops/composite.cpp
  ${NVFUSER_SRCS_DIR}/ops/indexing.cpp
  ${NVFUSER_SRCS_DIR}/ops/normalization.cpp
  ${NVFUSER_SRCS_DIR}/ops/utils.cpp
  ${NVFUSER_SRCS_DIR}/options.cpp
  ${NVFUSER_SRCS_DIR}/parallel_dimension_map.cpp
  ${NVFUSER_SRCS_DIR}/parallel_type_bitmap.cpp
  ${NVFUSER_SRCS_DIR}/polymorphic_value.cpp
  ${NVFUSER_SRCS_DIR}/predicate_compute.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/add_axioms.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/allocation_order_inference.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/consecutive_cast.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/exact_mapped_extent_substitution.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/insert_reshardings.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/make_resharding_contiguous.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/mark_aliases_prepare.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/move_pad.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/move_repeat_forward.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/move_split_cat.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/pre_segmenter.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/propagate_shardings.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/remove_bcast_squeeze.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/remove_empty.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/reorder_sharded_axis.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/segment_inplace_update.cpp
  ${NVFUSER_SRCS_DIR}/host_ir/pass/stream_parallel_type.cpp
  ${NVFUSER_SRCS_DIR}/host_ir/pass/insert_deallocations.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/translate_no_reduction_matmul_to_mul_squeeze.cpp
  ${NVFUSER_SRCS_DIR}/preseg_passes/translate_repeat_to_expand.cpp
  ${NVFUSER_SRCS_DIR}/rng.cpp
  ${NVFUSER_SRCS_DIR}/runtime/allocations.cpp
  ${NVFUSER_SRCS_DIR}/runtime/compiled_kernel.cpp
  ${NVFUSER_SRCS_DIR}/runtime/executor.cpp
  ${NVFUSER_SRCS_DIR}/runtime/executor_dispatch.cpp
  ${NVFUSER_SRCS_DIR}/runtime/executor_kernel_arg.cpp
  ${NVFUSER_SRCS_DIR}/runtime/executor_params.cpp
  ${NVFUSER_SRCS_DIR}/runtime/executor_utils.cpp
  ${NVFUSER_SRCS_DIR}/runtime/fusion_cache_utils.cpp
  ${NVFUSER_SRCS_DIR}/runtime/fusion_executor_cache.cpp
  ${NVFUSER_SRCS_DIR}/runtime/fusion_kernel_runtime.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/cache_policy_refiner.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/heuristic.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/mark_aliases.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/matmul.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/multi_matmul.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/ampere_multi_matmul.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/hopper_multi_matmul.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/matmul_heuristic_plugin.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/matmul_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/mma_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/no_op.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/communication.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/normalization_inner.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/normalization_inner_outer.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/normalization_inner_outer_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/normalization_inner_outer_tma_ws.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/normalization_inner_outer_multi_wave.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/normalization_outer.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/normalization_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/pointwise.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/pointwise_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/reduction.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/reduction_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/registry.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/registry_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/resize.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/runtime_info.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/scheduler_types.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/tools/domain_map.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/tools/inlining.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/tools/loop_domain_scheduler.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/tools/maxinfo_propagator.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/tools/resize_utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/tools/static_repeat.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/transpose.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/utils.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/vectorize_helper.cpp
  ${NVFUSER_SRCS_DIR}/scheduler/expr_eval_sched.cpp
  ${NVFUSER_SRCS_DIR}/serde/polymorphic_value.cpp
  ${NVFUSER_SRCS_DIR}/serde/utils.cpp
  ${NVFUSER_SRCS_DIR}/statement_guard.cpp
  ${NVFUSER_SRCS_DIR}/swizzle.cpp
  ${NVFUSER_SRCS_DIR}/sys_utils.cpp
  ${NVFUSER_SRCS_DIR}/tensor_metadata.cpp
  ${NVFUSER_SRCS_DIR}/tensor_view.cpp
  ${NVFUSER_SRCS_DIR}/tma.cpp
  ${NVFUSER_SRCS_DIR}/transform_iter.cpp
  ${NVFUSER_SRCS_DIR}/transform_replay.cpp
  ${NVFUSER_SRCS_DIR}/transform_rfactor.cpp
  ${NVFUSER_SRCS_DIR}/transform_view.cpp
  ${NVFUSER_SRCS_DIR}/type.cpp
  ${NVFUSER_SRCS_DIR}/type_promotion.cpp
  ${NVFUSER_SRCS_DIR}/utils.cpp
  ${NVFUSER_SRCS_DIR}/val_graph.cpp
  ${NVFUSER_SRCS_DIR}/val_graph_visitor.cpp
  ${NVFUSER_SRCS_DIR}/validator_utils.cpp
)

