diff --git a/paddle/infrt/dialect/infrt/infrt_ops.td b/paddle/infrt/dialect/infrt/infrt_ops.td
index ecd7093e72b8ad41a6385acb81dacfb6a0c4197b..e07a598d9bc631f416f6bcc0362b577267c5616a 100644
--- a/paddle/infrt/dialect/infrt/infrt_ops.td
+++ b/paddle/infrt/dialect/infrt/infrt_ops.td
@@ -18,6 +18,22 @@ def Infrt_KernelOp : Infrt_Op<"kernel", [NoSideEffect]> {
   let results = (outs Variadic<AnyType>);
 }
 
+def Infrt_ReturnOp : Infrt_Op<"return", [Terminator]> {
+  let summary = "host executor return operation";
+  let description = [{
+      The "infrt.return" operation represents a return operation within a function.
+
+        func @foo() : (i32, f8) {
+        infrt.return %0, %1 : i32, f8
+        }
+    }];
+
+  let arguments = (ins Variadic<AnyType>:$operands);
+
+  let builders = [OpBuilder<(ins),
+                  [{ build($_builder, $_state, llvm::None); }]>];
+}
+
 def Infrt_CvtTensorOp : Infrt_Op<"cvt_tensor", [NoSideEffect]> {
   let summary = "convert tensor type op";
   let description = [{convert tensor type op!}];
diff --git a/paddle/infrt/dialect/pd_ops.cc b/paddle/infrt/dialect/pd_ops.cc
index 338b04e001320289b71f6127318e7a073cefcacf..55ab174fcaf059d81f83e54e8f1e5864ef25b7e3 100644
--- a/paddle/infrt/dialect/pd_ops.cc
+++ b/paddle/infrt/dialect/pd_ops.cc
@@ -16,7 +16,6 @@
 
 #include <mlir/IR/Matchers.h>
 #include <mlir/IR/PatternMatch.h>
-#include "paddle/infrt/dialect/infrt/infrt_dialect.h"
 #include "paddle/infrt/dialect/infrt_base.h"
 
 #define GET_OP_CLASSES
diff --git a/paddle/infrt/dialect/pd_ops.h b/paddle/infrt/dialect/pd_ops.h
index b48c68060d42efdeecccd68abc5efb3045643370..41dd2ddd94eb161735568170a9a8bdc2ec259cdf 100644
--- a/paddle/infrt/dialect/pd_ops.h
+++ b/paddle/infrt/dialect/pd_ops.h
@@ -28,6 +28,7 @@
 #include <mlir/Interfaces/InferTypeOpInterface.h>
 #include <mlir/Interfaces/LoopLikeInterface.h>
 #include <mlir/Interfaces/SideEffectInterfaces.h>
+#include "paddle/infrt/dialect/infrt/infrt_dialect.h"
 
 namespace mlir {
 namespace pd {
diff --git a/paddle/infrt/dialect/tensorrt/trt_dilaect_types.h b/paddle/infrt/dialect/tensorrt/trt_dialect_types.h
similarity index 91%
rename from paddle/infrt/dialect/tensorrt/trt_dilaect_types.h
rename to paddle/infrt/dialect/tensorrt/trt_dialect_types.h
index efcf7dd5be19524293c8b58db26f35b9c11afb3d..0c3edcec1edb2d7774038e638f8c20f94a3e7166 100644
--- a/paddle/infrt/dialect/tensorrt/trt_dilaect_types.h
+++ b/paddle/infrt/dialect/tensorrt/trt_dialect_types.h
@@ -23,6 +23,8 @@ class EngineType
     : public mlir::Type::TypeBase<EngineType, mlir::Type, mlir::TypeStorage> {
  public:
   using Base::Base;
+  static EngineType get();
+  static EngineType get(mlir::MLIRContext *context);
 };
 
