diff --git a/paddle/fluid/operators/expand_op.cc b/paddle/fluid/operators/expand_op.cc
index 57f504f980df2233e5f11541234ee4c2dd25b59d..526c053b5b501dc20f58e0cd3376fe6c08006f14 100644
--- a/paddle/fluid/operators/expand_op.cc
+++ b/paddle/fluid/operators/expand_op.cc
@@ -114,7 +114,12 @@ class ExpandGradOp : public framework::OperatorWithKernel {
         ctx->Attrs().Get<std::vector<int>>("expand_times");
     auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
-    for (size_t i = 0; i < expand_times.size(); ++i) {
+    size_t start_pos = 0u;
+    if (!ctx->IsRuntime()) {
+      start_pos = 1u;
+    }
+
+    for (size_t i = start_pos; i < expand_times.size(); ++i) {
       PADDLE_ENFORCE_EQ(x_dims[i] * expand_times[i], out_dims[i],
                         "Each dimension size of Input(Out@GRAD) should be "
                         "equal to multiplication of crroresponding dimension "