add code for test resnet

test resnet
2022-10-21 14:46:42 +08:00 · 2022-10-20 16:36:55 +08:00
569 changed files with 2958 additions and 38695 deletions
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@ -1,81 +0,0 @@
-name: Build and test cpu
-on:
-  push:
-    paths-ignore:
-      - '**.md'
-      - 'LICENSE'
-  pull_request:
-    paths:
-      - '**.md'
-      - 'LICENSE'
-
-env:
-  protobuf-download: https://github.com/protocolbuffers/protobuf/releases/download/v21.12/protobuf-cpp-3.21.12.tar.gz
-  protobuf-version: "3.21.12"
-  python-version: "3.10"
-
-  resnet-download: https://github.com/InfiniTensor/InfiniTensor/releases/download/test-models/resnet18-v2-7.onnx
-  inception-download: https://github.com/InfiniTensor/InfiniTensor/releases/download/test-models/inception-v2-9.onnx
-  densenet-download: https://github.com/InfiniTensor/InfiniTensor/releases/download/test-models/densenet-12.onnx
-  efficientnet-download: https://github.com/InfiniTensor/InfiniTensor/releases/download/test-models/efficientnet-lite4-11.onnx
-
-jobs:
-  build:
-    name: Build
-    runs-on: ubuntu-latest
-    steps:
-    - uses: actions/checkout@v3
-      with:
-        submodules: recursive
-
-    - name: Set up Python ${{ env.python-version }}
-      uses: actions/setup-python@v4
-      with:
-        python-version: ${{ env.python-version }}
-
-    - name: Install libdw
-      run: sudo apt-get update && sudo apt-get install libdw-dev
-
-    # - name: Cache protobuf
-    #   id: cache-protobuf
-    #   uses: actions/cache@v3
-    #   with:
-    #     path: protobuf-${{ env.protobuf-version }}
-    #     key: protobuf-${{ env.protobuf-version }}
-
-    # - name: Download and compile protobuf
-    #   if: steps.cache-protobuf.outputs.cache-hit != 'true'
-    #   run: |
-    #     wget ${{ env.protobuf-download }}
-    #     tar xf protobuf-cpp-${{ env.protobuf-version }}.tar.gz
-    #     cd protobuf-${{ env.protobuf-version }}
-    #     ./autogen.sh
-    #     ./configure CFLAGS="-fPIC" CXXFLAGS="-fPIC"
-    #     make -j8
-
-    # - name: Install protobuf
-    #   run: |
-    #     cd protobuf-${{ env.protobuf-version }}
-    #     sudo make install
-    #     sudo ldconfig
-
-    - name: Build
-      run: make
-
-    - name: Test cpu
-      run: make test-cpp
-
-    - name: Install python-frontend
-      run: |
-        python -m pip install --upgrade pip
-        make install-python
-
-    - name: Download test models
-      run: |
-        wget ${{ env.resnet-download }}
-        wget ${{ env.inception-download }}
-        wget ${{ env.densenet-download }}
-        wget ${{ env.efficientnet-download }}
-
-    - name: Test onnx frontend
-      run: make test-onnx
--- a/.github/workflows/clang-format-check.yml
+++ b/.github/workflows/clang-format-check.yml
@ -1,14 +1,5 @@
 name: clang-format Check
-on:
-  push:
-    paths-ignore:
-      - '**.md'
-      - 'LICENSE'
-  pull_request:
-    paths:
-      - '**.md'
-      - 'LICENSE'
-
+on: [pull_request]
 jobs:
  formatting-check:
    name: Formatting Check
@ -20,7 +11,7 @@ jobs:
          - 'src'
          - 'test'
    steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v2
    - name: Run clang-format style check for C/C++/Protobuf programs.
      uses: jidicula/clang-format-action@v4.8.0
      with:
--- a/.gitignore
+++ b/.gitignore
@ -37,10 +37,4 @@ build_debug/
 .vscode/

 # python
-*.egg-info
-*.pyc
-
-# onnx model
-*.onnx
-*.pb
-*.npy
+*.pyc
--- a/.gitmodules
+++ b/.gitmodules
@ -10,9 +10,3 @@
 [submodule "3rd-party/backward-cpp"]
 	path = 3rd-party/backward-cpp
 	url = git@github.com:bombela/backward-cpp.git
-[submodule "example"]
-	path = examples/NNmodel
-	url = git@github.com:wanghailu0717/NNmodel.git
-[submodule "examples/distributed/onnxsim_large_model"]
-	path = examples/distributed/onnxsim_large_model
-	url = git@github.com:luchangli03/onnxsim_large_model.git
--- a/3rd-party/backward-cpp
+++ b/3rd-party/backward-cpp
@ -1 +1 @@
-Subproject commit 3bb9240cb15459768adb3e7d963a20e1523a6294
+Subproject commit f30744bcf726ea3735df7ecf9e9de9ddac540283
--- a/3rd-party/googletest
+++ b/3rd-party/googletest
@ -1 +1 @@
-Subproject commit b796f7d44681514f58a683a3a71ff17c94edb0c1
+Subproject commit e2239ee6043f73722e7aa812a459f54a28552929
--- a/3rd-party/nlohmann_json_cmake_fetchcontent
+++ b/3rd-party/nlohmann_json_cmake_fetchcontent
@ -1 +1 @@
-Subproject commit 13132dd361c8c5b5753983d5186cf54f689d90f9
+Subproject commit 6aebf09233951e4ce30a63919186a70b2b195756
--- a/3rd-party/pybind11
+++ b/3rd-party/pybind11
@ -1 +1 @@
-Subproject commit 0bd8896a4010f2d91b2340570c24fa08606ec406
+Subproject commit 1e3400b6742288429f2069aaf5febf92d0662dae
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@ -1,13 +0,0 @@
-# Changelog
-
-All notable changes to this project will be documented in this file.
-
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
-## Unreleased
-
-### Added
-
-### Modified
-
-### Fixed
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -1,65 +1,31 @@
-# Do not change these options in this file. Use cmake.config, cmake -DOPTION=VALUE, or ccmake to specify them.
-option(USE_CUDA "Support CUDA GPU" OFF)
-option(USE_BANG "Support BANG MLU" OFF)
-option(USE_KUNLUN "Support KUNLUN XPU" OFF)
-option(USE_INTELCPU "Support INTELCPU" OFF)
-option(USE_BACKTRACE "Print backtrace on exception and segmentation fault" ON)
-option(USE_PROTOBUF "Serialize and deserialize tensors" OFF)
-option(BUILD_NNET "Build nnet" OFF)
-option(BUILD_DIST "Build project for distributed running" OFF)
-option(BUILD_TEST "Build tests" OFF)
-
-if(USE_CUDA)
-    message("CMake 3.18 or higher is required for setting CUDAToolkit")
-    cmake_minimum_required(VERSION 3.18) # FindCUDAToolkit
-else()
-    cmake_minimum_required(VERSION 3.17)
-endif()
-
+cmake_minimum_required(VERSION 3.17) # FindCUDAToolkit
 include(CMakeDependentOption)
 project(InfiniTensor C CXX)

+# Do not change these options in this file. Use cmake.config, cmake -DOPTION=VALUE, or ccmake to specify them.
+option(USE_CUDA "Support CUDA GPU" ON)
+option(USE_BANG "Support BANG MLU" OFF)
+option(USE_BACKTRACE "Print backtrace on exception and segmentation fault" ON)
+option(USE_PROTOBUF "Serialize and deserialize tensors" ON)
+option(BUILD_TEST "Build tests" ON)
+
 cmake_dependent_option(BUILD_TEST_CORE "Build tests for core components" ON BUILD_TEST OFF)
 cmake_dependent_option(BUILD_TEST_PET "Build tests for PET" OFF BUILD_TEST OFF)
+cmake_dependent_option(BUILD_TEST_EINNET "Build tests for EINNET" OFF BUILD_TEST OFF)

 set(DEFAULT_BUILD_TYPE "RelWithDebInfo")
-# Build Type
-if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-    message("Configuring for Debug build.")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O0")
-    add_compile_definitions(DEBUG_MODE)
-elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
-    message("Configuring for Release build.")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O2")
-    add_compile_definitions(NDEBUG)
-elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo")
-    message("Configuring for RelWithDebInfo build.")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O2")
-else()
-    message("Build type not specified. Configuring for RelWithDebInfo build.")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O2")
-endif()
-
-
-if(EXISTS ${CMAKE_CURRENT_BINARY_DIR}/config.cmake)
-  message(STATUS "Using config.cmake in CMAKE_CURRENT_BINARY_DIR directory")
-  include(${CMAKE_CURRENT_BINARY_DIR}/config.cmake)
-else()
-  if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/config.cmake)
-    message(STATUS "Using config.cmake in CMAKE_CURRENT_SOURCE_DIR directory")
-    include(${CMAKE_CURRENT_SOURCE_DIR}/config.cmake)
-  endif()
-endif()

 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_EXTENSIONS OFF) # -std=gnu++11 when on, -std=c++11 when off
-add_compile_options(-Wno-error=unused-variable)
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -Wall -Werror -Wno-error=deprecated-declarations")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -UNDEBUG") # Enable assertion
+set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -UNDEBUG") # Enable assertion

 find_package(
  Python
  COMPONENTS Interpreter Development
  REQUIRED)
-
 # OpenMP
 find_package(OpenMP)
 if(OpenMP_C_FOUND)
@ -68,7 +34,6 @@ endif()
 if(OpenMP_CXX_FOUND)
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
 endif()
-
 #Protobuf
 if(USE_PROTOBUF)
  add_definitions(-D TENSOR_PROTOBUF)
@ -81,12 +46,14 @@ if(USE_PROTOBUF)
  set(PROTO_PATH "${CMAKE_CURRENT_SOURCE_DIR}/proto")
  file(GLOB PROTO_FILES "${PROTO_PATH}/data.proto")
  protobuf_generate_cpp(PROTO_SRCS PROTO_HDRS ${PROTO_FILES})
-  set_source_files_properties (${PROTO_SRCS} PROPERTIES COMPILE_FLAGS -Wno-unused-variable)
+  message(${PROTO_SRCS} "-----------" ${PROTO_FILES})
+  message(${PROTO_HDRS} "-----------" ${PROTO_FILES})
  add_library(tensor_proto SHARED ${PROTO_SRCS} ${PROTO_HDRS})
  target_link_libraries(tensor_proto PUBLIC ${PROTOBUF_LIBRARIES})
 endif()

 include_directories(include)
+
 # Pybind11
 add_subdirectory(3rd-party/pybind11)
 include_directories(3rd-party/pybind11/include)
@ -95,20 +62,6 @@ include_directories(3rd-party/pybind11/include)
 add_subdirectory(3rd-party/nlohmann_json_cmake_fetchcontent)
 include_directories(3rd-party/nlohmann_json_cmake_fetchcontent/single_include)

-# TVM backend
-if(BUILD_NNET AND BUILD_TEST)
-  # TVM and DMLC for invoking TVM packed functions
-  include_directories(${TVM_INCLUDE_DIR})
-  include_directories(${DMLC_INCLUDE_DIR})
-  include_directories(${DLPACK_INCLUDE_DIR})
-  if (TVM_INCLUDE_DIR AND DMLC_INCLUDE_DIR AND DLPACK_INCLUDE_DIR AND DLPACK_INCLUDE_DIR)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DDMLC_USE_LOGGING_LIBRARY=\\\<${TVM_INCLUDE_DIR}/tvm/runtime/logging.h\\\> ")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DINFINI_USE_TVM=1") # Enable TVM codegen kernels
-  else()
-    # message(FATAL_ERROR "TVM_INCLUDE_DIR, DMLC_INCLUDE_DIR, and DLPACK_INCLUDE_DIR must be set when BUILD_NNET AND BUILD_TEST is ON")
-  endif()
-endif()
-
 if(BUILD_TEST)
  set(BUILD_GMOCK
      OFF
@ -120,21 +73,8 @@ if(BUILD_TEST)
  include_directories(3rd-party/googletest/googletest/include)
 endif()

-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -Wall -Werror -Wno-error=deprecated-declarations -Wno-error=pointer-arith")
-set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -UNDEBUG") # Enable assertion
-set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -UNDEBUG") # Enable assertion
-
-
 # Source files
-file(GLOB_RECURSE SRC src/ffi/*.cc src/core/*.cc src/kernels/cpu/*.cc src/operators/*.cc src/utils/*.cc)
-
-if(BUILD_NNET)
-  add_compile_definitions(BUILD_NNET=1)
-  file(GLOB_RECURSE SRC_NNET src/nnet/*.cc)
-  list (APPEND SRC ${SRC_NNET})
-  # For locating resource files
-  set_source_files_properties(src/nnet/test.cc PROPERTIES COMPILE_OPTIONS "-DINFINI_PROJECT_HOME=${CMAKE_CURRENT_SOURCE_DIR}")
-endif()
+file(GLOB_RECURSE SRC src/ffi/*.cc src/core/*.cc src/kernels/cpu/*.cc src/nnet/*.cc src/operators/*.cc src/utils/*.cc)

 if(USE_CUDA)
  file(GLOB_RECURSE SRC_CUDA src/cuda/*.cc src/cuda/*.cu src/kernels/cuda/*.cc src/kernels/cuda/*.cu)
@ -146,16 +86,6 @@ if(USE_BANG)
  list (APPEND SRC ${SRC_BANG})
 endif()

-if(USE_KUNLUN)
-  file(GLOB_RECURSE SRC_KUNLUN src/kunlun/*.cc src/kernels/kunlun/*.cc )
-  list (APPEND SRC ${SRC_KUNLUN})
-endif()
-
-if(USE_INTELCPU)
-  file(GLOB_RECURSE SRC_INTELCPU src/intelcpu/*.cc src/kernels/intelcpu/*.cc )
-  list (APPEND SRC ${SRC_INTELCPU})
-endif()
-
 # Libraries
 add_library(InfiniTensor SHARED ${SRC})
 if(USE_PROTOBUF)
@ -164,15 +94,10 @@ endif()

 target_link_libraries(InfiniTensor pybind11::embed)

-# TVM backend
-if(BUILD_NNET AND BUILD_TEST AND TVM_LIB_DIR)
-  target_link_libraries(InfiniTensor ${TVM_LIB_DIR}/libtvm.so)
-endif()
-
 # Python bindings
 file(GLOB_RECURSE FFIS src/ffi/ffi_infinitensor.cc)
-pybind11_add_module(backend MODULE ${FFIS})
-target_link_libraries(backend PRIVATE InfiniTensor)
+pybind11_add_module(pyinfinitensor MODULE ${FFIS})
+target_link_libraries(pyinfinitensor PRIVATE InfiniTensor)

 if(USE_BACKTRACE)
  add_definitions(-D BACKWARD_TRACE)
@ -182,31 +107,6 @@ if(USE_BACKTRACE)
  target_link_libraries(InfiniTensor dw)
 endif()

-if(USE_INTELCPU)
-  add_compile_definitions(USE_INTELCPU=1)
-  find_package(MKL CONFIG REQUIRED)
-
-  # Refer to https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-link-line-advisor.html
-  target_link_libraries(InfiniTensor sycl OpenCL)
-
-  set(DNNL_CONFIGURATION "cpu_gomp")
-  find_package(dnnl CONFIG REQUIRED)
-  if(dnnl_FOUND)
-      add_compile_definitions(USE_MKL=1)
-      include_directories(BEFORE ${dnnl_DIR}/../../../cpu_gomp/include/)
-      link_directories(${dnnl_DIR}/../../../cpu_gomp/lib)
-      target_link_libraries(InfiniTensor   dnnl)
-  else()
-      message(FATAL_ERROR "dnnl library not found")
-  endif()
-  set(WNO_ERRORS "-Wno-error=unused-parameter -Wno-error=unused-function -Wno-error=unused-private-field -Wno-error=ignored-attributes -Wno-error=unused-const-variable -Wno-error=inconsistent-missing-override -Wno-error=unused-variable -Wno-error=tautological-constant-compare")
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMKL_ILP64 -qmkl=parallel -Werror  ${WNO_ERRORS}")
-  set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -DMKL_ILP64 -qmkl=parallel ${WNO_ERRORS}") # Enable assertion
-  set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -DMKL_ILP64 -qmkl=parallel ${WNO_ERRORS}") # Enable assertion
-
-  find_package(IntelDPCPP REQUIRED)
-endif()
-
 if(USE_CUDA)
  add_compile_definitions(USE_CUDA=1)
  # Since enable_language only executes once, rerun cmake is required if CMAKE_CUDA_HOST_COMPILER is wrong
@ -218,17 +118,9 @@ if(USE_CUDA)
  enable_language(CUDA)
  find_package(CUDAToolkit) # For nvrtc and cuda driver
  target_link_libraries(InfiniTensor cudnn CUDA::curand CUDA::cublas CUDA::nvrtc CUDA::cudart CUDA::cuda_driver)
-  if (BUILD_DIST)
-    message(STATUS "Add BUILD_DIST, use NCCL with CUDA")
-    list(APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake)
-    find_package(NCCL REQUIRED)
-    add_compile_definitions(INFINI_USE_NCCL=1)
-    target_link_libraries(InfiniTensor nccl)
-  endif()
 endif()

 if(USE_BANG)
-  add_compile_definitions(USE_BANG=1)
  include_directories(src/kernels/mlu/include)
  ################################################################################
  # Neuware Evironment
@ -242,13 +134,13 @@ if(USE_BANG)
    set(NEUWARE_HOME $ENV{NEUWARE_HOME} CACHE STRING "NEUWARE_HOME directory for Cambricon Neuware development")
  endif()
  message(STATUS "NEUWARE_HOME: ${NEUWARE_HOME}")
-
+  
  include_directories("${NEUWARE_HOME}/include")
  find_library(CAMBRICON_CNNL libcnnl.so "${NEUWARE_HOME}/lib64")
  find_library(CAMBRICON_CNRT libcnrt.so "${NEUWARE_HOME}/lib64")
  find_library(CAMBRICON_CNDRV libcndrv.so "${NEUWARE_HOME}/lib64")
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -lstdc++ -Wall -Werror")
-
+    
  if ((NOT DEFINED TARGET_CPU_ARCH) AND (NOT DEFINED ENV{TARGET_CPU_ARCH}))
    execute_process(COMMAND uname -m OUTPUT_VARIABLE _uname_m OUTPUT_STRIP_TRAILING_WHITESPACE)
    set(TARGET_CPU_ARCH "${_uname_m}" CACHE STRING "Target CPU ARCH")
@ -262,51 +154,10 @@ if(USE_BANG)
  ################################################################################
  # BangC Kernels
  ################################################################################
+  add_subdirectory(src/kernels/mlu)

-  if (BUILD_DIST)
-    find_library(CAMBRICON_CNCL libcncl.so "${NEUWARE_HOME}/lib64")
-    target_link_libraries(InfiniTensor ${CAMBRICON_CNCL} ${CAMBRICON_CNNL} ${CAMBRICON_CNRT} ${CAMBRICON_CNDRV} stdc++)
-    message(STATUS "Add BUILD_DIST, use CNCL with BANG")
-    add_compile_definitions(INFINI_USE_CNCL=1)
-  else()
-    target_link_libraries(InfiniTensor ${CAMBRICON_CNNL} ${CAMBRICON_CNRT} ${CAMBRICON_CNDRV} stdc++)
-  endif()
-endif()
-
-if(USE_KUNLUN)
-  add_compile_definitions(USE_KUNLUN=1)
-  if ((NOT DEFINED KUNLUN_HOME) AND (NOT DEFINED ENV{KUNLUN_HOME}))
-    message(FATAL_ERROR "KUNLUN_HOME is not defined from cmake or env")
-  elseif (DEFINED KUNLUN_HOME)
-          set(KUNLUN_HOME ${KUNLUN_HOME} CACHE STRING "KUNLUN_HOME directory for Kunlun development")
-  else()
-          set(KUNLUN_HOME $ENV{KUNLUN_HOME} CACHE STRING "KUNLUN_HOME directory for Kunlun development")
-  endif()
-  message(STATUS "KUNLUN_HOME: ${KUNLUN_HOME}")
-
-  include_directories("${KUNLUN_HOME}/include/")
-  find_library(KUNLUN_RT libxpurt.so "${KUNLUN_HOME}/lib64/")
-  find_library(KUNLUN_DNN libxpuapi.so "${KUNLUN_HOME}/lib64/")
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -lstdc++ -Wall -Werror")
-
-  if ((NOT DEFINED TARGET_CPU_ARCH) AND (NOT DEFINED ENV{TARGET_CPU_ARCH}))
-    execute_process(COMMAND uname -m OUTPUT_VARIABLE _uname_m OUTPUT_STRIP_TRAILING_WHITESPACE)
-    set(TARGET_CPU_ARCH "${_uname_m}" CACHE STRING "Target CPU ARCH")
-  elseif(DEFINED TARGET_CPU_ARCH)
-    set(TARGET_CPU_ARCH ${TARGET_CPU_ARCH} CACHE STRING "Target CPU ARCH")
-  else()
-    set(TARGET_CPU_ARCH $ENV{TARGET_CPU_ARCH} CACHE STRING "Target CPU ARCH")
-  endif()
-  message(STATUS "TARGET_CPU_ARCH: ${TARGET_CPU_ARCH}")
-
-  if (BUILD_DIST)
-    message(STATUS "Add BUILD_DIST, use XCCL with KUNLUN XPU")
-    list(APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake)
-    find_package(XCCL REQUIRED)
-    add_compile_definitions(INFINI_USE_XCCL=1)
-    target_link_libraries(InfiniTensor ${XCCL_LIBRARIES})
-  endif()
-  target_link_libraries(InfiniTensor ${KUNLUN_RT} ${KUNLUN_DNN} stdc++)
+  target_link_libraries(InfiniTensor ${CAMBRICON_CNNL} ${CAMBRICON_CNRT} ${CAMBRICON_CNDRV} stdc++)
+  target_link_libraries(InfiniTensor bangops)
 endif()

 # # Python bindings
@ -325,7 +176,6 @@ function(build_test files)
 endfunction()

 if(BUILD_TEST)
-  add_compile_definitions(BUILD_TEST=1)
  enable_testing()
  if(USE_TRACE)
    build_test(test/trace/*.cc)
@ -333,31 +183,17 @@ if(BUILD_TEST)
  if(BUILD_TEST_CORE)
    build_test(test/core/*.cc)
    build_test(test/operators/*.cc)
-    build_test(test/kernels/nativecpu/*.cc)
    if (USE_CUDA)
      build_test(test/kernels/cuda/*.cc)
-      build_test(test/cuda/*.cc)
    endif()
    if (USE_BANG)
      build_test(test/kernels/bang/*.cc)
-      build_test(test/bang/*.cc)
-    endif()
-    if (USE_KUNLUN)
-      build_test(test/kernels/kunlun/*.cc)
-      build_test(test/kunlun/*.cc)
-    endif()
-    if (USE_INTELCPU)
-      build_test(test/kernels/intelcpu/*.cc)
    endif()
  endif()
  if(BUILD_TEST_PET)
    build_test(test/pet/*.cc)
  endif()
-  if(BUILD_NNET AND BUILD_TEST)
+  if(BUILD_TEST_EINNET)
    build_test(test/nnet/test_*.cc)
-
-    # Build expression reader
-    add_executable(nnet_reader test/nnet/readlog.cc)
-    target_link_libraries(nnet_reader InfiniTensor)
  endif()
 endif()
--- a/77
+++ b/77
@ -1,77 +0,0 @@
-.PHONY : build clean format install-python test-cpp test-onnx
-
-TYPE ?= Release
-CUDA ?= OFF
-BANG ?= OFF
-KUNLUN ?= OFF
-INTELCPU ?= off
-BACKTRACE ?= ON
-TEST ?= ON
-DIST ?= OFF
-NNET ?= OFF
-DIST ?= OFF
-FORMAT_ORIGIN ?=
-# Docker build options
-DOCKER_NAME ?= infinitensor
-DOCKER_IMAGE_NAME ?= infinitensor
-DOCKER_FILE ?= infinitensor_ubuntu_22.04.dockerfile
-DOCKER_RUN_OPTION ?=
-
-# CUDA option.
-ifeq ($(CUDA), ON)
-	DOCKER_IMAGE_NAME = infinitensor_cuda
-	DOCKER_NAME = infinitensor_cuda
-	DOCKER_FILE = infinitensor_ubuntu_22.04_CUDA.dockerfile
-	DOCKER_RUN_OPTION += --gpus all -it --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 -v `pwd`:`pwd` -w `pwd`
-endif
-
-CMAKE_OPT = -DCMAKE_BUILD_TYPE=$(TYPE)
-CMAKE_OPT += -DUSE_CUDA=$(CUDA)
-CMAKE_OPT += -DUSE_BANG=$(BANG)
-CMAKE_OPT += -DUSE_KUNLUN=$(KUNLUN)
-CMAKE_OPT += -DUSE_BACKTRACE=$(BACKTRACE)
-CMAKE_OPT += -DBUILD_TEST=$(TEST)
-CMAKE_OPT += -DBUILD_DIST=$(DIST)
-CMAKE_OPT += -DBUILD_NNET=$(NNET)
-
-ifeq ($(INTELCPU), ON)
-	CMAKE_OPT += -DUSE_INTELCPU=ON -DCMAKE_CXX_COMPILER=dpcpp
-endif
-
-build:
-	mkdir -p build/$(TYPE)
-	cd build/$(TYPE) && cmake $(CMAKE_OPT) ../.. && make -j8
-
-clean:
-	rm -rf build
-
-format:
-	@python3 scripts/format.py $(FORMAT_ORIGIN)
-
-install-python: build
-	cp build/$(TYPE)/backend*.so pyinfinitensor/src/pyinfinitensor
-	pip install -e pyinfinitensor/
-
-test-cpp:
-	@echo
-	cd build/$(TYPE) && make test
-
-test-onnx:
-	@echo
-	python3 pyinfinitensor/tests/test_onnx.py
-
-test-api:
-	@echo
-	python3 pyinfinitensor/tests/test_api.py
-
-docker-build:
-	docker build -f scripts/dockerfile/$(DOCKER_FILE) -t $(DOCKER_NAME) .
-
-docker-run:
-	docker run -t --name $(DOCKER_IMAGE_NAME) -d $(DOCKER_NAME) $(DOCKER_RUN_OPTION)
-
-docker-start:
-	docker start $(DOCKER_IMAGE_NAME)
-
-docker-exec:
-	docker exec -it $(DOCKER_IMAGE_NAME) bash
--- a/README.md
+++ b/README.md
@ -1,75 +1,17 @@
 # InfiniTensor

-[中文项目简介](/README_CN.md) | Documentation | [中文文档](/docs/INDEX.md)
-
-[![Build](https://github.com/InfiniTensor/InfiniTensor/actions/workflows/build.yml/badge.svg?branch=master)](https://github.com/InfiniTensor/InfiniTensor/actions)
-[![issue](https://img.shields.io/github/issues/InfiniTensor/InfiniTensor)](https://github.com/InfiniTensor/InfiniTensor/issues)
-![license](https://img.shields.io/github/license/InfiniTensor/InfiniTensor)
-
-InfiniTensor is a high-performance inference engine tailored for GPUs and AI accelerators. Its design focuses on effective deployment and swift academic validation.
-
-## Get started
-
-### Make Commands
-
- `make`/`make build`: Builds the project;
- `make install-python`: Builds the project then install the python frontend;
- `make test-cpp`: Builds the project then run cpp unit tests;
- `make test-onnx`: Run python unit tests;
-
---
-
-> - Sets env: `TEST=OFF` to accelerate compiling.
-> - Sets env: `CUDA=ON` to enable cuda.
-> - Sets env: `BANG=ON` to enable bang.
-
-### CMake Options
-
-There are several configurable CMake options, see the [CMakeLists.txt](/CMakeLists.txt#L5) file.
-
- If `USE_BACKTRACE` is `ON`, `libdw-dev` have to be installed. See the README of [backward-cpp](https://github.com/bombela/backward-cpp) for details.
- If `USE_PROTOBUF` is `ON`, `protobuf` have to be installed. See the README of [protobuf](https://github.com/protocolbuffers/protobuf) for details.
- If `USE_CUDA` is `ON`, `cuda` have to be installed.
-
-## Roadmap
-
- [RefactorGraph](https://github.com/InfiniTensor/RefactorGraph) is a newly designed AI framework that is set to replace the current main branch.
- [EinNet](https://github.com/InfiniTensor/InfiniTensor/tree/NNET_e2e) is going to be merged into the main branch.
- Integration of [PET](https://github.com/thu-pacman/PET), a tensor program optimizer supporting partially equivalent transformations.
- Supported hardware
-  - ✔ NVIDIA GPU
-  - ✔ Cambricon MLU
-  - ✔ Kunlunxin XPU
-  - ⬜ Ascend NPU
+## Compilation on Lotus
+``` bash
+# Enter the root of InfiniTensor
+source test/script/env_lotus.sh 
+mkdir build && cd build
+cmake -DUSE_CUDA=ON .. && make -j 12
+```

