diff --git a/mindspore/ccsrc/runtime/framework/graph_compiler.cc b/mindspore/ccsrc/runtime/framework/graph_compiler.cc
index 2fe7caf1372..bb9560084c9 100644
--- a/mindspore/ccsrc/runtime/framework/graph_compiler.cc
+++ b/mindspore/ccsrc/runtime/framework/graph_compiler.cc
@@ -379,10 +379,6 @@ GraphId GraphCompiler::CompileGraph(const session::OpRunInfo &op_run_info, const
   MS_EXCEPTION_IF_NULL(device_context);
   device_context->OptimizeSingleOpGraph(graph);
 
-  MS_EXCEPTION_IF_NULL(session_);
-  session_->RunOpHideNopNode(graph);
-  session_->RunOpRemoveNopNode(graph);
-
   // Generate 'KernelMod' for kernel in graph.
   device_context->CreateKernel(graph->execution_order());
 
diff --git a/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc b/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc
index 6d34080a308..98a6821819a 100644
--- a/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc
+++ b/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc
@@ -314,6 +314,22 @@ void RunOpHardwareOptimize(const KernelGraphPtr &kernel_graph) {
   (void)optimizer->Optimize(kernel_graph);
   kernel_graph->SetExecOrderByDefault();
 }
+
+void RunOpHideNopNode(const KernelGraphPtr &kernel_graph) {
+  auto ms_context = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(ms_context);
+  if (!ms_context->get_param<bool>(MS_CTX_ENABLE_PYNATIVE_INFER)) {
+    opt::HideNopNode(kernel_graph.get());
+  }
+}
+
+void RunOpRemoveNopNode(const KernelGraphPtr &kernel_graph) {
+  auto ms_context = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(ms_context);
+  if (!ms_context->get_param<bool>(MS_CTX_ENABLE_PYNATIVE_INFER)) {
+    opt::RemoveNopNode(kernel_graph.get());
+  }
+}
 }  // namespace
 
 void GPUDeviceContext::OptimizeSingleOpGraph(const KernelGraphPtr &graph) const {
@@ -324,6 +340,9 @@ void GPUDeviceContext::OptimizeSingleOpGraph(const KernelGraphPtr &graph) const
   SetOperatorInfo(graph->execution_order());
 
   RunOpHardwareOptimize(graph);
+
+  RunOpHideNopNode(graph);
+  RunOpRemoveNopNode(graph);
 }
 
 void GPUDeviceContext::SetOperatorInfo(const std::vector<CNodePtr> &nodes) const {