[XPU] fix bug of AnalyseOpFuncType about xpu op : memcpy_d2d of xpu is actually async (#52042)

paddle/fluid/framework/new_executor/interpreter/data_transfer.cc

@@ -191,7 +191,9 @@ void DataTranferHelper::RunAndConstructOpFuncNode(
         (op_type == kMemcpyD2H ? OpFuncType::kGpuSync : OpFuncType::kGpuAsync);
   } else if (platform::is_xpu_place(place)) {
     // Memcpy in xpu is synchronous
-    new_op_func_node.type_ = OpFuncType::kGpuSync;
+    new_op_func_node.type_ = (op_type == kMemcpyD2H || op_type == kMemcpyH2D)
+                                 ? OpFuncType::kGpuSync
+                                 : OpFuncType::kGpuAsync;
   } else {
     // Memcpy in npu and custom devices is asynchronous
     new_op_func_node.type_ = OpFuncType::kGpuAsync;

paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc

@@ -336,6 +336,8 @@ OpFuncType AnalyseOpFuncType(const OpFuncNode& op_func_node,
   // and so that they would be dispatched to host thread.
   std::shared_ptr<OperatorBase> op = op_func_node.operator_base_;
   if (op->Type() == kCoalesceTensor &&
+      (!platform::is_xpu_place(place) ||
+       op->Attr<bool>("persist_output") == false) &&
       op->Attr<bool>("set_constant") == false &&
       op->Attr<bool>("copy_data") == false) {
     return OpFuncType::kGpuSync;