 }  // namespace trt
diff --git a/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.cc b/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.cc
index fa0095363c5fd34778162fb4e3204450ef1e7815..ad6b136463a71dcc2fcd9ce2b4e2da6f68e88dd2 100644
--- a/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.cc
+++ b/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.cc
@@ -53,9 +53,9 @@ bool reverseDfs(std::vector<mlir::Operation *> source,
 }
 
 // merge the first&second graph op to a new graph op.
-void mergeTwoAdjacentCreateEngineOp(mlir::OpBuilder &builder,  // NOLINT
-                                    CreateEngineOp first,
-                                    CreateEngineOp second) {
+void mergeTwoAdjacentGraphOp(mlir::OpBuilder &builder,  // NOLINT
+                             mlir::pd::GraphOp first,
+                             mlir::pd::GraphOp second) {
   // comput inputs and outputs
   ::llvm::SmallVector<mlir::Value, 4> inputs(first.getOperands()), outputs;
   for (mlir::Value input : second.getOperands()) {
@@ -84,8 +84,7 @@ void mergeTwoAdjacentCreateEngineOp(mlir::OpBuilder &builder,  // NOLINT
   // create the new graph op
   builder.setInsertionPoint(first);
   auto loc = first.getLoc();
-  auto graph_op =
-      builder.create<CreateEngineOp>(loc, return_types, inputs, true);
+  auto graph_op = builder.create<mlir::pd::GraphOp>(loc, return_types, inputs);
   mlir::Block *block = new mlir::Block;
   auto copy_range = second.getBody()->without_terminator();
   block->getOperations().splice(block->begin(),
@@ -98,7 +97,7 @@ void mergeTwoAdjacentCreateEngineOp(mlir::OpBuilder &builder,  // NOLINT
                                 copy_range.begin(),
                                 copy_range.end());
   builder.setInsertionPointToEnd(block);
-  builder.create<::infrt::dialect::ReturnOp>(loc, outputs);
+  builder.create<::infrt::ReturnOp>(loc, outputs);
   graph_op.body().push_back(block);
 
   // mapping the output
@@ -150,12 +149,13 @@ void TRTGraphFusePass::runOnFunction() {
   do {
     changed = false;
     for (auto &op : body) {
-      CreateEngineOp graph_op = ::llvm::dyn_cast_or_null<CreateEngineOp>(&op);
+      mlir::pd::GraphOp graph_op =
+          ::llvm::dyn_cast_or_null<mlir::pd::GraphOp>(&op);
       if (nullptr == graph_op) continue;
 
       for (auto user_op : op.getUsers()) {
-        CreateEngineOp user_graph_op =
-            ::llvm::dyn_cast_or_null<CreateEngineOp>(user_op);
+        mlir::pd::GraphOp user_graph_op =
+            ::llvm::dyn_cast_or_null<mlir::pd::GraphOp>(user_op);
         if (nullptr == user_graph_op) continue;
         // get all dst input nodes except src.
         std::vector<mlir::Operation *> source_nodes;
@@ -168,7 +168,7 @@ void TRTGraphFusePass::runOnFunction() {
         // Reverse DFS from the source_nodes.
         if (!reverseDfs(source_nodes,
                         [&op](const mlir::Operation *n) { return n == &op; })) {
-          mergeTwoAdjacentCreateEngineOp(builder, graph_op, user_graph_op);
+          mergeTwoAdjacentGraphOp(builder, graph_op, user_graph_op);
           changed = true;
           break;
         }
diff --git a/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.h b/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.h
index 350add905aac75c0ba8527aa6e9bc1510fab876c..803e53e3244f92134928e1105a8248e9f49e5432 100644
--- a/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.h
+++ b/paddle/infrt/dialect/tensorrt/trt_graph_fuse_pass.h
@@ -15,7 +15,6 @@
 #pragma once
 #include <mlir/Pass/Pass.h>
 #include "paddle/infrt/dialect/infrt_base.h"
-#include "paddle/infrt/dialect/tensorrt/trt_ops.h"
 
