diff --git a/paddle/fluid/inference/analysis/ir_pass_manager.cc b/paddle/fluid/inference/analysis/ir_pass_manager.cc
index 403ebfe72a..99611ce84b 100644
--- a/paddle/fluid/inference/analysis/ir_pass_manager.cc
+++ b/paddle/fluid/inference/analysis/ir_pass_manager.cc
@@ -67,10 +67,8 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set("max_batch_size", new int(argument->tensorrt_max_batch_size()));
       pass->Set("min_subgraph_size",
                 new int(argument->tensorrt_min_subgraph_size()));
-      pass->Set(
-          "program",
-          new framework::ProgramDesc *(
-              const_cast<framework::ProgramDesc *>(&argument->main_program())));
+      pass->Set("program",
+                new framework::ProgramDesc *(&argument->main_program()));
 
       bool enable_int8 = argument->tensorrt_precision_mode() ==
                          contrib::AnalysisConfig::Precision::kInt8;
diff --git a/paddle/fluid/operators/tensorrt/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt/tensorrt_engine_op.h
index e83247d39e..2ff35c7c6a 100644
--- a/paddle/fluid/operators/tensorrt/tensorrt_engine_op.h
+++ b/paddle/fluid/operators/tensorrt/tensorrt_engine_op.h
@@ -114,9 +114,9 @@ class TensorRTEngineOp : public framework::OperatorBase {
     framework::Executor executor(dev_place);
     auto *block = Attr<framework::BlockDesc *>("sub_block");
     auto *program = block->Program();
-    auto *scope_ptr = const_cast<framework::Scope *>(&scope);
+    auto &current_scope = scope.NewScope();
     auto ctx = executor.Prepare(*program, block->ID());
-    executor.RunPreparedContext(ctx.get(), scope_ptr, false, true, true);
+    executor.RunPreparedContext(ctx.get(), &current_scope, false, true, true);
   }
 
   void RunImpl(const framework::Scope &scope,