# We don't link CUPTI for MSVC
if(NOT MSVC)
  list(APPEND NVFUSER_SRCS
    ${NVFUSER_SRCS_DIR}/fusion_profiler.cpp
  )
endif()

if(BUILD_PYTHON)
  list(APPEND NVFUSER_SRCS
    ${NVFUSER_PYTHON_BINDINGS}/distributed_tensor.cpp
    ${NVFUSER_PYTHON_BINDINGS}/fusion_cache.cpp
    ${NVFUSER_PYTHON_BINDINGS}/fusion_definition.cpp
    ${NVFUSER_PYTHON_BINDINGS}/fusion_state.cpp
    ${NVFUSER_PYTHON_BINDINGS}/segmentation.cpp
    ${NVFUSER_PYTHON_BINDINGS}/translation.cpp
    ${NVFUSER_PYTHON_BINDINGS}/translation_utils.cpp
    ${NVFUSER_SRCS_DIR}/serde/fusion_record.cpp
  )
endif()

# We create both static and shared libraries.
#
# Shared libraries are what ships, but a large advantage of static libraries is
# that symbols are all visible. This allows us to test internal components
# inside our test or benchmark binaries, even if we do not want said components
# to be visible to the outside. If we used only shared libraries, then any API
# we invoked from test binaries would need to be marked as public, even if we
# did not want to expose it to users.
#
# Note technically we create an "OBJECT" library instead of a "STATIC" library.
# This is just a CMake quirk; an OBJECT library is a better way to implement a
# "private" (not installed) static library.
add_library(codegen_internal OBJECT ${NVFUSER_SRCS})

if(NOT MSVC)
  if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
    target_compile_options(codegen_internal PRIVATE
      -Wall -Wno-unused-function -Werror

      # These warnings are not treated as errors because of gcc 12.2 used in
      # manylinux image. consider enable this when we upgrade.
      # linking comment:
      # https://github.com/NVIDIA/Fuser/pull/3001#discussion_r1772551266
      -Wno-error=restrict -Wno-error=stringop-overflow -Wno-error=maybe-uninitialized)
  else()
    target_compile_options(codegen_internal PRIVATE
      -Wall -Wno-unused-function -Werror)
  endif()
endif()

target_compile_definitions(codegen_internal PRIVATE "-DTORCH_CUDA_BUILD_MAIN_LIB")
target_include_directories(codegen_internal PUBLIC ${NVFUSER_PYTHON_DIR})
target_include_directories(codegen_internal SYSTEM PUBLIC
  ${CMAKE_SOURCE_DIR}/third_party/flatbuffers/include
  PRIVATE
  ${CUDA_TOOLKIT_ROOT_DIR}/extras/CUPTI/include
  ${CUDA_INCLUDE_DIRS}
)
target_include_directories(codegen_internal PUBLIC
  "$<BUILD_INTERFACE:${NVFUSER_SRCS_DIR}>"
  "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/nvfuser>"
)
set_target_properties(codegen_internal PROPERTIES
  C_STANDARD ${NVFUSER_C_STANDARD}
  CUDA_STANDARD ${NVFUSER_CUDA_STANDARD}
  CXX_STANDARD ${NVFUSER_CPP_STANDARD}
  CXX_STANDARD_REQUIRED ON
  CXX_VISIBILITY_PRESET hidden

  # this is to find pip installed nvrtc.so
  INSTALL_RPATH
  "$ORIGIN/../../nvidia/cuda_runtime/lib:$ORIGIN/../../nvidia/cuda_nvrtc/lib:$ORIGIN/../../nvidia/cuda_cupti/lib:$ORIGIN/../../torch/lib"
  POSITION_INDEPENDENT_CODE Yes
  VISIBILITY_INLINES_HIDDEN Yes
)