 namespace infrt {
 namespace trt {
@@ -26,40 +25,37 @@ namespace trt {
  *
  * source func:
  *
- * func @main() -> tensor<?xf32> {
- *  %a = "pd.feed"()...
- *  %c = "trt.create_engine"(%a) {
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
+ *  %c = "pd.graph"(%a) {
  *     %m = "pd.conv2d"(%a)...
- *     "Infrt.return" %m
+ *     "infrt.return" (%m)
  *  } ...
- *  %d = "trt.create_engine"(%c) {
+ *  %d = "pd.graph"(%c) {
  *      %m = "pd.conv3d"(%c)...
- *      "Infrt.return" %m
+ *      "infrt.return" (%m)
  *  } ...
- *  %f = "trt.create_engine"(%a) {
+ *  %f = "pd.graph"(%a) {
  *      %m = "pd.conv2d"(%a)...
- *      "Infrt.return" %m
+ *      "infrt.return" (%m)
  *  } ...
- *  "pd.fetch" %d, %f
+ *  "infrt.return" (%d, %f)..
+ * }
  *
  * destination func:
- * func @main() -> tensor<?xf32> {
- *  %a = "pd.feed"()...
- *  %d, %f = "trt.create_engine"(%a) {
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
+ *  %d, %f = "pd.graph"(%a) {
  *     %m = "pd.conv2d"(%a)...
  *     %n = "pd.conv3d"(%m)...
  *     %s = "pd.conv2d"(%a)...
- *     "Infrt.return" %n, %s
+ *     "infrt.return" (%n, %s)
  *  } ...
- *  "pd.fetch" %d, %f
+ *  "infrt.return" (%d, %f)
  * }
  */
 class TRTGraphFusePass
     : public mlir::PassWrapper<TRTGraphFusePass, mlir::FunctionPass> {
  public:
-  void getDependentDialects(mlir::DialectRegistry &registry) const override {
-    registry.insert<TensorRTDialect, ::infrt::dialect::INFRTDialect>();
-  }
+  void getDependentDialects(mlir::DialectRegistry &registry) const override {}
   ::llvm::StringRef getName() const override { return "trtGraphFusePass"; }
   void runOnFunction() override;
 };
diff --git a/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.cc b/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.cc
index 5ee7b23213a0106d3491712c37d34940f7c15c58..e3a7b455024c65d40ccbafb28fba9e9b0ead0369 100644
--- a/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.cc
+++ b/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.cc
@@ -16,23 +16,23 @@
 
 #include <mlir/IR/Builders.h>
 #include "paddle/infrt/dialect/pd_ops.h"
-#include "paddle/infrt/dialect/tensorrt/trt_ops.h"
 
 namespace infrt {
 namespace trt {
 // Implementation of the trtGraphSplitPass。
 void TRTGraphSplitPass::runOnFunction() {
-  std::vector<CreateEngineOp> worklist;
+  std::vector<mlir::pd::GraphOp> worklist;
   mlir::Block& block = getFunction().front();
   for (auto& op : block) {
-    CreateEngineOp graph_op = ::llvm::dyn_cast_or_null<CreateEngineOp>(&op);
+    mlir::pd::GraphOp graph_op =
+        ::llvm::dyn_cast_or_null<mlir::pd::GraphOp>(&op);
     if (nullptr != graph_op &&
         graph_op.getBody()->getOperations().size() <= min_subgraph_size_) {
       worklist.push_back(graph_op);
     }
   }
   while (!worklist.empty()) {
-    CreateEngineOp graph_op = worklist.back();
+    mlir::pd::GraphOp graph_op = worklist.back();
     worklist.pop_back();
     mlir::Block* body = graph_op.getBody();
     auto return_op = body->getTerminator();
diff --git a/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.h b/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.h
index 28078e2bc2dbff46d6a9eaf5522b949f68785898..1c44a13cf9dfb65a1747a596dc1012e7f54d792e 100644
--- a/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.h
+++ b/paddle/infrt/dialect/tensorrt/trt_graph_split_pass.h
@@ -15,7 +15,6 @@
 #pragma once
 #include <mlir/Pass/Pass.h>
 #include "paddle/infrt/dialect/infrt_base.h"
-#include "paddle/infrt/dialect/tensorrt/trt_ops.h"
 