 ## Contributor Guide
-
 InfiniTensor development is based on the pull request on Github. Before requesting for merging, a PR should satisfy the following requirements
-
-1. Pass all tests.
-    1. Now CI on Github will test everything that can be tested in the ci environment, including code format. So, script `test/script/clang_format_inplace.sh` is for formatting all code.
-    2. Contributors should run `ctest` manually and copy its output to the PR. Use fenced code blocks (triple backquotes, i.e., `` ``` ``) to avoid referencing in Github. Otherwise, `#` in the output is interpreted as a Github reference. Do not directly paste the ctest output in commit messages either for the same reason.
-2. Receive at least one approval from reviewers.
+1. Pass all tests. 
+    1. Currently, CI on Github only checks code format. Script `test/script/clang_format_inplace.sh` is for formatting all code. 
+    2. Contributors should run `ctest` manually and copy its output to the PR. Use fenced code blocks (triple backquotes, i.e., `` ``` ``) to avoid referencing in Github. Otherwise, `#` in the output is interpreted as a Github reference. Do not directly paste the ctest output in commit messages either for the same reason.  
+2. Receive at least one approval from reviewers. 
 3. PR title should be concise since it is going to be the commit message in the main branch after merging and squashing.
-
-## Reference
-
-Please cite EinNet or PET in your publications if it helps your research:
-
-```plaintext
-@article{zheng2023einnet,
-  title={EINNET: Optimizing Tensor Programs with Derivation-Based Transformations},
-  author={Zheng, Liyan and Wang, Haojie and Zhai, Jidong and Hu, Muyan and Ma, Zixuan and Wang, Tuowei and Huang, Shuhong and Miao, Xupeng and Tang, Shizhi and Huang, Kezhao and Jia, Zhihao},
-  booktitle={17th USENIX Symposium on Operating Systems Design and Implementation (OSDI 23)},
-  pages={739--755},
-  year={2023}
-}
-
-@inproceedings{wang2021pet,
-  title={PET: Optimizing tensor programs with partially equivalent transformations and automated corrections},
-  author={Wang, Haojie and Zhai, Jidong and Gao, Mingyu and Ma, Zixuan and Tang, Shizhi and Zheng, Liyan and Li, Yuanzhi and Rong, Kaiyuan and Chen, Yuanyong and Jia, Zhihao},
-  booktitle={15th USENIX Symposium on Operating Systems Design and Implementation (OSDI 21)},
-  pages={37--54},
-  year={2021}
-}
-```
--- a/README_CN.md
+++ b/README_CN.md
@ -1,13 +0,0 @@
-# Infinitensor
-
-## 项目简介
-
-本项目是深度学习领域的一个编译器集合，本项目旨在缩小深度学习应用与后端硬件之间的鸿沟。本项目通过使用编译器超优化技术，对神经网络模型进行优化，从而获得更好的性能。同时，本项目与深度学习框架相互配合，为不同的硬件后端提供端倒端的编译，方便用户迁移部署。
-
-## 项目设计
-
-本项目的设计是前后端解耦合的，主要有三个模块，分别为：
-
- Runtime 模块：该模式负责对不同的加速卡后端进行包装与支持，支撑后端运行。另外提供统一的向上接口，方便上层建设。
- Compiler 模块：该模式负责对神经网络模型进行优化变换，获得更加高效的等价模型。
- Interface 模块：该模式负责给用户提供编程与交互的接口，方便用户使用本系统。
--- a/cmake/FindCNCL.cmake
+++ b/cmake/FindCNCL.cmake
@ -1,76 +0,0 @@
-SET(CNCL_LIB_SEARCH_PATHS $ENV{NEUWARE_HOME}/lib64)
-SET(CNCL_INCLUDE_SEARCH_PATHS $ENV{NEUWARE_HOME}/include)
-
-set(CNCL_INCLUDE_DIR $ENV{NEUWARE_HOME}/include)
-set(CNCL_LIB_DIR $ENV{NEUWARE_HOME}/lib64)
-set(CNCL_VERSION $ENV{CNCL_VERSION} CACHE STRING "Version of CNCL to build with")
-
-if ($ENV{CNCL_ROOT_DIR})
-  message(WARNING "CNCL_ROOT_DIR is deprecated. Please set CNCL_ROOT instead.")
-endif()
-list(APPEND CNCL_ROOT $ENV{CNCL_ROOT_DIR} ${MLU_TOOLKIT_ROOT_DIR})
-# Compatible layer for CMake <3.12. CNCL_ROOT will be accounted in for searching paths and libraries for CMake >=3.12.
-list(APPEND CMAKE_PREFIX_PATH ${CNCL_ROOT})
-
-find_path(CNCL_INCLUDE_DIRS
-  NAMES cncl.h
-  HINTS ${CNCL_INCLUDE_DIR})
-
-if (USE_STATIC_CNCL)
-  MESSAGE(STATUS "USE_STATIC_CNCL is set. Linking with static CNCL library.")
-  SET(CNCL_LIBNAME "CNCL_static")
-  if (CNCL_VERSION)  # Prefer the versioned library if a specific CNCL version is specified
-    set(CMAKE_FIND_LIBRARY_SUFFIXES ".a.${CNCL_VERSION}" ${CMAKE_FIND_LIBRARY_SUFFIXES})
-  endif()
-else()
-  SET(CNCL_LIBNAME "cncl")
-  if (CNCL_VERSION)  # Prefer the versioned library if a specific CNCL version is specified
-    set(CMAKE_FIND_LIBRARY_SUFFIXES ".so.${CNCL_VERSION}" ${CMAKE_FIND_LIBRARY_SUFFIXES})
-  endif()
-endif()
-
-find_library(CNCL_LIBRARIES
-  NAMES ${CNCL_LIBNAME}
-  HINTS ${CNCL_LIB_DIR})
-
-include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(CNCL DEFAULT_MSG CNCL_INCLUDE_DIRS CNCL_LIBRARIES)
-
-if(CNCL_FOUND)  # obtaining CNCL version and some sanity checks
-  set (CNCL_HEADER_FILE "${CNCL_INCLUDE_DIRS}/cncl.h")
-  message (STATUS "Determining CNCL version from ${CNCL_HEADER_FILE}...")
-  set (OLD_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
-  list (APPEND CMAKE_REQUIRED_INCLUDES ${CNCL_INCLUDE_DIRS})
-  include(CheckCXXSymbolExists)
-  check_cxx_symbol_exists(CNCL_VERSION_CODE CNCL.h CNCL_VERSION_DEFINED)
-
-  if (CNCL_VERSION_DEFINED)
-    set(file "${PROJECT_BINARY_DIR}/detect_cncl_version.cc")
-    file(WRITE ${file} "
-      #include <iostream>
-      #include <cncl.h>
-      int main()
-      {
-        std::cout << CNCL_MAJOR << '.' << CNCL_MINOR << '.' << CNCL_PATCH << std::endl;
-        int x;
-        CNCLGetVersion(&x);
-        return x == CNCL_VERSION_CODE;
-      }
-")
-    try_run(CNCL_VERSION_MATCHED compile_result ${PROJECT_BINARY_DIR} ${file}
-          RUN_OUTPUT_VARIABLE CNCL_VERSION_FROM_HEADER
-          CMAKE_FLAGS  "-DINCLUDE_DIRECTORIES=${CNCL_INCLUDE_DIRS}"
-          LINK_LIBRARIES ${CNCL_LIBRARIES})
-    if (NOT CNCL_VERSION_MATCHED)
-      message(FATAL_ERROR "Found CNCL header version and library version do not match! \
-(include: ${CNCL_INCLUDE_DIRS}, library: ${CNCL_LIBRARIES}) Please set CNCL_INCLUDE_DIR and CNCL_LIB_DIR manually.")
-    endif()
-    message(STATUS "CNCL version: ${CNCL_VERSION_FROM_HEADER}")
-  else()
-    # message(STATUS "CNCL version < 2.3.5-5")
-  endif ()
-  set (CMAKE_REQUIRED_INCLUDES ${OLD_CMAKE_REQUIRED_INCLUDES})
-
-  message(STATUS "Found CNCL (include: ${CNCL_INCLUDE_DIRS}, library: ${CNCL_LIBRARIES})")
-  mark_as_advanced(CNCL_ROOT_DIR CNCL_INCLUDE_DIRS CNCL_LIBRARIES)
-endif()
--- a/cmake/FindNCCL.cmake
+++ b/cmake/FindNCCL.cmake
@ -1,165 +0,0 @@
-# Copyright (c) 2021-2022, NVIDIA CORPORATION. All rights reserved.
-# 
-# From PyTorch:
-# 
-# Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
-# Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
-# Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
-# Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
-# Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
-# Copyright (c) 2011-2013 NYU                      (Clement Farabet)
-# Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
-# Copyright (c) 2006      Idiap Research Institute (Samy Bengio)
-# Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
-# 
-# From Caffe2:
-# 
-# Copyright (c) 2016-present, Facebook Inc. All rights reserved.
-# 
-# All contributions by Facebook:
-# Copyright (c) 2016 Facebook Inc.
-# 
-# All contributions by Google:
-# Copyright (c) 2015 Google Inc.
-# All rights reserved.
-# 
-# All contributions by Yangqing Jia:
-# Copyright (c) 2015 Yangqing Jia
-# All rights reserved.
-# 
-# All contributions by Kakao Brain:
-# Copyright 2019-2020 Kakao Brain
-# 
-# All contributions from Caffe:
-# Copyright(c) 2013, 2014, 2015, the respective contributors
-# All rights reserved.
-# 
-# All other contributions:
-# Copyright(c) 2015, 2016 the respective contributors
-# All rights reserved.
-# 
-# Caffe2 uses a copyright model similar to Caffe: each contributor holds
-# copyright over their contributions to Caffe2. The project versioning records
-# all such contribution and copyright details. If a contributor wants to further
-# mark their specific copyright on a particular contribution, they should
-# indicate their copyright solely in the commit message of the change when it is
-# committed.
-# 
-# All rights reserved.
-# 
-# Redistribution and use in source and binary forms, with or without
-# modification, are permitted provided that the following conditions are met:
-# 
-# 1. Redistributions of source code must retain the above copyright
-#    notice, this list of conditions and the following disclaimer.
-# 
-# 2. Redistributions in binary form must reproduce the above copyright
-#    notice, this list of conditions and the following disclaimer in the
-#    documentation and/or other materials provided with the distribution.
-# 
-# 3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America
-#    and IDIAP Research Institute nor the names of its contributors may be
-#    used to endorse or promote products derived from this software without
-#    specific prior written permission.
-# 
-# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
-# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
-# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
-# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
-# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
-# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
-# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
-# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
-# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
-# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
-# POSSIBILITY OF SUCH DAMAGE.
-# 
-# Find the nccl libraries
-#
-# The following variables are optionally searched for defaults
-#  NCCL_ROOT: Base directory where all NCCL components are foundHong Xu, 1 year ago: • Let CMake handle NCCL detection instead of ou…
-#  NCCL_INCLUDE_DIR: Directory where NCCL header is foundPieter Noordhuis, 3 years ago: • Bump gloo
-#  NCCL_LIB_DIR: Directory where NCCL library is found
-#
-# The following are set after configuration is done:
-#  NCCL_FOUND
-#  NCCL_INCLUDE_DIRS
-#  NCCL_LIBRARIES
-#
-# The path hints include CUDA_TOOLKIT_ROOT_DIR seeing as some folks
-# install NCCL in the same location as the CUDA toolkit.
-# See https://github.com/caffe2/caffe2/issues/1601
-
-set(NCCL_INCLUDE_DIR $ENV{NCCL_INCLUDE_DIR} CACHE PATH "Folder contains NVIDIA NCCL headers")
-set(NCCL_LIB_DIR $ENV{NCCL_LIB_DIR} CACHE PATH "Folder contains NVIDIA NCCL libraries")
-set(NCCL_VERSION $ENV{NCCL_VERSION} CACHE STRING "Version of NCCL to build with")
-
-if ($ENV{NCCL_ROOT_DIR})
-  message(WARNING "NCCL_ROOT_DIR is deprecated. Please set NCCL_ROOT instead.")
-endif()
-list(APPEND NCCL_ROOT $ENV{NCCL_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR})
-# Compatible layer for CMake <3.12. NCCL_ROOT will be accounted in for searching paths and libraries for CMake >=3.12.
-list(APPEND CMAKE_PREFIX_PATH ${NCCL_ROOT})
-
-find_path(NCCL_INCLUDE_DIRS
-  NAMES nccl.h
-  HINTS ${NCCL_INCLUDE_DIR})
-
-if (USE_STATIC_NCCL)
-  MESSAGE(STATUS "USE_STATIC_NCCL is set. Linking with static NCCL library.")
-  SET(NCCL_LIBNAME "nccl_static")
-  if (NCCL_VERSION)  # Prefer the versioned library if a specific NCCL version is specified
-    set(CMAKE_FIND_LIBRARY_SUFFIXES ".a.${NCCL_VERSION}" ${CMAKE_FIND_LIBRARY_SUFFIXES})
-  endif()
-else()
-  SET(NCCL_LIBNAME "nccl")
-  if (NCCL_VERSION)  # Prefer the versioned library if a specific NCCL version is specified
-    set(CMAKE_FIND_LIBRARY_SUFFIXES ".so.${NCCL_VERSION}" ${CMAKE_FIND_LIBRARY_SUFFIXES})
-  endif()
-endif()
-
-find_library(NCCL_LIBRARIES
-  NAMES ${NCCL_LIBNAME}
-  HINTS ${NCCL_LIB_DIR})
-
-include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
-
-if(NCCL_FOUND)  # obtaining NCCL version and some sanity checks
-  set (NCCL_HEADER_FILE "${NCCL_INCLUDE_DIRS}/nccl.h")
-  message (STATUS "Determining NCCL version from ${NCCL_HEADER_FILE}...")
-  set (OLD_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
-  list (APPEND CMAKE_REQUIRED_INCLUDES ${NCCL_INCLUDE_DIRS})
-  include(CheckCXXSymbolExists)
-  check_cxx_symbol_exists(NCCL_VERSION_CODE nccl.h NCCL_VERSION_DEFINED)
-
-  if (NCCL_VERSION_DEFINED)
-    set(file "${PROJECT_BINARY_DIR}/detect_nccl_version.cc")
-    file(WRITE ${file} "
-      #include <iostream>
-      #include <nccl.h>
-      int main()
-      {
-        std::cout << NCCL_MAJOR << '.' << NCCL_MINOR << '.' << NCCL_PATCH << std::endl;
-        int x;
-        ncclGetVersion(&x);
-        return x == NCCL_VERSION_CODE;
-      }
-")
-    try_run(NCCL_VERSION_MATCHED compile_result ${PROJECT_BINARY_DIR} ${file}
-          RUN_OUTPUT_VARIABLE NCCL_VERSION_FROM_HEADER
-          CMAKE_FLAGS  "-DINCLUDE_DIRECTORIES=${NCCL_INCLUDE_DIRS}"
-          LINK_LIBRARIES ${NCCL_LIBRARIES})
-    if (NOT NCCL_VERSION_MATCHED)
-      message(FATAL_ERROR "Found NCCL header version and library version do not match! \
-(include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES}) Please set NCCL_INCLUDE_DIR and NCCL_LIB_DIR manually.")
-    endif()
-    message(STATUS "NCCL version: ${NCCL_VERSION_FROM_HEADER}")
-  else()
-    # message(STATUS "NCCL version < 2.3.5-5")
-  endif ()
-  set (CMAKE_REQUIRED_INCLUDES ${OLD_CMAKE_REQUIRED_INCLUDES})
-
-  message(STATUS "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
-  mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
-endif()
--- a/cmake/FindXCCL.cmake
+++ b/cmake/FindXCCL.cmake
@ -1,27 +0,0 @@
-# Find the xccl libraries
-set(XCCL_INCLUDE_DIR $ENV{KUNLUN_HOME}/include CACHE PATH "Folder contains KUNLUN XCCL headers")
-set(XCCL_LIB_DIR $ENV{KUNLUN_HOME}  CACHE PATH "Folder contains KUNLUN XCCL libraries")
-
-list(APPEND CMAKE_PREFIX_PATH $ENV{KUNLUN_HOME})
-
-find_path(XCCL_INCLUDE_DIRS # ${XCCL_INCLUDE_DIR}
-  NAMES xpu/bkcl.h
-  HINTS XCCL_INCLUDE_DIR)
-
-find_library(XCCL_LIBRARIES # ${XCCL_LIB_DIR}
-  NAMES lib64/libbkcl.so
-  HINTS XCCL_LIB_DIR)
-
-message(STATUS "XCCL_INCLUDE_DIRS: ${XCCL_INCLUDE_DIRS}")
-message(STATUS "XCCL_LIBRARIES: ${XCCL_LIBRARIES}")
-
-include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(XCCL DEFAULT_MSG XCCL_INCLUDE_DIRS XCCL_LIBRARIES)
-
-if (XCCL_FOUND)
-  set (XCCL_HEADER_FILE "${XCCL_INCLUDE_DIRS}/xpu/bkcl.h")
-  message (STATUS "Determing XCCL version from ${XCCL_HEADER_FILE}...")
-  list (APPEND CMAKE_REQUIRED_INCLUDES ${XCCL_INCLUDE_DIRS})
-  message(STATUS "Found XCCL (include: ${XCCL_INCLUDE_DIRS}, library: ${XCCL_LIBRARIES})")
-  mark_as_advanced(XCCL_INCLUDE_DIRS XCCL_LIBRARIES)
-endif()
--- a/cmake/config_lotus_TVM.cmake
+++ b/cmake/config_lotus_TVM.cmake
@ -1,13 +0,0 @@
-set(TVM_HOME "/home/zly/Apps/tvm-v0.10.0")
-set(TVM_INCLUDE_DIR "${TVM_HOME}/include")
-set(TVM_LIB_DIR "${TVM_HOME}/build")
-set(DMLC_INCLUDE_DIR "${TVM_HOME}/3rdparty/dmlc-core/include")
-set(DLPACK_INCLUDE_DIR "${TVM_HOME}/3rdparty/dlpack/include")
-
-set(USE_CUDA ON)
-set(USE_BANG OFF)
-
-set(BUILD_TEST ON)
-set(BUILD_TEST_CORE ON)
-set(BUILD_TEST_PET OFF)
-set(BUILD_TEST_EINNET ON)
--- a/docs/INDEX.md
+++ b/docs/INDEX.md
@ -1,5 +0,0 @@
-# 项目文档
-
- [安装部署指南](INSTALL_GUIDE_CN.md)
- [硬件支持](SUPPORT_MATRIX_CN.md)
- [使用指南](USER_GUIDE_CN.md)
--- a/docs/INSTALL_GUIDE_CN.md
+++ b/docs/INSTALL_GUIDE_CN.md
@ -1,172 +0,0 @@
-# 安装部署指南
-
-## 目录
-
- [环境准备](#环境准备)
- [编译本项目](#编译本项目)
- [技术支持](#技术支持)
-
-## 环境准备
-
-目前的软硬件环境支持矩阵
-
-| Host CPU | Device        | OS            |  Support   |
-| -------- | ------------  | -----------   | ---------- |
-| X86-64   | Nvidia GPU    |  Ubuntu-22.04 |  Yes       |
-| X86-64   | Cambricon MLU |  Ubuntu-22.04 |  Yes       |
-
-推荐使用 X86-64 机器以及 Ubuntu-22.04，本文以此环境为例。
-
-1. 确认 GCC 版本为 11.3 及以上的稳定版本，如若您的机器 GCC 版本不满足此条件，请自行编译安装，下述方式二选一：
-
-   - [GCC 官方文档](https://gcc.gnu.org/onlinedocs/gcc-11.3.0/gcc/)
-
-   - [网友安装分享](https://zhuanlan.zhihu.com/p/509695395)
-
-2. 确认 CMake 版本为 3.17 及以上的稳定版本， 如若您的机器 CMake 版本不满足此条件，请自行编译安装，下述方式二选一：
-
-   - [CMake 官方文档](https://cmake.org/install/)
-
-   - [网友安装分享](https://zhuanlan.zhihu.com/p/110793004)
-
-3. 第三方加速卡软件资源安装，目前本项目已经适配了如下的第三方加速卡：
-
-   - 如您的第三方加速卡为英伟达 GPU，请参考英伟达官方文档进行：
-
-     > [驱动安装](https://www.nvidia.cn/geforce/drivers/)，
-     > [CUDA Toolkit 安装](https://developer.nvidia.com/cuda-toolkit)，
-     > [Cudnn 安装](https://developer.nvidia.com/rdp/cudnn-download)，
-     > [Cublas 安装](https://developer.nvidia.com/cublas)，
-     > 安装完成后请进行相应的环境变量配置，将可执行文件目录与库目录添加到操作系统识别的路径中，例如
-     >
-     > ```bash
-     > # 将如下内容写入到你的 bashrc 文件并 source 该文件
-     > export CUDA_HOME="/PATH/TO/YOUR/CUDA_HOME"
-     > export CUDNN_HOME="/PATH/TO/YOUR/CUDNN_HOME"
-     > export PATH="${CUDA_HOME}/bin:${PATH}"
-     > export LD_LIBRARY_PATH="${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}"
-     > # 如您不方便将上述环境变量配置到 bashrc 文件中进行长期使用，你也可以在我们提供的 env.sh 文件中进行正确配置并激活，作为临时使用
-     > source env.sh
-     > ```
-
-     我们强烈建议您规范安装，统一到一个目录下，以免不必要的麻烦。
-
-   - 如您的第三方加速卡为寒武纪 MLU，请参考寒武纪官方文档进行：
-     > [驱动安装](https://www.cambricon.com/docs/sdk_1.11.0/driver_5.10.6/user_guide_5.10.6/index.html)，
-     > [CNToolkit 安装](https://www.cambricon.com/docs/sdk_1.11.0/cntoolkit_3.4.1/cntoolkit_install_3.4.1/index.html)，
-     > [CNNL 安装](https://www.cambricon.com/docs/sdk_1.11.0/cambricon_cnnl_1.16.1/user_guide/index.html)，
-     > 安装完成后请进行相应的环境变量配置，将可执行文件目录与库目录添加到操作系统识别的路径中，例如
-     >
-     > ```bash
-     > # 将如下内容写入到你的 bashrc 文件并 source 该文件
-     > export NEUWARE_HOME="/usr/local/neuware"
-     > export PATH="${NEUWARE_HOME}/bin:${PATH}"
-     > export LD_LIBRARY_PATH="${NEUWARE_HOME}/lib64:${LD_LIBRARY_PATH}"
-     > # 如您不方便将上述环境变量配置到 bashrc 文件中进行长期使用，你也可以在我们提供的 env.sh 文件中进行正确配置并激活，作为临时使用
-     > source env.sh
-     > ```
-
-     我们强烈建议您规范安装，统一到一个目录下，以免不必要的麻烦。另外请注意，由于 MLU 上层软件建设适配程度有限，如您在其覆盖的机器，操作系统之外运行，需要在安装驱动之后使用上层软件的 Docker。
-
-4. 确认您安装了 make，build-essential， python-is-python3， python-dev-is-python3， python3-pip， libdw-dev，如您的机器没有上述基础依赖，请自行按需安装。
-
-   - 在使用 apt-get 工具情况下，您可以这样执行
-
-     ```bash
-     sudo apt-get install make cmake build-essential python-is-python3 python-dev-is-python3 python3-pip libdw-dev
-     ```
-
-5. 更新pip并切换到清华源
-
-   ```bash
-   python -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple --upgrade pip
-   pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
-   ```
-
-6. 安装一些不必要的项目（可选）
-
-   - 如您需要运行本项目下的 example 代码，您需要安装一些辅助项目。请注意这些项目不是必要的，若您不需要运行样例代码，这些项目无需安装。
-
-     > [Pytorch](https://pytorch.org/get-started/locally/)：业界内流行的神经网络编程框架
-     > [ONNX](https://onnx.ai/get-started.html)：业界内流行的神经网络模型存储文件与转换器
-     > [onnxsim](https://pypi.org/project/onnxsim/)：一个简化onnx模型的小工具
-     > [onnx2torch](https://github.com/ENOT-AutoDL/onnx2torch)：一个将onnx模型转换pytorch模型的小工具
-     > [tqdm](https://pypi.org/project/tqdm/)：一个显示程序运行进度条的小工具
-
-   - 如您需要使用本项目下的 InfiniTest 测试工具，你还需要安装如下的项目：
-
-     > [protobuf](https://github.com/protocolbuffers/protobuf)： 一种序列化文件的格式及其编译、序列化、解析工具
-
-## 编译本项目
-
-推荐使用 X86-64 机器以及 Ubuntu-22.04，本文以此环境为例。
-
-1. 配置环境
-
-   打开 env.sh 文件进行环境变量配置，之后执行
-
-   ```bash
-   source env.sh
-   ```
-
-2. 编译本项目并打包成 Python 库进行安装
-
-   我们提供了意见编译参数，您可以在项目根目录下执行下面的命令。第一次执行会同时安装 python 依赖库，耗时略长，请耐心等待。
-
-   仅编译 CPU 部分，不编译第三方计算卡：
-
-   ```bash
-   make install-python
-   ```
-
-   编译 CPU 部分，同时编译英伟达 GPU 部分：
-
-   ```bash
-   export CUDA_HOME=/path/to/your/cuda_home
-   make install-python CUDA=ON
-   ```
-
-   编译 CPU 部分，同时编译寒武纪 MLU 部分：
-
-   ```bash
-   export NEUWARE_HOME=/path/to/your/neuware_home
-   make install-python BANG=ON
-   ```
-
-   编译 CPU 部分，同时编译昆仑 XPU 部分：
-
-   ```bash
-   export KUNLUN_HOME=/path/to/your/kunlun_home
-   make install-python KUNLUN=ON
-   ```
-
-3. 使用方法
-
-   安装成功后，您就可以使用本项目的 Python 接口进行编码并运行。具体使用方式可以参考项目样例代码 example/Resnet/resnet.py 以及用户使用手册
-
-## Docker
-
-本项目也提供了 Docker 的环境，您可以使用 `make docker-build` 或 `make docker-build CUDA=ON` 命令启动并编译 Dockerfile，您可以通过添加编译选项或者修改 Makefile 变量修改 docker image 名称或者所选的 Dockerfile 文件。
- 
-由于在拉取 github repo 时需要将 ssh key 加入到 github profile 中，因此暂时注释掉拉取 repo 并编译项目的过程，由用户在进入 docker 后自己维护 ssh key（将 host 中的 ssh key 复制到 docker 中可能会遇到环境不一致的问题）。
-
-```shell
-# Build docker container.
-make docker-build
-# Run docker image.
-make docker-run
-# Execute docker image.
-make docker-exec
-```
-
-如果需要编译 CUDA 版，请使用如下命令：
-```shell
-# Build docker container.
-make docker-build CUDA=ON
-# Run docker image.
-make docker-run CUDA=ON
-```
-
-## 技术支持
-
-如遇到问题，请联系我们技术支持团队
--- a/docs/SUPPORT_MATRIX_CN.md
+++ b/docs/SUPPORT_MATRIX_CN.md
@ -1,30 +0,0 @@
-# 支持矩阵
-
-## 目录
-
-
- [环境支持](#环境支持)
- [神经网络支持](#神经网络支持)
- [技术支持](#技术支持)
-
-## 环境支持
-
-目前的软硬件环境支持矩阵
-
-| Host CPU | Device        | OS            |  Support   |
-| -------- | ------------  | -----------   | ---------- |
-| X86-64   | Nvidia GPU    |  Ubuntu-22.04 |  Yes       |
-| X86-64   | Cambricon MLU |  Ubuntu-22.04 |  Yes       |
-
-## 神经网络支持
-
-目前已经验证过的神经网络模型有
-
- [x] [ResNet18-v2](https://github.com/onnx/models/blob/main/validated/vision/classification/resnet/model/resnet18-v2-7.onnx)
- [x] [DenseNet-121-12](https://github.com/onnx/models/blob/main/validated/vision/classification/densenet-121/model/densenet-12.onnx)
- [x] [Inception-2](https://github.com/onnx/models/blob/main/validated/vision/classification/inception_and_googlenet/inception_v2/model/inception-v2-9.onnx)
- [x] [EfficientNet-Lite4](https://github.com/onnx/models/blob/main/validated/vision/classification/efficientnet-lite4/model/efficientnet-lite4-11.onnx)
-
-## 技术支持
-
-如若您遇到了本项目的问题，请联系我们的技术支持团队
--- a/docs/TODO.md
+++ b/docs/TODO.md
@ -1 +0,0 @@
-
--- a/docs/USER_GUIDE_CN.md
+++ b/docs/USER_GUIDE_CN.md
@ -1,203 +0,0 @@
-# 使用指南
-
-## 目录
-
- [使用方法](#使用方法)
- [python 前端应用指南](#python-前端应用指南)
-  - [导入 onnx 模型](#导入-onnx-模型)
-  - [优化](#优化)
-  - [导出 onnx 模型](#导出-onnx-模型)
-  - [执行推理](#执行推理)
-  - [样例代码](#样例代码)
- [技术支持](#技术支持)
- [测试](#测试)
-
-## 使用方法
-
-项目管理功能已写到 [Makefile](../Makefile)，支持下列功能：
-
- 编译项目：`make`/`make build`
- 清理生成文件：`make clean`
- 安装 python 库：`make install-python`
- 测试 c++ 后端：`make test-cpp`
- 测试 python 前端：`make test-onnx`
-
-并使用下列环境变量传递选项参数：
-
- `TYPE`：编译模式（`debug`/`release`），默认值为 `release`
- `CUDA`：是否编译 CUDA 后端，默认为 `OFF`，`ON` 打开
- `BANG`：是否编译寒武纪后端，默认为 `OFF`，`ON` 打开
- `KUNLUN`：是否编译昆仑后端，默认为 `OFF`，`ON` 打开
- `BACKTRACE`：是否启用栈回溯，默认为 `ON`，`OFF` 关闭，建议调试时打开
- `TEST`：是否编译 `googletest`，默认为 `ON`，`OFF` 关闭，只有 `test-cpp` 时必要
-
-## python 前端应用指南
-
-`make install-python` 会将项目的 python 前端以 `pyinfinitensor` 为名字安装到系统目录，可以直接 `import pyinfinitensor` 来使用。现阶段，项目的主要用法是从 onnx 导入模型进行优化，然后可以再导出优化后的模型到 onnx，也可以直接运行推理。
-
-### 导入 onnx 模型
-
-支持的模型：
-
- [x] [ResNet18-v2](https://github.com/onnx/models/blob/main/validated/vision/classification/resnet/model/resnet18-v2-7.onnx)
- [x] [DenseNet-121-12](https://github.com/onnx/models/blob/main/validated/vision/classification/densenet-121/model/densenet-12.onnx)
- [x] [Inception-2](https://github.com/onnx/models/blob/main/validated/vision/classification/inception_and_googlenet/inception_v2/model/inception-v2-9.onnx)
- [x] [EfficientNet-Lite4](https://github.com/onnx/models/blob/main/validated/vision/classification/efficientnet-lite4/model/efficientnet-lite4-11.onnx)
-
-```python
-import onnx
-from pyinfinitensor.onnx import OnnxStub
-from pyinfinitensor import backend
-
-stub = OnnxStub(onnx.load("model_file"), backend.cpu_runtime())
-```
-
-[`onnx.load`](https://onnx.ai/onnx/api/serialization.html#load-a-model) 是 onnx 提供的加载函数，将 onnx 文件读取为保存在内存中的 onnx 模型。
-
-`OnnxStub` 是 onnx 模型在项目中的表示，通过构造这个对象，将 onnx 模型导入到项目中。其构造器的第一个参数是 onnx 模型文件；第二个参数是模型运行的后端运行时，可以是 `backend.cpu_runtime()`、`backend.cuda_runtime()` 或 `backend.bang_runtime()`。
-
-构造出的 stub 对象可以用于操作项目中的模型和运行时。
-
-### 优化
-
-TODO
-
-### 导出 onnx 模型
-
-优化后的模型可以导出成 onnx 文件提供给其他运行时。
-
-```python
-with open("optimized.onnx", "wb") as f:
-    f.write(stub.to_onnx("optimized").SerializeToString())
-```
-
-`stub.to_onnx(<name>)` 将模型转换为 onnx 模型对象，`<name>` 将填写到 onnx 模型的 `name` 字段。序列化到文件的代码见[官方示例](https://onnx.ai/onnx/intro/python.html#model-serialization)。
-
-要可视化检查导出的模型文件，可以利用 [onnx 提供的功能](https://onnx.ai/onnx/api/shape_inference.html#infer-shapes)将所有的张量的形状推理出来再导出：
-
-```python
-from onnx.shape_inference import infer_shapes
-
-with open("optimized.onnx", "wb") as f:
-    f.write(infer_shapes(stub.to_onnx("optimized")).SerializeToString())
-```
-
-然后用 [Netron](https://netron.app/) 绘制计算图。
-
-### 执行推理
-
-也可以使用项目的运行时执行推理。
-
-第一步是将数据传入计算图。`OnnxStub.inputs` 是一个 `Dict[str, Tensor]`，保存着模型的所有输入的名字和对象。可以用 [`items()`](https://docs.python.org/zh-cn/3/library/stdtypes.html#dict.items) 来遍历。
-
-这个代码片段显示了如何打印出模型所有输入张量的名字、形状和对象指针：
-
-```python
-for name, tensor in stub.inputs.items():
-    print(name, tensor.shape(), tensor)
-```
-
-对于 [resnet18-v2-7.onnx](https://github.com/onnx/models/blob/main/validated/vision/classification/resnet/model/resnet18-v2-7.onnx)，会打印出：
-
-```plaintext
-data [1, 3, 224, 224] <backend.Tensor object at 0x7efeb828e3b0>
-```
-
-当然，地址是随机的。这个输出表明需要输入一个名为 “data”，形为 1×3×224×224 的数据。通常来说，这表示一张 224×224 的 rgb 图片。而这个模型是一个 1000 分类的图像分类模型。
-
-为了方便，这里我们向模型传入一个随机的数据。
-
-```python
-import numpy
-
-stub.init()
-for name, tensor in stub.inputs.items():
-    print(name, tensor.shape(), tensor)
-    input = numpy.random.random(tensor.shape()).astype(numpy.float32)
-    tensor.copyin_float(input.flatten().tolist())
-```
-
-`stub.init()` 为所有张量分配空间。空间是预分配的，所以不支持动态 size 的模型。
-
-`tensor.copyin_float(<data>)` 向张量传入数据。其参数必须是一个 `List[float]`，即压平的数据。类似的函数还有 `copyin_int32(<data>)` 和 `copyin_int64(<data>)`
-
-然后，调用 `stub.run()` 执行推理：
-
-```python
-stub.run()
-```
-
-最后，将结果拷贝出来，传入类似：
-
-```python
-stub.init()
-for name, tensor in stub.outputs.items():
-    print(name, tensor.shape(), tensor)
-    print(tensor.copyout_float())
-```
-
-### 样例代码
-
-您可以参照[resnet.py](https://github.com/wanghailu0717/NNmodel/blob/main/ResNet/resnet.py)的样例代码进行了解，并尝试运行。在这个文件中，我们使用了 Pytorch 构建了 resnet 网络。您可以查阅该脚本使用方式：
-
-```python
-python resnet.py -h
-```
-
-在样例代码中，我们对定义的网络进行了序列化操作，并存储为模型文件。之后加载该模型文件，并转换为本项目的模型进行优化操作，再进行推理。您可以关注一下代码中 242 行之后的代码。请注意，您可以按照您的需求来进行操作，通常来说，您所需要撰写的代码就是加载模型，转换为本项目的模型进行优化，推理运行。
-
-## 技术支持
-
-如若您遇到了本项目的问题，请联系我们的技术支持团队
-
-## 测试
-
-除了单元测试 `make test-cpp` 和 `make test-onnx` 之外，还可以用其他方式来测试单个模型导入导出和优化的正确性。
-
-这个脚本利用 onnxruntime 来测试导出的模型是否与导入的模型等价：
-
-```python
-import onnx
-import numpy
-import sys
-from onnx import ModelProto, ValueInfoProto
-from pyinfinitensor.onnx import OnnxStub
-from pyinfinitensor import backend
-from onnxruntime import InferenceSession
-
-
-def infer(model: ModelProto, input) -> dict:
-    collection = set()
-    for node in model.graph.node:
-        for output in node.output:
-            collection.add(output)
-    model.graph.output.extend([ValueInfoProto(name=x) for x in collection])
-    session = InferenceSession(model.SerializeToString())
-    i = session.get_inputs()[0].name
-    return dict(
-        zip(
-            [x.name for x in session.get_outputs()],
-            [x.flatten() for x in session.run(None, {i: input})],
-        )
-    )
-
-
-model0 = onnx.load(sys.argv[1])
-model1 = OnnxStub(model0, backend.cpu_runtime()).to_onnx("new")
-
-input_shape = [x.dim_value for x in model1.graph.input[0].type.tensor_type.shape.dim]
-input = numpy.random.random(input_shape).astype(numpy.float32)
-
-output0 = infer(model0, input)[model0.graph.output[0].name]
-output1 = infer(model1, input)[model1.graph.output[0].name]
-
-print("error =", sum((output1 - output0) ** 2) / len(output0))
-```
-
-要运行脚本，先安装 onnxruntime：
-
-```bash
-pip install onnxruntime
-```
-
-打印出的 `error = ...` 是两个模型输出张量的均方误差。对于不同的模型，这个误差最小为 0，最大不超过 1e-9。
--- a/env.sh
+++ b/env.sh
@ -1,38 +0,0 @@
-# 配置英伟达 CUDA 的 HOME 路径，请注意安装 CUDA Toolkit, CUDNN 并将路径配置到下述环境变量。
-export CUDA_HOME=/PATH/TO/YOUR/CUDA/HOME
-export CUDNN_HOME=/PATH/TO/YOUR/CUDNN/HOME
-export PATH="${CUDA_HOME}/bin:${PATH}"
-export LD_LIBRARY_PATH="${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}"
-
-# 配置寒武纪 BANG 的 HOME 路径，请注意 /usr/local/neuware 是寒武纪软件栈建议的，同时也是默认的安装路径。
-# 如若用户有其他的路径安装方式，请自行配置正确的路径。
-# 这里是 neuware 目录下一个可能的结构图，请参考。
-# .
-# ├── bin
-# ├── cmake
-# ├── data
-# ├── edge
-# ├── include
-# ├── lib
-# ├── lib64
-# ├── LICENSE
-# ├── mlvm
-# ├── README
-# ├── samples
-# ├── share
-# └── version.txt
-export NEUWARE_HOME=/usr/local/neuware
-export PATH="${NEUWARE_HOME}/bin:${PATH}"
-export LD_LIBRARY_PATH="${NEUWARE_HOME}/lib64:${LD_LIBRARY_PATH}"
-
-# 配置昆仑芯 XPU 的 HOME 路径，请注意 /usr/local/xpu 是昆仑芯软件栈提供的软件包路径。
-# 如若用户有其他的路径安装方式，请自行配置正确的路径。
-# 这里是 xpu 目录下一个可能的结构图，请参考。
-# .
-# ├── bin
-# ├── include
-# ├── lib64
-# ├── tools
-# ├── version
-# └── XTDK
-export KUNLUN_HOME=/usr/local/xpu
--- a/examples/NNmodel
+++ b/examples/NNmodel
@ -1 +0,0 @@
-Subproject commit 51d3105277f3774ed31c02ed4cd11fa92925af77
--- a/examples/distributed/README.md
+++ b/examples/distributed/README.md
@ -1,39 +0,0 @@
-# 分布式脚本
-
-## 英伟达平台运行方式
-
-#### 1. 运行pytorch模型并生成输入和标准输出，可选择导出onnx
-
-使用 `--export_onnx` 设置导出onnx的目录，默认为当前路径 `./`，不使用这个flag则只进行计算和生成输入输出。
-
-```bash
-python run_pytorch.py --model gpt2  --batch_size 1  --length 1 --export_onnx ./
-```
-
-会在当前目录下生成输入输出文件`test_inputs.npy` 和 `test_results.npy`，目前只支持单一输入输出。
-
-#### 2. 运行InfiniTensor分布式脚本
-
-```bash
-python cuda_launch.py --model "/XXX/XXX.onnx" --nproc_per_node 4 
-```
-
-## 寒武纪平台运行方式
-
-**将上述运行脚本 `run_pytorch.py` 以及 `cuda_launch.py` 针对寒武纪平台做了相应的适配，具体见 `run_pytorch_mlu.py` 以及 `bang_launch.py`。**
-
-#### 1. 运行pytorch模型并生成输入和标准输出，可选择导出onnx
-
-使用 `--export_onnx` 设置导出onnx的目录，默认为当前路径 `./`，不使用这个flag则只进行计算和生成输入输出。
-
-```bash
-python run_pytorch_mlu.py --model gpt2  --batch_size 1  --length 1 --export_onnx ./
-```
-
-会在当前目录下生成输入输出文件`test_inputs.npy` 和 `test_results.npy`，目前只支持单一输入输出。
-
-#### 2. 运行InfiniTensor分布式脚本
-
-```bash
-python bang_launch.py --model "/XXX/XXX.onnx" --nproc_per_node 4 
-```
--- a/examples/distributed/init.py
+++ b/examples/distributed/init.py
--- a/examples/distributed/bang/bang_launch.py
+++ b/examples/distributed/bang/bang_launch.py
@ -1,187 +0,0 @@
-import sys
-sys.path.append('../')
-
-import argparse
-import os
-import time
-import multiprocessing as mp
-from pyinfinitensor.onnx import OnnxStub, backend
-import onnx
-from onnx.external_data_helper import convert_model_to_external_data
-from onnx.shape_inference import infer_shapes_path
-import numpy as np
-from parallel_opt import parallel_model
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="launch distributed infinitensor")
-    parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
-    parser.add_argument(
-        "--nproc_per_node", type=int, default=1, help="number of processes per node"
-    )
-    parser.add_argument(
-        "--name", type=str, default="test", help="name of this instance."
-    )
-    parser.add_argument(
-        "--model", type=str, required=True, help="path to the ONNX model file."
-    )
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
-    parser.add_argument("--length", type=int, default=1, help="sequence length.")
-    parser.add_argument(
-        "--gen_std",
-        action="store_true",
-        help="whether to generate the standard results.",
-    )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], default="fp32", help="data type"
-    )
-    args = parser.parse_args()
-    print("arg setting: ", args)
-    return (
-        args.num_nodes,
-        args.nproc_per_node,
-        args.name,
-        args.model,
-        args.batch_size,
-        args.length,
-        args.gen_std,
-        args.type,
-    )
-
-
-def run_model(model, runtime, world_size=1, rank=0, n=10, data_type="default"):
-    stub = OnnxStub(model, runtime, matmul_compute_type=data_type)
-    load_inputs(stub, world_size, rank)
-    # stub.tune()
-    stub.run()
-    # get outputs
-    outputs = next(stub.outputs.values().__iter__()).copyout_numpy()
-
-    # bench
-    for _ in range(n):
-        stub.run()
-    begin = time.time()
-    for _ in range(n * 2):
-        stub.run()
-    end = time.time()
-    avg_time = (end - begin) / (n * 2)
-    print(f"average time: {avg_time}")
-    return outputs
-
-def load_inputs(stub, world_size=1, rank=0):
-    for i, (name, tensor) in enumerate(stub.inputs.items()):
-        input = np.load(f"./data/input_{i}.npy")
-        if all(x == y for x,y in zip(input.shape,tensor.shape())):
-            tensor.copyin_numpy(input)
-        else:
-            tensor.copyin_numpy(np.hsplit(input, world_size)[rank])
-
-
-def run_and_compare(name, model, runtime, world_size=1, rank=0, data_type="default"):
-    results = np.load(f"./data/output.npy")
-    outputs = run_model(model, runtime, world_size, rank, data_type=data_type)
-    print("outputs abs mean:", abs(outputs).mean())
-    print("max abs diff:", abs(outputs - results).max())
-
-def start_worker(
-    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto, data_type: str
-):
-    dist_name = name + "_dist"
-    model = parallel_model(model, world_size, rank)
-    extern_path = f"./{dist_name}_rank{rank}.pb"
-    if os.path.exists(extern_path):
-        os.remove(extern_path)
-    onnx.save_model(
-        model,
-        f"./{dist_name}_rank{rank}.onnx",
-        save_as_external_data=True,
-        location=extern_path,
-    )
-    #infer_shapes_path(f"./{dist_name}_rank{rank}.onnx")
-    runtime = backend.BangRuntime(local_rank)
-    # print("init comm")
-    runtime.init_comm(
-        dist_name,
-        world_size,
-        rank,
-    )
-    run_and_compare(name, model, runtime, world_size, rank, data_type)
