diff --git a/paddle/fluid/operators/tensor_array_to_tensor_op.cc b/paddle/fluid/operators/tensor_array_to_tensor_op.cc
index 620231eb2e298480665cf4eec316f034e0cf1d1c..eb20e1c2cd2748a5ab4db28df0c4798837c7bf21 100644
--- a/paddle/fluid/operators/tensor_array_to_tensor_op.cc
+++ b/paddle/fluid/operators/tensor_array_to_tensor_op.cc
@@ -250,8 +250,12 @@ class LoDTensorArray2TensorGradOp : public framework::OperatorBase {
     auto dout_name = Input(framework::GradVarName("Out"));
 
     std::vector<std::string> grad_names;
+    // NOTE(Aurelius84): Generating grad base name by Input("X") instead of
+    // fixed string to avoid incorrectly sharing same var's allocation in
+    // multi-thread that will cause wrong calculation result.
+    std::string grad_base_name = base_name + "_temp_grad_";
 
-    LodTensorVectorResizeFromLodTensorArray(scope, "grad_name", Input("X"),
+    LodTensorVectorResizeFromLodTensorArray(scope, grad_base_name, Input("X"),
                                             &grad_names);
 
     auto use_stack = Attr<bool>("use_stack");