 namespace infrt {
 namespace trt {
@@ -27,33 +26,29 @@ namespace trt {
  *
  * source func:
  *
- * func @main() -> tensor<?xf32> {
- *  %a = "pd.feed"()...
- *  %d, %f = "trt.create_engine"(%a) {
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
+ *  %d, %f = "pd.graph"(%a) {
  *     %m = "pd.conv2d"(%a)...
  *     %n = "pd.conv3d"(%m)...
  *     %s = "pd.conv2d"(%a)...
- *     "Infrt.return" (%n, %s)
+ *     "infrt.return" (%n, %s)...
  *  } ...
- *  "pd.fetch" (%d, %f)
+ *  "infrt.return" (%d, %f)...
  * }
  *
  * destination func:
- * func @main() -> tensor<?xf32> {
- *  %a = "pd.feed"()...
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
  *  %c = "pd.conv2d"(%a) ...
  *  %d = "pd.conv3d"(%c) ...
  *  %f = "pd.conv2d"(%a) ...
- *  "pd.fetch" (%d, %f)
+ *  "infrt.return" (%d, %f)...
  * }
  */
 class TRTGraphSplitPass
     : public mlir::PassWrapper<TRTGraphSplitPass, mlir::FunctionPass> {
  public:
   ::llvm::StringRef getName() const override { return "trtGraphSplitPass"; }
-  void getDependentDialects(mlir::DialectRegistry &registry) const override {
-    registry.insert<TensorRTDialect, ::infrt::dialect::INFRTDialect>();
-  }
+  void getDependentDialects(mlir::DialectRegistry &registry) const override {}
   void runOnFunction() override;
   explicit TRTGraphSplitPass(size_t min_subgraph_size = 3)
       : min_subgraph_size_(min_subgraph_size) {}
diff --git a/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc b/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc
index 8d81e739d9c72ebcaa57b927a360864db59d7e97..1be5f4dbc39d7699b6d8a36cfb3e164694e908c1 100644
--- a/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc
+++ b/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc
@@ -16,12 +16,64 @@
 #include <mlir/Transforms/DialectConversion.h>
 #include "paddle/infrt/dialect/infrt_base.h"
 #include "paddle/infrt/dialect/pd_ops.h"
+#include "paddle/infrt/dialect/tensorrt/trt_dialect_types.h"
 
