diff --git a/paddle/fluid/operators/layer_norm_op.h b/paddle/fluid/operators/layer_norm_op.h
index ff2935b0b450f5e42c77e5bd06396b4ab21fb8ca..5907d1d727835f2a19c025f2db3ba12f5fd87ea7 100644
--- a/paddle/fluid/operators/layer_norm_op.h
+++ b/paddle/fluid/operators/layer_norm_op.h
@@ -210,17 +210,35 @@ class LayerNormKernel : public framework::OpKernel<T> {
           ctx, &out, bias, /*axis*/ 1, AddFunctor<T>(), &out);
     }
 #else
-    PADDLE_ENFORCE_EQ(mean->numel(), left);
-    PADDLE_ENFORCE_EQ(var->numel(), left);
-    PADDLE_ENFORCE_EQ(scale->numel(), right);
-    PADDLE_ENFORCE_EQ(bias->numel(), right);
+    PADDLE_ENFORCE_EQ(mean->numel(), left,
+                      platform::errors::InvalidArgument(
+                          "mean's length (%d) is not equal with expected (%d).",
+                          mean->numel(), left));
+    PADDLE_ENFORCE_EQ(var->numel(), left,
+                      platform::errors::InvalidArgument(
+                          "var's length (%d) is not equal with expected (%d).",
+                          var->numel(), left));
+    if (scale) {
+      PADDLE_ENFORCE_EQ(
+          scale->numel(), right,
+          platform::errors::InvalidArgument(
+              "scale's length (%d) is not equal with expected (%d).",
+              scale->numel(), right));
+    }
+    if (bias) {
+      PADDLE_ENFORCE_EQ(
+          bias->numel(), right,
+          platform::errors::InvalidArgument(
+              "bias's length (%d) is not equal with expected (%d).",
+              bias->numel(), right));
+    }
 
     auto ker =
         jit::KernelFuncs<jit::LayerNormTuple<T>, platform::CPUPlace>::Cache()
             .At(right);
     ker(x.data<T>(), out.data<T>(), mean->data<T>(), var->data<T>(),
-        scale->data<T>(), bias->data<T>(), static_cast<int>(left),
-        static_cast<const float>(epsilon), right);
+        scale ? scale->data<T>() : nullptr, bias ? bias->data<T>() : nullptr,
+        static_cast<int>(left), static_cast<const float>(epsilon), right);
 #endif
   }
 };