-
-
-def start_single(name, model, data_type):
-    runtime = backend.BangRuntime(0)
-    run_and_compare(name, model, runtime, data_type=data_type)
-
-def generate_input_output(model):
-    os.makedirs(os.path.dirname("./data/"), exist_ok=True)
-    runtime = backend.BangRuntime(0)
-    stub = OnnxStub(model, runtime)
-    position_id = 0
-    for i, (name, tensor) in enumerate(stub.inputs.items()):
-        input = tensor.copyout_numpy()
-        if np.issubdtype(input.dtype, np.integer):
-            if input.size == 1:
-                # input = np.array([position_id])
-                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
-            else:
-                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
-        elif input.dtype == np.bool_:
-            input = np.random.randint(0,2,size=input.shape) > 0
-        else:
-            if i == 0:
-                input = np.ones(input.shape).astype(input.dtype)
-                position_id = input.shape[-1] - 1
-            else:
-                input = np.random.rand(*input.shape).astype(input.dtype)
-        tensor.copyin_numpy(input)
-        np.save(f"./data/input_{i}", input)
-    stub.run()
-    time.sleep(0.01)
-    output = next(stub.outputs.values().__iter__()).copyout_numpy()
-    if np.isnan(output).any():
-        print("Nan in output")
-    np.save(f"./data/output", output)
-
-
-def main():
-    nnodes, nproc_per_node, name, model_path, bs, length, gen_std, data_type = parse_args()
-    data_type = "default" if data_type == "fp32" else data_type
-    
-    model = onnx.load(model_path)
-
-    # generate standart output
-    if gen_std:
-        print(f"generate standard data for {name}.")
-        # a small vocabulary size to fit all LLM.
-        generate_input_output(model)
-        return
-
-    if nproc_per_node == 1:
-        # run single process.
-        # use standalone process to isolate bang.
-        print("run model by single MLU.")
-        # p = mp.Process(target=start_single, args=(name, model, data_type))
-        # p.start()
-        # p.join()
-        start_single(name, model, data_type)
-        return
-
-    # run distributed parallel.
-    world_size = nnodes * nproc_per_node
-    print(f"run model by {world_size} MLU in parallel.")
-    workers = [
-        mp.Process(
-            target=start_worker,
-            args=(name, world_size, rank, rank % nproc_per_node, model, data_type),
-        )
-        for rank in range(world_size)
-    ]
-
-    for w in workers:
-        w.start()
-
-    for w in workers:
-        w.join()
-
-
-if __name__ == "__main__":
-    main()
--- a/examples/distributed/bang/run_pytorch_mlu.py
+++ b/examples/distributed/bang/run_pytorch_mlu.py
@ -1,249 +0,0 @@
-import argparse
-import torch
-import torch_mlu
-from transformers import BertModel, BertConfig
-from transformers import GPT2Model, GPT2Config
-from transformers import OPTModel, OPTConfig
-from transformers import AlbertModel, AlbertConfig
-from transformers import LlamaModel, LlamaConfig
-import time
-import numpy as np
-import onnx
-import sys
-import os
-from onnx.external_data_helper import convert_model_to_external_data
-from onnxsim import simplify
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Run pytorch gpt2/bert/opt and optionally export onnx.")
-    parser.add_argument(
-        "--model", type=str, choices=["gpt2", "bert", "opt", "llama", "albert"], required=True, help="model type"
-    )
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
-    parser.add_argument("--length", type=int, default=1, help="sequence length.")
-    parser.add_argument(
-        "--export_onnx",
-        type=str,
-        nargs="?",
-        default=None,
-        const="./",
-        help="whether and where to export onnx file",
-    )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], required=True, help="model data type"
-    )
-    args = parser.parse_args()
-    print("arg setting: ", args)
-    return (
-        args.model,
-        args.batch_size,
-        args.length,
-        args.export_onnx,
-        args.type
-    )
-
-
-def get_model(modelname):
-    match modelname:
-        case "albert":
-            model = AlbertModel.from_pretrained("albert/albert-base-v2")
-            voc_size = AlbertConfig().vocab_size
-        case "bert":
-            model = BertModel.from_pretrained("bert-base-uncased", add_pooling_layer=False, hidden_act="gelu_new") # erf is not impl by infini
-            voc_size = BertConfig().vocab_size
-        case "gpt2":
-            model = GPT2Model.from_pretrained("GPT2")
-            voc_size = GPT2Config().vocab_size
-        case "opt":
-            model = OPTModel.from_pretrained("facebook/opt-125m")
-            voc_size = OPTConfig().vocab_size
-        case "llama":
-            model = LlamaModel.from_pretrained("meta-llama/Llama-2-7b-hf")
-            voc_size = LlamaConfig().vocab_size
-        case _:
-            raise KeyError(modelname)
-
-    model = model.eval()
-    return model, voc_size
-
-def run_pytorch(torch_model, voc_size, batchsize, len, dtype="fp32"):
-    data = np.random.randint(0, voc_size, (batchsize, len), dtype=np.int32)
-    os.makedirs(os.path.dirname("./data/"), exist_ok=True)
-    np.save("./data/input_0", data)
-    inputs = torch.from_numpy(data).to("mlu")
-    torch_model = torch_model.to("mlu")
-    if dtype == "fp16":
-        torch_model = torch_model.half()
-
-    n_iter = 20
-    with torch.no_grad():
-        for _ in range(10):
-            outputs = torch_model(inputs)
-    torch.mlu.synchronize()
-    begin = time.time()
-    with torch.no_grad():
-        for _ in range(n_iter):
-            torch.mlu.synchronize()
-            outputs = torch_model(inputs)
-            torch.mlu.synchronize()
-    torch.mlu.synchronize()
-    end = time.time()
-    
-    avg_time = (end - begin) / n_iter
-    outputs = outputs.last_hidden_state.to("cpu")
-    print("outputs abs mean:", abs(np.array(outputs)).mean())
-    print(f"average time: {avg_time}")
-    # torch.mlu.memory.empty_cache()
-    np.save("./data/output", np.array(outputs))
-    print("Save input & output into ./data.")
-
-
-def export_onnx(modelname, model, data, path, extern=False, dtype="fp32"):
-    data = data.to("mlu")
-    model = model.to("mlu")
-    if dtype == "fp16":
-        model = model.half()
-    torch.onnx.export(model, data, path, verbose=False, do_constant_folding=True)
-    if modelname != "llama":
-        # use onnxsim to simplify
-        onnx_model = onnx.load(path)
-        onnx_model, check = simplify(onnx_model, skipped_optimizers=['eliminate_duplicate_initializer'])
-        # onnx_model, check = simplify(onnx_model, skipped_optimizers=['fuse_qkv', 'eliminate_duplicate_initializer'])
-        assert check
-        add_value_info_for_constants(onnx_model)
-        onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
-        if extern:
-            extern_path = path.replace('.onnx', '.pb')
-            if os.path.exists(extern_path):
-                os.remove(extern_path)
-            extern_path = extern_path.split("/")[-1]
-            convert_model_to_external_data(
-                onnx_model,
-                all_tensors_to_one_file=True,
-                location=extern_path,
-                size_threshold=1024,
-                convert_attribute=False,
-            )
-        onnx.save(onnx_model, path)
-    else:
-        # use third party tool to simplify llama
-        # reference: https://github.com/luchangli03/onnxsim_large_model/
-        sys.path.append("onnxsim_large_model")
-        from onnx_utils import set_onnx_input_shape
-        from compress_model import SIZE_1MB, compress_onnx_model, uncompress_onnx_model
-
-        in_model_path = path
-        out_model_path = path
-        if not out_model_path:
-            out_model_path = in_model_path[:-5] + ".sim.onnx"
-        if os.path.isdir(out_model_path):
-            out_model_path = os.path.join(out_model_path, os.path.basename(in_model_path))
-
-        onnx_model = onnx.load(in_model_path)
-        print(f"load model from {in_model_path} success")
-
-        size_th_bytes = 1024 * 1024
-
-        onnx_model, removed_inits = compress_onnx_model(onnx_model, size_th_bytes=size_th_bytes)
-        print(f"compress model success")
-
-        onnx_model = set_onnx_input_shape(onnx_model, "")
-
-        tensor_size_threshold = f"1024KB"
-        skipped_optimizers = []
-        skipped_optimizers.append("eliminate_duplicate_initializer")
-        onnx_model, check = simplify(onnx_model, skipped_optimizers=skipped_optimizers,
-                                    tensor_size_threshold=tensor_size_threshold)
-        if not check:
-            raise ValueError(f"simplify compressed model {in_model_path} failed")
-
-        print(f"simplify model success")
-
-        onnx_model = uncompress_onnx_model(onnx_model, removed_inits)
-        print(f"uncompress model success")
-
-        add_value_info_for_constants(onnx_model)
-
-        onnx.save(onnx_model, out_model_path, save_as_external_data=True)
-
-
-def add_value_info_for_constants(model : onnx.ModelProto):
-    """
-    Currently onnx.shape_inference doesn't use the shape of initializers, so add
-    that info explicitly as ValueInfoProtos.
-    Mutates the model.
-    Args:
-        model: The ModelProto to update.
-    """
-    # All (top-level) constants will have ValueInfos before IRv4 as they are all inputs
-    if model.ir_version < 4:
-        return
-
-    def add_const_value_infos_to_graph(graph : onnx.GraphProto):
-        inputs = {i.name for i in graph.input}
-        existing_info = {vi.name: vi for vi in graph.value_info}
-        for init in graph.initializer:
-            # Check it really is a constant, not an input
-            if init.name in inputs:
-                continue
-
-            # The details we want to add
-            elem_type = init.data_type
-            shape = init.dims
-
-            # Get existing or create new value info for this constant
-            vi = existing_info.get(init.name)
-            if vi is None:
-                vi = graph.value_info.add()
-                vi.name = init.name
-
-            # Even though it would be weird, we will not overwrite info even if it doesn't match
-            tt = vi.type.tensor_type
-            if tt.elem_type == onnx.TensorProto.UNDEFINED:
-                tt.elem_type = elem_type
-            if not tt.HasField("shape"):
-                # Ensure we set an empty list if the const is scalar (zero dims)
-                tt.shape.dim.extend([])
-                for dim in shape:
-                    tt.shape.dim.add().dim_value = dim
-
-        # Handle subgraphs
-        for node in graph.node:
-            for attr in node.attribute:
-                # Ref attrs refer to other attrs, so we don't need to do anything
-                if attr.ref_attr_name != "":
-                    continue
-
-                if attr.type == onnx.AttributeProto.GRAPH:
-                    add_const_value_infos_to_graph(attr.g)
-                if attr.type == onnx.AttributeProto.GRAPHS:
-                    for g in attr.graphs:
-                        add_const_value_infos_to_graph(g)
-
-
-    return add_const_value_infos_to_graph(model.graph)
-
-
-def main():
-    torch.backends.mlu.matmul.allow_tf32 = False
-    torch.backends.cnnl.allow_tf32 = False
-    modelname, batchsize, seqlen, export_path, dtype = parse_args()
-    if dtype == "tf32":
-        torch.backends.mlu.matmul.allow_tf32 = True
-    else:
-        os.environ["CAMBRICON_TF32_OVERRIDE"] = "0"
-
-    model, voc_size = get_model(modelname)
-    if export_path is not None:
-        filename = "{}_{}_{}_{}.onnx".format(modelname, batchsize, seqlen, dtype)
-        path = os.path.join(export_path, filename)
-        if not os.path.exists(path):
-            param = torch.zeros((batchsize, seqlen), dtype=torch.int)
-            export_onnx(modelname, model, param, path, True, dtype)
-        else:
-            print("Onnx path exists, skipping export.")
-
-    run_pytorch(model, voc_size, batchsize, seqlen, dtype)
-
-if __name__ == "__main__":
-    main()
--- a/examples/distributed/cuda/cuda_launch.py
+++ b/examples/distributed/cuda/cuda_launch.py
@ -1,161 +0,0 @@
-import argparse
-import os
-import time
-import multiprocessing as mp
-from pyinfinitensor.onnx import OnnxStub, backend
-import onnx
-from onnx.external_data_helper import convert_model_to_external_data
-from onnx.shape_inference import infer_shapes_path
-import numpy as np
-from parallel_opt import parallel_model
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="launch distributed infinitensor")
-    parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
-    parser.add_argument(
-        "--nproc_per_node", type=int, default=1, help="number of processes per node"
-    )
-    parser.add_argument(
-        "--name", type=str, default="test", help="name of this instance."
-    )
-    parser.add_argument(
-        "--model", type=str, required=True, help="path to the ONNX model file."
-    )
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
-    parser.add_argument("--length", type=int, default=1, help="sequence length.")
-    parser.add_argument(
-        "--gen_std",
-        action="store_true",
-        help="whether to generate the standard results.",
-    )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], default="fp32", help="data type"
-    )
-    args = parser.parse_args()
-    print("arg setting: ", args)
-    return (
-        args.num_nodes,
-        args.nproc_per_node,
-        args.name,
-        args.model,
-        args.batch_size,
-        args.length,
-        args.gen_std,
-        args.type,
-    )
-
-
-def run_model(model, runtime, inputs, n=10, data_type = "default"):
-    stub = OnnxStub(model, runtime, matmul_compute_type=data_type)
-    for tensor, input in zip(stub.inputs.values(), inputs, strict=False):
-        tensor.copyin_numpy(input)
-    # stub.tune()
-    stub.run()
-    # get outputs
-    outputs = next(stub.outputs.values().__iter__()).copyout_numpy()
-
-    # bench
-    for tensor, input in zip(stub.inputs.values(), inputs, strict=False):
-        tensor.copyin_numpy(input)
-    begin = time.time()
-    for _ in range(n):
-        stub.run()
-    end = time.time()
-    avg_time = (end - begin) / n
-    print(f"average time: {avg_time}")
-    return outputs
-
-
-def run_and_compare(name, model, runtime, data_type):
-    input_ids = np.load(f"{name}_inputs.npy")
-    position_ids = np.arange(input_ids.shape[-1])
-    results = np.load(f"{name}_results.npy")
-    outputs = run_model(model, runtime, (input_ids, position_ids), data_type=data_type)
-    print("outputs abs mean:", abs(outputs).mean())
-    print("max abs diff:", abs(outputs - results).max())
-
-
-def start_worker(
-    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto, data_type: str
-):
-    dist_name = name + "_dist"
-    model = parallel_model(model, world_size, rank)
-    extern_path = f"./{dist_name}_rank{rank}.pb"
-    if os.path.exists(extern_path):
-        os.remove(extern_path)
-    onnx.save_model(
-        model,
-        f"./{dist_name}_rank{rank}.onnx",
-        save_as_external_data=True,
-        location=extern_path,
-    )
-    #infer_shapes_path(f"./{dist_name}_rank{rank}.onnx")
-    runtime = backend.CudaRuntime(local_rank)
-    # print("init comm")
-    runtime.init_comm(
-        dist_name,
-        world_size,
-        rank,
-    )
-    run_and_compare(name, model, runtime, data_type)
-
-
-def start_single(name, model, data_type):
-    runtime = backend.CudaRuntime(0)
-    run_and_compare(name, model, runtime, data_type)
-
-
-def gen_standard(name, model, voc_size, bs, len):
-    # generate standard results
-    input_ids = np.random.randint(0, voc_size, (bs, len))
-    position_ids = np.arange(len)
-    np.save(f"{name}_inputs", input_ids)
-    runtime = backend.CudaRuntime(0)
-    outputs = run_model(model, runtime, (input_ids, position_ids), 1)
-    print("outputs abs mean:", abs(outputs).mean())
-    np.save(f"{name}_results", outputs)
-
-
-def main():
-    nnodes, nproc_per_node, name, model_path, bs, length, gen_std, data_type = parse_args()
-    data_type = "default" if data_type == "fp32" else data_type
-    if data_type != "tf32":
-        os.environ["NVIDIA_TF32_OVERRIDE"] = "0"
-    model = onnx.load(model_path)
-
-    # generate standart output
-    if gen_std:
-        print(f"generate standard data for {name}.")
-        # a small vocabulary size to fit all LLM.
-        voc_size = 1000
-        gen_standard(name, model, voc_size, bs, length)
-        return
-
-    # run single process.
-    # use standalone process to isolate cuda.
-    print("run model by single GPU.")
-    p = mp.Process(target=start_single, args=(name, model, data_type))
-    p.start()
-    p.join()
-
-    # run distributed parallel.
-    world_size = nnodes * nproc_per_node
-    print(f"run model by {world_size} GPU in parallel.")
-    workers = [
-        mp.Process(
-            target=start_worker,
-            args=(name, world_size, rank, rank % nproc_per_node, model, data_type),
-        )
-        for rank in range(world_size)
-    ]
-
-    for w in workers:
-        w.start()
-
-    for w in workers:
-        w.join()
-
-
-if __name__ == "__main__":
-    main()
--- a/examples/distributed/cuda/launch_kvcache.py
+++ b/examples/distributed/cuda/launch_kvcache.py
@ -1,245 +0,0 @@
-import argparse
-import os
-import time
-import multiprocessing as mp
-from pyinfinitensor.onnx import OnnxStub, backend
-import onnx
-from onnx.external_data_helper import convert_model_to_external_data
-import numpy as np
-from parallel_opt import parallel_model
-
-
-os.environ["NVIDIA_TF32_OVERRIDE"] = "0"
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="launch distributed infinitensor")
-    parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
-    parser.add_argument(
-        "--nproc_per_node", type=int, default=1, help="number of processes per node"
-    )
-    parser.add_argument(
-        "--name", type=str, default="test", help="name of this instance."
-    )
-    parser.add_argument(
-        "--model1", type=str, required=True, help="path to the ONNX model file."
-    )
-    parser.add_argument(
-        "--model2", type=str, required=True, help="path to the ONNX model file."
-    )    
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
-    parser.add_argument("--length", type=int, default=1, help="sequence length.")
-    parser.add_argument(
-        "--gen_std",
-        action="store_true",
-        help="whether to generate the standard results.",
-    )
-    args = parser.parse_args()
-    print("arg setting: ", args)
-    return (
-        args.num_nodes,
-        args.nproc_per_node,
-        args.name,
-        args.model1,
-        args.model2,
-        args.batch_size,
-        args.length,
-        args.gen_std,
-    )
-
-
-def run_model(model1, model2, runtime1, runtime2, inputs1: np.array, inputs2: np.array, n=20):
-    ####################################
-    # run the first graph without kvcache
-    ####################################
-    stub1 = OnnxStub(model1, runtime1)
-    stub1.inputs['onnx::Reshape_0'].copyin_int32(inputs1.reshape(-1).tolist())
-    stub1.tune()
-    stub1.run()
-    kvcache_it1 = []
-    count = 0
-    for output in stub1.outputs.items().__iter__():
-        if count == 0:
-            logits_it1 = np.array(output[1].copyout_float(), dtype=np.float32)
-        else:
-            kvcache_it1.append(np.array(output[1].copyout_float(), dtype=np.float32))
-        count = count + 1
-        
-    # bench for stub1
-    next(stub1.inputs.items().__iter__())[1].copyin_int32(inputs1.reshape(-1).tolist())
-    begin = time.time()
-    for _ in range(n):
-        stub1.run()
-    end = time.time()
-    avg_time = (end - begin) / n
-    print(f"stub1 average time: {avg_time}")        
-        
-    ####################################
-    # run the second graph with kvcache
-    ####################################
-    i = 0
-    batchsize = 1
-    stub2 = OnnxStub(model2, runtime2)
-    past_kvcache_length = (i+2)*np.ones((batchsize, 1), dtype=np.int32)
-    # copyin input
-    stub2.inputs['onnx::Reshape_0'].copyin_int32(inputs2.reshape(-1).tolist())
-    stub2.inputs['input.3'].copyin_int32(past_kvcache_length.reshape(-1).tolist())
-    count = -1
-    for input in stub2.inputs.items().__iter__():
-        if count in range(24):
-            # print(count, input[0])
-            # print(np.dtype(kvcache_it1[count][0]), kvcache_it1[count].shape)
-            input[1].copyin_float(kvcache_it1[count].reshape(-1).tolist())
-        count = count + 1
-    stub2.tune()
-    stub2.run()
-    
-    # copyout output
-    count = 0
-    kvcache_it2 = []
-    for output in stub2.outputs.items().__iter__():
-        if count == 0:
-            logits_it2 = np.array(output[1].copyout_float(), dtype=np.float32)
-        else:
-            kvcache_it2.append(np.array(output[1].copyout_float(), dtype=np.float32))
-        count = count + 1     
-
-    # bench for stub2
-    # copyin input
-    stub2.inputs['onnx::Reshape_0'].copyin_int32(inputs2.reshape(-1).tolist())
-    stub2.inputs['input.3'].copyin_int32(past_kvcache_length.reshape(-1).tolist())
-    count = -1
-    for input in stub2.inputs.items().__iter__():
-        if count in range(24):
-            input[1].copyin_float(kvcache_it1[count].reshape(-1).tolist())
-        count = count + 1
-    begin = time.time()
-    for _ in range(n):
-        stub2.run()
-    end = time.time()
-    avg_time = (end - begin) / n
-    print(f"stub2 average time: {avg_time}")
-    return logits_it2
-
-
-def run_and_compare(name, model1, model2, runtime1, runtime2):
-    data1 = np.load(f"{name}_inputs1.npy")
-    data2 = np.load(f"{name}_inputs2.npy")
-    results = np.load(f"{name}_results.npy")
-    outputs = run_model(model1, model2, runtime1, runtime2, data1, data2)
-    print("outputs sum:", outputs.sum())
-    print("max abs diff:", abs(outputs - results).max())
-    print("max rel diff:", abs((outputs - results) / results).max())
-    # assert np.allclose(outputs, results, rtol=1e-3, atol=1e-6)
-
-
-def start_worker(
-    name: str, world_size: int, rank: int, local_rank: int, model1: onnx.ModelProto, model2: onnx.ModelProto
-):
-    dist_name = name + "_dist"
-    ####################################
-    # shard the first graph
-    ####################################
-    model1 = parallel_model(model1, world_size, rank)
-    extern_path = f"./{dist_name}_stub1_rank{rank}.pb"
-    if os.path.exists(extern_path):
-        os.remove(extern_path)
-    convert_model_to_external_data(
-        model1,
-        all_tensors_to_one_file=True,
-        location=extern_path,
-        size_threshold=1024,
-        convert_attribute=False,
-    )
-    onnx.save(model1, f"./{dist_name}_stub1_rank{rank}.onnx")
-    runtime1 = backend.CudaRuntime(local_rank)
-    runtime1.init_comm(
-        dist_name,
-        world_size,
-        rank,
-    )
-    
-    ####################################
-    # shard the second graph
-    ####################################    
-    model2 = parallel_model(model2, world_size, rank)
-    extern_path = f"./{dist_name}_stub2_rank{rank}.pb"
-    if os.path.exists(extern_path):
-        os.remove(extern_path)
-    convert_model_to_external_data(
-        model2,
-        all_tensors_to_one_file=True,
-        location=extern_path,
-        size_threshold=1024,
-        convert_attribute=False,
-    )    
-    onnx.save(model2, f"./{dist_name}_stub2_rank{rank}.onnx")
-    runtime2 = backend.CudaRuntime(local_rank)
-    # print("init comm")
-    runtime2.init_comm(
-        dist_name,
-        world_size,
-        rank,
-    )    
-    
-    # run the two graphs
-    run_and_compare(name, model1, model2, runtime1, runtime2)
-
-
-def start_single(name, model1, model2):
-    runtime1 = backend.CudaRuntime(0)
-    runtime2 = backend.CudaRuntime(0)
-    run_and_compare(name, model1, model2, runtime1, runtime2)
-
-
-def gen_standard(name, model1, model2, voc_size, bs, len):
-    # generate standard results
-    data1 = np.random.randint(0, voc_size, (bs, len), dtype=np.int32)
-    data2 = np.random.randint(0, voc_size, (bs, len), dtype=np.int32)
-    np.save(f"{name}_inputs1", data1)
-    np.save(f"{name}_inputs2", data2)
-    runtime1 = backend.CudaRuntime(0)
-    runtime2 = backend.CudaRuntime(0)
-    outputs = run_model(model1, model2, runtime1, runtime2, data1, data2, 1)
-    np.save(f"{name}_results", outputs)
-
-
-def main():
-    nnodes, nproc_per_node, name, model1_path, model2_path, bs, length, gen_std = parse_args()
-
-    model1 = onnx.load(model1_path)
-    model2 = onnx.load(model2_path)
-
-    # generate standart output
-    if gen_std:
-        print(f"generate standard data for {name}.")
-        # a small vocabulary size to fit all LLM.
-        voc_size = 1000
-        gen_standard(name, model1, model2, voc_size, bs, length)
-        return
-
-    # run single process.
-    # use standalone process to isolate cuda.
-    p = mp.Process(target=start_single, args=(name, model1, model2))
-    p.start()
-    p.join()
-
-    # run distributed parallel.
-    world_size = nnodes * nproc_per_node
-    workers = [
-        mp.Process(
-            target=start_worker,
-            args=(name, world_size, rank, rank % nproc_per_node, model1, model2),
-        )
-        for rank in range(world_size)
-    ]
-
-    for w in workers:
-        w.start()
-
-    for w in workers:
-        w.join()
-
-
-if __name__ == "__main__":
-    main()
--- a/examples/distributed/cuda/run_pytorch.py
+++ b/examples/distributed/cuda/run_pytorch.py
@ -1,188 +0,0 @@
-import argparse
-import torch
-from transformers import BertModel, BertConfig
-from transformers import GPT2Model, GPT2Config
-from transformers import OPTModel, OPTConfig
-import time
-import numpy as np
-import onnx
-import os
-from onnx.external_data_helper import convert_model_to_external_data
-from onnxsim import simplify
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Run pytorch gpt2/bert/opt and optionally export onnx.")
-    parser.add_argument(
-        "--model", type=str, choices=["gpt2", "bert", "opt"], required=True, help="model type"
-    )
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
-    parser.add_argument("--length", type=int, default=1, help="sequence length.")
-    parser.add_argument(
-        "--export_onnx",
-        type=str,
-        nargs="?",
-        default=None,
-        const="./",
-        help="whether and where to export onnx file",
-    )
-    parser.add_argument(
-        "--type", type=str, choices=["fp32", "fp16", "tf32"], default="fp32", help="data type"
-    )
-    args = parser.parse_args()
-    print("arg setting: ", args)
-    return (
-        args.model,
-        args.batch_size,
-        args.length,
-        args.export_onnx,
-        args.type,
-    )
-
-
-def get_model(modelname):
-    match modelname:
-        case "bert":
-            model = BertModel.from_pretrained("bert-base-uncased", add_pooling_layer=False, hidden_act="gelu_new") # erf is not impl by infini
-            voc_size = BertConfig().vocab_size
-        case "gpt2":
-            model = GPT2Model.from_pretrained("gpt2")
-            voc_size = GPT2Config().vocab_size
-        case "opt":
-            model = model = OPTModel.from_pretrained("./opt-125m")
-            voc_size = OPTConfig().vocab_size
-        case _:
-            raise KeyError(modelname)
-
-    model = model.eval()
-    return model, voc_size
-
-def run_pytorch(torch_model, voc_size, batchsize, len):
-    data = np.random.randint(0, voc_size, (batchsize, len), dtype=np.int32)
-    np.save("test_inputs", data)
-    inputs = torch.from_numpy(data).to("cuda")
-    torch_model = torch_model.to("cuda")
-
-    n_iter = 20
-    with torch.no_grad():
-        for _ in range(10):
-            outputs = torch_model(inputs)
-    torch.cuda.synchronize()
-    begin = time.time()
-    with torch.no_grad():
-        for _ in range(n_iter):
-            torch.cuda.synchronize()
-            outputs = torch_model(inputs)
-            # 
-            torch.cuda.synchronize()
-    torch.cuda.synchronize()
-    end = time.time()
-    
-    avg_time = (end - begin) / n_iter
-    outputs = outputs.last_hidden_state.to("cpu")
-    print("outputs abs mean:", abs(np.array(outputs)).mean())
-    print(f"average time: {avg_time}")
-    torch.cuda.memory.empty_cache()
-    np.save("test_results", np.array(outputs, dtype=np.float32))
-    print("Save input & output as test_inputs.npy and test_results.npy")
-
-
-def export_onnx(model, data, path, extern=False):
-    torch.onnx.export(model, data, path, verbose=False, do_constant_folding=True)
-    onnx_model = onnx.load(path)
-    onnx_model, check = simplify(onnx_model, skipped_optimizers=['eliminate_duplicate_initializer'])
-    #onnx_model, check = simplify(onnx_model, skipped_optimizers=['fuse_qkv', 'eliminate_duplicate_initializer'])
-    assert check
-    add_value_info_for_constants(onnx_model)
-    onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
-    if extern:
-        extern_path = path.replace('.onnx', '.pb')
-        if os.path.exists(extern_path):
-            os.remove(extern_path)
-        convert_model_to_external_data(
-            onnx_model,
-            all_tensors_to_one_file=True,
-            location=extern_path,
-            size_threshold=1024,
-            convert_attribute=False,
-        )
-    onnx.save(onnx_model, path)
-
-def add_value_info_for_constants(model : onnx.ModelProto):
-    """
-    Currently onnx.shape_inference doesn't use the shape of initializers, so add
-    that info explicitly as ValueInfoProtos.
-    Mutates the model.
-    Args:
-        model: The ModelProto to update.
-    """
-    # All (top-level) constants will have ValueInfos before IRv4 as they are all inputs
-    if model.ir_version < 4:
-        return
-
-    def add_const_value_infos_to_graph(graph : onnx.GraphProto):
-        inputs = {i.name for i in graph.input}
-        existing_info = {vi.name: vi for vi in graph.value_info}
-        for init in graph.initializer:
-            # Check it really is a constant, not an input
-            if init.name in inputs:
-                continue
-
-            # The details we want to add
-            elem_type = init.data_type
-            shape = init.dims
-
-            # Get existing or create new value info for this constant
-            vi = existing_info.get(init.name)
-            if vi is None:
-                vi = graph.value_info.add()
-                vi.name = init.name
-
-            # Even though it would be weird, we will not overwrite info even if it doesn't match
-            tt = vi.type.tensor_type
-            if tt.elem_type == onnx.TensorProto.UNDEFINED:
-                tt.elem_type = elem_type
-            if not tt.HasField("shape"):
-                # Ensure we set an empty list if the const is scalar (zero dims)
-                tt.shape.dim.extend([])
-                for dim in shape:
-                    tt.shape.dim.add().dim_value = dim
-
-        # Handle subgraphs
-        for node in graph.node:
-            for attr in node.attribute:
-                # Ref attrs refer to other attrs, so we don't need to do anything
-                if attr.ref_attr_name != "":
-                    continue
-
-                if attr.type == onnx.AttributeProto.GRAPH:
-                    add_const_value_infos_to_graph(attr.g)
-                if attr.type == onnx.AttributeProto.GRAPHS:
-                    for g in attr.graphs:
-                        add_const_value_infos_to_graph(g)
-
-
-    return add_const_value_infos_to_graph(model.graph)
-
-
-def main():
-    torch.backends.cuda.matmul.allow_tf32 = False
-    torch.backends.cudnn.allow_tf32 = False
-    modelname, batchsize, seqlen, export_path, data_type = parse_args()
-    if data_type == "tf32":
-        torch.backends.cuda.matmul.allow_tf32 = True
-    else:
-        os.environ["NVIDIA_TF32_OVERRIDE"] = "0"
-
-    model, voc_size = get_model(modelname)
-    if export_path is not None:
-        filename = "{}_{}_{}.onnx".format(modelname, batchsize, seqlen)
-        path = os.path.join(export_path, filename)
-        param = torch.zeros((batchsize, seqlen), dtype=torch.int)
-        export_onnx(model, param, path, True)
-
-    if data_type == "fp16":
-        model = model.half()
-    run_pytorch(model, voc_size, batchsize, seqlen)
-
-if __name__ == "__main__":
-    main()
--- a/examples/distributed/kunlun/export_onnx.sh
+++ b/examples/distributed/kunlun/export_onnx.sh
@ -1,14 +0,0 @@
