fix(core): fix output var replaced by optpass

GitOrigin-RevId: aea62de345

lite/test/test_network.cpp

@@ -510,7 +510,10 @@ void test_io_no_copy_ax(std::string model_name, int record = 1) {
     std::vector<std::vector<std::shared_ptr<Tensor>>> inputs;
     std::vector<std::vector<std::shared_ptr<Tensor>>> outputs;
 
-    std::shared_ptr<Network> network = std::make_shared<Network>();
+    Config config;
+
+    config.options.graph_opt_level = 0;
+    std::shared_ptr<Network> network = std::make_shared<Network>(config);
     network->load_model(model_path);
 
     input_names = network->get_all_input_name();
@@ -559,10 +562,10 @@ void test_io_no_copy_ax(std::string model_name, int record = 1) {
         outputs.push_back(net_outputs);
     }
 
-    Config config;
     config.options.force_output_use_user_specified_memory = true;
     config.options.comp_node_seq_record_level = record;
     config.options.const_shape = true;
+    config.options.graph_opt_level = 2;
 
     std::shared_ptr<Network> network_record = std::make_shared<Network>(config);
 

src/serialization/impl/serializer.cpp

@@ -10,6 +10,7 @@
  */
 
 #include "megbrain/serialization/serializer.h"
+#include "megbrain/gopt/inference.h"
 #include "megbrain/opr/utility.h"
 
 namespace mgb {
@@ -27,6 +28,35 @@ std::unique_ptr<cg::AsyncExecutable> GraphLoader::LoadResult::graph_compile(
     return ret;
 }
 
+void GraphLoader::LoadResult::graph_compile_ahead() {
+    //! when force_output_use_user_specified_memory is set, the output var may
+    //! be changed by gopt, then the var in LoadResult can not exist, so here
+    //! just do basic optimize_for_inference ahead, and replace the var in
+    //! LoadResult
+    if (graph->options().force_output_use_user_specified_memory) {
+        auto options = gopt::OptimizeForInferenceOptions{};
+        auto new_vars = gopt::optimize_for_inference(output_var_list, options);
+        output_var_list = new_vars;
+        output_var_map.clear();
+        for (auto& var : new_vars) {
+            output_var_map[var.node()->cname()] = var;
+        }
+        std::unordered_map<size_t, SymbolVar> var_map_id;
+        for (auto& var : new_vars) {
+            bool found = false;
+            for (auto& old_var_it : output_var_map_id) {
+                if (old_var_it.second.node()->name() == var.node()->name()) {
+                    found = true;
+                    var_map_id[old_var_it.first] = var;
+                }
+            }
+            mgb_assert(
+                    found, "can't find var name %s when optimize_for_inference. ",
+                    var.node()->cname());
+        }
+    }
+}
+
 GraphLoader::SharedTensorNameMap GraphLoader::shared_tensor_name_map() {
     SharedTensorNameMap ret;
     for (auto&& i : shared_tensor_id_map()) {

src/serialization/impl/serializer_oss.cpp

@@ -946,6 +946,7 @@ GraphLoader::LoadResult GraphLoaderOSS::load(const LoadConfig& config, bool rewi
     mgb_assert(fbs_end > cur);
     // Skip to Graph end
     m_file->skip(fbs_end - cur);
+    result.graph_compile_ahead();
     return result;
 }
 

src/serialization/include/megbrain/serialization/serializer.h

@@ -63,6 +63,14 @@ public:
          */
         MGE_WIN_DECLSPEC_FUC std::unique_ptr<cg::AsyncExecutable> graph_compile(
                 const ComputingGraph::OutputSpec& outspec);
+
+        /*!
+         * \brief after graph is loaded, do some basic optimized_for_inference,
+         * because some dest var maybe replaced, case error when optimize flag
+         * force_output_use_user_specified_memory is on
+         *
+         */
+        MGE_WIN_DECLSPEC_FUC void graph_compile_ahead();
     };
 
     //! helper to disable inplace arith graph optimization during