clone

lite/api/cxx_api.h

@@ -245,7 +245,7 @@ class LITE_API Predictor {
   std::map<TargetType, std::shared_ptr<void>> target_configs_;
   std::shared_ptr<cpp::ProgramDesc> program_desc_;
   std::shared_ptr<Scope> scope_;
-  Scope* exec_scope_;
+  Scope* exec_scope_{nullptr};
   std::shared_ptr<RuntimeProgram> program_;
   bool program_generated_{false};
   std::vector<std::string> input_names_;

lite/core/optimizer/mir/fusion/__xpu__multi_encoder_fuse_pass.cc

@@ -1906,7 +1906,7 @@ class XPUMultiEncoderFuser {
           if (is_qkv_already_fusion_) {
             end = i + 1;
           }
-          scope->NewTensor(update_tag);
+          scope->MutableParent()->NewTensor(update_tag);
           // Update weight, including tranpose\convert type\fuse qkv
           // weight\findmax.
           update_weight(scope,

lite/core/optimizer/mir/type_target_cast_pass.cc

@@ -274,8 +274,15 @@ void TypeTargetTransformPass::AddInputIoCopyInst(
   // So there will be a new Argument node and a new IoCopy Statement Node.
 
   CHECK(in->IsArg());
-  auto io_copy_output_name =
-      string_format("%s/target_trans", in->AsArg().name.c_str());
+  bool in_persist = in->AsArg().is_weight || in->AsArg().is_persist;
+  std::string io_copy_output_name;
+  if (in_persist) {
+    io_copy_output_name =
+        string_format("%s/target_trans_persistable", in->AsArg().name.c_str());
+  } else {
+    io_copy_output_name =
+        string_format("%s/target_trans", in->AsArg().name.c_str());
+  }
 
   if (copied_nodes->count(in->AsArg().name)) {
     // Remove the old link
@@ -292,7 +299,13 @@ void TypeTargetTransformPass::AddInputIoCopyInst(
     // TODO(MyPandaShaoxiang) should set same place with input?
     auto* io_copy_output_arg = graph->NewArgumentNode(io_copy_output_name);
     // Create the new var manually.
-    auto new_var = inst_node->AsStmt().op()->scope()->Var(io_copy_output_name);
+    Variable* new_var = nullptr;
+    if (in_persist) {
+      new_var = inst_node->AsStmt().op()->scope()->MutableParent()->Var(
+          io_copy_output_name);
+    } else {
+      new_var = inst_node->AsStmt().op()->scope()->Var(io_copy_output_name);
+    }
     // Set the place for io_copy_output_arg node, the target should be equal to
     // to.target()
     // The precision and layout should be equal to from.precision(),
@@ -316,7 +329,6 @@ void TypeTargetTransformPass::AddInputIoCopyInst(
     }
     auto* io_copy_inst = graph->NewInstructNode();
 
-    bool in_persist = in->AsArg().is_weight || in->AsArg().is_persist;
     std::string io_copy_type = in_persist ? "io_copy_once" : "io_copy";
     io_copy_output_arg->AsArg().is_persist = in_persist;
     // create Op and kernels.

lite/core/program.cc

@@ -718,7 +718,12 @@ void Program::PrepareWorkspace(
       // Create tensors or weights from variable description.
       if (!var_desc->Persistable()) {
         vars_.push_back(var_name);
-        auto* var = exec_scope_->Var(var_name);
+        Variable* var = nullptr;
+        if (var_name.find("/target_trans_persistable") != std::string::npos) {
+          var = scope_->Var(var_name);
+        } else {
+          var = exec_scope_->Var(var_name);
+        }
         if (var_type == lite::VarDescAPI::Type::LOD_TENSOR) {
           const auto& var_data_type =
               VarDescType2PrecisionType(var_desc->GetDataType());