diff --git a/.bashrc b/.bashrc
new file mode 100644
index 0000000000..aba1296dac
--- /dev/null
+++ b/.bashrc
@@ -0,0 +1,52 @@
+# .bashrc
+
+# User specific aliases and functions
+
+alias rm='rm -i'
+alias cp='cp -i'
+alias mv='mv -i'
+
+# Source global definitions
+if [ -f /etc/bashrc ]; then
+	. /etc/bashrc
+fi
+export http_proxy=http://child-prc.intel.com:913
+export https_proxy=http://child-prc.intel.com:913
+
+export PADDLEPADDLE_TP_CACHE="/home/guest/tp_cache"
+#export LD_LIBRARY_PATH=/usr/local/lib/python3.8/dist-packages/deepmd_kit-1.2.3.dev530+g8d8f289.d20220512-py3.8-linux-x86_64.egg/deepmd/op:$LD_LIBRARY_PATH
+#export LIBRARY_PATH=/usr/local/lib/python3.8/dist-packages/deepmd_kit-1.2.3.dev530+g8d8f289.d20220512-py3.8-linux-x86_64.egg/deepmd/op:$LIBRARY_PATH
+#export DEEP_MD_PATH=/usr/local/lib/python3.8/dist-packages/deepmd_kit-1.2.3.dev530+g8d8f289.d20220512-py3.8-linux-x86_64.egg/deepmd/op
+#export LD_LIBRARY_PATH=/usr/local/lib/python3.8/dist-packages/deepmd_kit-1.2.3.dev491+g1a06aa4-py3.8-linux-x86_64.egg/deepmd/op:$LD_LIBRARY_PATH
+#export LIBRARY_PATH=/usr/local/lib/python3.8/dist-packages/deepmd_kit-1.2.3.dev491+g1a06aa4-py3.8-linux-x86_64.egg/deepmd/op:$LIBRARY_PATH
+#export DEEP_MD_PATH=/usr/local/lib/python3.8/dist-packages/deepmd_kit-1.2.3.dev491+g1a06aa4-py3.8-linux-x86_64.egg/deepmd/op
+#
+#1.2.3.dev530+g8d8f289.d20220512
+export tensorflow_root=/home/tensorflowroot
+export deepmd_root=/home/deepmdroot
+# >>> conda initialize >>>
+# !! Contents within this block are managed by 'conda init' !!
+__conda_setup="$('/root/anaconda3/bin/conda' 'shell.bash' 'hook' 2> /dev/null)"
+if [ $? -eq 0 ]; then
+    eval "$__conda_setup"
+else
+    if [ -f "/root/anaconda3/etc/profile.d/conda.sh" ]; then
+        . "/root/anaconda3/etc/profile.d/conda.sh"
+    else
+        export PATH="/root/anaconda3/bin:$PATH"
+    fi
+fi
+unset __conda_setup
+# <<< conda initialize <<<
+export PATH=/home/jessie/cmake-3.21.0-linux-x86_64/bin:$PATH
+
+export PATH=/home/lammps-stable_29Oct2020/src:$PATH
+#export LD_LIBRARY_PATH=/home/paddle-deepmd/source/build/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/home/deepmd-kit/source/build/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/home/Paddle/build/paddle_inference_install_dir/paddle/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/home/Paddle/build/paddle_inference_install_dir/third_party/install/mkldnn/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/home/Paddle/build/paddle_inference_install_dir/third_party/install/mklml/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/home/Paddle/build/paddle/fluid/pybind/:$LD_LIBRARY_PATH
+#export LD_LIBRARY_PATH=/home/paddle-deepmd/source/build:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/home/deepmd-kit/source/build:$LD_LIBRARY_PATH
+
diff --git a/compile_deepmd.sh b/compile_deepmd.sh
index ec0480b743..2a8a2021df 100644
--- a/compile_deepmd.sh
+++ b/compile_deepmd.sh
@@ -1,6 +1,5 @@
 rm -rf /home/deepmdroot/ && mkdir /home/deepmdroot && deepmd_root=/home/deepmdroot
-cd /home/paddle-deepmd/source && rm -rf build && mkdir build && cd build
-#cmake -DTENSORFLOW_ROOT=$tensorflow_root -DCMAKE_INSTALL_PREFIX=$deepmd_root -DPADDLE_ROOT=/home/Paddle/build/paddle_inference_install_dir -DUSE_CUDA_TOOLKIT=FALSE ..
+cd /home/deepmd-kit/source && rm -rf build && mkdir build && cd build
 cmake -DPADDLE_ROOT=/home/Paddle/build/paddle_inference_install_dir -DUSE_CUDA_TOOLKIT=FALSE ..
 make -j 4 && make install
 make lammps
diff --git a/compile_lammps.sh b/compile_lammps.sh
index 7951f4a843..9d0f9e8fc7 100644
--- a/compile_lammps.sh
+++ b/compile_lammps.sh
@@ -3,6 +3,8 @@ cd /home
 rm -rf lammps-stable_29Oct2020/
 tar -xzvf stable_29Oct2020.tar.gz
 cd lammps-stable_29Oct2020/src/
-cp -r /home/paddle-deepmd/source/build/USER-DEEPMD .
+cp -r /home/deepmd-kit/source/build/USER-DEEPMD .
+#cp -r /home/paddle-deepmd/source/build/USER-DEEPMD .
 make yes-kspace yes-user-deepmd
-make serial -j 20
+#make serial -j 20
+make mpi -j 20
diff --git a/examples/water/lmp/in.lammps b/examples/water/lmp/in.lammps
index a501a5ab27..4dcce1466f 100644
--- a/examples/water/lmp/in.lammps
+++ b/examples/water/lmp/in.lammps
@@ -1,5 +1,5 @@
 # bulk water
-
+echo            screen
 units           metal
 boundary        p p p
 atom_style      atomic
@@ -11,7 +11,7 @@ read_data	water.lmp
 mass 		1 16
 mass		2 2
 
