diff --git a/paddle/cinn/hlir/dialect/CMakeLists.txt b/paddle/cinn/hlir/dialect/CMakeLists.txt
index 897e9ab8cfb445ec86939fd694c62b862f0988b1..d7c6d787a7fb0437c5727757c2f1a150851d7ea3 100755
--- a/paddle/cinn/hlir/dialect/CMakeLists.txt
+++ b/paddle/cinn/hlir/dialect/CMakeLists.txt
@@ -1,6 +1,53 @@
 # TODO(Aurelius84): new_ir_compiler depends on pd_dialect and could
 # not found under CINN_ONLY mode
 if(NOT CINN_ONLY)
-  cinn_cc_library(cinn_dialect SRCS runtime_dialect.cc jit_kernel_op.cc DEPS
-                  pd_dialect)
+  set(CINN_DIALECT_SOURCE_DIR "${PADDLE_SOURCE_DIR}/paddle/cinn/hlir/dialect")
+  set(CINN_DIALECT_BINARY_DIR "${PADDLE_BINARY_DIR}/paddle/cinn/hlir/dialect")
+
+  # Generate cinn_dialect files defining op using op_gen_file
+  set(cinn_op_gen_file
+      ${PADDLE_SOURCE_DIR}/paddle/fluid/ir/dialect/op_generator/op_gen.py)
+
+  set(cinn_op_compat_yaml_file
+      ${PADDLE_SOURCE_DIR}/paddle/phi/api/yaml/op_compat.yaml)
+
+  set(cinn_op_forward_yaml_file1
+      ${PADDLE_SOURCE_DIR}/paddle/cinn/hlir/dialect/cinn_ops.parsed.yaml)
+
+  set(cinn_op_yaml_files ${cinn_op_forward_yaml_file1})
+
+  set(cinn_op_namespace cinn,dialect)
+  set(cinn_dialect_name cinn)
+  set(cinn_op_header_file ${CINN_DIALECT_BINARY_DIR}/cinn_op.h)
+  set(cinn_op_source_file ${CINN_DIALECT_BINARY_DIR}/cinn_op.cc)
+  set(cinn_op_header_file_tmp ${cinn_op_header_file}.tmp)
+  set(cinn_op_source_file_tmp ${cinn_op_source_file}.tmp)
+
+  add_custom_command(
+    OUTPUT ${cinn_op_header_file} ${cinn_op_source_file}
+    COMMAND
+      ${PYTHON_EXECUTABLE} ${cinn_op_gen_file} --op_yaml_files
+      ${cinn_op_yaml_files} --op_compat_yaml_file ${cinn_op_compat_yaml_file}
+      --namespaces ${cinn_op_namespace} --dialect_name ${cinn_dialect_name}
+      --op_def_h_file ${cinn_op_header_file_tmp} --op_def_cc_file
+      ${cinn_op_source_file_tmp}
+    COMMAND ${CMAKE_COMMAND} -E copy_if_different ${cinn_op_header_file_tmp}
+            ${cinn_op_header_file}
+    COMMAND ${CMAKE_COMMAND} -E copy_if_different ${cinn_op_source_file_tmp}
+            ${cinn_op_source_file}
+    DEPENDS ${cinn_op_gen_file} ${cinn_op_forward_yaml_file1}
+            ${cinn_op_compat_yaml_file}
+    VERBATIM)
+
+  cinn_cc_library(
+    cinn_dialect
+    SRCS
+    runtime_dialect.cc
+    jit_kernel_op.cc
+    cinn_dialect.cc
+    cinn_op.cc
+    DEPS
+    pd_dialect)
+
+  target_include_directories(cinn_dialect PRIVATE ${CINN_DIALECT_BINARY_DIR})
 endif()
diff --git a/paddle/cinn/hlir/dialect/cinn_dialect.cc b/paddle/cinn/hlir/dialect/cinn_dialect.cc
new file mode 100644
index 0000000000000000000000000000000000000000..9bc6c6ff8c9c65b485ccdefb8e9928ab6480f1df
--- /dev/null
+++ b/paddle/cinn/hlir/dialect/cinn_dialect.cc
@@ -0,0 +1,42 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/cinn/hlir/dialect/cinn_dialect.h"
+// NOTE(chenxi67): File cinn_op.h is generated by op_gen.py, see details in
+// paddle/cinn/hlir/dialect/CMakeLists.txt.
+#include "paddle/cinn/hlir/dialect/cinn_op.h"
+
+namespace cinn {
+namespace dialect {
+
+CinnDialect::CinnDialect(::ir::IrContext* context)
+    : ::ir::Dialect(
+          name(), context, ::ir::TypeId::get<cinn::dialect::CinnDialect>()) {
+  this->initialize();
+}
+
+void CinnDialect::initialize() {
+  // NOTE(chenxi67): GET_OP_LIST is defined in cinn_op.h which is
+  // generated by op_gen.py, see details in
+  // paddle/cinn/hlir/dialect/CMakeLists.txt.
+  RegisterOps<
+#define GET_OP_LIST
+#include "paddle/cinn/hlir/dialect/cinn_op.h"  // NOLINT
+      >();
+}
+
+}  // namespace dialect
+}  // namespace cinn
+
+IR_DEFINE_EXPLICIT_TYPE_ID(cinn::dialect::CinnDialect)
diff --git a/paddle/cinn/hlir/dialect/cinn_dialect.h b/paddle/cinn/hlir/dialect/cinn_dialect.h
new file mode 100644
index 0000000000000000000000000000000000000000..77fb96863ad3714d43b22cebb8068f1df75bf3cb
--- /dev/null
+++ b/paddle/cinn/hlir/dialect/cinn_dialect.h
@@ -0,0 +1,35 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/ir/core/dialect.h"
+
+namespace cinn {
+namespace dialect {
+
+class CinnDialect : public ::ir::Dialect {
+ public:
+  explicit CinnDialect(::ir::IrContext* context);
+
+  static const char* name() { return "cinn"; }
+
+ private:
+  void initialize();
+};
+
+}  // namespace dialect
+}  // namespace cinn
+
+IR_DECLARE_EXPLICIT_TYPE_ID(cinn::dialect::CinnDialect)
diff --git a/paddle/cinn/hlir/dialect/cinn_ops.parsed.yaml b/paddle/cinn/hlir/dialect/cinn_ops.parsed.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6297be83e23986b871b29217738b90c258a7ac50
--- /dev/null
+++ b/paddle/cinn/hlir/dialect/cinn_ops.parsed.yaml
@@ -0,0 +1,78 @@