- export HF_ENDPOINT=https://hf-mirror.com
-
-models=("bert" "gpt2" "llama")
-batch_size=(1 32)
-seq_len=(100 500)
-nproc=(1 2 4)
-
-for model in "${models[@]}"; do
-    for bs in "${batch_size[@]}"; do
-        for len in "${seq_len[@]}"; do
-            python run_pytorch.py --model "$model" --batch_size "$bs" --length "$len" --export_onnx ../models/"$model" --export_only 
-        done
-    done
-done 
--- a/examples/distributed/kunlun/kunlun_launch.py
+++ b/examples/distributed/kunlun/kunlun_launch.py
@ -1,280 +0,0 @@
-import sys
-sys.path.append('../')
-
-import argparse
-import os
-import time
-import multiprocessing as mp
-from pyinfinitensor.onnx import OnnxStub, backend
-import onnx
-from onnx.external_data_helper import convert_model_to_external_data
-from onnx.shape_inference import infer_shapes_path
-import numpy as np
-from parallel_opt import parallel_model
-from functools import wraps
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="launch distributed infinitensor")
-    parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
-    parser.add_argument(
-        "--nproc_per_node", type=int, default=2, help="number of processes per node"
-    )
-    parser.add_argument(
-        "--name", type=str, choices=["gpt2", "bert", "llama"], help="name of model."
-    )
-    parser.add_argument(
-        "--model", type=str, default="", help="path to the ONNX model file."
-    )
-    parser.add_argument(
-        "--gen_std",
-        default=False,
-        action="store_true",
-        help="whether to generate the standard results.",
-    )
-    parser.add_argument(
-        "--run_single",
-        default=False,
-        action="store_true",
-        help="whether run model with single process with standard inputs"
-    )
-    parser.add_argument(
-        "--input_dir",
-        default="./",
-        help="path to save model input data"
-    )
-    parser.add_argument(
-        "--result_dir",
-        default="./",
-        help="path to save model standard output"
-    )
-    parser.add_argument(
-        "--internal_model_dir",
-        default="./",
-        help="path to save internal onnx model for parallel run"
-    )
-    args = parser.parse_args()
-
-    # check path, mkdir if not exist
-    check_exists(args.input_dir)
-    check_exists(args.result_dir)
-    check_exists(args.internal_model_dir)
-
-    print("arg setting: ", args)
-    return (
-        args.num_nodes,
-        args.nproc_per_node,
-        args.name,
-        args.model,
-        args.gen_std,
-        args.run_single,
-        args.input_dir,
-        args.result_dir,
-        args.internal_model_dir
-    )
-
-
-"""
-utils function for this scripts
-"""
-def check_exists(path: str):
-    if not os.path.exists(path):
-        os.makedirs(path)
-
-def np_assert(base, test, rtol=1e-2, atol=1e-1):
-    # np.testing.assert_allclose(test, base, rtol, atol)
-    print("max abs diff:", abs(base - test).max())
-
-
-"""
-Perf wrapper, run function n times
-then average
-"""
-def perf_it(n):
-    def decorator(func):
-        @wraps(func)
-        def wrapper(*args, **kwargs):
-            # warmup
-            for _ in range(n):
-                func(*args, **kwargs)
-
-            t_total = 0
-            for _ in range(n):
-                t0 = time.time()
-                func(*args, **kwargs)
-                t1 = time.time()
-                t_total += t1 - t0
-            avg_time = (t_total) / n
-            print(f"Avg runtime of {n} time is {avg_time:.6f} seconds")
-            return avg_time
-        return wrapper
-    return decorator
-
-
-"""
-Run InfiniTensor model with Standard input
-check=True: check with standard output gen by pytorch
-perf=True: run n times to get avg time
-"""
-def run_model(task_name,
-              model,
-              runtime,
-              world_size=1,
-              rank=0,
-              n=10,
-              check=True,
-              perf=True):
-
-    stub = OnnxStub(model, runtime,
-                    use_naive_allocator=True \
-                    if task_name == "llama" else False)
-
-    # load in Onnx model inputs
-    def load_inputs(stub: OnnxStub):
-        # check exists
-        inputs = []
-        for i, (name, tensor) in enumerate(stub.inputs.items()):
-            input_path = os.path.join(input_dir, \
-                                f"{task_name}_input_{i}.npy")
-            print(input_path)
-            if os.path.exists(input_path):
-                input = np.load(input_path)
-            else :
-                raise KeyError(f"{i} th input of model not exists")
-            # check shape
-            if all(x == y for x,y in zip(input.shape, tensor.shape())):
-                tensor.copyin_numpy(input)
-            else:
-                tensor.copyin_numpy(np.hsplit(input, world_size)[rank])
-
-    load_inputs(stub)
-    # stub.tune()
-    stub.run()
-    time.sleep(0.01)
-    output = next(stub.outputs.values().__iter__()).copyout_numpy()
-
-    # check output results with standard output
-    if check:
-        st_output_path = os.path.join(result_dir, \
-                                f"{task_name}_output.npy")
-        assert os.path.exists(st_output_path) , \
-                    "standard output not exists"
-        st_output = np.load(st_output_path)
-        if np.isnan(output).any():
-            print("Nan in output")
-            exit()
-        np_assert(st_output, output)
-
-    # perf
-    if perf:
-        @perf_it(n)
-        def perf_infinitensor(stub: OnnxStub):
-            stub.run()
-        perf_infinitensor(stub)
-
-    return output
-
-
-"""
-Start a worker in Parallel
-"""
-def start_worker(name: str,
-           world_size: int,
-           rank: int,
-           local_rank: int,
-           model: onnx.ModelProto):
-
-    dist_name = name + "_dist"
-    # partial a onnx model to world_size part
-    model = parallel_model(model, world_size, rank)
-    onnx.save(model, os.path.join(internal_model_dir, \
-                                    f"{dist_name}_rank{rank}.onnx"), save_as_external_data=True)
-    runtime = backend.KUNLUNRuntime(local_rank)
-    # print("init comm")
-    runtime.init_comm(
-        dist_name,
-        world_size,
-        rank,
-    )
-    run_model(name, model, runtime, world_size, rank)
-
-
-"""
-generate standard input/output with
-sigle card run
-"""
-def gen_standard(task_name: str, model: onnx.ModelProto):
-    runtime = backend.KUNLUNRuntime(0)
-    stub = OnnxStub(model, runtime)
-    position_id = 0
-    # generate random input for model
-    for i, (name, tensor) in enumerate(stub.inputs.items()):
-        input = tensor.copyout_numpy()
-        if np.issubdtype(input.dtype, np.integer):
-            if input.size == 1:
-                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
-            else:
-                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
-        elif input.dtype == np.bool_:
-            input = np.random.randint(0,2,size=input.shape) > 0
-        else:
-            if i == 0:
-                input = np.ones(input.shape).astype(input.dtype)
-                position_id = input.shape[-1] - 1
-            else:
-                input = np.random.rand(*input.shape).astype(input.dtype)
-        tensor.copyin_numpy(input)
-        np.save(os.path.join(input_dir, \
-                    f"{task_name}_input_{i}.npy"), input)
-    stub.run()
-    # print(stub.outputs)
-    output = next(stub.outputs.values().__iter__()).copyout_numpy()
-    if np.isnan(output).any():
-        print("Nan in output")
-        exit()
-    np.save(os.path.join(result_dir, f"{task_name}_output.npy"), output)
-
-
-def main():
-
-    global input_dir, result_dir, internal_model_dir
-
-    nnodes, nproc_per_node, task_name, \
-        model_path, gen_std, run_single, \
-            input_dir, result_dir, internal_model_dir = parse_args()
-
-    # load input onnx model
-    model = onnx.load(model_path)
-
-    # generate standart output
-    if gen_std:
-        print("Generate inputs and outputs.")
-        gen_standard(task_name, model)
-        return
-
-    if run_single:
-        print("Run model by one GPU card.")
-        runtime = backend.KUNLUNRuntime(0)
-        run_model(task_name, model, runtime)
-        return
-
-    # run distributed parallel.
-    world_size = nnodes * nproc_per_node
-    print(f"Run model by {world_size} GPU in parallel.")
-    workers = [
-        mp.Process(
-            target=start_worker,
-            args=(task_name, world_size, rank, rank % nproc_per_node, model),
-        )
-        for rank in range(world_size)
-    ]
-
-    for w in workers:
-        w.start()
-
-    for w in workers:
-        w.join()
-
-
-if __name__ == "__main__":
-    main()
--- a/examples/distributed/kunlun/launch.sh
+++ b/examples/distributed/kunlun/launch.sh
@ -1,36 +0,0 @@
-export HF_ENDPOINT=https://hf-mirror.com
-
-# models=("bert" "gpt2" "llama")
-models=("bert" "gpt2")
-batch_size=(1 32)
-seq_len=(100 500)
-nproc=(1 2 4)
-
-results_dir="results"
-
-if [ -d "$results_dir" ]; then
-    echo "directory ./$results_dir exists"
-else
-    mkdir -p "$results_dir"
-    echo "mkdir $results_dir, logs saved there"
-fi
-
-
-for model in "${models[@]}"; do
-    for bs in "${batch_size[@]}"; do
-        for len in "${seq_len[@]}"; do
-            # run pytorch model
-            echo "Run pytorch $model with batch_size=$bs length=$len ."
-            python run_pytorch.py --model "$model" --batch_size "$bs" --length "$len" #> results/"$model"_"$bs"_"$len"_pytorch
-            for n in "${nproc[@]}"; do
-                # run infinitensor 
-                echo "Run $n parallel infinitensor "$model" with batch_size=$bs and length=$len ."
-                python kunlun_launch.py --name "$model" --model ../models/"$model"/"$model"_"$bs"_"$len".onnx --nproc_per_node=$n # >> results/"$model"_"$bs"_"$len"_infini 
-                # delete internal files
-                find ./ -type f -name "*.onnx" -delete
-                find ./ -type f -name "*.pb" -delete
-            done
-            find ./ -type f -name "*.npy" -delete
-        done
-    done
-done
--- a/examples/distributed/kunlun/llama_launch.sh
+++ b/examples/distributed/kunlun/llama_launch.sh
@ -1,35 +0,0 @@
-export HF_ENDPOINT=https://hf-mirror.com
-
-# models=("bert" "gpt2" "llama")
-models=("llama")
-batch_size=(1 )
-seq_len=(100 500)
-nproc=(1 2 4)
-
-results_dir="results"
-
-if [ -d "$results_dir" ]; then
-    echo "directory ./$results_dir exists"
-else
-    mkdir -p "$results_dir"
-    echo "mkdir $results_dir, logs saved there"
-fi
-
-
-for model in "${models[@]}"; do
-    for bs in "${batch_size[@]}"; do
-        for len in "${seq_len[@]}"; do
-            echo "Run pytorch llama with batch_size="$bs" and length="$len""
-            python run_pytorch.py --model "$model" --batch_size "$bs" --length "$len"
-            for n in "${nproc[@]}"; do
-                    # run pytorch model
-                    echo "Run infinitensor llama with batch_size="$bs" and length="$len" and nproc="$n"."
-                    python kunlun_launch.py --name llama --model ../models/llama/llama_"$bs"_"$len"_fp32.onnx --nproc_per_node=$n
-                    # delete internal files
-                    find ./ -type f -name "*.onnx" -delete
-                    find ./ -type f -name "*0c" -delete
-            done
-            find ./ -type f -name "*.npy" -delete
-        done
-    done
-done
--- a/examples/distributed/kunlun/run_pytorch.py
+++ b/examples/distributed/kunlun/run_pytorch.py
@ -1,245 +0,0 @@
-import argparse
-import torch
-from transformers import BertModel, BertConfig
-from transformers import GPT2Model, GPT2Config
-from transformers import OPTModel, OPTConfig
-from transformers import LlamaModel, LlamaConfig
-import time
-import numpy as np
-import onnx
-import os
-import sys
-from onnx.external_data_helper import convert_model_to_external_data
-from onnxsim import simplify
-
-torch.backends.cuda.matmul.allow_tf32 = False
-torch.backends.cudnn.allow_tf32 = False
-def parse_args():
-    parser = argparse.ArgumentParser(description="Run pytorch gpt2/bert/opt and optionally export onnx.")
-    parser.add_argument(
-        "--model", type=str, choices=["gpt2", "bert", "opt", "llama"], required=True, help="model type"
-    )
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
-    parser.add_argument("--length", type=int, default=1, help="sequence length.")
-    parser.add_argument(
-        "--export_onnx",
-        type=str,
-        nargs="?",
-        default=None,
-        const="./",
-        help="whether and where to export onnx file",
-    )
-    parser.add_argument(
-        "--input_dir",
-        type=str,
-        default="./",
-        help="path to save pytorch model input data"
-    )
-    parser.add_argument(
-        "--result_dir",
-        type=str,
-        default="./",
-        help="path to save pytorch model output data"
-    )
-    parser.add_argument(
-        "--export_only",
-        action="store_true"
-    )
-    args = parser.parse_args()
-    print("arg setting: ", args)
-    return (
-        args.model,
-        args.batch_size,
-        args.length,
-        args.export_onnx,
-        args.input_dir,
-        args.result_dir,
-        args.export_only
-    )
-
-
-def get_model(modelname):
-    if modelname == "bert":
-        model = BertModel.from_pretrained("bert-base-uncased", add_pooling_layer=False, hidden_act="gelu_new") # erf is not impl by infini
-        voc_size = BertConfig().vocab_size
-    elif modelname == "gpt2":
-        model = GPT2Model.from_pretrained("gpt2")
-        voc_size = GPT2Config().vocab_size
-    elif modelname == "opt":
-        model = OPTModel.from_pretrained("./opt-125m")
-        voc_size = OPTConfig().vocab_size
-    elif modelname == "llama":
-        model = LlamaModel.from_pretrained("meta-llama/Llama-2-7b-hf")
-        voc_size = LlamaConfig().vocab_size
-    else :
-        raise KeyError(modelname)
-
-    model = model.eval()
-    return model, voc_size
-
-def run_pytorch(torch_model, voc_size, batchsize, len, model_name):
-    data = np.random.randint(0, voc_size, (batchsize, len), dtype=np.int32)
-    np.save(os.path.join(input_dir, f"{model_name}_input_0.npy"), data)
-    inputs = torch.from_numpy(data).to("cuda")
-    torch_model = torch_model.to("cuda")
-
-    n_iter = 10
-    with torch.no_grad():
-        for _ in range(10):
-            outputs = torch_model(inputs)
-    torch.cuda.synchronize()
-    begin = time.time()
-    with torch.no_grad():
-        for _ in range(n_iter):
-            torch.cuda.synchronize()
-            outputs = torch_model(inputs)
-            #
-            torch.cuda.synchronize()
-    torch.cuda.synchronize()
-    end = time.time()
-
-    avg_time = (end - begin) / n_iter
-    outputs = outputs.last_hidden_state.to("cpu")
-    print("outputs abs mean:", abs(np.array(outputs)).mean())
-    print(f"average time: {avg_time}")
-    torch.cuda.memory.empty_cache()
-    np.save(os.path.join(result_dir, f"{model_name}_output.npy"), \
-                                        np.array(outputs))
-    print(f"Save input & output as {model_name}_input_0.npy and {model_name}_output.npy")
-
-
-def export_onnx(model_name, model, data, path, extern=False):
-    # torch.onnx.export(model, data, path, verbose=False, do_constant_folding=True)
-
-    if model_name != "llama":
-        onnx_model = onnx.load(path)
-        onnx_model, check = simplify(onnx_model,
-                                 skipped_optimizers=['fuse_qkv', 'eliminate_duplicate_initializer'])
-                                 # skipped_optimizers=['fuse_qkv'])
-        assert check
-        add_value_info_for_constants(onnx_model)
-        onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
-        if extern:
-            extern_path = path.replace('.onnx', '.pb')
-            if os.path.exists(extern_path):
-                os.remove(extern_path)
-            convert_model_to_external_data(
-                onnx_model,
-                all_tensors_to_one_file=True,
-                location=extern_path.split("/")[-1],
-                size_threshold=1024,
-                convert_attribute=False,
-            )
-        onnx.save(onnx_model, path)
-    else:
-        sys.path.append("onnxsim_large_model")
-        from onnx_utils import set_onnx_input_shape
-        from compress_model import SIZE_1MB, compress_onnx_model, uncompress_onnx_model
-
-        in_model_path = path
-        out_model_path = in_model_path[:-5] + ".sim.onnx"
-
-        onnx_model = onnx.load(in_model_path)
-        print(f"load model from {in_model_path} success")
-
-        size_th_bytes = 1024 * 1024
-        onnx_model, removed_inits = compress_onnx_model(onnx_model, size_th_bytes=size_th_bytes)
-        print("compress model success")
-
-        onnx_model = set_onnx_input_shape(onnx_model, "")
-        tensor_size_threshold = f"1024KB"
-        skipped_optimizers = []
-        skipped_optimizers.append("eliminate_duplicate_initializer")
-        onnx_model, check = simplify(onnx_model, skipped_optimizers=skipped_optimizers,
-                                    tensor_size_threshold=tensor_size_threshold)
-        if not check:
-            raise ValueError(f"simplify compressed model {in_model_path} failed")
-
-        print(f"simplify model success")
-
-        onnx_model = uncompress_onnx_model(onnx_model, removed_inits)
-        print(f"uncompress model success")
-
-        add_value_info_for_constants(onnx_model)
-
-        onnx.save(onnx_model, out_model_path, save_as_external_data=True)
-
-
-def add_value_info_for_constants(model : onnx.ModelProto):
-    """
-    Currently onnx.shape_inference doesn't use the shape of initializers, so add
-    that info explicitly as ValueInfoProtos.
-    Mutates the model.
-    Args:
-        model: The ModelProto to update.
-    """
-    # All (top-level) constants will have ValueInfos before IRv4 as they are all inputs
-    if model.ir_version < 4:
-        return
-
-    def add_const_value_infos_to_graph(graph : onnx.GraphProto):
-        inputs = {i.name for i in graph.input}
-        existing_info = {vi.name: vi for vi in graph.value_info}
-        for init in graph.initializer:
-            # Check it really is a constant, not an input
-            if init.name in inputs:
-                continue
-
-            # The details we want to add
-            elem_type = init.data_type
-            shape = init.dims
-
-            # Get existing or create new value info for this constant
-            vi = existing_info.get(init.name)
-            if vi is None:
-                vi = graph.value_info.add()
-                vi.name = init.name
-
-            # Even though it would be weird, we will not overwrite info even if it doesn't match
-            tt = vi.type.tensor_type
-            if tt.elem_type == onnx.TensorProto.UNDEFINED:
-                tt.elem_type = elem_type
-            if not tt.HasField("shape"):
-                # Ensure we set an empty list if the const is scalar (zero dims)
-                tt.shape.dim.extend([])
-                for dim in shape:
-                    tt.shape.dim.add().dim_value = dim
-
-        # Handle subgraphs
-        for node in graph.node:
-            for attr in node.attribute:
-                # Ref attrs refer to other attrs, so we don't need to do anything
-                if attr.ref_attr_name != "":
-                    continue
-
-                if attr.type == onnx.AttributeProto.GRAPH:
-                    add_const_value_infos_to_graph(attr.g)
-                if attr.type == onnx.AttributeProto.GRAPHS:
-                    for g in attr.graphs:
-                        add_const_value_infos_to_graph(g)
-
-
-    return add_const_value_infos_to_graph(model.graph)
-
-
-def main():
-    global input_dir, result_dir
-
-    modelname, batchsize, seqlen, \
-        export_path, input_dir, result_dir, export_only = parse_args()
-
-    model, voc_size = get_model(modelname) # pytorch model
-
-    if export_path is not None:
-        os.makedirs(export_path, exist_ok=True)
-        filename = "{}_{}_{}.onnx".format(modelname, batchsize, seqlen)
-        path = os.path.join(export_path, filename)
-        param = torch.zeros((batchsize, seqlen), dtype=torch.int)
-        export_onnx(modelname, model, param, path, True) # export pytorch model to onnx model
-        if export_only:
-            return
-
-    run_pytorch(model, voc_size, batchsize, seqlen, modelname)
-
-if __name__ == "__main__":
-    main()
--- a/examples/distributed/onnxsim_large_model
+++ b/examples/distributed/onnxsim_large_model
@ -1 +0,0 @@
-Subproject commit cbcf3fbf985a00494b0f136c92eaccd42031bf65
--- a/examples/distributed/parallel.py
+++ b/examples/distributed/parallel.py
@ -1,103 +0,0 @@
-import onnx
-from onnx import (
-    ModelProto,
-    TensorProto,
-    NodeProto,
-    AttributeProto,
-)
-from onnx import helper, numpy_helper
-from typing import Dict, Any
-
-
-def parse_attribute(node: NodeProto, attrs: Dict[str, Any] = dict()) -> Dict[str, Any]:
-    for attr in node.attribute:
-        if attr.name in attrs:
-            if attr.type == AttributeProto.INT:
-                attrs[attr.name] = attr.i
-            elif attr.type == AttributeProto.INTS:
-                attrs[attr.name] = attr.ints
-            elif attr.type == AttributeProto.FLOAT:
-                attrs[attr.name] = attr.f
-            elif attr.type == AttributeProto.STRING:
-                attrs[attr.name] = attr.s
-            elif attr.type == AttributeProto.TENSOR:
-                attrs[attr.name] = attr.t
-            else:
-                assert False, "Unsupported Attribute Type: {}".format(attr.type)
-    return attrs
-
-
-def parallel_model(model: ModelProto, tp_world_size: int = 1, tp_rank: int = 0):
-    data = {init.name: init for init in model.graph.initializer}
-    nodes = list(model.graph.node)
-
-    def shard_tensor(tensor: TensorProto, dim: int):
-        array = numpy_helper.to_array(tensor)
-        if dim >= array.ndim:
-            dim = array.ndim - 1
-        assert array.shape[dim] % tp_world_size == 0
-        seg = array.shape[dim] // tp_world_size
-        array = array[tp_rank * seg : (tp_rank + 1) * seg]
-        return numpy_helper.from_array(array, name=tensor.name + f":sharded({dim})")
-
-    def shard_gemm(node: NodeProto):
-        attrs = parse_attribute(
-            node, {"alpha": 1.0, "beta": 1.0, "transA": 0, "transB": 0}
-        )
-        trans = [attrs["transA"], attrs["transB"]]
-        dim = 0
-        for i, (input, t) in enumerate(zip(node.input, trans)):
-            if input in data:
-                dim = i
-                sharded = shard_tensor(data[input], dim ^ t)
-                node.input[i] = sharded.name
-                data[input] = sharded
-        if len(node.input) > 2:
-            input = node.input[2]
-            sharded = shard_tensor(data[input], dim)
-            node.input[2] = sharded.name
-            data[input] = sharded
-
-        node.output[0] += f":sharded({dim})"
-        return dim
-
-    for i, node in enumerate(nodes):
-        if node.op_type == "Gemm":
-            output = node.output[0]
-            dim = shard_gemm(node)
-            gathered = [node.output[0] + f".{i}" for i in range(tp_world_size)]
-            # all_gather
-            nodes.insert(
-                i + 1,
-                helper.make_node(
-                    op_type="AllGather",
-                    inputs=[node.output[0]],
-                    outputs=gathered,
-                    name=node.name + "/allgather",
-                    # domain="infini", # shape inference fails for custom domain
-                ),
-            )
-            # concat
-            nodes.insert(
-                i + 2,
-                helper.make_node(
-                    op_type="Concat",
-                    inputs=gathered,
-                    outputs=[output],
-                    name=node.name + "/concat",
-                    axis=dim,
-                ),
-            )
-    graph = helper.make_graph(
-        nodes,
-        model.graph.name + f"_{tp_rank}",
-        model.graph.input,
-        model.graph.output,
-        data.values(),
-        doc_string=model.graph.doc_string,
-        value_info=model.graph.value_info,
-    )
-    model = helper.make_model(graph)
-
-    onnx.shape_inference.infer_shapes(model)
-    return model
--- a/examples/distributed/parallel_opt.py
+++ b/examples/distributed/parallel_opt.py
@ -1,247 +0,0 @@
-import onnx
-from onnx import ModelProto, NodeProto, TensorProto, ValueInfoProto
-from onnx import helper, numpy_helper
-from typing import Dict, List
-from placement import Placement, Replicate, Shard, _Partial
-import numpy as np
-
-
-def parallel_model(model: ModelProto, tp_world_size: int = 1, tp_rank: int = 0):
-    data = {init.name: init for init in model.graph.initializer}
-    vinfo = {info.name: info for info in model.graph.value_info}
-    vinfo.update({info.name: info for info in model.graph.input})
-    vinfo.update({info.name: info for info in model.graph.output})
-    output = {info.name: info for info in model.graph.output}
-    place: Dict[str, Placement] = {}
-    nodes: List[NodeProto] = []
-
-    def is_sharded(name: str):
-        return place[name].is_shard()
-
-    def shard_tensor(tensor: TensorProto, plc: Shard, groups: int = 1):
-        # print(f"shard {tensor.name} at dim {dim}")
-        assert plc.is_shard(), plc
-        ndim = len(tensor.dims)
-        if plc.dim < 0:
-            plc.dim += ndim
-        if tensor.dims[plc.dim] == 1:  # broadcast dim, no need to shard.
-            return tensor
-        array = numpy_helper.to_array(tensor)
-        assert array.shape[plc.dim] % tp_world_size == 0, array.shape[plc.dim]
-        dims = list(tensor.dims)
-        dims.insert(plc.dim, groups)
-        dims[plc.dim + 1] //= groups
-        array = array.reshape(dims)
-        seg = array.shape[plc.dim + 1] // tp_world_size
-        array = array.take(
-            indices=range(tp_rank * seg, (tp_rank + 1) * seg), axis=plc.dim + 1
-        )
-        dims = list(tensor.dims)
-        dims[plc.dim] //= tp_world_size
-        array = array.reshape(dims)
-        tensor = numpy_helper.from_array(array, name=tensor.name)
-        place[tensor.name] = plc
-        return tensor
-
-    def shard_gemm(node: NodeProto, groups: int = 1):
-        # print("gemm", node.name)
-        in_plc = place[node.input[0]]
-        w_plc = Shard(-1) if in_plc.is_replicate() else Shard(0)
-        transB = next((attr.i for attr in node.attribute if attr.name == "transB"), 0)
-        if transB:
-            w_plc.dim = ~w_plc.dim
-        input = node.input[1]
-        data[input] = shard_tensor(data[input], w_plc, groups)
-
-        output = node.output[0]
-        ndim = len(vinfo[output].type.tensor_type.shape.dim)
-        out_plc = Shard(ndim - 1) if in_plc.is_replicate() else _Partial()
-        place[node.output[0]] = out_plc
-
-    def shard_concat(node: NodeProto):
-        # hack for kvcache
-        in_plc = place[node.input[1]]
-        if in_plc.is_shard():
-            seq_len_dim = vinfo[node.input[0]].type.tensor_type.shape.dim.pop(1)
-            seq_len_dim.dim_value //= tp_world_size
-            vinfo[node.input[0]].type.tensor_type.shape.dim.insert(1, seq_len_dim)
-            place[node.input[0]] = in_plc
-            place[node.output[0]] = in_plc
-
-    def shard_binary(node: NodeProto, groups: int = 1):
-        # print("binary", node.name, node.input[0], place[node.input[0]])
-        a = node.input[0]
-        b = node.input[1]
-        if a in data:
-            a, b = b, a
-        place[node.output[0]] = place[a]
-        if is_sharded(a) and b in data and len(data[b].dims) == 1:  # broadcast
-            data[b] = shard_tensor(data[b], Shard(0), groups)
-
-    def shard_reshape(node: NodeProto):
-        # print("reshape", node.name, node.input[0], place[node.input[0]])
-        if not is_sharded(node.input[0]):
-            return
-        in_plc = place[node.input[0]]
-        s_dim = -1
-        in_dims = [d.dim_value for d in vinfo[node.input[0]].type.tensor_type.shape.dim]
-        tensor = data[node.input[1]]
-        out_dims = numpy_helper.to_array(tensor).copy()
-        if len(in_dims) == 3 and len(out_dims) == 4:
-            if in_plc.dim == 0:
-                s_dim = 1
-            elif in_plc.dim == 2:
-                s_dim = 2
-        if len(in_dims) == 4 and len(out_dims) == 3:
-            if in_plc.dim == 1:
-                s_dim = 0
-            elif in_plc.dim == 2:
-                s_dim = 2
-        if len(in_dims) == 2 and len(out_dims) == 3:
-            if in_plc.dim == 1:
-                s_dim = 2
-        if len(in_dims) == 4 and len(out_dims) == 2:
-            if in_plc.dim == 1:
-                s_dim = 0
-            elif in_plc.dim == 2:
-                s_dim = 1
-        if len(in_dims) == 3 and len(out_dims) == 2:
-            if in_plc.dim == 1:
-                s_dim = 0
-            elif in_plc.dim == 2:
-                s_dim = 1
-        assert s_dim != -1
-        assert out_dims[s_dim] % tp_world_size == 0, out_dims
-        out_dims[s_dim] //= tp_world_size
-        # if ONNX uses the same tensor for multiple Reshape Nodes, then rename it to distingush from others.
-        node.input[1] = node.output[0] + "_shape"
-        data[node.input[1]] = numpy_helper.from_array(out_dims, name=node.input[1])
-        place[node.output[0]] = Shard(s_dim)
-
-    def shard_split(node: NodeProto):
-        if not is_sharded(node.input[0]):
-            return
-        in_plc = place[node.input[0]]
-        split_tensor = data[node.input[1]]
-        split = numpy_helper.to_array(split_tensor).copy()
-        split //= tp_world_size
-        data[node.input[1]] = numpy_helper.from_array(split, name=node.input[1])
-        for output in node.output:
-            place[output] = in_plc
-
-    def shard_transpose(node: NodeProto):
-        plc = place[node.input[0]]
-        if plc.is_shard():
-            perm = next(attr.ints for attr in node.attribute if attr.name == "perm")
-            place[node.output[0]] = Shard(list(perm).index(plc.dim))
-
-    def shard_node(node: NodeProto):
-        if node.op_type in ["Relu", "Tanh", "Softmax", "Cast"]:
-            place[node.output[0]] = place[node.input[0]]
-        elif node.op_type in ["Where"]:
-            place[node.output[0]] = place[node.input[1]]
-        if node.op_type in {"Add", "Mul", "Div", "Max"}:
-            shard_binary(node)
-        elif node.op_type == "Reshape":
-            shard_reshape(node)
-        elif node.op_type == "Transpose":
-            shard_transpose(node)
-        elif node.op_type == "Split":
-            shard_split(node)
-        elif node.op_type == "MatMul":
-            assert (
-                place[node.input[0]] == place[node.input[1]]
-            ), f"{place[node.input[0]]} != {place[node.input[1]]}"
-            place[node.output[0]] = place[node.input[0]]
-        elif node.op_type == "Concat":
-            shard_concat(node)
-
-    def find_successor(op_type: str, idx: int, search_limit: int = 1):
-        for node in model.graph.node[idx + 1 : idx + 1 + search_limit]:
-            if node.op_type == op_type:
-                return node
-        return None
-
-    # all tensors are initially replicated.
-    for v in vinfo:
-        place[v] = Replicate()
-
-    for t in data:
-        place[t] = Replicate()
-
-    for index, node in enumerate(model.graph.node):
-        nodes.append(node)
-        # linear
-        if (node.op_type == "MatMul" or node.op_type == "Gemm") and any(
-            input in data for input in node.input
-        ):
-            # FIXME(constroy): the last MatMul should not be sharded as TP.
-            if (
-                node.output[0] in output
-                or (
-                    index + 1 < len(model.graph.node)
-                    and model.graph.node[index + 1].output[0]
-                )
-                in output
-            ):
-                continue
-            groups = 1
-            # If the Gemm or Matmul is followed by a split, then the inputs are concatinated by groups
-            split_node = find_successor("Split", index, search_limit=2)
-            if split_node is not None:
-                groups = len(split_node.output)
-            shard_gemm(node, groups)
-            plc = place[node.output[0]]
-            if plc.is_partial():
-                new_name = node.output[0] + f":{plc}"
-                place[new_name] = place[node.output[0]]
-                # insert all_reduce