 namespace infrt {
 namespace trt {
 
 #include "paddle/infrt/dialect/tensorrt/pd_lower_to_trt.cpp.inc"  // NOLINT
 
+struct PD2TRT_GraphLower : public ::mlir::RewritePattern {
+  PD2TRT_GraphLower(::mlir::MLIRContext *context)
+      : ::mlir::RewritePattern("pd.graph", 1, context, {"trt.create_engine"}) {}
+  ::mlir::LogicalResult matchAndRewrite(
+      ::mlir::Operation *op, ::mlir::PatternRewriter &rewriter) const override {
+    auto casted_op = ::llvm::dyn_cast<mlir::pd::GraphOp>(op);
+    ::mlir::Operation::operand_range inputs = casted_op.inputs();
+    auto ods_loc = rewriter.getFusedLoc(op->getLoc());
+    CreateEngineOp create_engine_op;
+    // inputs
+    ::mlir::SmallVector<::mlir::Value, 4> trt_inputs;
+    for (auto v : inputs) {
+      trt_inputs.push_back(v);
+    }
+    create_engine_op = rewriter.create<CreateEngineOp>(
+        ods_loc,
+        ::llvm::SmallVector<mlir::Type, 4>(1, EngineType::get()),
+        trt_inputs,
+        true /*run_once*/);
+    ::mlir::Block *block = new ::mlir::Block;
+    block->getOperations().splice(block->begin(),
+                                  casted_op.getBody()->getOperations(),
+                                  casted_op.getBody()->begin(),
+                                  casted_op.getBody()->end());
+    create_engine_op.body().push_back(block);
+
+    // trt.execute
+    // outputs
+    ::llvm::SmallVector<::mlir::Type, 4> execute_outputs_types;
+    for (auto v : casted_op.getODSResults(0)) {
+      execute_outputs_types.push_back(v.getType());
+    }
+    // inputs
+    ::mlir::SmallVector<::mlir::Value, 4> execute_inputs(
+        create_engine_op.getODSResults(0));
+    for (auto v : inputs) {
+      execute_inputs.push_back(v);
+    }
+    auto execute_op = rewriter.create<ExecuteOp>(
+        ods_loc, execute_outputs_types, execute_inputs);
+
+    ::llvm::SmallVector<::mlir::Value, 4> replace_values;
+    for (auto v :
+         ::llvm::SmallVector<::mlir::Value, 4>{execute_op.getODSResults(0)}) {
+      replace_values.push_back(v);
+    }
+    rewriter.replaceOp(op, replace_values);
+    return ::mlir::success();
+  }
+};
+
 void TRTOpConverterPass::runOnOperation() {
   // The first thing to define is the conversion target. This will define the
   // final target for this lowering.
@@ -36,6 +88,7 @@ void TRTOpConverterPass::runOnOperation() {
   // the set of patterns that will lower the TensorRT operations.
   ::mlir::RewritePatternSet patterns(&getContext());
   populateWithGenerated(patterns);
+  patterns.add<PD2TRT_GraphLower>(&getContext());
 
   // With the target and rewrite patterns defined, we can now attempt the
   // conversion. The conversion will signal failure if any of our `illegal`
diff --git a/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.h b/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.h
index a8128a585ee82dc60811c65b1105beb33e8c3b18..7550d8c84e19504fc0f41067c1194703a55410ba 100644
--- a/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.h
+++ b/paddle/infrt/dialect/tensorrt/trt_op_converter_pass.h
@@ -15,6 +15,7 @@
 #pragma once
 #include "mlir/IR/Dialect.h"
 #include "mlir/Pass/Pass.h"
+#include "paddle/infrt/dialect/infrt/infrt_dialect.h"
 #include "paddle/infrt/dialect/tensorrt/trt_ops.h"
 
 namespace infrt {
@@ -23,27 +24,26 @@ namespace trt {
  * trtOpConverterPass.
  *
  * source ir:
- * func @main() -> tensor<?xf32> {
- *   %a = "pd.feed"()...
- *   %d, %f = "trt.create_engine"(%a) {
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
+ *   %d, %f = "pd.graph"(%a) {
  *     %m = "pd.conv2d"(%a)...
  *     %n = "pd.conv3d"(%m)...
  *     %s = "pd.conv2d"(%a)...
- *     "Infrt.return" %n, %s
+ *     "infrt.return" (%n, %s)...
  *   } ...
- *   "pd.fetch" %d, %f
+ *   "infrt.return" (%d, %f)...
  * }
  *
  * destination ir:
- * func @main() -> tensor<?xf32> {
- *   %a = "pd.feed"()...
- *   %d, %f = "trt.create_engine"(%a) {
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
+ *   %engine = "trt.create_engine"(%a) ({
  *     %m = "trt.Convolution"(%a)...
  *     %n = "trt.Convolution"(%m)...
  *     %s = "trt.Convolution"(%a)...
- *     "Infrt.return" %n, %s
- *   } ...
- *   "pd.fetch" %d, %f
+ *     "infrt.return" (%n, %s)...
+ *   }){run_once = true} ...
+ *   %d, %f = "trt.execute"(%engine, %a)...
+ *   "infrt.return" (%d, %f)...
  * }
  */
 struct TRTOpConverterPass
diff --git a/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.cc b/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.cc
index 17e893a383a9cd3f893e80181858dc3cc2b0552b..13b7f1aee55d2a2d30822a878bbd50d385411f43 100644
--- a/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.cc
+++ b/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.cc
@@ -16,6 +16,7 @@
 