+- name: add
+  inputs:
+  - typename: Tensor
+    name: x
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  - typename: Tensor
+    name: y
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs: []
+  outputs:
+  - {typename: Tensor, name: out, optional: false, intermediate: false}
+  no_need_buffer: null
+  data_transform: null
+  infer_meta:
+    func: ElementwiseInferMeta
+    param: [x, y]
+  kernel:
+    func: [add]
+    param: [x, y]
+    backend: null
+    layout: null
+    data_type: null
+    dispatch: {add: null}
+    force_backend: null
+  inplace: {out: x}
+  view: null
+  backward: add_grad
+- name: add_grad
+  inputs:
+  - typename: Tensor
+    name: x
+    optional: false
+    no_need_buffer: true
+    data_transform: {}
+  - typename: Tensor
+    name: y
+    optional: false
+    no_need_buffer: true
+    data_transform: {}
+  - typename: Tensor
+    name: out_grad
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs:
+  - {typename: int, name: axis, default_value: '-1'}
+  outputs:
+  - {typename: Tensor, name: x_grad, optional: false, intermediate: false}
+  - {typename: Tensor, name: y_grad, optional: false, intermediate: false}
+  no_need_buffer: [x, y]
+  data_transform: null
+  infer_meta:
+    func: GeneralBinaryGradInferMeta
+    param: [x, y]
+  kernel:
+    func: [add_grad]
+    param: [x, y, out_grad, axis]
+    backend: null
+    layout: null
+    data_type: null
+    dispatch: {add_grad: null}
+    force_backend: null
+  inplace: {x_grad: out_grad}
+  view: null
+  composite: {func_name: add_grad, func_args: 'x, y, out_grad, axis, x_grad, y_grad'}
+  backward: add_double_grad
+  forward:
+    name: add
+    inputs:
+    - {name: x, typename: Tensor}
+    - {name: y, typename: Tensor}
+    attrs: []
+    outputs:
+    - {name: out, typename: Tensor}
diff --git a/paddle/fluid/ir/dialect/op_generator/op_build_gen.py b/paddle/fluid/ir/dialect/op_generator/op_build_gen.py
index f4d91d5c06821f29cd0788fce1e6751337748815..a59948bcb3c6d3677a72a24eadcf3b6b75bc7eda 100644
--- a/paddle/fluid/ir/dialect/op_generator/op_build_gen.py
+++ b/paddle/fluid/ir/dialect/op_generator/op_build_gen.py
@@ -191,7 +191,7 @@ def GenBuildAttributes(
 ):
     INTARRAY_STR_TEMPLATE = """  ir::Attribute attr_{attr_name} = {op_attribute_type}::get(ir::IrContext::Instance(), phi::IntArray({attr}));
 """
-    SCALAR_STR_TEMPLATE = """  ir::Attribute attr_{attr_name} = TransToIrAttribute({attr}, ir::IrContext::Instance());
+    SCALAR_STR_TEMPLATE = """  ir::Attribute attr_{attr_name} = paddle::dialect::TransToIrAttribute({attr}, ir::IrContext::Instance());
 """
     STR_TEMPLATE = """  ir::Attribute attr_{attr_name} = {op_attribute_type}::get(ir::IrContext::Instance(), {attr});
 """
@@ -286,7 +286,7 @@ def GenBuildOutputs(
     CREATE_INPUT_METATENSOR_TEMPLATE = """
   VLOG(4) << "Builder construction  dense_{name}";
   phi::DenseTensor dense_{name}(std::make_unique<paddle::experimental::DefaultAllocator>(paddle::platform::CPUPlace()).get(),
-                                phi::DenseTensorMeta(TransToPhiDataType({name}.dtype()),
+                                phi::DenseTensorMeta(paddle::dialect::TransToPhiDataType({name}.dtype()),
                                                      {name}.dims(),
                                                      {name}.data_layout(),
                                                      {name}.lod(),
@@ -297,7 +297,7 @@ def GenBuildOutputs(
     CREATE_INPUT_VEC_METATENSOR_TEMPLATE = """  std::vector<phi::DenseTensor> vec_dense_{name};
   for (size_t i=0; i < static_cast<size_t>({name}.size()); i++) {{
     vec_dense_{name}.push_back(phi::DenseTensor(std::make_unique<paddle::experimental::DefaultAllocator>(paddle::platform::CPUPlace()).get(),
-                                                phi::DenseTensorMeta(TransToPhiDataType({name}[i].dyn_cast<paddle::dialect::DenseTensorType>().dtype()),
+                                                phi::DenseTensorMeta(paddle::dialect::TransToPhiDataType({name}[i].dyn_cast<paddle::dialect::DenseTensorType>().dtype()),
                                                                      {name}[i].dyn_cast<paddle::dialect::DenseTensorType>().dims(),
                                                                      {name}[i].dyn_cast<paddle::dialect::DenseTensorType>().data_layout(),
                                                                      {name}[i].dyn_cast<paddle::dialect::DenseTensorType>().lod(),
@@ -430,13 +430,13 @@ def GenBuildOutputs(
     build_output_str += "\n  std::vector<ir::Type> argument_outputs;"
 