-                nodes.append(
-                    helper.make_node(
-                        op_type="ReduceSum",
-                        inputs=[new_name],
-                        outputs=[node.output[0]],
-                        name=node.name + "/all_reduce",
-                        noop_with_empty_axes=1,
-                        communicator=0,  # hack to treat ReduceSum as AllReduceSum
-                    )
-                )
-                place[node.output[0]] = Replicate()
-                node.output[0] = new_name
-            if len(node.input) > 2:  # split bias to add
-                prev = nodes[-1]
-                new_name = prev.output[0] + "_no_bias"
-                place[new_name] = place[node.output[0]]
-                bias = helper.make_node(
-                    op_type="Add",
-                    inputs=[new_name, node.input[2]],
-                    outputs=[prev.output[0]],
-                    name=node.name + "/bias",
-                )
-                node.input.pop()
-                prev.output[0] = new_name
-                shard_binary(bias, groups)
-                nodes.append(bias)
-            continue
-        shard_node(node)
-
-    new_input = []
-    for info in model.graph.input:
-        new_input.append(vinfo[info.name])
-
-    graph = helper.make_graph(
-        nodes,
-        model.graph.name + f"_{tp_rank}",
-        new_input,
-        model.graph.output,
-        data.values(),
-        doc_string=model.graph.doc_string,
-        # value_info=vinfo.values(),
-    )
-    for output in graph.output:
-        tt = output.type.tensor_type
-        if tt.HasField("shape"):
-            tt.ClearField("shape")
-    model = helper.make_model(graph)
-    #model = onnx.shape_inference.infer_shapes(model)
-    return model
--- a/examples/distributed/placement.py
+++ b/examples/distributed/placement.py
@ -1,64 +0,0 @@
-from typing import Optional
-
-
-class Placement:
-    # base class Placement type
-
-    # convenient utils to check for placement types
-    def is_shard(self, dim: Optional[int] = None) -> bool:
-        if dim is not None and isinstance(self, Shard):
-            return self.dim == dim
-        else:
-            return isinstance(self, Shard)
-
-    def is_replicate(self) -> bool:
-        return isinstance(self, Replicate)
-
-    def is_partial(self) -> bool:
-        return isinstance(self, _Partial)
-
-
-class Replicate(Placement):
-    def __eq__(self, other: object) -> bool:
-        if not isinstance(other, Replicate):
-            return False
-        return True
-
-    def __repr__(self) -> str:
-        """
-        machine readable representation of the Replicate placement
-        """
-        return "Replicate()"
-
-
-class Shard(Placement):
-    # shard placement, shard on a dim
-    def __init__(self, dim):
-        self.dim = dim
-
-    def __eq__(self, other: object) -> bool:
-        if not isinstance(other, Shard):
-            return False
-        return self.dim == other.dim
-
-    def __repr__(self) -> str:
-        """
-        machine readable representation of the Shard placement
-        """
-        return f"Shard(dim={self.dim})"
-
-
-class _Partial(Placement):
-    def __init__(self, reduce_op: str = "sum"):
-        self.reduce_op: str = reduce_op
-
-    def __eq__(self, other: object) -> bool:
-        if not isinstance(other, _Partial):
-            return False
-        return self.reduce_op == other.reduce_op
-
-    def __repr__(self) -> str:
-        """
-        machine readable representation of the Partial placement
-        """
-        return f"_Partial(reduce_op={self.reduce_op})"
--- a/examples/python/llama_kvcache_inference.py
+++ b/examples/python/llama_kvcache_inference.py
@ -1,145 +0,0 @@
-import os
-from pyinfinitensor.onnx import OnnxStub, backend
-import numpy as np
-import onnx
-import torch
-from transformers import LlamaModel, LlamaForCausalLM
-from tqdm import tqdm
-import onnx_graphsurgeon as gs
-from onnxsim import simplify
-import argparse
-
-parser = argparse.ArgumentParser(description='')
-parser.add_argument('--batchsize', dest='batchsize', type=int, default=1)
-parser.add_argument('--layer', dest='n_layers', type=int, default=2)
-parser.add_argument('--iter', dest='n_iter', type=int, default=1)
-parser.add_argument('--n_max_length', dest='n_max_length', type=int, default=1024)
-parser.add_argument('--pretrained_llama_path', dest='pretrained_llama_path', type=str, 
-                    default="/data0/shared/data/public/opensource_models/meta-llama/Llama-2-7b-hf/")
-parser.add_argument('--onnx_model_path', dest='onnx_model_path', type=str, 
-                    default="/data1/shared/llama")
-args = parser.parse_args()
-
-ONNX_MODEL_PATH = "{}/llama_bs{}_layer{}.onnx".format(args.onnx_model_path, args.batchsize, args.n_layers)
-ONNX_WEIGHT_PATH = "./llama_bs{}_layer{}.pb".format(args.batchsize, args.n_layers)
-
-def export_onnx(model: LlamaModel, ONNX_MODEL_PATH):
-    param = torch.zeros(
-        (args.batchsize, 1024), dtype=torch.long)
-    logits = model(param, past_key_values=None)
-    param_kvcache = torch.zeros((args.batchsize, 1), dtype=torch.long)
-
-    torch.onnx.export(model, (param_kvcache, {"past_key_values": logits.past_key_values,
-                                              "position_ids": param_kvcache}), ONNX_MODEL_PATH, verbose=False,
-                      do_constant_folding=True,)
-    onnx_model = onnx.load(ONNX_MODEL_PATH)
-    print("simplifing onnx model")
-    onnx_model, check = simplify(onnx_model, skipped_optimizers=[
-                                 'eliminate_duplicate_initializer'])
-    assert check
-    
-    onnx.save(onnx_model, ONNX_MODEL_PATH, save_as_external_data=True, location=ONNX_WEIGHT_PATH)
-    print("simlifing finished.")
-
-
-@gs.Graph.register()
-def replace_with_attention(self, inputs, outputs, inputs_added, outputs_removed):
-    for inp in inputs:
-        inp.outputs.clear()   
-    for out in outputs:
-        out.inputs.clear()
-    for inp in inputs_added:
-        inputs.append(inp)
-    for out in outputs_removed:
-        out.inputs.clear()
-    return self.layer(op="AttentionKVCache", inputs=inputs, outputs=outputs)
-
-
-def replace_onnx_with_attention_op():
-    graph = gs.import_onnx(
-        onnx.load(ONNX_MODEL_PATH))
-    tmap = graph.tensors()
-    for i in range(args.n_layers):
-        inputs = [
-            tmap["onnx::Concat_" + str((i+1)*2)],
-            tmap["onnx::Concat_" + str((i+1)*2+1)],
-            tmap["/model/layers." + str(i) + "/self_attn/Add_output_0"],
-            tmap["/model/layers." + str(i) + "/self_attn/Add_1_output_0"],
-            tmap["/model/layers." + str(i) + "/self_attn/Transpose_2_output_0"]]
-        outputs = [
-            tmap["/model/layers." + str(i) + "/self_attn/MatMul_1_output_0"]]
-
-        inputs_added = [graph.inputs[1]]
-        outputs_removed = []
-
-        graph.replace_with_attention(
-            inputs, outputs, inputs_added, outputs_removed)
-        
-    graph.outputs = [tmap[graph.outputs[0].name]]
-    graph.cleanup(True).toposort()
-    onnx.save(gs.export_onnx(graph), ONNX_MODEL_PATH, save_as_external_data=True)
-
-
-if __name__ == "__main__":
-    kvcache_torch = None
-    torch_model = LlamaForCausalLM.from_pretrained(
-        args.pretrained_llama_path, num_hidden_layers=int(args.n_layers)).eval()
-    
-    n_heads = torch_model.config.num_attention_heads
-    n_dims = torch_model.config.hidden_size // n_heads
-    
-    if not os.path.exists(ONNX_MODEL_PATH):
-        print("exporting onnx graph")
-        export_onnx(torch_model, ONNX_MODEL_PATH)
-        replace_onnx_with_attention_op()
-    else:
-        print("will use exsiting onnx graph")
-
-    onnx_model = onnx.load(ONNX_MODEL_PATH)
-    stub = OnnxStub(onnx_model, backend.cuda_runtime())
-
-    count_wrong = 0
-    for i in tqdm(range(0, args.n_max_length)):
-        query = np.random.randint(
-            torch_model.config.vocab_size, size=(args.batchsize, 1), dtype=np.int32)
-        position_id = i*np.ones((args.batchsize, 1), dtype=np.int32)
-
-        ####################################
-        # pytorch
-        ####################################
-        outputs_torch = torch_model(
-            torch.tensor(query), past_key_values=kvcache_torch)
-        logit_torch = outputs_torch['logits']
-        kvcache_torch = outputs_torch['past_key_values']
-
-        ####################################
-        # infinitensor
-        ####################################
-        # copyin input
-        (list(stub.inputs.items()))[0][1].copyin_int64(
-            query.reshape(-1).tolist())
-        (list(stub.inputs.items()))[1][1].copyin_int64(
-            position_id.reshape(-1).tolist())
-
-        stub.run()
-
-        ####################################
-        # validation
-        ####################################
-        # copyout output
-        logits_it = np.array((list(stub.outputs.items()))
-                                [0][1].copyout_float())
-        
-        try:
-            np.testing.assert_allclose(
-                logit_torch[:, -1, :].detach().cpu().numpy().flatten(), logits_it, rtol=1e-3, atol=1e-3)
-        except Exception as e: 
-            try:
-                np.testing.assert_allclose(
-                    np.argmax(logit_torch[:, -1, :].detach().cpu().numpy().flatten()), np.argmax(logits_it), rtol=1e-3, atol=1e-3)
-            except:
-                count_wrong = count_wrong + 1
-
-    result = "{}/{} failed.".format(count_wrong, args.n_max_length)
-    print(result)
-    del stub
--- a/examples/python/onnx_inference.py
+++ b/examples/python/onnx_inference.py
@ -1,29 +0,0 @@
-import sys
-import onnx
-import torch
-import numpy as np
-from pyinfinitensor.onnx import OnnxStub, backend
-
-if __name__ == '__main__':
-    args = sys.argv
-    if len(sys.argv) != 2:
-        print("Usage: python onnx_inference.py model_name.onnx")
-        exit()
-    model_path = sys.argv[1]
-    # print(model_path)
-
-    onnx_model = onnx.load(model_path)
-    onnx_input = onnx_model.graph.input[0]
-    input_shape = [[d.dim_value for d in _input.type.tensor_type.shape.dim]
-                   for _input in onnx_model.graph.input]
-    # Assume that there is only one input tensor
-    input_shape = input_shape[0]
-    # print(input_shape)
-    input_data = np.random.random(input_shape).astype(np.float32)
-
-    model = OnnxStub(onnx_model, backend.cuda_runtime())
-    next(iter(model.inputs.values())).copyin_numpy(input_data)
-    model.run()
-    outputs = next(iter(model.outputs.values())).copyout_numpy()
-    outputs = torch.tensor(outputs)
-    print(outputs.shape)
--- a/examples/python/paddle_densenet.py
+++ b/examples/python/paddle_densenet.py
@ -1,80 +0,0 @@
-
-import paddle
-import paddle.vision.transforms as T
-from paddle.vision.datasets import Cifar10
-from pyinfinitensor.onnx import OnnxStub, backend
-import onnx
-import itertools
-
-def run_cifar_train_and_infer():
-    
-    paddle.device.set_device("gpu")
-
-    transform = T.Compose(
-        [
-            T.Resize(224),
-            T.ToTensor(),
-            T.Normalize(
-                mean=[0.5, 0.5, 0.5],
-                std=[0.5, 0.5, 0.5],
-                to_rgb=True,
-            ),
-        ]
-    )
-    
-    # 下载数据集并初始化 DataSet
-    train_dataset = paddle.vision.datasets.Cifar10(mode='train', transform=transform)
-    test_dataset = paddle.vision.datasets.Cifar10(mode='test', transform=transform)
-
-    # 模型组网并初始化网络
-    densenet = paddle.vision.models.DenseNet(num_classes=10)
-    model = paddle.Model(densenet)
-
-    # 模型训练的配置准备，准备损失函数，优化器和评价指标
-    model.prepare(paddle.optimizer.Adam(parameters=model.parameters()), 
-                paddle.nn.CrossEntropyLoss(),
-                paddle.metric.Accuracy())
-
-    # 模型训练
-    model.fit(train_dataset, epochs=5, batch_size=64, verbose=1)
-    # 模型评估
-    model.evaluate(test_dataset, batch_size=64, verbose=1)
-
-    # export to ONNX
-    save_path = 'onnx.save/densenet' # 需要保存的路径
-    x_spec = paddle.static.InputSpec([1, 3, 224, 224], 'float32', 'x') # 为模型指定输入的形状和数据类型，支持持 Tensor 或 InputSpec ，InputSpec 支持动态的 shape。
-    paddle.onnx.export(densenet, save_path, input_spec=[x_spec], opset_version=11)
-
-    # 加载onnx模型并放到Infinitensor中
-    model_path = save_path + ".onnx"
-    onnx_model = onnx.load(model_path)
-    gofusion_model = OnnxStub(onnx_model, backend.cuda_runtime())
-    model = gofusion_model
-    model.init()
-
-    # 启动推理
-    cifar10_test = Cifar10(
-        mode="test",
-        transform=transform,  # apply transform to every image
-        backend="cv2",  # use OpenCV as image transform backend
-    )
-    batch_size = 1
-    total_size = 0
-    total_acc = 0.0
-    for data in itertools.islice(iter(cifar10_test), 10000):
-        images, labels = data
-        next(model.inputs.items().__iter__())[1].copyin_float(images.reshape([3*224*224]).tolist())
-        model.run()
-        outputs = next(model.outputs.items().__iter__())[1].copyout_float()
-        outputs = paddle.to_tensor(outputs)
-        outputs = paddle.reshape(outputs, (1, 10))
-        labels = paddle.to_tensor(labels)
-        labels = paddle.reshape(labels, (1,1))
-        acc = paddle.metric.accuracy(outputs, labels)
-        total_acc += acc
-        total_size += batch_size
-    print("test acc: {}".format(total_acc.numpy() / total_size))
-
-
-if __name__ == "__main__":
-    run_cifar_train_and_infer()
--- a/examples/python/paddle_inception.py
+++ b/examples/python/paddle_inception.py
@ -1,80 +0,0 @@
-import paddle
-import paddle.vision.transforms as T
-from paddle.vision.datasets import Cifar10
-from pyinfinitensor.onnx import OnnxStub, backend
-import onnx
-import itertools
-
-def run_cifar_train_and_infer():
-    
-    paddle.device.set_device("gpu")
-
-    transform = T.Compose(
-        [
-            T.Resize(224),
-            T.ToTensor(),
-            T.Normalize(
-                mean=[0.5, 0.5, 0.5],
-                std=[0.5, 0.5, 0.5],
-                to_rgb=True,
-            ),
-        ]
-    )
-    
-    # 下载数据集并初始化 DataSet
-    train_dataset = paddle.vision.datasets.Cifar10(mode='train', transform=transform)
-    test_dataset = paddle.vision.datasets.Cifar10(mode='test', transform=transform)
-
-    # 模型组网并初始化网络
-    inception = paddle.vision.models.InceptionV3(num_classes=10)
-    model = paddle.Model(inception)
-
-    # 模型训练的配置准备，准备损失函数，优化器和评价指标
-    model.prepare(paddle.optimizer.Adam(parameters=model.parameters()), 
-                paddle.nn.CrossEntropyLoss(),
-                paddle.metric.Accuracy())
-
-    # 模型训练
-    model.fit(train_dataset, epochs=5, batch_size=64, verbose=1)
-    # 模型评估
-    model.evaluate(test_dataset, batch_size=64, verbose=1)
-
-    # export to ONNX
-    save_path = 'onnx.save/inception' # 需要保存的路径
-    x_spec = paddle.static.InputSpec([1, 3, 224, 224], 'float32', 'x') # 为模型指定输入的形状和数据类型，支持持 Tensor 或 InputSpec ，InputSpec 支持动态的 shape。
-    paddle.onnx.export(inception, save_path, input_spec=[x_spec], opset_version=11)
-
-    # 加载onnx模型并放到Infinitensor中
-    model_path = save_path + ".onnx"
-    onnx_model = onnx.load(model_path)
-    gofusion_model = OnnxStub(onnx_model, backend.cuda_runtime())
-    model = gofusion_model
-    model.init()
-
-    # 启动推理
-    cifar10_test = Cifar10(
-        mode="test",
-        transform=transform,  # apply transform to every image
-        backend="cv2",  # use OpenCV as image transform backend
-    )
-    batch_size = 1
-    total_size = 0
-    total_acc = 0.0
-    for data in itertools.islice(iter(cifar10_test), 10000):
-        images, labels = data
-        next(model.inputs.items().__iter__())[1].copyin_float(images.reshape([3*224*224]).tolist())
-        model.run()
-        outputs = next(model.outputs.items().__iter__())[1].copyout_float()
-        outputs = paddle.to_tensor(outputs)
-        outputs = paddle.reshape(outputs, (1, 10))
-        labels = paddle.to_tensor(labels)
-        labels = paddle.reshape(labels, (1,1))
-        acc = paddle.metric.accuracy(outputs, labels)
-        total_acc += acc
-        total_size += batch_size
-    print("test acc: {}".format(total_acc.numpy() / total_size))
-
-
-
-if __name__ == "__main__":
-    run_cifar_train_and_infer() 
--- a/examples/python/paddle_model_dev.md
+++ b/examples/python/paddle_model_dev.md
@ -1,31 +0,0 @@
-## Description
-
-This is a doc to tell you how to run paddle*.py in your machine. If your model run on other machines except Nvidia, you may need to make some change.
-
-## What do we do in paddle*.py files?
-
-1. Train model and evalute model with Cifar10 dataset
-
-2. Export paddle model to onnx model
-
-3. Load onnx model, infer with InfiniTensor and calculate the inference accuracy
-
-## Command
-
-1. Go to `/examples/python` folder 
-
-2. Run the following command
-   
-   1. ```
-      python paddle_resnet.py
-      python paddle_densenet.py
-      python paddle_inception.py
-      ```
-
-## What should I do if I use other device(MLU, XPU, NPU)?
-
-You need to change this code:
-
-```
-paddle.device.set_device("gpu") # Change gpu to mlu, xpu or npu
-```
--- a/examples/python/paddle_resnet.py
+++ b/examples/python/paddle_resnet.py
@ -1,81 +0,0 @@
-
-import paddle
-import paddle.vision.transforms as T
-from paddle.vision.datasets import Cifar10
-from pyinfinitensor.onnx import OnnxStub, backend
-import onnx
-import itertools
-from paddle.vision.models.resnet import BasicBlock
-
-def run_cifar_train_and_infer():
-    
-    paddle.device.set_device("gpu")
-
-    transform = T.Compose(
-        [
-            T.Resize(224),
-            T.ToTensor(),
-            T.Normalize(
-                mean=[0.5, 0.5, 0.5],
-                std=[0.5, 0.5, 0.5],
-                to_rgb=True,
-            ),
-        ]
-    )
-    
-    # 下载数据集并初始化 DataSet
-    train_dataset = paddle.vision.datasets.Cifar10(mode='train', transform=transform)
-    test_dataset = paddle.vision.datasets.Cifar10(mode='test', transform=transform)
-
-    # 模型组网并初始化网络
-    resnet = paddle.vision.models.ResNet(BasicBlock, depth=18, num_classes=10)
-    model = paddle.Model(resnet)
-
-    # 模型训练的配置准备，准备损失函数，优化器和评价指标
-    model.prepare(paddle.optimizer.Adam(parameters=model.parameters()), 
-                paddle.nn.CrossEntropyLoss(),
-                paddle.metric.Accuracy())
-
-    # 模型训练
-    model.fit(train_dataset, epochs=5, batch_size=64, verbose=1)
-    # 模型评估
-    model.evaluate(test_dataset, batch_size=64, verbose=1)
-
-    # export to ONNX
-    save_path = 'onnx.save/resnet' # 需要保存的路径
-    x_spec = paddle.static.InputSpec([1, 3, 224, 224], 'float32', 'x') # 为模型指定输入的形状和数据类型，支持持 Tensor 或 InputSpec ，InputSpec 支持动态的 shape。
-    paddle.onnx.export(resnet, save_path, input_spec=[x_spec], opset_version=11)
-
-    # 加载onnx模型并放到Infinitensor中
-    model_path = save_path + ".onnx"
-    onnx_model = onnx.load(model_path)
-    gofusion_model = OnnxStub(onnx_model, backend.cuda_runtime())
-    model = gofusion_model
-    model.init()
-
-    # 启动推理
-    cifar10_test = Cifar10(
-        mode="test",
-        transform=transform,  # apply transform to every image
-        backend="cv2",  # use OpenCV as image transform backend
-    )
-    batch_size = 1
-    total_size = 0
-    total_acc = 0.0
-    for data in itertools.islice(iter(cifar10_test), 10000):
-        images, labels = data
-        next(model.inputs.items().__iter__())[1].copyin_float(images.reshape([3*224*224]).tolist())
-        model.run()
-        outputs = next(model.outputs.items().__iter__())[1].copyout_float()
-        outputs = paddle.to_tensor(outputs)
-        outputs = paddle.reshape(outputs, (1, 10))
-        labels = paddle.to_tensor(labels)
-        labels = paddle.reshape(labels, (1,1))
-        acc = paddle.metric.accuracy(outputs, labels)
-        total_acc += acc
-        total_size += batch_size
-    print("test acc: {}".format(total_acc.numpy() / total_size))
-
-
-if __name__ == "__main__":
-    run_cifar_train_and_infer()
--- a/examples/python/resnet_inference.py
+++ b/examples/python/resnet_inference.py
@ -1,24 +0,0 @@
-import sys
-import onnx
-import torch
-import numpy as np
-from pyinfinitensor.onnx import OnnxStub, backend
-import torchvision.models as models
-
-if __name__ == '__main__':
-    model_path = './resnet18.onnx'
-    tv_model = models.resnet50(weights=None)
-    input_shape = (1, 3, 224, 224)
-    param = torch.rand(input_shape)
-    torch.onnx.export(tv_model, param, model_path, verbose=False)
-
-    onnx_model = onnx.load(model_path)
-    model = OnnxStub(onnx_model, backend.cuda_runtime())
-    images = np.random.random(input_shape).astype(np.float32)
-    next(iter(model.inputs.values())).copyin_numpy(images)
-    model.run()
-    outputs = next(iter(model.outputs.values())).copyout_numpy()
-    outputs = torch.tensor(outputs)
-    outputs = torch.reshape(outputs, (1, 1000))
-    _, predicted = torch.max(outputs, 1)
-    print(predicted)
--- a/include/bang/bang_common.h
+++ b/include/bang/bang_common.h
@ -2,10 +2,6 @@
 #include "cnnl.h"
 #include "cnrt.h"
 #include "core/common.h"
-#include "core/data_type.h"
-#ifdef INFINI_USE_CNCL
-#include "cncl.h"
-#endif

 #define checkBangError(call)                                                   \
    {                                                                          \
@ -31,70 +27,4 @@ namespace infini {

 using BangPtr = void *;

-inline cnnlDataType_t cnnlDataTypeConvert(DataType dataType) {
-    if (dataType == DataType::Float32) {
-        return CNNL_DTYPE_FLOAT;
-    }
-    if (dataType == DataType::Float16) {
-        return CNNL_DTYPE_HALF;
-    }
-    if (dataType == DataType::Double) {
-        return CNNL_DTYPE_DOUBLE;
-    }
-    if (dataType == DataType::Int8) {
-        return CNNL_DTYPE_INT8;
-    }
-    if (dataType == DataType::Int32) {
-        return CNNL_DTYPE_INT32;
-    }
-    if (dataType == DataType::UInt8) {
-        return CNNL_DTYPE_UINT8;
-    }
-    if (dataType == DataType::BFloat16) {
-        return CNNL_DTYPE_BFLOAT16;
-    }
-    if (dataType == DataType::Int64) {
-        return CNNL_DTYPE_INT64;
-    }
-    if (dataType == DataType::Bool) {
-        return CNNL_DTYPE_BOOL;
-    }
-    IT_TODO_HALT_MSG("Data type " + dataType.toString() +
-                     " not supported in CNNL.");
-}
-
-#ifdef INFINI_USE_CNCL
-inline cnclDataType_t cnclDataTypeConvert(DataType dataType) {
-    if (dataType == DataType::Float32) {
-        return cnclFloat32;
-    }
-    if (dataType == DataType::Float16) {
-        return cnclHalf;
-    }
-    if (dataType == DataType::Int8) {
-        return cnclInt8;
-    }
-    if (dataType == DataType::Int16) {
-        return cnclInt16;
-    }
-    if (dataType == DataType::Int32) {
-        return cnclInt32;
-    }
-    if (dataType == DataType::UInt8) {
-        return cnclUint8;
-    }
-    if (dataType == DataType::UInt16) {
-        return cnclUint16;
-    }
-    if (dataType == DataType::UInt32) {
-        return cnclUint32;
-    }
-    if (dataType == DataType::BFloat16) {
-        return cnclBfloat16;
-    }
-    IT_TODO_HALT_MSG("Data type " + dataType.toString() +
-                     " not supported in CNCL.");
-}
-#endif
-
 } // namespace infini
--- a/include/bang/bang_element_wise.h
+++ b/include/bang/bang_element_wise.h
@ -0,0 +1,22 @@
+#pragma once
+#include "bang/bang_runtime.h"
+#include "bang_div.h"
+#include "operators/element_wise.h"
+namespace infini {
+
+void element_wise_kernel(const RuntimeObj *obj, const Operator &_op) {
+    auto op = as<ElementWiseObj>(_op);
+    float *const aData = (op->getInputs(0)->getRawDataPtr<float *>());
+    float *const bData = (op->getInputs(1)->getRawDataPtr<float *>());
+    float *const cData = (op->getOutput()->getRawDataPtr<float *>());
+
+    auto dim = op->getInputs(0)->getDims();
+    auto context = dynamic_cast<const BangRuntimeObj *>(obj);
+    int n = dim[0], c = dim[1], h = dim[2], w = dim[3];
+    if (op->getOpType() == OpType::Div)
+        div_kernel(context->cnnlHandle(), aData, bData, cData, n * c * h * w);
+    else
+        IT_TODO_HALT();
+}
+
+}; // namespace infini
--- a/include/bang/bang_runtime.h
+++ b/include/bang/bang_runtime.h
@ -7,35 +7,29 @@ namespace infini {
 class BangRuntimeObj : public RuntimeObj {
  private:
    cnnlHandle_t cnnl;
-    cnrtQueue_t queue;
-    std::unique_ptr<CommunicatorObj> comm;
    BangPtr workspace;
    size_t workspaceSize;
-    mutable size_t cursor;

  public:
-    explicit BangRuntimeObj(int deviceId = 0)
-        : RuntimeObj(Device::BANG, deviceId) {
-        cnInit(0);
-        CNdev dev;
-        cnDeviceGet(&dev, deviceId);
-        checkBangError(cnrtSetDevice(dev));
-        checkBangError(cnrtQueueCreate(&queue));
+    BangRuntimeObj() : RuntimeObj(Device::BANG) {
+        checkBangError(cnrtInit(0));
+        cnrtDev_t dev;
+        checkBangError(cnrtGetDeviceHandle(&dev, 0));
+        checkBangError(cnrtSetCurrentDevice(dev));
+        cnrtQueue_t queue;
+        checkBangError(cnrtCreateQueue(&queue));

        checkCnnlError(cnnlCreate(&cnnl));
        checkCnnlError(cnnlSetQueue(cnnl, queue));
        // 10GB for Longformer
        // size_t longformerNum = 3lu * (1 << 30);
        workspaceSize = 7ll << 30; // 7 GB
-        cursor = 0;
        workspace = alloc(workspaceSize);
    }
    virtual ~BangRuntimeObj() {
        dealloc(workspace);
-        checkBangError(cnrtQueueDestroy(queue));
        checkCnnlError(cnnlDestroy(cnnl));
    }
-    string toString() const override;

    void run(const Graph &graph, bool tune = false,
             bool profiling = false) const;
@ -50,15 +44,10 @@ class BangRuntimeObj : public RuntimeObj {
    void dealloc(void *ptr) override { checkBangError(cnrtFree(ptr)); }
    cnnlHandle_t cnnlHandle() const { return cnnl; }
    BangPtr getWorkspace(size_t size) const {
-        IT_ASSERT((cursor + size) <= workspaceSize);
-        cursor += size;
-        void *temp = workspace;
-        temp += (cursor - size);
-        return temp;
+        IT_ASSERT(size <= workspaceSize);
+        return workspace;
    }

-    void resetWorkspace() const { cursor = 0; }
-
    void copyBlobFromCPU(void *dst, const void *src,
                         size_t bytes) const override {
        checkBangError(cnrtMemcpy(dst, const_cast<void *>(src), bytes,
@ -76,9 +65,6 @@ class BangRuntimeObj : public RuntimeObj {
        checkBangError(cnrtMemcpy(dst, const_cast<void *>(src), bytes,
                                  CNRT_MEM_TRANS_DIR_PEER2PEER));
    }
-    void initComm(const string &name, int worldSize, int rank) final;
-    CommunicatorObj &getCommunicator() const override { return *comm; }
-    cnrtQueue_t getBangQueue() const { return queue; }

  private:
    void runWithoutSync(const Graph &graph, bool tune, bool profiling) const;
--- a/include/bang/cncl_communicator.h
+++ b/include/bang/cncl_communicator.h
@ -1,79 +0,0 @@
-#pragma once
-#include "bang_common.h"
-#include "core/communicator.h"
-#include <chrono>
-#include <cncl.h>
-#include <cnrt.h>
-#include <cstdlib>
-#include <filesystem>
-#include <fstream>
-#include <mutex>
-#include <thread>
-
-namespace infini {
-
-class CnclCommunicatorObj final : public CommunicatorObj {
-  private:
-    cnclComm_t *comms;
-
-  public:
-    CnclCommunicatorObj(const string &name, int worldSize, int rank)
-        : CommunicatorObj(worldSize, rank) {
-        const std::string filePath("./" + name + "_cncl_id.bin");
-        cnclCliqueId clique_id;
-        if (rank == 0) {
-            CNCL_CHECK(cnclGetCliqueId(&clique_id));
-            std::ofstream ofs(filePath, std::ios::binary);
-            ofs.write((char *)&clique_id, sizeof(cnclCliqueId));
-
-        } else {
-            auto begin = std::chrono::steady_clock::now();
-            while (!std::filesystem::exists(filePath)) {
-                auto now = std::chrono::steady_clock::now();
-                _IT_ASSERT_2(now < begin + std::chrono::seconds(10),
-                             "time limit (10s) exceeded.");
-                std::this_thread::sleep_for(std::chrono::milliseconds(100));
-            }
-            std::ifstream ifs(filePath, std::ios::binary);
-            ifs.read((char *)&clique_id, sizeof(cnclCliqueId));
-        }
-
-        int num_comms = 1;
-        int *dev_list = new int[num_comms];
-        int *rank_list = new int[num_comms];
-        comms = new cnclComm_t[num_comms];
-        uint32_t num_dev = 0;
-        checkBangError(cnrtGetDeviceCount(&num_dev));
-
-        for (int i = 0; i < num_comms; i++) {
-            rank_list[i] = rank;
-            dev_list[i] = rank_list[i] % num_dev;
-        }
-
-        CNCL_CHECK(cnclInitComms(comms, num_comms, dev_list, rank_list,
-                                 worldSize, &clique_id));
-
-        if (rank == 0) {
-            std::filesystem::remove(filePath);
-        }
-
-        delete[] dev_list;
-        delete[] rank_list;
-    }
-
-    ~CnclCommunicatorObj() {
-        CNCL_CHECK(cnclDestroyComms(comms, 1));
-        delete[] comms;
-    }
-
-    // Get the actual cnclComm_t
-    cnclComm_t getCnclComm() { return comms[0]; }
-
-    virtual string toString() const final {
-        std::ostringstream oss;
-        oss << "CNCL communicator";
-        return oss.str();
-    }
-};
-
-} // namespace infini
--- a/include/bang/operator_timer.h
+++ b/include/bang/operator_timer.h
@ -1,10 +0,0 @@
-#pragma once
-namespace infini {
-namespace opTimer {
-double getPerfConvCnnl(int n, int c, int h, int w, int f, int r, int s,
-                       int padh, int padw, int strideh, int stridew,
-                       int dilationh, int dilationw, int group,
-                       const char *name);
-double getPerfMatmulCnnl(int b, int m, int n, int k, const char *name);
-} // namespace opTimer
-} // namespace infini
--- a/include/core/common.h
+++ b/include/core/common.h
@ -39,18 +39,17 @@ using HashType = uint64_t; // compatible with std::hash
 #define _VA_SELECT(NAME, ...) _SELECT(NAME, _VA_SIZE(__VA_ARGS__))(__VA_ARGS__)

 // Assert: conditions should have no side effect
-#define _IT_ASSERT_2(condition, info)                                          \
-    static_cast<bool>(condition)                                               \
-        ? void(0)                                                              \
-        : throw ::infini::Exception(                                           \
-              std::string("[") + __FILE__ + ":" + std::to_string(__LINE__) +   \
-              "] Assertion failed (" + #condition + "): " + info)
-#define _IT_ASSERT_1(condition) _IT_ASSERT_2(condition, "")
+#define _IT_ASSERT_2(name, info)                                               \
+    (static_cast<bool>(name)                                                   \
+         ? void(0)                                                             \
+         : throw ::infini::Exception(                                          \
+               std::string("[") + __FILE__ + ":" + std::to_string(__LINE__) +  \
+               "] Assertion failed (" + #name + "): " + info))
+#define _IT_ASSERT_1(name) _IT_ASSERT_2(name, "");
 #define IT_ASSERT(...) _VA_SELECT(_IT_ASSERT, __VA_ARGS__)

 #define IT_TODO_HALT() _IT_ASSERT_2(false, "Unimplemented")
 #define IT_TODO_HALT_MSG(msg) _IT_ASSERT_2(false, msg)
-#define IT_ASSERT_TODO(condition) _IT_ASSERT_2(condition, "Unimplemented")
 #define IT_TODO_SKIP() puts("Unimplemented " __FILE__ ":" __LINE__)

 // Other utilities
@ -61,35 +60,21 @@ template <typename T> auto enum_to_underlying(T e) {
 }

 template <typename T> std::string vecToString(const std::vector<T> &vec) {
-    std::stringstream ss;
-    ss << "[";
-    for (size_t i = 0; i < vec.size(); ++i) {
-        ss << vec.at(i);
-        if (i < vec.size() - 1) {
-            ss << ",";
-        }
+    std::string ret;
+    ret.append("[");
+    for (auto d : vec) {
+        ret.append(std::to_string(d));
+        ret.append(",");
    }
-    ss << "]";
-    return ss.str();
-}
-
-template <typename T> std::string vecToString(const T *st, size_t length) {
-    std::stringstream ss;
-    ss << "[";
-    size_t i = 0;
-    for (i = 0; i < length; i++) {
-        ss << *(st + i);
-        if (i < length - 1) {
-            ss << ",";
-        }
-    }
-    ss << "]";
-    return ss.str();
+    if (!vec.empty())
+        ret.pop_back();
+    ret.append("]");
+    return ret;
 }

 double timeit(
    const std::function<void()> &func,
-    const std::function<void(void)> &sync = []() {}, int warmupRounds = 10,
-    int timingRounds = 10);
+    const std::function<void(void)> &sync = []() {}, int warmupRounds = 200,
+    int timingRounds = 200);

 } // namespace infini
--- a/include/core/communicator.h