# Ensure we don't link against libcuda; we'll dlopen it ourselves.
list(FILTER TORCH_LIBRARIES EXCLUDE REGEX "libcuda\.so")
target_link_libraries(codegen_internal PUBLIC
  dynamic_type
  ${LIBCUPTI}
  ${TORCH_LIBRARIES}
  dl
)

add_library(nvfuser_codegen SHARED $<TARGET_OBJECTS:codegen_internal>)

if(NVFUSER_BUILD_WITH_ASAN)
  target_compile_options(codegen_internal PRIVATE -fsanitize=address)
  target_link_options(codegen_internal PUBLIC -fsanitize=address)
  target_link_options(nvfuser_codegen PUBLIC -fsanitize=address)
endif()

target_include_directories(nvfuser_codegen PUBLIC
  "$<BUILD_INTERFACE:${NVFUSER_SRCS_DIR}>"
  "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/nvfuser>"
)
target_link_libraries(nvfuser_codegen PRIVATE
  flatbuffers
  ${CUDA_NVRTC_LIB}
  ${LIBCUPTI}
  ${TORCH_LIBRARIES}
  dl
)
set_target_properties(nvfuser_codegen PROPERTIES
  C_STANDARD ${NVFUSER_C_STANDARD}
  CUDA_STANDARD ${NVFUSER_CUDA_STANDARD}
  CXX_STANDARD ${NVFUSER_CPP_STANDARD}
  CXX_STANDARD_REQUIRED ON
  CXX_VISIBILITY_PRESET hidden
  INSTALL_RPATH
  "$ORIGIN/../../nvidia/cuda_runtime/lib:$ORIGIN/../../nvidia/cuda_nvrtc/lib:$ORIGIN/../../nvidia/cuda_cupti/lib:$ORIGIN/../../torch/lib"
  POSITION_INDEPENDENT_CODE Yes
  VISIBILITY_INLINES_HIDDEN Yes
)
install(TARGETS nvfuser_codegen EXPORT NvfuserTargets DESTINATION lib)

# We are keeping fusion_cache_generated.h for the submodule build because flatc is unavailable.
add_custom_command(
  OUTPUT
  ${NVFUSER_ROOT}/csrc/serde/fusion_cache_generated.h
  DEPENDS
  ${NVFUSER_ROOT}/csrc/serde/fusion_cache.fbs
  DEPENDS flatc
  COMMAND ${CMAKE_CURRENT_BINARY_DIR}/third_party/flatbuffers/flatc --scoped-enums -o ${NVFUSER_ROOT}/csrc/serde/ -c -b ${NVFUSER_ROOT}/csrc/serde/fusion_cache.fbs
  COMMENT "Generating fusion_cache_generated header from fusion_cache.fbs"
  VERBATIM
)
add_custom_target(build_flatbuffer_config ALL
  DEPENDS ${NVFUSER_ROOT}/csrc/serde/fusion_cache_generated.h)

if(NVFUSER_STANDALONE_BUILD_WITH_UCC)
  # User may need to set env vars UCC_DIR, UCX_DIR, UCC_HOME, UCX_HOME for CMake's Find_UCC to work.
  find_package(UCC REQUIRED)
  find_package(UCX REQUIRED)

  add_library(__nvfuser_ucc INTERFACE)
  set_target_properties(__nvfuser_ucc PROPERTIES
    C_STANDARD ${NVFUSER_C_STANDARD}
    CUDA_STANDARD ${NVFUSER_CUDA_STANDARD}
    CXX_STANDARD ${NVFUSER_CPP_STANDARD}
    CXX_STANDARD_REQUIRED ON
    CXX_VISIBILITY_PRESET hidden
    POSITION_INDEPENDENT_CODE Yes
    VISIBILITY_INLINES_HIDDEN Yes
  )
  target_link_libraries(__nvfuser_ucc INTERFACE ucx::ucs ucx::ucp ucc::ucc)
  target_include_directories(__nvfuser_ucc INTERFACE ${UCC_INCLUDE_DIRS})
  target_link_libraries(codegen_internal PRIVATE __nvfuser_ucc)
  target_compile_definitions(codegen_internal PRIVATE NVFUSER_BUILD_WITH_UCC)
endif()

add_dependencies(codegen_internal flatc build_flatbuffer_config)

# installing nvfuser headers
install(DIRECTORY "${NVFUSER_SRCS_DIR}/"
  DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/nvfuser"
  FILES_MATCHING
  PATTERN "*.h"
  PATTERN "csrc/C++20/type_traits"
  PATTERN "csrc/struct.inl")

# TODO guard including flatbuffers headers
# installing flatbuffers headers
install(DIRECTORY "${NVFUSER_THIRD_PARTY_DIR}/flatbuffers/include/flatbuffers/"
  DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/nvfuser/flatbuffers")

# installing dynamic_type headers
install(DIRECTORY "${NVFUSER_ROOT}/lib/dynamic_type/src/dynamic_type"
  DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/nvfuser")

