[Fix] Illegal memory access in GetOutputIndex with optional outputs

onnxruntime/core/graph/ep_api_types.cc

@@ -351,6 +351,10 @@ static Status GetOutputIndex(const EpNode& producer_node,
   gsl::span<const EpValueInfo* const> outputs = producer_node.GetOutputsSpan();
 
   for (size_t i = 0; i < outputs.size(); i++) {
+    if (outputs[i] == nullptr) {  // outputs == nullptr means the output is optional
+      continue;
+    }
+
     if (outputs[i]->GetName() == value_info_name) {
       index = i;
       found = true;

onnxruntime/test/ep_graph/test_ep_graph.cc

@@ -180,6 +180,13 @@ TEST(EpGraphTest, CheckModelExternalInitializers) {
   CheckGraphCApi(test_graph->GetGraphViewer(), test_graph->GetOrtGraph());
 }
 
+TEST(EpGraphTest, CheckModelOptionalIntermediateNodeOutputs) {
+  auto test_graph = TestGraph::Load(ORT_TSTR("testdata/skip_simplified_layer_normalization.onnx"));
+  ASSERT_NE(test_graph, nullptr) << "Failed to load test model";
+
+  CheckGraphCApi(test_graph->GetGraphViewer(), test_graph->GetOrtGraph());
+}
+
 static void RunConvQDQExtIni(const ORTCHAR_T* model_path, std::vector<float>& output_data) {
   auto memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU);
   Ort::SessionOptions sess_options;

onnxruntime/test/testdata/skip_simplified_layer_normalization.py

@@ -0,0 +1,65 @@
+from onnx import TensorProto, checker, helper, save, shape_inference
+
+batch_size = 1
+seq_len = 64
+hidden_size = 896
+
+input_vi = helper.make_tensor_value_info(
+    name="input",
+    elem_type=TensorProto.FLOAT,
+    shape=[batch_size, seq_len, hidden_size],
+)
+
+skip_vi = helper.make_tensor_value_info(
+    name="skip",
+    elem_type=TensorProto.FLOAT,
+    shape=[batch_size, seq_len, hidden_size],
+)
+
+output_vi = helper.make_tensor_value_info(
+    name="output",
+    elem_type=TensorProto.FLOAT,
+    shape=[batch_size, seq_len, hidden_size],
+)
+
+input_skip_bias_sum_vi = helper.make_tensor_value_info(
+    name="input_skip_bias_sum",
+    elem_type=TensorProto.FLOAT,
+    shape=[batch_size, seq_len, hidden_size],
+)
+
+gamma_init = helper.make_tensor(
+    name="gamma",
+    data_type=TensorProto.FLOAT,
+    dims=[hidden_size],
+    vals=[1] * hidden_size
+)
+
+node = helper.make_node(
+    op_type="SkipSimplifiedLayerNormalization",
+    inputs=["input", "skip", "gamma"],
+    outputs=["output", "", "", "input_skip_bias_sum"],
+    domain="com.microsoft",
+    epsilon=1e-6,
+    name="SkipLayerNorm",
+)
+
+graph = helper.make_graph(
+    nodes=[node],
+    name="SkipSimplifiedLayerNormGraph",
+    inputs=[input_vi, skip_vi],
+    outputs=[output_vi, input_skip_bias_sum_vi],
+    initializer=[gamma_init],
+)
+
+model = helper.make_model(
+    graph,
+    opset_imports=[
+        helper.make_operatorsetid("", 17),
+        helper.make_operatorsetid("com.microsoft", 1),
+    ],
+)
+
+model = shape_inference.infer_shapes(model)
+checker.check_model(model, True)
+save(model, "skip_simplified_layer_normalization.onnx")