+++ b/include/core/communicator.h
@ -1,22 +0,0 @@
-#pragma once
-#include "object.h"
-#include "ref.h"
-
-namespace infini {
-
-// base class
-class CommunicatorObj : public Object {
-  protected:
-    int worldSize;
-    int rank;
-
-  public:
-    CommunicatorObj(int worldSize, int rank)
-        : worldSize(worldSize), rank(rank) {}
-
-    virtual ~CommunicatorObj() = default;
-    virtual int getWorldSize() const { return worldSize; }
-    virtual int getRank() const { return rank; }
-};
-
-} // namespace infini
--- a/include/core/constants.h
+++ b/include/core/constants.h
@ -2,4 +2,4 @@

 namespace infini {
 constexpr double E_CONSTANT = 2.718281828459;
-}
+}
--- a/include/core/data_type.h
+++ b/include/core/data_type.h
@ -1,54 +1,13 @@
-#pragma once
 #include "core/common.h"

 namespace infini {

 class DataType {
  public:
-    // <https://onnx.ai/onnx/intro/concepts.html#element-type>
-    static const DataType Undefine;
    static const DataType Float32;
-    static const DataType UInt8;
-    static const DataType Int8;
-    static const DataType UInt16;
-    static const DataType Int16;
-    static const DataType Int32;
-    static const DataType Int64;
-    static const DataType String;
-    static const DataType Bool;
-    static const DataType Float16;
-    static const DataType Double;
    static const DataType UInt32;
-    static const DataType UInt64;
-    static const DataType BFloat16;
-    // "sizePerElement" show the DType to cpu_type
-    // DataType::Bool -> int8_t   DataType::Float16 -> uint16_t
-    static constexpr size_t sizePerElement[]{0,
-                                             sizeof(float),
-                                             sizeof(uint8_t),
-                                             sizeof(int8_t),
-                                             sizeof(uint16_t),
-                                             sizeof(int16_t),
-                                             sizeof(int32_t),
-                                             sizeof(int64_t),
-                                             sizeof(std::string),
-                                             sizeof(int8_t),
-                                             sizeof(uint16_t),
-                                             sizeof(double),
-                                             sizeof(uint32_t),
-                                             sizeof(uint64_t),
-                                             0,
-                                             0,
-                                             sizeof(uint16_t)};
-
-    static constexpr std::string_view names[]{
-        "Undefine",    "Float32", "UInt8",  "Int8",   "UInt16",
-        "Int16",       "Int32",   "Int64",  "String", "Bool",
-        "Float16",     "Double",  "UInt32", "UInt64", "PlaceHolder",
-        "PlaceHolder", "BFloat16"};
-
-    static constexpr int cpuType[]{-1, 0, 2, 3, 4, 5,  6,  7, -1,
-                                   3,  4, 9, 1, 8, -1, -1, 4};
+    static constexpr size_t sizePerElement[]{sizeof(float), sizeof(uint32_t)};
+    static constexpr std::string_view names[]{"Float32", "UInt32"};

  private:
    int index;
@ -61,43 +20,18 @@ class DataType {
    bool operator==(const DataType &rhs) const { return index == rhs.index; }
    bool operator<(const DataType &rhs) const { return index < rhs.index; }

-    template <typename T> static int get() {
+    template <typename T> static DataType get() {
        IT_TODO_HALT_MSG("Unsupported data type");
    }
    size_t getSize() const { return sizePerElement[index]; }
    string toString() const { return string(names[index]); }
-    int cpuTypeInt() const { return cpuType[index]; }
-    int getIndex() const { return index; }
 };

+inline const DataType DataType::Float32(0);
+inline const DataType DataType::UInt32(1);
 // Method definitions are out of the declaration due to GCC bug:
 // https://stackoverflow.com/questions/49707184/explicit-specialization-in-non-namespace-scope-does-not-compile-in-gcc
-template <> inline int DataType::get<float>() { return 0; }
-template <> inline int DataType::get<uint32_t>() { return 1; }
-template <> inline int DataType::get<uint8_t>() { return 2; }
-template <> inline int DataType::get<int8_t>() { return 3; }
-template <> inline int DataType::get<uint16_t>() { return 4; }
-template <> inline int DataType::get<int16_t>() { return 5; }
-template <> inline int DataType::get<int32_t>() { return 6; }
-template <> inline int DataType::get<int64_t>() { return 7; }
-template <> inline int DataType::get<uint64_t>() { return 8; }
-template <> inline int DataType::get<double>() { return 9; }
+template <> inline DataType DataType::get<float>() { return Float32; }
+template <> inline DataType DataType::get<uint32_t>() { return UInt32; }

-template <int index> struct DT {};
-template <> struct DT<0> { using t = bool; };
-template <> struct DT<1> { using t = float; };
-template <> struct DT<2> { using t = uint8_t; };
-template <> struct DT<3> { using t = int8_t; };
-template <> struct DT<4> { using t = uint16_t; };
-template <> struct DT<5> { using t = int16_t; };
-template <> struct DT<6> { using t = int32_t; };
-template <> struct DT<7> { using t = int64_t; };
-template <> struct DT<8> { using t = char; };
-template <> struct DT<9> { using t = int8_t; };
-template <> struct DT<10> { using t = uint16_t; };
-template <> struct DT<11> { using t = double; };
-template <> struct DT<12> { using t = uint32_t; };
-template <> struct DT<13> { using t = uint64_t; };
-template <> struct DT<16> { using t = uint16_t; };
-
-} // namespace infini
+} // namespace infini
--- a/include/core/dummy_mutator.h
+++ b/include/core/dummy_mutator.h
@ -1,15 +0,0 @@
-#pragma once
-#include "core/mutator.h"
-
-namespace infini {
-
-class DummyMutator : public Mutator {
-  public:
-    DummyMutator(int candidatesLimit) : Mutator(candidatesLimit){};
-
-    virtual vector<Graph> run(const Graph &inGraph) override;
-    virtual vector<Graph> mergeMultiBranch(const Graph &inGraph) override;
-    virtual bool isMultiBranchMergable(const Graph &inGraph) override;
-};
-
-} // namespace infini
--- a/include/core/graph.h
+++ b/include/core/graph.h
@ -1,5 +1,4 @@
 #pragma once
-#include "core/lazy_allocator.h"
 #include "core/operator.h"
 #include "core/tensor.h"

@ -9,69 +8,21 @@ class GraphObj : public Object {
  protected:
    Runtime runtime;
    TensorVec tensors;
+    TensorVec inputs;
+    TensorVec outputs;
    OpVec ops;
-    LazyAllocator allocator;

  public:
-    explicit GraphObj(Runtime runtime)
-        : runtime(runtime), allocator(runtime), sorted(false){};
-    GraphObj(Runtime runtime, OpVec ops_in);
+    GraphObj(Runtime runtime) : runtime(runtime){};
    string toString() const override;
-    Runtime getRuntime() const { return runtime; }

    Tensor addTensor(Shape dim, DataType dtype = DataType::Float32);
-    Tensor addTensor(const Tensor &tensor);
-    TensorVec addTensor(const TensorVec &tensors);
-    /**
-     * @brief Clone a tensor and add it to the graph.
-     */
    Tensor cloneTensor(const Tensor &tensor) {
-        return addTensor(tensor->clone(runtime));
+        auto ret = addTensor(tensor->getDims(), tensor->getDType());
+        ret->dataMalloc();
+        ret->copyData(tensor);
+        return ret;
    }
-    void removeOperator(Operator op) {
-        auto it = std::find(ops.begin(), ops.end(), op);
-        if (it != ops.end())
-            ops.erase(it);
-    }
-
-    void removeTensor(Tensor tensor) {
-        auto it = std::find(tensors.begin(), tensors.end(), tensor);
-        if (it != tensors.end())
-            tensors.erase(it);
-    }
-
-    void deleteConnection(Tensor tensor, Operator op);
-    void addConnection(Tensor tensor, Operator op);
-    void replaceConnection(Tensor oldInput, Tensor newInput, Operator op);
-
-    Operator cloneOperator(Operator op, TensorVec inputs, TensorVec outputs) {
-        auto opClone = op->clone(inputs, outputs);
-        addOperatorAndConnect(opClone);
-        return opClone;
-    }
-
-    const TensorVec &getTensors() const { return tensors; }
-    const OpVec &getOperators() const { return ops; }
-    OpVec getComputeOps() const;
-    Tensor getTensor(int) const;
-
-    /**
-     * Sort the nodes in topological order.
-     * It returns true if the sorting is successful.
-     * Otherwise false is returned, means that there are rings in the graph,
-     * so the topological sorting fails.
-     */
-    bool topo_sort();
-
-    void optimize();
-
-    void shape_infer();
-
-    void dataMalloc(bool useNaiveAllocator = false, size_t memPoolSize = 0);
-
-    Tensor cloneKV(Tensor &tensor);
-
-    void freeHeap();

    /**
     * @brief Add an operator and create its outputs. Output tensor arguments
@ -93,29 +44,15 @@ class GraphObj : public Object {
        return op;
    }

-    /**
-     * @brief Gets input tensors of this graph.
-     */
-    inline TensorVec getInputs() const {
-        TensorVec ret;
-        for (const auto &t : tensors)
-            if (!t->getSource())
-                ret.emplace_back(t);
-        return ret;
-    }
+    const TensorVec &getTensors() const { return tensors; }
+    const TensorVec &getInputs() const { return inputs; }
+    const TensorVec &getOutputs() const { return outputs; }
+    const OpVec &getOperators() const { return ops; }
+    OpVec getComputeOps() const;
+    // TensorVec &getInputs();
+    // TensorVec &getOutputs();

-    /**
-     * @brief Gets output tensors of this graph.
-     */
-    inline TensorVec getOutputs() const {
-        TensorVec ret;
-        for (const auto &t : tensors)
-            if (t->getTargets().empty())
-                ret.emplace_back(t);
-        return ret;
-    }
-
-    bool checkValid() const;
+    void dataMalloc();

  private:
    /**
@ -123,15 +60,9 @@ class GraphObj : public Object {
     */
    void addOperatorAndConnect(const Operator &op);

-    /**
-     * @brief If the nodes is sorted in topological order.
-     */
-    bool sorted;
-
-    /**
-     * @brief If the weight tensors are allocated.
-     */
-    bool weightAllocated = false;
+    // TODO: move to another class
+    // bool exportOnnx(const char *path);
+    // bool importOnnx(const char *net);
 };

 } // namespace infini
--- a/include/core/graph_handler.h
+++ b/include/core/graph_handler.h
@ -1,154 +0,0 @@
-#pragma once
-
-#include "core/graph.h"
-#include "core/runtime.h"
-#include <cstdint>
-#include <iostream>
-
-#ifdef USE_CUDA
-#include "cuda/cuda_runtime.h"
-#endif
-
-namespace infini {
-
-class GraphHandlerObj {
-    Graph g;
-
-  public:
-    GraphHandlerObj(Runtime runtime)
-        : g(make_ref<GraphObj>(std::move(runtime))) {}
-
-    Tensor tensor(Shape dims, int dtype);
-
-    //------ operators
-
-    inline OpVec operators() { return g->getOperators(); }
-
-    Tensor conv(Tensor input, Tensor weight, Tensor output, int ph, int pw,
-                int sh, int sw, int dh, int dw);
-    Tensor convTransposed2d(Tensor input, Tensor weight, Tensor output, int ph,
-                            int pw, int sh, int sw, int dh, int dw, int oph,
-                            int opw);
-    Tensor matmul(Tensor a, Tensor b, Tensor y, bool transA, bool transB,
-                  Tensor bias, ActType act,
-                  std::string matmul_compute_type = "default");
-    Tensor batchNormalization(Tensor input, Tensor output, Tensor mean,
-                              Tensor var, Tensor scale, Tensor bias,
-                              float momentum, float eps, bool training);
-    Tensor layerNormalization(Tensor input, Tensor scale, Tensor output,
-                              Tensor bias, float eps, int axis, int stash_type);
-    Tensor rmsNorm(Tensor input, Tensor weight, Tensor output);
-
-    Tensor maxPool(Tensor input, Tensor output, int kh, int kw, int dh, int dw,
-                   int ph, int pw, int sh, int sw, int ceilMode);
-    Tensor avgPool(Tensor input, Tensor output, int kh, int kw, int dh, int dw,
-                   int ph, int pw, int sh, int sw, int ceilMode);
-
-    Tensor add(Tensor a, Tensor b, Tensor c);
-    Tensor sub(Tensor a, Tensor b, Tensor c);
-    Tensor mul(Tensor a, Tensor b, Tensor c);
-    Tensor div(Tensor a, Tensor b, Tensor c);
-    Tensor pow(Tensor a, Tensor b, Tensor c);
-    Tensor min(Tensor a, Tensor b, Tensor c);
-    Tensor max(Tensor a, Tensor b, Tensor c);
-
-    Tensor relu(Tensor x, Tensor y);
-    Tensor silu(Tensor x, Tensor y);
-    Tensor gelu(Tensor x, Tensor y);
-    Tensor sigmoid(Tensor x, Tensor y);
-    Tensor hardSigmoid(Tensor x, Tensor y);
-    Tensor hardSwish(Tensor x, Tensor y);
-    Tensor tanh(Tensor x, Tensor y);
-    Tensor erf(Tensor x, Tensor y);
-    Tensor softmax(Tensor x, Tensor y, int axis);
-    Tensor abs(Tensor x, Tensor y);
-    Tensor sqrt(Tensor x, Tensor y);
-    Tensor neg(Tensor x, Tensor y);
-    Tensor shape(Tensor x, Tensor y);
-    Tensor identity(Tensor x, Tensor y);
-    Tensor flatten(Tensor s, Tensor y, int axis);
-    Tensor pRelu(Tensor x, Tensor slope, Tensor y);
-    Tensor clip(Tensor x, Tensor y, std::optional<float> min,
-                std::optional<float> max);
-    Tensor transpose(Tensor data, Tensor transposed, Shape perm);
-    Tensor reshape(Tensor data, Tensor reshaped, Shape shape);
-    Tensor resize(Tensor input, Tensor output,
-                  const std::optional<vector<int>> &axes, Tensor sizes,
-                  Tensor scales, Tensor roi, vector<uint32_t> sizes_,
-                  vector<float> scales_, vector<float> roi_, string mode,
-                  string ratioPolicy, string nearestMode,
-                  string coordTransMode);
-    Tensor squeeze(Tensor input, Tensor output, Shape axes);
-    Tensor unsqueeze(Tensor input, Tensor output, Shape axes);
-    Tensor concat(TensorVec inputs, Tensor output, int dim);
-    Tensor attentionKVCache(Tensor input_k_cache, Tensor input_v_cache,
-                            Tensor input_q, Tensor input_k, Tensor input_v,
-                            Tensor position_id, Tensor output_matmul);
-    Tensor RoPE(Tensor pos, Tensor input, Tensor output);
-    TensorVec split(Tensor input, std::optional<TensorVec> outputs, int axis,
-                    std::variant<int, vector<int>> numOrRatio);
-    Tensor gather(Tensor data, Tensor indices, Tensor output, int axis);
-    Tensor gatherElements(Tensor data, Tensor indices, Tensor output, int axis);
-    Tensor reduceMean(Tensor data, Tensor reduced,
-                      const optional<vector<int>> &axes, bool keepdims);
-    Tensor reduceSum(Tensor data, Tensor reduced,
-                     const optional<vector<int>> &axes, bool keepdims);
-    Tensor slice(Tensor input, Tensor output, const vector<int> &starts,
-                 const vector<int> &ends, const optional<vector<int>> &axes,
-                 const optional<vector<int>> &steps);
-    Tensor pad(Tensor input, Tensor output, const vector<int> &pads,
-               const optional<vector<int>> &axes);
-    Tensor cast(Tensor input, Tensor output, int to);
-    Tensor expand(Tensor input, Tensor output, Shape dims);
-    Tensor where(Tensor inputX, Tensor inputY, Tensor condition, Tensor output);
-    std::vector<int> getDims(Tensor x) { return x->getDims(); }
-
-    Tensor allReduceSum(Tensor input, Tensor output);
-    Tensor allReduceProd(Tensor input, Tensor output);
-    Tensor allReduceMin(Tensor input, Tensor output);
-    Tensor allReduceMax(Tensor input, Tensor output);
-    Tensor allReduceAvg(Tensor input, Tensor output);
-    TensorVec allGather(Tensor input, std::optional<TensorVec> outputs, int n);
-    Tensor broadcast(Tensor input, Tensor output, int root);
-    Tensor send(Tensor input, int source, int destination, Tensor output);
-    Tensor recv(Tensor output, int source, int destination, Shape dims,
-                int outputType, Tensor input);
-    Tensor depthToSpace(Tensor input, Tensor output, int blocksize,
-                        std::string mode);
-    Tensor lrn(Tensor input, Tensor output, float alpha, float beta, float bias,
-               int size);
-
-    //------ modifiers
-
-    inline bool topo_sort() { return g->topo_sort(); }
-
-    inline void optimize() { g->optimize(); }
-
-    inline void shape_infer() { g->shape_infer(); }
-
-    void change_shape(const vector<int> &shape, int tensorId);
-    //------ runtime
-
-    inline void data_malloc(bool useNaiveAllocator = false,
-                            size_t memPoolSize = 0) {
-        g->dataMalloc(useNaiveAllocator, memPoolSize);
-    }
-
-    inline Tensor clone_KV(Tensor &tensor) { return g->cloneKV(tensor); }
-
-    inline void free_heap() { g->freeHeap(); }
-
-    inline void tune() { g->getRuntime()->run(g, true); }
-
-    inline void run() { g->getRuntime()->run(g); }
-
-    inline double get_perf_time() { return g->getRuntime()->getPerfTime(g); }
-
-#ifdef USE_CUDA
-    inline void run_with_cudagraph() {
-        (as<CudaRuntimeObj>(g->getRuntime()))->runWithCudaGraph(g);
-    }
-#endif
-};
-
-} // namespace infini
--- a/include/core/graph_match.h
+++ b/include/core/graph_match.h
@ -1,108 +0,0 @@
-#pragma once
-
-#include "core/graph.h"
-namespace infini {
-class SubGraphObj : public GraphObj {
-    TensorVec ins;  // inputs from outer predecessors, orders are appointed.
-    TensorVec outs; // outputs to outer successors, orders are appointed.
-
-  public:
-    SubGraphObj(Runtime runtime, const TensorVec &inputs);
-    void setOutputs(const TensorVec &tensors) { outs = tensors; }
-    TensorVec getInputsFromOutside() const { return ins; }
-    TensorVec getOutputs2Outside() const { return outs; }
-    bool isInputFromOutside(Tensor t) const {
-        return std::find(ins.begin(), ins.end(), t) != ins.end();
-    }
-    bool isOutput2Outside(Tensor t) const {
-        return std::find(outs.begin(), outs.end(), t) != outs.end();
-    }
-    bool isHead(const Operator &op) const {
-        for (auto in : ins) {
-            auto ops = in->getTargets();
-            if (std::find(ops.begin(), ops.end(), op) != ops.end())
-                return true;
-        }
-        return false;
-    };
-    bool isTail(const Operator &op) const {
-        for (auto out : outs) {
-            if (op == out->getSource())
-                return true;
-        }
-        return false;
-    }
-};
-using SubGraph = Ref<SubGraphObj>;
-
-// Describe a match for subgraph replacement.
-class GraphMatchObj {
-    std::unordered_set<Operator> ops;
-    std::unordered_map<Operator, Operator> opMap;       // anchor->pattern
-    std::unordered_map<Operator, Operator> opMapRevese; // pattern->anchor
-    std::unordered_map<Tensor, Tensor> tensorMap;       // pattern->anchor
-    SubGraph pattern;
-
-  public:
-    GraphMatchObj(SubGraph pattern) : pattern(pattern) {}
-    Ref<GraphMatchObj> clone();
-    void addOp(const Operator &anchorOp, const Operator &patternOp);
-    bool hasContained(const Operator &op) const { return opMap.count(op) > 0; }
-    bool hasMatched(const Operator &op) const {
-        return opMapRevese.count(op) > 0;
-    }
-
-    Tensor getAnchorByPattern(const Tensor &t) {
-        IT_ASSERT(tensorMap.count(t) > 0);
-        return tensorMap.at(t);
-    }
-
-    Operator getAnchorByPattern(const Operator &op) {
-        IT_ASSERT(opMapRevese.count(op) > 0);
-        return opMapRevese.at(op);
-    }
-
-    TensorVec getInputs() const;
-    TensorVec getOutputs() const;
-    std::unordered_set<Operator> getOps() const { return ops; }
-    std::string toString() const;
-
-  private:
-    void recordOutsideTensorMap(const Operator &patternOp,
-                                const Operator &anchorOp);
-};
-using MatchGraph = Ref<GraphMatchObj>;
-
-class SubGraphRewriter {
-    SubGraph pattern;
-    Graph graph;
-
-  public:
-    SubGraphRewriter(Graph g) : graph(g) {}
-    vector<MatchGraph> findMatch(const SubGraph &pattern);
-    void replaceSubGraph(const SubGraph &pattern, const SubGraph &replacement);
-    TensorVec addSubGraph(const SubGraph &pattern, const TensorVec &inputs);
-
-  private:
-    void removeSubGraph(MatchGraph match);
-    bool MatchNode(const Operator &a, const Operator &b, bool isHead,
-                   bool isTail) const;
-    OpLists matchInCandidates(const OpVec &ops, const Operator &opDst,
-                              bool isHead, bool isTail);
-    bool findMatch(const MatchGraph &lastMatched, const Operator &opLastMatched,
-                   const Operator &opDst, vector<MatchGraph> &matched);
-    bool findMatch2(const MatchGraph &lastMatched,
-                    const Operator &opLastMatched, const Operator &opDst,
-                    vector<MatchGraph> &matched);
-    void updateMatchedGraph(const MatchGraph &lastMatched, OpLists &opMatched,
-                            vector<MatchGraph> &gMatched, Operator dst);
-
-    bool checkReplacement(const SubGraph &pattern, const SubGraph &other) const;
-    bool checkReplacement(const TensorVec &left, const TensorVec &right) const;
-    bool isReplacable(const Tensor &l, const Tensor &r) const;
-    bool checkOverlapsWithPreviousMatch(
-        const MatchGraph &match,
-        const std::unordered_set<Operator> &nodesToDelete) const;
-    bool checkMatchValid(const MatchGraph &match) const;
-};
-}; // namespace infini
--- a/include/core/hash.h
+++ b/include/core/hash.h
@ -15,4 +15,4 @@ template <typename T> inline HashType hashVector(const vector<T> &vec) {
    return ret;
 }

-} // namespace infini
+} // namespace infini
--- a/include/core/kernel.h
+++ b/include/core/kernel.h
@ -2,11 +2,10 @@
 #include "core/common.h"
 #include "core/operator.h"
 #include "core/tensor.h"
-#include "utils/operator_utils.h"
 #include <functional>
 #include <nlohmann/json.hpp>
-namespace infini {
 using json = nlohmann::json;
+namespace infini {

 class RuntimeObj; // Forward declaration for Kernel::compute

@ -30,6 +29,7 @@ class Kernel {
  public:
    Kernel() {}
    virtual ~Kernel() {}
+
    /**
     * @param op The operator to be executed.
     * @param record The parameters for kernel execution. If extra parameters
@ -102,9 +102,11 @@ class KernelRegistry {
    }
    Kernel *getKernel(const KernelAttrs &kernelAttrs) const {
        auto it = kernels.find(kernelAttrs);
-        IT_ASSERT(it != kernels.end(), "Kernel not found for key {" +
-                                           get_kernel_attrs_str(kernelAttrs) +
-                                           "}");
+        IT_ASSERT(it != kernels.end(),
+                  "Kernel not found for key {" +
+                      to_string(enum_to_underlying(std::get<0>(kernelAttrs))) +
+                      ", " + OpRegistry::getOpName(std::get<1>(kernelAttrs)) +
+                      ", " + std::get<2>(kernelAttrs).toString());
        return std::get<0>(it->second);
    }
    const KernelRecord &getKernelItem(const KernelAttrs &kernelAttrs) const {
@ -129,16 +131,15 @@ class CpuKernelWithoutConfig : public Kernel {

 } // namespace infini

-#define _REGISTER_KERNEL_1(device, opType, kernel, name, cnt)                  \
+#define _REGISTER_KERNEL_1(device, opType, dataType, kernel, name, cnt)        \
    namespace infini {                                                         \
    static const bool _CAT(_register_kernel_, cnt) =                           \
-        KernelRegistry::getInstance().registerKernel(KernelAttrs{device,       \
-                                                                 opType},      \
-                                                     new kernel(), name);      \
+        KernelRegistry::getInstance().registerKernel(                          \
+            KernelAttrs{device, opType, dataType}, new kernel(), name);        \
    }

-#define REGISTER_KERNEL(device, opType, kernel, name)                          \
-    _REGISTER_KERNEL_1(device, opType, kernel, name, __COUNTER__)
+#define REGISTER_KERNEL(device, opType, dataType, kernel, name)                \
+    _REGISTER_KERNEL_1(device, opType, dataType, kernel, name, __COUNTER__)

 #define _REGISTER_CONSTRUCTOR_1(type, constructor, cnt)                        \
    namespace infini {                                                         \
--- a/include/core/lazy_allocator.h
+++ b/include/core/lazy_allocator.h
@ -1,122 +0,0 @@
-#pragma once
-#include "core/runtime.h"
-#include "core/tensor.h"
-#ifdef BUILD_TEST
-#include "gtest/gtest.h"
-#endif
-#include <cstddef>
-#include <map>
-#include <unordered_set>
-
-namespace infini {
-
-class LazyAllocator {
-  private:
-#ifdef BUILD_TEST
-    FRIEND_TEST(LazyAllocator, testMergeFreeBlocks);
-
-    FRIEND_TEST(LazyAllocator, testAllocWithEndFreeBlock);
-#endif
-
-    Runtime runtime;
-
-    size_t used = 0;
-
-    size_t peak = 0;
-
-    size_t weightPeak = 0;
-
-    size_t heapPeak = 0;
-
-    size_t alignment;
-
-    bool hasMemPool = false;
-
-    size_t memPoolSize = 0;
-
-    // pointer to the memory actually allocated
-    void *ptr = nullptr;
-
-    // pointer to the weight memory space
-    void *weightPtr = nullptr;
-
-    // memory pool ptr
-    void *memPoolPtr = nullptr;
-
-    // // a cache designed for a batch size that has already occurred
-    // std::unordered_map<size_t, std::unordered_map<TensorObj *, size_t>>
-    // batchsizeToTensorOffset;
-
-    struct freeBlockInfo {
-        size_t addr;
-        size_t blockSize;
-    };
-
-    struct cmpFreeBlockInfo {
-        bool operator()(const freeBlockInfo &a, const freeBlockInfo &b) const {
-            return (a.blockSize != b.blockSize) ? (a.blockSize < b.blockSize)
-                                                : (a.addr < b.addr);
-        }
-    };
-
-    // free balanced tree, maintains all free memory blocks
-    std::set<freeBlockInfo, cmpFreeBlockInfo> freeBlocks;
-
-    // key: head address offset of the free memory block
-    // value: blockSize of the block
-    std::unordered_map<size_t, size_t> headAddrToBlockSize;
-
-    // key: tail address offset of the free memory block
-    // value: blockSize of the block
-    std::unordered_map<size_t, size_t> tailAddrToBlockSize;
-
-  public:
-    LazyAllocator(Runtime runtime);
-
-    virtual ~LazyAllocator();
-
-    void init();
-
-    void setMemPool(size_t memPoolSize);
-
-    bool getMemPoolStatus();
-
-    // function: simulate memory allocation
-    // arguments：
-    //     size: size of memory block to be allocated
-    // return: head address offset of the allocated memory block
-    size_t alloc(size_t size);
-
-    size_t allocWeight(size_t size);
-
-    size_t heapAlloc(size_t size);
-
-    void freeHeap();
-
-    // function: simulate memory free
-    // arguments:
-    //     addr: head address offset of memory block to be free
-    //     size: size of memory block to be freed
-    void free(size_t addr, size_t size);
-
-    // function: perform actual memory allocation
-    // return: pointer to the head address of the allocated memory
-    void *getPtr();
-
-    // void addCache(size_t batchsize, std::unordered_map<TensorObj *, size_t>);
-
-    // std::unordered_map<TensorObj *, size_t> getCache(size_t batchsize);
-
-    void *getWeightPtr();
-
-    void *getHeapPtr();
-
-    void info();
-
-  private:
-    // function: memory alignment, rouned up
-    // return: size of the aligned memory block
-    size_t getAlignedSize(size_t size);
-};
-
-} // namespace infini
--- a/include/core/mutator.h
+++ b/include/core/mutator.h
@ -8,28 +8,12 @@ class Mutator {
    int candidatesLimit;
    // // Statistical data
    // int numTotalCandidates;
-  protected:
-    Runtime runtime;

  public:
-    Mutator(int candidatesLimit,
-            Runtime runtime = NativeCpuRuntimeObj::getInstance())
-        : candidatesLimit(candidatesLimit), runtime(runtime){};
+    Mutator(int candidatesLimit) : candidatesLimit(candidatesLimit){};
    virtual ~Mutator(){};

    virtual vector<Graph> run(const Graph &in_graph) = 0;
-    /**
-     * @brief Merge a multi-branch graph into single branch graphs
-     *
-     * @param in_graph
-     * @return vector<Graph> Transformed graphs except the orignal one.
-     */
-    virtual vector<Graph> mergeMultiBranch(const Graph &in_graph) {
-        IT_TODO_HALT();
-    }
-    virtual bool isMultiBranchMergable(const Graph &in_graph) {
-        IT_TODO_HALT();
-    }
 };

 } // namespace infini
--- a/include/core/object.h
+++ b/include/core/object.h
@ -4,44 +4,27 @@

 namespace infini {

-using UidBaseType = int;
+using GuidBaseType = int;

-class Uid {
+class Guid {
  private:
-    UidBaseType uid;
+    GuidBaseType guid;

-  public:
-    Uid(UidBaseType uid) : uid(uid) {}
-    Uid &operator=(const Uid &rhs) = delete;
-
-    operator UidBaseType() const { return uid; }
-};
-
-class Guid : public Uid {
  private:
-    UidBaseType generateGuid() {
-        static UidBaseType guidCnt = 0;
+    GuidBaseType generateGuid() {
+        static GuidBaseType guidCnt = 0;
        return ++guidCnt;
    }

  public:
-    Guid() : Uid(generateGuid()) {}
-    Guid(const Guid &rhs) : Uid(generateGuid()) {}
-};
-
-/**
- * @brief Family unique ID. Cloned tensors shared the same FUID.
- */
-class Fuid : public Uid {
-  private:
-    UidBaseType generateFuid() {
-        static UidBaseType fuidCnt = 0;
-        return ++fuidCnt;
+    Guid() { guid = generateGuid(); }
+    Guid(const Guid &rhs) { guid = generateGuid(); }
+    Guid &operator=(const Guid &rhs) {
+        guid = generateGuid();
+        return *this;
    }

-  public:
-    Fuid() : Uid(generateFuid()) {}
-    Fuid(const Fuid &fuid) : Uid(fuid) {}
+    operator GuidBaseType() const { return guid; }
 };

 class Object {
@ -52,7 +35,7 @@ class Object {
    virtual ~Object(){};
    virtual string toString() const = 0;
    void print() { std::cout << toString() << std::endl; }
-    UidBaseType getGuid() const { return guid; }
+    GuidBaseType getGuid() const { return guid; }
 };

 inline std::ostream &operator<<(std::ostream &os, const Object &obj) {
@ -68,4 +51,4 @@ inline std::ostream &operator<<(std::ostream &os, const Ref<T> &obj) {
    return os;
 }

-} // namespace infini
+} // namespace infini
--- a/include/core/op_type.h
+++ b/include/core/op_type.h
@ -1,269 +0,0 @@
-#pragma once
-#ifndef OP_TYPE_H
-#define OP_TYPE_H
-
-#include <string>
-#include <unordered_set>
-
-namespace infini {
-
-struct OpType {
-    using underlying_t = uint16_t;
-
-    // Clang-format is ambiguous in formating of comment alignment.
-    // In order to disambiguate, it is necessary to comment all enum
-    // elements.
-    enum : underlying_t {
-        Unknown,
-        Abs,                // Unary
-        Acos,               // Unary
-        Acosh,              // Unary
-        Add,                // Binary
-        And,                // Binary
-        ArgMax,             //
-        Asin,               // Unary
-        Asinh,              // Unary
-        Atan,               // Unary
-        Atanh,              // Unary
-        AttentionKVCache,   // Fusion
-        AveragePool,        // Pool
-        BatchNormalization, //
-        Bernoulli,          //
-        BitShift,           // Binary
-        BitwiseAnd,         // Binary
-        BitwiseNot,         // Binary
-        BitwiseOr,          // Binary
-        BitwiseXor,         // Binary
-        BlackmanWindow,     //
-        Cast,               // Unary
-        CastLike,           //
-        Ceil,               // Unary
-        Celu,               //
-        CenterCropPad,      //
-        Clip,               // Unary
-        Col2lm,
-        Compress,
-        Concat,
-        ConcatFromSequence,
-        ConstantOfShape,
-        Conv,          // ComputationIntensive
-        ConvInteger,   // ComputationIntensive
-        ConvTranspose, // ComputationIntensive
-        Cos,           // Unary
-        Cosh,          // Unary
-        CumSum,
-        DFT,
-        DeformConv, // ComputationIntensive
-        DepthToSpace,
-        DequantizeLinear,
-        Det,
-        Div, // Binary
-        Dropout,
-        DynamicQuantizeLinear,
-        Einsum,
-        Elu,
-        Equal, // Compair
-        Erf,   // Unary
-        Exp,   // Unary
-        Expand,
-        EyeLike,
-        Flatten,
-        Floor, // Unary
-        GRU,
-        Gather,
-        GatherElements,
-        GatherND,
-        Gemm,
-        Gelu,              // Unary
-        GlobalAveragePool, // GlobalPool
-        GlobalLpPool,      // GlobalPool
-        GlobalMaxPool,     // GlobalPool
-        Greater,           // Compair
-        GreaterOrEqual,    // Compair
-        GridSample,
-        GroupNormalization,
-        HammingWindow,
-        HannWindow,
-        HardSigmoid,
-        HardSwish,
-        Hardmax,
-        Identity,
-        If,
