diff --git a/mindspore/ccsrc/device/kernel_runtime.cc b/mindspore/ccsrc/device/kernel_runtime.cc
index 7f3d31d8d08008fb876ae6dd062b069fe947abf2..d1a068b5841ddf241d74efcf765fe633f7c9e435 100644
--- a/mindspore/ccsrc/device/kernel_runtime.cc
+++ b/mindspore/ccsrc/device/kernel_runtime.cc
@@ -135,10 +135,11 @@ void KernelRuntime::AssignMemory(session::KernelGraph *graph) {
 }
 
 void KernelRuntime::RunOpAssignMemory(const std::vector<tensor::TensorPtr> &input_tensors,
-                                      const session::KernelGraph *graph) {
+                                      session::KernelGraph *graph) {
   MS_EXCEPTION_IF_NULL(graph);
   // assign memory for input nodes
   RunOpAssignInputMemory(input_tensors, graph);
+  AssignStaticMemoryValueNode(graph);
   for (const auto &cnode : graph->execution_order()) {
     // assign memory for output nodes
     RunOpAssignOutputMemory(cnode);
diff --git a/mindspore/ccsrc/device/kernel_runtime.h b/mindspore/ccsrc/device/kernel_runtime.h
index 61b43fd5c0a5fa1387941911baae16696054620e..8f4f769f551c15174cd7a4d925919c570ae46e59 100644
--- a/mindspore/ccsrc/device/kernel_runtime.h
+++ b/mindspore/ccsrc/device/kernel_runtime.h
@@ -46,7 +46,7 @@ class KernelRuntime {
   virtual ~KernelRuntime();
   virtual bool Init() = 0;
   virtual void AssignMemory(session::KernelGraph *graph);
-  void RunOpAssignMemory(const std::vector<tensor::TensorPtr> &input_tensors, const session::KernelGraph *graph);
+  void RunOpAssignMemory(const std::vector<tensor::TensorPtr> &input_tensors, session::KernelGraph *graph);
   virtual bool Run(session::KernelGraph *graph);
   virtual bool DumpData(session::KernelGraph *graph);
   virtual bool RunTask(const session::KernelGraph *graph);
diff --git a/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc b/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc
index 0de609f44131b345edd4cbff6a4bb72ff1a0459e..496a9b276febe63daba27d6d9e79fb014c079f82 100644
--- a/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc
+++ b/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc
@@ -222,6 +222,7 @@ void RunOpAscendBackendIRFusionOptimization(const std::shared_ptr<session::Kerne
   auto optimizer = std::make_shared<GraphOptimizer>();
   auto ir_fusion_pm = std::make_shared<PassManager>("ir_fusion_pm");
   ir_fusion_pm->AddPass(std::make_shared<BnSplit>());
+  ir_fusion_pm->AddPass(std::make_shared<TopKSplit>());
 
   optimizer->AddPassManager(ir_fusion_pm);
   (void)optimizer->Optimize(kernel_graph);
diff --git a/mindspore/ops/_op_impl/tbe/assign_add.py b/mindspore/ops/_op_impl/tbe/assign_add.py
index fbbb9a997f519fe56409571b184bafb0134af603..2b20a7781dd97818329752b19c27b2bd21a36f14 100644
--- a/mindspore/ops/_op_impl/tbe/assign_add.py
+++ b/mindspore/ops/_op_impl/tbe/assign_add.py
@@ -25,7 +25,7 @@ assign_add_op_info = TBERegOp("AssignAdd") \
     .partial_flag(True) \
     .input(0, "ref", False, "required", "all") \
     .input(1, "value", False, "required", "all") \
-    .output(0, "output_ref", False, "required", "all") \
+    .output(0, "ref", False, "required", "all") \
     .dtype_format(DataType.I8_Default, DataType.I8_Default, DataType.I8_Default) \
     .dtype_format(DataType.I8_5HD, DataType.I8_5HD, DataType.I8_5HD) \
     .dtype_format(DataType.U8_Default, DataType.U8_Default, DataType.U8_Default) \
diff --git a/mindspore/ops/operations/debug_ops.py b/mindspore/ops/operations/debug_ops.py
index 1d8fdedc260404d653d4ebbe3140dade5b16176d..97fa883bac2ced5285dce665b819c98caf17e210 100644
--- a/mindspore/ops/operations/debug_ops.py
+++ b/mindspore/ops/operations/debug_ops.py
@@ -210,6 +210,10 @@ class Print(PrimitiveWithInfer):
     def __init__(self):
         pass
 
+    def __call__(self, *args):
+        for arg in args:
+            print(arg)
+
     def infer_shape(self, *inputs):
         return [1]