 #include <mlir/IR/Builders.h>
 #include "paddle/infrt/dialect/basic_kernels.h"
+#include "paddle/infrt/dialect/infrt/infrt_dialect.h"
 #include "paddle/infrt/dialect/pd_ops.h"
 
 namespace infrt {
@@ -37,11 +38,11 @@ void TRTOpTellerPass::runOnFunction() {
     if (::llvm::dyn_cast_or_null<mlir::pd::FeedOp>(op)) continue;
     if (::llvm::dyn_cast_or_null<mlir::pd::FetchOp>(op)) continue;
     if (::llvm::dyn_cast_or_null<mlir::pd::GraphOp>(op)) continue;
-    if (::llvm::dyn_cast_or_null<CreateEngineOp>(op)) continue;
+    if (::llvm::dyn_cast_or_null<::infrt::ReturnOp>(op)) continue;
     builder.setInsertionPoint(op);
     auto loc = getFunction().getLoc();
-    auto graph_op = builder.create<CreateEngineOp>(
-        loc, op->getResultTypes(), op->getOperands(), true);
+    auto graph_op = builder.create<mlir::pd::GraphOp>(
+        loc, op->getResultTypes(), op->getOperands());
 
     ::llvm::SmallVector<mlir::Value, 4> tblgen_repl_values;
     for (auto v :
@@ -54,7 +55,7 @@ void TRTOpTellerPass::runOnFunction() {
     graph_op.body().push_back(block);
     op->moveBefore(block, block->begin());
     builder.setInsertionPointToEnd(block);
-    builder.create<::infrt::dialect::ReturnOp>(loc, op->getResults());
+    builder.create<::infrt::ReturnOp>(loc, op->getResults());
   }
 }
 }  // namespace trt
diff --git a/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.h b/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.h
index 471eafa9f9ba33dad4182ba7da55a607c2bf8f0d..b9e461c8633d906fd46e9f7d6799e8a157915048 100644
--- a/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.h
+++ b/paddle/infrt/dialect/tensorrt/trt_op_teller_pass.h
@@ -15,7 +15,6 @@
 #pragma once
 #include <mlir/Pass/Pass.h>
 #include "paddle/infrt/dialect/infrt_base.h"
-#include "paddle/infrt/dialect/tensorrt/trt_ops.h"
 