     CREATE_OUTPUT_DENSE_TENSOR_TEMPLATE = """
-  ir::Type {name}_dense_tensor_type = paddle::dialect::DenseTensorType::get(ir::IrContext::Instance(), TransToIrDataType(dense_{name}.dtype()), dense_{name}.dims(), dense_{name}.layout(), dense_{name}.lod(), dense_{name}.offset());
+  ir::Type {name}_dense_tensor_type = paddle::dialect::DenseTensorType::get(ir::IrContext::Instance(), paddle::dialect::TransToIrDataType(dense_{name}.dtype()), dense_{name}.dims(), dense_{name}.layout(), dense_{name}.lod(), dense_{name}.offset());
   argument_outputs.push_back({name}_dense_tensor_type);
 """
     CREATE_OUTPUT_VEC_DENSE_TENSOR_TEMPLATE = """
   std::vector<ir::Type> {name}_types;
   for (size_t i=0; i < static_cast<size_t>({output_size}); i++) {{
-    {name}_types.push_back(paddle::dialect::DenseTensorType::get(ir::IrContext::Instance(), TransToIrDataType(vec_dense_{name}[i].dtype()), vec_dense_{name}[i].dims(), vec_dense_{name}[i].layout(), vec_dense_{name}[i].lod(), vec_dense_{name}[i].offset()));
+    {name}_types.push_back(paddle::dialect::DenseTensorType::get(ir::IrContext::Instance(), paddle::dialect::TransToIrDataType(vec_dense_{name}[i].dtype()), vec_dense_{name}[i].dims(), vec_dense_{name}[i].layout(), vec_dense_{name}[i].lod(), vec_dense_{name}[i].offset()));
   }}
   ir::Type {name}_vector_type = ir::VectorType::get(ir::IrContext::Instance(), {name}_types);
   argument_outputs.push_back({name}_vector_type);
diff --git a/paddle/fluid/ir/dialect/op_generator/op_gen.py b/paddle/fluid/ir/dialect/op_generator/op_gen.py
index da1d7cbdde0900f4dd36225bf8acee6366408c86..423b47ae44ed0c81413c540b36b79bc6ac54ac71 100644
--- a/paddle/fluid/ir/dialect/op_generator/op_gen.py
+++ b/paddle/fluid/ir/dialect/op_generator/op_gen.py
@@ -97,8 +97,7 @@ op_n_attribute_declare_str = (
 # String Template for cc file code gen
 # =====================================
 CC_FILE_TEMPLATE = """// This file is generated by "paddle/fluid/ir/dialect/op_generator/op_gen.py"
-
-#include "paddle/fluid/ir/dialect/pd_op.h"
+#include "{h_file}"
 #include "paddle/fluid/ir/dialect/pd_type.h"
 #include "paddle/fluid/ir/dialect/pd_attribute.h"
 #include "paddle/ir/core/builtin_attribute.h"
@@ -115,7 +114,7 @@ CC_FILE_TEMPLATE = """// This file is generated by "paddle/fluid/ir/dialect/op_g
 #include "paddle/phi/infermeta/backward.h"
 #include "paddle/phi/api/lib/utils/allocator.h"
 #include "paddle/fluid/primitive/rule/vjp/vjp.h"
-#include "paddle/fluid/primitive/type/lazy_tensor.h"
+{def_primitive}
 #include "paddle/ir/core/op_base.h"
 
