diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index 991476dff55b9a4c7e8ff3c1093e18e0b3380fc1..eb7ad8ed94ec95e2ef8fabd85d816d949aa0953d 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -1603,11 +1603,11 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   }
 #endif
 
-  auto exe_ctx = ExecutionContext(*this, scope, *dev_ctx, *runtime_ctx);
   // using cache
   if (kernel_type_.get()) {
     dev_ctx = pool.Get(kernel_type_->place_);
   }
+  auto exe_ctx = ExecutionContext(*this, scope, *dev_ctx, *runtime_ctx);
 
 // TODO(Liu-xiandong): Now we are using too much if-else and hard code in XPU
 // device, it's ugly, and we will refactor in the future.