-        InstanceNormalization,
-        IsInf,
-        IsNaN,
-        LRN,
-        LSTM,
-        LayerNormalization,
-        LeakyRelu,
-        Less,        // Compair
-        LessOrEqual, // Compair
-        Log,         // Unary
-        LogSoftmax,
-        Loop,
-        LpNormalization,
-        LpPool,
-        MatMul,        // ComputationIntensive
-        MatMulInteger, // ComputationIntensive
-        Max,
-        MaxPool,
-        MaxRoiPool,
-        MaxUnpool,
-        Mean,
-        MeanVarianceNormalization,
-        MelWeightMatrix,
-        Min,
-        Mish,
-        Mod,         // Binary
-        Mul,         // Binary
-        Multinomial, //
-        Neg,         // Unary
-        NegativeLogLikelihoodLoss,
-        NonMaxSuppression,
-        NonZero,
-        Not, // Unary
-        OneHot,
-        Optional,
-        OptionalGetElement,
-        OptionalHasElement,
-        Or,            // Binary
-        PRelu,         //
-        Pad,           //
-        Pow,           // Binary
-        QLinearConv,   // ComputationIntensive
-        QLinearMatMul, // ComputationIntensive
-        QuantizeLinear,
-        RNN,
-        RandomNormal,
-        RandomNormalLike,
-        RandomUniform,
-        RandomUniformLike,
-        Range,
-        Reciprocal,
-        ReduceL1,        // Reduce
-        ReduceL2,        // Reduce
-        ReduceLogSum,    // Reduce
-        ReduceLogSumExp, // Reduce
-        ReduceMax,       // Reduce
-        ReduceMean,      // Reduce
-        ReduceMin,       // Reduce
-        ReduceProd,      // Reduce
-        ReduceSum,       // Reduce
-        ReduceSumSquare, // Reduce
-        Relu,            // Unary
-        Silu,            // Unary
-        Reshape,
-        Resize,
-        ReverseSequence,
-        RoiAlign,
-        RoPE,    // Fusion
-        Round,   // Unary
-        RMSNorm, // Fusion
-        STFT,
-        Scan,
-        Scatter,
-        ScatterElements,
-        ScatterND,
-        Selu,
-        SequenceAt,
-        SequenceConstruct,
-        SequenceEmpty,
-        SequenceErase,
-        SequenceInsert,
-        SequenceLength,
-        SequenceMap,
-        Shape,
-        Shrink,
-        Sigmoid,
-        Sign,
-        Sin,  // Unary
-        Sinh, // Unary
-        Size,
-        Slice,
-        Softmax,
-        SoftmaxCrossEntropyLoss,
-        Softplus,
-        Softsign,
-        SpaceToDepth,
-        Split,
-        SplitToSequence,
-        Sqrt,
-        Squeeze,
-        StringNormalizer,
-        Sub,  // Binary
-        Sum,  //
-        Tan,  // Unary
-        Tanh, // unary
-        TfIdfVectorizer,
-        ThresholdedRelu,
-        Tile,
-        TopK,
-        Transpose,
-        Trilu,
-        Unique,
-        Unsqueeze,
-        Upsample,
-        Where,
-        Xor, // Binary
-        // CUSTOM DEFINED
-        G2BMM,
-        GBMM,
-        MemBound,
-        // TODO
-        ConvTransNHWC,
-        ConvBackwardFilter,
-        ReluBackward,
-        SigmoidBackward,
-        TanhBackward,
-
-        Fill,
-        Extend,
-        MSELoss,
-        Hardtanh,
-        L2Loss,
-        Rsqrt,
-        FloorDiv,
-        FloorMod,
-        Square,
-        SquaredDifference,
-
-        // Communication Ops
-        AllReduceSum,
-        AllReduceProd,
-        AllReduceMin,
-        AllReduceMax,
-        AllReduceAvg,
-        AllGather,
-        Broadcast,
-        Send,
-        Recv,
-    } type;
-
-    constexpr OpType(decltype(type) t) : type(t) {}
-    constexpr explicit OpType(underlying_t val) : type((decltype(type))val) {}
-    constexpr underlying_t underlying() const { return type; }
-
-    bool operator==(OpType others) const { return type == others.type; }
-    bool operator!=(OpType others) const { return type != others.type; }
-    bool operator<(OpType others) const { return type < others.type; }
-
-    const char *toString() const;
-    bool isUnary() const;
-    bool isBinary() const;
-    bool isElementWise() const;
-    bool isCompair() const;
-    bool isPool() const;
-    bool isGlobalPool() const;
-    bool isMatMulOrConv() const;
-};
-
-enum class ActType {
-    None,
-    Relu,
-    Sigmoid,
-    Tanh,
-};
-
-} // namespace infini
-
-#endif // OP_TYPE_H
--- a/include/core/operator.h
+++ b/include/core/operator.h
@ -1,14 +1,109 @@
 #pragma once
-
-#include "core/op_type.h"
 #include "core/tensor.h"

 namespace infini {
-using KernelAttrs = std::tuple<Device, OpType::underlying_t>;
+
+enum class OpType {
+    Unknown = 0,
+    // linear
+    Conv = 100,
+    Matmul,
+    ConvTrans,
+    G2BMM,
+    GBMM,
+    Pad,
+    Slice,
+    Concat,
+    Split,
+    Transpose,
+    Extend,
+    MaxPool,
+    AvgPool,
+    Add,
+    Sub,
+    Mul,
+    Div,
+    Pow,
+    Gather,
+    ReduceMean,
+    Reshape,
+    Flatten,
+    Identity,
+    // element wise
+    BatchNorm = 200,
+    Softmax,
+    Activation,
+    Relu,
+    Sigmoid,
+    Tanh,
+    Abs,
+    Resize,
+    //
+    MemBound = 300,
+};
+
+using KernelAttrs = std::tuple<Device, OpType, DataType>;
+
+class OpRegistry {
+  public:
+    static std::string getOpName(OpType opType) {
+#define FOP(op)                                                                \
+    case OpType::op:                                                           \
+        return #op
+
+        switch (opType) {
+            FOP(Unknown);
+            // linear
+            FOP(Conv);
+            FOP(Matmul);
+            FOP(ConvTrans);
+            FOP(G2BMM);
+            FOP(GBMM);
+            FOP(Pad);
+            FOP(Slice);
+            FOP(Concat);
+            FOP(Split);
+            FOP(Transpose);
+            FOP(Extend);
+            FOP(MaxPool);
+            FOP(AvgPool);
+            FOP(Add);
+            FOP(Sub);
+            FOP(Mul);
+            FOP(Div);
+            FOP(Pow);
+            FOP(Gather);
+            FOP(ReduceMean);
+            FOP(Reshape);
+            FOP(Identity);
+            // element wise
+            FOP(BatchNorm);
+            FOP(Softmax);
+            FOP(Activation);
+            FOP(Relu);
+            FOP(Sigmoid);
+            FOP(Tanh);
+            FOP(Abs);
+            //
+            FOP(MemBound);
+        default:
+            IT_ASSERT(false);
+            break;
+        }
+#undef FOP
+    }
+};
+
+enum class ActType {
+    None,
+    Relu,
+    Sigmoid,
+    Tanh,
+};

 struct OpPerfKey {
    HashType hash;
-    OpType::underlying_t opType;
+    OpType opType;
    vector<int> attrs;

  public:
@ -16,7 +111,7 @@ struct OpPerfKey {
    // https://github.com/nlohmann/json#how-can-i-use-get-for-non-default-constructiblenon-copyable-types
    OpPerfKey() = default;
    OpPerfKey(HashType hash, OpType opType, vector<int> attrs = {})
-        : hash(hash), opType(opType.underlying()), attrs(attrs) {}
+        : hash(hash), opType(opType), attrs(attrs) {}
    bool operator==(const OpPerfKey &rhs) const {
        if (hash != rhs.hash)
            return false;
@ -42,10 +137,7 @@ struct OpPerfKey {
    }
 };

-class GraphObj;
 class OperatorObj : public Object {
-    friend class GraphObj;
-
  protected:
    OpType type;
    TensorVec inputs;
@ -55,7 +147,8 @@ class OperatorObj : public Object {

  public:
    OperatorObj(OpType opType, TensorVec inputs, TensorVec outputs);
-    virtual optional<vector<Shape>> inferShape(const TensorVec &inputs) = 0;
+    virtual optional<vector<Shape>>
+    inferShape(const TensorVec &inputs) const = 0;
    virtual vector<DataType> inferDataType(const TensorVec &inputs) const;
    /**
     * @brief Constructs outputs (if requried) and check whether the operator is
@ -72,7 +165,16 @@ class OperatorObj : public Object {
     */
    HashType hash() const;

-  public:
+  public: // check Op type
+    bool isLinearOp() const;
+    bool isElementWiseOp() const;
+    bool isSplitOp() const;
+    bool isConcatOp() const;
+    bool isComputeOp() const;
+    bool isTransposeOp() const;
+    bool isReshapeOp() const;
+    bool isMemBoundOp() const;
+
  public: // getter and setter
    const TensorVec &getInputs() const { return inputs; }
    const TensorVec &getOutputs() const { return outputs; }
@ -85,27 +187,18 @@ class OperatorObj : public Object {
        IT_ASSERT(i < outputs.size(), "Index exceeded");
        return outputs.at(i);
    }
+    void addPredecessors(const Operator &op) { predecessors.emplace_back(op); }
+    void addSuccessors(const Operator &op) { successors.emplace_back(op); }
    OpVec getPredecessors() const { return wrefs_to_refs(predecessors); }
    OpVec getSuccessors() const { return wrefs_to_refs(successors); }
    OpType getOpType() const { return type; }
    // HACK: set correct data type
    DataType getDType() const { return getInputs(0)->getDType(); }
-    DataType getOutDType() const { return getOutput()->getDType(); }
    virtual int numInputs() const = 0;
    virtual int numOutputs() const = 0;

-    /**
-     * @brief Clone this operator and replace its inputs and outputs.
-     *
-     * @param newInputs
-     * @param newOutputs
-     * @return Operator
-     */
-    virtual Operator clone(const TensorVec &newInputs,
-                           const TensorVec &newOutputs) const = 0;
-
  protected:
-    optional<vector<Shape>> inferShape();
+    optional<vector<Shape>> inferShape() const;
    vector<DataType> inferDataType() const;

  private:
@ -120,26 +213,8 @@ class OperatorObj : public Object {
     * and output shapes.
     */
    virtual vector<int> getWorkloadVector() const { IT_TODO_HALT(); }
-
-    void addPredecessors(const Operator &op) { predecessors.emplace_back(op); }
-    void addSuccessors(const Operator &op) { successors.emplace_back(op); }
-    void removePredecessors(const Operator &op);
-    void removeSuccessors(const Operator &op);
-    void replaceInput(Tensor t1, Tensor t2);
 };

-#define OP_CLONE(OpObj)                                                        \
-    virtual Operator clone(const TensorVec &newInputs,                         \
-                           const TensorVec &newOutputs) const override {       \
-        auto op = infini::make_ref<OpObj>(*this);                              \
-        op->inputs = newInputs;                                                \
-        op->outputs = newOutputs;                                              \
-        op->predecessors.clear();                                              \
-        op->successors.clear();                                                \
-        IT_ASSERT(op->checkValid(nullptr));                                    \
-        return op;                                                             \
-    }
-
 } // namespace infini

 namespace std {
--- a/include/core/perf_engine.h
+++ b/include/core/perf_engine.h
@ -2,8 +2,8 @@
 #include "core/graph.h"
 #include "core/kernel.h"
 #include <nlohmann/json_fwd.hpp>
-namespace infini {
 using json = nlohmann::json;
+namespace infini {

 class PerfEngine {
  public:
@ -49,4 +49,4 @@ class PerfEngine {
 void to_json(json &j, const PerfEngine &p);
 void from_json(const json &j, PerfEngine &p);

-} // namespace infini
+} // namespace infini
--- a/include/core/ref.h
+++ b/include/core/ref.h
@ -40,4 +40,4 @@ std::vector<Ref<T>> wrefs_to_refs(const std::vector<WRef<T>> &wrefs) {
    return refs;
 }

-} // namespace infini
+} // namespace infini
--- a/include/core/runtime.h
+++ b/include/core/runtime.h
@ -1,7 +1,5 @@
 #pragma once
 #include "core/common.h"
-#include "core/communicator.h"
-#include "core/op_type.h"
 #include "core/ref.h"
 #include <memory>

@ -12,37 +10,31 @@ class TensorBaseObj;
 class TensorObj;
 class OperatorObj;
 class GraphObj;
-class GraphHandlerObj;
 class RuntimeObj;
 class BlobObj;
-template <typename T> class WorkspaceObj;

 using TensorBase = Ref<TensorBaseObj>;
 using Tensor = Ref<TensorObj>;
 using Operator = Ref<OperatorObj>;
 using Graph = Ref<GraphObj>;
-using GraphHandler = Ref<GraphHandlerObj>;
 using Runtime = Ref<RuntimeObj>;
 using Blob = Ref<BlobObj>;
-template <typename T> using Workspace = Ref<WorkspaceObj<T>>;
+enum class OpType;

 using TensorVec = vector<Tensor>;
 using OpVec = vector<Operator>;
-using OpLists = list<Operator>;

 using VType = uint32_t;

-enum class Device { CPU = 1, CUDA, BANG, INTELCPU, KUNLUN };
+enum class Device { CPU = 1, CUDA, BANG };
 /***************** Forward declaration end *****************/

 class RuntimeObj : public std::enable_shared_from_this<RuntimeObj> {
  protected:
    Device device;
-    int deviceId;

  public:
-    explicit RuntimeObj(Device device, int deviceId = 0)
-        : device(device), deviceId(deviceId) {}
+    RuntimeObj(Device device) : device(device) {}
    RuntimeObj(RuntimeObj &other) = delete;
    RuntimeObj &operator=(RuntimeObj const &) = delete;
    virtual ~RuntimeObj() {}
@ -59,6 +51,7 @@ class RuntimeObj : public std::enable_shared_from_this<RuntimeObj> {
                     bool profiling = false) const = 0;
    virtual void *alloc(size_t size) = 0;
    virtual void dealloc(void *ptr) = 0;
+
    /**
     * @brief Get the execution time of each operator in performance record. No
     * execution happens.
@ -69,12 +62,9 @@ class RuntimeObj : public std::enable_shared_from_this<RuntimeObj> {
     */
    double getPerfTime(const Graph &graph, bool profiling = false) const;
    Blob allocBlob(size_t size);
-    bool isCpu() const {
-        return device == Device::CPU || device == Device::INTELCPU;
-    }
+    bool isCpu() const { return device == Device::CPU; }
    bool isCuda() const { return device == Device::CUDA; }
    bool isBang() const { return device == Device::BANG; }
-    bool isKUNLUN() const { return device == Device::KUNLUN; }
    void copyBlob(const TensorObj *dst, const TensorObj *src) const;
    // TODO: unify these copy APIs
    virtual void copyBlobFromCPU(void *dst, const void *src,
@ -83,12 +73,6 @@ class RuntimeObj : public std::enable_shared_from_this<RuntimeObj> {
                               size_t bytes) const = 0;
    virtual string toString() const = 0;

-    int getDeviceId() const { return deviceId; }
-
-    virtual void initComm(const string &name, int worldSize, int rank) = 0;
-
-    virtual CommunicatorObj &getCommunicator() const = 0;
-
  protected:
    void printProfilingData(double totTime,
                            const std::map<OpType, double> &opTime,
@ -99,36 +83,26 @@ class RuntimeObj : public std::enable_shared_from_this<RuntimeObj> {

 class CpuRuntimeObj : public RuntimeObj {
  public:
-    CpuRuntimeObj(Device dev) : RuntimeObj(dev) {}
+    CpuRuntimeObj() : RuntimeObj(Device::CPU) {}
+    static Ref<CpuRuntimeObj> &getInstance() {
+        static Ref<CpuRuntimeObj> instance = make_ref<CpuRuntimeObj>();
+        return instance;
+    }

    void run(const Graph &graph, bool tune = false,
             bool profiling = false) const override;
-
-    void copyBlobFromCPU(void *dst, const void *src,
-                         size_t bytes) const override;
-    void copyBlobToCPU(void *dst, const void *src, size_t bytes) const override;
-    void copyBlobInsideRuntime(void *dst, const void *src,
-                               size_t bytes) const override;
-    void initComm(const string &, int, int) override { IT_TODO_HALT(); }
-
-    CommunicatorObj &getCommunicator() const override { IT_TODO_HALT(); }
-};
-
-class NativeCpuRuntimeObj : public CpuRuntimeObj {
-  public:
-    NativeCpuRuntimeObj() : CpuRuntimeObj(Device::CPU) {}
-
-    static Ref<NativeCpuRuntimeObj> &getInstance() {
-        static Ref<NativeCpuRuntimeObj> instance =
-            make_ref<NativeCpuRuntimeObj>();
-        return instance;
-    }
    void dealloc(void *ptr) override { return free(ptr); };

    void *alloc(size_t size) override {
        return calloc((size + sizeof(uint64_t) - 1) / sizeof(uint64_t),
                      sizeof(uint64_t));
    };
+
+    void copyBlobFromCPU(void *dst, const void *src,
+                         size_t bytes) const override;
+    void copyBlobToCPU(void *dst, const void *src, size_t bytes) const override;
+    void copyBlobInsideRuntime(void *dst, const void *src,
+                               size_t bytes) const override;
    string toString() const override;
 };

--- a/include/core/search_engine.h
+++ b/include/core/search_engine.h
@ -1,80 +0,0 @@
-#pragma once
-
-#include "common.h"
-#include "graph.h"
-#include "mutator.h"
-
-#include <unordered_map>
-
-namespace infini {
-class SearchEngine {
-  private:
-    Runtime runtimeExec;
-    Ref<Mutator> mutator;
-
-  public:
-    SearchEngine(Runtime _runtime, Ref<Mutator> _mutator) {
-        runtimeExec = _runtime;
-        mutator = _mutator;
-    }
-    ~SearchEngine() {}
-
-  private: // Configurations
-    size_t partitionThreshold =
-        3;                  // cut nodes whose #in + #out >= partitionThreshold
-    size_t GRAPH_SIZE = 16; // num of best graphs.
-
-  private: // Composed objects
-    std::shared_ptr<Mutator> mutationEngine;
-
-  public:
-    std::shared_ptr<Mutator> getMutationEngine() { return mutationEngine; };
-    struct GroupEdge {
-        int v, next;
-        GroupEdge() = delete;
-    };
-
-    struct Candidate { // a graph with perf
-        std::shared_ptr<Graph> graph;
-        double perf = INFINITY;
-    };
-    class MetaGraph { // a graph of subgraphs, for searching.
-      public:
-        MetaGraph() {}
-        ~MetaGraph() {}
-        struct Node {
-            Graph graph;
-            std::vector<int> suc;
-            std::vector<int> pre;
-            int type, cnt;
-        };
-        std::vector<Node> nodes;
-    };
-
-    Graph run(const Graph graph);                  // entrance of search engine.
-    std::vector<Graph> search(const Graph &graph); // search for a partition.
-
-  private:
-    std::vector<Graph> partitionGraph(const Graph graph);
-    std::shared_ptr<MetaGraph> buildMetaGraphWithGraph(const Graph graph);
-    std::shared_ptr<MetaGraph>
-    buildMetaGraphWithPlan(const std::shared_ptr<MetaGraph> metaGraph,
-                           const std::vector<int> &plan);
-    // search horizontal merges
-    std::vector<std::shared_ptr<MetaGraph>>
-    searchMerge(std::shared_ptr<MetaGraph> &metaGraph);
-    void searchMergeDfs(std::shared_ptr<MetaGraph> &metaGraph,
-                        std::vector<int> &plan, std::vector<int> &frontier,
-                        std::vector<std::vector<int>> &plans,
-                        std::unordered_set<uint64_t> &planSet);
-    std::vector<Graph>
-    searchMutation(const std::shared_ptr<MetaGraph> &metaGraph);
-
-    void printMetaGraph(Ref<SearchEngine::MetaGraph> metaGraph);
-    /**
-     * @brief Check whether a multi-brach graph can be merged into a single
-     * branch.
-     */
-    bool isMultiBranchMergable(const Graph graph);
-};
-} // namespace infini
--- a/include/core/tensor.h
+++ b/include/core/tensor.h
@ -1,17 +1,7 @@
 #pragma once
 #include "core/tensor_base.h"
-#include "core/tensor_type.h"
-#include "utils/data_convert.h"
 #include <cmath>
-#include <cstring>
-#include <fstream>

-#if USE_CUDA
-#include "cuda/cuda_runtime.h"
-#endif
-#if USE_BANG
-#include "bang/bang_runtime.h"
-#endif
 namespace infini {

 // TODO: how to deal with this
@ -20,214 +10,83 @@ using Shape = vector<ShapeElem>;
 class TensorObj : public TensorBaseObj {
  private:
    Shape shape;
-    size_t _size; // Cache of Π(shape).
-    Fuid fuid;    // Cloned tensors share the same id. Tensors constructed from
-                  // scratch have a new id.
-    TensorType tensorType = TensorType::others;

  public:
-    TensorObj(Shape shape, DataType dtype, Runtime runtime);
+    TensorObj(const Shape &shape, DataType dtype, Runtime runtime);
    virtual ~TensorObj() {}
    string toString() const override;

-    size_t size() const { return _size; }
-    size_t getBytes() const { return _size * dtype.getSize(); }
+    size_t size() const;
+    size_t getBytes() const;

    Shape getDims() const { return shape; }
-    void setShape(Shape shape_);
-    size_t getRank() const { return shape.size(); }
-    Shape getStride() const;
-    size_t getOffset(const vector<int> &ds) const;
+    vector<size_t> getStride() const;
+    size_t getOffset(const Shape &ds) const;
+    using TensorBaseObj::getData;
+    VType getData(const Shape &pos) const;
    void dataMalloc();
-    UidBaseType getFuid() const { return fuid; }
-    bool isWeight() const { return tensorType == TensorType::weight; }
-    bool isInput() const { return tensorType == TensorType::input; }
-    bool isOutput() const { return tensorType == TensorType::output; }
-    bool isOthers() const { return tensorType == TensorType::others; }
-    void setWeight() { tensorType = TensorType::weight; }
-    void setInput() {
-        if (!this->isWeight()) {
-            tensorType = TensorType::input;
-        }
-    }
-    void setOutput() {
-        if (!this->isWeight()) {
-            tensorType = TensorType::output;
-        }
-    }
-    string tensorTypeToString() const {
-        switch (tensorType) {
-        case TensorType::weight:
-            return "weight";
-            break;
-        case TensorType::input:
-            return "input";
-            break;
-        case TensorType::output:
-            return "output";
-            break;
-        case TensorType::others:
-            return "others";
-            break;
-
-        default:
-            return "unknown tensor type";
-            break;
-        }
-    }

    void load(std::string file_path);
    void save(std::string file_path);

-    void copyin(const void *ptr, size_t size) {
-        runtime->copyBlobFromCPU(getRawDataPtr<void *>(), ptr, size);
-    }
-    void copyout(void *ptr, size_t size) const {
-        runtime->copyBlobToCPU(ptr, getRawDataPtr<void *>(), size);
+    template <typename T> void copyData(const T *dptr) {
+        IT_ASSERT(DataType::get<T>() == dtype);
+        IT_ASSERT(data != nullptr);
+        runtime->copyBlobFromCPU(getRawDataPtr<void *>(), dptr, getBytes());
    }

-    // Copy elements from `data`.
-    template <typename T> void copyin(const vector<T> &data) {
-        IT_ASSERT(DataType::get<T>() == dtype.cpuTypeInt());
-        IT_ASSERT(data.size() == _size);
-        copyin(data.data(), getBytes());
-    }
-    // Copy all the elements to a vector.
-    template <typename T> auto copyout() const {
-        IT_ASSERT(DataType::get<T>() == dtype.cpuTypeInt());
-        std::vector<T> ans(_size);
-        copyout(ans.data(), getBytes());
-        return ans;
-    }
-    // Copy the element at `pos`.
-    template <typename T> auto copyOne(const vector<int> &pos) const {
-        IT_ASSERT(DataType::get<T>() == dtype.cpuTypeInt());
-        auto offset = getOffset(pos);
-        auto bytes = dtype.getSize();
-        T ans;
-        runtime->copyBlobToCPU(
-            &ans, getRawDataPtr<uint8_t *>() + offset * bytes, bytes);
-        return ans;
+    template <typename T> void copyData(vector<T> dataVector) {
+        IT_ASSERT(DataType::get<T>() == dtype);
+        IT_ASSERT(dataVector.size() >= size());
+        copyData(dataVector.data());
    }

    void copyData(const TensorObj *src);
    void copyData(const Tensor &src) { copyData(src.get()); }
-
-    // TODO: Rename this function later, because it is confused that it will
-    // change the field data, but actually it generates data and maybe copy to
-    // device.
-    // FIXME: std::fucntion copies the generator instead of passing it by ref.
-    // Thus the internal state of generator cannot be updated.
    void setData(
-        std::function<void(void *, size_t, DataType)> const &generator) const;
-
-    void setDataBlob(const Blob &blob);
-
-    Tensor clone() const {
-        auto obj = make_ref<TensorObj>(*this);
-        obj->freeData();
-        obj->targets.clear();
-        obj->source.reset();
-        return obj;
-    }
-    Tensor clone(Runtime runtime) const {
-        auto obj = make_ref<TensorObj>(*this);
-        obj->runtime = runtime;
-        obj->freeData();
-        obj->targets.clear();
-        obj->source.reset();
-        if (hasData()) {
-            obj->dataMalloc();
-            obj->copyData(this);
+        const std::function<void(void *, size_t, DataType)> &generator) const {
+        IT_ASSERT(data != nullptr);
+        if (!runtime->isCpu()) {
+            IT_TODO_HALT();
        }
+        generator(data->getPtr<void *>(), size(), dtype);
+    }
+    Tensor clone(Runtime runtime) {
+        auto obj = make_ref<TensorObj>(shape, dtype, runtime);
+        obj->dataMalloc();
+        obj->copyData(this);
        return obj;
    }

    void printData() const;
-    void dumpData(std::ofstream &ofs) const;
-    bool equalData(const Tensor &rhs, double relativeError = 1e-6) const;
+    bool equalData(const Tensor &rhs) const;

    template <typename T> bool equalData(const vector<T> &dataVector) {
+        IT_ASSERT(DataType::get<T>() == dtype);
        IT_ASSERT(size() == dataVector.size());
-        if (dtype == DataType::Float16) {
-            return equalDataImpl_fp16(getRawDataPtr<uint16_t *>(),
-                                      (float *)dataVector.data(), size());
-        }
-        IT_ASSERT(DataType::get<T>() == dtype.cpuTypeInt());
        return equalDataImpl(getRawDataPtr<T *>(), dataVector.data(), size());
    }

    size_t getOffsetByBroadcastOffset(size_t bcOffset, Shape bcShape) const;

  private:
-    template <class T> string dataToString() const {
-        std::stringstream builder;
-        builder << "Tensor: " << guid << std::endl;
-
-        auto numDims = shape.size();
-        auto dimSzVec = vector<int>(numDims, 1);
-        auto ptr = data->getPtr<T *>();
-        dimSzVec[numDims - 1] = shape[numDims - 1];
-
-        for (int i = numDims - 1; i != 0; --i)
-            dimSzVec[i - 1] = dimSzVec[i] * shape[i - 1];
-
-        for (size_t i = 0, iEnd = size(); i < iEnd; ++i) {
-            for (size_t j = 0; j < numDims; ++j)
-                if (i % dimSzVec[j] == 0)
-                    builder << "[";
-
-            builder << ptr[i];
-            for (size_t j = 0; j < numDims; ++j)
-                if ((int)i % dimSzVec[j] == dimSzVec[j] - 1)
-                    builder << "]";
-
-            if (i != size() - 1)
-                builder << ", ";
-
-            auto column = (size_t)dimSzVec[numDims - 1];
-            if (i % column == column - 1)
-                builder << std::endl;
-        }
-        return builder.str();
-    }
+    void printDataFloat() const;
+    void printDataUint32_t() const;

    template <typename T>
-    bool equalDataImpl(const T *a, const T *b, size_t size,
-                       double relativeError = 1e-6) const {
+    bool equalDataImpl(const T *a, const T *b, size_t size) const {
        for (size_t i = 0; i < size; ++i) {
            if constexpr (std::is_integral_v<T>) {
                if (a[i] != b[i])
                    return false;
            } else if constexpr (std::is_floating_point_v<T>) {
-                if (std::min(fabs(a[i]), fabs(b[i])) == 0. &&
-                    fabs(a[i] - b[i]) > relativeError) {
-                    printf("Error on %lu: %f %f\n", i, a[i], b[i]);
-                    return false;
-                } else if (std::min(fabs(a[i]), fabs(b[i])) != 0. &&
-                           fabs(a[i] - b[i]) /
-                                   std::max(fabs(a[i]), fabs(b[i])) >
-                               relativeError) {
+                if (fabs(a[i] - b[i]) / std::max(fabs(a[i]), fabs(b[i])) >
+                    1e-6) {
                    printf("Error on %lu: %f %f\n", i, a[i], b[i]);
                    return false;
                }
-            } else {
+            } else
                static_assert(!sizeof(T), "Unsupported data type");
-            }
-        }
-        return true;
-    }
-
-    bool equalDataImpl_fp16(const uint16_t *a, const float *b,
-                            size_t size) const {
-        for (size_t i = 0; i < size; ++i) {
-            auto a_fp32 = fp16_to_float(a[i]);
-            auto b_fp32 = b[i];
-            if (fabs(a_fp32 - b_fp32) / std::max(fabs(a_fp32), fabs(b_fp32)) >
-                1e-6) {
-                printf("Error on %lu: %f %f\n", i, a_fp32, b_fp32);
-                return false;
-            }
        }
        return true;
    }
@ -248,8 +107,8 @@ class TensorObj : public TensorBaseObj {
    //         // std::cerr << "Init beginned " << std::endl;
    // #pragma omp parallel for
    //         for (size_t i = 0; i < iEnd; ++i)
-    //             data[i] = fastrand(random_seed[omp_get_thread_num() *
-    //             16]) % 10000;
+    //             data[i] = fastrand(random_seed[omp_get_thread_num() * 16]) %
+    //             10000;
    //         // std::cerr << "Init finished" << std::endl;
    //         computed = ComputedFull;
    //         return true;
@ -294,8 +153,8 @@ class TensorObj : public TensorBaseObj {
    //         auto nDim = dims.size();
    //         auto nBroadcastDim = ds.size() - nDim;
    //         for (size_t i = 0; i < nDim; ++i)
-    //             if (ds[nBroadcastDim + i] < 0 || ds[nBroadcastDim +
-    //             i] >= dims[i])
+    //             if (ds[nBroadcastDim + i] < 0 || ds[nBroadcastDim + i] >=
+    //             dims[i])
    //                 return (size_t)-1;
    //         size_t idx = 0;
    //         for (size_t i = 0; i < nDim; ++i)
@ -354,14 +213,12 @@ class TensorObj : public TensorBaseObj {
    //         return (g_seed >> 16) & 0x7FFF;
    //     }

-    //     std::vector<std::vector<int>> const *getSplittingPoints()
-    //     const {
+    //     std::vector<std::vector<int>> const *getSplittingPoints() const {
    //         assert(!splittingPoints.empty());
    //         return &splittingPoints;
    //     }

-    //     bool setSplittingPoints(std::vector<std::vector<int>> value)
-    //     {
+    //     bool setSplittingPoints(std::vector<std::vector<int>> value) {
    //         assert(!value.empty());
    //         splittingPoints = value;
    //         return true;
@ -383,7 +240,7 @@ class TensorObj : public TensorBaseObj {
    //     }

    //     void initSplittingPoints() {
-    //     splittingPoints.resize(getRank()); }
+    //     splittingPoints.resize(getDims().size()); }

    //     void printShape();
 };
--- a/include/core/tensor_base.h
+++ b/include/core/tensor_base.h
@ -3,11 +3,10 @@
 #include "core/data_type.h"
 #include "core/object.h"
 #include "core/runtime.h"
-namespace infini {
-class GraphObj;
-class TensorBaseObj : public Object {
-    friend class GraphObj;

+namespace infini {
+
+class TensorBaseObj : public Object {
  public:
    // enum TensorType {
    //     Input,
@ -20,8 +19,8 @@ class TensorBaseObj : public Object {
    int dim;

    DataType dtype;
-    vector<WRef<OperatorObj>> targets;
-    WRef<OperatorObj> source;
+    vector<WRef<OperatorObj>> inputOf;
+    WRef<OperatorObj> outputOf;
    Blob data;
    Runtime runtime;

@ -34,39 +33,23 @@ class TensorBaseObj : public Object {
        data = blob;
    }
    Blob getDataBlob() const { return data; }
-    bool hasData() const { return data != nullptr; }
-    void freeData() { data = nullptr; }
    template <typename T> T getRawDataPtr() const {
        static_assert(std::is_pointer_v<T>,
                      "Raw data pointer has a type of pointer");
        IT_ASSERT(data != nullptr);
        return data->getPtr<T>();
    }
+    VType getData(size_t offset) const;

    DataType getDType() const { return dtype; }
-    int getDTypeIndex() const { return dtype.getIndex(); }
    Runtime getRuntime() const { return runtime; }

+    void addInputOf(const Operator &op) { inputOf.emplace_back(op); }
+    void setOutputOf(const Operator &op) { outputOf = op; }
+    OpVec getInputOf() { return wrefs_to_refs(inputOf); }
+    Operator getOutputOf() { return outputOf.lock(); }
    //     std::pair<Operator *, int> getOutputOfWithIndex();

-    bool hasTarget() const { return !targets.empty(); }
-
-    OpVec getTargets() const { return wrefs_to_refs(targets); }
-    Operator getSource() const { return source.lock(); }
-
-  private:
-    void addTarget(const Operator &op) { targets.emplace_back(op); }
-    void setSource(const Operator &op) { source = op; }
-    void removeTarget(const Operator &op) {
-        for (auto itr = targets.begin(); itr != targets.end();) {
-            if (itr->lock() == op)
-                itr = targets.erase(itr);
-            else
-                ++itr;
-        }
-    }
-    //     std::pair<Operator *, int> getSourceWithIndex();
-
    //     bool setScalar(VType val) {
    //         if (data == nullptr || !dims.empty())
    //             return false;