 {input}
@@ -133,18 +132,14 @@ OpInfoTuple {op_name}::GetOpInfo() {{
   std::vector<paddle::dialect::OpInputInfo> inputs = {{ {inputs} }};
   std::vector<paddle::dialect::OpAttributeInfo> attributes = {{ {attributes} }};
   std::vector<paddle::dialect::OpOutputInfo> outputs = {{ {outputs} }};
-  paddle::dialect::OpRunTimeInfo run_time_info = OpRunTimeInfo("{infer_meta_func}", {{"{infer_meta_param}"}}, {{"{kernel_func}"}}, {{"{kernel_param}"}}, {{"{kernel_key_dtype}"}}, {{{inplace}}}, {{{view}}});
+  paddle::dialect::OpRunTimeInfo run_time_info = paddle::dialect::OpRunTimeInfo("{infer_meta_func}", {{"{infer_meta_param}"}}, {{"{kernel_func}"}}, {{"{kernel_param}"}}, {{"{kernel_key_dtype}"}}, {{{inplace}}}, {{{view}}});
 
   return std::make_tuple(inputs, attributes, outputs, run_time_info);
 }}
 """
-CONSTRUCT_INPUT_INFO_TEMPLATE = """OpInputInfo("{name}", "{typename}", {optional}, {no_need_buffer}, {is_mutable_attribute})"""
-CONSTRUCT_OUTPUT_INFO_TEMPLATE = (
-    """OpOutputInfo("{name}", "{typename}", {optional}, {intermediate})"""
-)
-CONSTRUCT_ATTRIBUTE_INFO_TEMPLATE = (
-    """OpAttributeInfo("{name}", "{typename}", "{data_type}")"""
-)
+CONSTRUCT_INPUT_INFO_TEMPLATE = """paddle::dialect::OpInputInfo("{name}", "{typename}", {optional}, {no_need_buffer}, {is_mutable_attribute})"""
+CONSTRUCT_OUTPUT_INFO_TEMPLATE = """paddle::dialect::OpOutputInfo("{name}", "{typename}", {optional}, {intermediate})"""
+CONSTRUCT_ATTRIBUTE_INFO_TEMPLATE = """paddle::dialect::OpAttributeInfo("{name}", "{typename}", "{data_type}")"""
 
