diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h
index a7d0fd4856edc74237151c64f286d468ad86e7ca..e20f417d76454ec9ba73dbe48506e7426ba1b79d 100644
--- a/paddle/fluid/operators/lookup_table_op.h
+++ b/paddle/fluid/operators/lookup_table_op.h
@@ -129,6 +129,7 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
           "must be either LoDTensor or SelectedRows");
     }
 
+    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
     bool is_sparse = context.Attr<bool>("is_sparse");
     // Since paddings are not trainable and fixed in forward, the gradient of
     // paddings makes no sense and we don't deal with it in backward.
@@ -187,10 +188,12 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
       memset(d_table_data, 0, d_table->numel() * sizeof(T));
 
       for (int64_t i = 0; i < ids->numel(); ++i) {
-        PADDLE_ENFORCE_LT(ids_data[i], N);
-        PADDLE_ENFORCE_GE(ids_data[i], 0);
-        for (int j = 0; j < D; ++j) {
-          d_table_data[ids_data[i] * D + j] += d_output_data[i * D + j];
+        if (ids_data[i != padding_idx) {
+          PADDLE_ENFORCE_LT(ids_data[i], N);
+          PADDLE_ENFORCE_GE(ids_data[i], 0);
+          for (int j = 0; j < D; ++j) {
+            d_table_data[ids_data[i] * D + j] += d_output_data[i * D + j];
+          }
         }
       }
     }