--- a/include/core/tensor_type.h
+++ b/include/core/tensor_type.h
@ -1,7 +0,0 @@
-#pragma once
-
-namespace infini {
-
-enum class TensorType { weight, input, output, others };
-
-} // namespace infini
--- a/include/core/workspace.h
+++ b/include/core/workspace.h
@ -1,42 +0,0 @@
-#pragma once
-#include "core/runtime.h"
-
-namespace infini {
-
-template <class T> class WorkspaceObj {
-  private:
-    T workspace;           // workspace pointer
-    size_t workspaceSize;  // Size of workspace
-    size_t workspaceAlloc; // currently use workspace size
-
-  public:
-    WorkspaceObj(T workspace_, size_t workspaceSize_)
-        : workspace(workspace_), workspaceSize(workspaceSize_) {
-        workspaceAlloc = 0;
-    }
-    virtual ~WorkspaceObj() {
-        // Dealloc workspace in RuntimeObj
-        // Set workspace = nullptr here
-        workspace = nullptr;
-    }
-    size_t getWorkspaceSize() const { return workspaceSize; }
-
-    T getWorkspace(size_t size) {
-        // Get unused workspace
-        IT_ASSERT(size + workspaceAlloc <= workspaceSize);
-        auto ret = (T)(static_cast<uint8_t *>(workspace) + workspaceAlloc);
-        workspaceAlloc += size;
-        return ret;
-    }
-    T getWorkspace() {
-        // Override getWorkspace in order to dealloc in runtime
-        return workspace;
-    }
-    void resetWorkspace() {
-        // Reset workspaceAlloc every time end kernel
-        workspaceAlloc = 0;
-    }
-    size_t getWorkspaceAlloc() const { return workspaceAlloc; }
-};
-
-} // namespace infini
--- a/include/cuda/cuda_attention_kvcache.h
+++ b/include/cuda/cuda_attention_kvcache.h
@ -1,17 +0,0 @@
-#pragma once
-#include "core/common.h"
-#include <cstdio>
-
-struct AttentionKVCacheMetadata {
-    int dimSize[4];
-    int stride[4];
-};
-
-namespace infini {
-void attention_kvcache_kernel(float *input_k_cache, float *input_v_cache,
-                              float *input_q, float *input_k, float *input_v,
-                              int *position_id, float *output_matmul,
-                              const AttentionKVCacheMetadata &compMeta,
-                              float *output_O_temp, float *output_sum_temp);
-
-} // namespace infini
--- a/include/cuda/cuda_clip.h
+++ b/include/cuda/cuda_clip.h
@ -1,9 +0,0 @@
-#pragma once
-
-#include "operators/unary.h"
-
-namespace infini {
-void clip_kernel(float *input, float *output, int num, float minValue,
-                 float maxValue);
-
-}; // namespace infini
--- a/include/cuda/cuda_common.h
+++ b/include/cuda/cuda_common.h
@ -5,13 +5,17 @@
 #include <cuda_profiler_api.h>
 #include <cudnn.h>
 #include <curand.h>
-#include <memory>

+// TODO: replace with Exception (IT_ASSERT)
 #define checkCudaError(call)                                                   \
-    if (auto err = call; err != cudaSuccess)                                   \
-    throw ::infini::Exception(std::string("[") + __FILE__ + ":" +              \
-                              std::to_string(__LINE__) + "] CUDA error (" +    \
-                              #call + "): " + cudaGetErrorString(err))
+    {                                                                          \
+        auto err = call;                                                       \
+        if (cudaSuccess != err) {                                              \
+            fprintf(stderr, "Cuda error in %s:%i : %s.\n", __FILE__, __LINE__, \
+                    cudaGetErrorString(err));                                  \
+            exit(EXIT_FAILURE);                                                \
+        }                                                                      \
+    }

 #define checkCUresult(call)                                                    \
    {                                                                          \
@ -19,8 +23,9 @@
        const char *errName;                                                   \
        if (CUDA_SUCCESS != err) {                                             \
            cuGetErrorString(err, &errName);                                   \
-            IT_ASSERT(err == CUDA_SUCCESS,                                     \
-                      (string("CU error: ") + string(errName)));               \
+            fprintf(stderr, "Cuda error in %s:%i : %s.\n", __FILE__, __LINE__, \
+                    errName);                                                  \
+            exit(EXIT_FAILURE);                                                \
        }                                                                      \
    }

@ -35,10 +40,14 @@
    }

 #define checkCudnnError(call)                                                  \
-    if (auto err = call; err != CUDNN_STATUS_SUCCESS)                          \
-    throw ::infini::Exception(std::string("[") + __FILE__ + ":" +              \
-                              std::to_string(__LINE__) + "] cuDNN error (" +   \
-                              #call + "): " + cudnnGetErrorString(err))
+    {                                                                          \
+        auto err = call;                                                       \
+        if (CUDNN_STATUS_SUCCESS != err) {                                     \
+            fprintf(stderr, "cuDNN error in %s:%i : %s.\n", __FILE__,          \
+                    __LINE__, cudnnGetErrorString(err));                       \
+            exit(EXIT_FAILURE);                                                \
+        }                                                                      \
+    }

 #define checkCurandError(call)                                                 \
    {                                                                          \
@ -112,20 +121,4 @@ inline const char *curandGetErrorString(curandStatus_t error) {

 using CudaPtr = void *;

-class CUDAStream {
-  public:
-    CUDAStream(const CUDAStream &) = delete;
-    CUDAStream(CUDAStream &&) = delete;
-    void operator=(const CUDAStream &) = delete;
-    void operator=(CUDAStream &&) = delete;
-    static cudaStream_t getCurrentStream() { return _stream; }
-    static void Init() { CUDAStream::_stream = 0; };
-    static void createStream() { checkCudaError(cudaStreamCreate(&_stream)); }
-    static void destroyStream() { checkCudaError(cudaStreamDestroy(_stream)); }
-
-  private:
-    CUDAStream(){};
-    static cudaStream_t _stream;
-};
-
 } // namespace infini
--- a/include/cuda/cuda_element_wise.h
+++ b/include/cuda/cuda_element_wise.h
@ -1,20 +1,6 @@
 #pragma once

 namespace infini {
-void div_kernel(int dtypeIndex, void *a, void *b, void *c, int a0, int a1,
-                int a2, int a3, int b0, int b1, int b2, int b3, int c0, int c1,
-                int c2, int c3);
-void add_kernel(int dtypeIndex, void *a, void *b, void *c, int a0, int a1,
-                int a2, int a3, int b0, int b1, int b2, int b3, int c0, int c1,
-                int c2, int c3);
-void pow_kernel(int dtypeIndex, void *a, void *b, void *c, int a0, int a1,
-                int a2, int a3, int b0, int b1, int b2, int b3, int c0, int c1,
-                int c2, int c3);
-void less_kernel(int dtypeIndex, void *a, void *b, void *c, int a0, int a1,
-                 int a2, int a3, int b0, int b1, int b2, int b3, int c0, int c1,
-                 int c2, int c3);
-
-void div_const_kernel(int dType, void *a, void *b, void *c, size_t n);
-
-void pow_const_kernel(int dType, void *a, void *b, void *c, size_t n);
+void div_kernel(float *a, float *b, float *c, int num);
+void pow_kernel(float *a, float *b, float *c, int num);
 }; // namespace infini
--- a/include/cuda/cuda_expand.h
+++ b/include/cuda/cuda_expand.h
@ -1,12 +0,0 @@
-#pragma once
-
-#include "operators/unary.h"
-#include "utils/small_array.h"
-namespace infini {
-void expandKernel(int dType, void *input, void *output, int nDims,
-                  int outputsize, SmallArray inputShape,
-                  SmallArray outputShape);
-
-void expandRowKernel(int dType, void *input, void *output, int n_rows,
-                     int row_len);
-}; // namespace infini
--- a/include/cuda/cuda_layernorm.h
+++ b/include/cuda/cuda_layernorm.h
@ -1,17 +0,0 @@
-#pragma once
-#include "operators/unary.h"
-
-namespace infini {
-void LaynormKernel(const float *input, const float *scale, const float eps,
-                   int size, int scaleSize, const int dimsize, const int stride,
-                   float *output, const float *bias, int biasSize);
-void LaynormKernel(const float *input, const float *scale, const float eps,
-                   int size, int scaleSize, const int dimsize, const int stride,
-                   float *output);
-void LaynormKernel(const half *input, const half *scale, const half eps,
-                   int size, int scaleSize, const int dimsize, const int stride,
-                   half *output, const half *bias, int biasSize);
-void LaynormKernel(const half *input, const half *scale, const half eps,
-                   int size, int scaleSize, const int dimsize, const int stride,
-                   half *output);
-}; // namespace infini
--- a/include/cuda/cuda_pad_slice.h
+++ b/include/cuda/cuda_pad_slice.h
@ -10,11 +10,10 @@ typedef struct {
    int wholeNDim[MAX_DIM];  // dim size after padding or before slicing
    int partNDim[MAX_DIM];   // dim size before padding or after slicing
    int partStride[MAX_DIM]; // stride before padding or after slicing
-    int DType;
 } TransMetaData;

 namespace infini {
-void pad_slice_kernel(void *partData, void *wholeData,
+void pad_slice_kernel(float *partData, float *wholeData,
                      const TransMetaData &metadata, int nDims, int num,
                      bool isPad);
 } // namespace infini
--- a/include/cuda/cuda_rmsnorm.h
+++ b/include/cuda/cuda_rmsnorm.h
@ -1,10 +0,0 @@
-#pragma once
-
-#include "operators/rms_norm.h"
-
-namespace infini {
-
-void rmsnorm_kernel(int dType, void *input, void *weight, void *output,
-                    int num_tokens, int hidden_size);
-
-}; // namespace infini
--- a/include/cuda/cuda_rope.h
+++ b/include/cuda/cuda_rope.h
@ -1,12 +0,0 @@
-#pragma once
-
-#include "operators/rope.h"
-#include "utils/small_array.h"
-
-namespace infini {
-
-void rope_kernel(int dType, int *pos, void *input, void *output, int size,
-                 int dim_model, int dim_head, int hidden_stride,
-                 int pos_stride);
-
-}; // namespace infini
--- a/include/cuda/cuda_runtime.h
+++ b/include/cuda/cuda_runtime.h
@ -1,9 +1,6 @@
 #pragma once
 #include "core/runtime.h"
 #include "cuda/cuda_common.h"
-#ifdef INFINI_USE_NCCL
-#include "cuda/nccl_communicator.h"
-#endif

 namespace infini {

@ -11,40 +8,34 @@ class CudaRuntimeObj : public RuntimeObj {
  private:
    cudnnHandle_t cudnn;
    cublasHandle_t cublas;
-    std::unique_ptr<CommunicatorObj> comm;
    CudaPtr workspace;
    size_t workspaceSize;
-    bool isCudaGraphCreated;
-    cudaGraph_t cudaGraph;
-    cudaGraphExec_t cudaGraphInstance;

  public:
-    explicit CudaRuntimeObj(int deviceId = 0)
-        : RuntimeObj(Device::CUDA, deviceId) {
+    CUdevice cuDevice;
+    CUcontext newContext;
+
+  public:
+    CudaRuntimeObj() : RuntimeObj(Device::CUDA) {
+        // Prepare for nvrtc. cuCtxCreate should be called befero others.
+        // Otherwise it will result in strange failure, such as cuBLAS failed on
+        // certian inputs.
+        checkCUresult(cuInit(0));
+        checkCUresult(cuDeviceGet(&cuDevice, 0));
+        checkCUresult(cuCtxCreate(&newContext, 0, cuDevice));

-        checkCudaError(cudaSetDevice(deviceId));
        checkCudnnError(cudnnCreate(&cudnn));
        checkCublasError(cublasCreate(&cublas));
        // 10GB for Longformer
        // size_t longformerNum = 3lu * (1 << 30);
        workspaceSize = 7ll << 30; // 7 GB
        workspace = alloc(workspaceSize);
-        isCudaGraphCreated = false;
-        CUDAStream::Init();
    }
    virtual ~CudaRuntimeObj() {
-        try {
-            if (isCudaGraphCreated) {
-                checkCudaError(cudaGraphExecDestroy(cudaGraphInstance));
-                checkCudaError(cudaGraphDestroy(cudaGraph));
-                CUDAStream::destroyStream();
-            }
-            dealloc(workspace);
-            checkCudnnError(cudnnDestroy(cudnn));
-            checkCublasError(cublasDestroy(cublas));
-        } catch (const std::exception &e) {
-            std::cerr << "Error in ~CudaRuntimeObj: " << e.what() << std::endl;
-        }
+        dealloc(workspace);
+        checkCudnnError(cudnnDestroy(cudnn));
+        checkCublasError(cublasDestroy(cublas));
+        checkCUresult(cuCtxDestroy(newContext));
    }
    string toString() const override;

@ -56,7 +47,6 @@ class CudaRuntimeObj : public RuntimeObj {
    CudaPtr alloc(size_t size) override {
        void *ptr;
        checkCudaError(cudaMalloc(&ptr, size));
-        // printf("cuda malloc: %p %lu bytes\n", ptr, size);
        return ptr;
    }
    void dealloc(void *ptr) override { checkCudaError(cudaFree(ptr)); }
@ -85,13 +75,6 @@ class CudaRuntimeObj : public RuntimeObj {

    void runWithoutSync(const Graph &graph) const;

-    void runWithCudaGraph(const Graph &graph);
-
-    // init communicator
-    void initComm(const string &name, int worldSize, int rank) final;
-
-    CommunicatorObj &getCommunicator() const final { return *comm; }
-
  private:
    void tune(const Graph &graph, bool profiling) const;
 };
--- a/include/cuda/cuda_softmax.h
+++ b/include/cuda/cuda_softmax.h
@ -1,8 +0,0 @@
-#pragma once
-#include "utils/small_array.h"
-namespace infini {
-void softmax_kernel(int num_blocks, float *input, float *output, int size,
-                    int dimsize, int stride);
-void softmax_kernel(int num_blocks, half *input, half *output, int size,
-                    int dimsize, int stride);
-} // namespace infini
--- a/include/cuda/cuda_split_concat.h
+++ b/include/cuda/cuda_split_concat.h
@ -3,13 +3,13 @@
 #include <cstdio>

 const int BATCH_SIZE = 32; // parallel tensor number.
-const int DIM_MAX_SIZE = 8;
+const int DIM_MAX_SIZE = 4;

 // Concat operator acts like element tensors composing to one big tensor,and
 // split operator acts like one big tensor being composed by element
 // tensors.
-template <typename T> struct ElementTensorMetadata {
-    T *data[BATCH_SIZE];
+struct ElementTensorMetadata {
+    float *data[BATCH_SIZE];
    int dimBgNo[BATCH_SIZE]; // the dimention begin no of the element tensor in
                             // the composed tensor.
    int dimSize[BATCH_SIZE]; // the dimention size of the element tensor.
@ -20,17 +20,16 @@ template <typename T> struct ElementTensorMetadata {
                   data[i], dimBgNo[i], dimSize[i], nElements[i]);
    }
 };
-template <typename T> struct ComposedTensorMetadata {
+
+struct ComposedTensorMetadata {
    int dimSize[DIM_MAX_SIZE];
    int stride[DIM_MAX_SIZE];
-    T *data;
+    float *data;
 };

 namespace infini {
-void split_concat_kernel(const ElementTensorMetadata<float> &eleMeta,
-                         const ComposedTensorMetadata<float> &compMeta, int dim,
-                         int batchSize, int nDims, bool isSplit);
-void split_concat_kernel(const ElementTensorMetadata<half> &eleMeta,
-                         const ComposedTensorMetadata<half> &compMeta, int dim,
+void split_concat_kernel(const ElementTensorMetadata &eleMeta,
+                         const ComposedTensorMetadata &compMeta, int dim,
                         int batchSize, int nDims, bool isSplit);
+
 } // namespace infini
--- a/include/cuda/cuda_transpose.h
+++ b/include/cuda/cuda_transpose.h
@ -1,11 +0,0 @@
-#pragma once
-
-#include "operators/transpose.h"
-#include "utils/small_array.h"
-
-namespace infini {
-
-void transpose_kernel(int dType, void *input, void *output, int nDims, int size,
-                      SmallArray strides, SmallArray outputShape);
-
-}; // namespace infini
--- a/include/cuda/cuda_unary.h
+++ b/include/cuda/cuda_unary.h
@ -3,22 +3,31 @@
 #include "operators/unary.h"

 namespace infini {
-template <typename T> void softmax_kernel(T *input, T *output, size_t num);
-template <typename T> void relu_kernel(T *input, T *output, size_t num);
-template <typename T> void silu_kernel(T *input, T *output, size_t num);
-template <typename T> void sigmoid_kernel(T *input, T *output, size_t num);
-template <typename T> void tanh_kernel(T *input, T *output, size_t num);
-template <typename T> void abs_kernel(T *input, T *output, size_t num);
-template <typename T> void sqrt_kernel(T *input, T *output, size_t num);
-template <typename T> void neg_kernel(T *input, T *output, size_t num);
-template <typename T> void gelu_kernel(T *input, T *output, size_t num);
-template <typename T> void erf_kernel(T *input, T *output, size_t num);
-template <typename T> void hard_sigmoid_kernel(T *input, T *output, size_t num);
-template <typename T> void hard_swish_kernel(T *input, T *output, size_t num);
+void softmax_kernel(float *input, float *output, int num);
+void relu_kernel(float *input, float *output, int num);
+void sigmoid_kernel(float *input, float *output, int num);
+void tanh_kernel(float *input, float *output, int num);
+void abs_kernel(float *input, float *output, int num);

-template <typename INPUT, typename OUTPUT>
-void cast_kernel(INPUT *input, OUTPUT *output, size_t num);
+void unary_kernel(const Operator &_op) {
+    auto op = as<UnaryObj>(_op);
+    float *const inputData = (op->getInputs(0)->getRawDataPtr<float *>());
+    float *const outputData = (op->getOutput()->getRawDataPtr<float *>());

-void unary_kernel(const Operator &_op);
+    auto dim = op->getInputs(0)->getDims();
+    int n = dim[0], c = dim[1], h = dim[2], w = dim[3];
+    if (op->getOpType() == OpType::Softmax)
+        softmax_kernel(inputData, outputData, n * c * h * w);
+    else if (op->getOpType() == OpType::Relu)
+        relu_kernel(inputData, outputData, n * c * h * w);
+    else if (op->getOpType() == OpType::Sigmoid)
+        sigmoid_kernel(inputData, outputData, n * c * h * w);
+    else if (op->getOpType() == OpType::Tanh)
+        tanh_kernel(inputData, outputData, n * c * h * w);
+    else if (op->getOpType() == OpType::Abs)
+        abs_kernel(inputData, outputData, n * c * h * w);
+    else
+        IT_TODO_HALT();
+}

 }; // namespace infini
--- a/include/cuda/cuda_utility.h
+++ b/include/cuda/cuda_utility.h
@ -1,29 +1,11 @@
-#pragma once
 #include "core/tensor.h"
-#include "cuda/cuda_common.h"

 namespace infini {

 void cudaPrintFloat(float *x, int len);

-void cudaPrintTensor(const Tensor &tensor);
+void cudaPrintTensor(const Tensor &tensor) {
+    cudaPrintFloat(tensor->getRawDataPtr<float *>(), tensor->size());
+}

-cudnnDataType_t cudnnDataTypeConvert(DataType dataType);
-cudaDataType cublasDataTypeConvert(DataType);
-
-template <int index> struct DT_CUDA {};
-template <> struct DT_CUDA<0> { using t = bool; };
-template <> struct DT_CUDA<1> { using t = float; };
-template <> struct DT_CUDA<2> { using t = unsigned char; };
-template <> struct DT_CUDA<3> { using t = char; };
-template <> struct DT_CUDA<4> { using t = unsigned short; };
-template <> struct DT_CUDA<5> { using t = short; };
-template <> struct DT_CUDA<6> { using t = int; };
-template <> struct DT_CUDA<7> { using t = long long; };
-template <> struct DT_CUDA<9> { using t = bool; };
-template <> struct DT_CUDA<10> { using t = half; };
-template <> struct DT_CUDA<11> { using t = double; };
-template <> struct DT_CUDA<12> { using t = unsigned int; };
-template <> struct DT_CUDA<13> { using t = unsigned long long; };
-template <> struct DT_CUDA<16> { using t = nv_bfloat16; };
-} // namespace infini
+} // namespace infini
--- a/include/cuda/cuda_where.h
+++ b/include/cuda/cuda_where.h
@ -1,17 +0,0 @@
-#pragma once
-#include "operators/unary.h"
-#include "utils/small_array.h"
-
-namespace infini {
-
-void whereKernel(const float *inputX, const float *inputY,
-                 const uint8_t *condition, float *output, int nDims,
-                 int outputsize, SmallArray inputXShape, SmallArray inputYShape,
-                 SmallArray conditionShape, SmallArray outputShape, int xSize,
-                 int ySize, int cSize);
-void whereKernel(const half *inputX, const half *inputY,
-                 const uint8_t *condition, half *output, int nDims,
-                 int outputsize, SmallArray inputXShape, SmallArray inputYShape,
-                 SmallArray conditionShape, SmallArray outputShape, int xSize,
-                 int ySize, int cSize);
-}; // namespace infini
--- a/include/cuda/gather.h
+++ b/include/cuda/gather.h
@ -1,61 +1,17 @@
 #pragma once
-#include "core/data_type.h"
-#include "core/operator.h"
-#include "operators/gather.h"
+
+typedef struct {
+    int *indexValue;
+    int axis;
+    int inNDim;
+    int outNDim;
+    int idxNDim;
+    int outDim[4];
+    int idxDim[4];
+    int idxStride[4];
+    int inStride[4];
+} GatherMetaData;

 namespace infini {
-struct GatherMetaData {
-    // Pointer to indices
-    void *indexValue;
-    // Type of index values
-    DataType indexType;
-    // Type of input and output data
-    DataType dataType;
-    // Axis of the gather operation
-    int axis;
-    // Rank of input
-    int inNDim;
-    // Rank of output
-    int outNDim;
-    // Rank of indices
-    int idxNDim;
-    // Shape of output
-    int outDim[4];
-    // Shape of indices
-    int idxDim[4];
-    // Strides of indices
-    int idxStride[4];
-    // Strides of input
-    int inStride[4];
-};
-
-inline void initGatherMetaData(GatherMetaData &metaData,
-                               const Ref<OperatorObj> &_op) {
-    memset(&metaData, 0, sizeof(metaData));
-    auto op = as<GatherBaseObj>(_op);
-    Ref<TensorObj> in = op->getInputs(0);
-    Ref<TensorObj> index = op->getInputs(1);
-    Ref<TensorObj> out = op->getOutput();
-    metaData.indexValue = index->getRawDataPtr<void *>();
-    metaData.indexType = index->getDType();
-    metaData.dataType = in->getDType();
-    metaData.axis = op->getAxis();
-    metaData.inNDim = in->getRank();
-    metaData.outNDim = out->getRank();
-    metaData.idxNDim = index->getRank();
-    for (int i = 0; i < metaData.outNDim; ++i)
-        metaData.outDim[i] = out->getDims()[i];
-    for (int i = 0; i < metaData.idxNDim; ++i) {
-        metaData.idxDim[i] = index->getDims()[i];
-        metaData.idxStride[i] = index->getStride()[i];
-    }
-    for (int i = 0; i < metaData.inNDim; ++i) {
-        metaData.inStride[i] = in->getStride()[i];
-    }
+void gather_kernel(float *in, float *out, GatherMetaData metaData, int num);
 }
-template <typename T>
-void gather_kernel(T *in, T *out, GatherMetaData metaData, size_t num);
-
-void gather_elements_kernel(void *in, void *out, GatherMetaData metaData,
-                            size_t num);
-} // namespace infini
--- a/include/cuda/nccl_communicator.h
+++ b/include/cuda/nccl_communicator.h
@ -1,70 +0,0 @@
-#pragma once
-#include "core/communicator.h"
-#include <chrono>
-#include <cstdlib>
-#include <filesystem>
-#include <fstream>
-#include <nccl.h>
-#include <thread>
-
-#define checkNcclError(call)                                                   \
-    {                                                                          \
-        auto err = call;                                                       \
-        if (ncclSuccess != err) {                                              \
-            fprintf(stderr, "NCCL error in %s:%i : %s.\n", __FILE__, __LINE__, \
-                    ncclGetErrorString(err));                                  \
-            exit(EXIT_FAILURE);                                                \
-        }                                                                      \
-    }
-
-namespace infini {
-
-class NcclCommunicatorObj final : public CommunicatorObj {
-  private:
-    ncclComm_t comm;
-
-  public:
-    NcclCommunicatorObj(const string &name, int worldSize, int rank)
-        : CommunicatorObj(worldSize, rank) {
-        const std::string filePath("./" + name + "_nccl_id.bin");
-        ncclUniqueId commId;
-        if (rank == 0) {
-            checkNcclError(ncclGetUniqueId(&commId));
-            std::ofstream ofs(filePath, std::ios::binary);
-            ofs.write((char *)&commId, sizeof(ncclUniqueId));
-
-        } else {
-            auto begin = std::chrono::steady_clock::now();
-            while (!std::filesystem::exists(filePath)) {
-                auto now = std::chrono::steady_clock::now();
-                _IT_ASSERT_2(now < begin + std::chrono::seconds(10),
-                             "time limit (10s) exceeded.");
-                std::this_thread::sleep_for(std::chrono::milliseconds(100));
-            }
-            std::ifstream ifs(filePath, std::ios::binary);
-            ifs.read((char *)&commId, sizeof(ncclUniqueId));
-        }
-        checkNcclError(ncclCommInitRank(&comm, worldSize, commId, rank));
-        if (rank == 0) {
-            std::filesystem::remove(filePath);
-        }
-    }
-
-    // Get the actual ncclComm_t
-    ncclComm_t getNcclComm() { return comm; }
-
-    void finalize() { checkNcclError(ncclCommFinalize(comm)); }
-
-    ~NcclCommunicatorObj() final {
-        finalize();
-        checkNcclError(ncclCommDestroy(comm));
-    }
-
-    virtual string toString() const final {
-        std::ostringstream oss;
-        oss << "NCCL communicator";
-        return oss.str();
-    }
-};
-
-} // namespace infini
--- a/include/cuda/resize.cuh
+++ b/include/cuda/resize.cuh
@ -1,21 +0,0 @@
-#pragma once
-#include "cuda/cuda_common.h"
-
-typedef struct {
-    int nDims;
-    int oDims[4];
-    int inDims[4];
-    int inStride[4];
-    float scale[4];
-    float roiS[4];
-    float roiE[4];
-} MetaData;
-
-namespace infini {
-void resize_kernel_nearest(float *in, float *out, const MetaData &metaData,
-                           size_t num, int coordinateMode, int nearestMode);
-void resize_kernel_linear(float *in, float *out, const MetaData &metaData,
-                          size_t num, int coordinateMode);
-void resize_kernel_cubic(float *in, float *out, const MetaData &metaData,
-                         size_t num, int coordinateMode);
-} // namespace infini
--- a/include/intelcpu/mkl_kernel_without_config.h
+++ b/include/intelcpu/mkl_kernel_without_config.h
@ -1,40 +0,0 @@
-#pragma once
-#include "core/kernel.h"
-#include "intelcpu/mkl_runtime.h"
-
-namespace infini {
-
-class MklKernelWithoutConfig : public Kernel {
-  public:
-    virtual void compute(const Operator &op, const PerfRecord &record,
-                         const RuntimeObj *_context) const override {
-        compute(op, _context);
-        auto context = dynamic_cast<const MklRuntimeObj *>(_context);
-        context->sync();
-    }
-    virtual void compute(const Operator &op,
-                         const RuntimeObj *context) const = 0;
-    // Premise: op is idempotent since it is called multiple times.
-    virtual PerfRecord tune(const Operator &op,
-                            const RuntimeObj *_context) const override {
-        auto context = dynamic_cast<const MklRuntimeObj *>(_context);
-        return make_ref<PerfRecordObj>(timeit([&]() { compute(op, _context); },
-                                              [&]() { context->sync(); }));
-    }
-
-  protected:
-    dnnl::memory::format_tag getUserFormatTag(int nDim) const {
-        if (nDim == 2)
-            return dnnl::memory::format_tag::nc;
-        else if (nDim == 3)
-            return dnnl::memory::format_tag::ncw;
-        else if (nDim == 4)
-            return dnnl::memory::format_tag::nchw;
-        else if (nDim == 5)
-            return dnnl::memory::format_tag::ncdhw;
-        else
-            IT_TODO_HALT();
-    }
-};
-
-} // namespace infini
--- a/include/intelcpu/mkl_runtime.h
+++ b/include/intelcpu/mkl_runtime.h
@ -1,35 +0,0 @@
-#pragma once
-#include "core/runtime.h"
-#include "dnnl.h"
-#include "oneapi/dnnl/dnnl.h"
-#include "oneapi/dnnl/dnnl.hpp"
-#include "oneapi/dnnl/dnnl_types.h"
-#include <dnnl_debug.h>
-#include <mkl.h>
-namespace infini {
-class MklRuntimeObj : public CpuRuntimeObj {
-    dnnl_engine_t engine;
-    dnnl_stream_t stream;
-
-  public:
-    MklRuntimeObj();
-    static Ref<MklRuntimeObj> &getInstance() {
-        static Ref<MklRuntimeObj> instance = make_ref<MklRuntimeObj>();
-        return instance;
-    }
-
-    virtual ~MklRuntimeObj();
-    void dealloc(void *ptr) override { return mkl_free(ptr); };
-
-    void *alloc(size_t size) override {
-        return mkl_calloc((size + sizeof(uint64_t) - 1) / sizeof(uint64_t),
-                          sizeof(uint64_t), 64);
-    };
-
-    string toString() const override { return "INTELCPU Runtime"; };
-    dnnl::engine getEngine() const { return dnnl::engine(engine, true); }
-    dnnl::stream getStream() const { return dnnl::stream(stream, true); }
-    void sync() const;
-};
-
-} // namespace infini
--- a/include/intelcpu/operator_timer.h
+++ b/include/intelcpu/operator_timer.h
@ -1,15 +0,0 @@
-#pragma once
-namespace infini {
-namespace opTimer {
-double getPerfConvMkl(int n, int c, int h, int w, int f, int r, int s, int padh,
-                      int padw, int strideh, int stridew, int dilationh,
-                      int dilationw, int group);
-
-double getPerfConvTransposed2dMkl(int n, int c, int h, int w, int f, int r,
-                                  int s, int padh, int padw, int strideh,
-                                  int stridew, int dilationh, int dilationw,
-                                  int oph, int opw, int group);
-
-double getPerfMatmulMkl(int b, int m, int n, int k);
-} // namespace opTimer
-} // namespace infini
--- a/include/kunlun/kunlun_act_type.h
+++ b/include/kunlun/kunlun_act_type.h
@ -1,23 +0,0 @@
-#include "core/op_type.h"
-#include "kunlun/kunlun_common.h"
-
-namespace infini {
-using KunlunActType = xdnn::Activation_t;
-KunlunActType parseActType(ActType act) {
-    switch (act) {
-    case ActType::None:
-        return KunlunActType::LINEAR;
-    case ActType::Tanh:
-        return KunlunActType::TANH;
-    case ActType::Sigmoid:
-        return KunlunActType::SIGMOID;
-    case ActType::Relu:
-        return KunlunActType::RELU6;
-    default:
-        fprintf(stderr, "Activation Type not support yet!\n");
-        break;
-    }
-    return KunlunActType::LINEAR;
-}
-
-}; // namespace infini
--- a/include/kunlun/kunlun_common.h
+++ b/include/kunlun/kunlun_common.h
@ -1,22 +0,0 @@
-#pragma once
-#include "core/common.h"
-#include "xpu/runtime_ex.h"
-#include "xpu/xdnn.h"
-
-namespace xdnn = baidu::xpu::api;
-
-#define checkKUNLUNError(call)                                                 \
-    {                                                                          \
-        auto err = call;                                                       \
-        if (XPU_SUCCESS != err) {                                              \
-            fprintf(stderr, "KUNLUN error in %s:%i : %s.\n", __FILE__,         \
-                    __LINE__, xpu_strerror(err));                              \
-            exit(EXIT_FAILURE);                                                \
-        }                                                                      \
-    }
-
-namespace infini {
-
-using KUNLUNPtr = void *;
-
-} // namespace infini
--- a/include/kunlun/kunlun_kernel_without_config.h
+++ b/include/kunlun/kunlun_kernel_without_config.h
@ -1,24 +0,0 @@
-#pragma once
-#include "core/kernel.h"
-#include "kunlun/kunlun_runtime.h"
-
-namespace infini {
-
-class KUNLUNKernelWithoutConfig : public Kernel {
-  public:
-    virtual void compute(const Operator &op, const PerfRecord &record,
-                         const RuntimeObj *context) const {
-        compute(op, context);
-    }
-    virtual void compute(const Operator &op,
-                         const RuntimeObj *context) const = 0;
-    // Premise: op is idempotent since it is called multiple times.
-    virtual PerfRecord tune(const Operator &op,
-                            const RuntimeObj *_context) const {
-        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
-        return make_ref<PerfRecordObj>(timeit([&]() { compute(op, _context); },
-                                              [&]() { context->sync(); }));
-    }
-};
-
-} // namespace infini
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
wanghailu	beba9c16c4	add code for test resnet	2022-10-21 14:46:42 +08:00
wanghailu	53594b2ebc	test resnet	2022-10-20 16:36:55 +08:00
				`@ -1 +0,0 @@`
				`Subproject commit 51d3105277f3774ed31c02ed4cd11fa92925af77`
				`@ -1 +0,0 @@`
				`Subproject commit cbcf3fbf985a00494b0f136c92eaccd42031bf65`