-pair_style	deepmd frozen_model.pb
+pair_style	deepmd /home/deepmd-kit/examples/water/lmp/model.pdmodel /home/deepmd-kit/examples/water/lmp/model.pdiparams
 pair_coeff	
 
 velocity        all create 330.0 23456789
diff --git a/examples/water/lmp/model.pdiparams b/examples/water/lmp/model.pdiparams
new file mode 100644
index 0000000000..033e12d0da
Binary files /dev/null and b/examples/water/lmp/model.pdiparams differ
diff --git a/examples/water/lmp/model.pdiparams.info b/examples/water/lmp/model.pdiparams.info
new file mode 100644
index 0000000000..de3a13a349
Binary files /dev/null and b/examples/water/lmp/model.pdiparams.info differ
diff --git a/examples/water/lmp/model.pdmodel b/examples/water/lmp/model.pdmodel
new file mode 100644
index 0000000000..922dcb7e0e
Binary files /dev/null and b/examples/water/lmp/model.pdmodel differ
diff --git a/source/CMakeLists.txt b/source/CMakeLists.txt
index e1d9fa749e..7bdcf35b07 100755
--- a/source/CMakeLists.txt
+++ b/source/CMakeLists.txt
@@ -5,7 +5,10 @@ set(CMAKE_LINK_WHAT_YOU_USE TRUE)
 # build cpp or python interfaces
 if (NOT DEFINED BUILD_CPP_IF) 
   set(BUILD_CPP_IF TRUE)
-endif (NOT DEFINED BUILD_CPP_IF)
+  message(WARNING "WARNING! -DON_INFER is set")
+  add_definitions("-DON_INFER")
+endif()
+
 if (NOT DEFINED BUILD_PY_IF) 
   set(BUILD_PY_IF FALSE)
 endif (NOT DEFINED BUILD_PY_IF)
@@ -72,72 +75,72 @@ if (USE_CUDA_TOOLKIT)
   add_definitions("-D GOOGLE_CUDA")
 endif()
 
-# find paddle
+if(BUILD_PY_IF)
+find_package(tensorflow REQUIRED)
+else()
 find_package(Fluid REQUIRED)
-
-# find tensorflow, I need tf abi info
-if (BUILD_PY_IF)
-  find_package(tensorflow REQUIRED)
-endif (BUILD_PY_IF)
+endif()
 