 namespace infrt {
 namespace trt {
@@ -26,30 +25,28 @@ namespace trt {
  *
  * source func:
  *
- * func @main() -> tensor<?xf32> {
- *  %a = "pd.feed"()...
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
  *  %c = "pd.conv2d"(%a) ...
  *  %d = "pd.conv3d"(%c) ...
  *  %f = "pd.conv2d"(%a) ...
- *  "pd.fetch" (%d, %f)
+ *  "infrt.return"(%d, %f) ...
  * }
  *
  * destination func:
- * func @main() -> tensor<?xf32> {
- *  %a = "pd.feed"()...
- *  %c = "trt.create_engine"(%a) {
+ * func @main(%a : tensor<?xf32>) -> tensor<?xf32> {
+ *  %c = "pd.graph"(%a) {
  *     %m = "pd.conv2d"(%a)...
- *     "Infrt.return" (%m)
+ *     "infrt.return" (%m)
  *  } ...
- *  %d = "trt.create_engine"(%c) {
+ *  %d = "pd.graph"(%c) {
  *      %m = "pd.conv3d"(%c)...
- *      "Infrt.return" (%m)
+ *      "infrt.return" (%m)
  *  } ...
- *  %f = "trt.create_engine"(%a) {
+ *  %f = "pd.graph"(%a) {
  *      %m = "pd.conv2d"(%a)...
- *      "Infrt.return" (%m)
+ *      "infrt.return" (%m)
  *  } ...
- *  "pd.fetch" (%d, %f)
+ *  "infrt.return" (%d, %f)
  * }
  * TODO(winter-wang): Supplementary how to judge the operators can be supported
  * by tensorrt.
@@ -57,9 +54,7 @@ namespace trt {
 class TRTOpTellerPass
     : public mlir::PassWrapper<TRTOpTellerPass, mlir::FunctionPass> {
  public:
-  void getDependentDialects(mlir::DialectRegistry &registry) const override {
-    registry.insert<TensorRTDialect, ::infrt::dialect::INFRTDialect>();
-  }
+  void getDependentDialects(mlir::DialectRegistry &registry) const override {}
   ::llvm::StringRef getName() const override { return "trtOpTellerPass"; }
   void runOnFunction() override;
 };
diff --git a/paddle/infrt/dialect/tensorrt/trt_ops.cc b/paddle/infrt/dialect/tensorrt/trt_ops.cc
index f179939e2320613bb5ce3ebd0f21e01f7abff2f0..d5222976625a2adece9a87c8952dba10137ae9ba 100644
--- a/paddle/infrt/dialect/tensorrt/trt_ops.cc
+++ b/paddle/infrt/dialect/tensorrt/trt_ops.cc
@@ -11,7 +11,6 @@
 // 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/infrt/dialect/tensorrt/trt_ops.h"
 #include <mlir/IR/DialectImplementation.h>
 #include <mlir/IR/Matchers.h>
@@ -19,11 +18,20 @@
 #include <mlir/IR/PatternMatch.h>
 #include <mlir/Interfaces/CallInterfaces.h>
 #include <mlir/Interfaces/SideEffectInterfaces.h>
-#include "paddle/infrt/dialect/tensorrt/trt_dilaect_types.h"
+#include "paddle/infrt/common/global.h"
+#include "paddle/infrt/dialect/tensorrt/trt_dialect_types.h"
 
 namespace infrt {
 namespace trt {
 
+EngineType EngineType::get() {
+  return Base::get(::infrt::Global::getMLIRContext());
+}
+
+EngineType EngineType::get(mlir::MLIRContext *context) {
+  return Base::get(context);
+}
+
 TensorRTDialect::TensorRTDialect(mlir::MLIRContext *context)
     : mlir::Dialect("trt", context, mlir::TypeID::get<TensorRTDialect>()) {
   addTypes<EngineType>();
diff --git a/paddle/infrt/dialect/tensorrt/trt_ops.h b/paddle/infrt/dialect/tensorrt/trt_ops.h
index 978b9906e5f52b69565673eae270c658b5232bd8..44444232915bad7d25b0ecedfa8e8427f4567e49 100644
--- a/paddle/infrt/dialect/tensorrt/trt_ops.h
+++ b/paddle/infrt/dialect/tensorrt/trt_ops.h
@@ -29,6 +29,8 @@
 #include <mlir/Interfaces/LoopLikeInterface.h>
 #include <mlir/Interfaces/SideEffectInterfaces.h>
 #include "paddle/infrt/dialect/basic_kernels.h"
+#include "paddle/infrt/dialect/infrt/infrt_dialect.h"
+#include "paddle/infrt/dialect/pd_ops.h"
 
 namespace infrt {
 namespace trt {
diff --git a/paddle/infrt/dialect/tensorrt/trt_ops.td b/paddle/infrt/dialect/tensorrt/trt_ops.td
index 31142a5157bfcd544128671fbdf22a993f1cc646..132a1d7805bdb85af8716e384ec29357a6ff68ad 100755
--- a/paddle/infrt/dialect/tensorrt/trt_ops.td
+++ b/paddle/infrt/dialect/tensorrt/trt_ops.td
@@ -7,14 +7,24 @@ include "mlir/Interfaces/CallInterfaces.td"
 include "mlir/IR/OpBase.td"
 include "paddle/infrt/dialect/tensorrt/trt_op_base.td"
 
