fix(mgb/opr): fix shape cache of NvOF

GitOrigin-RevId: 456ba478e9

src/opr/impl/misc.cpp

@@ -215,8 +215,9 @@ SymbolVar NvOf::make(
 
 void NvOf::scn_do_execute() {
     auto input_shape = this->input()[0]->shape();
+    std::vector<size_t> t_shape;
     for (size_t i = 0; i < 5; i++) {
-        vshape.push_back(input_shape[i]);
+        t_shape.push_back(input_shape[i]);
     }
     auto c = this->comp_node();
     //! comp_node may init on CUDA or CPU, eg: lar with --cpu
@@ -232,7 +233,8 @@ void NvOf::scn_do_execute() {
     //! create NvOF engine at same device id of comp_node, can not get
     //! comp_node device id, when NvOf:NvOf, so init at scn_do_execute
     std::lock_guard<std::mutex> lock(m_lock);
-    if (init_flag == false) {
+    if (init_flag == false || vshape != t_shape) {
+        vshape = t_shape;
         //! nvof sdk do not imp p2p copy, so init nvof engine on the same
         //! device with mgb comp_node
         nv_flow_extractor = std::make_shared<NVFlowExtractor>(