 # find threads
 find_package(Threads)
 
 # auto op_cxx_abi
-#if (NOT DEFINED OP_CXX_ABI)
-#  if (BUILD_PY_IF) 
-#    if (DEFINED TENSORFLOW_ROOT)
-#      set(FIND_ABI_CMD "import sys,os; sys.path.insert(0, os.path.join('${TENSORFLOW_ROOT}', '..')); import tensorflow; print(tensorflow.CXX11_ABI_FLAG if 'CXX11_ABI_FLAG' in tensorflow.__dict__ else tensorflow.sysconfig.CXX11_ABI_FLAG, end = '')" )
-#    else()
-#      set(FIND_ABI_CMD "import tensorflow; print(tensorflow.CXX11_ABI_FLAG if 'CXX11_ABI_FLAG' in tensorflow.__dict__ else tensorflow.sysconfig.CXX11_ABI_FLAG, end = '')")
-#    endif()
-#    execute_process(
-#      COMMAND ${PYTHON_EXECUTABLE} "-c" "${FIND_ABI_CMD}"
-#      WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}
-#      OUTPUT_VARIABLE PY_CXX_ABI
-#      RESULT_VARIABLE PY_CXX_ABI_RESULT_VAR
-#      ERROR_VARIABLE PY_CXX_ABI_ERROR_VAR
-#      )
-#    if (NOT ${PY_CXX_ABI_RESULT_VAR} EQUAL 0)
-#      message(FATAL_ERROR "Cannot determine cxx abi, error message: ${PY_CXX_ABI_ERROR_VAR}")
-#    endif()
-#    set(OP_CXX_ABI ${PY_CXX_ABI})
-#  endif()
-#  if (BUILD_CPP_IF)
-#    try_run(
-#      CPP_CXX_ABI_RUN_RESULT_VAR CPP_CXX_ABI_COMPILE_RESULT_VAR
-#      ${CMAKE_CURRENT_BINARY_DIR}/tf_cxx_abi
-#      "${CMAKE_CURRENT_SOURCE_DIR}/cmake/tf_cxx_abi.cpp"
-#      LINK_LIBRARIES ${TensorFlowFramework_LIBRARY}
-#      CMAKE_FLAGS "-DINCLUDE_DIRECTORIES:STRING=${TensorFlow_INCLUDE_DIRS}"
-#      RUN_OUTPUT_VARIABLE CPP_CXX_ABI
-#      COMPILE_OUTPUT_VARIABLE CPP_CXX_ABI_COMPILE_OUTPUT_VAR
-#      )
-#    if (NOT ${CPP_CXX_ABI_COMPILE_RESULT_VAR})
-#      message(FATAL_ERROR "Failed to compile: \n ${CPP_CXX_ABI_COMPILE_OUTPUT_VAR}" )
-#    endif()
-#    if (NOT ${CPP_CXX_ABI_RUN_RESULT_VAR} EQUAL "0")
-#      message(FATAL_ERROR "Failed to run, return code: ${CPP_CXX_ABI}" )
-#    endif()
-#    if (DEFINED PY_CXX_ABI)
-#      if (NOT (${CPP_CXX_ABI} EQUAL ${PY_CXX_ABI}))
-#	message (WARNNING "NOT consistent CXX_ABIs: python interface of tf uses ${PY_CXX_ABI}, while c++ interface of tf uses ${CPP_CXX_ABI}, we follow c++ interface ")
-#      endif()
-#    endif()
-#    set(OP_CXX_ABI ${CPP_CXX_ABI})
-#  endif()
-#  message (STATUS "Automatically determined OP_CXX_ABI=${OP_CXX_ABI} ")
-#else()
-#  message (STATUS "User set OP_CXX_ABI=${OP_CXX_ABI} ")  
-#endif()    
-# message (STATUS "No set OP_CXX_ABI=${OP_CXX_ABI} ")  
-# # message the cxx_abi used during compiling
-# if (${OP_CXX_ABI} EQUAL 0) 
-#   message (STATUS "Set GLIBCXX_USE_CXX_ABI=0 when compiling ops")
-# else ()
-#   set (OP_CXX_ABI 1)
-#   message (STATUS "Set GLIBCXX_USE_CXX_ABI=1 when compiling ops")
-# endif ()
+if (BUILD_PY_IF)
+if (NOT DEFINED OP_CXX_ABI)
+ if (BUILD_PY_IF) 
+   if (DEFINED TENSORFLOW_ROOT)
+     set(FIND_ABI_CMD "import sys,os; sys.path.insert(0, os.path.join('${TENSORFLOW_ROOT}', '..')); import tensorflow; print(tensorflow.CXX11_ABI_FLAG if 'CXX11_ABI_FLAG' in tensorflow.__dict__ else tensorflow.sysconfig.CXX11_ABI_FLAG, end = '')" )
+   else()
+     set(FIND_ABI_CMD "import tensorflow; print(tensorflow.CXX11_ABI_FLAG if 'CXX11_ABI_FLAG' in tensorflow.__dict__ else tensorflow.sysconfig.CXX11_ABI_FLAG, end = '')")
+   endif()
+   execute_process(
+     COMMAND ${PYTHON_EXECUTABLE} "-c" "${FIND_ABI_CMD}"
+     WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}
+     OUTPUT_VARIABLE PY_CXX_ABI
+     RESULT_VARIABLE PY_CXX_ABI_RESULT_VAR
+     ERROR_VARIABLE PY_CXX_ABI_ERROR_VAR
+     )
+   if (NOT ${PY_CXX_ABI_RESULT_VAR} EQUAL 0)
+     message(FATAL_ERROR "Cannot determine cxx abi, error message: ${PY_CXX_ABI_ERROR_VAR}")
+   endif()
+   set(OP_CXX_ABI ${PY_CXX_ABI})
+ endif()
+ if (BUILD_CPP_IF)
+   try_run(
+     CPP_CXX_ABI_RUN_RESULT_VAR CPP_CXX_ABI_COMPILE_RESULT_VAR
+     ${CMAKE_CURRENT_BINARY_DIR}/tf_cxx_abi
+     "${CMAKE_CURRENT_SOURCE_DIR}/cmake/tf_cxx_abi.cpp"
+     LINK_LIBRARIES ${TensorFlowFramework_LIBRARY}
+     CMAKE_FLAGS "-DINCLUDE_DIRECTORIES:STRING=${TensorFlow_INCLUDE_DIRS}"
+     RUN_OUTPUT_VARIABLE CPP_CXX_ABI
+     COMPILE_OUTPUT_VARIABLE CPP_CXX_ABI_COMPILE_OUTPUT_VAR
+     )
+   if (NOT ${CPP_CXX_ABI_COMPILE_RESULT_VAR})
+     message(FATAL_ERROR "Failed to compile: \n ${CPP_CXX_ABI_COMPILE_OUTPUT_VAR}" )
+   endif()
+   if (NOT ${CPP_CXX_ABI_RUN_RESULT_VAR} EQUAL "0")
+     message(FATAL_ERROR "Failed to run, return code: ${CPP_CXX_ABI}" )
+   endif()
+   if (DEFINED PY_CXX_ABI)
+     if (NOT (${CPP_CXX_ABI} EQUAL ${PY_CXX_ABI}))
+	message (WARNNING "NOT consistent CXX_ABIs: python interface of tf uses ${PY_CXX_ABI}, while c++ interface of tf uses ${CPP_CXX_ABI}, we follow c++ interface ")
+     endif()
+   endif()
+   set(OP_CXX_ABI ${CPP_CXX_ABI})
+ endif()
+ message (STATUS "Automatically determined OP_CXX_ABI=${OP_CXX_ABI} ")
+else()
+ message (STATUS "User set OP_CXX_ABI=${OP_CXX_ABI} ")  
+endif()    
+message (STATUS "No set OP_CXX_ABI=${OP_CXX_ABI} ")  
+# message the cxx_abi used during compiling
+if (${OP_CXX_ABI} EQUAL 0) 
+  message (STATUS "Set GLIBCXX_USE_CXX_ABI=0 when compiling ops")
+else ()
+  set (OP_CXX_ABI 1)
+  message (STATUS "Set GLIBCXX_USE_CXX_ABI=1 when compiling ops")
+endif ()
+endif()
 