 
 DEFINE_OP_TYPE_ID = """
@@ -751,17 +746,17 @@ def OpGenerator(
         op_kernel_map = op_info.kernel_map
         op_inplace_map = op_info.inplace_map
         op_view_map = op_info.view_map
-        op_interfaces = ["OpYamlInfoInterface"]
+        op_interfaces = ["paddle::dialect::OpYamlInfoInterface"]
         op_traits = []
 
         if op_info.infer_meta_func:
-            op_interfaces += ["InferMetaInterface"]
+            op_interfaces += ["paddle::dialect::InferMetaInterface"]
 
         if (
             op_info.backward_name
             and op_info.op_phi_name[0] in vjp_interface_declare_gen_op_list
         ):
-            op_interfaces += ["VjpInterface"]
+            op_interfaces += ["paddle::dialect::VjpInterface"]
         exclusive_interface_str = gen_exclusive_interface_str(op_info)
 
         # If op has inplace info, we will generate inplace op and non-inplace op.
@@ -779,7 +774,7 @@ def OpGenerator(
                 op_interfaces_str = "," + ",".join(op_interfaces)
 
             if op_name[-1] == "_":
-                op_traits += ["InplaceTrait"]
+                op_traits += ["paddle::dialect::InplaceTrait"]
 
             op_traits_str = ""
             if len(op_traits) > 0:
@@ -1079,7 +1074,13 @@ def OpGenerator(
                 ops_defined_list.append(build_func_with_muta_attr_is_input)
             ops_defined_list.append(op_verify_str)
             ops_defined_list.append(op_infer_meta_str)
-            ops_defined_list.append(op_vjp_str)
+            # NOTE(chenxi67)skip if dialect_name==cinn
+            if dialect_name == "cinn":
+                import logging
+
+                logging.warning("cinn is currently not support Vjp function")
+            else:
+                ops_defined_list.append(op_vjp_str)
 
     # (4) Generate head file str
     op_namespaces_prev = ""
@@ -1119,8 +1120,15 @@ def OpGenerator(
     for op in ops_name_with_namespace_list:
         define_type_id_str += DEFINE_OP_TYPE_ID.format(op_name=op)
 
+    # NOTE(chenxi67) Skip include this header file if dialect_name == cinn
+    # otherwise we may get compile error when compile with "ncclDataType_t"
+    def_primitive_str = "#include \"paddle/fluid/primitive/type/lazy_tensor.h\""
+    if dialect_name == "cinn":
+        def_primitive_str = ""
+
     source_file_str = CC_FILE_TEMPLATE.format(
         h_file=op_def_h_file[:-4],
+        def_primitive=def_primitive_str,
         input=source_file_str,
         define_type_id=define_type_id_str,
     )  # Add head
diff --git a/paddle/fluid/ir/dialect/utils.h b/paddle/fluid/ir/dialect/utils.h
index 2cc1c653fc0effc0e1a40b2d43997a40cbbdcef6..9b5a79c7a4433f4d24c5a19798a8fef7395439be 100644
--- a/paddle/fluid/ir/dialect/utils.h
+++ b/paddle/fluid/ir/dialect/utils.h
@@ -14,7 +14,7 @@
 
 #pragma once
 
-#include "paddle/fluid/framework/convert_utils.h"
+// #include "paddle/fluid/framework/convert_utils.h"
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/ir/dialect/pd_attribute.h"
 #include "paddle/fluid/ir/dialect/pd_type_storage.h"