# -----------------------------
# build nvfuser python library
# -----------------------------
if(BUILD_PYTHON)
  # nvfuser python API sources
  set(NVFUSER_PYTHON_SRCS)
  list(APPEND NVFUSER_PYTHON_SRCS
    ${NVFUSER_PYTHON_BINDINGS}/multidevice_bindings.cpp
    ${NVFUSER_PYTHON_BINDINGS}/python_bindings.cpp
    ${NVFUSER_PYTHON_BINDINGS}/python_bindings_extension.cpp
    ${NVFUSER_PYTHON_BINDINGS}/schedule_bindings.cpp
  )

rdspring1 · 2025-04-26T16:18:33Z

@xwang233 For jit_examples_sinh_libtorch_20, the test should be changed:

Current:

import nvfuser.utils; import torch.utils; print(nvfuser.utils.cmake_prefix_path, torch.utils.cmake_prefix_path, sep=";")

New:

import nvfuser_common.utils; import torch.utils; print(nvfuser_common.utils.cmake_prefix_path, torch.utils.cmake_prefix_path, sep=";")

tests/python/test_python_frontend.py

xwang233 · 2025-04-29T17:31:49Z

I updated the test script for jit_examples_sinh_libtorch for --dev CI branch, re: #4323 (comment)

rdspring1 · 2025-05-05T16:52:06Z

!test --dev

jjsjann123

sorry this falls off my radar earlier. LGTM

rdspring1 · 2025-05-05T23:27:19Z

!test --dev

xwang233

Thanks for working on this!

rdspring1 added build Direct Bindings Python extension with direct mapping to NvFuser CPP objects. labels Apr 26, 2025

rdspring1 changed the title ~~Create nvfuser_common package~~ Create nvfuser_common package in build system Apr 26, 2025

rdspring1 requested review from jjsjann123 and xwang233 April 26, 2025 16:20

jjsjann123 reviewed Apr 28, 2025

View reviewed changes

tests/python/test_python_frontend.py Outdated Show resolved Hide resolved

rdspring1 force-pushed the next_pt0 branch from 0348281 to 650f6de Compare April 29, 2025 16:00

rdspring1 added 4 commits April 29, 2025 12:14

create NVFUSER_PYTHON_BINDINGS

54b5ff7

Move common from python/nvfuser to python/nvfuser_common

3074e51

Update python tests

9ab0014

revert nvfuser_version

6972414

rdspring1 force-pushed the next_pt0 branch 2 times, most recently from 4aaade9 to 23b5223 Compare May 1, 2025 23:51

Merge branch 'main' of https://github.com/nvidia/fuser into next_pt0

ccad78d

rdspring1 force-pushed the next_pt0 branch from 23b5223 to ccad78d Compare May 2, 2025 00:00

rdspring1 added 2 commits May 5, 2025 09:29

Merge branch 'main' of https://github.com/nvidia/fuser into next_pt0

7638eb7

update gen_nvfuser_version

0ff76fb

jjsjann123 approved these changes May 5, 2025

View reviewed changes

Merge branch 'main' into next_pt0

a59964c

xwang233 approved these changes May 6, 2025

View reviewed changes

xwang233 merged commit 29d89f0 into main May 6, 2025
53 checks passed

xwang233 deleted the next_pt0 branch May 6, 2025 18:12

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Create nvfuser_common package in build system#4323

Create nvfuser_common package in build system#4323
xwang233 merged 8 commits intomainfrom
next_pt0

rdspring1 commented Apr 26, 2025 •

edited

Loading

Uh oh!

github-actions bot commented Apr 26, 2025 •

edited

Loading

Uh oh!

rdspring1 commented Apr 26, 2025 •

edited

Loading

Uh oh!

Uh oh!

xwang233 commented Apr 29, 2025

Uh oh!

rdspring1 commented May 5, 2025

Uh oh!

jjsjann123 left a comment

Uh oh!

rdspring1 commented May 5, 2025

Uh oh!

xwang233 left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

rdspring1 commented Apr 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

CI updates for jit_examples_sinh_libtorch_20

Uh oh!

github-actions bot commented Apr 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Changes walkthrough 📝

PR Reviewer Guide 🔍

Uh oh!

rdspring1 commented Apr 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Current:

New:

Uh oh!

Uh oh!

xwang233 commented Apr 29, 2025

Uh oh!

rdspring1 commented May 5, 2025

Uh oh!

jjsjann123 left a comment

Choose a reason for hiding this comment

Uh oh!

rdspring1 commented May 5, 2025

Uh oh!

xwang233 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

rdspring1 commented Apr 26, 2025 •

edited

Loading

CI updates for `jit_examples_sinh_libtorch_20`

github-actions bot commented Apr 26, 2025 •

edited

Loading

rdspring1 commented Apr 26, 2025 •

edited

Loading