 # define USE_TTM
 if (NOT DEFINED USE_TTM)
@@ -193,7 +196,7 @@ endif (BUILD_CPP_IF)
 
 # include
 include_directories(${DeePMD_INCLUDE_DIRS})
-#include_directories(${TensorFlow_INCLUDE_DIRS})
+include_directories(${TensorFlow_INCLUDE_DIRS})
 
 # define names of libs
 set (LIB_DEEPMD		"deepmd")
@@ -216,6 +219,7 @@ endif (BUILD_CPP_IF)
 # add_subdirectory (op/)
 add_subdirectory (lib/)
 if (BUILD_PY_IF)
+  add_subdirectory (op/)
   add_subdirectory (config/)
   # add_subdirectory (tests/)
 endif (BUILD_PY_IF)
diff --git a/source/cmake/FindFluid.cmake b/source/cmake/FindFluid.cmake
index 39e60c0d2e..fcafb90349 100755
--- a/source/cmake/FindFluid.cmake
+++ b/source/cmake/FindFluid.cmake
@@ -127,6 +127,7 @@ function(third_party_library TARGET_NAME TARGET_DIRNAME)
   set(PADDLE_THIRD_PARTY_LIBRARIES ${PADDLE_THIRD_PARTY_LIBRARIES} ${local_third_party_libraries} PARENT_SCOPE)
 endfunction()
 