-def TRT_CreateEngineOp : TRT_Op<"create_engine", [SingleBlockImplicitTerminator<"::infrt::dialect::ReturnOp">]> {
-  let summary = "trt Graph Op";
+
+def TRT_CreateEngineOp : TRT_Op<"create_engine", [SingleBlockImplicitTerminator<"::infrt::ReturnOp">]> {
+  let summary = "trt CreateEngine Op";
   let description = [{
     Describe a tensorrt subgraph.
   }];
   let regions = (region SizedRegion<1>:$body);
   let arguments = (ins Variadic<TRT_Tensor>:$inputs, DefaultValuedAttr<BoolAttr, "true">:$run_once);
-  let results = (outs Variadic<TRT_Tensor>:$outputs);
+  let results = (outs TRT_EngineType:$output);
+}
+
+def TRT_ExecuteOp : TRT_Op<"execute", [NoSideEffect]> {
+  let summary = "trt execute Op";
+  let description = [{
+    Describe a tensorrt runtime.
+  }];
+  let arguments = (ins TRT_EngineType:$engine, Variadic<TRT_Tensor>:$inputs);
+  let results = (outs Variadic<TRT_Tensor>:$output);
 }
 
 def TRT_ActivationOp : TRT_Op<"Activation", [NoSideEffect]> {
diff --git a/paddle/infrt/tests/dialect/trt_ops.mlir b/paddle/infrt/tests/dialect/trt_ops.mlir
index 49510bc542dc0409067b5d61cb189dfab8b6601f..6d25044d139f32c0a29adefb44c8fd2640cadd82 100644
--- a/paddle/infrt/tests/dialect/trt_ops.mlir
+++ b/paddle/infrt/tests/dialect/trt_ops.mlir
@@ -1,13 +1,6 @@
 // RUN: trt-exec %s
 // CHECK-LABEL: @main
-func @main() -> tensor<?xf32> {
-  %bias = "pd.feed"() {name="input0"} : () -> tensor<?xf32>
-  %c = "pd.feed"() {name="input1"} : () -> tensor<?xf32>
-  %b1 = "pd.feed"() {name="input2"} : () -> tensor<?xf32>
-  %b2 = "pd.feed"() {name="input3"} : () -> tensor<?xf32>
-  %bias1 = "pd.feed"() {name="input4"} : () -> tensor<?xf32>
-  %bias2 = "pd.feed"() {name="input5"} : () -> tensor<?xf32>
-
+func @main(%bias:tensor<?xf32>, %c:tensor<?xf32>, %b1:tensor<?xf32>, %b2:tensor<?xf32>, %bias1:tensor<?xf32>, %bias2:tensor<?xf32>) -> tensor<?xf32> {
   %d = "pd.elementwise_add"(%c, %bias) {axis=-1:si32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
   %e = "pd.relu6"(%d) {} : (tensor<?xf32>) -> tensor<?xf32>
 
@@ -19,5 +12,5 @@ func @main() -> tensor<?xf32> {
   %d2 = "pd.elementwise_add"(%c2, %bias2) {axis=-1:si32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
   %e2 = "pd.relu"(%d2) {} : (tensor<?xf32>) -> tensor<?xf32>
   
-  "pd.fetch"(%e2) {name="output"} :(tensor<?xf32>)->()
+  "infrt.return"(%e2) : (tensor<?xf32>)->()
 }
diff --git a/tools/infrt/custom_pdop.td b/tools/infrt/custom_pdop.td
index 2139fbd8155bb067c542f1e54f135c828e06cc58..f754767259563f2cd64bac92adf76249b18af11f 100644
--- a/tools/infrt/custom_pdop.td
+++ b/tools/infrt/custom_pdop.td
@@ -33,7 +33,7 @@ def PD_ReturnOp : PD_Op<"return", [Terminator]> {
   let arguments = (ins Variadic<PD_Tensor>:$inputs);
 }
 
-def PD_GraphOp : PD_Op<"graph", [SingleBlockImplicitTerminator<"ReturnOp">]> {
+def PD_GraphOp : PD_Op<"graph", [SingleBlockImplicitTerminator<"::infrt::ReturnOp">]> {
   let summary = "paddle graph Op";
   let description = [{
     Describe a paddle graph or subgraph.