diff --git a/onnxruntime/core/optimizer/layer_norm_fusion.cc b/onnxruntime/core/optimizer/layer_norm_fusion.cc
index 4b30a42e31..a4c543df56 100644
--- a/onnxruntime/core/optimizer/layer_norm_fusion.cc
+++ b/onnxruntime/core/optimizer/layer_norm_fusion.cc
@@ -218,25 +218,36 @@ Status LayerNormFusion::ApplyImpl(Graph& graph, bool& modified, int graph_level)
     }
     nodes_to_remove.push_back(last_add_node);
 
-    // Get the inputs for the new LayerNormalization node
+    // Get the inputs for the new LayerNormalization node.
     NodeArg* scale = nullptr;
     NodeArg* bias = nullptr;
     for (size_t i = 0; i < mul_node.MutableInputDefs().size(); i++) {
       if (graph_utils::NodeArgIsConstant(graph, *(mul_node.MutableInputDefs()[i])) || 
           graph_utils::IsGraphInput(graph, mul_node.MutableInputDefs()[i])) {
-        scale = mul_node.MutableInputDefs()[i];
+        // Scale must be 1d.
+        if (mul_node.MutableInputDefs()[i]->Shape()->dim_size() == 1) {
+          scale = mul_node.MutableInputDefs()[i];
+        }
       }
     }
 
     for (size_t i = 0; i < last_add_node.MutableInputDefs().size(); i++) {
       if (graph_utils::NodeArgIsConstant(graph, *(last_add_node.MutableInputDefs()[i])) ||
           graph_utils::IsGraphInput(graph, last_add_node.MutableInputDefs()[i])) {
-        bias = mul_node.MutableInputDefs()[i];
+        // Bias must be 1d.
+        if (last_add_node.MutableInputDefs()[i]->Shape()->dim_size() == 1) {
+          bias = last_add_node.MutableInputDefs()[i];
+        }
       }
     }
     if (scale == nullptr || bias == nullptr) {
       continue;
     }
+
+    // Scale and bias must have the same dimension. 
+    if (scale->Shape()->dim(0).dim_value() != bias->Shape()->dim(0).dim_value()) {
+      continue;
+    }
     const std::vector<NodeArg*> layer_norm_input_defs{reduce_mean_node.MutableInputDefs()[0], scale, bias};
     Node& layer_norm_node = graph.AddNode(graph.GenerateNodeName("LayerNormalization"),
                                           "LayerNormalization",
diff --git a/onnxruntime/test/optimizer/graph_transform_test.cc b/onnxruntime/test/optimizer/graph_transform_test.cc
index 39e07d43f1..81e19a1751 100644
--- a/onnxruntime/test/optimizer/graph_transform_test.cc
+++ b/onnxruntime/test/optimizer/graph_transform_test.cc
@@ -839,6 +839,21 @@ TEST(GraphTransformationTests, LayerNormFusionTest) {
   ASSERT_TRUE(op_to_count["Pow"] == 0);
   ASSERT_TRUE(op_to_count["Sqrt"] == 0);
   ASSERT_TRUE(op_to_count["LayerNormalization"] == 1);
+
+  for (const Node& node : graph.Nodes()) {
+    if (node.OpType() == "LayerNormalization") {
+      // LayerNormalization should have three inputs.
+      EXPECT_EQ(node.InputDefs().size(), 3) << "LayerNormalization number of inputs does not equal to 3. Got:" << node.InputDefs().size();
+      // LayerNormalization input "scale" and "bias" should have the same dimension.
+      const TensorShapeProto* scale_shape = node.InputDefs()[1]->Shape();
+      const TensorShapeProto* bias_shape = node.InputDefs()[2]->Shape();
+      EXPECT_EQ(scale_shape->dim_size(), 1) << "LayerNormalization scale should be 1D. Got: " << scale_shape->dim_size();
+      EXPECT_EQ(bias_shape->dim_size(), 1) << "LayerNormalization bias should be 1D. Got: " << bias_shape->dim_size();
+      EXPECT_EQ(scale_shape->dim(0).dim_value(), bias_shape->dim(0).dim_value());
+    } else {
+      EXPECT_TRUE(false) << "Unexpected node " << node.Name();
+    }
+  }
 }
 
 TEST(GraphTransformationTests, LayerNormWithSubDupFusionTest) {
@@ -860,6 +875,21 @@ TEST(GraphTransformationTests, LayerNormWithSubDupFusionTest) {
   ASSERT_TRUE(op_to_count["Pow"] == 0);
   ASSERT_TRUE(op_to_count["Sqrt"] == 0);
   ASSERT_TRUE(op_to_count["LayerNormalization"] == 1);
+
+  for (const Node& node : graph.Nodes()) {
+    if (node.OpType() == "LayerNormalization") {
+      // LayerNormalization should have three inputs.
+      EXPECT_EQ(node.InputDefs().size(), 3) << "LayerNormalization number of inputs does not equal to 3. Got:" << node.InputDefs().size();
+      // LayerNormalization input "scale" and "bias" should have the same dimension. 
+      const TensorShapeProto* scale_shape = node.InputDefs()[1]->Shape();
+      const TensorShapeProto* bias_shape = node.InputDefs()[2]->Shape();
+      EXPECT_EQ(scale_shape->dim_size(), 1) << "LayerNormalization scale should be 1D. Got: " << scale_shape->dim_size();
+      EXPECT_EQ(bias_shape->dim_size(), 1) << "LayerNormalization bias should be 1D. Got: " << bias_shape->dim_size();
+      EXPECT_EQ(scale_shape->dim(0).dim_value(), bias_shape->dim(0).dim_value());
+    } else {
+      EXPECT_TRUE(false) << "Unexpected node " << node.Name();
+    }
+  }
 }
 #endif