+if(NOT BUILD_PY_IF)
 set(OP_DIR "${PROJECT_SOURCE_DIR}/op/paddle_ops/srcs")
 if(USE_CUDA_TOOLKIT)
   file(GLOB CUSTOM_OPERATOR_FILES ${OP_DIR}/*.cc)
@@ -135,6 +136,7 @@ else()
   # set(CUSTOM_OPERATOR_FILES "${OP_DIR}/pd_prod_env_mat_multi_devices_cpu.cc;${OP_DIR}/pd_prod_force_se_a_multi_devices_cpu.cc;${OP_DIR}/pd_prod_virial_se_a_multi_devices_cpu.cc;")
 endif()
 add_library(pd_infer_custom_op SHARED ${CUSTOM_OPERATOR_FILES})
+endif()
 
 third_party_library(mklml ${THIRD_PARTY_ROOT}/install/mklml/lib libiomp5.so libmklml_intel.so)
 third_party_library(mkldnn ${THIRD_PARTY_ROOT}/install/mkldnn/lib libmkldnn.so.0)
diff --git a/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cpu.cc b/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cpu.cc
index 7affd494e6..334aa0f523 100755
--- a/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cpu.cc
+++ b/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cpu.cc
@@ -1,5 +1,9 @@
-//#include "paddle/extension.h"
+#ifdef ON_INFER
 #include "paddle/include/experimental/ext_all.h"
+#else
+#include "paddle/extension.h"
+#endif
+// #include "paddle/include/experimental/ext_all.h"
 #include "utilities.h"
 #include "coord.h"
 #include "region.h"
diff --git a/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cuda.cc b/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cuda.cc
index ee714a2d1b..1f8f1e358a 100644
--- a/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cuda.cc
+++ b/source/op/paddle_ops/srcs/pd_prod_env_mat_multi_devices_cuda.cc
@@ -1,5 +1,9 @@
 #define GOOGLE_CUDA 1
+#ifdef ON_INFER
+#include "paddle/include/experimental/ext_all.h"
+#else
 #include "paddle/extension.h"
+#endif
 #include "utilities.h"
 #include "coord.h"
 #include "region.h"
diff --git a/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cpu.cc b/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cpu.cc
index f2a6482772..efccc010c2 100755
--- a/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cpu.cc
+++ b/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cpu.cc
@@ -1,8 +1,11 @@
 #include <assert.h>
 #include "prod_force.h"
 #include "prod_force_grad.h"
-//#include "paddle/extension.h"
+#ifdef ON_INFER
 #include "paddle/include/experimental/ext_all.h"
+#else
+#include "paddle/extension.h"
+#endif
 
 #define CHECK_INPUT(x) PD_CHECK(x.place() == paddle::PlaceType::kCPU, #x " must be a CPU Tensor.")
 #define CHECK_INPUT_READY(x) PD_CHECK(x.is_initialized(), #x " must be initialized before usage.")
diff --git a/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cuda.cc b/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cuda.cc
index 752646d5f2..3d36d8cc8e 100644
--- a/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cuda.cc
+++ b/source/op/paddle_ops/srcs/pd_prod_force_se_a_multi_devices_cuda.cc
@@ -1,6 +1,10 @@
 #define GOOGLE_CUDA 1
 #include "prod_force.h"
+#ifdef ON_INFER
+#include "paddle/include/experimental/ext_all.h"
+#else
 #include "paddle/extension.h"
+#endif
 #include <assert.h>
 
 #define CHECK_INPUT(x) PD_CHECK(x.place() == paddle::PlaceType::kGPU, #x " must be a GPU Tensor.")
diff --git a/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cpu.cc b/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cpu.cc
index c4095d1985..3f76426dfc 100755
--- a/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cpu.cc
+++ b/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cpu.cc
@@ -1,8 +1,11 @@
 #include <assert.h>
 #include "prod_virial.h"
 #include "prod_virial_grad.h"
-//#include "paddle/extension.h"
+#ifdef ON_INFER
 #include "paddle/include/experimental/ext_all.h"
+#else
+#include "paddle/extension.h"
+#endif
 
 
 #define CHECK_INPUT(x) PD_CHECK(x.place() == paddle::PlaceType::kCPU, #x " must be a CPU Tensor.")
diff --git a/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cuda.cc b/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cuda.cc
index 8163e636f3..da32032b30 100644
--- a/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cuda.cc
+++ b/source/op/paddle_ops/srcs/pd_prod_virial_se_a_multi_devices_cuda.cc
@@ -1,7 +1,11 @@
 #define GOOGLE_CUDA 1
 #include <assert.h>
 #include "prod_virial.h"
+#ifdef ON_INFER
+#include "paddle/include/experimental/ext_all.h"
+#else
 #include "paddle/extension.h"
+#endif
 
 #define CHECK_INPUT(x) PD_CHECK(x.place() == paddle::PlaceType::kGPU, #x " must be a GPU Tensor.")
 #define CHECK_INPUT_DIM(x, value) PD_CHECK(x.shape().size() == value, #x "'s dim should be " #value ".")